Bioreaction Engineering

Springer
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo
K. Schiigerl . K.-H. Bellgardt (Eds.)

Bioreaction
Engineering
Modeling and Control

With 225 Figures and 70 Tables

Springer
Prof. em. Dr. Dr. h. c. Karl Schiigerl
e-mail: [email protected]
Prof. Dr.-Ing. Karl-Heinz Bellgardt
e-mail: [email protected]

University of Hannover
Institute of Technical Chemistry
Callinstrasse 3
D-30167 Hannover
Germany

TSBN-13: 978-3-642-64103-9 Springer-Verlag Berlin Heidelberg New York

Library of Congress Cataloging-in-Publication Data

Bioreaction engineering: modeling and control I K. Schiigerl, K.-H. Bellgardt (eds.)

p.cm.
Includes bibliographical references and index.

ISBN-13: 978-3-642-64103-9 e-ISBN-13: 978-3-642-59735-0

001: 10.1007/978-3-642-59735-0

1. Bioreactors. I. Schiigerl, K. (Karl) II. Bellgardt, K.-H.

TP 248.25.B55 B549 2000
660'.2832-dc21 00-026598

This work is subject to copyright. All rights are reserved, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in other ways, and storage in data
banks. Duplication of this publication or parts thereof is only permitted under the
provisions of the German Copyright Law of September 9, 1965, in its current version, and
permission for use must always be obtained from Springer-Verlag. Violations are liable for
prosecution under the German Copyright Law.

Springer-Verlag Berlin Heidelberg New York

a member of BertelsmannSpringer Science+ Business Media GmbH

Springer-Verlag Berlin Heidelberg 2000

Softcover reprint ofthe hardcover I st edition 2000

The use of general descriptive names, registered names, trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.

Product liability: The publisher cannot guarantee the accuracy of any information about
dosage and application contained in this book. In every individual case the user must check
such information by consulting the relevant literature.

Cover Design: design & production, Heidelberg

Typesetting: MEDIO GmbH, Berlin

Printed on acid-free paper SPIN 10665072 52/3020 - 5 4 3 2 1 0

Preface

The book is intended to present various examples for reactor and process modeling
and control as well as for metabolic flux analysis and metabolic design at an ad-
vanced level.
In Part A, General principles and techniques with regard to reactor and process
models, process control, and metabolic flux analysis are presented. In addition the
accuracy, precision, and reliability of the measured data are discussed which are ex-
tremely important for process modeling and control. A virtual bioreactor system is
presented as well, which can be used for the training of students and operators of
industrial plants and for the development of advanced automation tools.
In Part B, the General principles are applied for particular bioreactor models. It
covers the application of the computational fluiddynamic (CFD) technique to stirred
tank and bubble column bioreactors. Different solution methods are presented: the
Reynolds-averaging of the turbulent Navier-Stokes equations and modeling of the
Reynolds stresses with an appropriate turbulence (k-ee) model, and the Euler (two
fluid model), as well as the Euler-Langrange approaches.
In Part C, Application of general principles for process models including control are
discussed with regard to the production of baker's yeast, beer, lactic acid, recombi-
nant protein, and ~-lactam antibiotics. Various models and control strategies are
used for baker's yeast production with regard to the optimization of productivity,
product quality and process economy. Hybrid models are applied for the state pre-
diction, optimization, and process supervision for beer production. Kinetic models
are presented for the production of lactic acid which take into account the various
metabolic paths of lactic acid bacteria. A kinetic model is developed for high density
cultivation of E. coli and a control strategy is applied for the recombinant protein
production. Various models are discussed, which take the fungal morphology into
account and are used for optimization of the production of ~-lactam antibiotics.
In Part D, examples are given for Metabolite flux analysis and metabolic design.
They discuss the metabolic balancing, isotopic labeling combined with NMR spec-
troscopy methods, and discuss their application to the flux analysis in Saccharo-
myces cerevisiae, Zymomonas mobilis, Corynebacterium glutamicum, and mamma-
lian cells.

July 2000 Karl Schugerl and Karl-Heinz Bellgardt

Contents

List of Abbrevations and Symbols XVIII

Introduction
Karl-Heinz Bellgardt
The Need for Modeling and Control in Biotechnical Processes
Some Modeling Basics . . . . . . . . . . . . . . . . . 3
Structure and Operation of Biotechnical Plant . . . . . . . . . 5
Types and Structure Elements of the Bioreactor . . . . . . 8
The Stirred Tank Reactor as an Example for Reactors with
Mechanical Energy Input; Reactors with Energy Input by
Compressed Air; Membrane Reactors for Bubble Free Aeration;
Liquid-Phase; Gas-Phase; Solid-Phase; Biotic Phase; Modes
of Operation of a Bioreactor; Batch Cultivation; Fed-Batch
Cultivation; Continuous Cultivation; Cultivation with Cell
Retention; Repeated or Cyclic Batch or Fed-Batch Cultivation;
Aerobic Processes; Anaerobic Processes; Micro-Aerobic Processes;
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Part A General Principles and Techniques

Bioreactor Models
Karl Schugerl, Karl-Heinz Bellgardt 21
1.1 Introduction . . . . . . . . . . . . . 21
1.2 Interrelations Between the Cells and Their Physical!
Chemical Environment . . . . . . . . . . . . . . . . 21
l.3 Stirred Tank (ST) Reactors . . . . . . . . . . . . . . 22
l.3.1 Description of the Physical Processes in the Reactors 22
l.3.2 Reactor Models . . . . . . . . . . . . . . . . . . . . . 23
l.3.2.1 Model for the Ideal Stirred Tank Reactor. . . . . . . 23
1.4 Bubble Column (BC) and Airlift Tower Loop (ATL) Reactors. 34
l.4.1 Description of the Physical Processes in the Reactors . . . . . 34
Contents VII

1.4.2 Flow Models . . 35

1.4.3 Reactor Models 36
1.5 Conclusions 41
References . 41

2 Bioprocess Models
Karl-Heinz Bellgardt 44
2.1 Introduction . . . . . . . . . . . 44
2.1.1 Intracellular Structure Elements 44
2.1.2 Regulation of the Metabolism 46
2.1.2.1 Bottle-Neck Principle . . . . . . 46
2.1.2.2 Optimality Principle . . . . . . 47
2.1.3 Kinetics of Growth and Product Formation 47
2.1.4 General Model Structure for Biotechnical Processes 51
2.1.5 Transport in Microbial Aggregates . . . . 54
2.2 Unstructured Models . . . . . . . . . . . . 56
2.2.1 Kinetics of Growth and Substrate Uptake . 57
2.2.2 Endogenous and Maintenance Metabolism 66
2.2.3 Product Formation . . . . . . . . . . 67
2.2.4 Other Parameters Influencing Growth 70
2.3 Structured Models . . . . . . . . . . . 73
2.3.1 The Constitutive Equations . . . . . . 74
2.3.2 Some Applications of Structured Models 77
2.3.3 Cybernetic Models of the Compartment Type . 81
2.3.4 Cybernetic Models of the Metabolic Regulator Type 83
2.4 Segregated Models . . . . . . . . . . . . . . . . 86
2.4.1 Simple Segregated Models . . . . . . . . . . . . . . 86
2.4.2 Segregated Models for Physiological Properties . . . 87
2.4.3 A Model for Spatial Segregation by Wall Attachment 90
2.4.4 Segregated Models for Morphological Differentiation,
Morphologically Structured Models . . . . . . . 91
2.4.5 Segregated Models for Recombinant Organisms . 95
2.4.6 Population Balance Models 100
References . . . . . . 102

3 Metabolic Flux Analysis

Maria I. Klapa, Gregory Stephanopoulos 106
3.1 Introduction . . . . . . . . . 106
3.2 Flux Quantification Methods 108
3.2.1 Metabolite Balancing .. . . 108
3.2.2 Isotopic-Tracer Techniques . 112
3.3 Applications of Metabolic Flux Analysis in the Elucidation
of Metabolic Networks . . . . . . . . . . . . . . . . . . . . . . . . . 118
VIII Contents

3.4 Conclusions 121

References 122

4 Accuracy and Reliability of Measured Data

Bernd Hitzmann . . . . . . . . . . . . . 125
4.1 Accuracy and Reliability of Measured Data 125
4.1.1 Accuracy and Precision of Measurements 126
4.1.2 Accuracy ........ . 126
4.1.3 Precision ................. . 128
4.2 Measurement Reliability . . . . . . . . . . 132
4.2.1 Assessment of Measured Data Reliability by Means
of a Knowledge-Based System . . . . . . . . 133
4.2.2 Numerical and Statistical Tests Performed
by the Knowledge-Based System . . . . . . 135
4.2.3 Knowledge-Based Module . . . . . . . . . . 139
4.2.4 Methodology of the Knowledge-Based System 141
4.3 Conclusions 142
References 143

s Bioprocess Control
D. Dochain, M. Perrier 145
5.1 Introduction . . . . . . 145
5.2 Bioprocess Control: Basic Concepts 146
5.2.1 Disturbances . . . . 147
5.2.2 Stability . . . . . . . 148
5.2.2.1 Equilibrium Points . 149
5.2.2.2 Stability Analysis 149
5.2.3 Regulation vs Tracking 150
5.3 Bioprocess Control: Basic Ingredients 150
5.3.1 Dynamical Model 150
5.3.2 Feedback . . . . . . 151
5.3.3 Proportional Action 151
5.3.4 Integral Action . . . 151
5.3.5 Feedforward Action 151
5.3.6 Linear Control vs Nonlinear Control 152
5.3.6.1 Linear Control . . . . . . . . . . . . 152
5.3.6.2 Nonlinear Control . . . . . . . . . . 153
5.3.7 Adaptive Control vs Non-Adaptive Control 154
5.3.8 Other Approaches . . . . . . . . . . . . . . 156
5.4 Adaptive Linearizing Control of Bioprocesses 156
5.4.1 General Dynamical Model . . . . 157
5.4.1.1 Example 1: Anaerobic Digestion 157
5.4.1.2 Example 2: Animal Cell Culture 158
5.4.2 Model Reduction. . . . . . . . . 159
Contents IX

5.4.2.1 Singular Perturbation Technique for Low Solubility Products 159

5.4.2.2 A General Rule for Order Reduction 159
5.4.2.3 Example 1: Anaerobic Digestion . . . . . . . . . . 160
5.4.3 Control Design . . . . . . . . . . . . . . . . . . . 160
5.4.3.1 The Monitoring Tool 1: An Asymptotic Observer. 161
5.4.3.2 The Monitoring Tool 2: The Parameter Estimation 162
5.4.3.3 The Control Tool: The Adaptive Linearizing Controller 164
5.4.4 Experimental Results 165
References . . . . ... . . . . . . . . . . . . . . . . . . . 166

6 On-Line Simulation Techniques for Bioreactor Control Development

Reiner Luttmann, Klaus-Uwe Gollmer 167
6.1 Introduction . . . . . . . . . . . . . . . . 167
6.2 Application . . . . . . . . . . . . . . . . 168
6.2.1 Application in the Biochemical Industry. 168
6.2.1.1 Plant Set Up 168
6.2.1.2 Economy. 169
6.2.1.3 Quality . . . 169
6.2.1.4 Validation . 169
6.2.1.5 Complexity. 169
6.5.1.6 Training .. 170
6.2.2 Application in Education 170
6.3 General Architecture of On-Line Simulation Systems 170
6.3.1 Components of Simulation Systems 172
6.3.1.1 Models . . . . . . . 172
6.3.1.2 Numerical Methods . . . . . . . . . 172
6.3.1.3 User Interface . . . . . . . . . . . . 173
6.4 Full Scope Model of the Fermentation Process 173
6.5 Submodels of the Bioreactor Process . 175
6.5.1 Engineering Components . . 175
6.5.1.1 Temperature Control System 175
6.5.1.2 Pressure Behavior . . . . . . 179
6.5.1.3 Aeration Behavior . . . . . . 180
6.6 Mass Balances of the Complete Aerobic Growth Process 180
6.6.1 Gas Phase Balances . . . . . . . . . . 181
6.6.2 The Oz- and COz- Transfer Equations 183
6.6.3 The kLa Correlation . . . . . . . . .. 184
6.6.4 The Liquid Phase Balances . . . . . . 185
6.6.5 The Feed and Titration Vessels System 187
6.7 The pH Model . . . . 188
6.8 The Reaction Model . 189
6.9 Application Examples of On-Line Simulation Techniques 194
6.9.1 Training with Virtual Reaction Processes . . . . 194
6.9.2 Development of a High Cell Density Cultivation . . . . . 195
x Contents

6.9.2.1 The JL-Stat Problem . . . . . . . . . . . . . 195

6.9.2.2 Observation of Cell-Specific Growth Rate . 198
6.9.2.3 Course and Testing of Processing Strategies 199
6.10 Summary. 202
References 202

Part B Application of General Principles for Reactor Models

7 Application of Computational Fluiddynamics (CFD)

to Modeling Stirred Tank Bioreactors
Matthias Reuss, Sven Schmalzriedt, Marc Jenne 207
7.1 Introduction . . . . . . . . . . . . . . . . . . 207
7.2 Modeling and Simulation of Gas/Liquid Flow
in Stirred Tank Reactors 210
7.3 Single Phase Flow . . . . . . . . . . . . . . . 213
7.3.1 Transport Equations . . . . . . . . . . . . . . 213
7.3.2 Simulations and Comparison with Experimental Observations . 217
7.4 Multiple Impellers 223
7.5 Gas-Liquid Flow . 225
7.5.1 Interfacial Forces . 225
7.5.1.1 Drag Force . . . . 226
7.5.1.2 Virtual Mass Force . 227
7.5.2 Turbulence Model 227
7.5.3 Impeller Model . . . 228
7.5.4 Simulation Results . 229
7.6 Application of CFD to Simulations of Mixing
and Biotechnical Processes . . . . 232
7.6.1 Methodology . . . . . . . . . . . . . . . . . . 232
7.6.2 Simulation of Tracer Experiments . . . . . . 235
7.6.3 Simulation of Substrate Distributions in Fed Batch Fermentations 239
7.6.4 Production of Acetoin/Butanediol with Bacillus subtilis . 241
References . . . . . . . . 244

8 Bubble Column Bioreactors

Andreas Lubbert . 247
8.1 Introduction . . 247
8.2 Phenomenology 248
8.3 Basic Equations of Motion 251
8.3.1 Fundamental Laws of Fluid Motion . 251
8.3.1.1 Mass Conservation 252
8.3.1.2 Conservation of Momentum . . . . 252
8.3.1.3 Navier-Stokes Equation System .. . 252
8.3.1.4 Problems with Solving the Equations of Motion 253
Contents XI

8.3.1.5 Numerical Aspects 256

8.3.2 Two-Fluid Model . 257
8.3.3 Euler-Lagrange Approach. 258
8.3.3.1 Dynamics of the Dispersed Gas-Phase 258
8.3.3.2 Effective Viscosity . . . . . . . . . . . 259
8.3.3.3 Mass Transfer and Chemical Reaction . 259
8.3.3.4 Mixing Due to the Bubble Rise . . . . . 260
8.3.3.5 Problem of Bubble Coalescence and Redispersion 262
8.3.3.6 Rating of the Euler-Lagrange Representation . . . 262
8.4 Modeling of Particular Aspects of Bubble Column Reactors 264
8.4.1 Velocity Patterns in Bubble Column Reactors . . . . . .. 264
8.4.2 Fate of Individual Cells in the Bubble Column Bioreactor 265
8.4.3 Influence of Tilted Columns . . . . . . . . 267
8.4.4 Oxygen Distribution in a Yeast Fermenter . 267
8.5 Conclusions 272

References . 273

Part ( Application of General Principles for Process Models Including (ontrol

9 Baker'S Yeast Production

Karl-Heinz Bellgardt 277
9.1 Introduction . . . . . . 277
9.1.1 Metabolic Types of Yeast Growth and Regulatory Effects. 278
9.1.2 The Asymmetric Propagation of Yeast 279
9.2 Growth Modeling . . . . . . . . . . . 283
9.2.1 Stoichiometric Model . . . . . . . . . 284
9.2.2 Cybernetic Modeling of Metabolic Regulation. 287
9.2.3 Application of the Model for Simulation of Batch, Fed-Batch,
and Continuous Cultivations . . . . . . 291
9.3 Growth in Airlift Tower-Loop Reactors ............ 293
9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 296
9.4.1 Age Distribution Model of Yeast for Batch
and Fed-Batch Processes . . . . . . . . . . . . . . . . . . . . . . 297
9.4.2 Age Distribution Model for Data Analysis of Stable Synchronous
Oscillations in a Chemostat . . . . . . . . 303
9.5 Considerations for Process Optimization 307
9.5.1 Optimization of Product Quality. . . . . 308
9.5.2 Economic Optimization. . . . . . . . . . 311
9.6 Automatic Control of Fed-Batch Processes 313
9.6.1 General Remarks . . . . . . . . . . . . 313
9.6.2 Examples for Applied Control Systems 315
References . . . . . . . . . . . . . . . . 319
XII Contents

10 Modeling of the Beer Fermentation Process

Andreas Liibbert . 321
10.1 Introduction . . . 321
10.2 Process Optimization 323
10.2.1 Different Knowledge Representation Techniques 323
10.2.1.1 Classical Approach . . . . . . . . . . . . . . 324
10.2.1.2 Heuristic Approach . . . . . . . . . . . . . . 325
10.2.1.3 Alternative Methods to Describe the Kinetics 326
10.2.2 State Prediction for Process Optimization 327
10.2.3 Remarks on Hybrid Models . . . . . . . . . . 331
10.3 Process Supervision . . . . . . . . . . . . . . 331
10.3.1 On-Line Measurement are Difficult to Perform 332
10.3.2 Estimation of the Extract Degradation 334
10.3.2.1 Simple Mathematical Model . . . . . . 334
10.3.2.2 Estimation of the Extract Degradation
by Artificial Neural Networks . . . . . . . . . . . . . . . . . . . 335
10.3.2.3 Hybrid Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 337
10.3.3 Kalman Filters, and an Advanced Method for State Estimation 338
10.4 Process Control . . . . . . . . . . . . . . . . . . . . . . . . . . 339
10.4.1 Controllers that Consider the Dynamics of the Fermenters . . . 341
10.4.2 Reduction of Energy Costs by Temperature Profile Optimization
and Control in a Production-Scale Brewery .. 343
10.5 Conclusion . . . . . . . . . . . . . . . . . . . . 345
10.5.1 Summary of the Application of the Techniques
to Beer Fermentation 346
References . . . . . . 347

11 Lactic Acid Production

John Villadsen 349
11.1 Introduction . 349
11.2 Classification of Lactic Bacteria 350
11.3 Sugar Metabolism of LAB . . . . 350
11.3.1 An Example Showing the Functioning of PTS Systems 351
11.3.2 Sugar Uptake by LAB in General . . . . . 354
11.3.3 Homolactic vs Heterolactic Fermentation 355
11.4 Nitrogen Uptake and Metabolism . . . . . 359
11.5 Growth Kinetics and Product Formation Kinetics . 363
11.6 Lactic Acid Production on the Industrial Scale . 368
11.7 Process Technology in Lactic Acid Fermentation 369
References . . . . . . . . . . . . . . . . . . . . . 372
Contents XIII

12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

Clemens Posten, Ursula Rinas 374
12.1 Introduction . . . . . . . . . . . . . . . 374
12.2 Basic Modeling of a Fed-Batch Strategy 374
12.2.1 The Physiological Model .. . . . . . . 375
12.2.2 The Reactor Model . . . . . . . . . . . 377
12.3 Growth Rate Control via Substrate Feeding 378
12.4 Growth Rate Control via Oxygen Supply .. 381
12.5 Considerations for Improved Observation and Control. 386
12.6 A Case Study: Kinetics of Acetate Formation
and Recombinant Protein Synthesis in HCDC . 387
References . . . . . . . . . . . . . . . . . . . . 389

13 ~-Lactam Antibiotics Production with Penicillium chrysogenum and

Acremonium chrysogenum
Karl-Heinz Bellgardt 391
13.1 Introduction . . . . . . . 391
13.2 Modeling of Penicillin Production 395
13.2.1 Unstructured and Simple Segregated Models 395
13.2.2 Biosynthesis Model of Penicillin V . . . . . . 400
13.2.3 Morphologically Structured Models for Growth of Hyphae 402
13.2.4 Models for Growth of Fungal Pellets . . . 406
13.2.5 Models for Growth of Pellet Populations . 409
13.3 Modeling of Cephalosporin C Production 412
13.3.1 Biosynthesis of Cephalosporin . . . . . . 413
13.3.2 Simple Cybernetic Model for Growth and Production
on Sugar and Soy-Oil . . . . . . . . . . . . . . . . . . 415
13.3.3 Segregated Models Describing Morphological Differentiation 416
13.4 Process Control and Optimization . . . . . . . . . . 421
13.4.1 Problems and Possibilities . . . . . . . . . . . . . . 421
13.4.2 Example for Dynamic Optimal Control of Fed-Batch
Antibiotics Production . . . . . . . . . . . . . . . . 424
13.4.3 Economic Optimization for Mycelia Fed-Batch Cultivation. 425
References . . . . . . . . . . . . . . 429

Part D Metabolite Flux Analysis, Metabolic Design

14 Quantitative Analysis of Metabolic and Signaling Pathways

in Saccharomyces cerevisiae
Klaus Mauch, Sam Vaseghi, Matthias Reuss 435
14.1 Introduction . . . . . . . 435
14.2 Metabolic Flux Analysis . 436
XIV Contents

14.2.1 Metabolite Balancing in Compartmented Systems . 437

14.2.2 Stoichiometric Model . 438
14.2.3 Computational Aspects . . . . . . . . . . 439
14.2.4 Results . . . . . . . . . . . . . . . . . . . 440
14.3 Measurement of Intracellular Compounds 445
14.3.1 Measurement of Intracellular Metabolites and Signals -
General Tools . . . . . . . . . . . . . . . . . . . . . . . 445
14.3.2 Dynamic Response of Metabolite Pools of Glycolysis . . 448
14.3.3 Dynamics of the Pentose Phosphate Pathway - an Example
for in vivo Diagnosis of Intracellular Enzyme Kinetics . . . 453
14.4 Quantitative Analysis of Glucose Induced Signal Transduction . 456
14.4.1 Measurement of Intracellular cAMP 459
14.4.2 Measurement of the PFK2 Activity . . . . . . . . . . 460
14.4.3 Measurement of F2,6bP . . . . . . . . . . . . . . .. 461
14.5 Comparison Between in vitro and in vivo Kinetics -
Illustrated for the Enzyme PFKl (Phosphofructokinase 1) 462
References .. . . . . . . . . . . . . . . 475

1S Metabolic Analysis of Zymomonas mobilis

Albert A. de Graaf . 478
15.1 Introduction . . . . . . . . . . . 478
15.1.1 Zymomonas mobilis . . . . . . . 478
15.1.2 Substrate Spectrum Engineering 479
15.1.3 Purpose . . . . . . . . . . . . . . 480
15.2 Methods for Metabolic Analysis 481
15.2.1 Introduction . . . . . . . . . . 481
15.2.2 Metabolite Pool Determination 481
15.2.2.1 Invasive Approaches 481
15.2.2.2 In vivo Techniques . . . 482
15.2.2.3 Rapid Sampling . . . . 484
15.2.3 Metabolic Flux Analysis 485
15.2.3.1 Basic Carbon Balancing 485
15.2.3.2 Metabolite Balancing . 486
15.2.3.3 Stable Isotope Labeling 486
15.2.3.4 NMR Magnetization Transfer 487
15.2.4 Metabolic Modeling . . . . . 488
15.3 Metabolic Analysis of Zymomonas mobilis . 489
15.3.1 Introduction . . . . . . . . . . 489
15.3.2 Enzymatic Studies . . . . . . . 490
15.3.3 Metabolite Pool Measurements 491
15.3.3.1 Overview . . . . . . . . . 491
15.3.3.2 Glycolytic Intermediates . 493
15.3.3.3 Sugars 495
15.3.3.4 Ethanol . . . . 495
15.3.4 Flux Analyses 496
Contents xv

15.3.4.1 Overview . . . . . . . . . . . . . . 496

15.3.4.2 Metabolite Balancing . . . . . . . 496
15.3.4.3 NMR and Stable Isotope Labeling 498
15.3.5 Summary . . . . . . 500
15.4 Concluding Remarks 501
References . . . . . . 501

16 Metabolic Flux Analysis of Corynebacterium glutamicum

Albert A. de Graaf . 506
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 506
16.2 Fundamentals of Intracellular Metabolic Flux Analysis
in Corynebacterium glutamicum 509
16.2.1 Metabolite Balancing . . . . . . 509
16.2.1.1 Biomass Composition . . . . . . 509
16.2.1.2 Condensed Bioreaction Network 513
16.2.1.3 Approaches to Resolve Network Underdeterminacy 514
16.2.1.4 Theoretical Lysine Selectivity. . . . . . . . . . . . . 515
16.2.1.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . 517
16.2.2 Isotopic Labeling Combined with NMR Spectroscopy 518
16.2.2.1 Isotopic Atom Balancing . . . . . . . . . . . . . . . . 518
16.2.2.2 Resolving Glycolysis and Pentose Phosphate Pathway 519
16.2.2.3 Resolving the Parallel Lysine Biosynthetic Pathways . 520
16.2.2.4 Resolving Anaplerosis, Citric Acid Cycle, and the Glyoxylate Shunt 522
16.2.2.5 Resolving the Principal Ammonium-Assimilatory Pathways 523
16.2.2.6 Influence of Reaction Reversibility . . . 525
16.2.2.7 Isotopomers . . . . . . . . . . . . . . . 527
16.2.2.8 Sources of Isotopic Measurement Data 528
16.2.2.9 A Comprehensive Modeling Framework 529
16.3 Metabolite Balancing Studies . . . . . . 530
16.3.1 Overview . . . . . . . . . . . . . . . . . 530
16.3.2 Comparison of Fluxes During Growth and Lysine Production 531
16.3.3 The Search for Yield-Limiting Flux Control Architectures 532
16.3.3.1 The Pyruvate Branch Point . . . . . . . 532
16.3.3.2 The Glucose-6-Phosphate Branch Point 532
16.3.4 Growth Rate-Dependent Modulation
of the Central Metabolic Fluxes 534
16.3.4.1 Growth on Lactate . 534
16.3.4.2 Growth on Glucose . . . . . . . 536
16.3.5 Summary . . . . . . . . . . . . 536
16.4 Studies Based on Isotopic Labeling and NMR . 538
16.4.1 Overview . . . . . . . . . . . . . . . . . . . 538
16.4.2 The Dual Pathways of Lysine Biosynthesis . 539
16.4.2.1 Correlation with Lysine Production . . . . 539
16.4.2.2 Correlation with Culture Parameters . . . . 540
16.4.3 Distinct Metabolic Modes: Growth, Glutamate Production,
and Lysine Production . . . . . . . . . . . . . . . . . . . . 541
XVI Contents

16.4.3.1 Comparing Isogenic Strains in Continuous Cultures 541

16.4.3.2 Comparing Different Strains in Batch Cultures 545
16.4.4 Perturbations of the Redox Metabolism 546
16.4.5 The Ammonium-Assimilating Fluxes. 550
16.4.6 Summary . . . . . . . 551
16.5 Concluding Remarks 552
References . . . . . . 552

17 Analysis of Metabolic Fluxes in Mammalian (ells

Neil S. Forbes, Douglas S. Clark, Harvey W. Blanch . 556
17.1 Applications of Metabolic Flux Analysis in Mammalian Cells 556
17.1.1 Optimization of Protein Production .. . . . . . 557
17.1.2 Metabolic Regulation in Transformed Cells . . . 557
17.1.3 Metabolic Regulation in Non-Transformed Cells 558
17.2 Experimental Techniques . . . . . . . . . . . . . 559
17.2.1 Direct Measurement of Extracellular Production and
Consumption Rates .. . . . . . 560
17.2.1.1 Continuous Suspension Culture . 561
17.2.1.2 Perfused Culture . . . . . . . . . 561
17.2.1.3 Batch Culture . . . . . . . . . . 562
17.2.2 Detection of Isotope Distribution by 13C-NMR 562
17.2.2.1 Measuring Fractional Enrichments 562
17.2.2.2 Measuring Isotopomer Fractions 563
17.2.2.3 In Vivo NMR . . . . . . . . . . . . 564
17.2.2.4 Extraction NMR . . . . . . . . . . 564
17.2.3 Radio-Isotope Tracer Studies and Enzyme Activity Assays 565
17.3 Mathematical Descriptions to Quantify Fluxes
in Metabolic Models . . . . . . . . . . . . . . . . . . . . 565
17.3.1 Determining Fluxes Using Cometabolite Measurements 566
17.3.1.1 Solution of the Stoichiometric Matrix . 568
17.3.1.2 The Objective Function . . . . . . . . . 570
17.3.2 General Principles of Isotope Balancing 570
17.3.2.1 Steady State Flux Analysis . . . . . . . . 570
17.3.2.2 The Isotope Balance . . . . . . . . . . . 571
17.3.3 Least Squares Fitting of the Algebraic Form 573
17.3.4 Atom Mapping/Transition Matrices 573
17.3.5 Isotopomer Mapping Matrices . . . . 575
17.3.6 Transient NMR Measurement . . . . . 578
17.3.7 Errors in the Determination of Fluxes 578
17.3.7.1 Errors in Linear Models . . . . . . . . 578
17.3.7.2 Errors in Non-linear Models . . . . . 579
17.4 Biochemical Pathway Model Formulation and Reduction 580
17.4.1 Reduction of Comprehensive Models . . . . . . . . . . 580
17.4.2 Pathway Inclusion and Reduction Assumptions . . . . 582
17.5 Observed Metabolic Flux Patterns in Mammalian Cells 588
Contents XVII

17.5.1 Linkage of Glycolysis to the Tricarboxylic Acid Cycle 588

17.5.2 Reducing Equivalents 589
17.5.3 Glutaminolysis . . . . . . . . 589
17.5.4 Pyruvate Carboxylase . . . . 590
17.5.5 Pentose Phosphate Pathway . 591
17.5.6 Tumors as Nitrogen Sinks . 591
17.5.7 Oxidative Glycolysis in the Rat Brain 592
17.6 Specific Uses of Flux Pattern Information 592
References 593

Subject Index . . . . . 595

List of Abbrevations and Symbols

Chapter 1

A cross section area of the flow, interfacial area, m 2

a specific interfacial area, m- I
ATL airlift tower loop
BC bubble column
Bo Bodenstein number of axial dispersion, 1
C concentration, kg m- 3
CLi molar concentration of i in the liquid phase, kmole m- 3
CLX cell mass concentration in the liquid phase, kg m- 3
CPR CO 2 production rate, kg m- 3 S-I
CSTR continuous stirred tank reactor
D dilution rate, S-I
Di liquid diffusivity of substance i, m 2 S-I
DL liquid phase axial dispersion coefficient, m 2 S-I
Dm molecular diffusivity, m 2 S-I
Ds diameter of the column, m
EG gas hold up, 1
E(t) residence time distribution
F flow rate m 3 S-I
f(t) weighting function
g acceleration of gravity, m S-2
H Henry constant, bar- I
Hj height of the section j (j=r, d, k)
iTR transfer rate of substance i, kg m- 3 S-I
K equilibrium constant
k fluid consistency factor, kgm- I s-m
kH Henry constant, m 2 S-2
kL mass transfer coefficient, m 2 S-I
kLa volumetric mass transfer coefficient, S-I
m mass, kg
M molecular mass, kg kmole- I
n flow behavior index
n particle number
List of Abbrevations and Symbols XIX

OTR oxygen transfer rate, kg m- 3 S-l

OUR oxygen uptake rate, kg m- 3 S-l
p pressure, bar
p pressure, Pa
Po power input, kg m Z S-3
P power input, kW
paz percentage of Oz saturation
Q gas throughput, m 3 S-l
Q reaction rate, kgm- 3 s- 1
qi specific mass fraction transport rate of i
R gas constant, 8.314 J mole- 1 K- 1
r model constant
RQ respiratory quotient
RTD distribution of residence times
rG gas recirculation ratio
ST stirred tank
s Laplace transform variable
T temperature, K
time, s
te circulation time, s
UB bubble rise velocity, m S-l
UBT bubble terminal velocity, m S-l
Ue liquid circulation velocity, m S-I
UG superficial gas velocity, m S-l
UGD fraction weighted drift velocity, ms- 1
Us slip velocity of bubbles, m S-l
UL superficial liquid velocity, m S-I
V volume, m- 3
w gas throughput, S-I
x mole fraction
X cell concentration, kg m- 3
YXIO yield coefficient of growth with respect to oxygen consumption
z axial coordinate
£ volume fraction, gas holdup
11 dynamic viscosity of the medium kg m- I S-I
p density of the medium, kg m- 3
cr surface tension, N m -I
v kinematic viscosity m Z S-I
't mean residence time, s

Subscripts

C COz
D dispersed gas phase
d down comer
E ethanol
e entrance
G gas phase
H head space
 substance index
j stage index
k head section
L liquid phase
m measured
M molar
N nitrogen
0 oxygen
p product
r rIser
S substrate
W water
X cell mass

Superscripts

reactor inlet
N normal (standard) conditions
o reactor exit
Q reaction
R reservoir
TR transfer

Chapter 2

A area, m 2
c intrinsic concentration, -
c vector of intrinsic concentrations, -
C concentration, kg m- 3
C vector of concentrations
D dilution rate, S-I
Di diffusion coefficient of substance i, m 2 S-I
E activation energy, J mole- I
f function
F fraction of cells
g function
J performance index of metabolic coordinator
k constant
kLa volumetric mass transfer coefficient, S-I
K kinetic constant
L thickness, m
m specific maintenance coefficient, S-I
m vector of specific maintenance coefficients
M matrix of molecular weights, kg mole- I
OTR oxygen transfer rate, kg m -3 S-I
P partitioning function
p parameter vector
pH pH-value
q specific reaction rate, S-1
List of Abbrevations and Symbols XXI

q vector of specific reaction rates

Q volumetric reaction rate, kgm- 3 S-1
Q vector of volumetric reaction rates
r radius, m
r intrinsic reaction rate, S-1
r vector of intrinsic reactions
r normalized kinetic rate expression
R gas constant, 8.314] mole- 1 K- 1
5 fraction of sugars
t time, s
T temperature, K
TR vector of mass transfer rates
u cybernetic variable
x normalized variable
x vector of spatial coordinates
v volume fraction
V volume, m 3
X concentration related to the biotic phase, kg m- 3
X vector of concentrations of the biotic phase
Y yield coefficient
Y matrix of yield coefficients
z spatial coordinate, m
z vector of cellular state variables
Z vector of concentrations of the abiotic phase
(~ fraction of viable cells, or probability for plasmid loss
(3 probability for mutation
77 transcription efficiency
v activity controlling variable
iL specific growth rate, S-1
p density, kg m -3
T cell age, s
E, translation efficiency

Subscripts

B antibiotic
crit critical
D death or decay
E endogenous or enzyme
G growth or gene
component index
I inhibitor
j component index
L liquid phase
min minimum value
max maximum value
M metabolite
0 oxygen
P product
XXII List of Abbrevations and Symbols

S substrate
V volumetric
X biomass, biotic phase
Z abiotic phase
+ plasmid bearing
plasmid free

Superscripts

R reservoir
T transpose of a matrix

Chapter 3

MFA Metabolic Flux Analysis

S Stoichiometric matrix of metabolic network
.Y: vector of fluxes
! vector of metabolite extracellular accumulation rates
Vi flux of reaction i
ri extracellular accumulation rate of metabolite i
LP Linear Programming
TCA Tricarboxylic Acid Cycle
~ vector of weight factors of fluxes in the objective function of the linear
programming problem
Ci weight factor of flux Vi in the objective function of the linear program-
ming problem
Pl shadow price (dual variable) of stoichiometric constraint i
z optimal value of the objective function of the linear programming pro-
blem
NMR Nuclear Magnetic Resonance spectroscopy
ATP Adenosine triphosphate
NADH Nicotinamide adenine dinucleotide
GC [LC]-MS Gas (or Liquid) Chromatography-Mass Spectrometry
H4D tetrahydrodipicolinate
DAP diaminopimelate
PEP phospho-enol-pyruvate
G6P glucose-6-phosphate
CHO Chinese Hamster Ovary
MCA Metabolic Control Analysis
FCC Flux Control Coefficient
C;; Flux control coefficient of reaction i (as expressed through the activity
of enzyme i) on the flux of reaction j
activity of enzyme i of the network
steady-state flux through reaction j (symbol used in MCA)
Group Flux Control Coefficient
List of Abbrevations and Symbols XXIII

Chapter 5

A state matrix
AD matrix of the state transformation
B input matrix
D dilution rate (h -I)
f state mapping
F feedrate vector
influent flow rate (l/h)
g input mapping
gt (go) gain matrix of the RLS algorithm (initial value)
G carbon dioxide
Gt gain matrix of the RLS algorithm
kij yield coefficient (1=1-9, j=1-4)
km maintenance coefficient
K yield coefficient matrix
K[ inhibitor constant (gIl)
Kp controller gain
Ks Monod constant (gil)
L lactate
M number of reactions
N number of process component
OHPA Obligate Hydrogen Producing Acetogens
P product concentration (gIl)
Psat saturation outflow rate vector
PI Proportional Integral
Q gaseous outflow rate vector
r reaction rate vector
RLS recursive least squares
S substrate concentration (gIl)
S' desired value of S (gIl)
time (h)
u input vector
x state vector
X biomass concentration (gIl)
Y, output variable
Y desired value of Y
Y yield coefficient
Z auxiliary variable
I forgetting factor
'1,'2
E
gains of the observer-based estimator
singular perturbation variable
A controller again
J-L specific growth rate (h- I )
J-Lmax maximum specific growth rate (h- 1 )
II singular perturbation variable
integral (or reset) time
process component concentration vector
XXIV List of Abbrevations and Symbols

n steady state coefficient

Subscripts

in influent

Superspcripts

steady-state
estimate
/\

Chapter 6

AJK heat transfer area between thermostat subsystems J and K [mz)

Av vessel cross-sectional area [mz)
CIK concentration of component I in subsystem K [g rl)
CK specific heat capacity in thermostat subsystem K [WhK-Ig- I)
CIK molar concentration of component I in subsystem K [molrl)
CkLa fitted kLa-constant [I N-zh -I)
CK volumetric heat capacity in thermostat subsystem K [W h K- I)
dz stirrer diameter [m]
DK reciprocal mean residence time in thermostat subsystem K [h-I]
FK flow rate from (in) subsystem K [lh- I)
FnG aeration rate at normalized conditions [I h -I)
Fnl aeration rate of gas component I at normalized conditions [lh- I)
HI Henry coefficient of component I [Nm-Zlg- I)
ITR volumetric mass transfer rate of component I from gas phase [g I-I h -I]
k lji inhibition constant of substrate j by substrate i [g I-I]
kJ Monod limitation constant of component J [g I-I]
kLa volumetric Ortransfer coefficient [h -I)
KAlvol evaporation constant of ammonia [-)
KJi dissociation constant i of component J [mol r I]
Kw ion product of water [moe rZ]
K HSt proportional gain of stirrer heat generation [W min 3 1- I)
KHM proportional gain of microbial heat generation [W h g-I)
mJ exchange mass in thermostat subsystem J [g]
mlK mass of component I in subsystem K [g)
m mass flow [gh- I)
mOT oxygen mass transfer rate from gas phase [gh- I)
MI mole mass of component I [g mor l )
N st stirrer agitation speed [min-I]
Ne Newton number [-)
OUR oxygen uptake rate [gl-Ih- I)
Pi valency of an anion [-)
PK total pressure in subsystem K [N m -z)
PIK partial pressure of component I in subsystem K [N m -z)
pH pH-value [-)
PH electrical heating power [W)
PSt stirrer power input [W)
List of Abbrevations and Symbols xxv

valency of a cation [-]

cell specific reaction rate of component I [g g-I h- I ]
supply and desupply rate of component J [g 1-1 h -1]
thermal power of the stirrer [W]
thermal power of the microorganisms [W]
volumetric reaction rate of component I [g 1-1 h -1]
universal gas constant [N m K- I kmol- I ]
cooling water entry heat resistance [K W- I ]
overall heat transmission resistance between thermostat subsystem J and K
[KW- I ]
RT thermostat circulation heat resistance [K W- I ]
t time [h]
TK absolute temperature of subsystem K [K]
UG superficial gas velocity [m s-1]
VK volume of subsystem K [l]
VnM gas mole volume [l mol-I]
XIAIR mole fraction of component I in AIR [-]
XIG mole fraction of component I in the gas phase [-]
YX/I cell mass yield coefficient of component I [g g-I]
CY stirrer power input kLa-correlation parameter [-]
CYJ heat transfer coefficient in fluid J [W m -2 K- 1 ]
(3 aeration rate kLa-correlation parameter [-]
"( viscosity kLa-correlation parameter [-]
DJK wall thickness between thermostat subsystems J and K [m]
DC/o ratio of CO 2 /O T transfer coefficients [-]
T}eff effective viscosity [N s m -2]
'19 K temperature in subsystem K [0C]
iiI environmental growth control function of parameter I [-]
AJK thermal conductivity of the wall between thermostat subsystems J and K
[Wm-IK- I]
specific growth rate [gg-I h- I ]
density of the medium in vessel K [g I-I]
thermal transport time constant between thermostat subsystems J and K [h]

Indices

Ac Acid
AIR air
Al alkali (base)
C cooling water system
C,C02 carbon dioxide
D subsystem double jacket
eff effective
E reactor environment
gr growth fraction
G subsystem gas phase
H heating, harvest
H+ hydrogen ion
I inhibition
XXVI List of Abbrevations and Symbols

m input
II] component I per component]
L subsystem liquid phase
m maintenance fraction
max maximum value
min minimum value
M microbial, molar
n normalized gas conditions
N,N2 nitrogen
opt optimal value
out output
0,02 oxygen
Pi product i
R subsystem feed tank (reservoir)
S sample
Si substrate i
St stirrer
tot total amount of a dissociable variable
T thermostat
Tl subsystem titration tank 1 (acid)
T2 subsystem titration tank 2 (base)
Tc thermostat cooling system
Th thermostat heating system
vol volatile
V evaporation of water, vessel
w set point
W water, reactor wall
X dry biomass

Superscripts
,. equilibrium with gas phase
derivated over the time
/\ estimated value

Chapter 7

a specific interfacial surface area, m- 1

array of virtual acceleration, m S-2
sectional area of a bubble, m 2
concentration of species i, mol m- 3
drag coefficient of a bubble, -
Cvrn , Cvma coefficients of virtual mass force, -
C·
0, oxygen concentration at the gas liquid interface, mol m- 3
dlO average bubble diameter, m
d32 Sauter mean bubble diameter, m
db bubble diameter, m
di impeller diameter, m
Deft effective turbulent dispersion coefficient, m 2 S-1
List of Abbrevations and Symbols XXVII

diffusion coefficient of oxygen, m 2 S-1

tank diameter, m
correction factor of drag coefficient, -
force, N
gravitational acceleration, m S-2
liquid height, m
inhomogeneity, -
turbulent kinetic energy, m 2 S-2
mass transfer coefficient, m S-1
volumetric mass transfer coefficient, S-1
half saturation constant, g r 1
Eulerian macro length scale, m
resultant Eulerian macro length scale, m
impeller speed, S-1
ni number of impellers, -
nt number of tanks, -
p pressure, Pa
po, partial pressure of oxygen in the gas phase, Pa
P power input of the impeller, W
Pk production of turbulent kinetic energy, m 2 S-3
q relative number density, -
qo, oxygen transfer rate, mol m- 3 S-1
QL liquid pumping rate of impeller, m 3 S-1
r radial coordinate, m
ri reaction rate of species i, mol m- 3 S-1
Sc source of species c, mol m- 3 S-1 or gm- 3 s- 1
Si interfacial coupling term, N m- 3
Sc Schmidt number, -
SCturb turbulent Schmidt number, -
t time, s
tc circulation time, s
Ui mean velocity component, m S-1
Ut rise velocity of a bubble, m S-1
Utip. impeller tip velocity, ms- 1
uc D superficial gas velocity, m S-1
V volume, m 3
w impeller blade height, m
X coordinate, m
x0 2 concentration fraction of oxygen in the liquid phase, -
Yo, molar fraction of oxygen in the gas phase, -
y yield coefficient, -
z axial coordinate, m

Greek letters

{j Kronecker symbol, -
c energy dissipation rate, m 2 S-3
cc void fraction, -
cp tangential coordinate, 0
XXVIII List of Abbrevations and Symbols

dynamic viscosity, Pa s
Vturb turbulent viscosity, m 2 S-I
() mixing time, s
p density, kg m- 3
(J surface tension, N m- I
Td dissipation-range time scale, s
T;j laminar deformation tensor
Tp production-range time scale, s

Subscripts

0 without gassing
Ac acetoin
Bu butanediol
G gas phase
L liquid phase
m modified
O2 oxygen
P product
X biomass
S substrate

Dimensionless numbers

Morton number
Reynolds number of a bubble
Schmidt number
turbulent Schmidt number
Weber number

Chapter 9

A area, m 2
ACD acetaldehyde
ACY acetyl-coA
ADP adenosine-diphosphate
ADY active dry yeast
ATP adenosine-triphosphate
b(i) discrete budded cell number density in age interval i
B budding phase
c intrinsic concentration, mole g-I
C vector of concentrations
C concentration, kg m- 3
CCM cell cycling model
CPR carbon dioxide production rate, kg m- 3 S-I
d diameter, m
d(i) discrete daughter cell number density in age interval i
D dilution rate, 5- 1
D daughter phase
List of Abbrevations and Symbols XXIX

DL dispersion coefficient of the liquid phase, m 2 S-I

E gas holdup
EtOH ethanol
i population density variable
f shape function of population density
f input function of regulation model
P flow rate, m 3 S-I
F system function of regulation model
PRC fraction of budding cells
PDP fructose-diphosphate
P6P fructose-6-phosphate
G growth or intermediate phase
GLU glucose
G6P glucose-6-phosphate
H height, m
I mode number of the parent cycle
f mode number of the daughter cycle
f profit or performance criterion
k coefficient
kLa volumetric mass transfer coefficient, 5- 1
K constant
m specific maintenance coefficient, mole(g sri
m vector of specific maintenance rates
M mitotic phase
M molecular weight
MAC macromolecular cell constituent
n number of discrete cell cycle intervals
N cell number
NAD nicotine-amid-dinucleotide
OUR oxygen uptake rate, kg(m 3 sri
P specific price, kg- I
pU) discrete parent cell number density in age interval i
p parent phase
P price
PYR pyruvate
PIO effectiveness of oxidative phosphorylation
q vector of specific reaction rates
q specific reaction rate, mole(g sri
r intrinsic reaction rate, mole(g sri
r vector of intrinsic reaction rates
R volumetric reaction rate, g(m 3 sri

R(x) switch function, R(x) = {xo for x >0

:::;0
RQ respiratory quotient
s synthetic phase
t time, s
T average cell number doubling time or length of cell cycle phase, s
TIJ oscillation period for mode If, s
XXX List of Abbrevations and Symbols

Tee tricarboxylic-acid cycle

u superficial velocity, m S-1
V volume, m 3
XN cell number concentration, m- 3
y intrinsic state vector of regulation model
Y yield coefficient
Y matrix of yield coefficients
z normalized spatial coordinate
a stoichiometric coefficient
all decrement of oscillation for mode IJ
specific growth rate, S-1
T cell age, s

Subscripts

B budded phase
e carbon dioxide
d downcomer
D daughter phase
E ethanol
ex external
F fermentation
G gas phase
j index of reactor subsystem
k head
L liquid phase
min minimum value
max maximum value
o oxygen
P parent phase, or interval of reactor preparation
r riser
S substrate, or synthetic phase
set setpoint
tot total
T total
X biomass
o initial value

Superscripts

F final condition
R reservoir
o initial condition

Chapter 12

CTR=QC02 volumetric carbon dioxide transfer rate [gil h]

ePR=Rc02 volumetric carbon dioxide production rate [g/lh]
Cc concentration of compound e
List of Abbrevations and Symbols XXXI

CS,f substrate concentration in feed medium [gIl]

eE,e mass fraction of element E in compound C [-]
f volume flow rate [l/h]
kLa volumetric gas transfer coefficient [lIh]
me mass of compound C in the reactor [g]
MW molecular weight [g/mol]
OTR=Q02 volumetric oxygen transfer rate
OUR=R02 volumetric oxygen uptake rate [gIl h]
q mass flow rate [gIl h]
Q volumetric mass transfer rate [gil h]
r specific turnoverlreaction rate [gIg h]
R volumetric turnoverlreaction rate [gIl h]
sv,p fully relative sensitivity of variable v to parameter p [-]
t time [h]
tc cultivation time [h]
VR reactor volume [1]
X gas mole fraction [-]
Yx,e differential yield coefficient of compound X from compound C [-]
=rx specific growth rate [lth]
m maintenance term
1\
estimated value
o (subscript) starting (fed-batch) conditions
o (superscript) nominal value of variable
saturation value

Chapter 13

A mean culture age, s

Ap mean projected area of a pellet, m 2
a model parameter
AAA L-a-aminoadipic acid
ACV L-a-aminoadipyl-L-cysteinyl- D-valine
APA aminopenicillanic acid
AT acyltransferase
b model parameter
c intrinsic concentration
c vector of intrinsic concentrations
C concentration, kg m- 3
C vector of concentrations
CLS corn steep liquor
CPC Cephalosporin C
D dilution rate, S-1
Di diffusion coefficient of substance i, m 2 S-1
d diameter, m
DAOC deacetoxy-cephalosporin C
e total number of hyphal elements
f number density function
F flow rate, m 3 S-1
XXXII List of Abbrevations and Symbols

g spore germination frequency, S-I

h loss function in population balance, S-I
IDP Iterative Dynamic Programming
IPN Isopenicillin N
J performance index
k constant
K constant
1 length, m
m mass, kg
m specific maintenance coefficient, S-I
m vector of specific maintenance coefficients
n number of hyphal tips
N stirrer speed, S-I
OTR oxygen transfer rate, kg m- 3 S-I
P partitioning function
P probability or price
PAA phenylacetic acid
Pen Penicillin
POA phenoxyacetic acid
p02 percentage of dissolved oxygen saturation
q specific reaction rate, S-I
q vector of specific reaction rates
Q volumetric reaction rate, kg m- 3 S-I
r radius, m
r intrinsic reaction rate, S-I
r vector of intrinsic reactions
r normalized kinetic rate expression
t time, s
V volume, m 3
x variable
X constituent of the biotic phase
Y yield coefficient
Y matrix of yield coefficients
vector of process states
linear extension rate, ms- I
Dirac delta function
branching frequency, S-I
specific rate of fragmentation, S-I
specific growth rate, S-I
P density, kg m- 3
't time, s
AShear shear parameter in probability distribution

Subscripts

av average value
bra branching
bre breakage
C carbon dioxide
List of Abbrevations and Symbols XXXIII

crit critical
cys cysteine
Cycl cyclase
D death or decay
e effective
E enzyme
eq equilibrium
Exp expandase
f final
fra fragmentation
g germination
gro growth
hgu hyphal growth unit
Hyd hydrolase
component index
inhibitor
L liquid phase
lag lag-phase
ly lysis
m maintenance
min minimum value
max maximum value
0 oxygen
Oil soy oil
P product
p pellet
PM pharma medium
pr preparation
R averaged variable in radial layers
S substrate
Stir Stirrer
t total
thr threshold
tip hyphal tip
val valine
X biomass

Superscripts

R reservoir

Chapter 14

Metabolites
1,3-DPG 1,3-Diphosphoglycerate
3-GP 3-Glycerolphosphate
ADP Adenine diphosphate
AMP Adenine monophosphate
XXXIV List of Abbrevations and Symbols

ATP Adenine triphosphate

cAMP cyclic AMP
Fl,6bP Fructose-l,6-bisphosphate
F2,6bP Fructose-2,6- bisphosphate
F6P Fructose-6-phosphate
FADHz Flavinadenine dinucleotide (reduced)
G6P Glucose-6-phosphate
GAP Glyceraldehyde-3-phosphate
GDP Guanosine-diphosphate
GTP Guanosine-triphosphate
NAD Nicotinamide adenine-dinucleotide (oxidized)
NADH Nicotinamide adenine-dinucleotide (reduced)
NADP Nicotinamide adenine-dinucleotide phosphate (oxidized)
NADPH Nicotinamide adenine-dinucleotide phosphate (reduced)
PEP Phosphoenole-pyruvate
T6P Trehalose-6-phosphate

Enzymes

6PG-DH 6-Phosphogluconate-dehydrogenase
C-PKA Catalytic subunit of PKA
FBPasel Fructose-bisphosphatase-l
G6P-DH Glucose-6-phosphate-dehydrogenase
GAP-DH Glyceraldehyde-3-phophate-dehydrogenase
GDH Glutamate-dehydrogenase
HK Hexokinase
PFKl Phosphofructokinase-l
PFK2 Phosphofructokinase-2
PK Pyruvate-kinase
PKA Proteinkinase A
R-PKA Regulatory subunit of PKA
TAL Transaldolase
TKL2 Transketolase-2

Other abbreviations

DNA Deoxyribonucleic acid

EMP Embden-Mayerhof-Parnas pathway
HPLC High performance liquid chromatography
P/O P/O ratio
PCA Perchloric acid
PPP Pentose-phosphate pathway
Rl Relaxed conformation
R2 Relaxed subconformation
RNA Ribonucleic acid
Tl Tensed conformation
T2 Tensed sub conformation
List of Abbrevations and Symbols xxxv

Mathematical Symbols

U Index for compartment

Ui Non-exclusive binding coefficient for metabolite i
~ Index for compartment
c Exclusive binding coefficient
Ci Concentration of metabolite i
<;:j Vector of concentration for the reaction j
<;:j Vector of steady state concentration for the reaction j
Ci Approximation function for the time course of metabolite i
D Dilution rate
E Elemental composition matrix
t Error criteria
11 Efficiency of oxidative phosphorylation
trel,i Relative error square for metabolite i
jj Kinetic rate expression of reaction j
r Reaction matrix
k Maintenance coefficient
Ki Affinity with respect to metabolite i
L General allostery coefficient
La Allostery coefficient in the absence of ligands
Growth rate
N N function
Ni Michaelis-Menten term for metabolite i
Pj Parameter vector for reaction j
Q Effector function
Q Vector of net-conversion rates
r Vector of reaction rate
r"'j Maximal rate of reaction j
rj Steady state flux of reaction j
rDeg,F2,6bP Rate of degradation of F2,6bP
rm,ATP Maintenance
t Time
T Transport matrix
t Transport vector
V Volume
YX1S Coefficient of yield
Z Z function

Chapter 17

null vector
ATP adenosine triphosphate
Uj,A reaction coefficient for pathway j and metabolite A
AMM atom mapping matrix
A stoichiometric matrix
AT aminotransferase
BHK baby hamster kidney (cells)
l3C carbon, isotope 13
XXXVI List of Abbrevations and Symbols

14C carbon, isotope 14

Ci steady state concentration of i
eM intracellular concentration of M
Ci,O feed concentration of i
CO 2 carbon dioxide
CHO Chinese hamster ovary (cells)
CS citrate synthase
() tolerance
DHAP dihydroxyacetone phosphate
E redundancy matrix
fj flux through pathway j
f column vector of all fluxes
f vector of best fit fluxes
F volumetric flow rate
F6P fructose 6 phosphate
FADH2 flavin adenine dinucleotide, reduced
G6P glucose 6 phosphate
GDH glutamate dehydrogenase
GTP guanosine triphosphate
HK hexokinase
I identity matrix
mv isoptopomer distribution vector
IMM isotopomer mapping matrix
LDH lactate dehydrogenase
"- label flux
LAi concentration of A labeled on carbon i
MAi fractional enrichment of metabolite A, carbon i
MDH malate dehydrogenase
ME malic enzyme
N cell number within a perfusion device
NADH nicotinamide adenine dinucleotide, reduced
NADPH nicotinamide adenine dinucleotide phosphate, reduced
NH3 ammonia
NMR nuclear magnetic resonance (spectroscopy)
N0 3 nitrate
OAA oxaloacetate
OUR oxygen uptake rate
P permutation matrix
PC pyruvate carboxylase
PDH pyruvate dehydrogenase
PEPCK phosphenolpyruvate carboxykinase
PK pyruvate kinase
PPP pentose phosphate pathway
PSSA pseudo-steady-state approximation
rA rate of accumulation of metabolite A

;2
r
1
column vector of all measured fluxes
variance of measurement i
TA transaldolase
TCA tricarboxylic acid (cycle)
List of Abbrevations XXXVII

TK transketolase
V chemstat liquid volume
V intracellular volume
X cell concentration
Ylac/gluc molar yield of lactate from glucose

Metabolites in Figs. 17.1 and 17.7

A acetyl-CoA
AL alanine
C citrate
D dihydroxyacetone phosphate
DPG diphosphglycerate
E erythrose
F fructose
FU fumarate
G glucose
GA glyceraldehyde
K a-ketoglutarate
L lactate
M malate
N glutamine
0 oxaloacetate
P pyruvate
PE phosphoenolpyruvate
PG phosphoglycerate
R ribose
Ru ribulose
S succinate
Se sedoheptulose
T glutamate
X xylulose
Contributors

Karl-Heinz Bellgardt
Institut fUr Technische Chemie der Universitat Hannover,
Callinstrasse 3, 30167 Hannover, Germany
e-mail: [email protected], Fax: +49 511 762 3004
Harvey ~ Blanch
Department of Chemical Engineering, 201 Gilman Hall, University of California,
Berkeley, CA 94720, USA
e-mail: [email protected], Fax: + 1 510 643 1228
Douglas S. Clark
Department of Chemical Engineering, 201 Gilman Hall, University of California,
Berkeley, CA 94720, USA
D. Dochain
Senior Research Associate FNRS, Cesame, Universite Catholique de Louvain, Bat.
Euler, 4-6 avo G. Lemaitre, 1348 Louvain-La-Neuve, Belgium
e-mail: [email protected], Fax: +32 10 472180
Neil S. Forbes
Department of Chemical Engineering, 201 Gilman Hall, University of California,
Berkeley, CA 94720, USA
e-mail: [email protected]
K. Gollmer
University of Applied Sciences Trier, Environment Campus, 55761 Birkenfeld,
Germany
e-mail: [email protected]
Albert A. de Graaf
Institute of Biotechnology I, Research Centre Jillich, 52425 Jillich, Germany
e-mail: [email protected], Fax: +49 2461 612710
Bernd Hitzmann
Institut filr Technische Chemie, Universitat Hannover, Callinstrasse 3, 30167
Hannover, Germany
e-mail: [email protected]
Contributors XXXIX

Marc Jenne
Institut fUr Bioverfahrenstechnik, University of Stuttgart, Allmandring 31, 70569
Stuttgart, Germany
Maria I. Klapa
Department of Chemical Engineering, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA
e-mail: [email protected]
Andreas Liibbert
Martin-Luther-University, Halle-Wittenberg, 06099 Halle/Saale, Germany
e-mail: [email protected], Fax: +49 345 5527 260
Reiner Luttmann
University of Applied Sciences Hamburg, Research Center of Bioprocess Engineering
and Analytical Techniques, Lohbrugger Kirchstrasse 65, 21033 Hamburg, Germany
e-mail: [email protected]
Klaus Mauch
Institut fur Bioverfahrenstechnik, University Stuttgart, Allmandring 31, 70569
Stuttgart, Germany
M. Perrier
Departement de Genie Chimique, Ecole Poly technique de Montreal, Succursale
"Centre Ville", CP 6079, Montreal H3C 3A7, Canada
Clemens Posten
Institut fur Mechanische Verfahrenstechnik und Mechanik, Universitat Karlsruhe
(TH), 76128 Karlsruhe, Germany
e-mail: [email protected], Fax: +49 721 6086
Matthias Reuss
Institut fur Bioverfahrenstechnik, University of Stuttgart, Allmandring 31, 70569
Stuttgart, Germany
e-mail: [email protected], Fax: +49 711 685 5164
Ursula Rinas
GBF, National Research Center for Biotechnology Ltd., Dept. of Biochemical
Engineering, Mascheroder Weg 1, 38124 Braunschweig, Germany
e-mail: [email protected], Fax: +49 531 6181 111
Sven Schmalzriedt
Institut fur Bioverfahrenstechnik, University of Stuttgart, Allmandring 31, 70569
Stuttgart, Germany
Karl Schiigerl
Institut fur Technische Chemie Universitat Hannover, Callinstr. 3, 30167 Hannover,
Germany
e-mail: [email protected], Fax: +49 511 7622253
Gregory Stephanopoulos
Department of Chemical Engineering, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA
e-mail: [email protected], Fax: +1 6172533122
XL Contributors

Sam Vaseghi
Institut fUr Bioverfahrenstechnik, University Stuttgart, Allmandring 31, 70569
Stuttgart, Germany
John Villadsen
Department of Biotechnology, Block 223, Technical University of Denmark, 2800
Lyngby, Denmark
e-mail: [email protected], Fax +45 4588 4148)
Introduction
Karl-Heinz Bellgardt

The Need for Modeling and Control in Biotechnical Processes

The kinetics of biotechnological processes are determined by the properties of the

microorganisms, the construction of the reactor, as well as by the cultivation condi-
tions and media. The metabolic flexibility of the cells in connection with inhomo-
geneities in the reactor often leads to very complex growth dynamics, which make it
difficult to ensure high operational stability and reproducibility of the process, as
well as constant product quality and yield. The situation is further complicated by
the fact that technical substrates for industrial processes are rather undefined. The
widely established on-line analytical methods are mostly not sufficient or fast en-
ough to characterize the state of the running process. Nevertheless, before the back-
ground of increased international competition, reduced profit margins, and rigorous
safety and environmental regulations, there is clearly a need for improved process
control and optimization based on advanced analytical methods for substrates, pro-
ducts and state of the cells [1,2].
Mathematical models can take an important part in this task although biotechnical
processes are rather complex systems which still can be described only in a roughly
simplifying way. The complicated structure of the metabolism and the mechanisms
of its regulation are still not fully understood. Beside the intracellular processes, also
the variation of the local condition in the bioreactor, caused by the fluid dynamics of
the multiphase system, have to be looked at in a very simplified way. Nevertheless,
the mathematical models, which naturally must be incomplete and inaccurate to a
certain degree, can still be very useful and effective tools to describe those effects
which are of great importance for control, optimization, or our understanding of the
process [3-6]. Mathematical models provide a functional interrelation of the input
variables, output variables, and inner variables of the process that can easily evalu-
ated by computer. Thus numerical solution of the models is the fundament for the
development of economic and powerful methods in the fields of process design,
plant design, scale-up, optimization and automatic control [7-9] Most bioprocesses
are operated under non-stationary conditions in batch or fed-batch mode. This leads
to complicated optimal time profiles for the control variables, which are sometimes
impossible to be determined purely experimentally. Here, mathematical or numerical
optimization methods for determination of control variables and parameters can
advantageously be applied to reduce the experimental effort and the required time
for optimization.
The general principles of modeling of bioprocesses and related techniques are in-
troduced in Part A. The mathematical modeling of biotechnical processes is an
2 Introduction

extremely wide field that covers all important kinds of processes with many different
microorganisms or cells of plants and animals. The existing biological models are
aimed at various levels of the biotic phase, beginning with a description of the com-
plex intracellular reaction network, of the metabolic regulation on the reaction level,
of the processes of replication and translation of genes, the events during the cell
replication cycle, up to models for morphological processes and description of po-
pulation dynamics (Chapters 2 and 3). No less manifold are the reactor models,
spanning from simple homogenous single-phase systems (Chapter 1), up to complex
structured models of multiphase systems and more sophisticated, detailed hydrody-
namic models for description of transport and mixing processes in bioreactors
(Part B).
Due to the inherent nonlinear kinetics of a bioprocess, under non-stationary op-
eration there is an enduring change in the time constants of the system. Further-
more, most bioprocesses are essentially multi-variable systems with strong inner
couplings. From these facts arise serious problems for conventional methods of auto-
matic control, such as single-input/single-output PID-control. Here, control theory
provides the advanced methods of adaptive, non-linear, and multi-variable control,
which are often directly based on mathematical models. It is self-evident that the
more measurements are available, the better can be the automatic control. But even
with the most advanced methods of process monitoring, it will not be possible to
completely determine the actual reaction rates or the intracellular state of the cells as
an important subset of the state of the process. This difficulty can be overcome by
the concept of model-based estimators or observers for the not measurable states
and time-variable parameters. An introduction to the problems of automatic control
is given in Chapter 5. A closely related field is the detection of failures or undesired
process states: e.g., defects in sensors or actuators, infections by other microorgan-
isms. The application of methods of artificial intelligence in this area is outlined in
Chapter 4.
Process scheduling, supervision, automatic control, and documentation in modern
biotechnical plants is done by computerized process control systems (PCS), where all
the functions are implemented in software (See Fig. 1). The effort for software devel-
opment can account for a great or even major percentage of the total costs for con-
trol equipment. Here the application of model simulations for testing the PCS and
training of the operators can significantly reduce the time for setting to work the
plant. More details of such methods are given in Chapter 6.
Part C gives examples for the modeling of selected processes of industrial im-
portance. There, the general methods are applied to describe the kinetics of batch
and fed-batch cultivations of bacteria, yeasts, and filamentous fungi. The method of
metabolic flux analysis as introduced in chapter 3 is an important tool for evaluation
of the intracellular reaction network, and for optimization of cultivation conditions
and substrate composition. Its connection with genetic engineering for directed
modification of the cellular reaction network is known as metabolic design. Some
applications are given in Part D for yeast and bacteria.
Introduction 3

Some Modeling Basics

A definition of a model can be given as follows:

A model is an image of a real system that shows analogous behavior in the impor-
tant properties, and that allows within a limited region a prediction of the behavior
of the original system.
The experimental study of the original system can then be replaced by the model.
This can have several advantages:
Economy, cost factors: The model system is simpler, smaller, cheaper or faster then
the original system
Research and development: The model is less complex than reality, certain effects
can be emphasized or suppressed. This helps one to obtain a clear view of the
process.
There are generally three different groups of models, physical, mathematical, and
verbal models. Physical models can be realizations of the original system in a differ-
ent (usually smaller) scale or with structural modifications. A second type of physi-
cal models is obtained by turning to a different physical system, e.g. from the origi-
nal biotechnical process to electrical circuits. This is then called an analogous com-
puter. Verbal models give a linguistic representation of our knowledge about the
system, usually as rules (e.g. if this or that happens then the system reacts by ... )
[10, 11]. They are widely applied in the area of artificial intelligence, namely expert
systems. Mathematical models are not based on the real existence of a physical mod-
el system, but describe the behavior of the original system by mathematical equa-
tions. This is the highest level of abstraction. The experimental investigation is re-
placed by manipulation and solution of the model equations. A link between verbal
models and mathematical models is established by fuzzy logic, which allows one to
translate qualitative and rule-based knowledge into mathematical equations.
Mathematical models can be classified further depending on the mathematical
formalism or the methods for model building. In this book, we will focus on deter-
ministic theoretical models of biotechnical processes. Theoretical models as me-
chanistic models are based on physical and chemical laws and our knowledge about
the inner structure and function of the system, e.g. the flows of mass and energy.
They can, therefore, provide far-reaching predictions of the system behavior. In con-
trast, experimental or non-mechanistic models try to give - without looking into the
inner of the system - a description of the observed reaction of the system in re-
sponse to a certain forcing signal. These types of models are called black box mod-
els. Many "mechanistic" models in biotechnology are actually due to their over-sim-
plification quite closer to black-box models than to mechanistic models, although
mostly mechanistic interpretations are given. Very typical black-box models are Ar-
tificial Neural Networks, which have also found wide application in biotechnology
[12, 13].
The universal mathematical tool of modeling are differential equations which are
obtained by the balances for conservative quantities, such as mass of reactants, en-
ergy and impulse. These have the general form:
storage element convective transport diffusive transport
r-----''' " "
'local change in the' local local
reaction volume inflow - outflow inflow - outflow
4 Introduction

The models include input variables (e.g. Flow rate of substances), that are not
determined by the model itself, output variables (e.g., oxygen uptake rate, OUR)
and state variables (e.g. concentrations). The state variables represent storage ele-
ments for mass or energy of the system. The model can be seen as a calculation rule,
that relates a looked for pattern of the output variables (e.g. time course) to a certain
known or given pattern of input variables. Another type of model variables, the
parameters, are fIxed entities of the system. They can have a direct physical mean-
ing. Some parameters are known or can be measured, while others have to be deter-
mined from experimental data.
The discrimination in parameters, state variables, or input variables is not an im-
manent property of a system, but a specifIc view by the developer of the model,
which can be changed in relation to the mode of operation of the process or the
simplifying assumptions of the model. The temperature, for example, can be con-
trolled very tightly at a constant value, so it may be considered as a parameter in
one cultivation. In a second one, the operator may set different values certain times
or the control follows a given profIle; this time temperature would be an input vari-
able to the model. If the temperature varies freely it should be calculated by an en-
ergy balance, and thus became a state variable.
A process model as a mathematical model should provide a coherent description
of the entire process on the level of plant operation, including reactor and further
processing steps. The degree of complexity of the model or the possibility for a sim-
plifIed modeling of some parts of the plant is determined by the intended applica-
tion of the model. The model may be used in numerical simulations to answer sev-
eral interesting questions: What will be the output of product per unit of time for a
given input of raw material and primary energy? What are the costs of production
and of treatment of waste? What is the optimum mode of operation for the reactors?
Under which dynamic control of manipulating variables is the product obtained
with high productivity and a certain quality? How can the profIt be maximized? Very
simplifIed balances may be suffIcient as a model for the calculation of the total con-
version of the process, e.g. yield of cell mass or product. Greater modeling effort and
a more detailed description of the most important unit operations is generally re-
quired for a model based process optimization. One should be aware of the fact that
the result of such theoretical optimizations may be signifIcantly influenced by the
model accuracy, especially for the determination of an optimal dynamic control.
To develop an accurate and complete model is not an easy task for several reasons
[14]. Sometimes there are only very simplifIed models available for parts of the
plant, e.g. cultivation. And mostly not all interdependencies between the elementary
units of the process are qualitatively and quantitatively known: The cultivation is
influenced in a complicated manner by medium composition, by substrate quality
and preprocessing, and by the inoculum preparation. The cultivation itself may in-
fluence the downstream processing by changing rheology of the broth and varying
product properties. The situation for model building is further complicated because
for batch and fed-batch operation the process is never in steady state and the model
must consider the process kinetics. Therefore, the modeling of the biological system,
as outlined in chapters 2 and 3, is obviously an important aspect of process models
for biotechnological systems.
Model building is always a combination of theoretical studies and practical experi-
ments in a very iterative sequence. Since the problems in parameter identifIcation
and model verifIcation increase rapidly with the model complexity, one should begin
Introduction 5

Table 1. Step sequence of model building

Step Action

1 Running typical experiments

2 Define the modeling goal
3 Analysis of the system and determination of structural elements
4 Simplifying assumptions (e.g. about mixing, process structure and dynamics, me-
tabolism, kinetics)
5 Choice of important process variables: parameters, input variables, and states
6 Establishing the model by use of balances, physical laws and empirical equations
7 Simulation of the model; parameter identification to fit it to experimental data
8 Evaluation of the model quality; repeat with step 1

with as far as possible simplified assumptions and withdraw them step by step, in
the case that the model quality is not sufficient. In this way, a most simple initial
model grows step by step in complexity and accuracy, without becoming too com-
plicated. Modeling includes not only the selection of correct model structure, but
also the quantitative fitting to the experimental data by determining the model para-
meters. Unfortunately, their values are often not or only not exactly known in ad-
vance and must determined from experimental data by methods of parameter iden-
tification. The repeated steps of model building are summarized in Table 1.

Structure and Operation of Biotechnical Plant

A biotechnical production line can be roughly subdivided into three sections, the
preparation and preprocessing of the input raw materials, the cultivation, and the
downstream processing of the cultivation broth. An example for the general layout
of a biotechnical process is shown schematically in the upper part of Fig. 1. Each of
the sections is built of many elementary steps, the unit processes and unit opera-
tions, of which the cultivations of microorganisms or cells in the bioreactors are
naturally the most important ones [15].
In the preparation section, the primary energy is converted into process energy in
form of steam and compressed air. The substrate for the cultivations is prepared
from raw substrate. Here, several steps can be involved, such as filtration, hydrolysis
of higher carbon compounds, and sterilization. Usually further essential or growth
supplementing compounds like mineral salts or yeast extract are added. In aerobic
processes, sterile compressed air has to be provided for aeration of the bioreactors.
In the preculture, the inoculation material for the first seed reactor is prepared in
shaking flasks from deep-frozen cell material that is coming from the strain main-
tenance. The production starts with cleaning and sterilization of reactors, filling the
reactors with substrate, and inoculation with the preculture.
In the cultivation section, here assumed as batch plant, the cells or microorgan-
isms are propagated in the series of reactors. To avoid low cell concentrations and
low volumetric productivity, the reactors usually have an increasing volume, begin-
ning with the smallest seed tank in 10-1001 scale, up to the final production tank in
10-100 m 3 scale. When the cell concentration in a certain stage has reached its max-
imum, the fermentation broth is inoculated to the next stage. In the last production
6 Introduction

Downstream
Preparation Cultivation processing

Cooling water

Primary
energy

Raw
substrate
Additi-
yes

Stra in

Seed
tank

Air

Process Com pressed air

. . . . . . . . . . . . .... . . . . . . . ". . . . . . . . . ·. .t. ·". .·,. . t. .·. . ·.·. .,. .·$. . ". . . . t. . . . ,. . . . . . .·
~ control units
Field level

Process
computer
Local area network

Plant control
level

Operator term inal

Main computer

Fig.1. The general layout of a biotechnical production process.

stage, the major amounts of cell mass and product are formed, and emphasis is put
for the process control on a high product concentration, respectively productivity.
During the transfer from one reactor to the next, the broth can be further treated
and might be temporarily stored to meet the scheduling for the entire plant. In big
production plants, several seed or production tanks are often operated in parallel,
Introduction 7

but with a certain time shift. This renders an effective and more continuous usage of
the equipment for preparation and downstream processing.
In the down-stream processing section, a wide variety of processing steps can be
found, depending on the properties of the product. Gaseous products can be sepa-
rated in situ from the exhaust gas stream. Non-gaseous products have to be sepa-
rated from the broth after the final cultivation in the production tank. For intracel-
lular products the cells have to be disrupted at first. The cell material is then sepa-
rated by filtration or centrifugation. The product purification might include filtra-
tion, extraction, adsorption, dialysis, chromatography, distillation, and other meth-
ods. Often the raw product is further modified in purely chemical steps to obtain the
final product.
The treatment of waste is another important part of a biotechnological process.
Waste material arises from preparation of the substrate, cleaning of reactors, and
downstream processing. If it includes viable cells, the waste must be sterilized, and
the load of organic material must be reduced. Furthermore, depending on the down-
stream processing it may be necessary to remove solvents or other reactants.
In modern production plants, the process is coupled to a computer process control
system (PCS) that is shown schematically in the lower part of Fig. 1. The entire PCS
can be roughly subdivided into the field level and the process control level. The main
features of the different levels are summarized in Table 2. In decentralized systems,
the different units are independent and spatially separated. This can improve the
stability of the control system in the case of hardware failures. For smaller pilot
and laboratory plants, due to economic reasons the functions of field level and pro-
cess control level may be realized within only one centralized computer system.
In a decentralized system, the field level is built of small independent microcom-
puter control units which are directly connected to the measuring equipment and the
controlling devices of the plant. Each control unit serves only for a few variables or
control loops. Usually, simple signal processing functions like filtering, linearization,
and normalization as well as the basic control loops, such as for temperature, flow,
pressure, and pH are realized on this level. The control units also perform fixed
automation sequences via Programmable Logic Control (PLC), e.g. sterilization,
medium transfer. The field level provides only simple functions for displaying the
values of measured and controlling variables and for direct manual operation. The
control units can communicate to each other and to the process computer on the
next hierarchical level via the field bus. Prefiltered data are sent to the process com-
puter, set points and digital control commands are received from it.
The process control level covers the functions for automatic operation of the entire
plant and comfortable manual operation and supervision in the control room. Gra-
phical displays provide all information about the process and give access to the con-
trolling variables. The entire production run is automatically controlled by recipes,
which include all information on the different process phases, scheduling of the bior-
eactors and the other equipment, control strategies to be used, set-points during the
different phases of the process, profiles for controlling variables, and composition of
substrates. This is a very special feature of biotechnical processes which are mostly
not operated continuously, but in batch or fed-batch mode. Another important point
is that, different from chemical processes, many important variables which are es-
sential for an optimal control can only be measured off-line. Therefore, the PCS must
provide functions for handling of off-line data and for its use in the automatic con-
trol [16].
8 Introduction

Table 2. Specific functions and features on the different levels of process and control

Level Features Signals and variables

Process Mass flow Physical and chemical

Energy flow
(Bio-}Chemical reactions
Physical processes
Field Measuring, analog-digital conversion Electrical,
Local low level control by simple control loops optical,
Programmable logic control (PLC) digital data
Signal filtering
Short term data storage
Linearization and normalization
Local display
Direct manual operation
Process Common features:
control Man-machine interface via menus and graphics Digital data
Supervision, alarm handling
Balancing
Data processing and data reduction
Long term data storage
Documentation and protocolling
Advanced features:
Handling of off-line analysis
Recipes, scheduling
High level control
State and parameter estimation
Simulation, prediction
Optimization

On the process control level also some advanced functions can be realized that
require a high computing power: Balancing of mass and energy flows, estimation
of non-measurable quantities or of time-variable parameters, sophisticated control
of the cultivation that considers the state of the biological system, usage of model
simulations for prediction of the future course of the process, and on line optimiza-
tion. Another important task is complete proto colling of all process variables and all
process events, including long term data storage for future reference or further ana-
lysis of the production runs. Data reduction and calculation of economic indices
provides information for the next hierarchical level of factory control or manage-
ment.

Types and Structure Elements of the Bioreactor

The design of the bioreactor - as an important part of the entire production process
- together with its mode of operation with respect to substrate supply, air supply and
medium exchange has to ensure the optimal conditions for growth and product for-
mation of the microorganisms, or for cells of higher organisms, such as animals and
Introduction 9

plants. As a first step of process analysis aimed at mathematical modeling and con-
trol, the main structure elements of the bioreactor and its mode of operation will be
shortly summarized.
The bioreactor has to provide the proper physical and chemical environment -
e.g., temperature, pH, substrate concentration - for the cells, and to ensure fast
transport of substrates and products between gas phase, bulk medium and cells with
as low as possible effort in material, energy, and mechanics. Different types of bior-
eactors have been developed to optimize transport properties, homogeneity of the
liquid phase, gas dispersion, and mechanical stress. Design goals are good mixing
properties resulting in low gradients of dissolved reactants and temperature, good
heat exchange, and low costs of investment.

Acid/Base

Steam

Substrate
~

~
Anti-foam Medium outflow
agent

Cooling water

TIC Liquid phase with

disperged gas phase
and cells

Cooling water
Air sparger

,, Harvest
Air
,,
j
,-
Steam

Fig. 2. Schematic diagram of a stirred tank reaktor and basic control functions: control and
indication for flow rate (FIC), temperature (TIC), pressure (PIC), speed (SIC), quality variables
like pH (QIC), level (LIC).
10 Introduction

The characterization of the bioreactor as a multi-phase system means that, beside

the common inner-phase transport processes of convection and diffusion, the inter-
phase transport phenomena solid-liquid and liquid-gas play an important part for
the over-all kinetics. Since these phenomena can be influenced by the design of the
reactor, the effective minimization of inter-phase transport resistance is another cen-
tral goal for reactor design. With respect to the gas-phase, the effective supply of
oxygen and removal of gaseous products is desired. For a good oxygen usage the size
and residence time distributions of the bubbles have to be controlled to proper va-
lues. Both can be influenced by the design of air spargers and mechanical power
input. Since the oxygen transfer is often the key factor for the productivity of the
process, effective methods for the dispersion of the gas stream are required [17, 18].
The bioreactor, as shown for a STR in Fig. 2, is embedded into a set of equipment
to establish and maintain optimum conditions for the cells: heat exchangers and
related control loops for temperature, air compressors to ensure sufficient supply of
oxygen, control loops for the regulation of pH by addition of acid or base, foam
control by addition of antifoam agent, and last but not least the supply of substrates
and other medium additives. In several positions, controlled by a number of valves,
steam can be directed into the reactor and the connecting tubes for sterilization of
the entire system.
The bioreactors can be classified according to the employed methods for supply of
air and mechanical energy for mixing, and are shortly introduced in the following.
Beside the basic types, there exist a great number of variants.

The Stirred Tank Reactor as an Example for Reactors with Mechanical Energy Input

A typical stirred tank reactor (STR) as shown schematically in Fig. 2 has a height-to-
diameter ratio of usually less than 4:l. Mechanical energy for mixing is supplied by
an rotating impeller, which is also responsible for the gas dispersion. In small scale,
the STRs show a high specific energy input, and in consequence, good mixing and a
high possible oxygen transfer capacity. They can then usually be taken as ideally
mixed. With increasing size, the energy requirement grows over-proportional. Other
disadvantages are the great base area of the reactors, and the high mechanical peak-
stress near to the impeller. In the outer zones of big STRs the medium can be almost
stagnant. Therefore, practicable volumes of STRs range from 1 dm 3 to several 10 m 3 •

Reactors with Energy Input by Compressed Air

For very big reactors up to 500 m 3 , the usage of the expansion energy of compressed
air as an power input is most effective. The uprising gas bubbles lead also to accep-
table mixing of the liquid phase. To obtain a sufficient gas residence time, the height-
to-diameter ratio must be greater than 10:1. As examples, in Fig. 3, a bubble column
reactor and an airlift reactor with inner loop are shown schematically. In bubble
columns, the liquid flow pattern is rather undefined, and the liquid velocity is low.
This results in unsatisfactory mixing. In airlift reactors, there is a guided liquid cir-
culation with up-flow in the gassed region (Riser) and down-flow in the degassed
region (Down comer). The low gas-holdup in the down comer may lead to oxygen
limitation. The liquid velocity is relatively high, resulting in good mixing of the li-
quid phase. Advantages of this type of reactors are the small base area, the simple
design, and the low energy requirements for mixing.
Introduction 11

n 1\ n f"\
1\
r.
r /
I 11 I

I1 I I

I 1I I

Fig. 3. Schematic diagrams of a bubble column (left) and an airlift ractor with inner loop
(right).

( , ( '\
Tube
membrane

Fig. 4. Schematic diagram of membrane reactors for bubble-free aeration.

Membrane Reactors for Bubble Free Aeration

In the cultivation of animal cells the mechanical shear stress must be kept very low
to avoid damage of the cells. One origin of the shear forces are small uprising bub-
bles, because of the high velocity gradients around the bubbles. In membrane reac-
tors, such as shown in Fig. 4, the gas exchange with the liquid phase is via a micro-
porous tubular membrane. In this system bubble formation can be completely
avoided. The maximum oxygen transfer capacity is quite low, but sufficient for the
very slowly growing animal cells. The final end of the aeration membrane can be
either closed or open. In the first case, the oxygen usage is most effective, but degas-
sification of carbon dioxide can be only via the liquid surface. Therefore, an addi-
tional gas stream through the head-space of the reactor is necessary. In the second
case, the mass transfer capacity for carbon dioxide via the membrane is comparable
12 Introduction

to that for oxygen. Aeration of the head space is not necessarily required, but can be
used to increase the mass transfer capacity.
A bioreaction system is typically a multi-phase system. The incorporated phases
are:

Liquid-Phose
In a typical bioreactor, all substances and substrates which are required for growth
or product synthesis are dissolved or suspended in the liquid medium that sur-
rounds the cells. The medium contains organic and inorganic material that is taken
up and metabolized by the cells to provide cell material and chemical energy. The
main components are the sources of carbon (e.g. sugars, alcohols), nitrogen (ammo-
nium, amino acids), phosphorus (phosphate), vitamins, and oxygen. Also the pro-
ducts excreted by the cells are usually dissolved in the liquid phase.

Gas-Phose
The gas phase can be subdivided into the dispersed gas phase that is contained as
small gas bubbles in the liquid phase, and the gas phase in the head space of the
reactor. With the exception of reactors with bubble free aeration, only the first is of
much relevance for the biological reaction due to its high interfacial area. Neverthe-
less, for a detailed dynamic analysis of the growth kinetics also the head gas phase
has to be considered, because it increases the total gas residence time, and the delay
of exhaust gas measurements. Beside the inert components that might be contained
in the inlet gas stream (e.g. nitrogen in supplied air), the gas-phase consists of gas-
eous substrates such as oxygen, gaseous products of the metabolism such as carbon
dioxide, or evaporated components of the liquid phase, mainly water.

Solid-Phase
There can be several sources of solid particles in a bioprocess. Many substrates (eg.
casein, peanut powder, cellulosic material) contain insoluble fractions. Often, the
cells are immobilized by solid material to facilitate the retention of biomass in the
reactor. The cells can be attached to the surface of solid particles (eg. sand, polymer-
ous microcarriers), they can be immobilized within solid particles (porous beads of
clay, amorphous glass, cellulosic material, or alginate), or they can be encapsulated
within spherical membranes (microencapsulation, e.g. by polymers, cellulose). Even
in carrier-free systems when the cells form agglomerates or pellets, it might be ne-
cessary to consider the cell-flocs as a solid phase. The presence of the solid particles
can strongly influence the hydrodynamics, and the transport processes of substrates
and products due to additional diffusion resistance. In the case of dense and big
particles in the range of millimeters the reactor is operated as a fluidized or a fixed
bed. For low particle concentrations and small particles with a density comparable
to water, the solid phase might be neglected.
From the biological view-point, the gas phase, liquid phase, and solid phase may
be summed up as the abiotic phase that surrounds or is in contact with the cells. The
biological activity takes place in the biotic phase.

Biotic Phose
The entire population of cells in a bioreactor forms the biotic phase. Usually the
corpuscular character of the cells is neglected and the biotic phase is characterized
Introduction 13

in a volume element of the liquid phase by averaged quantities such as mass of total
cells or cell constituents, which are often expressed as concentrations. This is justi-
fied in the case of single cells or small cell flocs, since the cell size is only in the
range of {tm and the cell density is close to that of water. As biocatalysts, the cells
use substrates supplied in the cultivation medium for their own survival and cell
propagation, and by this synthesize a number of products among which also the
desired product of the bioprocess can be found. The different functional groups of
cellular metabolism for substrate degradation, energy production and biosynthesis
of cell material are closely connected to each other by the network of metabolic
reactions. A superseded system of control loops coordinates the reactions, connects
them to the genetic level, and ensures a coordinated metabolism and cell division.
The construction of the reactor and its mode of operation have to provide a proper
stream of substrates - most important are carbon sources and oxygen - to the cells
to guarantee a maximum productivity. Thus the coupling of biotic phase, liquid
phase and gas phase on one side, and the dynamics of inner phase transport and
reaction on the other side are important aspects of any bioprocess. The dynamics of
the process have time scales on rather different magnitude: quite fast phenomena
with time constants below seconds are local mixing due to turbulent flow and cata-
bolic reactions in the cells. The circulation time through the entire reactor lies in the
range of seconds up to minutes, while the growth process and regulatory adaptation
of the cells proceeds in the range of hours.
For the modeling and process design, the very fast phenomena are usually ne-
glected. This does not mean that they are not important but that it is very difficult
to develop proper measuring techniques for the fast mixing phenomena and meta-
bolic reactions, and that also their simulation requires an enormous effort. When the
cells pass along their way in the reactor through regions of different concentrations
the fast catabolic reactions may trigger a long lasting regulatory response on the
metabolism that cannot be completely understood when looking only at quantities
averaged in time or space. This complex interaction between mixing and transport,
fast metabolic reactions, and cellular growth and regulation can only be clarified on
the basis of very detailed dynamic models of the reactors and of the cells.

Modes of Operation of a Bioreactor

Depending on the flow of medium to or from the reactor, or the supply of oxygen,
the operational mode of a biotechnical processes can be classified into several
groups.

Batch Cultivation

In a batch cultivation there is no exchange of liquid medium (see Fig. 5). All the
substrates are contained in the medium from the beginning on, therefore, their in-
itial concentration is quite high. After inoculation, the cells are growing uncon-
trolled until an essential medium component is exhausted or the accumulation of
inhibiting products ceases the growth. There is a lasting change of concentrations in
the reactor. The advantage of batch cultivation is the low effort for process control. It
can be applied advantageously when high substrate concentrations have no negative
effect on the desired biological reactions. Batch cultivation is also most suited when
only relatively small amounts of the product have to be produced. In practice, a pure
14 Introduction

Initial
substrate supply Substrate inflow Substrate in flow Medium outflow

Increasing
volume
t

a Batch b Fed-batch c Continuous

Fig. 5. The principal operational modes of bioreactors.

batch cultivation is seldom found. Often some components are fed continuously, e.g.
oxygen, nitrogen source.

Fed-Batch Cultivation

In a fed-batch cultivation, medium components are continuously fed to the reactor,

but no medium is taken out. This means that the liquid volume is increasing during
the process. The fed-batch mode of operation is used when the substrate concentra-
tion must be kept low for optimum growth or product formation, e.g., in the case of
substrate inhibition or catabolite repression of product synthesis. Since the cell mass
is continuously increasing during the cultivation, the flow rate for the added med-
ium components must be adjusted properly, either by a feedback control loop or by a
predetermined flow-profile. The advantage of improved possibilities of controlling
the biological reaction by the substrate flow is opposed to the greater effort for
equipment.
Both, batch and fed-batch operation are preferred for most industrial processes, to
avoid problems with strain stability and sterility that may arise during prolonged
cultivations. The disadvantage of the previous discontinuous modes of operation is
the low productivity, mainly by two reasons: after each cultivation, the reactor has to
be emptied, cleaned and refilled again. This causes long unproductive intervals.
Furthermore, since the reactor is inoculated with low cell densities, the maximum
cell concentration and production rate is only reached close to the end of the culti-
vation.

Continuous Cultivation

In a continuous cultivation there is a permanent inflow of substrate to and outflow of

medium including cells from the reactor, usually both with the same rate so that a
steady state is reached. The cell concentration can be maintained at high values com-
parable to the maximum values in batch operation. Similar to the fed-batch mode,
continuous operation provides good possibilities to control the' biological reaction
by setting a proper residence time via the flow rate. According to the criterion for
controlling the flow rate, several types of steady continuous operation are obtained.
The most important is the chemostat that is operated at a fixed inflow rate. In the
turbidostat, the inflow rate is controlled by the measured cell concentration (e.g. via
optical density of the medium). In the pH-auxostat, the respective control variable is
Introduction 15

the pH. These control methods can stabilize the operation at high flow rates close to
wash-out of the reactor. In all cases, the outflow rate can be controlled in an under-
lying feedback loop via the liquid level or the weight of the reactor.
The advantage a continuous process is that it usually can reach the highest pro-
ductivity, since the repeated cycle of reactor preparation, cleaning, and cell propaga-
tion is not required. By using reactor cascades with a proper choice of the residence
times, the continuous process can be optimally adapted to many biological systems.
On the other hand, failures in the equipment, infection by other microorganisms, or
aging effects of the cell population limit the maximum operation time. Therefore,
and because of the higher effort for control and continuous medium preparation,
most processes are operated in discontinuous mode. Common exceptions where
continuous operation is preferred are processes that can be operated under insterile
conditions, e.g. the production of alcohols or solvents, and waste water treatment.

Cultivation with Cell Retention

In the simple continuous operation, cells are taken out from the reactor together
with the medium. For products that are not synthesized in direct coupling to the cell
growth, the productivity of the continuous operation can be increased by keeping
the biocatalyst in the reactor (types II and III in Table3). Three common methods
for cell retention are schematically shown in Fig. 6. For large scale production, the
immobilization of cells and usage of fixed- or fluidized-bed reactors is most practic-
able. For suspension cultures, cell recycle reactors with centrifugation or filtering in
the outflow, and membrane reactors with integrated retention can be applied. A low
rate of cell harvest may be of advantage, to avoid a too high percentage of old and
inactive cells.

Repeated or Cyclic Batch or Fed-Batch Cultivation

This is an extension of the discontinuous operation. Here, the cleaning-inoculation
cycle is not entered after each batch or fed-batch run. Instead, only a part of the

Cell-free medium Substrate

Cell·free
medium

Substrate Cell-free medium Substrate

a Fluidized b Membrane c Cell recycle

bed reactor reactor reactor

Fig. 6. Methods of cell retention in continuous operation: fixed or fluidized bed reactor with
immobilized cells (a), membrane reactor (b), reactor with external sparator (c).
16 Introduction

medium and cells is released and the cultivation then continued at once, eventually
after adding new substrate. By this method, the productivity is increased because the
length of the unproductive interval of the plant is reduced and the average cell con-
centration can be higher.
The previous modes of operation were concerned with the liquid phase and the
supply of substrates. Another important substance in biotechnical processes is oxy-
gen, and the following modes of operation with respect to it can be distinguished:

Aerobic Processes
In aerobic processes for the cultivation of strict or facultative aerobic microorgan-
isms, the bioreactor is supplied with air that even might be supplemented by pure

Table 3. Examples for the preferred mode of operation and kinetic type of important indus-
trial biotechnical processes

Product Organism Mode of operation Type"

Substrate Oxygen

Ethanol Saccharomyces cerevisiae, continuous or anaerobic/

Zymmomonas mobilis (repeated) batch microaerobic
Acetone/Butanol Clostridium acetobutylicum continuous anaerobic
Baker's yeast Saccharomyces cerevisiae fed-batch aerobic
Penicillin Penicillium chrysogenum fed-batch aerobic III
Cephalosporin Acremonium chrysogenum fed-batch aerobic III
Tetracycline Streptomyces strains fed-batch aerobic III
Alkaline protease Bacillus lycheniformis batch/fed-batch aerobic II
Amylases Aspergillus oryzae batch/fed-batch aerobic II
Pectinases Aspergillus niger batch/fed-batch aerobic II
Cellulases Trichoderma resii batch/fed-batch aerobic II
Invertases Saccharomyces cerevisiae batch/fed-batch aerobic
Lactic acid Lactobacillus bulgaricus batch anaerobic
Propionic acid Propionibacterium shermanii batch anaerobic/
micro-aerobic
Acetic acid Gluconobacter suboxidans batch aerobic
Citric acid Aspergillus niger batch aerobic II
Polysaccharides Xanthomonas campestris batch aerobic
Amino acids Corynebacterium glutamicum batch aerobic II
Single cell protein Candida utilis, S. cerevisiae, batch/fed-batch aerobic
Methylophilus methylotrophus
Riboflavin Eremothecium ashbyi batch aerobic II
Vitamin B12 Pseudomonas denitrificans, batch anaerobic/ II
Propionibacterium aerobic micro aerobic
Alkaloids Claviceps paspali batch aerobic III
Recombinant Saccharomyces strains batch/fed-batch aerobic
proteins Escherichia coli
Immunglobulines Hybridoma cells continuous aerobic III
monoclonal
antibodies
Interferones Mammalian cells continuous aerobic III

" according to the classification of [19], see also chapter 2

Introduction 17

oxygen. This supply normally has to cover the entire oxygen demand of the cells,
since the solubility of oxygen in aqueous solutions is very low. In a typical aerobic
process, e.g. the production of baker's yeast, low concentrations of oxygen have ne-
gative effects on the productivity. Therefore, if possible one tries to keep the concen-
tration of dissolved oxygen at a relatively high percentage of the saturation value to
avoid local oxygen limitation in less mixed regions of the reactor. In a large produc-
tion tank, the oxygen supply is often the limiting factor for the attainable cell mass
and the costs for aeration are a great part of the total production costs. Here, by
economic reasons one has to accept low oxygen concentrations, or even oxygen lim-
ited conditions.

Anaerobic Processes

In anaerobic processes it is tried to keep the reactor oxygen-free. This is essential for
strictly anaerobic microorganisms where oxygen is toxic. But also in the cultivation
of facultative aerobic microorganisms, anaerobic operation may lead to the highest
productivity or substrate turnover. An example is ethanol production by yeast. In the
anaerobic case, the gas phase of the reactor consists mainly of the gaseous products
of metabolism, mostly carbon dioxide.

Micro-Aerobic Processes

Some microorganisms (e.g. Propionibacteria) tolerate only very low oxygen concen-
trations, or the production rate is maximum under defined oxygen limited condi-
tions (e.g. for Myxobacteria). In these cases one has to limit the oxygen transfer to
the reactor by controlling the aeration rate, the stirrer speed, or the inlet oxygen
concentration, for example by mixing air with nitrogen. This mode of operation is
called micro-aerobic. Sometimes, typical anaerobic processes are also operated un-
der micro-aerobic conditions. For the ethanol production, the viability of the yeast
and in turn the productivity can be improved in this way.
Several examples for important industrial biotechnical processes and their modes
of operation are given in Table 3. Selected processes are dealt with in detail in Parts
C and D of this book.

References
1. Schiigerl K (1998) Bioreaction Engineering III. Wiley, Chichester, Amsterdam
2. Rehm HJ, Reed G (1991) Biotechnology, vol 4. VCH, Weinheim
3. Bastin G, Dochain D (1990) On-line Estimation and Adaptive Control of Bioreactors,
Elsevier, Amsterdam
4. Dunn IT, Heinzle E, Ingham J, Prenosil JE (1992) Biological Reaction Engineering. VCH,
Weinheim
5. Moser A (1988) Bioprocess Technology: Kinetics and Reactors. Springer, Berlin Heidel-
berg New York
6. Roels JA (1982) J Chern Tech Biotechnol 32:59
7. Sonnleitner B, Cheruy A (1997) J Biotechnol 53:173
8. Sonnleitner B (1997) J Biotechnol 52:175
9. Bailey JE (1998) Biotechnol Prog 14:8
10. Siimes T, Linko P, von Numers C, Nakajima M, Endo I (1995) Biotechnol Bioeng 45:135
11. Konstantinov KB, Yoshida T (1992) Biotechnol Bioeng 39:479
12. Glassey J, Ignova M, Ward AC, Montague GA, Morris AJ (1997) J Biotechnol 52:201
18 Introduction

l3. Montague G, Morris J (1994) Tibtech 12:312

14. Moser A (1993) Chern Biochem Eng Q 7(1):21
15. Bailey JE, Ollis DF (1986) Biochemical Engineering Fundamentals. McGraw-Hill, New
York
16. Royce PN (1993) Crit Rev Biotechno!l3(2):179
17. Schiigerl K (1991) Bioreaction Engineering 1. Wiley, Chichester, Amsterdam
18. Schiigerl K (1992) Bioreaction Engineering II. Wiley, Chichester, Amsterdam
19. Gaden, EL (1959) J Biochem Microbio! Techno! Eng 1(4), 413
 Part A
General Principles and Techniques
1 Bioreactor Models
Karl Schiigerl, Karl-Heinz Bellgardt

1.1
Introduction

A common submerse bioreactor consists of a vessel provided with a mechanically

moved agitator for power input, which is necessary for the intensification of the
transfer processes. Various types and sizes of impellers are applied for agitator de-
pending on the properties of the biological system, the cultivation medium, and the
size of the reactor. The exact description of the fluid movement by a simple model is
not possible because the main liquid flow caused by the stirrer is overlapped by
turbulence fluctuations. The situation is made more complex by the presence of
two or more phases. The accurate description of the biological, chemical and physi-
cal processes and their interrelation in stirred tank (ST) reactors is impossible, and
therefore considerable abstraction is necessary. The same holds true for other sub-
merse reactors, such as the bubble column (BC) and airlift tower loop (ATL) reac-
tors, in which the power input is accomplished with the expansion of compressed gas
and with a liquid pump, respectively. The prerequisite for the abstraction of the re-
actors is the knowledge of the influence of various parameters on the process perfor-
mance. Therefore, in Sect. 1.2 the interrelation between the growth of microbial and
animal cells and their physical and chemical environment is considered. Experimen-
tal investigations allow the description of the physical processes in the reactors,
which are discussed in Sects. 1.3.1 and 1.4.1. The description of a simplified fluid
dynamics in BC and ATL reactors is possible and considered in Sect. 1.4.2, in con-
trast to ST reactors, which is discussed later in Sect. 7.2. In Sects. 1.3.2 and 1.4.3
examples are given for the most common mathematical models and in Sect. 1.5 the
connection is made to chapters 7 and 8 of this book.

1.2
Interrelations Between the Cells and Their Physical/Chemical Environment

The momentum, mass, and heat transfer processes have to be intensified by turbu-
lence in the cultivation medium of high performance submersed reactors. High mo-
mentum transfer is needed for dispersion of the gas phase and increase of the inter-
face between the gas and liquid phases in an aerated reactor. A high turbulence is
necessary for high mixing intensity, i.e., quick dispersion of nutrients, dissolved oxy-
gen, and acid/base added to the reactor, as well as for enhancing the exchange rate of
gaseous and volatile components across the interface of the phases. The improved
heat transfer is needed for uniform temperature in the reactor and to remove the
22 1 Bioreactor Models

heat produced by the living microorganisms and by the turbulent energy dissipa-
tion. Oxygen limitation of the growth is caused by low specific gas-liquid interfacial
area and low exchange rate across the interfacial area. Non-uniform distributions of
nutrient and dissolved oxygen in the reactor due to low mixing intensity can cause
strong local concentration fluctuations, inhibition of the growth at high nutrient
concentrations, and growth limitation at low nutrient and oxygen concentrations,
respectively. Large pH fluctuations can impair the physiology of the cells as well.
Temperature control can be a problem at low heat transfer rates. However, at high
stirrer speed and excessive power input the cells can be damaged.
On the other hand, high cell density, substrate concentration, and highly viscous
products, respectively, increase the viscosity of the cultivation medium and reduce
the intensity of the turbulence and the rate of transfer processes. At low energy dis-
sipation rate, large bubbles with short contact time are formed and eventually wall
growth can occur. In addition, at low turbulence intensity local growth limitation
and non-stirred dead water regions with exhausted nutrient and dissolved oxygen,
zones appear in which cell lysis occurs.

1.3
Stirred Tank (S1) Reactors

1.3.1
Description of the Physical Processes in the Reactors

All agitator types (with axial as well as with tangential/radial main flow) cause cir-
culation flow patterns in the reactor below and above the impeller plane, if the liquid
rotation is hindered by baffles. This main flow is overlapped by turbulent fluctua-
tions. Therefore, the flow path is chaotic and the circulation time is a stochastic
quantity. In aerated liquid, the bubbles are dispersed and re-dispersed by the stirrer.
Large bubbles quickly escape from the reactor, and small bubbles are dragged along
with the liquid and recirculated. The higher the bubble recirculation rate, the higher
is the gas hold up. At high cell concentrations, the oxygen consumption rate is high,
oxygen is gradually exhausted in the bubbles along the recirculation loop. The oxy-
gen concentrations in the gas and liquid phases depend on the local oxygen transfer
rate (OTR) from the gas into the liquid phase, the age distribution of the bubbles in
the reactor, and the oxygen consumption rate of the cells.
Only in the stirrer region is the oxygen transfer rate high. Some elements quickly
return to the stirrer with fairly high oxygen concentration, others return later with
low oxygen content, because the oxygen is consumed during the passage of the liquid
elements along the different recirculation paths in the reactor, but only a low amount
of oxygen is transferred from the gas phase into the liquid during this passage. The
nutrient is added to the reactor close to the stirrer. It has to be quickly distributed to
avoid local growth inhibitions and limitations, respectively. Its distribution in the
reactor can be described by the liquid circulation as well. The nutrient concentration
in the liquid depends on its consumption rate and the recirculation time distribu-
tion. Therefore, the process performance is influenced by the nutrient and oxygen
consumption rates, the volumetric mass transfer coefficient, and the circulation time
distributions of the phases.
1.3 Stirred Tank (ST) Reactors 23

To compare the key variables, they are converted into quantities with common
dimension: characteristic times [1). This comparison indicates that the mass transfer
processes and the oxygen consumption rate are the key process variables. The spe-
cific power input (PolV), the volumetric gas flow (QIV), and the cell concentration
are the main control variables which influence these process variables.

1.3.2
Reactor Models

The ideally mixed continuous stirred tank reactor (CSTR) model assumes that in the
reactor no local variation of the process variables occur. Therefore, the reactor can
be described by position independent (lumped) parameters. Prerequisites are the
ideally mixed gas and liquid phases, which can be tested by the determination of
the residence time distributions (RTDs) of the phases in continuous reactors and
by measuring-mixing time in batch reactors with global stimulus/response techni-
ques. This model usually holds true only for small laboratory reactors with low visc-
osity cultivation media.

1.3.2.1
Model for the Ideal Stirred Tank Reactor

The reactor model dynamically describes the concentrations in the gas and liquid
phase of the reactor in dependence on initial conditions, manipulating variables,
and biological reactions of the microorganisms. For simplicity it is assumed that
both phases are ideally mixed and impulse balances need not be considered. There-
fore, the corresponding sub-models include only ordinary differential equations. A
third sub-model describes the mass exchange between gas and liquid phase. For the
ideal stirred tank reactor, the formal balance [2, 3)

storage element convective transport diffusive transport

~ ,---_-'A~_"""

global change global , global ' global

in the reactor inflow - outflow inflow - outflow + reaction

yields ordinary differential equations as given below. A schematic diagram of a stir-

red tank reactor and the employed model variables are shown in Fig. 1.1. The reactor

Gas phase (head) VGH

.................. ··A~ Gas outflow

FG O xjo TO pO

~ Gas

nr~:IS~j~~~;1
Reactor Liquid phase phase
VL h Cj : (dispersed)

VR ---~-~-~~-w" l
FOCj

Fig. 1.1. Schematic diagram of a stirred tank reactor and related model variables
24 1 Bioreactor Models

volume can be divided into three parts: liquid volume, VL; volume of the dispersed
gas phase, VGD; and volume of the head space gas phase, VGH:
(1.1)

VG = VGD + VGH (1.2)

which are all described by corresponding balances. By using the relative gas (cG) and
liquid fractions (CL), the volume can be expressed as
(1.3)

(1.4)
Mostly, the difference between dispersed and head gas phase is neglected, and
both are considered as a sole gas volume.

Liquid-Phase Model

Here, the mass balances for a general compound i shall be considered, which are

dmd;(t) = m;(t) = pI (t)mf(t) - pO(t)m;(t) + m?(t) + (1.s)

t ~ ~ '-v-"
inflow outflow reaction gas-liquid transfer

The final balances shall be given in terms of concentrations, which are related to
mass by
m;(t) = Vdt)C;(t) (1.6)
and therefore
dm;(t) = VL(t) dC;(t) + c-(t) dVdt) (1.7)
dt dt' dt
Introducing these substitutions, together with the volumetric rate of biological
reaction, Qi (a positive value means production of compound i), and of mass ex-
change gas-liquid, iTR, into the above balance, Eq. (1.5), gives the model equation
of the liquid phase:
dC;(t) _ pI(t) R() _ pO(t) .() _ C;(t) dVdt) .() . ()
d t - VL() + Q, + zTR (1.8)
t C; t VL(t ) c, t VL()t dt t t

where the change of volume is

dVL(t) = pI(t) _ pO(t) (1.9)

dt
By substitution of this into Eq. (1.8), together with the definition of the dilution
rate

D(t) = pI(t) (l.l0)

Vdt)
the concentration balance becomes for a compound i

+,
dC·(t)
= D(t) (cf(t) - C;(t)) + Q;(t) + iTR(t) (1.11)
1.3 Stirred Tank (ST) Reactors 25

Table 1.1. Modes of operation of a bioprocess with respect to the liquid phase

Mode pI pO D VL

Batch 0 0 0 constant
Ped-batch >0 0 pl/V L increasing
Chemostat >0 =pI pI/V L constant
Continous >0 >0 pl/V L variable

This can be written for a number of compounds in vector notation as

dC(t)
dt = D(t) (CR(t) - C(t)) + Q(t) + TR(t) (1.12 )

The different modes of operation of a bioreactor, depending on the rates of inflow

to and outflow of the reactor, are summarized in Table 1.1. The only difference be-
tween fed-batch and continuous operation is that for the first there is no outflow of
medium, FO=O [4,5]. For multistage reactor cascades one has to set [6]

Cij = Cij-l, C~ = Cf, j = stage index

The model equations for usual compounds are summarized in Table 1.3. In the
balance of dissolved gases (d and e), the transport term by liquid exchange,
D(t)[CiR(t) - Ci(t)], is usually neglected because it is small compared to Qi and iTR.

Gas-Phase Model

The gas-phase model is an important part of any model for a biotechnical process, if
oxygen limited growth has to be considered, or measurements in the exhaust gas are
used for process control. The gas phase can be further divided into dispersed gas
phase and head space of the reactor, which are connected in series. The balance
equations for both are identical, but with different parameters. Here, only the gen-
eral form is given. The main components of the gas phase are oxygen, carbon diox-
ide, nitrogen, water (i=Oz, COz, N z, HzO), and additional gaseous products or sub-
strates; they are advantageously described by their mole fractions, Xo, xc, XN, and xw,
respectively, which can be easily measured. To begin with, the balance for the num-
ber of particles is

(1.13)
inflow outflow gas-liquid transfer

The volume or mole fractions are

x;(t) = n;(t) = V;(t) = p;(t) (1.14)

ntot(t) VG(t) PG(t)
By using the gas law

(1.15)
26 1 Bioreactor Models

the particle number becomes

pYG
ni(t) = RT Xi(t) (1.16)

pYGdXi(t)
--- ( 1.17)
RT dt
With V=P the particle flow at gas inlet can be expressed as
.I() pIxf(t) I rJ I( ) (1.18)
ni t = ~PG = ~MYLXi t

where the inlet flow FG1 is often specified under normal conditions, i.e.,
·I
ni =
pWx;
yN ' VN = 22.5 lmole
-1
(1.19)

The molar volumetric flow rate as used in Eq. (1.18) is defined by

PPG
Ck = RTYL (1.20)

Similarly, it follows for the outlet stream:

OxO(t)
iz0(t)
, = P_'_po
RTo G = QOM YL XO(t)
, (1.21)

and the molar exchange flow with the liquid phase is

izTR(t) = iTl~~;YL (1.22)

where the positive direction of the transfer flux is defined into the liquid phase.
Using the above substitutions, Eqs. (1.18), (1.21), and (1.22) in Eq. (1.13) gives the
differential equations for the mole fractions in the gas phase
dxf(t) pITOP&(t) I() Pg(t) O() RTOyL . ()
--= x· t - - - x · t - ITR t (1.23)
dt pOTIyG ' YG ' MiPoYG
The resulting outlet gas stream, FG 0, is usually not measured but it can be calcu-
lated since the summation of all mole fractions must equal one,

(1.24)

By summing up the above differential equations, Eq. (1.23), for all compounds it
follows
d1 = 0 = pI TOP&(t) 1 _ Pg(t) 1 _ RTOyL"" iTR(t)
( 1.25)
dt pOTIYG YG pOYG ~, Mi
and the outlet flow can be calculated as

(1.26)
1.3 Stirred Tank (ST) Reactors 27

Table 1.2. Modes of operation of the gas-phase

Mode Aeration Inlet mole fractions required balances

Aerobic, aeration by air xo'=O.2096 Oz, CO 2 , Nz,

Xc'=O.0003 Water
xN'=O.7901
xw'=O
Aerobic, aeration by oxygen Xo'=l Oz, CO 2, Water
Xc'=O
xw'=O
Anaerobic, no aeration CO 2 , Water
Anaerobic, aeration by nitrogen FG'=F GN ' xN'=l CO 2 , Water
Xc'=O
Xw'=O
Additional gaseous products i X/=O products i (in
in any mode addition)

Using this expression in Eq. (1.23) leads to the final model equations for the gas-
phase

Some special cases for the modes of operation of the gas phase are summarized in
Table 1.2. A further simplification can be obtained by considering inertial compound
as being in quasi steady-state. This is justified because for such compounds there is
no reaction, Qi=O, which results in a very small mass transfer only when the concen-
trations in the gas phase are changing. The error is very small if one assumes that at
any time the inflow of inertial compounds equals the outflow of the reactor, and thus
neglects completely the balance equations. The mole fractions of the inertial com-
pounds can then be eliminated from the remaining equations by using Eq. (1.24).
In many situations the gas phase can be taken as quasi-stationary, since its resi-
dence time is much lower compared to the liquid phase. This means mass storage
(d/dt) can be neglected and, therefore, the reaction rates equal the transfer rates.
This method can also be used for the calculation of characteristic biological para-
meters, such as oxygen uptake rate, OUR, carbon dioxide production rate, CPR, and
respiratory quotient, RQ, by only using exhaust gas measurements. From Eq. (1.27) it
follows with d/dt=O:

(1.28 )
28 1 Bioreactor Models

If only the compounds i=(Oz, COz) need to be considered, the "measured" oxygen
uptake rate becomes after some calculations
I ~(1 - xg) - x8(1 - x~)
OURm = -QOm ~ OTRm ~ QM 0 0 M02 (1.29)
"-v-' 1- Xo - Xc
F~
Vi'TVr
and the "measured" carbon dioxide production rate
Ai 4(1 - x8) - xg(l - xlo)
CPRm = QCm ~ -CTRm ~ -l.l.M 0 0 MC02 (1.30)
1- Xo - Xc

The respiratory quotient is an important parameter for evaluation of the growth

kinetics:
CPR M02
RQ=-~ (1.31)
OURMc02
This can be approximated by the measured respiratory quotient
4(1- x8) - xg(1- xb)
RQm = ~ RQ (1.32 )
xb ( 1-xg ) -x8 ( 1-x~ )
when substituting Eqs. (1.29) and (1.30) into Eq. (1.31).

Mass Transfer Gas-Liquid

The performance of aerobic cultivations with the need for continuous supply of oxy-
gen is influenced significantly by the mass-transfer phenomena gas-liquid. In large
production scale reactors mostly the oxygen transfer to the liquid phase is the limit-
ing factor for the overall productivity. Many operating variables can effect the oxy-
gen transfer in bioreactors, and the interactions between these variables are complex:
bubble size distribution, medium composition, oxygen solubility, cell concentration,
temperature, viscosity, power input [7, 8]. During a cultivation, all factors cannot
usually be measured and analyzed separately. Therefore the complete modeling is
quite difficult, and only a rough correlation with limited predictive power can be
derived. The volumetric mass transfer coefficient kLa describes the nature of the
mass transfer within the reactor and serves as an index of its performance. The
entire transport pass, e.g., of oxygen from the gas phase to the microorganisms, in-
volves a number of steps in series and in parallel. The overall mass transfer is deter-
mined by several factors:
1. Gas film resistance for transport between the bulk of the gas phase and the gas-
liquid interface. It can be neglected compared to the resistance on the liquid side
2. Resistance of the gas-liquid interfacial area and liquid film resistance for transport
to or from the bulk liquid phase. This is usually the major limiting step
3. Resistance within the bulk of the liquid phase. This is usually neglected due to
eddy diffusion
4. Liquid film resistance around the cell and in the cell boundary. This can usually be
neglected
5. Transport resistance through aggregates of cells. For flocculating or pellet-forming
microorganisms, this can be a main limiting factor
6. The intracellular resistance with parallel reactions is very difficult to be quantified
1.3 Stirred Tank (ST) Reactors 29

Beside these steps in series there can be a parallel path by direct contact of cells with
the gas-liquid interface. The adherence of the microorganisms at the bubble surface
enhances the overall oxygen transfer.
The mass transfer rate between gas phase and liquid phase is proportional to the
concentration gradient in the interfacial area and to the volumetric mass transfer
coefficient. For oxygen and carbon dioxide the major mass transfer resistance is
located at the liquid side and the mass transfer, either by adsorption or desorption,
can be calculated by the single film model [9]:
iTR(t) = kiLa(C; - Ci ) (1.33)
where * indicates the saturation concentration in the gas-liquid interface. The trans-
fer flux is proportional to the driving force, the concentration difference film-bulk
liquid, and to the product of mass transfer coefficient, kL' and specific interfacial
area,
A
a=- (1.34)
VL
where A is the total contact area between gas and liquid phase. The saturation con-
centration for the dissolved gases depends on the mole fractions in the gas phase.
After Henry's law,
( 1.35)
the mole fraction in the liquid phase at saturation is proportional to the partial
pressure in the gas phase. The mole fraction in the liquid phase can be approximated
by
ni ni
XiL=-;;::;;;- (1.36)
ntot nw
With
mi
ni=- (1.37)
Mi
it follows
mi MWVL Ci Mw
XiL = - - - - = - - - (1.38)
mw Mi VL Pw Mi
and with the mole fraction of the gas phase as

Xi=-
Pi (1.39)
P
the saturation concentration is finally

C* = HiPWMiP XO = C'(P)x O (1.40 )

1 Mw ! ! !

It depends on the operating pressure of the reactor and on the hydrostatic pressure
of the liquid phase. For small reactors, p=po can be taken, while for very high reac-
tors the pressure at half of the reactor height can be used as an approximate average
value. C[ is the saturation concentration for a gas phase of a pure compound. For
oxygen and carbon dioxide, these values are about Cb;;::;;;37 mg r l and Ch;;::;;;1400 mg r l
at 37 DC and 1 bar, but these also depend on the ion concentration in the medium. In
30 1 Bioreactor Models

biotechnology, the concentration of dissolved oxygen is often specified as percentage

of the saturation value:
Co
pOz = - ,-[ 100% (1.41 )
Coxo

The mass transfer resistance depends on the hydrodynamics in the reactor and the
physical properties of the fermentation broth. Of special interest for further model-
ing is a correlation of the mass transfer coefficient to the molecular diffusion con-
stant, which connects the parameters for different substances [10]:
(1.42 )
The exponent can be taken in the range r1=0.5 (Higbie's model for small bubbles)
to r1=0.66 (model of Dankwerts for large bubbles). With this equation the volumetric
mass transfer coefficient for a substance i can be calculated, while that for oxygen is
known as
Di)ri
(Do (1.43 )
kLi = kw

The specific interfacial area, a, depends on the geometry of the system, the type of
the stirrer, the power input, and the aeration rate [11-l3]. The following correlation
can be used [7]:

VL (p)rz (1.44)
a F~ = r3 F~

For stirred tank reactors the specific power input is a function of the stirrer speed;
in bubble columns it is determined by the expansion energy of the pressurized gas-
inflow. This gives a final correlation, including Eqs. (1.43) and (1.44), of

(1.45)

The parameters r1 to r4 are difficult to predict and must be identified from experi-
mental data. This correlation, together with Eqs. (1.11) and (1.33), gives the final
model equation for dissolved oxygen. For the modeling of dissolved carbon dioxide,
one has in addition to consider its dissociation according to [14-16]:

(1.46)
Only the free, not dissociated, COz contributes to the driving force for gas-liquid
exchange. Bicarbonate can be mostly neglected. Nevertheless, under usual condi-
tions for bioprocesses (pH=6.8 and T=25°C), 75% of total carbon dioxide may be
dissociated. From Eq. (1.46), the total carbon dioxide can be calculated as
( 1.47)

The dissociation reaction is very fast and always in the equilibrium given by

[HC0 3] . [CO~-] = Kc
( 1.48)
[Cdfree
1.3 Stirred Tank CST) Reactors 31

Substitution of this equation into Eq. (1.47) leads to

(1.49)

and

[C] = [Cc]tot[H+] (1.50)

C free [H+] + Kc

where the proton concentration can be expressed as

(1.51)

This leads to the balance of total carbon dioxide as shown in Table 1.3, Eq. (e).
The lumped parameter models renounce the structure of the flow and consider
only the averaged process variables. This allows to use global relationships between
the process and control variables. Another typical example for the relationship of the

Table 1.3. Complete example for a CSTR model without inflow of cell mass and product

Substance Balance equation

a) Substrate dC;;t) = D(t) (cf(t) - Cs(t») + Qs(t)

b) Cell mass dC;?) = -D(t)Cx(t) + Qx(t)

c) Product

OTR(t)
dCo(t) ,
----:it =
A ,

d) Dissolved oxygen Qo(t) + kwa(Coxo(t) - Co(t»

CTR(t)

e) Dissolved (total)
carbon dioxide
dCc( t)
- d - = Qc(t)
t
+ ,kw a (Dc)
-
Do
T1 (
Ccxc(t) -
10
lO-pH
_H
p + Kc cc(t)
)'

f) Oxygen in the dxo(t) = pITOF~(t)

dt pOTI VG
(x!a (t) - Xo
(t»)
gas phase + RToVL (CTR(t) x (t) _ OTR(t) (1 - x (t»)
pOVG Me C Mo a

g) Carbon dioxide in the dxc(t) _ pIToF~(t) ( I (t) _ (t»)

dt - pOTIVG Xc Xc
gas phase + RToV L (OTR(t) X (t) _ CTR(t) (1 - x (t»)
pOVG Mo a Me C

The superscript 0 for the outlet mole fraction was omitted in Eqs. d) to g) for simplicity
32 1 Bioreactor Models

volumetric mass transfer coefficient as a function of the control variables in low

viscosity media is

(l.52)

which was developed by Schluter et al. [17] for Trichosporon cutaneum cultivation
medium. Here the constants are: r4=7.94x 10-4 , rl=0.62, and r2=0.23 for the flat
bladed disc turbine, and r4=5.89x 10-4 , rl=0.62, and r2=0.19 for the INTERMIG im-
peller.
In highly viscous non-Newtonian media a more complex relationship is valid. E.g.,
Kawase and Moo-Young [18] recommended the following relationship for xanthan
production cultivation media, the validity of which was confirmed by Herbst et al.
[19]:

kLa = 0.675Do.s
p3/S{PO/Vp}[9+4n j /[I+n j {u ~
}{7)ff} (l.53)
7)w
_e_
m {k/p}[l+nj/2a3/4 UB

Here uG=uB=0.265 m s-l=bubble rise velocity, lleff=effective viscosity of the culti-

vation medium and llw=viscosity of water.
A large number of similar relationships have been recommended by various
authors (for review see [8]).
By means of the relationship
(1.54)

the growth rate of the cells Ox can be calculated by the measured volumetric mass
transfer coefficient kLa and the dissolved oxygen concentration Co in the reactor if
the yield coefficient of the growth with respect to the oxygen consumption YXIO is
known and the oxygen saturation at the interface is assumed.

Recirculation and Compartment Models

Recirculation time models assume the oxygen uptake in the stirrer region (micro
mixer) and the oxygen consumption along the circulation loop (macro mixer) ac-
cording to Bajpai and Reuss (Fig. 1.2) [20, 21]. In batch reactors the fluid elements
enter from the micro mixer into the macro mixer and stay here for different periods

a IlItr~nia, i;t" ",

TCltulWlf: dClllcna
4 _.

· · ·~i;iI;:""
Micro · HacromiuJ

,~"...,...... ~
miur

t~ 1/9
Oisl ri bul ion
and
conversion
of mau
ur \ IIIt
",,,,.
llin" 1I1lbf,.iCIQl"litu U

~ 1~r G"l--_ _- ;

1n~.llo.lplJl Dr \. tlunl/
\.
~
~

l~bLLr.l .. InHGf ....... -

Val J~t tt~tat
~rodutl
III ,~e UfQ

Fig. 1.2. Recirculation model of Bajpai and Reuss [20,21]

1.3 Stirred Tank (ST) Reactors 33

Fraction (I-q)
(2)
Fraction q

Fig. 1.3. Two compartment model of Sinclair and Brown [22]

of times before they return into the micro mixer. In continuous reactors, part of the
liquid elements returns into the micro mixer, another part leaves the reactor.
In batch reactors, the recirculation time distribution can be determined by stimu-
lus/response technique or by flow follower. In continuous reactors, the recirculation
time is measured by a flow follower and the share of the fluid elements which recir-
culates and which leaves the reactor by stimulus response technique.
Only small bubbles recirculate in the cultivation medium many times. With in-
creasing bubble size, the frequency of recirculation decreases and large bubbles do
not recirculate at all, but leave the reactor immediately. Their residence time in the
macro mixer is very short. Therefore, the distribution of residence time of the gas
phase in the macro mixer is usually broad and depends on the bubble size distribu-
tion.
A two-region mixing model was recommended by Sinclair and Brown (Fig. 1.3)
[22].(1-q) is the share of the fluid which does not enter into the recycle-region 2.
The share q enters into the recycle-region 2; it recycles n times, before it returns into
the region 1 and leaves the reactor. In the reactor with a single impeller, the weight-
ing function of the reactor f(t) is given by the convolution (®) integral:

L
n=oo
f(t) = (1 - q) qnf1(t) ® (f1(t) ®fi(tt* ( I.SS)
n=O

where
f1(t) is the weighting function of region 1
f2 (t) is the weighting function of region 2 without recycle
f1 (t) ® f2 ( t) is the weighting function of a single recycling
(f1(t) ® f1(t))n* is the weighting function ofn-recycling through regions 1 and 2.

In reactors with m-stage impellers, each of the impeller regions can be described
with a convolution integral with n-recycling in the m-regions.
Oosterhuis and Kossen [23] published a model for reactors with a two-stage im-
peller system and Bader [24, 2S] for reactors with a multistage impeller system.
These models were extended by Singh et al. [26] and by Mayr et al. [27].
Ragot and Reuss [28] developed a multiphase compartment model, which consid-
ers a recirculation in the liquid and gas including the mass transfer between the
phases. This model is discussed in chapter 7 of this book.
Another model assumes three regions for the gas phase in a stirred tank reactor
with a single impeller (Fig. 1.4) [29]:
1. The impeller region with gas re-dispersion (micro mixer), region 2
2. The recirculation region of small bubbles with long residence time (macro mixer),
because the small bubbles follow the liquid flow pattern), region 1
3. the recirculation region of large bubbles with short residence time, region 3
34 1 Bioreactor Models

Gas phase dynamics in stirred tonks

Degassing

Impeller region

Aeration -===..J
Fig. 1.4. Schematic display of the regions of gas dispersion and exchange (region 2), stagnant
gas hold up of small bubbles (region 1) and dynamical gas hold up of large bubbles (region 3)
of Ruffer et al. [29].

The RTD probability function is given by

(1.56a)

with
A = wa(51 + d) B = wa(52 + d) c = wa(53 + d) (1.56b)
(51 - 52)(51 - 53) , (5 - 5d(52 - 53) , (53 - 5d(53 - 52)
Here s=Laplace transform variables, SI for region 1 (sl=-l/'td, S2 for region 2
(S2=-1/1:2)' and S3 for region 3 (S3=-1/1:3) and 1:1> 1:2' 1:3 the mean residence times in
regions 1, 2, and 3; a=exchange rate from region 2 to region 3; d=exchange rate from
region 1 to region 2, and w=gas throughput.
In a 1 m 3 stirred tank reactor with penicillin V production by Penicillium chrY50-
genum, the volume share of the regions varied with the time. The average values
were: 64% (region 1), 33% (region 3), and 3% (region 2).
More detailed calculations were carried out by several authors [30-40], which are
discussed in chapter 7 of this book.

1.4
Bubble Column (BC) and Airlift Tower Loop (All) Reactors

1.4.1
Description of the Physical Processes in the Reactors

In bubble column and airlift tower loop reactors the power input is accomplished
with the expansion of compressed gas across a gas distributor or in various two-
phase nozzles [8, 41, 42]. The bubbles rising in the column cause turbulence and
mixing. They playa more important role in these reactors than in stirred tank ones.
On account of the buoyancy forces of the bubbles, the bubble swarm moves upwards
in the column. The same is valid for the riser of ATL reactors. In the down comer of
ATL reactors with external loop, only the small bubbles (with buoyancy forces less
than their resistance forces) are dragged along with the liquid and therefore the gas
1.4 Bubble Column (BC) and Airlift Tower Loop (ATL) Reactors 35

Fig. 1.5. Multiple circulation cells in a bubble column according to Joshi and Sharma [43]

bubbles follow the liquid flow with nearly plug flow character. The separate descrip-
tion of the riser and the down comer of ATL reactors with internal loop is only
possible for large units.
In a batch Be, the bubbles rise in the column center and transport the liquid up-
wards. Close to the wall, the liquid moves downward and drags along the small bub-
bles. By this means, multiple liquid circulation patterns consisting of axially superim-
posed circulation cells with diameter to height ratio of unity are formed (Fig. 1.5) [43].
In ATL reactors the flow pattern in the riser depends on the transport capacity of
the down comer. No back flows close to the wall are formed in the riser if the liquid
transported to the top can return to the bottom through the down comer. At low
transport capacity of the down comer, back flow and circulation patterns are formed
similar to BC reactors.
With fed-batch operation, the liquid volume increases during the cultivation. The
ATL reactor is already laid out to maintain the liquid circulation through the riser-
down comer system at the beginning of the cultivation. A head part is installed to
take up the increasing medium volume. Such reactors consist of a riser, down comer,
and a head at the top of the column.

1.4.2
Flow Models

Three flow regimes can be distinguished in BC and ATL reactors. At low superficial
gas velocity homogeneous (bubbly) flow prevails in columns of broad diameter
36 1 Bioreactor Models

range. The flow in down comer is always in this bubbly flow regime. By increasing
superficial gas velocity above the flooding point large bubbles formed by coalescence
are stabilized by the wall and fill the entire column diameter in laboratory columns,
the homogeneous flow changes into slug flow. In large columns, large bubbles are
formed by coalescence and bubble clusters move predominantly at the center line
of the column and rise in a swarm of small bubbles with high velocity. The bubble
size and velocity distribution are broad in this heterogeneous churn-turbulent flow,
which is characteristic for industrial BC and in ATL reactors. The slip velocity Us
[44] and the drift flux FDF [45] are only valid for bubbly flow:
UG UL
Us = - ± - - , FDF = usEG(l - EG) = uG(l - EG) ± uLEG ( 1.57)
EG 1 - EG
where minus and plus signs represent the co-current and counter-current flows. The
model of Zuber and Findlay [46] applies for both bubble and churn-turbulent flow:

(1.58)

where UGD is the volume fraction weighted average drift velocity of the gas phase and
UBT is the bubble terminal velocity, which is a function of the bubble size. Co is a
constant.
Ueyama and Miyauchi [47] published a relationship for the flow pattern. Joshi and
Sharma [43] developed a multiple-cell circulation model (Fig. 1.5). The average cir-
culation velocity Ue is given by

Uc = 1.31 [gDs (UG ± 1E~~G - EGUBT) ] 1/3, (1.59)

and for the liquid phase axial dispersion coefficient DL :

(1.60 )
was obtained, which is in close agreement with the experimental results.
This relationship indicates that there is a direct relationship between the liquid
circulation velocity U e and the axial dispersion coefficient DL in the liquid phase.
According to Eq. (1.60 ), the superficial gas and liquid velocities, UG and UL, the gas
hold up EG , and the column diameter Ds influence the process performance.
Verlaan et al. [48] used the drift flux model of Zuber and Findlay to model an ATL
reactor.

1.4.3
Reactor Models

The ideally mixed BC and ATL reactor can be described by position independent
(lumped) parameters. These prerequisites are fulfilled if the variation of the process
parameters along the column (in BC) and riser and down comer (in ATL with inter-
nal loop) can be neglected. In continuous reactors the distribution of residence
times is measured by stimuluslresponse technique. The ideal mixing in the liquid
phase is assumed if it follows an exponential function. This model can only be ap-
plied for small laboratory reactors with internal loop and low viscosity cultivation
medium. It is not suitable for describing the behavior of tall ATL reactors and reac-
tors with external loop, because of the low intensity of the axial dispersion in liquid
1.4 Bubble Column (BC) and Airlift Tower Loop (ATL) Reactors 37

phase of the down comer. The flow has nearly plug flow character in the down co-
mer.
The lumped parameter models allow the evaluation of the relationships between
the process and control variables, e.g., for the volumetric mass transfer coefficient of
oxygen. A typical example was developed by Akita and Yoshida [49] and confirmed
by Kataoka et al. [50] in a 5.5 m diameter bubble column reactor:

(1.61 )

The growth rate can be calculated by Eq. (1.54).

Recycling models are often used for the interpretation of the residence time dis-
tributions in ATL reactors [51-56]. However, apart from the investigations of Froh-
lich et al. [54, 55], who characterized and modeled a 5_m3 pilot plant ATL reactor
with 10 m height and internal loop, in which baker's yeast was produced in contin-
uous operation, most of the models were tested only on laboratory reactors.
In the following, a recycle model is presented, which was tested in another 5_m3
pilot plant ATL reactor with 23 m overall height including a head section, in which
baker's yeast was produced by fed-batch operation. The reactor fluid dynamics was
characterized by following measurements: global gas residence time distribution
(RTD) of the entire system and separate local gas RTDs in the riser, in the down
comer, and in the head sections; liquid recirculation time and mixing times in the
entire systems and local liquid mixing times in the riser, down comer, and head
sections; radial bubble velocity profiles in the head section and radial liquid velocity
profiles in the riser; radial and axial liquid dispersion by heat pulse technique and
fractal analysis [56].
The residence time distribution E(t) in an ATL reactor with a head section for fed-
batch operation is given by [57]:

E () r (BOOTO) [BOO(t -
t = { --0.5 - - exp -
TO)Z] + (1.62 )
r - q 7ft 7ft

~ n (BonTo) 0.5 [Bon(t-To)2]} (l-r)/To

+ f=t r 0.5 -7f-t- exp - 4tTo - -;-(1---r7-;)[c-r/-:-:(-'r-'----q--,-),-]+-r

Here n=number of circulation, Boo=Bodenstein number in the riser-head section,

Boc=Bodenstein number in the riser-down comer section, 'to=mean residence time
riser-head section, tc=circulation time, r=fraction of the gas which is recirculated
(small bubbles), q=fraction of the gas which does not recirculate (large bubbles),
( 1.62a)

(1.62b)

Xn = [ (2BO;:1 + 8Bo;:Z)t~
Z + (-1 -Z) TOz] . [TO (1 + 2BoO-1) + ntc ]-1 (1.62c)
2BoO + 8BoO
(1 + 2Bo~l)
The identification of the model parameters was performed by a least-square fit-
gradient method. The sum formation was broken off at n=5, since no higher signals
could be observed.
38 1 Bioreactor Models

On the other hand, microbial cultivation in ATL reactors can be simulated by axial
dispersion models. Luttmann et al. [58, 59] modeled the cultivation of Hansenula
polymorpha in an ATL reactor of 60 I volume with external loop 60 I by an axial dis-
persion model.
As an example, the modeling of the fed-batch cultivation of Saccharomyces cerevi-
siae on molasses in the 5_m3 pilot plant ATL reactor with a head on the top by means
of an axial dispersion model is considered [60]. In addition to the fluid dynamic
parameters listed above, the following cultivation variables are monitored: tempera-
ture, pH, pOz (in two different positions), the concentrations of cell mass, sugar, and
ethanol in the cultivation medium, Oz and COz in off-gas.
The following assumptions were made:
- the reactor is isothermal
- quasi steady state prevails
- the concentrations of cell mass, sugar, and ethanol in the ATL reactor is uniform
(experimentally confirmed)
- cells have the density of the cultivation medium
- the ideal gas law is valid
- mass transfer resistance in the gas film can be neglected
- for each of the three segments individual axial dispersion coefficients are valid
- the radial variation of the process variables can be neglected (uniform radial li-
quid velocity profile was confirmed)
- axial variation of the dissolved oxygen concentration is taken into account
- axial variations of the cross sections of the flow are taken into account
- boundary conditions between the segments were experimentally determined
- the metabolisms of the yeast is described by the Bellgardt [61] model, but the
Monod model was used for the growth instead of the Monod-Blackman model
Figure 1.6 shows the scheme of the reactor system.
Regarding the one-dimensional mass transport equations for the gas phase:

- ~ [A-EG-dZ Pj XOG-J - ~ [A-UG- Pj X

at J J RT J az J J RT
-J dz -
OG} (1.63)
- Aj(l - EG)dz(kLa)j(kHjPjxOGj - CWj ) = 0

Here j=r, d, k (r=riser, d=down comer and k=head).

Regarding the one-dimensional mass transport equations for the liquid phase:

- ~ [A-(1- EG)dzCc -] + ~ [A-(1 -

at J I} az J
EG-)Dc aCLij ] dz
J } az

a
- az [AjULjCLij] dz + Aj(l - EGj)dz(kLa)ij (kHPjxGj - CLij )- (1.64)
dz dz
- Aj(l - EGj)dzCLXjqij + OrFj H F-CLie - OrZj H CLij = 0
1 J

Here CLi=molar concentration of i in the liquid phase, Hj=height of the section j,

HFj=height of the liquid entrance section j; i=O, S, E, X (O=oxygen, S=substrate,
E=ethanol, X=cell mass). Mass transfer gas-liquid for S and X is nil. DLi=axial dis-
persion coefficient of i in the liquid phase, CLx=cell mass fraction in the liquid
phase, qi=specific mass fraction transport rate of i, QLP=volumetric rate of the feed-
ing, CLie=concentration of i in the feed, QLz=volumetric rate of the total liquid.
1.4 Bubble Column (BC) and Airlift Tower Loop (ATL) Reactors 39

. . .·_···_·_··l QgN (t), Xgia(t)

,

Head(k)

,
Xgik(z,t)

f CLil/Z,t) rg(t)
.......... _................ _.......•_..._..._......•. _._.,
z

!
! T(t) ~
Riser (r) VL(t) Doumcomer (d)
km(t)
dlz), HT dlz), Hd
QLF (t), Ugr(z,t), uLiz,t) ...
QLZ(t), c Lie P/z,t),e/z,t) ...
Edt)
...
...
(kuJ.lt) z
xg<iz,t)
xgl/z,t)
cLU/(z,t)
cL./z,t)

,. t
i !
!
QgN (t), Xg~(tJ ,
..._._._ ........._........._._.~_. ____._._._._._.__.____._____________...__ i

Fig. 1.6. System scheme of the pilot plant ATL reactor with the riser, down comer, head regions
modeled by Strauss [60)

From the balances in the gas and liquid phase the axial dispersion model for the
coupled riser, down comer, head system, a second order non-linear coupled partial
differential equation system, is obtained.
For the gas phase (i=O; j=r, d, k):
40 1 Bioreactor Models

With regard to the boundary conditions, in the gas phase the dispersion can be
neglected. Therefore, differential equations are of first order, and only a single
boundary condition is needed: The state of the entrance (0) corresponds to the state
of the exit (H) of the section positioned before: Following boundary condition is
valid for i=O; (O=oxygen):
(1.67a)

(1.67b)
Here rG=the gas recirculation ratio, Hd=exit of the down corner, Hr=exit of the
riser.
The experiments indicated that the exchange between the liquid phases of the
three sections of the reactors are negligible. Therefore, the following boundary con-
ditions were assumed at the entrance of the three sub-systems:

(1.68a)

(1.68b)

(1.68c)

For the exit is was assumed that the dispersion coefficients are negligible, which
was also proved by experiments:

- ArULr a~~ir + Ar(l - EGr ) (kLa)ir(kHiPrXGir - CLir )

- Ar ( 1 - EGr ) CLXrqir + QLPrCLie = 0

HPr (1.69)
aCLid
- AdULd ----az + Ad(l - EGd) (kLa)id(kHiPdXGid - CLid)

- Ad(l - EGd)CLXdqid = 0

At the top of the head section

OCLik
--=0
OZ
is valid.
Initial conditions are
(j=r,d,k) :XGOj(z,O),CLOj(z,O) and CLsj(z,O) in steady state,
(1.70)
CLEj(Z,O) = CLEO,CLXj(Z,O) = CLXO

The program package SLDGL, based on the self-adaptive difference method, was
used [48,49]. The computation was performed in the Regional Computer Centre of
Lower Saxony (RRZN) Hannover. The agreement between the calculated and mea-
sured process variables was very good.
References 41

1.S
Conclusions

The lumped parameter models are suitable to model the performance in laboratory
reactors. The liquid circulation models can be used for pilot plant reactors up to few
cubic meters in volume. This holds true for the application of the axial dispersion
model for BC and ATL reactors as well. However, they are unsuitable for modeling of
large industrial reactors above 10 m 3 in volume. Only Computational Fluid Dy-
namics (CDF) is appropriate for modeling large stirred tank reactors [31-40, 62-
66), bubble column and airlift tower loop reactors [67-73]. These methods are pre-
sented in chapter 7 for ST and chapter 8 for BC reactors in this book.

References
1. Sweere APJ, Luyben KChAM, Kossen NWF (1987) Enzyme Microb Technol9:386
2. Dunn IT, Heinzle E, Ingham J, Prenosil JE (1992) Biological Reaction Engineering, VCH,
Weinheim
3. Schiigerl K (1991) Bioreaction Engineering 1. Wiley, Chichester, Amsterdam
4. Dunn IJ, Mor JR (1975) Biotechnol Bioeng 17:1805
5. Lim HC, Chen BJ, Creagan CC (1977) Biotechnol Bioeng 19:425
6. Durado A, Goma G, Albuquerque U, Sevely Y (1987) Biotechnol Bioeng 29:187
7. Mersmann A, Schneider G, Voit H, Wenzig E (1990) Chern Eng Technol l3:357
8. Schiigerl K (1992) Bioreaction Engineering II. Wiley, Chichester, Amsterdam
9. Yagi H, Yoshida F (1977) Biotechnol Bioeng 19:801
10. Aiba S, Furuse H (1990) Biotechnol Bioeng 36:534
11. Kawase Y, Halard B, Moo-Young M (1992) Biotechnol Bioeng 39:1l33
12. Ogut A, Hatch RT (1988) Can J Chern Eng 66:79
l3. Vardar-Sukan F (1986) Process Biochemistry 4:40
14. Royce PNC, Thornhill NF (1991) AIChE Journal 37:1680
IS. Pirt (1975) Principles of microbe and cell cultivation. Blackwell Scientific, Oxford
16. Noorman HJ, Luijkx GCA, Luyben K, Heinjen JJ (1992) Biotechnol Bioeng 39:1069
17. Schliiter V, Yonse! S, Deckwer WD (1992) Chern Ing Techn 64:474
18. Kawase Y, Moo-Young M (1988) Chern Eng Res Dev 66:284
19. Herbst H, Schumpe A, Deckwer WD (1992) Chern Eng Technol15:425
20. Bajpai RK, Reuss M (1982) Can J Chern Eng 60:384
21. Reuss M, Bajpai RK (1991) Stirred tank models. In: Schiigerl K (ed) Biotechnology 2nd
edn, vol 4. VCH, Weinheim, p 299
22. Sinclair CG, Brown DE (1970) Biotechnol Bioeng 12:1001
23. Oosterhuis NMG, Kossen NWF (1984) Biotechnol Bioeng 26:546
24. Bader FG (1987) Improvements in multiturbine mass transfer models. In: Ho CS, Oldshue
JY (eds) Biotechnology processes: scale up and mixing. AIChE, New York, p 96
25. Bader FG (1987) Biotechnol Bioeng 30:37
26. Singh V, Fuchs R, Constantinides A (1987) A new method for fermentor scale up incor-
porating both mixing and mass transfer effects. 1. Theoretical basis. In: Ho CS, Oldshue
JY (eds), Biotechnology processes: scale up and mixing. AIChE, New York, p 200
27. Mayr B, Horvat P, Nagy E, Moser A (1993) Bioprocess Eng 9:1
28. Ragot F, Reuss M (1991) A multi-phase compartment mode! for stirred bioreactors incor-
porating mass transfer and mixing. In: Reuss M, Gilles ED, Knackmuss HJ (eds) Biochem-
ical engineering-Stuttgart. Gustav-Fischer, Stuttgart, p 184
29. Riiffer HM, Pethii A, Schiigerl K, Liibbert A, Ross A, Deckwer WD (1994) Bioproc Eng
11:145
30. Spalding DB (1985) Computer simulation of two-phase flow with special reference to nu-
clear reactor systems. In: Lewis RW, Morgan K, Johnson lA, Smith WR (eds.) Computa-
tion techniques in heat transfer. Pineridge Press, p 1
42 1 Bioreactor Models

31. Harvey PS, Greaves M (1982) Trans Inst Chern Eng 60:195
32. Placek J, Tavlarides LL (1985) AIChEJ 31:1113
33. Placek J, Tavralides LL Smith GW, Fort 1(1986) AIChE J 32:1771
34. Ju SY, Mulvahill TM, Pike RW (1990) Can J Chern Eng 68:3
35. Issa RI, Gosman AD (1981) The computation of three-dimensional turbulent two-phase
flows in mixer vessel. Numerical methods in laminar and turbulent flow. Preneridge
Press, Swansea, p 829
36. Gosman AD, Lekakou C, Politis S, Issa RI, Looney MK (1992) AIChE J 38:1946
37. Bakker A, Akker HE (1991) A computational study on dispersion gas in a stirred reactor.
Proc 7th European Congress on Mixing, p 189
38. Traghard C (1988) A hydrodynamic model for the simulation of an aerated agitated fed
batch fermentor. Proc 2nd Int Conf Bioreactor Fluid Dynamics, Cambridge, p 117
39. Jenne M, Reuss M (1996) Chern Ing Techn 68:295
40. Hjertager BH (1996) Computational fluid dynamics (CFD) modelling and simulation of
bioreactors. In: Larson G, Forberg C (1996) (eds) Bioreactor engineering, EFB, WP bior-
eactor performance and measuring and control. Saltsjobaden
41. Deckwer WD (1991) Bubble column reactors. Wiley, Chichester
42. Chisti Y (1989) Airlift bioreactors. Elsevier Sci Publ, Amsterdam
43. Joshi JB, Sharma MM (1979) Trans I Chern E 57:244
44. Lapidus L, Elgin JC (1957) AIChE J 3:63
45. Wallis GB (1969) One-dimensional two-phase flow. McGraw Hill, New York
46. Zuber M, Findlay, JA (1965) Trans ASME J. Heat Transfer 87c:453
47. Ueyama K, Miyauchi T (1979) AIChE J 25:258
48. Verlaan P, Tramper H, van't Riet K, Luyben KChMA (1986) Chern Eng J 33:B43
49. Akita K, Yoshida F (1974) Ind. Eng, Chern Proc Des Dev 13:84
50. Kataoka H, Takeushi H, Nakao K, Yagi H, Tadahi T, Otake, T, Miauchi, T, Washini, K,
Watanabe, K, Toshida, F (1979) J Chern Eng Japan 12:105
51. Fields PR, Slater NKH (1983) Chern Eng Sci 38:647
52. Warnecke HJ, Priiss J, Langemann H (1985) Chern Eng Sci 40:2321
53. Warnecke HJ, Priiss J, Leber L, Langemann H (1985) Chern Eng Sci 40:2327
54. Frohlich S, Lotz M, Korte T, Liibbert A, Schiigerl K, Seekamp M (1991) Biotechnol Bioeng
38:43
55. Frohlich S, Lotz M, Korte T, Liibbert A, Schiigerl K, Seekamp M (1991) Biotechnol Bioeng
37:910
56. Wan Liwei (1990) Detaillierte Untersuchungen des Mischverhaltens realer Mehrphasen-
reaktoren. VDI Verlag Reihe 8:Nr. 228
57. Riiffer HM, Wan Liwei, Liibbert A, Schiigerl K (1994) Bioproc Eng 11:153
58. Luttmann R, Thoma M, Buchholz H, Schiigerl K (1983) Comput and Chern Eng 7:43, 51
59. Luttmann R, Munack A, Thoma M (1985) Adv in Biochem Eng 32:95
60. Strauss G (1990) Modellbildung des Wachstums von Backhefe (Saccharomyces cerevisae)
im Fed-Batch-Airliftreaktor. Diss University Hannover
61. Bellgardt KH, Kuhlmann W, Meyer HD (1982) Deterministic growth model of Saccharo-
myces cerevisiae. Parameter identification and simulation. In: Halme A (ed) Modelling
and control of biotechnological processes, IFAC, Helsinki. Pergamon, Oxford, p 67
62. Schonauer W, Raith K (1980) Skizze der Arbeitsweise des SLDGL-Programmpaketes, In-
ternal Report No. 18/80, Computer Centre, University of Karlsruhe
63. Schonauer W, Raith K, Glotz G (1981) The principle of the difference of the difference
quotients as a key of the self-adaptive solution of non-linear partial difference equations,
Comput Methods Appl Mech Eng 28:327
64. Patanakar SV (1980) Numerical heat transfer and fluid flow. McGraw Hill, New York
65. Fokema MD, Kresta SM (1994) Can J Chern Eng 72:177
66. Abid M, Xuereb C, Bertrand J (1994) Can J Chern Eng 72:184
67. Lapin A, Liibbert A (1994) Chern Eng Sci 49:3661
68. Becker S, Sokolichin A, Eigenberger G (1994) Chern Eng Sci 49:5747
69. Sokolichin A, Eigenberger G (1994) Chern Eng Sci 49:5735
70. Devanathan N, Dudukovic MP, Lapin A, Liibbert A Chern Eng Sci (1995) 50:2661
References 43

71. Liibbert A, Paasch en T, Lapin A (1996) Biotechnol Bioeng 52:248

72. Sokolichin A, Eigenberger G, Lapin A, Liibbert A (1997) Chern Eng Sci 52:611
73. Delnoij E, Kuipers JAM, van Swaaij WPM (1997) Chern Eng Sci 52:3623
2 Bioprocess Models
Karl-Heinz Bellgardt

2.1
Introduction

In this chapter, principles of modeling of biotechnical processes will be introduced.

The focus is set onto models of the biological system within the entire bioprocess, as
its most important steps are growth and product formation during the cultivation.
The general modeling techniques will then be applied and outlined further in
Parts C and D, where the methods will be applied to various processes. This chapter
is structured as follows: an overview of the different types of models is given first.
The models of the biotic phase are then introduced with increasing complexity, be-
ginning with simple formal-kinetic models over structured models up to segregated
population models.
For an analysis and mathematical modeling of biotechnical processes, one must
look in somewhat more detail into the biological system, i.e., the microbe, plant, or
animal cells that constitute the biotic phase of the bioreactor. It is well known that
the cell metabolism and the resulting kinetics of growth and product formation are
rather complex. Therefore, the interesting question is why this complex system often
results in relatively simple growth kinetics, and can be described by only a few math-
ematical equations. One reason is that the functional blocks of metabolism operate
together - coordinated by a network of metabolic regulation and of exchange of
mass, charge and energy - to ensure the survival and reproduction of the organism.
Another reason is the tremendous number of cells in the population in a bioreactor
that hides individual variations in their growth and leads to a smoothed average
behavior.

2.1.1
Intracellular Structure Elements

The cell of a living organism contains in a very simplified view several. functional
blocks and central pools of substances [1]. In the catabolism, the extracellular sub-
strates, e.g., sugars, that were transported into the cell are degraded to smaller com-
pounds, the so-called catabolites. Closely connected to the catabolism is the inter-
mediary metabolism, in which a number of organic acids and phosphate-esters are
synthesized. Well known catabolic systems that are found in many microorganisms
are the glycolysis from glucose to pyruvate, and the citric-acid-cycle from pyruvate
to carbon dioxide. Another important task of the catabolism is to provide chemical
energy by substrate phosphorylation. This energy is required to drive energy-con-
2.1 Introduction 45

suming reactions and to synthesize the macromolecular compounds of the cells. The
energy yield of substrate phosphorylation is much lower compared to the respira-
tion, but in anaerobic processes the substrate phosphorylation is often the only en-
ergy producing step of the metabolism. Many products released from microorgan-
isms - the primary metabolites - arise directly from the catabolism, often in con-
nection with the electron metabolism for regeneration of NADH. Examples are, be-
side the common product COz, typical "fermentation products", such alcohols, gly-
cerol, lactate, butyrate, butanol, butane-diol, methane, and also products of aerobic
"incomplete oxidation", e.g., acetic acid, gluconic acid. The formation of these pro-
ducts is often caused by an only partially functioning citric-acid-cycle.
In the anabolism, the catabolite and intermediate pool provided by the catabolism
and intermediary metabolism, together with pools of chemical energy (ATP) and
reduction equivalents (NADH), is used to synthesize higher-molecular cell material
and structural compounds of the cell, like cell wall, membranes, organelles, and
macro-molecules, such as carbohydrates, proteins, RNA, DNA, lipids. The anabolism
is the main energy sink of the metabolism.
Regarding the energy-metabolism and ATP-ADP-AMP-pool, the cellular metabo-
lism as a whole consumes energy that is produced by catabolism and respiration.
The energy released in these processes is conserved in phosphorylated compounds,
mainly in adenosine-tri-phosphate (ATP), the universal energy carrier of the cell.
The less phosphorylated compounds are adenosine-di-phosphate (ADP) and adeno-
sine-mono-phosphate (AMP). A second energy pool of minor importance is formed
by the guanosine-phosphates (GTP, GDP, GMP).
With regard to the electron-metabolism and NAD-NADH-pool, the growth process is
essentially an oxidation process with electron-transfer from the substrates to the final
products. The biological oxidation is done by dehydrogenation via nicotine-amide-
adenine-di-nucleotide (NAD) as a central proton acceptor. In some electron transfer
reactions nicotine-amide-adenine-di-nucleotide-phosphate (NADP) and flavine-ade-
nine-di-nucleotide (FAD) are involved. The reduction equivalents produced in the
catabolism have to be continuously regenerated in parallel, either by the respiration,
or by transferring hydrogen to the fermentative end-products of the metabolism.
The respiration is under aerobic conditions the most effective way for both regen-
eration of NADH to NAD and production of chemical energy as ATP by oxidative
phosphorylation. The respiratory chain, located in the mitochondrial membrane,
performs the biological oxy-hydrogen reaction:

!02 + NADH + H+ + PIO· ADP --> H2 0 + NAD+ + PIO· ATP

The number of phosphorylations per atom oxygen (PiO-ratio) can be up to three,

but is mostly found in the region from one to two.
The analysis of the flows of mass, energy, and electrons through the cell by the
metabolic reactions is the basis for the mathematical modeling of the stoichiometry
of the growth process. The simple fact that chemical elements, mass, energy, and
charge must be conserved in any reaction, and in the metabolism as a whole, puts
strong restrictions onto the microbial reaction network [2]. A few reactions can al-
ready determine the rates of many others. By using the related balances, far-reaching
predictions on the yields of cell mass and products, especially primary products, can
be obtained [6]. When the main substrates and final products of the growth are
46 2 Bioprocess Models

known, this prediction is often even possible with incomplete information on the
involved intracellular reaction network.

2.1.2
Regulation of the Metabolism

In a living system as an open system, the steady flow of substances through meta-
bolic reactions is the precondition for its lasting survival. Therefore, the cell depends
greatly on the ability to stay far away from the chemical equilibrium where no reac-
tion is possible. Most of the reactions of the metabolism are specifically catalyzed by
enzymes [8]. This provides the cell with the possibility to adjust actively in a con-
trolled manner the intracellular reaction rates, in response to the extracellular con-
ditions (e.g., the available substrates), or depending on the intracellular state (e.g.,
the position in the cell propagation cycle).
There are two important mechanisms for this regulation.
Enzyme inhibition or activation is a very fast ad-hoc regulation on the reaction
level. The regulatory substances - often substrates or products of the regulated reac-
tion chain - form a complex with the enzyme and in this way modify immediately
its catalytic activity in negative (inhibition) or positive direction (activation).
Induction or repression as a regulation on the genetic level influences the concen-
tration of the enzymes by modifying the translation efficiency of genes coding for
the enzyme. Compared to the previous type, this regulation is relatively slow since
the synthetic capacity of the cells is rather limited. The time for full induction of
enzymes is typically in the magnitude of the doubling time of the cell. The process
of enzyme induction or repression often manifests itself in the so-called lag-phases,
adaptation phases with time-varying growth activity of the cells. By these mechan-
isms the cells avoid the surplus synthesis of enzymes that are not required under the
actual growth conditions. This frees energy, material, and cellular resources for
synthesis of other proteins.
At the first look, with the complex regulation kept in mind, the growth kinetics
seem to be too complicated for a mathematical modeling, i.e., for an exact quantita-
tive description of the most important cellular processes. The situation becomes
more difficult if one considers that structural details and the exact kinetics of many
cellular regulation systems are still not known. Fortunately, as a matter of fact, both
kinds of cellular regulation proceed well-coordinated in a hierarchical way to obtain
a most effective metabolism. The regulation on the genetic level may be considered
as the upper level of the hierarchy that is controlled rather roughly. On the lower
reaction level, the fine-control with short response time is performed. For most bio-
technical processes, as a very good approximation, the entire metabolism may be
taken as being determined by only a few key-reactions. This may be formulated in
the following rules which often can help to simplify the modeling.

2.1.2.1
Bottle-Neck Principle

The rate of the entire growth process is determined by the slowest reaction, which is
called the rate-limiting step or the bottle-neck reaction.
In a linear reaction-chain, there can be only one bottle-neck at a time. In a com-
plex branched reaction network as the cell is, there may be several bottle-necks in
2.1 Introduction 47

different branches of the network. The bottle-neck principle is sustained by the con-
servation laws of elements, mass, charge, and energy, by which rather independent
system of the metabolism such as catabolism, anabolism, or respiration are tightly
coupled. Therefore, quite often the modeling of the growth can be restricted to only
a very few or even one rate-limiting reaction.

2.1.2.2
Optimality Principle

The metabolic regulation coordinates the metabolism in an optimal way. The optim-
ality criterion is often the maximization of the growth rate.
This principle has direct consequences for the regulation on both the reaction and
the genetic level. At first it means that intermediates or unnecessary end-products of
the metabolism are not excessively accumulated, even if the bottle-neck is not the
first reaction of the pathway. All fast reactions are down-regulated to meet the rate
of the slowest reaction; they need not be considered in the model. For the modeling,
the metabolism can be taken in the major parts as a quasi-stationary flow system
without internal storage.
The second consequence of the optimality principle is that often, by formally sol-
ving the optimization problem for a suitable optimization criterion, an accurate and
predictive model for the regulation on both hierarchical levels can be developed even
without knowledge of the underlying mechanisms which are an outcome of the evolu-
tion by the competition of different species for substrates and biotopes. This property
is used in the cybernetic models (Sects. 2.3.3 and 2.3.4). The main tasks of modeling
are then to construct a very general model frame and to formulate the optimization
criterion that mathematically represents the optimality principle.

2.1.3
Kinetics of Growth and Product Formation

The principal growth kinetics for a batch process are shown schematically in Fig. 2.1.
After starting the process by inoculation, there is a maximum substrate concentra-
tion and low cell concentration. During the following cultivation with cell propaga-
tion, several different phases can be distinguished, although not all may be always
present.
The lag-phase directly after inoculation is a period with an almost constant cell
number, and only minor increase of cell mass. This can have two reason. The so-
called apparent lag-phase is caused by a very low fraction of viable cells which first
has to overgrow the inviable fraction to lead to a remarkable increase in cell number.
During the true lag-phase, the regulatory adaptation of the cells to the growth con-
ditions in the bioreactor take place, usually different from those in the preculture.
There can also be morphological changes such as spore germination. The metabolic
adaptation is performed by induction and repression of enzymes, or formation of
cell organelles like ribosomes, and is accompanied by a steady increase of the growth
rate.
During the following exponential growth-phase, cell mass and cell number in-
crease exponentially, and the growth rate has reached its highest value for the given
cultivation conditions and medium. The cellular metabolism is in a quasi-steady-
48 2 Bioprocess Models

i
In
C
o
~
1:
~
c
uo ...._ _

II III~I N v
I Substrate uptake I I I ~
i I
I
I I
I I
I
I
I I I I

,
In
I Production I I I
~c I I
o Growth I I
t5 I I
'~" I I
()
I I
j
~
()
Q)
c.
f/)
I
I :\ I
I
I I I
time _

Fig.2.la,b. Schematic plot of a type-I bioprocess. I: Initial lag phase; II: Exponential growth
phase; III: Transition phase; IV: Stationary phase; V: Declining phase: a kinetics of growth and
production; b rate patterns

state, i.e., growth is balanced. The length of this phase depends mainly on the initial
amounts of substrate and cell mass.
In the transition phase, the growth becomes limited by the low substrate concen-
tration or accumulation of a toxic substance. In a batch process, the transition phase
to complete exhaustion of the substrate is usually very short and no important part
of the entire process. In a fed-batch process, this phase can be almost arbitrarily
prolonged by substrate feeding.
In the stationary phase, there is no net growth, either due to exhaustion of sub-
strate, or do to a balance of growth and lysis processes. The concentrations of cells
and product remain constant. In a production run, the cultivation is normally
stopped before this phase. The length of the stationary phase depends greatly on
the microorganism strain and the cultivation condition.
In the declining phase, when there is no external substrate available, the cells can
maintain their structural integrity by degradation of storage material, which reduces
the cell mass. After exhaustion of the storage material, the cell viability decreases,
followed by cell damage and autolysis. These processes not only reduce the total cell
mass, but also the cell number.
Besides this pattern of the development of cell mass, the interrelation of growth
rate and product synthesis is an important characteristic of any biotechnical process.
Gaden [3] classified biotechnical processes according to their growth and production
kinetics roughly into three different types, which all demonstrate a fairly distinctive
rate pattern, as shown in Figs. 2.1-2.3.
2.1 Introduction 49

en
c
o
~c
~
c
o
()

time ------->-

Fig.2.2a,b. Schematic plot of a type-II bioprocess. I: Initial lag phase; II: First exponential
growth phase with high growth rate; III: Lag or transition phase; IV: Production phase; V:
Transition phase; VI: Stationary and declining phase: a kinetics of growth and production; b
rate patterns

In processes of type I (production of primary metabolites), the main products

appear directly as final products of catabolism, electron-metabolism, or energy me-
tabolism. The production is in direct coupling with the growth. The rates of sub-
strate uptake, cell growth, and product formation are virtually coincident, and the
concentrations of cells and product show a very similar pattern. Hence these pro-
ducts are called growth-associated. There is basically only one key reaction or bot-
tle-neck to be considered in a mathematical model. Examples for type-I processes
are cultivations where the cells itself are the product, or anaerobic alcoholic fermen-
tation.
To processes of type II (over-production of primary metabolites) belong such pro-
cesses in which the formation of the main product is also coupled to catabolism or
energy metabolism, but the product is not a final end-product of the pathway. It may
be an essential intermediate of the metabolism or result from some side-reactions;
therefore, the production may be partially decoupled from the growth. Under opti-
mal growth conditions the product formation is low. Maximum product formation
proceeds on the cost of growth rate, although there may still be a positive correlation
of growth and product formation. Often the enhanced production is due to limita-
tion of a certain growth factor, regulatory deficiencies of the cell, or it can be in-
itiated by addition of a precursor. The production phase - eventually with exponen-
tial growth at slow speed - can be clearly distinguished from the first exponential
phase of fast growth. The production phase might also be preceded by an additional
50 2 Bioprocess Models

CJ)
c
o
""~
c
~
c
o
o

II III N

Substrate
uptake

Growth

time~

Fig.2.3a,b. Schematic plot of a type-III bioprocess. I: Initial lag phase; II: Exponential growth
phase; III: Lag or transition phase; IV: Production phase; V: Stationary and declining phase: a
kinetics of growth and production; b rate patterns

transient or lag phase. The over-excretion of citric acid or amino acids belongs to
this type, but also the production of some secondary metabolites. In this type of
process the stoichiometry changes significantly and one has always to consider more
than one rate limiting reaction in a model.
In processes of type III (non-growth associated products), the main product does
not result from central growth processes - such as catabolism or energy metabolism
- and it may have no obvious function in the metabolism. Antibiotic synthesis is
again an example of this type. Often these products are synthesized under impaired
growth conditions or in connection with morphological differentiation of the cells
[4, 9]. In this type of processes there is no direct and simple coupling of growth rate
and product formation; growth and production phases can be clearly distinguished.
Typically, at high growth rates there is practically no product synthesis, while the
maximum product formation only proceeds at very low or even without growth. In
a normal batch process on a single substrate, the production phase is usually very
short and the final product concentration low. The production phase can be ex-
tended by fed-batch operation, or in batch operation when the cells are grown on a
mixture of two substrates. Then, in the first growth phase the substrate leading to
high growth rate is used and growth is exponential. After exhaustion of the first
substrate, the cells switch to the second substrate and low growth rates, where the
production is induced. The production may continue into the stationary or declining
phase. For this type of process, mostly structured or segregated models are appro-
priate (see Sects. 2.3 and 2.4).
2.1 Introduction 51

2.1.4
General Model Structure for Biotechnical Processes

For analysis and design of biotechnical processes one must consider two aspects -
the biological reactions catalyzed by the microorganisms, and the numerous chemi-
cal and physical processes which precede, accompany, and follow them. Within the
bioreactor, the most important physical processes, which are intimately bound to the
biological reactions, are associated with the transport of material to and from the
surface of the microbial cells. The dependency of the biological reactions on the
microenvironment around the cells is called micro kinetics. Due to the transport phe-
nomena, the microenvironment of the cells will vary along different locations in the
bioreactor. For a continuous or fed-batch process the concentration of substrate will
be higher near to the substrate inlet, or in an aerobic process the oxygen concentra-
tion is high near to the air sparser or around the stirrer. The modeling of these
mixing effects is shown in Part B, while the modeling principles of transport effects
in films or flocs of microorganisms are presented in Sect. 2.1.5. The integral de-
scription of microkinetics in connection with the transport processes that may be
included in a reactor model, is called the macro kinetics. Ideally, one would always
attempt to model the transport phenomena and the microkinetics of the process
separately, because this model could be combined with proper reactor models to
predict the macrokinetics of different bioreactors. Unfortunately, at present, the ef-
fort for simulation of detailed reactor models is rather high and the methods for
analyzing the microenvironment of the cells are still rare. Therefore, what is usually
observed - and what is modeled, even for a stirred tank reactor - is always a kind of
macrokinetics. In the following, inhomogeneities in the reactor will not be explicitly
considered further.
As well as the transport phenomena not being modeled, the modeling of the bio-
logical system is also incomplete. The metabolism of a microorganism is a very com-
plicated process, not only because the intracellular reaction network may be quite
complex, but also due to the large number of control loops on the genetic and reac-
tion level that are overlaid to coordinate the elementary reactions. So even when the
microkinetics could be measured exactly, it would be impossible to establish a cor-
rect model in every detail because one has to introduce rigorous simplifications.
Another reason for uncertainty about the microkinetics is the inhomogeneity of the
microorganism population. The metabolism varies during the cycle of cell growth
and replication. Furthermore, there might be morphological differentiation of cells
accompanied by changes in the cell metabolisms. But what can be observed is only
an averaged behavior over a great number of cells in different states. To emphasize
all these facts and our limited knowledge about the process, the description of the
biological reaction may be called the formalkinetic model. This means a formal ap-
plication of model equations that represent actually microkinetics to a process,
where only averaged macrokinetics can be measured. As an immediate consequence
the model parameters will change when the operating conditions of the reactor are
changed, or much more if a different reactor is used. This necessarily limits the
predictive power of models for biotechnical processes.
The general structure of a model for a bioreactor with gas phase and liquid phase,
and including the biotic phase can be drawn as in Fig. 2.4. The models for bioreac-
tors (see Chap.!) provide the concentrations of substrates, products, and cells as
input variables to the models of the biotic phase. To accomplish this, the reactor
52 2 Bioprocess Models

X;'

C,
Reactor
Model

Cell

~ _:···········
_ _ _______········-BioliC
_ ___phase····················
___________·: J

Fig. 2.4. General model structure for biotechnical processes

model must know the actual reaction rates which are the output of the biological
models as described in this chapter. According to the level of the modeling and its
degree of detail, different types of models of the biotic phase can be distinguished
[5,12].
Unstructured models (Sect.2.2) have the greatest degree of simplification. The po-
pulation of microorganisms or cells is viewed as homogenous biomass with their
properties being time-invariant. The propagation is taken as continuous, i.e., the
discontinuous processes of cell division are neglected. There are also no intracellular
elements or states considered, and the model does not include inner balances. So the
metabolism is assumed to be in a balanced state. The only determining variable is
the cell mass or its concentration. Their properties are described by kinetics for
substrate uptake, growth, and product formation that all only depend on the concen-
trations in the reactor [10].
Structured models (Sect. 2.3) consider the internal structural elements of the cells
that may be metabolites, enzymes, or other cell constituents. By interaction of these
elements, which are described by inner balances and related intrinsic reactions, the
properties of the cells may become variable in time. With this approach it is possible
to model the dynamics of metabolic regulation, e.g., during lag-phases of growth.
But structured models also take the growth process as continuous and the popula-
tion as homogenous; all cells are assumed to be identical and in the same state.
Simple segregated models (Sect.2.4.1) discriminate inhomogeneous populations
into a few classes of cells with different properties. The classes may represent differ-
2.1 Introduction 53

ent species, or within one species differences of physiological, morphological, genet-

ic, or genealogical nature. It is assumed that the cell classes themselves are homoge-
nous and thus can be represented by unstructured or structured models. Segregated
distribution models (Sect. 2.4.6) include - besides - a second or further continuous
independent variables. These can be, for example, mass, volume, or age of a single
cell. The model equations are then partial differential equations. They are applied for
very inhomogeneous populations with a continuous variation in some properties,
which may be viewed as an infinite number of cell classes. Since the solution of such
models requires a great numerical effort, the cell models are kept simple, e.g., un-
structured models. Possible applications are descriptions of the variation of the me-
tabolism during the cell cycle, or of the stability of plasmids in recombinant popula-
tions.
In any of the above types of models, the main task of modeling is essentially the
establishment of consistent rate expressions for the main compounds of the process,
cell mass, substrates, and products. Gaden [3] had already proposed in 1955 the
concept of volumetric rates and specific rates. The volumetric rate Q is the rate of
change of unit mass of reactant per unit volume of reactor as it appears in the bal-
ance equation, while the specific rate is the rate of change of unit mass of reactant
per unit cell mass (here in vector notation for several compounds):

q(C(t) p(t)) = Q(C(t),p(t)) (2.1 )

, Cx(t)

The specific rates can be viewed as the reaction rates of a single cell. Beside the
concentrations, other time varying operating parameters can also influence the mi-
crobial reactions. This is considered by the vector p(t). The definition of the specific
reaction rates is useful since the biomass plays the role of a catalyst, i.e., the volu-
metric reaction rates are indeed proportional to cell mass, and usually the specific
reaction rates are independent of it. The specific rates are preferred for the formula-
tion of models for the biological system.
The net reaction rate q as exchange rate between cells and environment is mostly
the result of several independent intracellular processes and reactions, which are
called the intrinsic reaction rates, r. Due to conservation laws and stoichiometric
laws, the relation between the net reaction rates and the intrinsic specific reaction
rates, catalyzed by and within the microorganisms, can be written as

q (C) (t), P (t)) = y. r (C (t), p (t)) (2.2)

where Y is the matrix of stoichiometric or yield coefficients. When a sufficient num-

ber of kinetic expressions are specified for the intrinsic reactions, all of the net reac-
tions can be calculated as dependent variables of the model. The stoichiometric ma-
trix reflects the structure of the metabolic pathways within the cells. For simple
models, as in the following sections, the yield coefficients are taken as formal para-
meters without looking at the details of the underlying reaction network. The analy-
sis of this is done by flux analysis as shown in Chap. 3 and Part D. Another approach
for determination of the yield coefficients are elemental balances [2, 10] and thermo-
dynamic considerations [7, 11], which are not reported here.
Usual variables in growth models are as elements of the vector q, the specific sub-
strate uptake rate qs, and the specific growth rate qx . The latter is often referred to
54 2 Bioprocess Models

by the symbol JL. This will only be used here when the specific growth rate is speci-
fied as an intrinsic reaction.
Mostly, the kinetics of intrinsic reactions are of saturation type, i.e., for high sub-
strate concentrations the reaction rate approaches a constant maximum value, and it
vanishes if no substrate is available. The substrate uptake rate, for example, can then
be written as

rs = rSmaxr(CS) where r(Cs) -+ {~ for Cs -+ { ~ (2.3)

Therefore, it is convenient to present the kinetics in normalized form, r(C). The

saturation characteristics of the kinetics greatly simplifies the modeling since it
means that substrates or other growth factors being in excess mostly do not influ-
ence the growth and need not be considered in the model.
The maximum reaction rates, r max, in the kinetics can also depend on the "phy-
siological state" of the cells, which might be seen as a representation of the actual
metabolic activity in terms of history of the growth process, concentrations of me-
tabolites and enzymes, or the position in the cell propagation cycle. In the view of a
simple unstructured model all these factors would lead to a variation of the maxi-
mum reaction rates in time, which only can be modeled by the more complex ap-
proaches.

2.1.5
Transport in Microbial Aggregates

Inhomogeneities in the liquid phase of the reactor or the near environment of the
cells can have different causes:
- Insufficient mixing in the reactor, as in non-ideal STR, bubble column or airlift
reactors. This leads to the development of strong spatial concentration profiles,
mainly for substrates and products while the biomass mostly is more uniformly
distributed in the reactor.
- Formation of relatively dense cell aggregates of more or less spherical geometry,
such as clumps, floes, or pellets. The dense biomass hinders the convective and
diffusive transport within the aggregates, and causes spatial concentration gradi-
ents for substrates and products in their inner by an increased transport resis-
tance. As a consequence, the metabolic activity of the cells, e.g., the growth rate,
also shows spatial variations. This, together with the immobilization effect for the
cells within the aggregates, gives rise to the development of an inhomogeneous
biomass density.
- Growth of biofilms by attachment of cells to surfaces. The phenomena are similar
to the previous case, but here the biomass is fixed in the reactor and the geometry
of the film is flat. A simple segregated model will be given in Sect. 2.4.
- Immobilization of cells in carrier particles. This technique is used to facilitate re-
tention of cells, the biocatalyst, in continuously operated bioreactors. This case is
quite analogous to microorganism pellets although the transport resistance may be
much higher.
In all of these cases an accurate modeling has to consider the transport effects and
spatial variation of the concentrations. Models for transport phenomena are partial
differential equations which require a much higher effort for simulation than lumped
2.1 Introduction 55

parameter models. Therefore, unstructured or simple segregated models are pre-

ferred for the description of the biological reaction in the transport models, because
they do not increase the model complexity by introducing further state variables as
in structured models. When applying structured models for an inhomogeneous en-
vironment, special care of its mathematical validity has to be taken as will be dis-
cussed in Sect. 2.3.
When convective flow is neglected and biomass is taken as a continuum, transport
of substrates and products in aggregates of microorganisms can be described by the
diffusion equation

aCi(X,
at t) . grad (
= dlV Di(x,)
t Ci ())
X, t + Qi (X,)t (2.4)
= /::"(Di(X, t)Ci(x, t)) + Qi(X, t)
where x is the spatial coordinate in three dimensions, and Di the diffusion coeffi-
cient. Usually the following further simplifying assumptions are introduced for the
modeling of low-density biofilms or pellets:
- The diffusion coefficient Di is constant
- The system is highly symmetric and the concentration varies only along one spa-
tial coordinate

A more detailed treatment can be found in [13]. Biofilms are usually taken as homo-
geneous along the supporting surface; thus transport is only in the perpendicular
direction, z, and the balance equation simplifies to

(2.5)

where the reaction term Qi is determined by the biological model. For a unique
solution of this equation, boundary conditions at the attachment surface, z=O, where
the transport flux vanishes

(2.6)

and at the liquid side boundary layer, z=L, have to be specified, where L is the thick-
ness of the biofilm:

(2.7)

Here an additional transport resistance and concentration gradient within the

bulk liquid phase is neglected. Furthermore, an initial condition

(2.8)

that must also fulfill the balance equation, Eq. (2.5), and boundary conditions,
Eqs. (2.6) and (2.7), has to be given to solve a specific problem. Similarly, for a floc
56 2 Bioprocess Models

or pellet with spherical symmetry, the variation is only along the radial coordinate
and the model becomes
aCi(r, t) _ . (a2Ci(r, t) 2 aCi(r, t)) _ Q.( )
at D, a r2 +r ar - , r, t (2.9)

aCi(O, t)
or
_-c:'----'-. = ° (2.10)

Ci(R, t) = CiBulk(t) (2.11)

Ci(r,O) = CiO(r) (2.12)

where r is the running radius and R the outer radius of the particle.
When the biomass density is high, the cells occupy a significant volume of the
aggregate. This reduces the available volume for transport and increases the effective
diffusion coefficient. The effect can be modeled by introducing the following depen-
dency on the cell concentration into Eq. (2.4) [14, 15]:

Diejf = (1 - Cx(x, t))Di

CXmax
(2.13)

Mostly the biomass concentration is assumed as constant over time, i.e., growth of
cells equals cell lysis. For a description of net biomass growth the balance equation,
Eq. (2.4), can be used, where the diffusion coefficient is replaced by a factor being
proportional to the growth rate of the cell [16, 17]:
DXejf = kxqx (2.14)
The outgrowing biomass leads to a spatial extension of the cell aggregate. The
balance equation then has to be solved as a moving boundary problem, or the outer
boundary has to be extended to the bulk liquid.

2.2
Unstructured Models

In unstructured models, the biological reaction depends directly and only on macro-
scopic variables, the conditions in the bioreactor. Therefore, unstructured models
are essentially combinations of elementary kinetics that mainly describe the influ-
ence of substrate and product concentrations or other variables, such as pH or tem-
perature. The only biological state variable is the cell mass concentration Cx [10].
Nevertheless, many phenomena in biotechnological processes can be covered by this
type of model. Beside the cell mass, only those other process variables that show
great variations during the fermentation and have significant influence on the mi-
crobial behavior have to be considered in the model. Also some limited variations in
the yield coefficients that can be incorporated by the unstructured models will be
summarized here.
The most simple models, like the Logistic Law or the Cube-Root Law, do not in-
clude substrate limitation kinetics, but depend only on the cell mass [18]. They have
found applications in the modeling of natural populations or filamentous microor-
ganisms, and in such biotechnical processes where no information about substrates
is available [22] or where growth is not restricted by the available substrate [19-21].
2.2 Unstructured Models 57

For more detailed process models this type of model is less interesting and not re-
ported here.

2.2.1
Kinetics of Growth and Substrate Uptake

The simplest unstructured model describes the growth of a homogenous cell popu-
lation, characterized by the cell concentration Cx, depending only on one limiting
substrate with concentration Cs. When the specific growth rate is chosen as the only
intrinsic rate, the specific rates and yield coefficients in the stoichiometric model,
Eq. (2.2), become for constant parameters

q = (qx)
qs y = (_
xs-1)
1
Y
r = s)
jjmaxr( C (2.15)

where the normalized kinetics r( Cs ) can be chosen from Table 2.1 to suit best the
experimental data. The preferred application area of the different kinetics is also
mentioned in the table. Especially in batch processes, the differences among them
are less relevant if one keeps in mind measuring errors and remaining modeling
errors. Therefore, the Monod-kinetics is mostly chosen since it is simple and has
continuous derivatives with respect to Cs, which is of advantage for numerical simu-
lation procedures.
Beside the substrate limitation, i.e., a reduction in substrate uptake rate with de-
creasing substrate concentration, often the reverse phenomenon is found in biotech-
nological processes, a reduction in growth, substrate uptake, or product formation

Table 2.1. Normalized growth kinetics for a single substrate [18]

Name Remark Normalized kinetics r( Cs )

a Monod Only the substrate uptake step is rate limiting Cs

Ks + Cs
b Blackman Another rate limiting step besides substrate min(1, Ks Cs )
uptake determines the maximum rate
c Teissier Empirical equation 1- e- KsCs

d Moser Substrate uptake with higher order of reaction, CsN

e.g. for gaseous substrates Kf +Cf
e Contois Diffusion layer around the cell Cs
KsCx + Cs
f Powel Considers back diffusion of inner substrate Cs - K\r(Cs )
Ks + Cs - K\r(Cs )
g Mason Parallel uptake by transport and diffusion Cs
and Millis KC+K\Cs
s+ 5
h Vavilin Extension of d) for initial inactivation C5N
by toxic substrates Kf M C~t~O) + cf
58 2 Bioprocess Models

Monod-
model

Substrate
inhibition

Product
inhibition

time

Fig.2.S. Typical patterns of substrate and product inhibited growth in batch culture in com-
parison to non-inhibited growth

with increasing concentration of an effecting substance. This is usually referred to

inhibition of growth or product formation by substrates or products. An example for
such inhibitions is shown in Fig. 2.5 for a batch cultivation. The substrate inhibition
prolongs the initial phase of the process while the product inhibition reduces the
growth rate mainly in the final phase. This is due to the opposite variation in the
substrate and product concentration. The latter increases during the process while
the first is catabolized by the cells.
Table 2.2 contains a list of normalized pure substrate inhibition kinetics and Ta-
ble 2.3 of normalized general inhibition kinetics that can be applied to products or
substrates. Most of the inhibition kinetics are extensions of the Monod-equation
and were derived from basic mechanisms of enzyme inhibition. Besides such true
inhibitions, there may be other mechanisms leading to similar kinetics: the sub-
strate or product might be toxic for the cell, for instance by denaturing enzymes
or causing damage to the cell wall, membranes, and other cellular systems. Many of
the kinetics cannot predict zero growth for a finite inhibitor concentration. There-
fore, other empirical equations were proposed having such property that is mostly
connected to toxic substrates (Table 2.2, h or i, and Table 2.3, g-k). The kinetics can
be generalized for a joint modeling of multiple substrate limitations and inhibitions
by combining substrate uptake kinetics in Table 2.1 or 2.2 and normalized inhibi-
tion terms in Table 2.3 to a product with several factors. Such generalized kinetics
of different types were experimentally investigated by several authors, e.g, for anae-
robic growing yeast [34, 36, 40] and for bacteria [35, 41]. The validity of different
substrate inhibition kinetics for growth of yeast strains on alcohols was also suc-
cessfully tested [37, 39].
As an example for combined substrate uptake/inhibition kinetics, diauxic growth
on a mixture of two substrates is considered, where the uptake of one substrate is
inhibited or even repressed by another, the preferred substrate. Here, the inhibition
phenomenon is related to the regulation of the metabolism. A generalized kinetic
expression is used for the uptake step of the second substrate (Cd that is inhibited
by the preferred substrate (CS1 ) and, furthermore, by a toxic product Cp By choosing
the Monod-kinetics (Table 2.1, a) for the uptake of the second substrate, non-compe-
2.2 Unstructured Models 59

Table 2.2. Substrate uptake kinetics including inhibition [18]

Name Normalized kinetics r( Cs)

Cs
a Haldane (uncompetitive)a
Ks +C
s (1 + Jt )

Cs
b Ierusalimsky (non-competitive)a
Ks +C
s (1 + Jt )

C
- -s- e K,
-~
c Aiba, Edwards
Ks + Cs
Cs
d Yano (generalized uncompetitive)

e Teissier type

f Webb
C(1 +~)
S

Ks + Cs ( 1 + Jt)

cf
g Hill (allosteric)

h Wayman and Tseng (toxic substrate) Cs

----KImm. (CSI - Cs,O )
Ks + Cs

Chen (toxic substrate) ( Cs) Cs

1 - KI Ks + Cs - (KszCd

M (C s ) j

Tan et al. [42] ZK;;

1=1

a Uncompetitive and non-competitive substrate inhibition are of the same type of function

titive inhibition by the first substrate (Table 2.3, c), and a Levenspiel-type inhibition
for the product (Table 2.3, i), the resulting model becomes

qS2 = QS2max CS2 Kn ( 1- (cp)N)

- (2.16)
KS2 + CS2 Kn + CS1 K12

For a given practical problem, one should not expect that any of the kinetics gives
a true mechanistic description. Instead, an equation with as few as possible para-
meters should be chosen that fits well to the experimental data.
A further extension of the single substrate uptake kinetics is necessary for the
description of growth on mixtures of several substrates. Various attempts have
been made to develop unstructured models for multiple limitations that are sum-
marized in Table 2.4. The single substrate terms can be specified by the elementary
60 2 Bioprocess Models

Table 2.3. Inhibition kinetics for single inhibitor [18]

Name Normalized kinetics 1'( Cs )

Cs
a Competitive a
KS(I+~)+CS
Cs
b Uncompetitive a

c Ierusalimsky (non-competitive)

d Yano and Koya (generalized uncompetitive)

e Aiba, Edwards

f Teissier type a e-~ - e- KsCs

C1
g Ghose and Tyagi, Dagley and Hinshelwood 1--
KI

h Bazua and Wilke (1 -KICI)t

-

Levenspiel

Han and Levenspiel

k Luong [38]

a These kinetics include the substrate uptake step

Table 2.4. Kinetics for the specific growth rate by multiple limitation of N substrates [18]

Type Kinetics for the specific growth rate Remarks

a Interacting model for Il( CSl>CSZ,CS3,· .. CSN) =IlmaxrSl (Cstl

essential substrates rsz( Csz)rd CS3 ) .•. rSN( CSN )
b Non-interacting model Il( CS1,CSZ,CS3,· .. CSN) =min(lls1 (Cstl, IlSi( CSi ) = IlSimaxrSi( CSi )
for essential substrates Ilsz( CSZ ),IlS3( CS3 )'· .. ,IlSN( CSN))
c Model for growth- Il( CSl>CSZ,CS3,· .. CSN) = IlS1 (CS1 ) + IlSi( CSi ) = IlSimaxrSi( CSi )
enhancing and Ilsz( Csz ) + IlS3( CS3 ) + ... + IlSN( CSN)
alternative substrates
2.2 Unstructured Models 61

kinetics as presented before. There are several possibilities for the interrelation of
growth rate and concentrations of the substrates, depending on the kind of sub-
strate and the microorganisms [46]. All substrates can be essential; that means
there is no growth if only one of the substrates is lacking. Here, in the interactive
model, all substrates together determine the growth rate. In the non-interacting
model, at any time only the substrate with the strongest limitation governs the
growth rate. This leads to a selection of the minimum growth rate allowed among
all substrates. Equations (a) and (b) of Table 2.4 can be translated into the general
vector form of the stoichiometric model, Eq. (2.2), by choosing the specific growth
rate as intrinsic reaction. The vectors of the specific reaction rates, including cell
mass formation and substrate uptake, and the vector of yield coefficients, are then
for essential substrates

qx
qSl -Yxs\
qS2 -YXl2
q= qS3 y= -YXl3 r = JL (2.17)

qSN -Yxs~
A typical example for the presence of two essential substrates is growth of obligate
aerobic microorganisms under limitation of the carbon source and oxygen. Simula-
tions of steady states in a chemostat after [23] are shown in Fig. 2.6 for both the
interacting and non-interacting substrate uptake kinetics. The model is summarized
in Table 2.5. At low dilution rates, where growth is limited by the carbon source, the
cell mass decreases due to endogenous or maintenance metabolism, as explained in
the following section. This can be well described by the model of Herbert, Eq. (2.20).
Above a dilution rate of D=0.16h-I, growth switches from glycerol to oxygen limited,
as can be seen by the very low oxygen concentration and increasing concentration of
the carbon source. Near the critical dilution rate of about D=0.54 h -1, both substrates
are present in high concentrations. There, a second switch in the limiting compound

'"
ex
" '" "0 *,
,
0
* , ,,
* * *...-........
00

.~

Co Cs ex D [h-']
[gm -3J[kgm -3J[kgm-'l

Fig. 2.6. Simulation and experimental data for carbon and oxygen limited growth in a chemo-
stat after Sinclair and Ryder [23]
62 2 Bioprocess Models

Table 2.5. Example for an unstructured model for growth on two essential substrates, glycerol
and oxygen, in continuous cultivation [23]

Description Model equations

a General vector notation,

see Eqs. (1.12) and (2.1)

b Specialized vectors, r considers

endogeneous metabolism

c Intrinsic rate expression based on Interacting

Monod-kinetics for carbon source Cs Co
and Contois-kinetics for oxygen J.L = J.Lmax Ks + CsKoCx + Co
Non - interacting

J.L = J.Lmaxmin(Ks ~ Cs ' KoC~: cJ

d Explicit model equations for the dCx(t)
liquid phase, for the gas phase model, ~ = -DCx(t) + J.L(Cs)Co)Cx(t) - J.LECX(t)
see Table 1.3
dC~;t) = D(t)(cf(t) _ Cs(t)) _ J.L(Cs , ~~;Cx(t)
dCo(t) = kLOa(C~xo(t) _ Co(t)) _J.L(Cs , Co)Cx(t)
dt, v ' Yxo
OTR

OTR "' ....

DO
c,
o :4'! t,.- - --
I
"II
,
{
,I

6
OTR Co Cs ex t [hi
Igm-'h'l [gm~ [kgm-~ [kgm-~

Fig. 2.7. Simulation and experimental data for carbon and oxygen limited growth of Methylo-
monas MIS
2.2 Unstructured Models 63

can be found for the non-interacting model. A theoretical investigation on oxygen-

limited growth by a non-interacting model was presented by Liden et al. [43] and
general comments on its applicability were given by Baltzis and Fredrickson [47]. A
simulation of an oxygen limited bacterial batch cultivation with methanol as carbon
source is shown in Fig. 2.7. The model resembles that of Table 2.5, but with p, given
as an interactive model with double Monod-kinetics. In the oxygen limited period
from t=3.5-4.8 h, growth proceeds only linearly in time due to the almost constant
oxygen transfer (OTR) into the reactor.
Another case for growth on multiple substrates is when one of the substrates is
already sufficient for growth and others are used up in parallel or sequential. For
modeling of this kind of phenomena the entire kinetics can be constructed by the
sum of elementary kinetics (Table 2.4, c). This model can be advantageously trans-
formed into the general form by using each summand as an inner reaction:
qx 1 1
P,SI
qSl -Yxs\ 0 0 0
0 0 0 P,S2
qS2 -Yxs~
q= y= 0 0 0 r= P,S3 (2.18)
qS3 -YXl3

0 0
P,SN
qSN 0 -YXlN
An example of growth-enhancing substrates is given in Fig. 2.8. In the simulation
of the cultivation of P. vulgaris on glucose and citric acid the model developed by
Tsao and Yang [24] as listed in Table 2.6 was used. The reduction of the specific
growth rate can clearly be seen after exhaustion of citric acid at t=15 h.
In any case, compared to the simple kinetics, the multi-substrate models have to
consider much more information about the structure of the metabolism, and these
elementary models will only match simple cases. A number of different models for
growth of yeast on the alternative, growth-enhancing substrates glucose and fruc-
tose, ranging from simple Monod-kinetics up to extended kinetics including inhibi-
tion terms, were presented by Barford et al. [44]. A satisfactory simulation of the
entire cultivation could only be obtained by introducing a switch-condition for cer-

c,

C S2 C S1 ex
[h~ll [kgm~31 [kgm~31 [kgm~31

Fig. 2.8. Batch cultivation of P. vulgaris on glucose and citric acid, simulation and experimen-
tal data for growth enhancing substrates after Tsao and Yang [24]
64 2 Bioprocess Models

Table 2.6. Example for an unstructured model for growth on two alternative substrates, glu-
cose and citrate, in batch cultivation [24]

Description Model equations

a General vector notation,

see Eqs. (1.12) and (2.1) dC(t) = D(CR - C) + Q where Q = qCx
dt

b Specialized vectors
c=

y=

c Intrinsic rate expressions based

on Monod-kinetics Ih = /Lm", I KSI + CS1 /L2 = /Lm", 2 KS2 + CS2

d Explicit model equations dCx(t)

for the liquid phase -----;{t = (/Ll(CSJ) + /L2(CS2 ))CX(t)
dCS1(t) /Ll(CSJ)CX(t)
dt YXS1
dCS2 (t) /L2(CS2 )CX(t)
dt YXS2

tain model parameters. Another case for alternative substrates is the uptake of sugar
anomers, as was investigated by Benthin et al. [45] for Lactococcus cremoris, Escher-
ichia coli and Saccharomyces cerevisiae. While in the yeast the uptake of glucose
anomers proceeds according to competitive inhibition, in the bacteria independent
uptake systems exist that are not inhibited by the other anomer. Another model for
alternative substrates that uses a special type of inhibition function for every sub-
strate component i (si=fraction of sugars in the feed):

(2.19)

was presented by Lendenmann and Egli [50] for growth of E. coli on six different
substrates in a chemostat.
While the models for essential substrates usually give a satisfactory description of
the observed phenomena, this is not the case for alternative substrates. Here, and for
more complicated situations with combinations of alternative and essential sub-
strates, the model should consider the dynamic regulatory response of the cells and
also be based on a proper description of the cellular reaction network. A sequential
uptake is often found under excess of substrate. As long as the preferred substrate
can support a sufficient growth rate, it suppresses and inhibits the uptake system of
other substrates. After its limitation the enzymes for the substrate with the next low-
er preference are induced or derepressed. The result is a remarkable lag-phase for
2.2 Unstructured Models 65

the adaptation to the next substrate. Furthermore, a purely sequential uptake is sel-
dom found, but the substrates can be used up in parallel if the concentrations are low
enough, and these alternative substrates, e.g., different sugars, can all be affected by
other essential substrates like oxygen. This means a lot of regulatory phenomena
have to be considered for a model of multi-substrate kinetics, which are very closely
associated with the metabolic pathways and their regulation. The methodologies of
structured or cybernetic models (Sect. 2.3) fit better to the modeling of such compli-
cated growth processes.
To establish such a global model for the entire course of a cultivation that covers
lag-phases, growth phases - possibly on a number of different substrates - as well as
production phases, as shown in Figs. 2.2 and 2.3, is quite troublesome, and the re-
sults obtained are not always satisfactory. A much simpler task is to build several
individual so-called local models for each of the phases or physiological states of a
cultivation and to connect them together in an ordered time sequence to provide a
full description of the process. This concept - having some analogies to the Physio-
logical State Control proposed by Konstantinov and Yoshida [59] - was introduced
by Halme [57, 58] as functional state models. To establish the correct sequence,
proper switching conditions have to be found that determine the transition from
one model to the next [60]. For example, in a multi-substrate cultivation, each sub-
model will cover only the growth on a single limiting substrate. Naturally, the mod-
els for each phase will be relatively simple; usually unstructured ones should be suf-
ficient, even for complicated processes. These will also have only a few parameters
with great sensitivity in the respective phase. This makes functional state models
well-suited for on-line application and automatic control, because at any time they
contain only a minimum set of states and parameters. It also facilitates the model
adaptation by on-line estimation, resulting in increased model accuracy.
Zhang et al. [61] presented a functional state model for fed-batch cultivations of
baker's yeast. The possible sequence of sub-models and the related switching condi-
tions can be represented by a state transition diagram with four states as shown in
Fig. 2.9. The switching conditions - expressed in terms of respiratory quotient and
dissolved oxygen concentration - worked well in the simulation of a fed-batch culti-
vation. But the difficulties for finding proper switching conditions that are robust
against measuring and modeling errors must not be underestimated. The methods

Sugar
State 1: limitation
State 2:
respiro ------7 mixed
fermentative oxidative
growth growth

S ugar
excess
r dEthanol
limitation
~ Oxygen
limitation

State 3: State 4:
oxidative oxygen-limited
growth on ------7 fermentative
sugar Oxygen growth
limitation

Fig. 2.9. State transition diagram of the functional state model for baker's yeast [61]
66 2 Bioprocess Models

of fuzzy sets and artificial neural networks can help in this field [62]. A similar
modeling concept with emphasis on the time sequence of phases during a cultivation
was developed by Ruenglertpanyakul [63]. He presented models for diauxic growth
on two substrates, cultivations with inhibitory products, and production of antibio-
tics. Local unstructured models were assigned to lag-phases, growth-phases and
production-phases.

2.2.2
Endogenous and Maintenance Metabolism

In many processes, the so far presented unstructured models agree fairly well with
experimental data for a wide range of specific growth rates. But notable exceptions
are known, which often can be accounted for as arising from variations in the yield
coefficient, i.e., the ratio of growth rate and substrate uptake rate. This variation in
the efficiency of the substrate usage with growth rate can be explained as follows.
The endogenous metabolism - the standing requirement for an expenditure of en-
ergy and metabolites for cell repair and maintenance that is independent of growth -
reduces the yield at low growth rates. At strong limitation there can even be sub-
strate uptake without growth. Without any external substrate being available, inter-
nal storage material can be degraded by the cells for maintaining their integrity. This
can for instance be observed during the stationary/declining phase of a batch culture
by a reduction in the cell mass, as shown in Fig. 2.1. Another reason for this effect
can be cell lysis, when the cell repair mechanisms fail. Very satisfactory descriptions
of these phenomena can be obtained by simple unstructured models.
The model of Herbert [25] took this effect into account by assuming that the ob-
served net growth rate qx(Cs) is lower than a fictive true growth rate /lG(Cs). The
constant difference is the specific rate of endogenous metabolism /lE:
(2.20)
An example for this model has already been presented in Fig. 2.6. At strong growth
limitation, the observed growth rate can become less than zero. So, the rate /lE also
can be viewed as a specific rate of cell lysis or degradation rate of intracellular sto-
rage material. The true growth rate, which may be expressed by any of the kinetics
given before, determines the rate of substrate uptake by a constant yield coefficient,
YXSG :

qs( Cs ) = _ /lG( Cs ) (2.21 )

YXSG
The model can be translated into the usual vector notation by taking both, /lG and
/lE, as inner reactions:
1
Y= ( -1 (2.22)
YXSG

The vector of specific reaction rates, q, includes the specific growth rate and sub-
strate uptake rate. The variation of yield with specific growth rate in this model
results in the following hyperbolic function

Yxs(qx) = - qx = YxSG-q,,--x- (2.23)

qs /lE + qx
2.2 Unstructured Models 67

Pirt [26] referred the same phenomena of the decreasing yield to an additional
need for substrate consumption for maintaining the cell structure, expressed by the
maintenance coefficient ms, being a positive constant. Thus, if the uptake rate of
substrate used only for growth is rSG, then the total substrate uptake qs is
(2.24)
Remember that qs, rSG<O, since the material flux for production was defined as a
positive direction. The energetic aspects of the maintenance concept were further
discussed by Esener et al. [10]. By taking the specific growth rate
(2.25)
together with ms as intrinsic reactions, Pirt's model becomes in vector notation
1
Y= ( _\ (2.26)
Y XSG

and the variation in yield is given by

qx qx
Yxs (qx) = - - = YXSG ----=---- (2.27)
qs mSYXSG + qx

which is the same type of function as in Herbert's model, Eq. (2.20). A drawback of
Pirt's concept is for model simulations that in batch culture the substrate concentra-
tion can become less than zero, due to the zero order maintenance reaction. But in
regions of normal growth, both models are equivalent.

2.2.3
Product Formation

Modeling of product formation is, together with the modeling of cell growth, the
most important part of a process model because production is just the purpose of
the process. If one follows the classification of Gaden [3] (see Figs. 2.1-2.3), the ty-
pical application area of unstructured models are growth-associated products, as can
be found in processes of mainly type I, but also type II with partial growth coupling.
The production of primary metabolites is directly coupled with the central catabolic
and anabolic reactions which proceed at a speed that is more or less proportional to
the growth rate. Therefore, it can be described by a simple model, known as the
Luedeking-Piret equation. For a multi-product process of N products the model
can be written in vector notation for the specific rate of product formation as
qp=Ypx!-l+mp (2.28)
where YPX is the vector of yield coefficients for the growth associated production
and mp is the vector of specific production rates due to maintenance metabolism or
non-growth-associated processes.
Since in this model all production rates are linearly dependent on the specific
growth rate, it cannot be applied to processes where there is a great variation in the
product spectrum with growth rate as in processes of type II. Such variation means
there is a switch in the fluxes of the reaction network of the cell that can, for in-
stance, be triggered by the availability of different substrates. Then the model must
contain, besides or instead of the specific growth rate, several independent intrinsic
reactions that determine the product formation. This more general case is already
68 2 Bioprocess Models

covered by the stoichiometric model, Eq. (2.2). If one assumes that the system of
equations is partitioned according to growth rate, substrate uptake rates for N com-
pounds, and product formation rates for M compounds, it can be written for a given
set of K independent intrinsic reactions as

growth { qx YX1 YX2

--------------
YXK
qSl -YSll -YS12 -YS1K
substrate
uptake
j qSN -YSN1 -YSN2
--------------
-YSNK r (2.29)
qPl YPll YPl2 YP1K
product
formation
I qPM
'-v---"
q
YPM1 YPM2
v
y
YPMK

and after separation of the equations for growth, substrate uptake, and product for-
mation as
qx = Yxr
qs = Ysr (2.30)
qp = Ypr
The last line in Eq. (2.30) is a generalization of the Luedeking-Piret-equation,
Eq. (2.28). As for Herbert's and Pirt's model, the endogenous or maintenance meta-
bolism can be modeled here by a single constant element in the vector of intrinsic
reactions. There is also evidence in the literature that the product formation is not
always proportional to the growth rate, even for typical primary products as ethanol.
This phenomenon can, for instance, be modeled by including the specific produc-
tion rate into the vector r and using an exclusive kinetic expression for the depen-
dency on substrate and product concentration [36].
Somewhat different notations of the model of product formation in connection
with the substrate uptake rates can also be found in the literature:
(2.31)

(2.32)
These can be directly derived from Eq. (2.30) by a proper substitution of rand Y.
Nevertheless, Eqs. (2.30) or (2.31) should be preferred, since there the model para-
meters, i.e., yield coefficients, of growth and substrate uptake are completely de-
coupled from those of product formation [51]. This facilitates greatly the model
building and parameter identification, because in the early modeling phase one can
concentrate on the pure growth process without looking at the products. When the
growth model is satisfactory, in a second model building phase, the product model
can be added without further changes in the growth model.
The modeling of non-growth associated products is mainly concerned with the
processes of type III (see Fig. 2.3) where there is by definition no direct coupling to
central metabolic processes or to the growth rate. The inhibition, repression or in-
activation of enzymes by catabolism of a rapidly used carbon source, e.g., glucose,
does not hamper the production of primary metabolites. But the synthesis of sec-
2.2 Unstructured Models 69

ondary metabolites is mostly repressed under such conditions. There are two strate-
gies to obtain maximum productivity for secondary metabolites, and to avoid the
repression: batch cultivation with slowly used carbon source, such as lactose, or
fed-batch cultivation with rapidly used carbon sources. In both cases the growth rate
of the microorganisms is controlled to such low values that production becomes
maximal. Another established fact is that the production of secondary metabolites
is often related to morphological differentiation of the cells. In this case, simple un-
structured models cannot give an adequate description of the growth kinetics, be-
cause they assume homogeneous biomass. To consider the effects of morphology
one has to turn to segregated models, as described in Sect. 2.4. Since the morpholo-
gical differentiation is also controlled by the available carbon source, its direct reg-
ulatory influence on the production of secondary metabolites is understandable. The
regulatory effects can also give rise to lag-phases in growth or product formation,
which should be represented by structured models.
If one nevertheless chooses to use unstructured models, it will be hard to put some
mechanistic ideas into the "formal-kinetics". There is no straightforward way for the
modeling of this more complicated type of process. One possible solution is that the
product formation is still modeled as being directly dependent on the specific
growth rate, but by a non-linear function, f1, [2, 33]
(2.33)
that may have a maximum at a certain low growth rate or be of saturation type, as
shown in Fig. 2.10. Another idea sees the production as being related to dynamic
transients of growth as a consequence of regulatory phenomena. This is expressed
by a model
_ i'.
QP - J2 dt
(d1l) (2.34)

where f2 is again some non-linear function. Finally, both could be combined to a

more general form
(2.35)

Primary
metabolites
(Types I and II)

............ ::::::: Secondary

metabolites
~------r-----~------r-----~~~

Fig. 2.10. Schematic diagram of the coupling of specific production rate and the specific
growth rate for different types of metabolites
70 2 Bioprocess Models

In any case, these are pure black-box models that must be constructed for a given
data set by a proper fitting procedure.

2.2.4
Other Parameters Influencing Growth

In biotechnological processes, growth and product formation depend beside the

composition of the medium on many other variables and operating parameters, like
temperature or pH. Temperature is an important environmental factor affecting the
growth of microorganisms. In cultivations of genetically modified cells, the expres-
sion of the foreign protein is often initiated by a temperature shift. Also in the pro-
duction of antibiotics the productivity can be increased by controlling the tempera-
ture to follow a certain profile. Since the kinetics of the total product proceeds ac-
cording to the specific production rate multiplied by the cell mass, the temperature
dependence must not only be included in the product model, but also in the growth
model. The response of a culture to variation of temperature can be very complex.
The composition of microorganisms with respect to DNA, RNA, protein, and lipids
varies significantly with temperature [32). The modeling of pH-influence is of special
interest for cultivations without pH-control. Like temperature, growth or production
is deterred when the conditions deviate from the optimum.
A convenient way is to consider the additional influencing variables, the parameter
vector p(t) in Eq. (2.2), only in the kinetic and stoichiometric coefficients, while
maintaining the form of the kinetics. When, for instance, looking at the tempera-
ture- and pH-dependence, p=(T(t), pH(t)) , a Monod-model for substrate uptake
may be written as

(c T H) _ -Mmax(T,pH) Cs
(2.36)
qs s, ,p - Yxs(T,pH) Ks(T,pH) + Cs

where the elements of p may be given by polynomials, e.g., for the maximum specific
growth rate,

Mmax(T,pH) = ko + kl T + k2T2 + k3T3 + k4 pH + kspH2 + k6 pH3+

(2.37)
+ k7TpH + ksT2pH + ...

A polynomial degree of two is sufficient when the parameter, here Mmax, has a
single extremum and no inflexion point. Otherwise, a higher degree has to be used.
Similar polynomials can be specified for the other model parameters. All the many
coefficients k of the polynomials have then to be determined from experimental data
sets, which cover a sufficiently great number of combinations of different tempera-
tures and pH-values in the region of interest. This is a tremendous experimental and
numerical task. Several models of this type that are rather specific to a certain strain
or process can be found in the literature. As a general finding, the yield coefficients
and limitation "constants" seem to be less sensitive to variation of the growth con-
ditions compared to the maximum reaction rates.
2.2 Unstructured Models 71

Eroshin et al. [27] investigated the variation of the yield with pH and temperature
when Saccharomyces cerevisiae was cultivated on ethanol as carbon source. The fol-
lowing polynomial correlation was found:
Y xs = 0.63 - 0.071xl - 0.012x2 + 0.0045xIX2 - 0.026xi - 0.053x~
pH - 4.5 T - 29 (2.38)
Xl = 2 pH X2 = 2 --T- , T in Celsius

Tayeb et al. [28] gave the following linear correlation in the range of 5.5<pH<6.5
and 33 C<T<44 C for the maximum specific growth rate of Streptococcus thermophi-
Ius:

J1max = (446 - 15.9 (T - 37) + 92 (pH - 6.5)) 10- 4 min- 1 (2.39)

and Lactobacillus bulgaricus:
J1max = (340 + 29.7 (T - 44) - 115 (pH - 5.5)) 10- 4 min- l (2.40)
An analogous correlation for temperature only was presented by Kluge et al. [56]
for the dependency of the specific growth and production rate of Penicillium chryso-
genum. Chu and Constantinides [29] correlated kinetic parameters to temperature
and pH for a Cephalosporin C process. For the cell yield from glucose the following
function was derived:
Yxs = 0.495 - 0.015~T - 0.194~pH2 - 0.144~pH3 - 0.057~T~pH
+ 0.039~T~pH2 + 0.134~T~pH3 + 0.006~T2~pH - 0.006~T2~pH2
- 0.018~T2~pH3 (2.41)
where ~T = T - 301K and ~pH = pH - 6.8
Reuss et al. [30] developed a model for the gluconic acid production by Aspergilus
niger. The dependency of the maximum specific production rate on pH was approxi-
mated by
qPmax =qPmax (pH= 5.5) (-4.07 + 1.84pH - 0.167pH2 ) (2.42)
A similar model was given by Venkatesh et al. [52] for growth of Lactobacillus
bulgaricus. Brown and Halsted [31] investigated the influence of pH on the growth
of Trichoderma viride on glucose medium. In a range of 2.5<pH<4.0 a linear corre-
lation of the maximum growth rate and the hydrogen ion concentration was found:
J1max = 0.1 (1- 309[H]+) h- 1 (2.43)

For pH-values lower than 2.48 no growth was possible at all. This model is in
agreement with the postulation that the growth rate reduction with increased hydro-
gen ion concentration is associated with a membrane diffusion limitation of the sub-
strate-permease complex. Starzak et al. [34] modeled the pH-dependency of the spe-
cific growth rate of anaerobic growing yeast Saccharomyces cerevisiae by a variation
of the half-saturation constant ([H]t) is the initial hydrogen ion concentration):

1 + kl [Ht + ([~'t-)2
Ks = + k2 (2.44)
1 + kl [Hlo + ([HI~)2
72 2 Bioprocess Models

A usual function to describe the variation of the specific growth rate with pH is
[48,54]

(2.45)

According to this model, Zeng et al. [35] investigated the growth of Clostridium
butyricum and Klebsiella pneumoniae, Akerberg et al. [49] of Lactococcus lactis, and
Ohara et al. [53] of Streptococcus faecalis. Tan et al. [55] presented a model based on
a statistical thermodynamic approach that resembles the form of Table 2.2, j, when
substituting Cs by [Ht. This may be taken as a generalization of Eqs. (2.44) and
(2.45).
More mechanistic models can be established when only looking at the tempera-
ture effects. The microorganisms can grow only over a restricted range of tem-
peratures. The metabolism is made of a complex sequence of enzymatic reactions
whose rates are individually related to temperature by activation or inactivation
that should generally follow a function of the Arrhenius-type. Beside the general
effect on the rates, the temperature can also exert highly selective effects on meta-
bolic pathways, e.g., by repression or induction of particular protein synthesis
(e.g., heat-shock proteins), or by irreversible denaturation of proteins and nucleic
acids. According to Franks et al. [64] the general dependency of growth on tem-
perature can be divided into several regions as shown schematic in Fig.2.11. In
the Arrhenius region, the growth rate increases exponentially with temperature up
to the optimum, Topt• In the superoptimal region, growth decreases by inactivation
effects. Balanced growth is possible up to a maximum temperature, Tmax. In the
supermaximal region, cells can only grow temporarily before they become com-
pletely inactivated. A segregated model for this transient effect is given in
Sect. 2.4. At low temperatures the growth rate falls more quickly with temperature
as expected by the Arrhenius-relation.
Topiwala and Sinclair [32] analyzed the balanced growth of Aerobacter aerogenes
in continuous culture on glucose mineral medium by means of Herbert's model.
Monod-kinetics were assumed for glucose uptake. While the yield decreased only
slightly with temperature, the other kinetic parameters varied strongly according to

109(/1)

Arrhenius

superoptimal
region

supermaximal
region

Fig.2.ll. Principal dependency of specific growth rate on temperature [64]

2.3 Structured Models 73

Arrhenius relationships. The following correlation was found in the range

293 K <T <323 K:

JiB = 2.7 . 105 e -~i'0

Ks = 3.3 . 10- 11 e'~~o (2.46)
10 -3400 23 -7860
Jimax = 2.5 . 10 e"'RT - 1.4 . 10 e"'RT

The maximum specific growth rate showed a maximum at 313 K. Esener et al. [10]
describes the influence of temperature on the microbial activity as superposition of
activation and deactivation effects on the key-enzymes of the reaction. The resulting
model for temperature dependence of the maximum specific growth rate with the
parameter values from [18] is

_ 6 .74· 10 14 e86800
RT h- 1
(2.47)
Jimax - 1 + 1. 6 . 1048 e28;~00

The Arrhenius-equation can also be used to model the kinetics of temperature

induced cell death: the number of viable cells decreases exponentially according to
dCx ~
- - = -kDe RT Cx (2.48)
dt
where ED is the "activation energy" for cell death and kD the inactivation rate con-
stant.

2.3
Structured Models

In unstructured cell models, the biotic phase is viewed as a homogenous component

with the only property, cell mass. In contrast, structured models provide informa-
tion about the physiological state of the microorganisms, their composition, and
regulatory adaptation to the environment. For this, the cell mass is structured into
several intracellular compounds and functional groups which are connected to each
other and to the environment by fluxes of material and information. The functional
groups may, in one extreme, consist of detailed reaction networks considering as
many as possible of known reactions or, at the other extreme, be built of roughly
lumped or idealized paths of the reaction network that may represent, for instance,
in only one step the entire biosynthetic system of the cell. While such a quite ab-
stract view leads to the low-complexity compartment models, the former approach
can be found in the so-called single-cell models that try to use as much as possible
a-priori information on mechanisms and kinetics of the metabolic reactions. Both
mark the field in which all of the structured models for the biotic phase can be
found: multi-compartment models, genetically structured models, and biochemi-
cally structured models. In any case, a structured model is of higher complexity than
an unstructured one because it is just the intention to give a better representation of
the intracellular processes by including balances for intracellular compounds. Never-
theless, as a custom of good modeling practice, a structured model should be nor-
mally constructed of as few as possible functional groups only for the most relevant
intracellular processes that provide a sufficient description for the desired applica-
tion of the model. This guideline keeps the model simple and the number of state
74 2 Bioprocess Models

variables and parameters as small as possible. Otherwise the difficulties for the ex-
perimental model verification and parameter identification become insuperable.

2.3.1
The Constitutive Equations

To be mathematically correct, all structured models must fit into a certain frame-
work of balance equations that is summarized in the following. A structured model
is built of balances for intracellular substances, metabolites, storage material, en-
zymes, cell wall material, RNA, DNA, and so on. These balances are coupled together
and to the outer liquid phase of the reactor by a number of reactions, as shown in
Fig.2.4. For the following, it is convenient to separate the concentrations of extra-
cellular compounds in the liquid phase of the reactor (the abiotic phase)

(2.49)

from the concentration of the biotic phase, which may represent intracellular meta-
bolites, storage material, enzymes, RNA, DNA, and so on:

CrnetaboUte )
Cenzyrne
X = (Xi) = ( . (2.50)

CDNA

The entire concentration vector can be written as

(2.51 )

The first constitutive equation of structured models is obtained from the condi-
tion that the total cell mass Cx is just the sum of all its components:
(2.52)

In an unstructured model, the composition of a cell was taken as constant, Xi=

constant; then the biotic phase is already uniquely characterized by Cx. The intrinsic
reactions and related kinetics are the main target of structured modeling. These ki-
netics are not only governed by the concentrations in the bulk liquid phase outside
the cell but also by the intracellular concentrations. Therefore it makes sense to write
down the balance equations for structured models in terms of intrinsic concentra-
tions. When doing this, one has to take special care of the models consistency: the
balance over the entire reactor must be equivalent to the balances over the biotic
phase [2,95,96]. The volumetric intrinsic concentrations are defined in the average
as mass of the substance in the cell per volume of the cell:
Xi VL +- total mass of substance i in all cells in the reactor
(2.53)
CiV = Cx VLPx:1 +- total volume of all cells
2.3 Structured Models 75

Since the cell density Px is almost constant and its remaining variations difficult to
be measured, it is more convenient to use cell mass related intrinsic concentrations:
x
c=- (2.54)
Cx
From Eq. (2.52) then follows the second constitutive equation for the sum of all
intrinsic components,

Lei = I Tc = 1 (2.55)

and therefore,
d(I T c)
--=0 (2.56)
dt
The derivation of the balances for intrinsic concentrations starts with the balance
equation of the reactor, Eq. (1.12), where the gas-liquid transfer here need not be
considered. By partitioning the rate vectors and matrix of yield coefficients in ana-
logy to the concentrations, Eq. (2.51),

Q=(~) q= (~) Y= (~;) (2.57)

the balances for compounds in the outer and inner of the cell can be written down
separately:
Qz
,----...
'Iz
dZ (t) ,.-"--0..
----;{t = Yzr(t) Cx(t) +D(t) (ZR(t) - Z(t)) (2.58)

dXd(t) = Yxr(t) Cx(t) +D(t) (XR(t) - X(t)) (2.59)

t "-v--"
'Ix
'----v---'
Qx

In the following, only the elements related to the biotic phase, Qx and Yx, will be
further considered. By substituting Eq. (2.54) into Eq. (2.59), together with
(2.60)

the balance for the cellular compounds is obtained as

dCx(t)c(t) _ () dc(t) ( ) dCx(t)
dt - Cx t dt + c t dt (2.61 )
= Yxr(t)Cx(t) + D(t)c(t) (cf(t) - Cx(t))
Multiplying 1T to each summand and using the conditions of Eqs. (2.55) and (2.56)
yields

dC;?) = IT(Yx r(t))Cx (t) + D(t)(cf(t) - Cx(t)) (2.62)

76 2 Bioprocess Models

which is just the balance equation of the total cell mass, and the factor
I T(y x r(t)) = J.t(t) (2.63)
must equal the specific growth rate. This is the third constitutive equation of struc-
tured models. The specific growth rate is never independent of the other intrinsic
reaction rates. When substituting Eqs. (2.62) and (2.63) into Eq. (2.61), the transport
term due to inflow and outflow of the reactor is canceled out and the balance equa-
tion for the intrinsic concentrations, the fourth constitutive equation of structured
models, follows as
I'(t)
dc(t) ~
----at = Yxr(t) - IT(YXr(t)) c(t) (2.64)

The second summand in this equation is due to dilution of intracellular material

by the cell growth, i.e., the expansion of the cell volume. The four constitutive equa-
tions form the general frame of every structured model that may be established by
different sets of rand Yx .
For biochemically structured models it is convenient to write down molar bal-
ances instead of mass balances. By the substitutions
r=Mr Yx = MYx c = Me with IT(Me) = I (2.65)
where A denotes molar quantities and M is the diagonal matrix of molecular
weights, the balance equation, Eq. (2.64), becomes

de( t) =
----at Yxr t - J.t t c t
A A ( ) ( ) A ( )
(2.66)

and the specific growth rate is then

J.t(t) = ITM(Yxr(t)) (2.67)

The consumption of metabolites for maintenance metabolism can be modeled in

analogy to unstructured models by considering a zero-order reaction on the rate
vector r, as given by Eq. (2.26). If one prefers an explicit modeling of maintenance
turnover, the balance equation, Eq. (2.64), becomes
dc( t) T
----at = Yxr(t) - I (Yxr(t))c(t) - m (2.68)

and the growth rate is then

I'G I'E
..----... ,--A-..
J.t(t) = IT(YXr(t) -m) = ITYxr(t) - ITm (2.69)
This equation is a unification of the models of Herbert, Eq. (2.20), and Pirt,
Eq. (2.24), since application of the latter to intrinsic variables leads to an endogenous
growth rate J.tE=1 Tm.
Some more comments on the area of validity of the above structured model and
the special condition, Eq. (2.60), for the inflow from the reservoir have to be given.
The cell suspension is essentially a macro-fluid. This has serious consequences for
the global mass balances and reaction kinetics, and restricts the application area of
structured models. Consider a continuous process with sudden cell inflow. Then,
2.3 Structured Models 77

generally, the in-flowing cells will have a state c different from the cells already being
in the reactor. Because the intrinsic substances are not exchanged between the cells
ad hoc, two clearly distinguishable cell populations with different states will keep
growing and the observed average reaction rate for the entire reactor will be differ-
ent from the reaction rate calculated from the kinetic expressions, r, by using aver-
aged concentrations for intrinsic compounds. The only exception - without practical
relevance in biological systems - is the case of linear first order kinetics. Therefore,
structured models are only valid for reactors without inflow of cells, or with in-flow-
ing cells that have at any time an exactly identical state to the cells in the reactor.
This is only the case for cell-recycle from the same reactor. The more general case
can only be covered by distributed population balance models [120). An equivalent
problem arises when using segregated population models (Sect. 2.4) in connection
with structured growth models. There, the "inflow" of cells to a population results
from transitions of another popUlation growing in the same reactor. Such models are
only mathematically correct when both populations have identical intrinsic states.
This is not always considered in the works published in the literature.
The situation in reactors with non-ideal mixing is similar to the ideal CSTR with
inflow of cells. When the cells pass through regions with different concentrations,
e.g., of substrates or oxygen, their intrinsic state is altered slightly. Since the passes
of cells through the reactor follow a stochastic regime, cells of different state become
neighbors in a small volume element. This means, the population is always inhomo-
geneous and structured models must be applied with care in connection with dis-
tributed reactor models. The best way to avoid mathematical difficulties is to assume
(if justified) that the recirculation time is much lower than the time constant of in-
tracellular processes. The biological model can then be taken as a lumped one, as-
suming homogeneous biomass in the entire reactor having everywhere an identical
intrinsic state. Then, averaged concentrations of the abiotic phase have to be used
for calculation of the intrinsic reaction rates.

2.3.2
Some Applications of Structured Models

Compartment models introduced by Williams [65) as low complexity structured

models are obtained by lumping biological systems of similar function and dy-
namics together into a few pools. The average behavior of the i-th pool is modeled
by a single variable that can be a representative concentration for a group of meta-
bolites or enzymes. These models are mainly used to describe very general features
of the biomass, such as variation of the RNA or DNA content during the cultivation
or the regulation of complete functional groups of the metabolism, e.g., pathways for
certain substrates, or products. Into this area fits the modeling of lag-phases of
growth being observed during growth on multiple substrates. An example for dia-
uxic growth is given below. Esener et al. [96) developed a two-compartment model
for growth of Klebsiella pneumoniae. One compartment was assigned to genetic ma-
terial, proteins, and structural compounds, the other to RNA and small metabolites.
The model was able to describe lag-phases of growth on a complex substrate. Nielsen
et al. [97, 98) presented a two-compartment model for the lactic acid fermentation
by Streptococcus cremoris with similar structural pools as in the above model. The
model was verified by measurements of intracellular RNA and by dynamic experi-
ments. The model was extended to four compartments by Nikolajsen et al. [99) for
78 2 Bioprocess Models

cultivations on mixtures of different sugars. Mundry and Kuhn [102] published a

two-compartment model for phosphate-limited growth of Streptomyces tendae in
fed-batch culture. Korte et al. [66,67] investigated the growth of Escherichia coli on
complex substrates including a variety of amino acids and developed a respective
multi-compartment model. A model for antibiotics production by different Strepto-
myces strains that considers, besides glucose, the uptake of ammonium and phos-
phate was presented by King [100, 101]. The model, including six compartments
for DNA, RNA, protein, amino acids, nucleotides, and structural compounds, was
successfully applied for simulation of fed-batch and shift experiments.
Compartment models have also interesting applications for genetically modified
cells, since the system for foreign gene expression is usually inducible. The model
of Nielsen et al. [68, 69] for run-away cultivations of Escherichia coli includes four
compartments for RNA, host genome, plasmid genome, and foreign protein. Bentley
and Compala [70] consider further in their 8-compartment model amino acids, nu-
cleotides, lipids and native protein. By choosing proper kinetics, the models could
describe the variation of total plasmid or RNA content with growth rate. Lee and
Ramirez [71] modeled the expression dynamics of foreign protein as superposition
of shock reaction after temperature shift and subsequent recovery of the cells. The
model could mirror the effect of IPTG-inducer on production and cell growth. Cop-
pella and Dhurjati [108] presented a model of recombinant yeast for production of
human epidermal growth factor, including seven intrinsic balances. The model gave
a reasonable fit to experimental data.
On the level of molecular mechanisms, the kernels of the models for regulation of
substrate uptake by the operon and for foreign gene expression are quite similar.
Both include the two steps of transcription of a particular gene with intrinsic con-
centration CG into mRNA with intrinsic concentration CmRNA, and subsequent transla-
tion of the mRNA to the key enzyme for substrate uptake, or to the foreign product
(cp). The according balance with consideration of first-order deactivation or decay -
with rate constants kDJ and kD2 - can be formulated as follows:
dCmRNA
~= kl~CG -kD1CmRNA - ~CmRNA (2.70)

(2.71)

The transcription efficiency ~ accounts for modulation of transcription by an op-

erator system. It is a complex function of the concentrations of effectors, e.g., sub-
strates, repressors, and inducers. Often the transcription is the rate-limiting step, so
the mRNA balance has to be included in a model for a proper description of the
induction dynamics. In more simplified models the mRNA is assumed as being in
quasi-steady-state and the mRNA-concentration can be eliminated from Eq. (2.71).
The translation efficiency ~ in the synthesis rate of the protein describes the affinity
of the ribosome binding site to the given nucleotide sequence, and the availability of
ribosomes. The kinetic expressions for both factors are crucial parts of the models
and were targeted in many of the papers mentioned here. For genetically modified
cells, where the foreign gene is coded on a plasmid, the concentration CG may also
vary. Then DNA synthesis and plasmid replication or loss must be modeled in a
proper way. Such models are known as genetically structured models [76, 79-82,
109-115].
2.3 Structured Models 79

Table 2.7. Example for a two-compartment model for growth on two alternative substrates
[18]

Description Model equations

a General vector notation, see dC(t)

Eqs. (1.12), (2.1), and (2.64) ----;Jt = Q = qCx = YrCx
dc(t)
----;[f = Yxr(t) - Jl(t)c(t)

b Specialized vectors and matrices

c=

Y=

c Total cell mass Cx = CEl + CEl 1 = eEl + eE2

d Intrinsic rate expressions rSl = rSmaxl KSI + CS1

CS2 Kn

CS1 CS2 Kn
e Specific growth rate, Eq. (2.63) Jl = rSmaxl + rS max 2 eE2
KSI + CS1 KS2 + CS2 Kn + CS1
f Explicit model equations
for the liquid phase

,,
,,
,
\
,,
I
I
I
I
I
I

,,
I

... - ... _------_ ........

Fig. 2.12. Structure of the two-compartment model for diauxic growth of Klebsiella terrigena
on glucose and maltose
80 2 Bioprocess Models

,,
t,.\,
,
4,
,
16 20
C S2 C S1 ex t[h]
[kgm ~3] [kgm ~3] [kgm ~3]

Fig. 2.13. Diauxic growth of Klebsiella terrigena on glucose and maltose, experimental data
and simulation by a two-compartment model [18]

The most complex group of detailed mechanistic models tries to give a complete
image of the biochemical processes of the cells, including replication of the genome
and cell division. Although they must consider an enormous number of metabolites
and reaction steps, they are in a certain aspect the most handy models: the elemen-
tary reactions have naturally the most simple kinetics, and the model parameters
have direct physical or chemical meaning; both can often be examined by in vitro
experiments. The single-cell growth models for Escherichia coli of Domach et al.
[72], Ataai and Shuler [73], Shu and Shuler [74], and Peretti and Bailey [75] served
as basis for model studies on genetically modified strains. One focus of model ap-
plication was in the area of host-vector-interaction and mechanistic description of
plasmid replication [77, 78, 116-118]. These models can provide theoretical predic-
tions on a number of parameters that are of interest for simple segregated models,
e.g., the variation of maximum growth rate, yield coefficients, or plasmid copy num-
ber with the properties of the vector or with growth conditions.
As an example of a low-complexity structured model, a two-compartment model
for diauxic growth of Klebsiella terrigena on the substrates glucose (CS1 ) and maltose
(CS1 ), is summarized in Table 2.7 and its structure is shown in Fig. 2.12. Glucose is
the preferred substrate that inhibits and represses the uptake of the second substrate
maltose. The inducible enzymes of the maltose uptake system are represented by the
Ez-compartment, while EJ-compartment stands for all of the remaining metabolism.
When glucose becomes limited and maltose is present in the medium, E2 is synthe-
sized with increased rate by derepression on rS2 and degradation of Ej • The diauxic
lag-phase is modeled by the resulting transient on CE2' One of the balances in Table
2.7 f is obsolete since, due to condition c, the concentration of either Ej or E2 can be
calculated from the other. A simulation result for a batch cultivation is given in
Fig. 2.13.
2.3 Structured Models 81

2.3.3
Cybernetic Models of the Compartment Type

A growing microbial cell breaks down high molecular weight carbon and energy
sources, brings the smaller derivatives into the cell, catabolizes these further to smal-
ler molecules, converts them to amino acids, nucleotides, carbohydrates, fatty acids,
and vitamins, and finally builds cell constituents such as proteins, nucleic acids,
polysaccharides, and lipids. Hundreds of enzymes have to be synthesized and must
then act together in a coordinated way. Thus, regulatory mechanisms have evolved
that enable a species to compete successfully with other forms of life in nature: feed-
forward regulation (e.g., substrate induction and activation), feedback regulation by
repression or inhibition (e.g., by catabolites or primary metabolites), or energy
charge regulation of biosynthetic pathways. These control mechanisms direct the
substrate and metabolite flux to proper pathways. The rate of an enzyme-catalyzed
reaction is controlled by the enzyme concentration itself, and by the regulation on
the reaction level by activation/inhibition to meet the requirements of metabolism
and to ensure optimal growth. An ideal cell does not overproduce extremely large
amounts of metabolites or enzymes, no matter what are the growth conditions. This
coupling of independent reactions by the network of metabolic regulation often
makes the understanding of the metabolism and its modeling so difficult. It is never
enough to look at a few reactions; instead one has to adopt a more global viewpoint.
In a very radical manner this is realized by the approach of cybernetic models for
multi-substrate growth kinetics, introduced by Ramkrishna [83], and modified in
several papers afterwards [84-88, 90, 93, 94).
Sequential and parallel uptake of more than one substrate at a time is accompa-
nied by a complex metabolic regulation, which cannot be included satisfactorily into
simple kinetics. Also usual structured models often will not give a simple and use-
able description in this situation, because they have to contain many variables, com-
plicated mechanisms, and kinetics for the metabolic regulation. In the cybernetic
approach, the mechanisms by which the microorganisms can coordinate their me-
tabolic reactions to obtain optimum control are not questioned. The concept relies
on the assumption that the evolution had selected the mechanisms performing just
optimal control, and that it is sufficient to identify the optimality criterion. When
this criterion and some general restrictions on the growth process - or more exactly,
the substrate uptake steps and the biosynthetic capacity for enzymes - are known, it
is possible to calculate at any time the rates of reactions by a mathematical optimi-
zation procedure without looking at the underlying mechanisms. In this way, cyber-
netic models can sometimes predict the sequential and parallel substrate consump-
tion, without any assumption about coupled kinetics, by extrapolation from experi-
ments on single substrates. The modeling proceeds in three steps:
l. Assigning an optimal control motive to the cell
2. Establishing kinetic rate expressions for uptake of a single substrate
3. Modeling of the slow dynamics of regulation by induction and repression, as de-
termined by the synthetic capacity of the cell

The metabolic control action in cybernetic models is twofold: first, by adjusting the
level of key enzymes for each substrate uptake system according to the optimality
criterion, and second, by controlling the activity of the key-enzymes of reaction in
analogy to a fast mechanism of activation or inhibition. Each key-enzyme forms its
82 2 Bioprocess Models

own compartment. In this view, the microorganisms are always able - by allocating
their internal resources - to choose those substrates which fit their requirements
best. For formulation of the model, the microbial growth on each substrate compo-
nent Si is represented by a reaction

Key-enzyme i
X + Si ' (1 + YXSi)X (2.72)

and the substrate uptake is given by a kinetic expression that depends on the con-
centration of the substrate, CSi , and of the related key enzyme, CEi:

(2.73)

The enzymes are synthesized according to an optimum strategy given later. The
specific growth rate on substrate i, J-li, is a product of three factors, the intrinsic
concentration of the key enzyme, CEi, a kinetic expression, J-lmaxr(CSi), chosen for
instance from Table 2.1, and an activity controlling variable, Vi:

(2.74)

The total growth rate for N substrates is

N
J-li(CS1 )'" CSN ) t) = I>max;ri(Csi(t))Vi(t)CEi(t) (2.75)
i=1

The partition of the control action into one part for long term response by induc-
tion and repression, modeled by cElt), and another part for immediate response by
inhibition and activation, as described by vlt), mirrors the two hierarchical levels of
regulatory mechanisms. These may become visible at diauxic growth for example:
the cells may switch to the preferred substrate at once when this is suddenly added
to the medium, even if the less preferred substrate is present at high concentrations
and the related enzymes are induced. The uptake of the less preferred substrate is
then immediately inhibited, followed by a slower degradation of the enzymes. The
induction-repression dynamics of key-enzymes are described by balances for their
intrinsic concentrations,

(2.76)

where KEi is the degradation rate constant. The synthesis rate of the key enzymes is
controlled by the cybernetic variables u;(t) and by a kinetic expression depending on
the substrate concentration. A small constitutive formation rate rEiD guarantees a
minimum activity of the pathway. To complete the model, the control variables have
to be specified. The cybernetic variables

o S; Ui(t) S; 1 with LUi(t) = 1 (2.77)

fully repressed fully induced
2.3 Structured Models 83

are determined through Herstein's matching law, which maximizes the total profit,
here taken as J.L, by allocating fractions of a fixed resource to alternative pathways i,
according to the expected maximum relative profit for the i-th pathway, given by J.Li,
expected maximum profit of pathway i
"
Pmax.iTi( CSi (t) )CEi( t)'
Ui ( t) = ----:N-=------'------'--'---'-- (2.78)
L J.Lmaxi) (CSj(t) )CEj(t)
j=l
'~----~v~------~
total maximum profit

The fixed resource here is the total amount of enzymes in the cell, which cannot be
synthesized in arbitrary amounts because the number of ribosomes is limited.
Therefore, an increased formation of one enzyme is going on the costs of another.
The activity control variables
o S; Vi(t) S; 1 (2.79)
fully inhibited fully activated

for fast regulation of the key enzymes by inhibition and activation are allocated
according to a heuristic strategy, in the way that pathways yielding a higher growth
rate are preferred:
expected maximum profit of pathway i
"
Pmax;l;i(CSi(t))CEi(t)'
Vi ( t ) = --'----7"''''-'----'-'---=-'-'-'----=-:--'-:--c___- (2.80)
max (J.Lmax/j (CSj( t) )CEj( t))
)
, v-----"
maximum attainable profit among all pathways

Obviously, the above optimization strategy is a local one. It only considers the
actual growth conditions without trying to optimize the response over a certain in-
terval in the future.
A formalism similar to the above can be used to consider maintenance metabo-
lism. Another extension of the concept was given by Straight and Ramkrishna [89,
91, 92] for the modeling of connected metabolic pathways, and Yoo and Kim [103]
have proposed a variant for non-growth-associated product synthesis (poly-iJ-hy-
droxybutyric acid). Doshi et al. [105] and Venkatesh et al. [106] suggested a simpli-
fied procedure for determination of the cybernetic variables for sequential substrate
utilization. Chetan et al. [119] presented a cybernetic model for growth of baker's
yeast on glucose, galactose, and melibiose.

2.3.4
Cybernetic Models of the Metabolic Regulator Type

Another modeling method with similarities to the above cybernetic models is the
metabolic regulator approach [18]. The ideas behind cybernetic modeling were out-
lined in the previous section. Again, the microorganism is modeled as an optimal
strategist, which tries to use the different metabolic pathways optimally. This model
concept is not restricted to substrate uptake steps, but includes the general model of
the reaction network as given by Eq. (2.2), including further intracellular reactions
and product synthesis. Here, the role of the net reaction rates q as dependent vari-
84 2 Bioprocess Models

abies of the intrinsic reaction rates r is not fixed for any growth condition, but ele-
ments of q and r can change their role in different growth situations. Furthermore,
the intrinsic reactions are not solely determined by kinetic expressions, but also by
the optimization strategy behind the metabolic regulation. Therefore, it is conveni-
ent to rewrite the stoichiometric model in the form
Yr(t) = met) (2.81)
without explicitly distinguishing between net and intrinsic reactions. Here, r is an
extended reaction vector also containing elements of q, and m is a vector of M
growth independent specific reaction rates, like maintenance terms or rates of con-
stitutive reactions. But mostly the elements of m will be zero. In the above stoichio-
metric model the number of reactions, N, will usually be larger than the number of
stoichiometric equations M, otherwise the microorganisms would have no freedom
to regulate the relative activity of their pathways. The way to derive more conditions
and finally to solve Eq. (2.81) uniquely is to introduce a cybernetic criterion, the
metabolic coordinator J, for determination of the optimum metabolic regulation:
r(t)
J(r(t)) ---7 optimum (2.82)
under consideration of boundary conditions given by Eq. (2.84). The optimization
strategy is again a local one, since J depends only on the actual reaction rates. Simi-
lar to the cybernetic models of the compartment type, the optimization strategy
followed by the microorganisms is often successfully modeled by maximization of
the specific growth rate,
r(t)
J ;= f-t(t) ---7 maximum (2.83)
Since the stoichiometric model, Eq. (2.81), is underdetermined, additional condi-
tions have to be considered to obtain a meaningful solution of the optimization pro-
blem. These are possible rate limiting steps of the metabolism, which can be formu-
lated as inequalities,
(2.84)
as further explained by Table 2.8. Remember that rj<O means uptake and rj>O pro-
duction. Such possible rate-limiting steps might be an inherent maximum biosyn-
thetic capacity for macro-molecules and cell material (a) or kinetics of substrate

Table 2.8. Possible boundary conditions in Eq. (2.84)

Kinetics Expression for rmin or f max Meaning

a Zero order r min = constant or r max = constant Saturation

b Single substrate uptake rminr( Cs) or rmaxr( Cs) Substrate limitation
c Time function, as given rmin(t) = rEj(t) or rmax(t) = rEj(t) Regulation of the key-enzyme
by Eq. (2.85) of the pathway by induction
or repression
d Infinite ±oo Never limiting
e Combinations of max(rmin), min(rmax) Switch between different
(a) to (c) mechanisms
2.3 Structured Models 85

uptake (b). By using case (a) with ro=O on one side ofEq. (2.84), the rates can also be
restricted to either uptake or production. Of special importance for the modeling of
microbial adaptation are rate-limiting steps due to regulation of key enzymes by
induction and repression (c). Beside this slow regulation, implicit in the model there
is a fast regulation on the level of enzyme activity, since, as an outcome of the opti-
mization, the turnover of some reactions may be lower than the maximum turnover
determined by the boundary conditions. The vectors of possible rate-limiting steps
must contain constraints (finite rates) on at least N-M reactions; by the optimization,
at least N-M elements of r will always equal a constraint, i.e., the corresponding ele-
ment of r min or r max. Such reactions are said to be actually rate limiting. Recently,
Bonarius et al. [104] have discussed the problem of finding the inherent constraints
of metabolic networks that could be presented by a model as given above. They also
suggested an experimental procedure for their determination.
The regulation on reaction level by inhibition/activation is completely covered by
the above model without further assumptions. When condition (c) in Table 2.8 is not
rate-limiting or when such conditions are not used, the model describes the fully
adapted, optimum state of growth, in which all inducible enzymes are present in
sufficient amount. Then the model is essentially an unstructured one that uses ki-
netics of the non-interacting type and a refined reaction network. For consideration
of the dynamics of adaptation, balance equations for key-enzymes have to be speci-
fied which determine rEj(t). In this approach, a strategy is used which is local to the
regulated pathway. Their properties are as follows: as long as the coordinator J uses
fully the capacity of the pathway up to the actual maximum rate rEj(t), the concen-
tration of the key enzyme has to be increased by further induction. If the actual rate
allocated to the pathway is not rate-limiting, then the concentration of the enzyme
should be reduced by stronger repression to manage the resources economically.
These conditions can be automatically met by a tracking controller, a metabolic reg-
ulator, for the concentration of the key enzyme, which is regulated according to the
actual requirements rj(t):

(2.85)

where rEj is proportional to a fictive intrinsic concentration of key enzyme for the
pathway j and rOj is a small constitutive level of enzyme synthesis. The regulation has
sufficient gain and the required tracking properties when

(2.86)

The balance at Eq. (2.85) for synthesis of a key enzyme is under growth limitation
of an autocatalytic nature, since then the actual rate rj is proportional to rEj and to
the enzyme concentration. Thus, during limitation, the regulator is unstable and
therefore increases the enzyme concentration until the regulated pathway has suffi-
cient capacity. The importance of autocatalytic enzyme synthesis for modeling of
growth dynamics on mixtures of substrates was recently discussed by Narang et al.
[107].
An example of this type of model is given in Chap. 9. By this concept, growth on
multiple substrates with sequential and parallel uptake, including essential com-
pounds, as well as shifts in the product spectrum can be successfully described.
86 2 Bioprocess Models

2.4
Segregated Models

Unstructured and structured models presumed a homogenous population of cells

and only one species in the bioreactor. Many phenomena cannot be described by
these unsegregated models:
- Irreversible alteration or disturbances in physiology and cell metabolism
- Morphological differentiation of the cells
- Mutations in the genome with related variations in the metabolism
- Spatial segregation of growth regions, attachment of cells to surfaces, or aggrega-
tion of cells
- Growth of more than one species in the bioreactor by intentional mixed cultures
or infections
Simple segregated models for these phenomena can be built on the basis of ordinary
differential equations by discriminating several classes of independent cells. Each
cell class is mostly described by unstructured models, but with great mathematical
care structured models can also be used. Here, only models with a finite number of
cell classes will be considered, while segregated models with a continuous variation
in cell properties that have to be based on partial differential equations are referred
to in Sect. 2.4.6. The applications of segregated models are quite widespread and
manifold so that it is difficult to construct a general and useful framework as for
structured models. So, this chapter will rely mainly on some selected examples.
Modeling of growth of multiple species that interact only indirectly, for instance by
competing for substrate, is beyond the scope of this chapter. A review on this topic is
given in [IJ.

2.4.1
Simple Segregated Models

For the general formulation of simple segregated models, which are also schemati-
cally represented by Fig. 2.4, the vector of concentrations of the liquid phase is simi-
larly to structured models (see Eq. 2.51) split into two parts: the first, Z, contains the
concentrations of the abiotic phase such as substrates and cell-external products,
and the second, X, the concentrations of the biotic phase. When restricting the mod-
els for the N cell classes of the biotic phase to the unstructured type, the vector X
includes only the biomass concentrations for each class:

X = (~;~) (2.87)

CXN
Now every cell class or species is described by a stoichiometric model, as given by
Eq. (2.2):
qi(C(t» =Yiri(C(t» (2.88)
where i denotes the index of the cell class. The vector r of intrinsic reactions now not
only includes kinetic expressions as shown in Sect. 2.2, but also describes the transi-
2.4 Segregated Models 87

tion of cells from class i to the other classes by appropriate rate equations. In the
balance equation of the liquid phase, the contributions of all classes have to be con-
sidered. Equation (1.12) for a CSTR is then modified to
dC(t) N
-----;Jt = ~~ + D(t) (CR(t) - C(t)) + TR(t) (2.89)
Qi(t)

When all net reactions qi of one cell class are concatenated together to a matrix of
net reactions for all cell classes
(2.90)
the notation of the previous balance is simplified to
dC(t)
-----;Jt = q(t)X(t) + D(t) (CR(t) - C(t)) + TR(t) (2.91 )

The possible transition of cells between classes causes some problems when struc-
tured models are employed for the cell classes. For the target class, the transition
means just the same as an inflow of cells from a reservoir in a continuous cultiva-
tion. Therefore the same restrictions are valid as discussed for structured models in
Sect. 2.3: the application of such models is only correct when both the source and
target cell classes have identical state at any time. Therefore, structured segregated
models with cell transition from one class to another are only valid when cells of
both classes have been grown from a single initial population, and when their differ-
ent properties do not lead to different intrinsic states afterwards. These restrictions
are not always observed in structured-segregated models in the literature.

2.4.2
Segregated Models for Physiological Properties

Many factors can harm microorganisms or cells and change irreversibly their phy-
siology: heat, radiation, extreme pH-values, long periods of starvation, and so on.
The damages caused under those circumstances often alter the structure of cell con-
stituents such as membranes, cell wall, proteins, or DNA, but the cell as a whole may
survive without being completely destroyed. Such extreme treatment of the cell po-
pulation may prevent some of its parts from further regular cell growth, the popula-
tion becomes inhomogeneous with respect to its growth and propagation activity.
Tempest et al. [121] found, during continuous culture of Aerobacter aerogenes, only
40% of the cells were active at strong limitation due to low dilution rates
(D=0.004 h -1), while under better substrate supply at high dilution rates (D>O.S h -1)
the percentage increased to more than 90%. This variation has great influence on the
yield and productivity of the process. Another related phenomenon is a reduced but
clearly non-zero cell concentration at very low dilution rates. Under such conditions
the observed cell yield can be much higher than that predicted by the simple main-
tenance models, Eqs. (2.20) and (2.24). Sinclair and Topiwala [123] presented a sim-
ple segregated model (see Table 2.9) that discriminates for the entire population two
cell classes: viable (or active) cells, Xl> and non-viable (or dormant, dead, inactive)
cells without metabolic activity, X2 :
(2.92)
88 2 Bioprocess Models

Table 2.9. Segregated model including viable and non-viable cells [18]

Description Model equations

a Specialized vectors, r considers

endogenous metabolism and
maintenance 1 (~I
y =
YXSI
~1 -1
0
~OI )

Y2 = r2 = 0

b Explicit model equations dCx1

-- = (I"I - I"E - KI2 - D)Cx1
for the liquid phase dt
dCX2
dt = K 12 Cx1 - DCX2

dCs -I"I
- = --CX1 - msCx1 + D(CsR - Cs )
dt YXS1

By the fraction of viable cells,

CX ! CX !
a=-= <1 (2.93)
Cx CX ! +
CX2 -

which is under steady-state conditions a function of the specific growth rate, respec-
tive dilution rate in a chemostat,
D
a=--- (2.94)
Kv+ D
the cell concentrations of both classes can be expressed as
CX ! = (1- a)Cx CX2 = aCx (2.95)
In steady-state, the true specific growth rate {ll of the viable cell population is then
increased compared to the observed growth rate qx of the entire population:
qx
{l! = - ( - ) + {lE {l2 = 0 (2.96)
a qx
where {lE is the rate of endogenous metabolism. In the model, {llCs) can be chosen
as Monod-Kinetics; K12 is the specific formation rate of inviable cells. Veilleux [124]
applied such a model for simulation of a two-stage continuous cultivation. In both
stages the concentrations of viable and non-viable cells could be nicely fitted. In
addition to the original model [123], maintenance requirement of the viable cell
fraction was introduced for the substrate uptake by the specific rate ms [18]. This
extended model also covers Pirt's concept of dormant cells [122]. For steady-states in
a chemostat, the observed yield can then be calculated as
Kv+ D
Yxs = YXs ! ---'----- (2.97)
+ +
Kv D msYXS !
2.4 Segregated Models 89

At high dilution rates, observed yield and the fraction of viable cells attain their
maximum values. For D-+O the viable cells disappear completely, while the observed
yield - and the observed biomass - clearly remains at non-zero values.
The influence of temperature on the growth can often be described by unstruc-
tured models by means of a functional variation in the model parameters, as dis-
cussed in Sect. 2.2. But the temperature influence can be more complex. For growth
of Streptococcus cremoris it was found that anabolism already breaks down at lower

Fig.2.14. Structure of the segregated model for growth at superoptimal temperatures [64]

Table 2.10. Segregated model for temperature inactivation [64]

Description Model equations

a Specialized vectors

b Explicit model equations dCXI

dt = (ILl - Kl - K2 )CXI
dC
X2
dt = KICXI - K2 CX2
dCX3
dt = K2 CXl + K2 CX2
dCs -ILl
- = - - CX! - m S CX2
dt Yxs l
90 2 Bioprocess Models

temperatures than catabolism, the latter being decoupled from the first. The anabolic
inactive cells are still able to catabolize substrate and secret metabolites, but are un-
able to grow. Furthermore, after temperature-shifts cells can grow for a while at
temperatures that lead, after a prolonged time, to a complete inactivation of all cells,
thus preventing balanced growth. Franks et aI. [64] developed a segregated model for
this phenomenon of temperature inactivation. The cell population can be repre-
sented by three classes - fully active cells, Xl, catabolic active/anabolic inactive cells,
X2, and catabolic inactive cells, X3 • The structure of the model is shown in Fig. 2.14
and the model equations are summarized in Table 2.10. The rate constants were as-
sumed to have Arrhenius-form:

(2.98)
The model could successfully describe the experimental results for balanced
growth at various temperatures, as well as transient growth at supermaximal tem-
peratures according to Fig. 2.11.

2.4.3
A Model for Spatial Segregation by Wall Attachment

Attachment of cells to surfaces is an often observed phenomenon in biotechnology.

In continuously operated bioreactors it leads to the effect that cells do not wash out
even when the dilution rate is above the critical value. The reason is a continuous
inoculation of the suspended population by cells detaching from the wall. Kreiken-
bohm and Stephan [125] published a segregated model for anaerobic growth of Pe-
lobaeter acidigalliei on gallic acid, containing two cell classes, the first, Xl, is homo-
geneously suspended, the second, X 2 , forms a thin homogenous biofilm on the reac-
tor wall with small constant thickness as shown in Fig. 2.15. The biofilm is assumed
as being well supplied with substrate. Such a model can help to estimate the rate
constants of cell attachment and detachment. After introducing the volume fraction
v of cells attached to the wall,
AL
v=- (2.99)
VL
where VL is the reactor liquid volume, A the surface of the wall and L the thickness
of the biofilm, the model can be noted as in Table 2.11. There, K12 and K21 are the
rate constant of adsorption and desorption, respectively, KlL and K2L are specific
rates of cell lysis.

Reactor

Suspended cells
CX1 VL

-4~
Biofilm i/A
Cx2 / ... _Iiiioooo_....
Fig. 2.15. Schematic diagram for a suspension culture that exhibits wall growth
2.4 Segregated Models 91

Table 2.11. Segregated model for continuous cultivations with cell attachment to the reactor
wall [125]

Description Model equations

C") C -I) (;,; )

a Specialized vectors -1
0 0 V-I
c= CX2 Yl= rl =
-1
Cs YXSI
0 0 K12

C
0
Y2 = 1
-v ~I
YXS2
) ,,~(;:,)
-1
0 K21

b Explicit model equations dCX1

dt = (Jll - K12 - KIL - D)CX1 + K21 CX2 v
dC K C
-dt-
X2
= (Jl2 - K21 - K 2L )CX2 +12-v -
X1

dCs Jll Jl2 R

- = ---CX1 ---vCX2 +D(Cs - Cs )
dt YXSI Y XS2

By a steady-state analysis of the model, the following approximate condition for

the critical dilution rate can be derived for high substrate concentrations in the med-
ium:

+ (2.100)
K21
"
+ K2L - /12 max
~
Deritl
!:lDcrit
Critical dilution rate of suspended cells
Increase by wall growth

It is increased by !:lDerit over the value Dcritl where a culture without wall attach-
ment would wash out totally. Here, above Dcritl only the concentration of suspended
cells begins to fall, but the steady states remain stable up to Deril> where the final
wash-out occurs. This is a major difference to the model of Topiwala and Hamer
[126]. There the biofilm was assumed to form a constant monolayer, CX2 =constant,
which results for high dilution rates in a hyperbolic relation

CXl -_ /12max CX2 (2.101)

D-/1lmax

without wash-out.

2.4.4
Segregated Models for Morphological Differentiation, Morphologically Structured Models

Growth of filamentous microorganisms proceeds quite differently from symmetri-

cally dividing bacteria or yeast. The hyphae, long tubular cells, only grow at free tips.
Cell propagation is due to branching, i.e., formation of new tips, and segmentation
of the hyphae, usually without separation of the cells. Therefore, big filamentous
floes or even dense pellets of more or less spherical form may be formed that contain
92 2 Bioprocess Models

a great number of cells with a wide distribution of age and morphological state. Cells
without growing tips within the mycelia can have a metabolic activity that is differ-
ent from those having free tips. Growth conditions, substrate supply and cell age as
well have great influence on morphology of filamentous organisms, and in turn the
morphological form is an important factor for product synthesis. The hyphal form is
preferred at optimal growth while, at impaired conditions or stress, morphological
differentiation to more resistant forms or (arthro)spores can be observed. Therefore,
the growth kinetics of filamentous organisms are rather complex and it is difficult to
identify the key factors to control growth and production. In processes for second-
ary metabolites the optimum conditions for growth differ substantially from those
for biosynthesis of the desired product. This is not only true for the substrate supply
but also temperature and pH. Mathematical models can help to clarify the view of
the process and offer means for its optimization. The methodology of simple segre-
gated models, when describing morphological differentiation, also called morpholo-
gically structured models, provide the appropriate tools for modeling: the clearly
distinguishable forms within the population, e.g., growing and non-growing, hyphae
and spores, young and old, may be lumped into respective cell classes [1, 146].
In the production of Cephalosporin C by Cephalosporium acremonium (Acremo-
nium chrysogenum) three clearly distinguishable forms can be found: slim hyphae,
thick-walled or swollen hyphal fragments, and arthrospores. Under optimal growth
conditions in the early phase of the cultivation, mainly slim hyphae can be found
which represent the growing fraction of the population. With time hyphae differenti-
ate into highly septated thick-walled cells which differentiate further into arthros-
pores under impaired growth conditions or limitation. Maximum production of the
antibiotic is correlated with a high percentage of swollen hyphal fragments.
The model of Chu and Constantinides [144] describes growth of Cephalosporium
acremonium on glucose and sucrose (Table 2.12), as well as its pH and temperature
dependency (not shown). Product synthesis is repressed by glucose, but not during
growth at slowly catabolizible substrates such as sucrose, maltose, or lactose. This
property can be used to optimize productivity in batch culture: the process is started
on a mixture of glucose and sucrose. The growth is then diauxic with a first phase on
glucose where at high growth rates mainly slim hyphae are formed. In a second
phase on sucrose with low growth rates, the morphology changes to the high-produ-
cing thick-walled cells. Two cell types are discriminated in the model: slim hyphae,
Xl> and producing thick-walled hyphal fragments, Xl> as schematically drawn in
Fig. 2.16. Slim hyphae can develop from both forms. The ring expansion enzyme in
the product synthesis chain is synthesized by the slim hyphae during metamorpho-
sis. Product synthesis by thick-walled cells proceeds proportional to the enzyme
concentration that is modeled by a respective balance. The model was successfully
applied for optimization of pH and temperature profiles during batch cultivations,
leading to a doubled product concentration.
Matsumura et al. [145] published an extended model of Cephalosporin C produc-
tion with additional consideration of metabolic inactive arthrospores, X3 , that are
formed at limitation (see Table 2.13). The structure of the model is shown in
Fig.2.17. Furthermore, balances for intracellular methionine are included into the
model. Methionine stimulates product formation and differentiation of slim hyphae
to thick-walled forms. Synthesis of ring-expansion enzyme and product is assumed
to take place only in thick-walled cells. In the model, the accumulation of intracellu-
lar methionine is calculated for all cell types. The sum is compared to experimental
2.4 Segregated Models 93

Table 2.12. Excerpts of the segregated model for growth of Cephalosporium acremonium in-
cluding morphological differentiation of thin hyphae to swollen hyphae [144]

Description Model equations

a Slim hyphae, lysis at glucose limitation dCXl (KDlmaxKn)

at = /Ll(CXl + CX2 ) - /L12 + Kn + C SI CXl

b Thick-walled hyphae, dC
X2
lysis by first -order reaction at = /L12 CXl - KD2 CX2

c Glucose dCSl -/LSI

- d = --(CX!
t YXSI
+ CX2 )

d Succrose, used for maintenance

when glucose is limiting
dCS2
dt
= _ (~+
YXS2 KS2
mS2 CS2 c ) (CXl
+ CS2 + Ki;
+
CX2 )

e Product Cephalosporin C
(CE: intrinsic concentration
of ring-expansion enzyme)

f Specific growth rate /Ll = /LSI + /LS2

/LlmaxCSl /L2 max (t)CS2
/LSI = KSI + C SI /LS2 = KS2 + CS2
g Metamorphosis reaction for slim
to thick-walled cells

Succrose

Cephalosporin C

Fig. 2.16. Structure of the segregated model with slim and swollen hyphae for growth of Ce-
phalosporium acremonium [144]

data. An intrinsic balance for ring-expansion enzyme is established under the as-
sumption that enzyme synthesis is induced after a lag-phase by methionine and im-
mediately repressed by glucose. Product synthesis rate in swollen hyphae is propor-
tional to the intracellular enzyme concentration. The model was used for evaluation
of fed-batch strategies for production of the antibiotic, by which the product concen-
tration could be increased almost twice.
Simulations of the Penicillin production by Penicillium chrysogenum in fed-batch
and repeated fed-batch cultivations were carried out by Zangirolami et al. [147). The
model, summarized in Table 2.14, discriminates three morphological forms, begin-
94 2 Bioprocess Models

Table 2.13. Excerpts of the segregated model for growth of Cephalosporium acremonium in-
cluding arthrospores [145]

Description Model equations

dCxI
a Slim hyphae dt = (Jil - Jil2 - KDl)CXI

dCX2
b Thick-walled cells (swollen hyphea) dt = Jil2 CXI - Ji23 CX2 - KD2 CX2

dCX3
c Inactive arthrospores dt = Ji23 CX2 - KD3 CX3

d Glucose, used for growth of slim dCsI -Jil mSI CSI

- - = --C XI - CX2
hyphae and maintenance metabolism dt Yxs I KSI + CSI
of swollen hyphae
dCs2 QS21max CS2 C + QSZZm.x CS2 C
e Methionine, uptake by the active forms
dt KS2 + CS2 XI KS2 + CS2 X2

f Product formation
(CE: intrinsic enzyme concentration)
Jil m.xCSI
g Specific growth rate on glucose Jil = KSI + CSI
K12CS2) CSI
h Metamorphosis rate to swollen hyphae Jil2 = ( Kll + --=---=--
KS2 + CS2 KSI + CSI
KzzKsI
Metamorphosis rate to arthrospores Ji23 = K21 + KSI + CSl

Table 2.14. Excerpts of the model for Penicillium chrysogenum including three morphological
forms [147]

Description Model equations

dCXl
a Apical cells dt = (Jil - K 12 )CxI + K21 CX2 - DCxI

dCX2
b Subapical cells dt = (Ji2 - K21 - Ji23)CX2 + K12 CxI - DCX2

c Hyphal cells, dCX3

(K3: active fraction of the cell) dt = Ji3 K3CX3 + Ji23 CX2 - DCX3

dCp CSI
d Penicillin formation d = Kp c' (CX2 + K3CX3 ) - DCp
t CSI + KSI + K S1
II

e Specific growth rates

f Metamorphosis rate to hyphal cells

2.4 Segregated Models 95

Cephalosporin C
Fig. 2.17. Structure of the segregated model for growth of Cephalosporium acremonium in-
cluding arthrospores [145]

Substrate

Penicillin
Fig. 2.18. Structure of the segregated model for hyphal growth of Penicillium chrysogenum
[147]

ning from the tip: apical cells of the hyphae, subapical cells, and hyphal cells. The
metamorphosis reactions are shown schematically in Fig. 2.18. During their life cy-
cle, cells pass through the sequence of the three forms, but differentiation to hyphal
cells runs only under limitation. By branching, new tips with apical cells can grow
out of subapical cells. Product synthesis, inhibited by glucose, is assumed to take
place in subapical and hyphal cells. A similar model that was also fitted to experi-
mental data from cultivations of Getrichun candium and Streptomyces hygroscopicus
was published by Nielsen [148]. There, the average properties of the hyphal elements
were derived from a distributed population model. Another example of segregated
modeling of Penicillin production will be given in Chap. 13.

2.4.5
Segregated Models for Recombinant Organisms

In recombinant organisms, the ability to produce foreign proteins is often mediated

by introducing a relatively short circular DNA-sequence, the plasmid, into the host
or wild-strain. The number of plasmids per cell is usually in the range of a few
dozen. The plasmid-bearing, producing cells are denoted in the following as X+.
During cell division, the plasmids from the mother cell are randomly distributed to
96 2 Bioprocess Models

the daughter cells. Therefore, there is a certain probability the case will happen that
one daughter cell receives all the plasmids and the other none (segregational in-
stability). The further fate of the plasmid-free non-producing cells, i.e., the host
strain, here denoted as X_, depends on the applied selection mechanism. In the case
without selection force, the plasmid-free cells have a growth-advantage because they
do not carry the burden of propagating the plasmids and synthesizing the foreign
protein. This results in a higher growth rate compared to the recombinant cells and
the percentage of the X_ -population will increase with time. This effect lowers
greatly the yield and productivity for the foreign protein. In continuous culture, the
X_-population eventually overgrows the X+-population [127].
By applying a selection mechanism that prefers the plasmid-bearing cells, the sta-
bility of the plasmids can be increased to some degree. In the case where the plasmid
is coding for a resistance against some antibiotic, the plasmid-free cells cannot grow
at all when the medium contains the antibiotic. Nevertheless, it is possible that non-
producing cells appear which still have resistance against the antibiotics. The me-
chanism can be an incorporation of the resistance-marker of the plasmid into the
genome, or a mutation in the product-coding region of the plasmid (mutational in-
stability). Again these cells have a growth-advantage compared to the producing
cells. Another selection mechanism is based on host strains that are unable to
synthesize an essential growth factor. The plasmid then contains a complementing
gene that is coding for an enzyme by which the essential component can be synthe-
sized. If the medium lacks the essential component, plasmid-free cells are again un-
able to propagate normally. The causes of instability in this case are identical to the
antibiotics-resistance.
Therefore, in cultivations of genetically modified cells, a genetically inhomoge-
neous population is usually found. Segregated models can describe the stability of
the plasmid-coded foreign DNA and the interference of growth and plasmid synth-
esis by introducing at least two cell classes, plasmid-free and plasmid-bearing
[128]. The kinetics of the entire mixed population are then determined by the
transition or mutation probabilities between the classes, and by the differences in
their growth parameters such as maximum specific growth rate or yield. These can
be estimated from experimental data [129, 130]. The plasmid content together with
the percentage of plasmid-bearing cells determines the foreign protein productivity
of the culture.
A minimal model for growth of recombinant cells on non-selective media consid-
ering only segregational instability of the plasmid includes the two cell classes X+
and X_. With the probability a for loss of plasmid the balances for the cell concen-
trations in a continuous cultivation are

(2.102)

(2.103)

Ollis and Chang [131] studied batch-growth kinetics based on this segregated
model using Monod-expressions for the specific growth rates. The Luedeking-Piret-
equation described product formation by the X+-population. Such a model was used
by Park et al. [132] to investigate optimal control policies for batch, fed-batch, and
continuous cultivation of Escherichia coli with cloned trp operon. Satyagal and Agra-
2.4 Segregated Models 97

wal [136] introduced an inhibition by the plasmid concentration into the kinetics for
{l+ to consider the additional burden for plasmid synthesis. Mosrati et al. [137] used
this model with slight modification for an accurate analysis of experimental data.
They conclude that a increases with growth rate.
Parulekar et al. [134] published the results of theoretical studies for continuous
cultures of recombinant methylotrophs. Methanol is an inhibitory substrate. Plasmid
stability can therefore be improved when it contains a gene that removes this inhibi-
tion. For this case, the following kinetics were used for the specific growth rates:
Cs
{l+ = {l+ max Ks+ + Cs (2.104)

Cs
{l- = {l- max c2 (2.105)
Ks- + Cs +i;
The model simulations showed regions of coexistence of both cell types at various
dilution rates during steady states. Cycling of the feeding could improve the stability.
This was also found by Ryder and DiBiasio [135].
A study on continuous cultivation of recombinant yeast in air-lift reactors was
carried out by Zhang et al. [133]. Based on the model the maximum specific growth
rates and plasmid loss probability were estimated. Coppella and Dhurjati [108] ap-
plied the segregated model, Eqs. (2.102) and (2.103), to recombinant yeast for pro-
duction of human epidermal growth factor in connection with a structured growth
model. Cell growth, plasmid segregation, and gene product synthesis were success-
fully predicted for three reactor configurations: batch, fed-batch, and a hollow-fiber
bioreactor. For production of Hepatitis B virus antigen by recombinant yeast a mod-
el was presented by Shi et al. [138]. Their unstructured growth model also took the
effects of ethanol and leucine into account. The model was used to construct a Kal-
man-filter for state estimation.
Chang and Lim [139] published a model for continuous cultivation of antibiotic-
resistant recombinant strains of Escherichia coli and Proteus mirabilis carrying a
resistance marker. For modeling of antibiotic-induced death, the balance equation
of plasmid-free cells was extended by a death rate {lrn
x
dC -
---;[t = {l+aCx+ +( )
{l- - {lD Cx - - DCx _ (2.106)

with the following kinetic expression:

CB
{lD = {lDmax KB + CB (2.107)

where CB is the antibiotic concentration. Model studies were carried out for optimi-
zation of the antibiotic level in the medium.
Scrienc et al. [140] developed two models for growth of a recombinant yeast mu-
tant with an impaired URA3 gene on selective media, an age distribution model and
a simple segregated model. The employed yeast strain is unable to form an enzyme
that catalyzes the synthesis of the essential growth factor uracil. The complementing
gene for this enzyme, which is called the complementing gene product, is also con-
tained on the plasmid. The authors stress that an important and general aspect for
all selection systems is that it is not the selection gene itself that endows the host cell
98 2 Bioprocess Models

with the selective growth phenotype but the product of that gene. This fact has ma-
jor significance for the growth of unstable recombinant cultures. Since the number
of enzyme-molecules is relatively high, one can assume that the enzymes are always
regularly partitioned between the daughter cells. After loss of the plasmid, the
daughter cells can grow further for a while, because they still contain a certain
amount of the enzyme coding for the essential growth factor. The age distribution
model is compared with the simple segregated model, Eqs. (2.102) and (2.103), as-
suming constant average growth rate of plasmid-free cells, JL-=[L. This is related to
the fraction of the plasmid-bearing cells, F +, obtained by the distribution model
under stationary conditions:

(2.108)

where

(2.109)

In case the X_ -population is not overgrown, the frequency of the segregational

instability can be estimated for stationary exponential growth from the observed
over-all growth rate, JL(oo), as
JL(OO)
n=l--- (2.110)
[L

The model of Sardonini and DiBiasio [141] for growth of recombinant Sacchar-
omyces cerevisiae on selective media was aimed to explain the unexpectedly high
number of plasmid-free cells in the population. The employed selection mechanism
is also based upon the incapability of the host to synthesize an essential metabolite
(M). Ideally, plasmid-free cells should be unable to propagate. Plasmid-bearing cells
have the complementing gene and synthesize the essential metabolite for their own
growth. But since the metabolite is also released into the media, it can support
growth of plasmid-free cells to some extent, which increases their percentage in
the entire population. The explanation of the effect that the X_ -cells not only ori-
ginate from segregational instability but also from independent growth here is
slightly different from the previous model. An extended Monod-kinetics describes
the dependency of the specific growth rate of plasmid-free cells on the essential
metabolite:
Cs CM
(2.111)
JL- = JL-max Ks + CSKM + CM
The simulation showed that 67% of the plasmid-free cells originated in indepen-
dent growth and only 33% form the unequal plasmid partitioning.
An extension of the models for growth on non-selective media by another cell
class "plasmid-carrying with mutation in the product gene" for consideration of
the mutational/structural stability causes formally no problems [79, 142], although
the determination of the model parameters raises some difficulties. The model for
both segregational and mutational instability of recombinant yeast is based on three
cell classes as shown in Fig. 2.19: the non-producing host, X_, the producing plasmid
bearing cells, X+' and plasmid-bearing but non-producing cells, X_ *. When the
2.4 Segregated Models 99

Fig.2.19. Structure of the model for segregational and mutational/structural instability of re-
combinant yeast [79]

probability for segregationalloss of plasmid is denoted as a, and for mutational loss

of only the product-coding gene as (3, the model becomes
dCx+
----;It = fl+(1 - a - (3)Cx+ (2.112)

(2.113)

(2.114)

For the modeling of the interference of plasmid copy number, cloned gene product
synthesis, and growth, they are assumed as being separable into independent factors.
A Monod-kinetics for substrate uptake and inhibition kinetics by the plasmid con-
centration, CG, and product concentration, CII are chosen:

fl+ = flmax (1- CG)'1 ( 1 -

-- --
cp) '2 - -
Cs
- (2.11S)
CGmax Cpmax Ks + Cs
Cs
fl- = flmax Ks + Cs (2.116)

fl*+ = ji'max ( 1 - --
CG )'3 ---
Cs
(2.117)
CGmax Ks + Cs
The formation of the product that is accumulated within the cells is modeled ac-
cording to Eqs. (2.70) and (2.71) with specialized expressions for T} and ~.
Schwartz et al. [143] developed a segregated model for plasmid instability of yeast
grown in selective and non-selective media. The plasmid contains the gene for anti-
biotic resistance. In the model, the effects of plasmid loss and mutation are consid-
ered. By mutation (or crossover) the yeast cells can acquire antibiotic resistance.
Four cell classes are considered: plasmid bearing cells, X+ that produce foreign pro-
tein and do have antibiotics resistance; non-producing plasmid free cells without
100 2 Bioprocess Models

antibiotics resistance, X_; plasmid bearing cells that have acquired antibiotics resis-
tance by mutation, X\; and plasmid-free cells with acquired resistance, X*_. The
model balances are
dCx+
----;Jt = J-l+(1 - a - ;3)Cx+ (2.118)

(2.119)

(2.120)

(2.121 )

The influence of the antibiotic, CB in mg tl, given by the following inhibition ki-
netics,
0.41CB )
J-l+ = J-lmax
( 1 - 159 + CB (2.122)

is analyzed and compared to experimental data for the case of constant maximum
growth rate. The experimental findings could be explained that in non-selective
media after 50 generations, 80% of the cells still exhibited the desired phenotype,
and that only marginal improvement of plasmid stability by addition of antibiotics
could be observed due to mutations.

2.4.6
Population Balance Models

The unstructured or structured models took the cell population in the bioreactor as
a homogenous biomass. Segregated models divided an inhomogeneous population
into a few cell classes that again were each assumed as being homogeneous. The
advantage of this simplified view is that the models are based on ordinary differen-
tial equations which are easy to handle and solve numerically. On the other hand,
these models may be inadequate or inaccurate in several situations as it was already
discussed in the related sections. The simple segregated modeling approach may be
inaccurate when there is a great continuous variation of cell properties that cannot
be matched to few discrete classes. This case can, for instance, be found for plasmid
copy number and product synthesis in cultivations of unstable recombinant organ-
isms [152-154].
The situation can also be met that the unstructured, structured, or segregated
models usually work fine but fail only under some circumstances, for instance, when
the inoculation procedure or the process control scheme was changed. The reason
can be a synchronization of the growth of a majority of cells in the division cycle.
Usually, growth and division of a single cell is independent of the others in the po-
pulation. This, together with a slight variation in the lengths of the cell cycles, results
in a uniform distribution of the division events over time. On average, the popula-
tion seems then to be homogeneous and the cell number increases steadily. But
growth may become synchronized by a sudden change in conditions. A typical case
2.4 Segregated Models 101

is a preculture that was running into substrate limitation. Under starvation most of
the cells rest at a similar point in the cell cycle, e.g., shortly before initiation of divi-
sion. When a bioreactor is inoculated with this preculture, all cells resume growing
at the same time, and subsequently also enter division synchronously. This may last
for several generations until the synchrony is lost by individual variations in the
cycle lengths. Under synchronous growth the modulation of the metabolism over
the cell cycle becomes visible on the measured variables of the bioreactor, e.g., as
oscillations on the concentrations or a step-like increase in cell number. When there
exists some feed-back mechanism from such an oscillating variable in the reactor to
the cell metabolism, e.g., by a regulatory influence, this can even stabilize synchro-
nous growth for a long time.
The above-mentioned cases can be described to some extent by population bal-
ance models. These models, based on partial differential equations, not only look at
the time-dependence of the system but consider its development as governed by
further independent variables z representing the state of the cell. These can be the
age of a cell, its position in the cell cycle, its total mass or volume, the mass of cell
constituents, or other properties. The related balance equation is for symmetric cell
division [149]

af~~z) + a(if~~,Z)) = 2 Jp(Z,Z/)g(Z/)f(t,Z/)dzl - g(z)f(t,z) (2.123)

with initial condition

f(O, z) =fo(z) (2.124)

where z is the vector of cellular state variables, i is the mean growth rate vector of z,
f is the density function of the distribution in the population state, g is the division
rate or probability of division of a cell, and p is the partitioning function, i.e., the
conditional probability that a mother cell in state z' will divide and form two new
cells in state z.
The partitioning function must have the following properties:

p(z, Z/) ° for z > Zl

J
=

p(z, z/)dz = 1 (2.125)

p(z, Z/) = p(ZI - Z, Z/) for z < Zl

The solution of the balance equation is, in general, rather difficult, but it can be
significantly simplified by an iterative procedure based on successive generations of
cells [150]. In age distribution models, the state of the population is characterized by
the time since birth of the cell, i.e., the cell age T. The balance equation then becomes

(2.126)

with initial condition

f(O,T) = fo( T) (2.127)

102 2 Bioprocess Models

and boundary condition

J
00

f(t,O) = 2 g(r)f(t, r)dr' (2.128)

o

Equation (2.126) is the so-called M'Kendrick-von Foerster equation for the devel-
opment of cell number density [151]. An application of age distribution models for
simulation of Baker's yeast production is given in Chap. 9.

References
1. Bailey JE, Ollis DF (1986) Biochemical engineering fundamentals, 2nd edn. McGraw-
Hill, New York
2. Roels JA (1983) Energetics and kinetics in biotechnology. Elsevier Biomedical Press,
Amsterdam
3. Gaden EL (1959) J Biochem Microbiol Technol Eng 1:413
4. Aharonowitz Y, Demain AL (1980) Biotechnol Bioeng 22, Suppl 1:5
5. Bailey JE (1998) Biotechnol Prog 14:8
6. Heijnen JJ (1994) Tibtech, 12:483
7. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
8. Toda K (1981) J Chern Tech Biotechnol 31:775
9. Calam CT (1979) Folia Microbiol 24:276
10. Esener AA, Roels JA, Kossen NWF (1983) Biotechnol Bioeng 25:2803
11. Noorman HJ, Heijnen JJ, Luyben KCAM (1991) Biotechnol Bioeng 38:603
12. Nielsen J, Villadsen J (1992) Chern Eng Sci 47:4225
13. Wood BD, Whitaker S (1998) Chern Eng Sci 53:397
14. van Suijdam JC, Hols H, Kossen NWF (1982) Biotechnol Bioeng 24:177
15. de Gooijer CD, RH Wijffels, Tramper J (1991) Biotechnol Bioeng 38:1991
16. Buschulte TK, Gilles ED (1990) Modelling and simulation of hyphal growth, metabolism
and mass transfer in pellets of streptomyces. In: Christiansen C, Munck L, Villadsen J
(eds) Proceedings of the 5th European Congress on Biotechnology, p 279
17. Buschulte TK (1992) PhD thesis, Universitat Stuttgart
18. Bellgardt KH (1991) Cell models. In: Rehm HI, Reed G (eds) Biotechnology, vol 4. VCH,
Weinheim
19. Cui QH, Lawson L, Gao D (1984) Biotechnol Bioeng 26:682
20. Frame KK, Hu WS (1988) Biotechnol Bioeng 32:1061
21. Monbouquette HG (1992) Biotechnol Bioeng 39:498
22. Weiss RM, Ollis RF (1980) Biotechnol Bioeng 22:859
23. Sinclair CG, Ryder DH (1975) Biotechnol Bioeng 17:375
24. Tsao GT, Yang CM (1976) Biotechnol Bioeng 18:827
25. Herbert D (1959) In: Tenvall D (ed) Recent progress in microbiology. Alquist and Wik-
sell, Stockholm
26. Pirt SJ (1965) Proc R Soc Ser B 163:224
27. Eroshin VK, Utkin IS, Ladynichev SV, Samoylov Vv, Kuvshinnikov VD, Skryabin GK
(1976) Biotechnol Bioeng 18:289
28. Tayeb YJ, Bouillanne C, Desmazeaud MJ (1984) J Ferment Techno162:461
29. Chu WB, Constantinides A (1988) Biotechnol Bioeng 32:277
30. Reuss M, Frohlich S, Kramer B, Messerschmidt K, Pommerening G (1986) Bioproc Eng
1:79
31. Brown DE, Halsted DJ (1975) Biotechnol Bioeng 17:1199
32. Topiwala H, Sinclair CG (1971) Biotechnol Bioeng 13:785
33. Bellgardt KH (1998) Adv Biochem Eng Biotechnol 60:153
34. Starzak M, Krzystek L, Nowicki L, Michalski H (1994) Chern Eng J 54:221
References 103

35. Zeng AP, Ross A, Biebl H, Tag C, Giinzel B, Deckwer WD (1994) Biotechnol Bioeng
44:902
36. Dourado A, Goma G, Albuquerque U, Sevely Y (1987) Biotechnol Bioeng 29:187
37. Kowda M, Wasungu M, Simard RE (1982) Biotechnol Bioeng 24:1125
38. Luong JHT (1885) Biotechnol Bioeng 27:280
39. Luong JHT (1987) Biotechnol Bioeng 29:242
40. Han K, Levenspiel 0 (1988) Biotechnol Bioeng 32:430
41. Kosaric N, Ong SL, Duvnjak Z, Moser A (1984) Acta Biotechnol4:153
42. Tan Y, Wang ZX, Marshal KC (1996) Biotechnol Bioeng 52:602
43. Liden G, Franzen CJ, Niklasson C (1994) Biotechnol Bioeng 44:419
44. Barford JP, Phillips PJ, Orlowski JH (1992) Bioproc Eng 7:297
45. Benthin S, Nielsen J, Villadsen J (1992) Biotechnol Bioeng 40:137
46. Pavlou S, Fredrickson AG (1989) Biotechnol Bioeng 34:97l
47. Baltzis BC, Fredrickson AG (1988) Biotechnol Bioeng 31:75
48. Sinclair CG (1989) Microbial process kinetics. In: Bu'Lock J, Kristiansen B (eds) Basic
biotechnology. Academic Press, London, p 102
49. Akerberg C, Hofvendahl K, Zacchi G, Hahn-Hagerdal (1998) Appl Microbiol Biotechnol
49:682
50. Lendenmann U, Egli T (1998) Biotechnol Bioeng 59:99
51. Lam JC, Ollis DF (1981) Biotechnol Bioeng 23:1517
52. Venkatesh KV, Okos MR, Wankat PC (1993) Process Biochemistry 28:231
53. Ohara H, Hiyama K, Yoshida T (1992) Appl Microbiol Biotechnol 38:403
54. Jackson JV, Edwards VH (1975) Biotechnol Bioeng 17:943
55. Tan Y, Wang ZX, Marshal KC (1998) Biotechnol Bioeng 59:724
56. Kluge M, Siegmund D, Diekmann H, Thoma M (1992) Appl Microbiol Biotechnol36:446
57. Halme A (1989) Expert system approach to recognize the state of fermentation and to
diagnose faults in bioreactors. In: Fish NM, Fox RI, Thornhill NF (eds) Computer appli-
cation in fermentation technology. Elsevier Applied Science Publishers, London, p 159
58. Halme A, Visala A, Kankare J (1992) Functional state concept in modelling of biotech-
nical processes. In: Stephanopoulos G, Karim MN (eds) IFAC Symposium on Modelling
and Control of Biotechnical Processes, Pergamon Press, New York, p 153
59. Konstantinov K, Yoshida T (1989) Biotechnol Bioeng 33:1145
60. Zhang XC, Visala A, Halme A, Linko P (1994) Process Control 4:1
61. Zhang XC, Visala A, Halme A, Linko P (1994) J Biotechnol 27:1
62. Linko P, Siimes T, Kosola A, Zhu YH, Eerikainen T (1994) Hybrid fuzzy knowledge-
based and neural systems for fed-batch baker's yeast bioprocess control. In: Proceedings
of the 1st Asian Control Conference, 27-30 July 1994, Tokyo:491
63. Ruenglertpanyakul W (1996) PhD Thesis, Universitat Hannover
64. Franks PA, Hall RJ, Linklater PM (1980) Biotechnol Bioeng 22:1465
65. Williams FM (1967) J Theor BioI 15:190
66. Korte G, Rinas U, Kracke-Helm HA, Schiigerl K (1991) Appl Microbiol Biotechnol35:185
67. Korte G, Rinas U, Kracke-Helm HA, Schiigerl K (1991) Appl Microbiol Biotechnol35:189
68. Nielsen J, Emborg C, Halberg K, Villadsen J (1989) Biotechnol Bioeng 34:478
69. Nielsen J, Pedersen AG, Strudsholm K, Villadsen J (1991) Biotechnol Bioeng 37:802
70. Bentley WE, Kompala DS (1989) Biotechnol Bioeng 33:49
71. Lee J, Ramirez WF (1992) Biotechnol Bioeng 39:635
72. Domach MM, Leung SK, Cahn RE, Cocks GG, Shuler ML (1984) Biotechnol Bioeng
26:203
73. Ataai MM, Shuler ML (1985) Biotechnol Bioeng 27:1027
74. Shu J, Shuler ML (1989) Biotechnol Bioeng 33:1117
75. Peretti SW, Bailey JE (1986) Biotechnol Bioeng 28:1672
76. Lee SB, Bailey JE (1984) Biotechnol Bioeng 26:66
77. Peretti, SW, Bailey JE (1987) Biotechnol Bioeng 29:316
78. Ataai MM, Shuler ML (1987) Biotechnol Bioeng 30:389
79. Lee SB, Seressiotis A, Bailey JE (1985) Biotechnol Bioeng 27:1699
80. Lee SB, Bailey JE (1984) Biotechnol Bioeng 26:1372
104 2 Bioprocess Models

81. Lee SB, Bailey JE (1984) Biotechnol Bioeng 26:1383

82. Betenbaugh MJ, Dhurjati PA (1990) Biotechnol Bioeng 36:124
83. Ramkrishna D (1982) A cybernetic perspective of microbial growth. In: Papoutsakis E,
Stephanopoulos GN, Blanch HW (eds) Foundations of biochemical engineering kinetics
and thermodynamics in biological systems. American Chemical Society, Washington
DC
84. Kompala DS, Ramkrishna D, Tsao GT (1984) Biotechnol Bioeng 26:1272
85. Dhurjati P, Ramkrishna D, Flickinger MC, Tsao GT (1985) Biotechnol Bioeng 27:1
86. Kompala DS, Ramkrishna D, Jansen NB, Tsao GT (1986) Biotechnol Bioeng 28:1044
87. Turner BG, Ramkrishna D, Jansen NB (1988) Biotechnol Bioeng 32:46
88. Turner BG, Ramkrishna D (1988) Biotechnol Bioeng 31:41
89. Straight JV, Ramkrishna D (1990) Biotechnol Bioeng 37:895
90. Baloo S, Ramkrishna D (1990) Biotechnol Bioeng 38:1337
91. Straight JV, Ramkrishna D (1994) Biotechnol Prog 10:574
92. Straight JV, Ramkrishna D (1994) Biotechnol Prog 10:588
93. Ramakrishna R, Ramkrishna D, Konopka AE (1997) Biotechnol Bioeng 54:77
94. Narang A, Konopka A, Ramkrishna D (1997) Chern Eng Sci 52:2567
95. Fredrickson AG (1976) Biotechnol Bioeng 18:1481
96. Esener AA, Veerman T, Roels JA, Kossen NWF (1982) Biotechnol Bioeng 24:1749
97. Nielsen J, Nikolajsen K, Villadsen J (1991) Biotechnol Bioeng 38:1
98. Nielsen J, Nikolajsen K, Villadsen J (1991) Biotechnol Bioeng 38:11
99. Nikolajsen K, Nielsen J, Villadsen J (1991) Biotechnol Bioeng 38:24
100. King R (1997) J Biotechnol 52:219
101. King R, Biidenbender C (1997) J Biotechnol 52:235
102. Mundry C, Kuhn KP (1991) Appl Microbiol Biotechnol 35:306
103. Yoo S, Kim WS (1994) Biotechnol Bioeng 43:1043
104. Bonarius HPJ, Schmid G, Tramper J (1997) Tibtech 15:308
105. Doshi P, Rengaswamy R, Venkatesch KV (1997) Process Biochem 32:643
106. Venkatesch KV, Doshi P, Rengaswamy R (1997) Biotechnol Bioeng 56:635
107. Narang A, Konopka A, Ramkrishna D (1997) J Theor Bioi 184:301
108. Coppella SJ, Dhurjati P (1990) Biotechnol Bioeng 36:356
109. Ray NG, Vieth WR, Venkatasubramanian K (1987) Biotechnol Bioeng 29:1003
110. Lee SB, Bailey JE (1984) Plasmid 11:151
111. Lee SB, Bailey JE (1984) Plasmid 11:166
112. Ryu DY, Park SH (1987) Ann NY Acad Sci 506:396
113. Chen W, Bailey JE, Lee SB (1991) Biotechnol Bioeng 38:679
114. Koh BT, Yap MGS (1993) Biotechnol Bioeng 41:707
115. Axe DD, Bailey JE (1994) Biotechnol Bioeng 43:242
116. Kim BG, Shuler ML (1991) Biotechnol Bioeng 1076
117. Laffend L, Shuler ML (1994) Biotechnol Bioeng 43:388
118. Laffend L, Shuler ML (1994) Biotechnol Bioeng 43:399
119. Chetan J, Gadgil P, Bhat PJ, Venkatesh RV (1996) Biotechnol Prog 12:744
120. Fredrickson AG (1992) AIChE Journal 38:835
121. Tempest DW, Herbert D, Phipps PJ (1967) Microbial physiology and continuous culture.
HMSO, London, p 240
122. Pirt SJ (1987) J Ferment Technol 65:173
123. Sinclair CG, Topiwala HH (1970) Biotechnol Bioeng 12:1069
124. Veilleux BG (1980) European J Appl Microbiol Biotechnol 9:165
125. Kreikenbohm R, Stephan W (1985) Biotechnol Bioeng 27:296
126. Topiwala HH, Hamer G (1971) Biotechnol Bioeng 13:919
127. Summers DK (1991) Tibtech 9:273
128. Imanaka T, Aiba S (1981) Ann NY Acad Sci 396:1
129. Park SH, Ryu DDY, Lee SB (1991) Biotechnol Bioeng 37:404
130. Shoham Y, Demain AL (1991) Biotechnol Bioeng 37:927
131. Ollis DF, Chang HT (1982) Biotechnol Bioeng 24:2581
132. Park TH, Seo JH, Lim HC (1989) Biotechnol Bioeng 34:1167
References 105

133. Zhang Z, Scharer JM, Moo-Young M (1997) J Biotechnol 55:31

134. Parulekar SJ, Chang YK, Lim HC (1987) Biotechnol Bioeng 29:911
135. Ryder DF, DiBiasio D (1984) Biotechnol Bioeng 26:942
136. Satyagal VN, Agrawal P (1989) Biotechnol Bioeng 34:265
137. Mosrati R, Nancib N, Boudrant J (1993) Biotechnol Bioeng 41:395
138. Shi Y, Ryu DY, Yuan WK (1993) Biotechnol Bioeng 41:55
139. Chang YK, Lim HC (1987) Biotechnol Bioeng 29:950
140. Scrienc F, Campbell JL, Bailey JE (1986) Biotechnol Bioeng 18:996
141. Sardonini CA, DiBiasio D (1987) Biotechnol Bioeng 29:469
142. Hjortso MA, Bailey JE (1984) Biotechnol Bioeng 26:528
143. Schwartz LS, Jansen NB, Ho NWY, Tsao GT (1988) Biotechnol Bioeng 32:733
144. Chu WB, Constantinides A (1988) Biotechnol Bioeng 32:277
145. Matsumura M, Imanaka T, Yoshida T, Taguchi H (1981) J Ferment Technol 59:115
146. Nielsen J (1992) Adv Biochem Eng Biotechnol 46:187
147. Zangirolami TC, Johansen CL, Nielsen J, J0rgensen SB (1997) Biotechnol Bioeng 56:595
148. Nielsen J (1993) Biotechnol Bioeng 41:715
149. Fredricksen AG, Ramkrishna D, Tsuchiya HM (1967) Math Biosci 1:327
150. Liou J], Scrienc F, Fredrickson AG (1997) Chern Eng Sci 52:1529
151. Webb GF (1986) J Math Biology 23:269
152. Seo JH, Bailey JE (1985) Biotechnol Bioeng 27:156
153. Wittrup KD, Bailey JE (1988) Biotechnol Bioeng 31:304
154. Srienc F, Campbell JL, Bailey JE (1986) Biotechnol Bioeng 18:996
3 Metabolic Flux Analysis
Maria I. Klapa, Gregory Stephanopoulos

3.1
Introduction

Metabolic pathways are sequences of enzyme-catalyzed reaction steps, converting

substrates to a variety of products to meet the needs of the cell. Metabolic pathways
can interact to create complex networks. Manipulation of metabolic pathways to im-
prove the cellular properties is an old concept in biological sciences. This approach
initially relied on random mutagenesis and creative selection techniques to identify
superior strains with respect to a certain objective. Despite impressive successes in
many biological areas, the mutations that played important role in achieving the
desired cellular properties were poorly characterized and the strain development
process remained random, combining science with elements of art [1]. Recombinant
DNA technology introduced a new dimension to pathway manipulation, because it
allowed precise modifications of enzymatic reactions in metabolic pathways. Meta-
bolic engineering emerged then as the scientific field aiming at the directed modifi-
cation of the enzymatic, regulatory, or transport activities of the cell to improve
cellular properties, with the use of recombinant DNA technology [2-4].
One could argue that metabolic engineering is just the technological manifestation
of applied molecular biology and there is no justification for the introduction of a
new scientific field. However, metabolic engineering is indeed a new scientific field
[2] based on the rational, as opposed to the ad hoc, process to identify the targets for
pathway manipulation [1]. To determine the required biological changes, an en-
hanced understanding of metabolism and cellular functions must be obtained. Thus,
the analysis and correct interpretation of the structure and control of metabolic net-
works constitute the first critical task of metabolic engineering towards the goal of
rational pathway manipulation.
One important novel aspect of metabolic engineering is the emphasis it places on
the concept of metabolic networks, as opposed to individual reactions. By consider-
ing integrated networks rather than reactions or parts of networks in isolation from
one another, an enhanced understanding of metabolism can be obtained [4]. The
complexity though of cellular metabolic networks, involving large numbers of meta-
bolites and enzymes under the control of intricate regulatory mechanisms, and the
difficulty of assessing enzyme kinetic properties in vivo pose two key obstacles in the
analysis of metabolic networks. Metabolic fluxes emerge then as the preferred para-
meter to determine the physiological state of the cell. Perhaps the most significant
contribution of metabolic engineering is the emphasis it places on in vivo metabolic
fluxes and their control [5] for the investigation of metabolic networks.
3.1 Introduction 107

Flux is defined as the rate at which material is processed through a metabolic

pathway. Consequently, the flux of a linear pathway at steady state is equal to the
steady-state rates of all the individual reactions. The value of the flux does not carry
information about the activity of the corresponding pathway enzymes, but it defines
the extent to which these enzymes participate in a conversion process. Thus, meta-
bolic fluxes provide a measure of the degree of engagement of the various pathways
in the overall cellular functions and metabolic processes at specific genetic and en-
vironmental conditions, constituting a fundamental determinant of cell physiology
[1]. Fluxes are also useful for the determination of maximum theoretical yields [6-
10] and the observation of the function of metabolic pathways in vivo [11]. Flux
determination can play an important role in elucidating metabolic pathways to a
finer detail [1, 12]. By comparing and analyzing flux distributions throughout a me-
tabolic network at various biological conditions, information concerning the control
of flux and the regulatory structure of metabolic networks can be extracted [3, 12].
Therefore, the accurate quantification and analysis of metabolic fluxes is an essential
step in the study of the structure and control of metabolic networks. The quantifica-
tion of metabolic fluxes is termed Metabolic Flux Analysis (MFA) and is a major
objective of metabolic engineering.
The accurate quantification of flux distribution throughout a metabolic network is
not a trivial task, considering the complexity of metabolic networks and the scarcity
of useful experimental data to describe the intracellular reactions. As flux determi-
nation is coupled with reaction stoichiometries, as shown in Sect. 3.2, the accuracy
of flux quantification depends on the degree to which the assumed intracellular bio-
chemistry describes the actual in vivo structure of metabolic networks. Unfortu-
nately, many reaction steps are unknown, even qualitatively. It is critical then that
the available experimental measurements and other biochemical constraints provide
sufficient redundancy to validate the accuracy of measurements used [13-15], as well
as to test the consistency of the assumed intracellular biochemistry with the overall
data [6, 16]. Barring experimental error, any identified inconsistency of the bio-
chemistry with the existing data indicates the presence of errors in the assumed
structure of the metabolic network. In this regard, issues such as pathway modifica-
tion (e.g., considering reaction reversibility [17, 18], or alternate conversion routes
[8, 19]) and discovery of new pathways [19-21] are an important part of the meta-
bolic flux determination process. The complexity of the metabolic flux quantification
problem should not discourage researchers from flux analysis, as it is an essential
part of the analysis of intracellular metabolism, which can greatly improve the effi-
cacy of future genetic manipulations.
In the next section, we present methods that have been employed to date for flux
quantification. Their advantages and weaknesses, the latter leading to the introduc-
tion of new methods or to combinations of several existing techniques, will be dis-
cussed in an effort to provide a systematic way to flux quantification. The last sec-
tion of this chapter will include examples of the use of fluxes for the elucidation of
metabolic pathways, emphasizing the importance of Metabolic Flux Analysis as a
major task of metabolic engineering.
108 3 Metabolic Flux Analysis

3.2
Flux Quantification Methods

The starting point of MFA is the reaction network stoichiometry for the conversion
of substrates to metabolic products and biomass constituents. The extent to which
intracellular flux estimates can be accepted as reliable measures of the actual in vivo
fluxes depends on the validity of the assumed intracellular biochemistry. Therefore,
flux determination methods require a high degree of redundancy to check the accu-
racy of experimental measurements and the consistency of the assumed biochemis-
try. This is a critical point that must be considered before the selection of a particu-
lar flux determination method to be applied on a metabolic network.

3.2.1
Metabolite Balancing

Metabolite balancing is a very powerful methodology and usually is the first step in
the determination of fluxes. Intracellular fluxes are determined as functions of the
measurable extracellular fluxes, using a stoichiometric model for the major intracel-
lular reactions and applying a mass balance around each intracellular metabolite.
The method is readily applied and does not require enzyme kinetics information.
The only assumption is the pseudo-steady state approximation for the intracellular
metabolites, which is justified in most cases, as there is a very high turnover rate of
the pools of most metabolites [1]. Experimentally a metabolic steady-state can be
achieved in a continuous flow reactor. The resulting set of equations can be ex-
pressed in a matrix form as:

S ·Y=r (3.1)

where, S is the stoichiometric matrix of the metabolic network. The number of rows
is equal to the number of metabolites in the network and the number of columns is
equal to the number of unknown fluxes at steady state. y is the vector of unknown
fluxes and r is the vector of metabolite extracelluar accumulation rates. The elements
of r corresponding to intracellular metabolites are zero, whereas those correspond-
ing to extracellular metabolites are equal to the measured accumulation rates.
Since the number of reactions is, in general, greater than the number of metabo-
lites, the system of Eq. (3.1) is in general underdetermined. Additionally, there are
some structural singularities that increase the degrees of freedom of the system, like
pathways that branch apart and rejoin later in the network, reversible reactions, or
metabolic cycles, that cannot be resolved by metabolite balancing alone Pl. In math-
ematical terms, these structures introduce linearly dependent equations in the stoi-
chiometric matrix, decreasing its rank (and consequently increasing the degrees of
freedom). Figure 3.1 illustrates this point for the case of an artificial network con-
taining two competing pathways and one reversible reaction; the forward and the
backward directions correspond to linearly dependent columns decreasing the rank
of the matrix by 1. Using metabolite balancing, one may be able to determine the net
flux of the reversible reaction, but the forward and the backward fluxes cannot be
differentiated.
In these cases, where metabolic networks cannot be resolved by metabolite balan-
cing constraints alone, additional constraints - at least as many as the degrees of
3.2 Flux Quantification Methods 109

~D
rA A~ B
~v3+
VS!
V3-~C ~E

System of Mass Balances:

v3net

vI v2 v3 + v3 - v4 v5
vI = -rA
v2 = vl-v3net = vl-a
v4 = rC -v3net = rC-a
v5 = vl- rD- a
v5 =rE + rC- a
a = [(v3+)-(v3-)]

Fig. 3.1. Using only extracellular accumulation rate measurements, structural singularities like
reversible reactions cannot be resolved. The forward and backward directions of a reversible
reaction correspond to linearly dependent columns of the stoichiometric matrix, as it is shown
for reversible reaction 3 in the network above. Note that, in the depicted network, another
source of singularity are the two competing pathways B->C and B->D->E->C. The flux split
ratio at B cannot be determined from metabolite balancing (rank(S) = 4, 2 degrees of freedom
and one redundant equation).

freedom of the system of Eq. (3.1) - are needed before the system can be solved for
the unknown fluxes. If more constraints than the degrees of freedom are introduced,
then the system becomes overdetermined and the redundant equations can be used
to test the consistency of the overall balances and the validity of the pseudo-steady
state assumption [1]. These additional constraints can be obtained from the follow-
ing sources:
- Theoretical assumptions concerning the intracellular biochemistry [7, 8, 16, 21,
22]. Some reactions introducing singularities could be lumped together, consid-
ered irreversible, or even neglected, if there is no experimental evidence of the
presence or high activity of the corresponding enzyme [6, 7, 23, 24]. Information
about the relative flux ratio of two pathways could be obtained by in vitro enzyme
activity measurements [21, 24, 25]. This information should be considered with
caution, because in vitro enzyme activity measurements are found to bear little
relationship to actual in vivo flux distributions. It should also be taken into con-
sideration that most of the theoretical assumptions are usually based on observa-
tions of the wild-type microorganisms in their native environment. Thus, they
might not hold true for highly engineered organisms and/or under extreme envir-
onmental conditions [26].
As the reliability of flux estimates is based strongly on the assumed biochemis-
try, the validity of the theoretical assumptions must be tested through satisfaction
of redundancies. Any conclusions about the physiological state of the cell based on
the determined flux distribution must be evaluated in the context of the assump-
tions that have been made as part of the analysis. When comparing results from
various analyses, they should be based on similar set of assumptions.
110 3 Metabolic Flux Analysis

- Linear programming (LP) [27-29]. The use oflinear programming for the analysis
of metabolic networks dates back to the mid-1980s. It has been applied to deter-
mine the maximum yields of fermentation products from glucose in butyric and
propionic acid bacteria [30,31], to adipocyte metabolism [32], and to analysis of
tricarboxylic acid (TeA) cycle and acetate overflow in Escherichia coli [33]. With
this approach, metabolic fluxes are determined such as to be stoichiometrically
feasible (subject to the metabolite balance constraints) and optimize a certain ob-
jective function. In mathematical terms this is equivalent to the solution of the
following LP problem [27]:

minimize <::Y = L CiVi

(3.2)
subject to S . Y = r
where <:: is the vector of the weight factors (or cost vector [34]) of fluxes in the objec-
tive function. The weights Ci are selected in such a way as to allow the formulation of
physiologically meaningful objective functions [27]. For example, if the objective
function is to maximize the production rate of a metabolite, then the weight factor
corresponding to the flux of the reaction(s) producing this metabolite should be -1
and the weight factors of the other fluxes should be o. Objective functions that have
been used for flux estimation, as well as their interpretation, are listed in Table 3.1, as
it has been adopted from [35]. By imposing an objective function, one tests the lim-
its of a cell in carrying out a certain task (such as growing, utilizing ATP, etc.) sub-
ject only to the constraints imposed by the assumed biochemistry under specific
genetic and environmental conditions. The LP problem of Eq. (3.2) is usually solved
using the simplex method [34, 36].
Linear optimization solutions tend, by definition, to occur at extreme points ('cor-
ners') of the feasible domain [34]. When LP is applied to a metabolic network, the
obtained flux distribution is an end point of the stoichiometrically feasible domain
[24,28,35], termed metabolic genotype [29,35]. In general, the feasible domain is a

Table 3.1. Objective functions that can be imposed on the flux distribution under specific ge-
netic and environmental conditions to study the limits of cellular function in carrying out a
certain task (adopted from [35])

Question Objective Reference

What are the biochemical Maximize metabolite production Varma et al. [41]
production capabilities? rate
What is the maximal growth Maximize growth rate Varma and Palsson [39];
rate and biomass yield? Varma and Palsson [94]
How efficiently can Minimize the Euclidean norm Bonarius et al. [42]
metabolism channel meta-
bolites through the network?
How energetically efficient Minimize ATP production rate Majewski and Domach [33];
can metabolism operate? or minimize nutrient uptake rate Savinell and Palsson [27};
Fell and Small [32]
What is the tradeoff between Maximize biomass production Varma et al. [41]
biomass production and rate for a given metabolite
metabolite overproduction? production rate
3.2 Flux Quantification Methods 111

Stoichiometric constraint

Flux V2

Optimal solution of a particular LP problem

Flux VI

Fig. 3.2. In two-dimensional space, the stoichiometric constraints correspond to lines and the
feasible domain is a plane region bounded by them. The linear optimization solutions tend to
occur at extreme points (,corners') of the feasible region

polyhedron [34] formed by the intersection of hyperplanes [34] corresponding to

the stoichiometric constraints imposed on the metabolic fluxes as these are ex-
pressed by the rows of the system of Eq. (3.1). Figure 3.2 illustrates the notion of
the stoichiometrically feasible domain in the case of a network with two unknown
fluxes, VI and Vz. In two-dimensional space, the stoichiometric constraints corre-
spond to lines and the feasible domain is a plane region bounded by them.
Application of LP to flux determination is a very useful technique, since it pro-
vides boundaries for the in vivo achievable flux distributions. The technique shares
all the advantages of metabolite balancing: it is simple and does not require any
kinetic information. It can be used to test the metabolic flexibility of a microorgan-
ism [35, 37], i.e., how the cell readjusts its metabolism to meet a specific objective
after mutations of its genotype. In addition, through duality theory [34], it is possi-
ble to determine the effect of the presence, absence, or modification of a stoichio-
metric constraint on the objective function [27, 28, 38-41]. For example, if the ob-
jective function is to maximize the production rate of metabolite j, then the dual
variables (,shadow prices')
oz*
Pi=- (3.3)
ori

reflect the effects that changes in the extracelluar accumulation rate of metabolite i
have on the theoretical yield of metabolite j. z* is the optimal value of the objective
function, which, in this example, is equal to the theoretical yield of metabolite j.
Savinell and Palsson [27,38] used LP to examine mammalian cell growth in culture.
They were able to compare the value of different nutrients with respect to their over-
all growth supporting ability and also their ability to satisfy energy or other require-
ments of the cell.
The drawbacks of the application of LP to flux quantification derive from the fact
that (1) it is not certain that the intracellular biochemistry remains unchanged under
various cellular "objectives" or that the rest of the metabolism, besides the mutation
site, is not affected after a mutagenesis, and (2) that the obtained flux distribution
may not be the one achievable in vivo, because of "optimization" criteria used by the
112 3 Metabolic Flux Analysis

cells in vivo are not the same as the considered ones. Bonarius et al. [42] calculated
the flux distribution in hybridoma cells imposing the minimum-norm constraint,
i.e., efficient channeling of metabolites through the metabolic pathways, considering
it as the situation most likely to exist in vivo. However, using data from !3C-nuclear
magnetic resonance (NMR) spectroscopy experiments [43], the in vivo flux distribu-
tion was actually found to be relatively similar to that obtained if one of the objective
functions 'maximize ATP' or 'maximize NADH' was imposed to the system instead of
the 'minimum Euclidean norm' constraint, that they had initially considered.
- Inclusion of additional measurements, such as those obtained from isotopic-tracer
techniques. Such methods provide additional information about the in vivo meta-
bolic fluxes, decreasing the number of required assumptions and increasing the
observability of the system. Combination of these techniques with metabolite bal-
ancing can elucidate the in vivo flux distribution in finer detail and provide en-
ough redundancy to check the validity of estimated fluxes and the consistency of
the assumed intracellular biochemistry. They will be discussed in more detail in
the next session.

Aiba and Matsuoka [19] could be considered the first to present the use of extracel-
lular metabolite measurements in combination with intracellular reaction stoichio-
metry for the determination of metabolic fluxes. Their work, as well as the work of
Papoutsakis [44] on solvent production, was focused on the validation of the bior-
eaction network responsible for product formation through the estimation of intra-
cellular fluxes. Vallino and Stephanopoulos [3, 6, 16, 45, 46] were the first to use the
extracellular fluxes and metabolite balancing technique for the elucidation of intra-
cellular regulatory mechanisms. Considering changes in the intracellular fluxes and
flux split ratios at key branch points of Corynebacterium glutamicum under different
genetic and environmental conditions, they were able to derive useful conclusions
about the control of flux. Computer programs have been written to facilitate flux
determination using metabolite balancing in a user friendly environment [6, 47,
48]. We discuss other applications of metabolite balancing in elucidating metabolic
networks in Sect. 3.3.

3.2.2
Isotopic-Tracer Techniques

Isotopic tracer techniques have been used extensively in biochemistry for the eluci-
dation of metabolic pathways [49-57]. In isotopic-tracer techniques, cells are pro-
vided with a substrate labeled with a detectable isotope at (a) specific atom(s). It is
assumed that the metabolic system is not disturbed by the introduction of the iso-
tope. One class of isotopic-tracer methods is the use of substrates labeled with J3 C or
14C at specific carbon locations. !3C isotopes have the advantage of being stable, non-
radioactive, and detectable by Nuclear Magnetic Resonance (NMR).
The isotope distribution among the metabolites of the network for specific labeled
substrate(s) and known biochemistry is a function of the in vivo metabolic fluxes.
Information on the isotope distribution of various extracellular and/or intracellular
metabolites can be obtained by measuring the label enrichment distribution by
NMR, studying the fine structure of NMR spectra or measuring the mass isotopo-
mer distribution by Gas (or Liquid) Chromatography-Mass Spectrometry (GC [LC]-
MS). Analysis of these experimental measurements can be used to obtain additional
3.2 Flux Quantification Methods 113

constraints on in vivo metabolic fluxes beyond those of metabolite balancing. We

show how data from isotopic-tracer experiments can be used systematically to re-
solve parts of the network that cannot be elucidated by metabolite balancing alone.
Flux quantification through isotopic tracer techniques requires either transient or
steady-state isotope intensity measurements. Information about the transient ap-
proach, which uses radioactive isotopes and can be used to small networks, can be
found in [1]. It should be noted that the flux determination analysis presented here is
based on the second approach and it is valid only if the experimental data are ob-
tained at metabolitic and isotopic steady-state. The chemostat is the preferred sys-
tem to conduct metabolitic steady-state experiments. In addition the labeled sub-
strate has to be introduced to the system in such a way (I) that the metabolitic
steady-state is not disturbed and (2) to allow isotopic steady state to be reached. It
is possible to assume metabolitic and isotopic steady-state for a limited duration
during a transient experiment in a batch reactor. It is important though, to seek for
internal checks providing evidence for the validity of the metabolitic and isotopic
steady-state assumption, as inappropriate application of metabolite balancing- and
isotopic-tracer analysis will produce incorrect and misleading results.
Whether the measured isotopic distribution of metabolites in the network contains
information to elucidate parts of networks unresolved by metabolite balancing only
depends on (1) the assumed biochemistry, (2) the type and labeling of substrate used,
and (3) the extracellular metabolite(s), whose degree of label enrichment and isoto-
pomer distribution is measured [1]. The reason for this is that, based on the assumed
structure of the network, some metabolites are affected differently to others from the
introduction of a particular labeled substrate. Figure 3.3 shows the splitting pathway
at the tetrahydrodipicolinate (H4D) branch point, which is part oflysine biosynthesis,
a metabolic pathway that cannot be elucidated using material balances as the only
constraints. These pathways, that branch apart and rejoin later, can be differentiated
through the use of isotopic labels, provided that they introduce asymmetries in the
distribution of carbon atoms of intermediate metabolites, generating different label-
ing patterns in the common secreted metabolite [1]. The measurement of the isotopic
distribution of the common metabolite provides information to determine the com-
peting pathways. In the case of Fig. 3.3, the asymmetry is introduced by the epimerase
reaction in the last step of the four-step pathway, so that the flux ratio can be deter-
mined from the difference in the l3C label recovered on the 3rd and 5th carbon atom of
lysine, if [I-l3C] glucose or [3-l3C] pyruvate is used as substrate [58,25].
The redundant information obtained from the isotopic-tracer techniques can be
used to confirm the flux estimates obtained by material balancing and test the accu-
racy of the assumptions made about intracellular biochemistry. In the first use of
l3C-Iabeled glucose with hybridoma cultures [59-62] it was possible to confirm the
flux estimates that have been independently obtained through extracellular metabo-
lite measurements and material balancing. In this method, the labeling pattern of
secreted lactate was predicted from the fluxes estimated by material balancing and
found to agree well with the isotope enrichments measured by NMR spectroscopy.
This method uses isotope enrichment measurements (henceforth referred to as 'car-
bon enrichment analysis') to determine metabolic fluxes by applying isotope mass
balance around every carbon atom of the metabolic network. Because of isotopic
steady-state assumption, the label flux into a carbon atom should be equal to the
label flux out. The isotope mass balances provide additional constraints to the meta-
bolite balancing ones of Eq. (3.1) and could enhance the observability of the system.
114 3 Metabolic Flux Analysis

Oxaloacetate

Pyruvate

y (l-y)

l".~
(I_Yl/2[ (I-Yl/21

I O[ 0 2 0 3 04 P3 P2
Lysine Lysine Lysine

Fig. 3.3. Template for lysine labeling from pyruvate and oxaloacetate through the splitting
pathway at the H4D branch point, which is part of lysine biosynthesis. The four-step succiny-
lase pathway (right) contains the enzymatic sequence N-succinyl-2,6-ketopimelate synthase, N-
succinylaminoketopimelate:glutamate aminotransferase, N-succinyl-diaminopimelate desucci-
nylase and diaminopimelate epimerase. The one-step pathway (left) involves the action of
meso-diaminopimelate (meso-DAP) dehydrogenase. Subscripts on 0 and P indicate the original
carbon positions in oxaloacetate and pyruvate. y denotes the fraction of the flux out of the
H4D node that goes via the one-step pathway

In Fig. 3.1, we showed that metabolite balancing cannot differentiate between for-
ward and backward flux of a reversible reaction. In Fig. 3.4 we show how this singu-
larity can be resolved in the particular network of Fig. 3.1 using carbon enrichment
analysis. We need to note that the examples in this section refer to carbon enrich-
ment analysis, but the same analysis can be applied if isotope of another element
(such as phosphorus or oxygen) is used to resolve the structural singularities of the
metabolic network.
In order to get a matrix representation for carbon enrichment balancing, as in
Eq. (3.1), and facilitate the use of the method for the analysis of complex networks,
Zupke and Stephanopoulos [63] introduced the concept of atom mapping matrices.
Marx et al. [64,65] used the carbon enrichment analysis to elucidate the overall net-
work of C. glutamicum. The method is simple and, in combination with metabolite
balancing, very powerful, but its complexity increases with the size of the network,
since the carbon enrichments are non-linear function of metabolic fluxes. Addition-
ally, the resolution of l3e NMR prevents the accurate determination of label enrich-
ment at many carbon atoms of intracellular metabolites, because of low intracellular
3.2 Flux Quantification Methods 115

Isotope Mass Balance around carbon:

D t : V2Bl:;:: (rn+Vs)D1 => Bz= DJ (because of the mass balance around D-see Pig. 3.1)
D 2 : vIR] = (rO+vS)D2 => B]= D z (because of the mass balance around D)
E 1: Vs Dl = (fE + V4) E1 => DJ = E\ (because of the mass balance around E)
E 2: Vs Dz = (fE + V4) E2 => Dz = ~ (because of the mass balance around E)

Bz=Dl =E 1
Bl ;D,;E,

Isotope Mass Balance around carbon:

!
B 1: V2 AJ + V3- C] = (vz + v] +) Bl
B 2: Vz Az + V3' CZ = (V2 + v] +) B2
C 1: V3 + B1 + V4 EJ = (V3- + rC) C 1 => V3 + BI + V4 B2 = (V3' + rC) C 1
C 2: V3 + Hz + V4 ~ ;: (V3- + rC) Cz => V3 + Hz + V4 BI :;:: (V3- + rC) Cz
Solve for v3+ and v3
Net flux of reactions 2 and 4 (see Fig. 3.1)

VI =-rA-(v; -v;-}
v4 ;:: rC - (v; -v;-)

Fig. 3.4. Differentiation of the forward and backward fluxes of a reversible reaction using car-
bon enrichment measurements. Assuming that all network metabolites have 2 carbon atoms
and the transfer of label through the network above is done as shown with the black and white
carbon atoms (i.e. Al -+ B], A2 -+ B2, B] -+ D2, B2 -+ Db B] -+ Cb B2 ---t C2, D] ---t Eb D2 ---t E2>
E] ---t Cl> E2 ---t C2), then measuring the label enrichment ofD1 and D2 (or E1, E2 or D1, E2 or
D2, El) will allow the determination of the forward and backward flux of reaction 3 of the
network. The flux split ratio at B (see caption of Figure 3.1) is simultaneously determined.
Since the system is now fully resolved, any additional label enrichment measurement provides
redundant information that can be used in consistency analysis

concentrations, especially in microorganisms. This reduces the amount of informa-

tion obtained about the system and there are parts of networks that cannot be ob-
served with carbon enrichment analysis because of lack of data [64]. In addition,
carbon enrichment analysis cannot elucidate to a detailed extent networks contain-
ing metabolic cycles [1,51].
To maximize the amount of information that can be extracted from experiments
with 13 e labeled substrates, one needs to account for all isotopomers [51,55,56,66]
that arise in a pathway for a certain labeled substrate. Isotopomers are the molecules
of the same metabolite with different labeling patterns. Isotopomer distribution ana-
lysis is based on the formation of steady-state isotopomer balances [66, 67] around
every metabolite of the network. The isotopomer balances allow the determination
of the isotopomer population as a function of metabolic fluxes.
116 3 Metabolic Flux Analysis

Isotopomer distribution analysis can use information from carbon enrichment

measurements, fine structure of NMR spectra, and mass isotopomer distribution
measurements from GC-MS. The degree of label enrichment at any carbon atom of
the network can be derived from the isotopomer distribution of the corresponding
metabolite as the sum of the relative populations of metabolite isotopomers labeled
at this particular carbon atom [66, 68]. Therefore, the carbon enrichment analysis is
an inherent component of isotopomer distribution analysis. The observed multiplet
pattern of NMR spectra at various carbon atoms is the result of the superposition of
all metabolite isotopomers that are labeled at those carbon atoms [66-67, 69-70].
The line splitting is due to l3C_l3C coupling between adjacent carbon atoms. There-
fore, the ratio of lines in a particular multiplet is equal to the ratio of isotopomers
contributing to the formation of these lines (for more details see [51, 55, 56, 66]).
Mass isotopomer distribution measurements obtained from GC [LCl-MS can provide
further information on the population of isotopomers with different molecular
weight [66, 67]. Thus, all three kinds of measurements can, through the isotopomer
balances, provide additional constraints on the metabolic fluxes, besides those im-
posed from metabolite balances.
The general flow of calculations for the determination of metabolic flux distribu-
tion is illustrated in Fig. 3.5. Although the flux quantification methods are in princi-
ple rather simple, their implementation on integrated complex networks can be in-
volved and computationally intensive. It requires the solution of steady state bal-
ances for all metabolites and their isotopomers in the network, which is a demand-
ing task for realistic networks, since the isotopomer balances are non-linear func-
tions of fluxes. They have to be solved in an iterative way to allow estimation of the
flux distribution that is consistent with the imposed constraints. The number of un-
knowns increases with the complexity of the system and the reaction reversibility.
Whereas reaction reversibility is not important for the application of metabolite bal-
ancing technique, it has to be considered in the isotopic tracer techniques, since it
affects labeling scrambling through the network. Neglecting it will lead to wrong
estimates of metabolic flux distribution and wrong conclusions about the intracellu-
lar biochemistry [17, 18,65,71].
A great degree of redundancy is built into the calculations if all the available mea-
surements presented above are considered. This allows one to test the accuracy of
measurements used and the consistency of the assumed intracellular biochemistry
with the overall data through satisfaction of redundancies. Identification of inconsis-
tencies between the theoretical predictions and the overall data will be an indication
of the presence of errors either in the measurements or in the assumed biochemical
pathways. If some measurements are suspected of containing gross errors, either
they are eliminated and the flux estimation process is repeated with the rest of the
measurements or a new set of measurements is used (for more details see [13]). If
none of the measurements is identified as containing gross errors and the inconsis-
tency remains, this is an indication of errors in the assumed biochemistry. Then the
flux quantification process can be repeated by altering the biochemical pathways,
until a satisfactory fit is obtained (see schematic diagram in Fig. 3.5). However, be-
cause of the nonlinearity of the isotopomer balancing equations with respect to the
metabolic fluxes, classical statistical methods for the analysis of the quality of the
experimental measurements [13] and the estimated fluxes [6, 14, 16] cannot be read-
ily applied. The development and application of more sophisticated statistical tech-
niques to include the nonlinear case of carbon label systems is required (see [70,
3.2 Flux Quantification Methods 117

Initial as.sumplion~ about the

inlr3Ctllular biochernisuy

Biochemlsuy

fu;dicJ;
Fomlulatc • tubel c:nriehmcnu
Sol\'( (or mct:abolitc
mctOlbohtc • ----+ • l.sotopumcr molecular
iliiotopomer pocJl!o
iscxopOmer "eight dl~di~ltIbution
• Fine MmclUrt of NMR
t botlanccs t

Initial flux ------+ Readjust nux

No
Is the agreeme nt
t
satisfactory']

Yes

Check for
gross errors In
No nW:3..IiUremenlS
and lJ~~unled
biochtmis.tl")'

Change se l of
! Yes

experimental data and Yes Gro So errors In

repeat the nux measuremenl O!
dclemllnallon process

Modify
inlraccllulou
DlochcnmM)'

Fig. 3.5. Schematic diagram of the flux determination process based on information obtained
from isotope enrichments, fine structure of NMR spectra, and molecular weight distribution
by GC [LCl-MS. It consists of two parts:l) the iterative process leading to the determination of
flux estimates that satisfy the non-linear system of constraints for an assumed biochemistry (it
is illustrated by the part of the diagram in the box) and 2) consistency analysis to identify
gross errors in the measurements or the assumed biochemistry (see text) (it is illustrated by
the part of the diagram outside the box)

71]). In addition, suitable computer programs that facilitate the analysis of label
transfer and calculation of isotopomer distribution through the metabolic network
(e.g., [72,73]) can be very useful for the selection of the type oflabeled substrate(s)
that facilitates the elucidation of a particular metabolic network. In the beginning of
this section we noted that, in order to delineate the flux distribution in a part of the
network using isotopic-tracer techniques, it is important to select carefully the label-
ing of the substrate and to be able to measure the isotope distribution of this (or
those) metabolite(s) that contain information about this part of the network. Still
we lack a systematic way of experimental design. To date it is not clear, before the
actual experiment is carried out, if and to what extent the selected labeling of sub-
strate, or if and which experimental measurements, can enhance the observability of
118 3 Metabolic Flux Analysis

the metabolic system and delineate structural singularities that cannot be resolved
using only extracellular accumulation rate measurements.
In addition to the development of sophisticated and systematic computational
methods for accurate analysis of the experimental data, the interest should also be
focused in the enhancement of the resolution of the experimental analytical techni-
ques. The low resolution of l3C NMR for low intracellular metabolite concentrations,
which is mostly the case in microorganisms, may be the source of gross errors in the
experimental measurements. Recently, the isolation and hydrolysis of intracellular
proteins and nucleic acids [66, 74] enabled the collection of high precision carbon
enrichment measurements for intracellular metabolites that were previously unat-
tainable [18,65,66]. These analytical methods require that the system is at metabo-
lite and isotopic steady state, so that the measured enrichment of the macromole-
cules reflects the actual in vivo enrichment of their precursors. Recently, two-dimen-
sional NMR, using small amounts of uniformly labeled substrates, was successful as
an additional method for the determination of metabolic fluxes [26, 70, 75, 76].

3.3
Applications of Metabolic Flux Analysis in the Elucidation
of Metabolic Networks

A large number of publications on the applications of metabolic flux determination

to the analysis of metabolic networks can be found in the recent literature. In the
short period of time since fluxes became a central focus of metabolic engineering
as fundamental determinants of cell physiology, their quantification at various ge-
netic or environmental conditions has provided valuable information about many
biological systems. The examples below were selected to illustrate how metabolic flux
analysis can upgrade the information contained in experimental measurements to
provide further insights into the cellular metabolism.
Metabolic flux distribution at specific genetic and environmental conditions repre-
sents a 'snapshot' [76] of metabolic function. It shows the relative activity of the
various pathways under these conditions. It is, though, through the comparison of
many 'snapshots' that information about the control of flux and a better understand-
ing of the intracellular metabolism can be obtained. Utilizing different carbon
sources, mutants and other perturbations, and the corresponding flux distributions,
Vallino and Stephanopoulos [45, 46] analyzed the rigidity of the two 'principal
nodes' of lysine production network of C. glutamicum. The branch points of the net-
work that have the most direct impact on product yield are defined as 'principal' or
critical [3, 4]. Rigidity is the 'inherent resistance to flux alterations' [3]. Thus, if a
node is found to be rigid, the flux distribution around this node is expected to re-
main relatively insensitive to perturbations of the enzymes surrounding it [3]. Ana-
lysis of the rigidity of principal nodes can provide insights about the control of the
network. In the lysine production network, the phosphoenolpyruvate (PEP)/pyru-
vate node was found to be rigid, as opposed to the glucose-6-phosphate (G6P) node,
which was identified as flexible. These results suggest that genetic modifications to
enzymes around the PEP/pyruvate node are more promising with respect to the im-
pact on product yield than those around the G6P node.
Another example of MFA in providing insights about network structure is the
work of Park et al. [20]. Through the use of metabolite balancing combined with
3.3 Applications of Metabolic Flux Analysis in the Elucidation 119

isotopomer distribution analysis in selected mutants, they obtained direct evidence

of the presence of pyruvate carboxylating activity in C. glutamicum. The pyruvate
carboxylase gene was indeed cloned later in C. glutamicum [77]. This is an impor-
tant finding as the PEP/pyruvate node is one of the principal nodes of the C. gluta-
micum network responsible for the production of glutamate and lysine - two main
products of this microorganism. In another analysis, Nissen et al. [21] studied the
anaerobic growth of Saccharomyces cerevisiae under various growth conditions, con-
sidering a compartmentalized intracellular network. By adding or ignoring reactions
corresponding to the function of a particular isoenzyme and testing the feasibility of
the flux distribution for each network configuration, they were able to analyze the
possible role of various isoenzymes in the network. Nyberg et al. [78, 79] investi-
gated potential causal factors of glycosylation variability in cultures of Chinese Ham-
ster Ovary (CHO) cells producing recombinant human interferon-y (IFN-y) using
metabolite balancing analysis. A correlation between glycosylation site occupancy
and the TCA cycle activity in glucose limited chemos tats was identified, suggesting
the possible limitation of glycosylation by the availability of activated sugar precur-
sors, verified experimentally later [78, 79]. Metabolic flux analysis was also applied
to the analysis of Bacillus subtilis [26,80]' E. coli [81], and Penicillium chrysogenum
[8,82] among other studies. The pentose phosphate pathway and the pyruvate shunt
were identified as major pathways of glucose catabolism in riboflavin-producing B.
subtilis [26]. Analyzing the intracellular flux distribution of two phenylalanine-pro-
ducing recombinant E.coli strains, Chen et al. [83] studied the effects of using differ-
ent glucose uptake systems in phenylalanine production. Analyzing the penicillin
production by P. chrysogenum, J0rgensen et al. [8] and Henriksen et al. [82] found
a correlation between the flux split ratio at the G6P node and the yield of penicillin
on glucose.
Varma and Palsson [28, 39] studied the metabolic capabilities of E. coli using lin-
ear optimization analysis. This allowed them to identify the limiting factors in the
production of the biosynthetic precursors and the relative importance of the biosyn-
thetic precursors to biomass generation. Sensitivity analysis enabled the characteri-
zation of the effect of factors such as changes of the active pathways, metabolic de-
mands, P/O ratio or maintenance requirement on the maximum biomass yield. Sauer
et al. [10] used linear programming to investigate the metabolic capacity of B. sub-
tilis for the production of purine pathway related biochemicals. This analysis is im-
portant because it provides valuable insights of the stoichiometric limitations of ba-
cilli bacteria, a strain commonly used in industry. Their results showed that the
maximum theoretical yields of purine pathway related biochemicals are limited by
pathway stoichiometry, providing the host generates sufficient energy. PEP was iden-
tified as the bottleneck for the maximum yield when glucose is the substrate. Knowl-
edge of this limitation allowed the determination of potential ways to overpass it.
Linear programming was also used for the analysis of growth of S. cerevisiae on
various glucose ethanol/mixtures [9].
It was mentioned before that an important goal of metabolic engineering is to
elucidate the parameters responsible for the control of flux. We also presented some
examples of how metabolic fluxes, and in particular their variations at different con-
ditions, can be used to study the interaction between different pathways, as well as to
identify the principal nodes in a metabolic network and examine their rigidity. How-
ever, the analysis of flux distribution at different genetic and environmental condi-
tions can provide qualitative measures of the control of flux. Only in the proper
120 3 Metabolic Flux Analysis

framework can flux analysis be used to quantify the regulation of a metabolic net-
work. Metabolic Control Analysis (MCA) [84,85], developed in 1970s for the quan-
tification of the degree of control exercised by specific enzymes, metabolites, and
effectors in a network upon each network flux, provides such a framework. MCA
relates system variables, such as the flux, to the system parameters, i.e., enzyme ac-
tivities, through a set of control coefficients. The most important of these coeffi-
cients is the 'flux control coefficient' (FCC) defined as

(3.4)

where Ij is the steady-state flux through reaction j and Ej is the activity of the i-th
enzyme of the network [1]. FCCs quantify the effect of each enzyme of the network
on each of the fluxes through the different reactions. It should be noted that experi-
mental determination of individual FCCs requires the measurement of flux changes
through a reaction j following a known change in the activity of the enzyme i.
Hence, accurate determination of flux control coefficients, which is essential for the
reliability of the MCA analysis, depends on the accuracy of flux determination pro-
cess and the validity of the measured flux distribution.
An extension of control analysis through individual FCCs, suitable for the analysis
of complex metabolic networks, is to collect individual reactions around various
branch points in groups and then describe the degree of control exercised by each
group of reactions through the introduction of group control coefficients [86]. Meth-
ods have been developed for the determination of group control coefficients from
the measurement of fluxes and flux changes caused by the introduction of specific
genetic and environmental perturbations to the system [87]. Figure3.6 shows the
calculation of such control coefficients for the three groups of reactions formed
around the PEP/pyruvate branchpoint of C. glutamicum, along with the list of the
various perturbations that were applied for this purpose [88]. The magnitude of
these coefficients suggests that the main control in the production oflysine lies with-
in the group of reactions in the lysine biosynthetic pathway, while a moderate effect
is expected from reactions in the TCA cycle.
MCA is one of several means to analyze metabolism. Kinetic models have also
been used to quantify the control of flux. The kinetic models describe reaction rates
in the network as a function of the parameters that affects them [12], such as meta-
bolite concentrations, enzyme activities, effector concentrations, etc. (e.g., [88,89]).
Schlosser et al. [91] were able to determine sensitivity parameters characterizing
metabolic control based on the formulation of a linear kinetic model. A more so-
phisticated (log)-linear kinetic model [92], which takes into consideration the non-
linearity of enzyme kinetics, has been used to identify the optimal regulatory struc-
ture to achieve a certain objective in a metabolic network [93]. It should be noted
that both MCA and metabolic optimization through kinetic models rely on the
knowledge of flux. The degree to which both approaches can describe the regulation
of a metabolic network depends on the amount of information that can be extracted
from the metabolic system based on the analysis of fluxes.
It is thus seen that metabolic fluxes are indeed fundamental determinants of cell
physiology. Knowledge of intracellular fluxes is essential in resolving the metabolic
state of the cells to a finer detail. Consequently, flux quantification supports the ana-
3.4 Conclusions 121

• Perturbations:
• Glucose concentration
• GPlmutant
• Gluconate addition
• Fluoropyruvate
• Threonine inhibition

1
Perturbed branch

A B C
Affected A 0.07 0.51 0.42 gFCC's
branch B 0.09 1.22 -0.31
C 0.02 -0.34 1.32
Fig. 3.6. Determination of group control coefficients (gFCCs) for the groups of reactions
formed around the PEP/pyruvate branch point in the lysine biosynthesis pathway of C. gluta-
micum. gFCCs were determined from flux experimental data obtained after the indicated ge-
netic and environmental perturbations. The gFCCs of the table represent the consensus esti-
mates from all these perturbations. Rows of the gFCC table refer to the affected branch of
reactions, and columns to the perturbed branch of reactions. For example, the element of the
first column and third row (0.09) is the gFCC of branch A on branch C and measures the effect
of branch A on the flux of branch C

lysis of the regulatory mechanisms of metabolic networks [5, 12, 26], a major objec-
tive of metabolic engineering towards the goal of rational pathway manipulation.

3.4
Conclusions
Metabolic fluxes are the fundamental determinants of cell function and carry valu-
able information about the metabolism, its structure, and its control. It has been
shown how experimental information can be upgraded through metabolic flux ana-
lysis to provide insight about the metabolic state of the cells, about their stoichio-
metric limitations and theoretical yields. They can be used to identify intracellular
regulatory mechanisms and calculate control coefficients or equivalent sensitivity
parameters. The accurate quantification of metabolic fluxes is a major part of the
analysis of metabolic networks. Such knowledge forms the basis for directed mod-
ifications of metabolic networks to achieve a certain objective. It has to be empha-
sized that the accuracy of any metabolic control models, which are used to unravel
the intracellular regulatory structure and examine the metabolic flexibility of the
cells, is strongly connected with the accuracy of the flux estimates. Therefore, the
flux becomes a focal point of metabolic engineering and justifies further research
into the improvement of methods for flux quantification, even though the latter is
not a trivial task and it requires extensive instrumentation along with expensive
media, as well as sophisticated analytical and computational techniques.
122 3 Metabolic Flux Analysis

Although the emphasis has been on metabolic fluxes and metabolic pathways, the
concepts and methods of metabolic engineering can be used, in general, for the ana-
lysis of any bioreaction network. Defining the flux as the rate at which material is
processed through a metabolic pathway, all the flux determination methods de-
scribed above are based on the law of mass conservation, expressed either as meta-
bolite or isotope balances. We can introduce the same concept for pathways transdu-
cing fluxes of energy and their determination will be based on the law of energy
conservation. In a similar way, the field of metabolic engineering can be broadened
and its principles can be applied in the analysis of signal transduction or other path-
ways. When what today is called "signal" is identified in terms of physicochemical
laws, the metabolic flux analysis methods could be readily applied for the elucidation
of signal transduction pathways or any corresponding networks. The only difference
will be the use of 'signal' balances instead of mass or energy ones.

References
1. Stephanopoulos G, Nielsen J, Aristidou A (1998) Metabolic engineering. Academic Press,
San Diego
2. Bailey JE (1991) Science 252:1668
3. Stephanopoulos G, Vallino JJ (1991) Science 252:1675
4. Stephanopoulos G, Sinskey AJ (1993) TIBTECH 11:392
5. Stephanopoulos G (1998) Metabolic Engineering 1:1
6. Vallino JJ (1990) PhD Thesis, Massachusetts Institute of Technology, Cambridge, MA,
USA
7. Vallino JJ, Stephanopoulos G (1993) Biotechnol Bioeng 41:633
8. J0rgensen H, Nielsen J, Villadsen J, M011gaard H (1995) Biotechnol Bioeng 46:117
9. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
10. Sauer U, Cameron DC, Bailey JE (1998) Biotechnol Bioeng 59:227
11. Takiguchi N, Shimizu H, Shioya S (1997) Biotechnol Bioeng 55:170
12. Bailey JE (1998) Biotechnol Prog 14:8
13. Wang NS, Stephanopoulos G (1983) Biotechnol Bioeng 25:2177
14. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KCAM (1994) Biotech-
nol Bioeng 43:3
15. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KCAM (1994) Biotech-
nol Bioeng 43:11
16. Vallino JJ, Stephanopoulos G (1989) Flux determination in cellular bioreaction networks.
In: Sikdar SK, Bier M, Todd P (eds) Frontiers in bioprocessing. CRC Press, Boca Raton,
p205
17. Wiechert W, de Graaf AA (1997) Biotechnol Bioeng 55:101
18. Follstad BD, Stephanopoulos G (1998) Eur J Biochem 252:360
19. Aiba S, Matsuoka M (1979) Biotechnol Bioeng 21:1373
20. Park SM, ShawReid C, Sinskey AI, Stephanopoulos G (1997) Appl Microbiol Biot 47:430
21. Nissen TL, Schulze U, Nielsen J, Villadsen J (1997) Microbiology 143:203
22. Park SM, Sinskey AJ, Stephanopoulos G (1997) Biotechnol Bioeng 55:864
23. Papoutsakis ET (1984) Biotechnol Bioeng 26:174
24. Bonarius HPJ, Schmid G, Tramper J (1997) TIBTECH 15:308
25. Shaw-Reid CA (1997) PhD Thesis, Massachusetts Institute of Technology, Cambridge,
MA, USA
26. Sauer U, Hatzimanikatis V, Bailey JE, Hochuli M, Szyperski T, Wuthrich K (1997) Nat
Biotechnol 15:448
27. Savinell JM, Palsson BO (1992) J Theor BioI 154:421
28. Varma A, Palsson BO (1993) J Theor BioI 165:477
29. Varma A, Palsson BO (1994) Bio-TechnoI12:994
30. Papoutsakis ET, Meyer CL (1985) Biotechnol Bioeng 27:50
References 123

31. Papoutsakis ET, Meyer CL (1985) Biotechnol Bioeng 27:67

32. Fell DA, Small JR (1986) Biochem J 238:781
33. Majewski R, Domach M (1990) Biotechnol Bioeng 35:732
34. Bertsimas D, Tsitsiklis IN (1997) Introduction to linear optimization. Athena Scientific,
Belmont, Massachusetts
35. Edwards JS, Palsson BO (1998) Biotechnol Bioeng 58:162
36. Dantzig GB (1963) Linear programming and extensions. Princeton University Press, Prin-
ceton
37. Pramanik I, Keasling JD (1997) Biotechnol Bioeng 50:398
38. Savinell JM, Palsson BO (1992) J Theor BioI 154:455
39. Varma A, Palsson BO (1993) J Theor BioI 165:503
40. Varma A, Boesch BW, Pals son BO (1993) Appl Environ Microbiol 59:2465
41. Varma A, Boesch BW, Pals son BO (1993) Biotechnol Bioeng 42:59
42. Bonarius HPJ, Hatzimanikatis V, Meesters KPH, de Gooijer CD, Schmid G, Tramper J
(1996) Biotechnol Bioeng 50:299
43. Bonarius HPJ, Timmerarends B, de Gooijer CD, Tramper J (1998) Biotechnol Bioeng
58:258
44. Papoutsakis ET (1984) Biotechnol Bioeng 26:1741
45. Vallino JJ, Stephanopoulos G (1994) Biotechnol Prog 10:320
46. Vallino JJ, Stephanopoulos G (1994) Biotechnol Prog 10:327
47. Pissarra PN (1996) PhD Thesis, University of London, UK
48. Pissarra PN, Henriksen CM (1998) FLUXMAP. A visual environment for metabolic flux
analysis of biochemical pathways. In:Yoshida T, Shioya S. (eds) Preprints of the 7th Inter-
national Conference on Computer Applications in Biotechnology, May 31-June 4, 1998,
Osaka, Japan, p 339
49. Shulman RG, Brown TR, Ugurbil K, Ogawa S, Cohen SM, den Hollander JA (1979) Science
205:160
50. Dickinson JR, Dawes IW, Boyd ASF, Baxter RL (1983) Proc Nat! Acad Sci USA 80:5847
51. Chance EM, Seeholzer SH, Kobayashi K, Williamson JR (1983) J BioI Chern 258:13,785
52. Walsh K, Koshland DE (1984) J BioI Chern 259:9646
53. den Hollander JA, Ugurbil K, Brown TR, Bednar M, Redfield C, Shulman RG (1986) Bio-
chemistry 25:203
54. Crawford A, Hunter BK, Wood JM (1987) App Environ Microbiol 53:2445
55. Malloy CR, Sherry AD, Jeffrey FMH (1988) J BioI Chern 263:6964
56. Malloy CR, Sherry AD, Jeffrey FMH (1990) Am J Physiol 259:H987
57. Di Donato L, Des Rosiers C, Montgomery JA, David F, Garneau M, Brunengraber H (1993)
J BioI Chern 268:4170
58. Sonntag K, Eggeling L, de Graaf AA, Sahm H (1993) Eur J Biochem 213:1325
59. Mancuso A, Sharfstein ST, Tucker SN, Clark DS, Blanch HW (1994) Biotechnol Bioeng
44:563
60. Sharfstein ST, Tucker SN, Mancuso A, Blanch HW, Clark DS (1994) Biotechnol Bioeng
43:1059
61. Zupke C, Sinskey AI, Stephanopoulos G (1995) Appl Microbiol Biotechnol 44:27
62. Zupke C, Stephanopoulos G (1995) Biotechnol Bioeng 45:292
63. Zupke C, Stephanopoulos G (1994) Biotechnol Prog 10:489
64. Marx A, de Graaf AA, Wiechert W, Eggeling L, Sahm H (I996) Biotechnol Bioeng 49: 111
65. Marx A, Striegel K, de Graaf AA, Sahm H, Eggeling L (1997) Biotechnol Bioeng 56:168
66. Klapa MI, Park SM, Sinskey AJ, Stephanopoulos GN (1999) Biotechnol Bioeng 62:375
67. Park SM (1996) PhD Thesis, Massachusetts Institute of Technology, Cambridge, MA, USA
68. Wiechert W, de Graaf AA (1996) In vivo stationary flux analysis by 13C labeling experi-
ments. In: Sahm H, Wandrey C (eds) Advances in biochemical engineering biotechnology.
Springer, Berlin Heidelberg New York, p 109
69. Park SM, Klapa MI, Sinskey AI, Stephanopoulos GN (1999) Biotechnol Bioeng 62:392
70. Schmidt K, Nielsen J, Villadsen J (1999) J Biotech 71:175
71. Wiechert W, Siefke C, de Graaf AA, Marx A (1997) Biotechnol Bioeng 55:118
72. Schmidt K, Carlsen M, Nielsen J, Villadsen J (1997) Biotechnol Bioeng 55:831
124 3 Metabolic Flux Analysis

73. Zupke C, Tompkins R, Yarmush M (1997) Anal Biochem 247:287

74. Szyperski T (1995) Eur J Biochem 232:433
75. Szyperski T, Bailey JE, Wuthrich K (1996) TIBTECH 14:453
76. Nielsen J (1998) Biotechnol Bioeng 58:125
77. Koffas MAG, Ramamoorthi R, Pine WA, Sinskey AJ, Stephanopoulos G (1998) Appl Mi-
crobiol Biotechnol 50:346
78. Nyberg GB, Balcarcel R, Follstad BD, Stephanopoulos G, Wang DIC (1999) Biotechnol
Bioeng 62:336
79. Nyberg GB (1998) PhD Thesis, Massachusetts Institute of Technology, Cambridge, MA,
USA
80. Sauer U, Hatzimanikatis V, Hohmann HP, Manneberg M, vanLoon APGM, Bailey JE
(1996) Appl Environ Microb 62:3687
81. Chen R, Bailey JE (1994) Biotechnol Progr 10:360
82. Henriksen CM, Christensen LH, Nielsen J, Villadsen J (1996) J Biotechnol 45:149
83. Chen RZ, Hatzimanikatis V, Yap WMGJ, Postma PW, Bailey JE (1997) Biotechnol Progr
13:768
84. Kaeser H, Burns JA (1973) Symp Soc Exp Bioi 27:65
85. Heinrich R, Rapoport TA (1974) Eur J Biochem 42:39
86. Brown GC, Hafner RP, Brand MD (1990) Eur J Biochem 188:321
87. Stephanopoulos G, Simpson TW (1997) Chern Eng Sci 52:2607
88. Simpson TW, Shimizu H, Stephanopoulos G (1998) Biotechnol Bioeng 58:149
89. Savageau MA (1969) J Theor Biology 25:365
90. Savageau MA (1969) J Theor Biology 25:370
91. Schlosser OM, Holcomb T, Bailey JE (1993) Biotechnol Bioeng 41:1027
92. Hatzimanikatis V, Floudas C, Bailey JE (1996) AIChE Journal 42:1277
93. Hatzimanikatis V, Emmerling M, Sauer U, Bailey JE (1998) Biotechnol Bioeng 58:154
94. Varma A, Palsson BO (1994) Appl Environ Microbiol 60: 3724
4 Accuracy and Reliability of Measured Data
Bernd Hitzmann

4.1
Accuracy and Reliability of Measured Data

All measurements have the ultimate goal of creating information. Measurements are
also the basis for new bioprocess development. The study of cell metabolism and
regulation would be impossible without reliable analytical methods. Furthermore,
process optimization and control are also based on reliable measurements. Biopro-
cess analyzers have the fundamental task to provide the information for optimizing
microbial growth, yields, and product quality. Therefore, measurements are not only
a key issue in modern process development, but they also open the door to process
supervision and control, and avoid restricted views when analyzing processes just
through a keyhole.
Increasingly sophisticated equipment has been developed due to the growing de-
mand on high quality measurements. Almost no measurement performed in bio-
reaction engineering is direct, i.e., none is obtained by immediate comparison with a
reference quantity. Most measurements are achieved by means of some specific phy-
sical or chemical property of the analyte, often hidden in the very complex sample
matrix common to bioprocesses. This very complex matrix, which consist of the
numerous substances surrounding the analyte and their mutual influences, makes
the task of measuring difficult and leads to the development of complicated measure-
ment systems. Validation of these systems under real process conditions requires
even more ingenuity than their development. However, the greater the measuring
system complexity, the more difficult is it to ensure the accuracy and the reliability
of the collected data. The need for accuracy and reliability increases even more if the
data of a process analyzer is to be used as input for a controller. If unusual measure-
ments are made the system operator is required to decide if the fault is from the
analyzer or from the bioprocess itself.
An additional typical requirement is that the operation of the analyzer and its
service must be as simple as possible. The aim of using such a measurement system
is to get more information, not to increase the chances of malfunction of the whole
process. Due to the complex nature of many process analyzers this requirement can
only be met if the degree of automation of such a system is very high. To achieve this
goal, the automation system must not only collect and evaluate the measured data
and generate direct actions if necessary (i.e., control of valves) but also make deci-
sions based on sophisticated data analysis. Making such complex decisions requires
an additional module. This decision-making module must be able to supervise the
analyzer in a similar way to the operator. Causes of faults must be identified and
126 4 Accuracy and Reliability of Measured Data

fixed, if possible. The operator (and potentially other entities) must be informed
using a high level of communication. The satisfactory operation of such a module
demands knowledge-based systems.
The accuracy and reliability of measured data will be discussed in this contribu-
tion. Recommendations are given to guarantee the accuracy and reliability of mea-
sured data in the complex environment that characterizes the field of bioprocess en-
gineering.

4.1.1
Accuracy and Precision of Measurements

When the terms accuracy and precision are applied to measurements in a colloquial
way, they are frequently used as synonyms. In a more scientific way, they have totally
different meanings. In this work, the words accuracy and precision are used as sug-
gested in the literature (i.e., Guide for the use of terms in reporting data in analytical
chemistry, Anal Chim 54(1982):157). The different meaning of these two terms can
be expressed by a cause-effect relationship. The accuracy of a measurement is influ-
enced by the method bias or systematic errors, which are caused by malfunctioning
parts of the measurement system, inhibitors in the sample, unstable reagents, or an
inappropriate sampling point or system and will be recognized as a shift of the mea-
surement results. On the other hand, the precision of a measurement will be affected
by random errors, which will cause a distribution of measurements and can be taken
into account by means of statistical methods. Therefore, measurements can be any
combination of precise/imprecise and accurate/inaccurate.

4.1.2
Accuracy

A new analyzer must be validated, before use, by means of an alternative method

that guarantees the required accuracy. Otherwise, the complex sample matrix can
influence the results significantly without the user's knowledge. Recovery rates can
be used as a measurement of the accuracy of the measurement system. Measure-
ments of samples of known analyte concentration can also be obtained. The mea-
surement system should work under realistic conditions, as similar as possible to the
ones prevailing on the actual system. A common fault is not to use a routinely used
sampling device and procedure during accuracy tests. However, sampling devices
and procedures are a tremendous source of systematic errors and must be calibrated
in the same manner as the process analyzer [1]. Schugerl [2] presents examples of
errors that are produced by the sampling system during a cultivation process. The
sample used during validation of a data collection system typically should have a
matrix which is representative of the real sample. However, this requirement is not
directly applicable to bioprocess engineering, due to the almost unavoidable time
variability of sample composition. This intrinsic complexity of living systems should
not be used as an excuse to use unrealistic samples consisting of pure solutions of
the analyte when validating a measurement system. A more realistic sample matrix
that mimics a real bioprocess environment is, for instance, a yeast extract solution.
Such a solution has been used by Weigel et al. [3] to compare different PIA systems
for sucrose measurement. For validation purposes the absolute accuracy of an indi-
vidual measurement has to be calculated. This parameter is given by
4.1 Accuracy and Reliability of Measured Data 127

Xi-X (4.1)
In this expression, Xi is the measured value and X is the estimate of the true value
of the measured quantity (i.e., the concentration of a standard solution). For repli-
cate measurements, the mean value x is used which yields (for the measure of accu-
racy)
X-X (4.2)
However, it is unlikely that the mean value will be exactly equal to the true value,
even if no systematic error is involved in the corresponding measurement. Of course,
the reason for this is the influence of small, unavoidable random errors. Therefore, a
statistical test is necessary to decide if the observed difference is significant. In that
case, it could be caused by a systematic error, or by random errors. In fact, the null
hypothesis tested is that - apart from random variation - there is no significant dif-
ference between observed and known values. As a threshold, the probability of 5% is
often used. If the probability that the difference is produced by chance is small, the
null hypothesis is rejected. For example, if the probability is less than 5%, the null
hypothesis is rejected and the difference between the observed value and its estimate
is said to be significant at the 5% (or 0.05) level. Sometimes the number of measured
values is small. In that case, the t-distribution has to be applied instead of the normal
(or Gaussian) distribution. Even when the assumption of approximating a set of data
to a normal distribution is not strictly valid, it is still considered an acceptable ap-
proximation. A testing value TV is calculated and compared with a threshold value of
t from a table, considering the significance level as well as the degrees of freedom:

TV= vn 1x - xl
5
(4.3)

In this equation n is the number of measurements and s the standard deviation,

calculated as follows:

i=I
5= (4.4)
n-1
If TV is lower than the tabulated critical value of t with n-1 degrees of freedom
and a significance level of 0.05 (which can be found in most statistical textbooks, for
example Miller and Miller [4], Massart et al. [5], Funk et al. [6]), then no significant
difference can be demonstrated, and the null hypothesis cannot be rejected.
As an example, 15 glucose measurements are presented in Table4.1, by using a
standard solution of 0.25 g r l and a Yellow Springs analyzer. The mean value is
0.255 g I-I and the standard deviation is 0.010 g rl.

Table 4.1. Glucose measurements performed on a standard solution ofO.25gr l

Glucose measurements [g rl]

0.278 0.235 0.261 0.256 0.262

0.247 0.267 0.253 0.255 0.252
0.258 0.253 0.252 0.244 0.249
128 4 Accuracy and Reliability of Measured Data

The calculated TV=1.85 is smaller than the tabulated value of 2.14. No differences
between the expected and measured value can be proved and therefore the null hy-
pothesis cannot be rejected. If, however, the null hypothesis would have been re-
jected due to a higher TV value, then there is a risk of 5% that the hypothesis has
been rejected, even though it is true (this is called type I error). This risk can be
reduced by reducing the significance level to 1% or even to 0.1 %. It is also possible
to retain the null hypothesis even if it is false (what consists of a type II error). To
avoid this error, an alternative hypothesis has to be postulated, for which a new value
- the true value, related to the tested mean value - must be defined.

4.1.3
Precision

To determine how precise a measurement is, a statistic evaluation of a measurement

set of repeated measurements Xi (i=l, ... n) has to be performed and the standard
deviation s of this measurement set is calculated. The mean measured value x and
its standard deviation s are considered estimates of the true values of these quanti-
ties. The relative standard deviation Srel or the variance var(x) are also used to char-
acterize the precision,
S
Sre/ = -=X (4.5)

var(x) =S2 (4.6)

For a set of data that consists of a few measurements, then the mean deviation d or
its relative value d rel can be calculated as a measurement of precision,
n
I: IXi - xl
d=-=--i=--=l_ __ (4.7)
n

d
drel =-=x (4.8)

While these formulas apply to a set of data, there is no universal convention to

estimate the precision of a single measurement. A common practice in analytical
chemistry is to present estimates in the form
x±Ax=x±s (4.9)

or
x±Ax=x±d (4.10)
The standard error of the mean is calculated less often than the standard deviation
S
x±~x=x± v'n (4.11)

Also the 95% confidence limits are frequently used:

x ± ~x = x ± to.95 _s_
n-l v'n
(4.12 )
4.1 Accuracy and Reliability of Measured Data 129

In this equation, the value t~~~ represents the corresponding value of the t-distri-
bution with a confidence level of 95% and n-l degrees of freedom. This confidence
level provides a high probability of inclusion of the true value in the interval. Such
values can be found in most statistical textbooks as mentioned above. All these ex-
pressions can be easily transformed, given that the number of measurements is
known.
If a calibration model (regression line) is used to calculate the concentration of a
sample solution, then the precision depends not only on the experimental error, as
discussed above, but also on the confidence limits of the calibration model Yeon!>
which are given by the expression

(x - X)2
-+----:::-'-----'- (4.13)
n t
i=1
(Xi - X)2

Here bo and b l are the regression coefficients, x is the concentration at which the
confidence limit is calculated, t~_2 is the t value corresponding to a given signifi-
cance level p and n-2 degrees of freedom, and XI and x are the concentrations of the
calibration solution and the mean value, respectively. The term SRsd is the residual
standard deviation with respect to the regression line, which is calculated as

n
I: (yi,mea - Yi,lin.modf
i=1
SRsd = (4.14)
n-2
Here Yi,mea is the i-th measurement datum, and Yi,lin.mod is its corresponding pre-
dicted value using the linear model. It is apparent from the equation to calculate the
confidence limit, as well as from Fig. 4.1, that the larger the difference between X and
x, the larger the term containing the square root becomes. This widens the confi-
dence interval at the edges of the calibration line. The most precise value is obtained
in the center of the measurement range. Another aspect which can be derived from
the equation to calculate the confidence interval is the fact that the term of the

0,9
0 ,8
~ 0 ,7
<l>
::> 0 ,6
"iij
>
c: 0 ,5
'"
.
E 0,4
<l>
:;
0 ,3
co
<l>
::;: 0,2
0,1
0,0
0,5 1,0 1 ,5 2,0 2,5 3,0

Glucose [gil]

Fig.4.1. A glucose calibration function (solid line) and its 95% confidence interval (dashed
lines)
130 4 Accuracy and Reliability of Measured Data

square root is reduced if the number of measurements n is large or if 2:7~1 (Xi - X)2
is large. For a constant n, the summation term is directly proportional to the differ-
ence between Xi and x. This means that the values Xi must be far away from the edges
of the measurement range in order to increase the sum and to decrease the error of
the prediction. In practice, however, this can only be done if the appropriateness of a
linear calibration model is assured. Otherwise, the calibration points should be
evenly spaced over the measurement range of interest and a test should be per-
formed to assure that a linear model is really more appropriate than a quadratic
one. The way that such a test is performed will be described below.
Since the regression line cannot be calculated exactly, it is obvious that the pre-
dicted concentration is subject to error too. The confidence limits for the value of
the analyte concentration can be estimated by the following formula:

Y - bo P SRsd
xconf = ~b-l- ± t n-2 ~ (4.15)

where Ym is the mean value of the measured values for which the confidence limits
will be determined, and m is the number of replicates. If just one measurement has
been performed, it will be Ym and m=l. Y is the mean value of the measurements of
the calibration set. The discussion of the formula is analogous to the one for the
confidence limits of the calibration model. For the latter, a narrow confidence band
requires large values for n, m, b, and the difference between x and Xi. As mentioned
before, this condition can only be achieved if the adequate fit of a linear model is
guaranteed.
As has been mentioned before in this work, the complex sample matrix character-
istic of bioprocess systems will often influence the precision as well as the accuracy
of the measurement. The confidence limit of the predicted concentration will not be
very helpful under this condition. The matrix of the calibration solution should be as
similar as possible to the matrix of the sample solution. However, this condition
cannot be guaranteed in bioprocess techniques because of the time-changing matrix.
Substances that interfere, either inhibiting or increasing the detection reaction, can
experience changes in concentration during a bioprocess run and will therefore in-
fluence the determination of analyte concentration. A procedure which can account
for this consideration is the standard addition method. Using this method, the ori-
ginal sample solution is divided into several samples of the same volume (aliquots).
Variable known amounts of the analyte are added to these samples and all are di-
luted to the same volume. The measured signals of all samples are determined and a
regression line is calculated, for which as usual the independent variable (x-axis) is
the amount of analyte added and the dependent variable (y-axis) is the signal ob-
tained with the corresponding sample. The absolute value of the root of the regres-
sion line is the concentration of the original sample.
The disadvantage of this procedure is that a large amount of sample solution is
required and several measurements have to be performed. As described, the method
uses an extrapolation model, which is less precise than an interpolation technique.
Furthermore, it is difficult to automate.
To compensate this disadvantage, the technique can be transferred to FIA [7-9]. A
multiple injection FIA system is required, where three different solutions are injected
in a fast sequence. The solutions are a standard solution, the sample solution, and a
4.1 Accuracy and Reliability of Measured Data 131

3,0

2,5

'"" 2,0
;g
ic 1,5

'"
iii 1,0

0,5

0,0 +-_"L...,-__,-__,-__,-_--=;:=_--,-
10 30 50 70 90 110 130
Time [sec]

Fig. 4.2. Typical measurement signal obtained by a fast injection of three solutions containing
0.35, 0.55, and 0.6 g rl glucose, respectively. The values marked by the bold line style were

used for network evaluation

second standard solution. Due to dispersion inherent in FIA a mixing of the three
solutions will occur. A typical measurement signal of such a triplicate injection
(L1t1njection time=20 s) is presented in Fig. 4.2.
Partial least squares regression [10) as well as neural networks can be used for the
evaluation of such signals. The later technique will be discussed in this work in more
detail. The neural network was trained to predict the concentration of the three in-
jected solutions, using the individual measurements as input values. As an example,
if the enzyme activity is changing due to effects hidden in the sample matrix, the
signal intensity of the whole superimposed signal will change. The known analyte
concentration of the first and last injected standard solutions allow the reliable cal-
culation of the concentration of the injected sample by linear interpolation. Solving
for the sample concentration yields
(;2 - (;1
C2 = -A--A-(C3 - Cd + C1 (4.16)
C3 - C1

In this equation (;1' (;2> and (;3 are the predicted analyte concentrations of samples
1, 2, and 3 respectively. C1 and C3 are the known analyte concentrations of the two
standard solutions (which have been injected first and last). C2 is the corrected ana-
lyte concentration of the solution injected in second place. Due to the mixture of the
three solutions, the matrix effects of the bioprocess sample will also affect the signals
obtained from the standards. Therefore, these effects will be compensated due to the
application of linear interpolation. The analyte concentration is reliably determined.
Results of the evaluation of multiple injection signals can be seen in Table 4.2. In the
first three columns the concentrations of the injected solutions (standard, sample,
and second standard) are presented. Signals from some of these measurements have
been used for the training of a feed forward back-propagation neural network (con-
sisting of 20 input neurons, 4 hidden neurons and 3 output neurons). The other
signals were applied as a test set to evaluate the training. The relative mean deviation
(called error) is used to express the precision of the evaluation as mentioned above.
The best results obtained can be seen in the fourth and fifth column, where the mean
error (out of four repeated measurements) of the uncorrected prediction for the
132 4 Accuracy and Reliability of Measured Data

Table 4.2. Mean % errors (based on the relative mean deviation) of glucose prediction from
FIA signals of a threefold injection system (for more information see the text)

Mean error of prediction of sample concentration [%)

Standard 1 Sample Standard 2 Predicted directly Applying a linear interpolation
correction
Glucose [g 1-1) Training Test set Training Test set Validation
set set set

0.10 0.15 0.35 6.0 4.7 8.9

0.10 0.20 0.35 2.4 4.6 4.2
0.10 0.25 0.35 2.8 2.5 6.7
0.10 0.30 0.35 1.2 1.3 6.7
0.35 0.40 0.60 1.9 2.2 0.9
0.35 0.45 0.60 0.8 0.8 2.4
0.35 0.50 0.60 1.8 1.4 1.8
0.35 0.55 0.60 6.1 0.9 0.3
0.60 0.65 0.85 3.2 3.7 1.5
0.60 0.70 0.85 1.0 1.0 2.7
0.60 0.75 0.85 1.1 1.1 1.4
0.60 0.80 0.85 1.0 1.1 2.7
Mean values 2.4 2.5 2.2 2.1 3.3

sample concentration (given by the neural network) is shown. The overall mean
error of the training set and the test set are 2.40/0 and 2.50/0, respectively. Similar
results have been obtained for the prediction of the concentration of the two stan-
dards. The correction using linear interpolation has been applied to the correspond-
ing values, although it is not necessary. As can be seen from the results, presented in
column six and seven of Table 4.2, the errors of prediction (2.20/0 for the training,
2.1 % for the test set) are slightly lower.
For the validation of the matrix effect on the compensation technique, the tem-
perature of the enzyme reactor thermostat was reduced from 30°C to 26.5 °C, and
all samples were measured again. The signal intensity was reduced by more than
200/0 due to the temperature shift. This value also represent a rough estimation of
the measurement error if no calibration was performed. Using the proposed com-
pensation technique, a correction is inherent in the evaluation. The results of an
evaluation can be seen in the last column of Table 4.2. Although, the mean overall
error of 3.30/0 is slightly larger than the ones presented before, it is still a good result
and demonstrates the usefulness of the compensation technique.

4.2
Measurement Reliability

The reliability of measured data is especially of utmost importance if the data are to
be used for process control. An operator must use his experience and judgment
when supervising a process analyzer. For example, he can make an assessment of
the order of magnitude of the data collected during a special process state and how
different measurements can be correlated to each other. Furthermore, during a (re)-
calibration procedure the operator knows the way the raw measured data should
4.2 Measurement Reliability 133

look. Re can check if the calculated calibration model is only changing marginally
and, moreover, if this change is continuous rather than abrupt. The process operator
is able to distinguish between unusual process measurements due to a faulty analy-
zer or due to a fault in the production process itself. Also, a fault can happen so that
the operator has to follow the whole process as well as the analyzers very closely.
This, however, is unrealistic for industrial applications because it will be excessively
time-consuming, resulting in additional expense.
The calibration model will not be calculated correctly if a fault occurs during the
measurement of calibration data. Therefore, such a fault will have a tremendous in-
fluence on all further measurements. Because of this essential significance of the
calibration, a system is required to validate the calibration measurements as well as
the calibration model on-line. Such a system must contain the knowledge about mea-
surement interpretation, as mentioned above. Therefore a knowledge-based system
that performs self-testing and self-correcting procedures is required.
The significance of knowledge-based systems in analytical chemistry has been
pointed out by Wunsch and Gansen [11]. These researchers built various expert sys-
tems employing commercial available shells, and demonstrated that analytical deci-
sions can be automated in this way. Schoenmakers [12] presented an overview of
knowledge-based systems developed for analytical chemistry. So far, only a few ap-
plications have been published in which knowledge-based systems were used for the
analysis of on-line process analyzers. Wolters et al. [13] presented a generally applic-
able validation program called VALID. This system was based on an expert system
and evaluates the calibration procedure, the drift of the analytical system, and the
effect of the sample matrix in an interactive way with the user. Peris et al. [14] re-
ported using an expert system developed in Modula-2 for the specification and in-
telligent control of an FIA system for bioprocess monitoring. This expert system
runs in real time, in contrast to VALID. The tasks of this knowledge-based system
are to control the analysis procedure, to select the measurement that will be pro-
cessed next, and to identify faults in the operation. Brandt and Ritzmann [15] re-
ported a knowledge-based system that has been applied to the fast detection and
diagnosis of general faults in FIA systems. This system also runs under real-time
condition without a dialog with a user, in the same way as the system described by
Peris et al. [14]. The system presented by Brandt and Ritzmann [15] has been pro-
grammed by using Vax-OPS5, a tool for the development of production systems. It
runs on a Vax computer and obtains the FIA measurements from another computer,
which has been used to automate fully the FIA system.

4.2.1
Assessment of Measured Data Reliability by Means of a Knowledge-Based System

The knowledge-based system was applied to supervise the FIA system, as a process
analyzer. This FIA system has been used for glucose measurements during various
bioprocesses, as described in detail by Dullau and Schugerl [16]. The glucose con-
centration was converted to glucono lactone using an enzyme reactor with immobi-
lized glucose oxidase. Oxygen is utilized during this reaction. The oxygen uptake was
measured by means of an oxygen probe. The typical peak shaped signal was evalu-
ated using the peak height. In the following paragraph the peak height is just called
"measurement" and can be correlated with the glucose concentration. Eighteen sam-
ples per hour have been processed with this FIA system.
134 4 Accuracy and Reliability of Measured Data

- CAFCA-
Computer Assisted Flow Control & Analysis

Main tasks:
Measuremcnt uptake I II
Conlrol of valves I
Evalualion of dala I
Visllali~.a t ion of results I
Storage of data I I'
Knowledge-ba ed analy 'is

- - to eva lu ate re liablc

.. - to delect faults
- 10 correct fault.
- to relieve the operator
,1 ,

I.

~ MS -DOS -PC

Flo\\
InJecliun
Analy:!ois
Sy~ICIl1

Fig. 4.3. The tasks of the automation system CAFCA containing the knowledge-based analysis
of measurements

An automated data evaluation and management system called CAFCA (Computer

Assisted Flow Control & Analysis) was developed [17] for the control of the FIA
system. This system is already commercially available as a professional version
(ANASYSCON, Hannover). CAFCA was programmed by using Borland Pascal 7.0
and two toolboxes. One toolbox has been used to guarantee the real time require-
ments (RTKernel Version 4.0, On Time Informatik, Hamburg) and the other for the
development of a user-friendly graphic output. CAFCA runs on standard PC compu-
ters using the operating system MS-DOS, and processes all the tasks in a multi-task
manner. As can be seen in Fig. 4.3, one of the tasks is the knowledge-based proce-
dure called XP-CAFCA. All the implemented numerical procedures implemented in
XP-CAFCA have been validated previously [18] and complemented with heuristic
rules. Various statistical tests are executed by XP-CAFCA, which can be obtained
from any statistical textbook (for example Miller and Miller [4], Massart et al. [5],
Funk et al. [6]).
Information required for the development of the knowledge-based system [19] was
obtained by observing the FIA system during several cultivation processes. Depend-
ing on the running time of these bioprocess (which lasted from a few days up to
several weeks), a (re)calibration was performed almost every day. As described else-
where [17] CAFCA can perform a user-activated re-calibration whenever it is re-
quired, or after a user-specified number of measurements. Since a fault can occur
any time during calibration and data collection, supervision of the system would be
very time-consuming. A systematic error analysis of FIA system is presented by
4.2 Measurement Reliability 135

Chen and Zeng [20]. The knowledge-based system was developed to relieve the op-
erator of this task. For this, all the faults collected during a run of the analyzer had to
be collected. Examples of typical faults were plugging of injector, selector valve, or
tubes; standard solutions of incorrect concentration; microbial contamination of en-
zyme cartridges; empty vial; fluctuating flow rate; air bubbles; disconnection of tube;
malfunction of valves; failure of the pump; rupture of a tube fitting; changing of
temperature; reduced sensitivity; damaged detector system; electronic disturbance;
baseline drift; wrong AID-range; hardware breakdown; and wrong measurement
range.
An experienced operator is able to identify these faults, using his judgment. The
knowledge-based system XP-CAFCA was developed following this same principle. It
consists of a numerical-statistical and a knowledge-based module. The numerical-
statistical module is composed of several functions for calculating the characteristics
of the calibration. These characteristics are employed for the judgment of the cali-
bration, using the knowledge-based module. Faulty measurements are selected and
eliminated for modeling purposes. After XP-CAFCA has identified a fault, it reports
the decisions and actions taken to the operator. The main tests and analysis of the
numerical-statistical and knowledge-based module are described in detail in the fol-
lowing section.

4.2.2
Numerical and Statistical Tests Performed by the Knowledge-Based System

The numerical-statistical module uses validated algorithms for time series and sta-
tistical processing, and for data management. Every algorithm of the numerical
module calculates parameters which are used by the knowledge-based module for
decision-making purposes. A model-based approach (called Projective Reference
Evaluation, or PRE) was in order to increase the reliability of raw measurement eva-
luation. Another model-based approach was given by Wu and Bellgardt [21], which
uses a fault detection algorithm that relies on a recursive parameter identification.
Other evaluation methods presented for a reliable evaluation of FIA measurements
are digital filters [22, 23], Fourier transformation [24], wavelet transform [25], as
well as neural networks [26-28]. However, PRE allows not only unreliable measure-
ments to be recognized, but even distorted measurements can be evaluated reliably.
This new evaluation technique will be described in the following section. A more
detailed presentation is available [29].
In FIA, each element of the dispersed sample zone has a different concentration
ratio of sample and carrier solution. Usually the shape of all signals is equal if sig-
nals obtained from different concentrations are normalized so that the peak maxi-
mum has a value of 1.0. However, the information relevant to the way that the ana-
lysis system responds to a measurement is hidden in that peak shape. The theory
underneath this response is too complex to provide simple modeling alternatives.
Therefore, additional information resulting from the evaluation of disturbed signals
is not readily available, but can be exploited in the following way.
A typical FIA-measurement of an unknown sample can be represented as a vector
m. For this purpose, each individual measurement of the peak shaped signal is con-
sidered a component of such a vector. This vector can be related to a reliable refer-
ence measurement r (which can also be obtained by averaging many signals) as
follows:
l36 4 Accuracy and Reliability of Measured Data

m=Ar+e (4.17)
For this calculation the measurement signals are transformed so that the baseline
is equal to zero. The residual vector e represents those contributions to m which
cannot be explained by Ar. From this equation A can be estimated (using the least
squares estimator) by the formula
A=mr/rr (4.18)
and can be used as a latent variable - instead of the peak height or area - in a
(linear) regression model to calculate the concentration of the unknown sample in
a very reliable fashion. It is apparent from the above equation that the higher parts
of a peak, such as the part around the maximum, contribute more to A than the
lower parts (i.e., the peak tail). Because the components of r that belong to the base-
line are equal to zero, they do not contribute to A.
Using this procedure, it is possible to validate the quality of A by inspecting e. If
some components of e are higher than a specific threshold - such as, for example,
more than twice the standard deviation - it is most likely that the corresponding
part of the measurement signal is faulty. Therefore, it should not be used for the
estimation of A. So, A can be estimated again (neglecting those components of m
and r which correspond to the components of e whose values are too high) and e
can be inspected iteratively, until all the components of the residual vector e are
small. The following paragraph provides an example of the usefulness of this tech-
nique, even if the original measurement signal is faulty. In this example, inherent
information contained on r is used to perform the correction mentioned above.
Various patterns of distortions to raw measurement signals of a sucrose FIA sys-
tem can be added to demonstrate the ability of PRE as a reliable signal evaluation
technique. One original measurement (a), as well as three different distortions (b, c,
and d) are presented in Fig. 4.4.
Figure 4.4b represents a distortion by which the measurements between 50 sand
60 s are set to 3.0 units. In Fig.4.4c, each second measured value between 30 sand
85 s was set to a random number from the interval (-0.5,0.5). Figure 4.4d represents
a distortion obtained by adding a minus function to the measured values between
45 sand 65 s. All three distortions represent typical observed faults and have been
applied to signals of sucrose measurements. The precision of this evaluation is ex-
pressed as the relative deviation to the evaluation of the signals without the distor-
tion. The mean value of the five individual measurements is called average error and
is presented in Table 4.3. The measured signal shown in Fig. 4.4 was obtained from a
standard concentration of 2.6 g 1-1. As a result, the relative percentage of the distor-
tions is much bigger for lower concentrations. It can also be observed that the error
is lower than 6.8%, even for severely distorted signals. The average error of all dis-
tortions is 1.7%, which demonstrates the usefulness of the PRE technique.
The knowledge-based system XP-CAFCA uses this evaluation technique to deter-
mine reliably the concentration of each measurement, as well as the calibration data
used for the regression model. Furthermore, it uses the number of rejected raw mea-
surements and the sum of the remaining error (calculated by means of PRE) as a
measure of the quality of the analyte determination. This can also be used as a mea-
sure of the actual performance of the analyzer itself.
Further tests were carried out in order to obtain more evidence of reliable analy-
zer performance. XP-CAFCA does a simple monotony-test as a simple method to
4.2 Measurement Reliability 137

2,0 2,0
a b

1,5 1,5

.l!! .l!!
1,0
'"
::J
1,0
'"
::J

0,5 0,5

0,0 0,0
0 50 100 150 0 50 100 150
Time [sec) Time [sec)

2,0 2,0
C d

1,5 1,5

.l!! .l!!
1,0 1,0
'"
::J '"
::J

0,5 0,5

0,0 -r--=cJ-,---'T-===l
o 50 100 150 o 50 100 150
Time [sec) Time [sec)

Fig.4.4.a Measured signal of a sample containing 2.6 g 1-1 sucrose. b-d Distortions applied
toA

Table 4.3. Average error of distorted signals

Sucrose Cone. Distortion b Distortion c Distortion d

[grl] Error [%] Error [%] Error [%]

0.6 6.8 0.6 6.1

1.0 4.0 0.4 3.1
1.4 1.5 0.3 1.2
1.8 0.7 0.1 0.5
2.2 1.1 0.1 0.9
2.6 1.6 0.1 0.8
3.0 3.4 0.1 2.4
Ave. error 2.7 0.2 2.1

show if a fault occurred during the calibration. The test counts the local maximum
and minimum points contained in the actual calibration data, which are sorted by
standard solution concentrations. If no fault occurred, the calibration data should
include the minimum measurement of the lowest concentration standards, as well
as the maximum measurement of the highest concentration standards. It is expected
that the calibration measurements are monotonically climbing or falling. The char-
acteristic value returned is zero if the measurements of the calibration are mono-
l38 4 Accuracy and Reliability of Measured Data

tonic without any local maximum or minimum in the mid-range. Otherwise the
number of local extremes is returned. If the value is not equal to zero, then the
knowledge-based module uses this information to select the tests that will be further
performed. The advantage of this simple test is that no additional information about
the method or previous calibrations is necessary to identify a fault.
XP-CAFCA stores the results of each numerical procedure of all calibrations in a
file for further testing. The knowledge-based system checks the deviation of the ac-
tual calibration parameters from the previous ones during the validation of a new
calibration. It also performs a linear regression with all previous calibration para-
meters vs time to make predictions of the actual values. The next step is to inspect,
if the measured values for each standard fit into an interval set by experience, using
the preceding calibration model parameters. The measurements are marked as faulty
if some values are out of the chosen range. These marked values will be further
processed by the knowledge-based system. For example, it decides if a test on
swapped standards has to be performed.
The detection of outliers in a calibration data set is very important. For instance,
outliers can be caused by plugged tubes and valves, as well as by standard solutions
of incorrect concentration. XP-CAFCA uses three different methods (outlier tests A,
B, and C) to detect outliers included in the calibration data set.
If the number of multiple determinations is equal to or higher than three, then the
Outlier-A-Test is used. It calculates the standard deviation and the average standard
deviation of all measurements of a standard solution. A measurement is marked as
an outlier if its deviation is more than twice the average standard deviation. This
results in a 5% error probability, considering a normal distribution. An example of
an outlier included on a multiple determination is presented in Fig.4.5a.

i:L:=: I:k:::
° 0,2 0,4
a

0,6 0,8 1,2

.. 2
c
~ 1

° 0,2
0

0,4
b

O+---t---+---+---+--+--j
0,6 0,8 1,2
Glucose [gil] Glucose [gil]

~:k:=:c ·'I . + '

i:~
=2 0
~ 0
~ 1
0+--+---+---+---+--+---1 I
° 0,2 0,4 0,6 0,8 1,2 ° 0,2 0,4 0,6 0,8 1,2
Glucose [gil] Glucose [gil]

Fig.4.Sa-d. Examples of disturbed calibration data: a outlier of a multiple determination; b

outlier of multiple measurements of a standard; c swapped standard solutions; d calibration
invalid. (+=reliable measurements; O=faulty measurements)
4.2 Measurement Reliability 139

Outlier-B-Test is based on a cross-validation procedure. First, the calibration mod-

el with the standard deviation is calculated using all data. A single measurement is
selected and eliminated out of the data set, and a new calibration model is calculated
using the remaining data. The residual standard deviations of both models are F-
tested for a significant difference. For this, a test-value (TV) is calculated:
(nl - 2)SRsdl - (n2 - 2)SRsd2
TV = 2 (4.19)
SRsd2

TV is the test-value, nt, nz are the numbers of measurements, and SRsdl> SRsd2 are
the residual standard deviations of the calibration model using all measurements
and all measurements except one, respectively. The calculated TV is compared with
the corresponding value obtained from the tabulated F-Test values

P(f[ = l,fz = nz - 2, P =99 %) (4.20)

which represents the F value calculated by the Fisher-Test with f[ and fz degrees of
freedom and a significance level of 99%. TV is considered an outlier if its value is
greater than the corresponding tabulated value. If no outliers are present, no signifi-
cant difference will be detected. This procedure is repeated with all measurements.
This test can be performed even when single determinations of each standard are
carried out.
The Outlier-B-Test and Outlier-C-Test are similar. The main difference is that on
the Outlier-B-Test not only one measurement is eliminated but so are all the values
measured for a special standard. This test is used to detect outliers caused by faulty
standard solutions, for example. Figure 4.Sb shows a calibration data that includes
one outlier of this type.
XP-CAFCA calculates the confidence intervals for all standard concentrations ap-
plied to the actual calibration model as a further test for reliability, as described
above. The knowledge-based system checks if the measurements of a standard solu-
tion marked by the above-mentioned trend-analysis fits into the confidence interval
of another marked standard measurement. The results are used in the knowledge-
based module to decide if standard solutions have been swapped by the operator.
Figure 4.5c shows a calibration data with swapped standards.
XP-CAFCA performs weighted least-squares fits on first-order, second-order, and
third-order regression model, depending on the number of standards. A lack-of-fit
test is applied to all models. The residual standard deviations of all three models are
F-tested for a significant difference for this purpose. The lowest order regression
model is used, provided that the next higher order model does not provide a signif-
icantly improved fit. This criterion guarantees an optimal calibration model.

4.2.3
Knowledge-Based Module

The calculated characteristics of the numerical-statistical module are judged using

the knowledge-based module. The knowledge implemented in this module is repre-
sented in logical rules of the form IF AND THEN ELSE . The conditional part of
these rules compares the characteristics returned by the numerical-statistical mod-
ule with experience-based limits. These rules are used to decide whether a fault oc-
curred or not, and also to decide which rule will be performed next. The simplified
140 4 Accuracy and Reliability of Measured Data

Start validation

Number of multiple YES

Set of Rules for
determinations
Outlier-A-test
is equal or greater 3 ?

YES NO YES

Write final report

Call the attention of the operator

Fig. 4.6. Main decisions of the knowledge-based module for calibration analysis

flow diagram of the overall process is shown in Fig. 4.6. Every node of the network
represents a set of rules. Every set of rules is partitioned into subsets. Beginning
with the knowledge-based analysis the characteristics returned by the numerical-
statistical procedures are inspected. The subset rules are activated to validate or to
reject this lack of consistence in case an unusual deviation between calculated and
expected values occurs. The marked values as well as the calculated trends of the
trend-analysis are used for this and other considerations.
4.2 Measurement Reliability 141

If a fault is detected, XP-CAFCA generates a list of actions that depends on the

magnitude of the fault. The actions are different for the first and subsequent calibra-
tions because the first calibration is used as a reference to analyze new calibrations.
XP-CAFCA stores all faults found and the corresponding required actions during a
validation. Then it returns the highest priority action to CAFCA. This action can
only be generated during the first calibration and will cause CAFCA to repeat auto-
matically the calibration and inform the operator by means of a light signal. XP-
CAFCA decides to use the previous validated calibration for the evaluation of the
new measurements if a subsequent calibration is validated and a serious fault is
found. When faults susceptible to correction by the knowledge-based system are
detected, such as outliers or swapped standard solutions, XP-CAFCA performs these
corrections and informs the operator by a text message and a light signal. The actual
values are confirmed if they are not found faulty during the calibration performed by
the knowledge-based system.

4.2.4
Methodology of the Knowledge-Based System

The methodology on which the XP-CAFCA is based can be explained by an example.

Table 4.4 contains the calibration data obtained from the FIA system. The same data
is presented in Fig. 4.7, in which two faults are apparent. One of these is caused by
the operator, by mistakenly swapping two standard solutions. The other is caused by
a plugged valve and results in the outlier that corresponds to a glucose concentration
of 0.8 g tl. The system knows that a linear calibration model is required due to the
user selection of the method, and also due to previous calibrations obtained with the
FIA system and validated by XP-CAFCA.
The first step during the investigation is the Outlier-A-Test, which eliminates the
outlier that corresponds to 0.8 g tl standard solution. The trend-analysis finds that
the measurements of two standards (0.2 g I-I and 0.4 g 1-1) deviate significantly from
the previous calibrations. Therefore, XP-CAFCA decides to perform a test on
swapped standards. By these considerations, the knowledge-based system recognizes
that the measurements of both marked standards have been swapped. The last step
of the validation is the test for the optimum calibration model. This is performed
because a loss in enzyme activity can cause a shift of the model from first to second
order. XP-CAFCA determines that the fit provided by second or third order regres-
sions is not improved, finishes the investigation, and generates the action list saying
that the validated calibration is chosen for evaluation of further measurements to

Table 4.4. Measured values and calibration model parameters

Standard solution [g rl] Multiple determinations [units]

0.2 1.022 1.032 1.048

0.4 0.534 0.571 0.558
0.6 1.556 1.541 1.503
0.8 2.059 2.603 2.042
1.0 2.511 2.577 2.579

Function: y=O.1653+2.3618x
Rest.Std.Dev=0.4088
142 4 Accuracy and Reliability of Measured Data

.,2

..'"
c
2-
c
'"
iii 1
+
+
o+-----+-----+-----+-----+-----+---~

o 0,2 0,4 0,6 0,8 1,2

Glucose [gil]

Fig. 4.7. Calibration data with swapped standards and an outlier

Table 4.5. Validated values and calibration model parameters

Standard solution [grlj Multiple determinations [units j

0.2 0.534 0.571 0.558

0.4 1.022 1.032 1.048
0.6 1.556 1.541 1.503
0.8 2.059 2.042 outlier
1.0 2.511 2.577 2.579
Function: y=O.0398+ 2.5096x
Rest.Std.Dev=O.0122

CAFCA. The calibration data and the calibration model parameters obtained after
the validation are presented in Table 4.5. The final report of the validation is saved
to disk.
This example demonstrates that the reliability of the measurements of the process
analyzer has been significantly increased. The advantages of the knowledge-based
system are that it is as helpful as the human expert, except that it will always be
ready, attentive, and will perform at the same level. Using XP-CAFCA during several
measurements was very helpful. Some faults have been corrected by the system, but
still the operator must be called when some other faults occur. Although the building
of a real-time knowledge-based system requires much effort, the benefits of the auto-
mated supervision provide a positive balance.

4.3
Conclusions
The accuracy and reliability of measurements in bioprocess technique are discussed
in this contribution. The terms accuracy and precision are explained and common
examples in analytical chemistry are provided. The application of these concepts to
bioprocess engineering is not a simple task, due to the higher variability of living
systems. A method that enables one to compensate the complex sample matrix typi-
cal to bioprocesses is presented. However, not only the effects due to sample compo-
References 143

sition have to be considered, but also the effects of the sampling itself. The latter are
related to positioning of the sample module and sample preparation as well, and can
have tremendous influence on the reliability of measured data. The concentration of
the analyte should be determined by means of two independent measurement meth-
ods if a new analyzer system has to be validated. The results from this validation
have to be analyzed using statistical tests.
The way that the reliability of measurements can be guaranteed is shown by the
application of a knowledge-based system to the supervision and validation of a com-
plex process analyzer for glucose measurements performed during bioprocess mon-
itoring. The application of the knowledge-based system demonstrated its usefulness
to reduce problems presented during process automation. The description of the
entire analytical process in terms of models and heuristic rules allowed the applica-
tion of the knowledge-based system. In this manner, the reliability of the analyzer
system as well as the whole bioprocess can be significantly enhanced. The analyzer
operating under this system requires reduced supervision, and the operator is re-
lieved from repetitive work. The reliability of measurements can be improved not
only by the application of analyzer-related knowledge but also by the application of
knowledge related to the bioprocess itself. Balance equations can be tested and even
approximately applicable models should be used to test the consistency of the data.
Data are the ultimate basis for process analysis, supervision, and control. Therefore,
their reliability is of utmost importance. The supervision of other analyzers (such as
HPLC) and whole processes can be applied in the same manner as presented for the
FIA system. The expense of such a system can be justified by the improved perfor-
mance of the analyzer or the process.

References

l. Mattiasson B, Hakanson, H. (1993) Tibtech 11:136

2. Schiigerl K (1993) J Biotechnol 31:241
3. Weigel B, Hitzmann B, Kretzmer G, Schiigerl K, Huwig A, Gifforn F (1996) J Biotechnol
50:93
4. Miller JC, Miller IN (1993) Statistics for analytical chemistry. Ellis Horwood, Chichester
5. Massart DL, Vandeginste BGM, Deming SN, Michotte Y, Kaufman L (1988) Chemo-
metrics: a textbook. Elsevier, Amsterdam
6. Funk W, Dammann V, Donnevert G (1995) Quality assurance in analytical chemistry.
VCH Verlagsgesellschaft mbH, Weinheim
7. Ruzicka J, Hansen EH (1988) Flow injection analysis, 2nd edn. Wiley, New York
8. Valcarcel M, Luque de Castro MD (1987) Flow-injection analysis, principles and applica-
tions. Ellis Horwood, Chichester
9. Scheller F, Schmid RD (eds) (1992) Biosensors: fundamentals, technologies and applica-
tions. GBF Monographs, vol 17. VCH, Weinheim
10. Schiingarth K, Hitzmann B (1998) Anal Chim Acta 363:183
1l. Wiinsch G, Gansen M (1989) Fresenius Z Anal Chern 333:607
12. Schoenmakers P (1993) Knowledge-based systems in chemical analysis. In: Buydens LMC,
Schoenmakers PJ (eds) Intelligent software for chemical analysis. Elsevier, Amsterdam,
p13
13. Wolters R, van Opstal MAT, Kateman G (1990) Anal Chim Acta 133:65
14. Peris M, Maquieira A, Puchades R, Chirivella V, Ors R, Serrano J, Bonastre J (1993) Che-
mometrics Intell Lab Syst 21 :243
15. Brandt J, Hitzmann B (1994) Anal Chim Acta 291:29
144 4 Accuracy and Reliability of Measured Data

16. Dullau T, Schugerl K (1991) Process analysis and control by enzyme-FIA-systems. In:
Schmid RD (ed) Flow injection analysis (PIA) based on enzyme or antibodies. GBF
Monographs, vol 14. VCH, Weinheim, p 27
17. Hitzmann B, Lohn A, Reinecke M, Schulze B, Scheper T (1995) Anal Chim Acta 313:55
18. Press WH, Flannery BP, Teutolsky SA, Vetterling WT (1987) Numerical recipes - the art of
scientific computing. Cambridge University Press, Cambridge
19. Van Leeuwen H (1993) Developing expert systems. In: Buydens LMC, Schoenmakers PJ
(eds) Intelligent software for chemical analysis. Elsevier, Amsterdam, p 79
20. Chen D, Zeng Y (1990) Anal Chim Acta 235:337
21. Wu X, Bellgardt KH (1995) Anal Chim Acta 313:161
22. Szostek B, Trojanowicz M (1992) Anal Chim Acta 261:509
23. Trojanowicz M, Szostek B (1992) Anal Chim Acta 261:521
24. Szostek B, Trojanowicz M (1994) Anal Chim Acta 22:221
25. Bos M, Hoogendam E (1992) Anal Chim Acta 267:73
26. Hartnett M, Diamond D, Barker PG (1993) Analyst 118:347
27. Hitzmann B, Kullick T (1994) Anal Chim Acta 294:243
28. Hitzmann B, Ritzka A, Ulber R, SchOngarth K, Broxtermann 0 (1998) J Biotechnology
65:15
29. Hitzmann B, Lohn A, Arndt M, Ulber R, Muller C (1997) Anal Chim Acta 348:161
5 Bioprocess Control
D. Dochain, M. Perrier

5.1
Introduction

Industrial-scale biotechnological processes have progressed vigorously over the last

decades. Generally speaking, the problems arising from the implementation of these
processes are similar to those of more classical industrial processes and the need for
monitoring systems and automatic control in order to optimize production effi-
ciency, improve products quality, or detect disturbances in process operation is ob-
vious. Nevertheless, automatic control of industrial biotechnological processes is
clearly developing very slowly. There are two main reasons for this:
1 The internal working and dynamics of these processes are as yet badly grasped
and many problems of methodology in modeling remain to be solved. It is diffi-
cult to develop models taking into account the numerous factors which can influ-
ence the specific bacterial growth rate. The modeling effort is often tedious and
requires a great number of experiments before producing a reliable model. Repro-
ducibility of experiments is often uncertain due to the difficulty in obtaining the
same environmental conditions. Moreover, as these processes involve living organ-
isms, their dynamic behavior is strongly nonlinear and non-stationary. Model
parameters cannot remain constant over a long period - they will vary, e.g., due
to metabolic variations of biomass or to random and unobservable physiological
or genetic modifications. It should also be noted that the lack of accuracy of the
measurements often leads to identifiability problems.
2 Another essential difficulty lies in the absence, in most cases, of cheap and reliable
instrumentation suited to real-time monitoring. To date, the market offers very
few sensors capable of providing reliable on-line measurements of the biological
and biochemical parameters required to implement high performance automatic
control strategies. The main variables, i.e., biomass, substrate, and synthesis pro-
duct concentrations, generally need determining through laboratory analyses. The
cost and duration of the analyses obviously limit the frequency of the measure-
ments.

Figure 5.1 shows a schematic view of a computer controlled bioreactor. In the illu-
strated situation, the influent flow rate is modulated via a control algorithm imple-
mented in the computer. As will be explained below, the control algorithm combines
different informations about the process provided by on-line measurements and some
knowledge about the process dynamics. The on-line measurements can sometimes
undertake some mathematical treatment by software sensors in order to provide ex-
146 5 Bioprocess Control

Control
action

bioreactor Sensors : analysers

and measurement
devices

Fig. 5.1. Schematic view of a computer controlled bioreactor

tra information about some key process parameters (like specific growth rates) and
unmeasured variables.
The chapter is organized as follows. Section 5.2 is dedicated to the basic concepts
of automatic control of bioprocesses with a particular emphasis on the notions of
stability and disturbances. Section 5.3 introduces the basic ingredients of automatic
control (dynamical model, feedback, proportional, integral, and feedforward ac-
tions), and briefly compares linear control vs nonlinear control, and adaptive control
vs non adaptive control. Finally Sect. 5.4 is concerned with the design of adaptive
linearizing control schemes.

5.2
Bioprocess Control: Basic Concepts

The objective of automatic control when applied to any process (including biopro-
cesses) is to run and maintain the process in stable optimal operating conditions in
spite of disturbances.
Let us first give some more insights about the concepts that are covered by the
words stable, optimal, and disturbances. And let us consider the following simple
example to illustrate these concepts: a simple microbial growth reaction in a stirred
tank reactor. The microbial growth reaction is characterized by the following reac-
tion scheme:

S-+X (5.1)
5.2 Bioprocess Control: Basic Concepts 147

with S the limiting substrate and X the biomass. Mass balances for both components
S and X leads to the following set of dynamical equations:
dS 1
- = -DS - -/LX + DS in (5.2)
dt Yx /s

dX
-= -DX+/LX (5.3)
dt
where D is the dilution rate (h- l ) (i.e., the ratio of the influent flow rate Fin (lIh) over
the volume V (1) occupied by the reacting medium in the reactor), YX/s is the yield
coefficient, /1 is the specific growth rate (h- l ), and Sin is the influent substrate con-
centration (gil). Note that:
1. The above model is a dynamical model in order to capture the time evolution of
the process: this is an essential ingredient of automatic control (we shall go back to
this in Sect. 5.3).
2. The model is valid whatever the operation mode is: continuous (then the reactor is
sometimes called a chemostat or a CSTR (Continuous Stirred Tank Reactor»,
batch (then D is equal to zero), fed-batch (then the volume is not constant and
its time variation is given by the equation ~~ = Fin.
3. We use the same notation for the components (in the reaction scheme at Eq. 5.1)
and their concentration (in Eqs. 5.2 and 5.3).
Let us now consider that the specific growth rate here is of the Haldane type:

(5.4)

with Ks the Monod constant, KJ the inhibitor constant, and /L' a constant, which is
indeed connected to the maximum specific growth rate /Lmax as follows:

/L = /Lmax(1 +2 jf;
s
-)
KJ
(5.5)

Let us also consider that the objective is to control the process by using the dilu-
tion rate D (or more precisely, the influent flow rate Fin) or the influent substrate
concentration Sin as the control action.

5.2.1
Disturbances

Let us illustrate the notion of disturbances with the above example. If the dynamical
model (Eqs. 5.2-5.4) is assumed to be perfect, and if the dilution rate D is the only
manipulated input, then the influent substrate concentration Sin is a disturbance:
variations of Sin will disturb the process by inducing variations in the time evolution
of the variables Sand X.
Assume now that the dynamical model is not perfect, e.g., that some of the para-
meters /1 *, K" and KJ of the Haldane model (Eq. 5.4) are changing with time (e.g.,
due to micro-organisms adaptation): these variations will also be disturbances with
respect to the considered process model. Another similar disturbance may come
from the presence of a side-reaction (e.g., maintenance or micro-organisms death)
that has been neglected in the above dynamical model but which may not be not
148 5 Bioprocess Control

negligible under some operating conditions. Then, in case of maintenance for in-
stance, the true dynamical model of the process is described by Eq. (5.3) and the
following modified mass balance equation for S:

dS 1
- = -DS--IIX - k X +DS' (5.6)
dt yx/st"" m In

where k m is the maintenance coefficient.

Finally, typical disturbances may be due to process failures (e.g., actuator, sensor,
or even stirring system). Even if those disturbances are inherently different, it is
important to consider all the disturbances that are likely to take place during the
process operation when analyzing and testing the performances of a designed con-
troller that has to be implemented on a real bioprocess.

5.2.2
Stability

The notion of stability has been largely considered and studied in dynamical systems
and automatic control, and has led to different mathematical definitions and tests
(see, e.g., [17]). Here we shall concentrate on Lyapunov stability. Qualitatively the
notion of stability can be roughly explained as follows. Let us consider that the pro-
cess is at some equilibrium point (or steady state). The equilibrium point will be
stable if the process does not go far from this state except for small deviations. It
will be unstable if the variables of the process moves away from the equilibrium
point and if their variations becomes larger and larger as time increases. Stability
is obviously of vital concern in automatic control: it is essential to design controllers
that keep processes in stable conditions on the one hand, and that are capable of
stabilizing unstable processes on the other.
Let us illustrate the concept of stability on the above microbial growth example,
and how to test it.
The Haldane model is represented in Fig. 5.2. What we shall show now is that
values of the substrate c~entration on the left hand side of the maximum specific
growth rate JL* /(1 + 2V~) correspond to stable equilibrium points, while those on
the right hand side correspond to unstable ones.

Jl max = 0,3

I
I
I
I
I
I
I

: S2

0,0 it
+-%,---t"=--.----,-~.%....----~..--~__r_~___i~
o 3
Substrate Concentration S
Fig. 5.2. The Haldane model
5.2 Bioprocess Control: Basic Concepts 149

5.2.2.1
Equilibrium Points

An equilibrium point is, by definition, a constant state (denoted here [S, Xl) which
satisfies the equation of the dynamical model, i.e., the following algebraic equations
in our example:
dS -- 1 - - --
- = 0 =} -DS - -JL(S)X + DSin = 0 (5.7)
dt Yx /s

dX ----
dt = 0 =} - DX + JL( S)X = 0 (5.8)

for given constant values of D and Sin- If we consider the Haldane model, it is
straightforward to check that there are three possible equilibrium points:
l. Operational equilibrium point number 1: SI <::: ylK2KI' XI = Yx/s(Sin - Sd
2. Operational equilibrium point number 2: S2 > ylKsKJ,X2 = YX/S(Sin - S2)
3. Wash-out equilibrium point: X 3=0, S3=Sin.

The wash-out may occur for any value of D and Sin- It is also the only possible equi-
librium point for D> JLmax. It is obviously an undesirable steady-state. The values of S
of the first two equilibrium points corresponds indeed to the two solutions of the
second order algebraic equation obtained by introducing the Haldane equation
(Eq. 5.4) into the equilibrium point equation of X (Eq. 5.8):

D S2 + (D - JL*)S + DKs = 0 (5.9)

K]

5.2.2.2
Stability Analysis

For the stability analysis, let us consider the linear approximation of the model
(Eqs. 5.2-5.4) around the equilibrium points (i.e., the linearized tangent model).
Whatever the input u (D, Sin> or both), the linearized tangent model around each
operational equilibrium point is written as follows:
dx
dt = Ax+Bu (5.10)

where x is the vector of the deviations of S and X from the equilibrium values:

x- _[S-S]
X-X
- (5.11)

and the matrix A is given by the following expression:

(5.12 )
150 5 Bioprocess Control

with
K)
IL*X(KS - KI
n= (5.13)
(K S + S+~)2
KI

The stability of the equilibrium points is determined by the eigenvalues of the

matrix A: the equilibrium point will be stable if the real parts of all the eigenvalues
are negative; it will be unstable if one eigenvalue has a positive real part. In our
example, the eigenvalues are equal to (-D) and (- i-
[2). If the first eigenvalue is
always negative, the second one is negative for the t/;st equilibrium point (51) Xl)
and is positive for the second one (52, X2 ). This means that (51) Xl) is a stable equi-
librium point, and that (52) X2 ) is an unstable one.

5.2.3
Regulation vs Tracking

Two typical situations are usually considered in automatic control: constant set point
control (regulation) and trajectory tracking. Regulation is typical of the control of
continuous reactors, where the objective is usually to maintain the process in an a
priori chosen steady state despite the disturbances. But in many instances like in
batch and fedbatch reactors but also for start-up of reactors and in case of produc-
tion grade changes, regulation is not appropriate: the objective is then to drive the
process from some initial state to some final desired state by following some prede-
termined trajectory. In the above simple microbial growth example, if we consider a
fedbatch reactor and if the objective is to optimize the production of biomass, then
there is an optimal trajectory which is an "exponential" profile for X: the controller
has then to be designed so as to maintain the process as close as possible to the
optimal profile.

5.3
Bioprocess Control: Basic Ingredients

5.3.1
Dynamical Model

The first essential basic ingredient for the design of automatic control algorithms is a
dynamical model of the process. The need for dynamical models in the control de-
sign is motivated by the intrinsic need to handle transients, either for transferring
the process from one state to another, from one set point to another, for following a
desired trajectory, or for rejecting or reducing the effect of disturbances. In the latter,
disturbances influence the dynamical behavior of the process: good knowledge of
the process dynamics, and in particular of the dynamical effect of the disturbance,
will be helpful to select the approriate strategy to reduce and possibly reject their
negative influence on the process operation.
The dynamical models to be considered for control design can take different
forms. The preceding sections suggest that the most obvious model forms are the
mass balance models. But many other forms are possible. These include linear mod-
els deduced, e.g., by linearization of the mass balance models, or by identification
from input-output data (e.g., [8]) (such models are also called black-box models).
5.3 Bioprocess Control: Basic Ingredients 151

Fig. 5.3. Feedback loop for the control of S in a bioprocess

The input-output models may be nonlinear; neural network models are presently a
quite popular version of nonlinear black box models (e.g., [9]). A possibly attractive
form of dynamical models of bioprocesses are hybrid models, which are an inter-
mediate form between mass balance and neural network models (e.g., [6]). In the
following, mainly for reasons of conciseness and pedagogical reasons, we concen-
trate on mass balance models.

5.3.2
Feedback

Another essential basic ingredient of automatic control is the introduction of a feed-

back loop: on-line (real-time) information about the process (typically on-line mea-
surement of the variable to be controlled) is provided to the controller which then
provides the control action to the process. This concept of feedback loop is illu-
strated in Fig. 5.3.

5.3.3
Proportional Action

Most feedback control algorithms (except very simple controllers like on-off control-
lers) modulate the amplitude of the control action proportionally to the control error,
i.e., the difference between the desired value of the controlled variable and its mea-
sured value: this is known as the proportional action.

5.3.4
Integral Action

The proportional action is usually not sufficient to guarantee a control error equal to
zero in steady-state: this motivates the introduction in the control scheme of an in-
tegral action, i.e., a term proportional to the integral of the control error.

5.3.5
Feedforward Action

If the effects of a disturbance on the bioprocess dynamics are known, the controller
may also be designed by incorporating a feedforward action in order to anticipate
the possibly negative effects of the disturbance. We shall illustrate below how to in-
corporate feedforward action when the influent substrate concentration Sin is a mea-
sured disturbance.
152 5 Bioprocess Control

5.3.6
Linear Control vs Nonlinear Control

The controller can be designed on the basis either of a linear model, or of a non-
linear model of the process. As above, let us illustrate the design with a simple ex-
ample: the control of the substrate concentration S by acting on the dilution rate D
(or more precisely the influent flow rate Fin) in a stirred tank reactor with a simple
microbial growth reaction.
Let us start with the first option: linear control. The control design is based on a
linear model of the process, e.g., the linearized tangent model (Eqs. 5.10 and 5.11)
around some steady state (S, X, (D, Sin)' If we consider the control of S at a pre-
scribed value S* (e.g., equal to S) a linear PI (Proportional Integral) regulator will
then have typically the following form (see e.g., [15]):

D=D+Kp[(S*-S)+~ t(S*-S(T))dT],Kp>O,Ti>O (5.14)

TiJo
where Kp is the controller gain, and Ti is the integral time or reset time.
Indeed a similar PI regulator can be computed on the basis of a nonlinear model
of the process, e.g., the mass balance equation of the substrate concentration S
(Eq. 5.2). The control algorithm will then be written as follows:
)'1(S* - S) +).2 J~ (S* - S(T))dT - /-LX/Yx / s
D = ,).1 > 0,).2 > 0 (5.15)
Sin - S
The above controller equation could have been obtained by combining the mass
balance equation (Eq. 5.2) with the following desired dosed-loop second-order dy-
namical equation:
d(S* - S)
dt
*
= -).I(S - S) -).2
t *
Jo (S - S(T))dT (5.16)

for S· constane. In simple terms, Eq. (5.16) imposes that, if at some time instant the
substrate concentration S is different from its desired value S*, the controller will
force S to converge to S* with a rate determined by the design parameters ).1 and ).2'
Let us present at this point the advantages and drawbacks of the approaches in-
troduced above. Before doing so, it is important to note that the second algorithm
(Eq. 5.15) is more sophisticated than the first one (Eq. 5.14). In particular the sec-
ond controller requires the knowledge (on-line measurement) of the biomass con-
centration X and of the influent substrate concentration Sim and the knowledge of
the specific growth rate model. When comparing both approaches, it is important to
have these differences in mind.

5.3.6.1
Linear Control

The linear controller, if it is correctly tuned (see, e.g., [15] for tuning rules), will
most probably do a good job if it is used for regulation around a steady-state. It also

Indeed by differentiating Eq. (5.16) once with respect to time, one obtains the usual second
order linear equation ~:~ + )1) ~ + A1X = 0 with x=S*-S
5.3 Bioprocess Control: Basic Ingredients 153

1.2

1 ---

0.8
~ 0.6
(/J

0.4

0.2

10 15 20
time (h)

Fig. 5.4. Numerical simulation of a linear PI control of a fedbatch reactor

has the advantage of being easily implemented via standard industrial PI regulators.
However, its performance (in particular its transient performance) might be degrad-
ing if the process is moving away from the steady-state for which it has been cali-
brated and tuned: this happens typically for start-up and grade changes in CSTRs,
and this might even be more critical for (fed- )batch reactors in which the process
state typically follows large variations (just think of the "exponential" growth of the
biomass in this type of reactor!). This is illustrated in Fig. 5.4 where the PI regula-
tion of S has been numerically simulated with the following model parameter values
for Eqs. (5.2)-(5.4) and under the following conditions:
p, * = 5 h-l, Ks = 10 gtl, K[= 0.1 gtl, Yx /s = 1

S(O) = 0 gtl, (X(O) = 0.1 gtl, V(O) = 101, V(tj) = 20 1, Sin = 10 gtl

S* = 19t1, Kp=2, 7[=0.5

The values of the design parameters of the PI have been chosen so as to corre-
spond to a closed loop dynamics close to the mean open loop dynamics. We note
that the controller gain is too high at the beginning (and it induces large variations
of the controlled output S), and then too small (the controller is not able to maintain
the controlled output S to its desired value during the second part of the "exponen-
tial" growth phase).

5.3.6.2
Nonlinear Control

The main advantage of the nonlinear controllers is directly related to the above-men-
tioned drawback of the linear ones: if correctly designed and tuned, it will be able to
handle a larger spectrum of the operating domain of the bioprocess. In the above
example, the main source of nonlinearity is the term p,X/YXIS ' The presence of this
term in the controller will be particularly important in the control of fedbatch reac-
tors: as has been pointed out in [1], the integration of a term which characterizes the
"exponential" biomass growth is a key factor for the efficiency of the control action.
The other additional term Sin-S also plays an important role: if Sin is a disturbance in
the context of the control of the bioprocess, then this term is indeed a feedforward
term that allows one to anticipate the effect of variations of Sin on the performance of
the closed-loop system (process+controller).
154 5 Bioprocess Control

However, first the implementation of the control algorithm will require a computer
instead of a standard industrial PI (but can this still be considered as a drawback
today?). Secondly (as for any controller) the performance will depend greatly on the
quality and reliability of the model on which the design has been based. In other
words, in the above example this means that we assume that the hydrodynamics in
the reactor are fairly represented by completely mixed conditions, and that the sim-
ple growth reaction is largely dominant with regard to other possible side reactions.
Besides, the implementation as such of the controller (Eq. 5.15) requires the values of
X and also of Sin' In many cases one may expect that if S is measured on-line it
should not be too difficult to measure Sin too (in many instances you do not even
need to measure it since the influent substrate concentration is known by user's
choice). But the situation is usually different for the biomass concentration: very
often X is not measured on-line, or let us say it differently, S and X are often not
available at the same time for on-line measurement. Finally, the nonlinear controller
(Eq.5.15) relies on the knowledge of (the model of) the specific growth rate /1>: as
suggested previously, this is a major source of uncertainty in bioprocess models.

5.3.7
Adaptive Control vs Non-Adaptive Control

One possible way to handle the above difficulties (lack of on-line measurements,
uncertainties on the kinetics models) is to use adaptive (linearizing) control (see
[2] and Sect. 5.4). In the above example this means that the control law will be im-
plemented by considering on-line estimates of X and /1>. An on-line estimate of the
biomass concentration can be provided by an asymptotic observer (for a systematic
design, see [2] and Sect. 5.4). In our example, it will take the following form. We first
define an auxiliary variable Z:
S
Z=S+~ (5.17)
YXjS

The dynamics of Z are readily derived from the mass balance equations of X and S:
dZ
- =DS· -DZ
dt In
(5.18)

The on-line estimate X of X is then given by computing Z from Eq. (5.18) and by
rewriting Eq. (5.17):

(5.19)
or equivalently, one can replace the term X/YX1S in the controller equation by Z-S.
Let us now consider the on-line estimation of /1>. Different methods can be used to
give an on-line estimate {1 of /1>. One possible estimator is the observer-based esti-
mator (see [2]), which specializes in our example as follows:

dS
dt = -(1(Z - S) + DSin - DS + /1 (S -
'
S), /1 > 0 (5.20)

(5.21 )
5.3 Bioprocess Control: Basic Ingredients 155

S is an "estimate" of S whose role is to drive the estimation of the unknown value of

the specific growth rate IL: this is achieved by comparing S to the measured value S
(this is the role of the correction terms S-S in Eqs. (5.20) and (5.21). YI and Y2 are
design parameters that have to be calibrated to obtain the best on-line estimates (see
[2)).
Since the parameter estimation introduces an integral action in the control loop
([2, 11]), the term 12 J~ (S* - S( T)) dT in the control law (Eq. 5.15) is not necessary
any more. Then in our example, the adaptive linearizing controller is given by the
following equations:
)'1(S* - S) - {1(Z - S)
D= ,)'1>0 (5.22)
Sin - S
with Z and {1 given by Eqs. (5.18), (5.20), and (5.21), respectively.
Different applications of adaptive linearizing control to bioprocesses have been
described, e.g., in [3, 12-14, 16]. One of the key feature of the adaptive linearizing
can be illustrated on the basis of the above example: it allows one to incorporate the
well-known characteristics of the process dynamics, while keeping the usual features
of classical controllers:
- proportional action: via the term Al (S* -S)
- integral action: via the adaptation mechanism (Eqs. 5.20 and 5.21)
- feedforward action: via the presence of Sin' The controller is capable of anticipat-
ing the effect of a variation of the influent substrate concentration: for instance, an
increase of Sin results in a decrease of the control input D, inversely proportional
to this increase.

Besides, as for the non-adaptive nonlinear PI, the controller contains a state "esti-
mate" via the term JLXIYxls (or more precisely {1(Z-S). This is particularly important
in fedbatch fermentation since it gives the exponential biomass growth term, and
also in continuous operation since it provides a means to modulate the control ac-
tion in connection with the level of biomass (e.g., if the process is in good working
conditions, in particular if the biomass concentration is high, it is possible to treat
high amounts of substrate, since the control action D is proportional to X).
Finally, note that, beside the integral action, the estimation of "physical" para-
meters (here JL) has the further advantage of giving useful information that can be
used for monitoring the process, and possibly also for analyzing the internal work-
ing of the process. Figure 5.5 illustrates the performance of the adaptive linearizing
controller in numerical simulation. The same model parameters as in Fig. 5.4 have
been used. The following design parameters and initial values have been considered:
2
S* = 19lf, CI = 2, II = 5, 12 = 4(Z'~ S) (5.23)

{1(0) =0.1 h- I , S(O) =S(O), Z(O) =S(O) + O.5X(O), S(O) =0.5 gil (5.24)

The initial value of S has been set to 0.5 gil: this gives the opportunity to see that
the controller has no problem to force S to converge to its desired value (see
Fig.5.5a). Figure 5.5b,c exhibits the exponential profile of the biomass X, and the
control input Fin' Finally Fig. 5.5d gives the on-line estimation of the specific growth
rate JL.
156 5 Bioprocess Control

6r-------~--~------,
S' .1 gil
1 5

0.8
( a
4
b

~0.6 ~3
rn x
0.4

0.2 I
I~
0 °OC=~~5--~10----1-5---2~0--~
0 5 10 15 20

2.5 0.5 r-------~--~-----_,

- : estimated Ii
2 c 0.4
- - : sirrulated Ii
d

:21.5 _-0.3
~
S ____ _ =-=_=-c_~_--"

u..:= 1 ::1.0.2

0.5 0.1

0 °0L--~5-~'0----'-5-~~~
0 5 10 15 20
lime (h) time (h)

Fig. S.Sa-d. Adaptive linearizing controller of a fedbatch reactor

5.3.8
Other Approaches

The (adaptive) nonlinear controller presented in the preceding section belongs in-
deed to a large class of control strategies, often labelled model-based control. This
includes control approaches like model predictive control in which constraints (like
input constraints) may be handled explicitly in the design (see, e.g., [15]). We have
restricted the presentation to mass balance models, but obviously any dynamical
model of the bioprocess (e.g., hybrid model), as long as it is reliable, can be consid-
ered in the control design.

5.4
Adaptive Linearizing Control of Bioprocesses

In this section we shall concentrate on the design of adaptive linearizing controllers

on a systematic basis (via the General Dynamical Model introduced below). The
motivation for this choice of control approach has already been suggested above,
and can be further justified. As has already been said, the dynamics of bioprocesses
are usually described by material balance equations. If the hydrodynamics and the
underlying reaction mechanisms of the bioprocess are often rather well character-
ized, the kinetics of these reactions are usually badly characterized or defined.
Symptomatically there exist many (heuristic) models available in the literature for
the specific growth rate (more than 60, see, e.g., [2]). Beside the difficulty of choos-
ing an appropriate model, the parameters of the selected model are often practically
unidentifiable, i.e., they cannot be given unique values from the available experimen-
5.4 Adaptive Linearizing Control of Bioprocesses 157

tal data (see, e.g., [2, 7]). The adaptive linearizing control approach has the basic
following features: it incorporates the well-known features about the process dy-
namics (basically, the reaction network and the material balances) in the control
algorithms which are moreover capable of dealing with the process uncertainty (in
particular of the reaction kinetics) via an adaptation scheme.

5.4.1
General Dynamical Model

A biotechnological process can be defined as a set of M biochemical reactions invol-

ving N components. The dynamical model of a bioprocess in a stirred tank reactor
can be deduced from mass balance considerations and written in the following com-
pact form:
d~
-=-D~+Kr+F-Q (5.25)
dt
where ~ is the vector of the bioprocess component (gIL) (dim(~)=N), D is the dilu-
tion rate (h- l ), K is the yield coefficient matrix (dim(K)=NxM), r is the reaction rate
vector (g/(L.h)) (dim (r)=M), F (g/(L.h)) is the feed rate vector, and Q (g/(L.h)) the
gaseous outflow rate vector (dim(F)=dim( Q)=N). The model equation (Eq. 5.25) has
been called the General Dynamical Model for stirred tank bioreactors (see [2]). As
already mentioned in Sect. 5.2, note that the model covers the different types of op-
erating conditions: batch, fed-batch (as in example 2 below) or continuous reactors
(as in example 1 below).

5.4.1.1
Example 1: Anaerobic Digestion

Four metabolic paths can be identified in anaerobic digestion [10]: two for acidogen-
esis and two for methanization. In the first acidogenic path (Path 1), glucose is de-
composed into fatty volatile acids (acetate, propionate), hydrogen, and inorganic car-
bon by acidogenic bacteria. In the second acidogenic path (Path 2), OHPA (Obligate
Hydrogen Producing Acetogens) decompose propionate into acetate, hydrogen, and
inorganic carbon. In a first methanization path (Path 3), acetate is transformed into
methane and inorganic carbon by acetoclastic methanogenic bacteria, while in the
second methanization path (Path 4), hydrogen combines with inorganic carbon to
produce methane under the action of hydrogenophilic methanogenic bacteria. The
process can then be described by the following reaction network:
(5.26)

(5.27)

(5.28)

(5.29)
where 51> 52, 53, 54, 55' Xl, X 2 , X 3 , X 4 , and PI are respectively glucose, propionate,
acetate, hydrogen, inorganic carbon, acidogenic bacteria, OHPA, acetoclastic metha-
no genic bacteria, hydrogenophilic methanogenic bacteria, and methane. The dyna-
158 5 Bioprocess Control

mical model of the anaerobic digestion process (N =10, M=4) in a stirred tank reac-
tor can be described within the above formalism (Eq.5.25) by using the following
definitions:

XI 0 0 0 0
0
SI -k21 0 0 0 0
DS in
X2 0 0 0 0
0
S2 k41 -k42 0 0 0 0
X3 0 0 0 0 0
~= , K= , F= ,Q=
53 k61 k62 -k63 0 0 0 (5.30)
X4 0 0 0 1 0 0
S4 kSI kS2 0 - kS4 0 QI
SS 0 Qz
k91 k92 k93 -k94
0 Q3
PI 0 0 k03 k04

(5.31)

where /1>1' /1>2, /1>3' /1>4 are the specific growth rates (h- I ) of the reactions at Eqs. (5.26)-
(5.29), respectively, and Sin> QI, Q2, and Q3 represent respectively the influent glucose
concentration and the gaseous outflow rates of Hz, CO 2 , and CH4 •

5.4.1.2
Example 2: Animal Cell Culture

Let us consider one animal cell culture process: a human embryo kidney (HEK-293)
cell culture [16]. It is characterized by the following reaction network:

S+C->X+G (5.32)

S->X+L (5.33)

where S, C, X, G, and L represent the glucose, dissolved oxygen, cells, carbon dioxide
and lactate, respectively. Reactions at Eqs. (5.32) and (5.33) are an oxidation (respira-
tion) reaction on glucose, and a glycolysis (fermentation) on glucose. The dynamics
of the process in a stirred tank reactor are given in the matrix format (Eq.5.25) by
the following vectors and matrices:

o
DSinj
Qin o
o ,Q= o (5.34)
o QI
o o
5.4 Adaptive Linearizing Control of Bioprocesses 159

kj 0=1-9) are yield coefficients, Sin is the influent glucose concentration (gil), Qin is
the oxygen feedrate (gil/h), QJ is the CO2 outflow rate (gil/h), and Iti (i=R and F) are
the specific growth rates (l/h) associated to each growth reaction.

5.4.2
Model Reduction

The above anaerobic digestion example illustrates that the dynamical model may be
fairly complex with a large number of differential equations. But there are many
practical applications where a simplified reduced order model is sufficient from an
engineering viewpoint. Model simplification can be achieved by using singular per-
turbation, which is a technique for transforming a set of n+m differential equations
into a set of n differential equations and a set of m algebraic equations. It is suitable
when neglecting the dynamics of substrates and of products with low solubility in
the liquid phase. The method will be illustrated with one specific example (low so-
lubility product) before stating the general rule for order reduction.

5.4.2.1
Singular Perturbation Technique for Low Solubility Products

Let us consider a biochemical reaction with a volatile product P which gives off in
gaseous form and has low solubility in the liquid phase. The dynamical equation of P
is as follows:
dP
-= kr-DP- Q (5.35)
dt
The consistency of this model requires that the product concentration P be pro-
portional to a saturation concentration representative of the product solubility,
which is expressed as P=ITPsat where Psat is the saturation concentration which is
constant in a stable physico-chemical environment. The model (Eq. 5.35) is rewritten
in the standard singular perturbation form, with £=Psat :
dIT
e - = kr - eDIT - Q (5.36)
dt
If the solubility is very low, we obtain a reduced order model by setting £=0 and
replacing the differential equation (Eq.5.36) by the algebraic one:
Q=kr (5.37)

5.4.2.2
A General Rule for Order Reduction

The above example shows that the rule for model simplification is actually very sim-
ple and that an explicit singular perturbation analysis is not really needed. Consider
that, for some i, the dynamics of the component ~i are to be neglected. The dynamics
of ~i are described by Eq. (5.25):
d~i
dt = -D~i + Ki r + Fi - Qi (5.38)
160 5 Bioprocess Control

where Ki is the row of K corresponding to the component I;i' The simplification is

achieved by setting I;i and dl;;ldt to zero, Le., by replacing the differential Eq. (5.38)
by the following algebraic equation:
Kir=-Fi+ Qi (5.39)
It has been shown that the above model order reduction rule is not only valid for
low solubility products but also for bioprocesses with fast and slow reactions. Then
the above order reduction rule (Eq.5.39) applies to substrates of fast reactions (as
long as they intervene only in fast reactions).

5.4.2.3
Example 1: Anaerobic Digestion

Let us see how to apply the above model order reduction rule to the anaerobic diges-
tion. First of all, it is well-known that methane is a low solubility product. Therefore
the above procedure applies. Furthermore, assume that for instance the first acido-
genic path (reaction at Eq.5.26) is limiting, i.e., that the last three reactions
(Eqs.5.27-5.29) are fast and the reaction at Eq. (5.26) is slow. We can then apply
the model order reduction rule to the propionate concentration 52, the acetate con-
centration 53' the hydrogen concentration 54, and the dissolved methane concentra-
tion Pl' By setting their values and their time derivatives to zero, we reduce their
differential equations to a set of algebraic equations. Let us further consider that
the gaseous hydrogen outflow rate Ql is negligible. If one considers the expressions
of the reaction rate rl which can be drawn from these algebraic equations, and in-
troduces them in the dynamical equation of the glucose concentration 51> this one
can be rewritten as follows:
d5 1
-
dt
= -D5 J - klQ3 + DS· In
(5.40)

where kl is defined as follows:

k 1_- k21k42 k63 kS4

(5.41)
k03 kS4(k6Jk42 + k62 k41 ) + k04 k63 (kSI k42 + kS2k4J)
Note that the coefficient kl is a nonlinear combination of the yield coefficients kij.

5.4.3
(ontrol Design

Let us now concentrate on the design of adaptive linearizing controllers. It is based

on the dynamical material balances of the process (Eq.5.25). More specifically, the
control design will be based on the dynamical equation that relates the controlled
variable and the control variable. If we denote the controlled variable (typically,
some process component concentration(s» by y and the control variable (typically
flow rate(s) or influent substrate concentration(s» by u, then the dynamical equa-
tion between y and u is given from Eq. (5.25) (possibly after model reduction) by the
following expresion:

d; = f(~) + g(~)u (5.42)

5.4 Adaptive Linearizing Control of Bioprocesses 161

where f and g are functions of the process component concentration. In the follow-
ing, we shall illustrate the control design with the animal cell culture. In this exam-
ple, if the control objective is to control the glucose concentration S (y=S) by acting
on the influent flow rate Fin (U=Fin)' then Eq. (5.42) before model reduction specia-
lizes as follows:
dS Fin Fin
dt = -kl/.LRX - k4/.LFX - -yS + -ySin (5.43)

In our example, we consider that CO2 is a low solubility product. Then the dyna-
mical equation (Eq.5.43) becomes:

(5.44)

i.e.,
k\ Sin - S
f = - k3 Q\ - k4/.LFX,g = - V - (5.45)

An interesting feature of the above dynamical equation is its independence with

respect to the specific growth rate /.LR: the largely uncertain or unknown specific
growth rate /.LR has been replaced by a variable usually easy to measure: the gaseous
outflow rate of CO2 , Q\.
The design of the adaptive linearizing controllers is also based on the following
(realistic) assumptions:
Ai. Some of the (combinations of) the yield coefficients are known
A2. The kinetics are unknown
A3. The influent flow rates and substrate concentration, and the gaseous outflow
rates are known (either by measurement or by user's choice)

5.4.3.1
The Monitoring Tool 7: An Asymptotic Observer

The key result of this section is the use of a state transformation by which part of the
dynamical model (Eq. 5.25) becomes independent of the reaction kinetics r (see [2]).
This transformation will playa very important role since it will allow us in our ex-
ample to implement a software sensor called asymptotic observer for the (unmea-
sured) biomass based on the general dynamical model independently of the unmea-
sured variables. In other instances (see [4, 5]), it allows to implement a controller
independent of unmeasured variables (see also Sect. 5.3.7).
The transformation is defined as follows. Let us denote rank(K)=p, dim(~)=n
(number of process components), dim(r)=m (number of reactions), and consider a
state partition:

~ = [~:] (5.46)

where ~b contains p (arbitrarily chosen) process variables and ~a the others, but such
that the corresponding submatrix Ka is full rank (rank(Ka)=r). Let us define the state
transformation Z (dim(z)=n-r):
(5.47)
162 5 Bioprocess Control

where Ao is solution of the matrix equation:

(5.48)
The dynamical equation of Z can be readily derived from the material balance
equations (Eq.5.25):
dZ
dt = -DZ + Ao(Fa - Qa) + Fb - Qb (5.49)

where the indices Fa-Qa and Fb-Qb correspond to the feed rates and gaseous outflow
rates of Sa and Sb, respectively. Equation (5.49) is independent of the reaction ki-
netics T. An important special case of the above state transformation is J..L=m (inde-
pendent irreversible reactions), Ao is equal to:
(5.50)
Equation (5.49) is a system linear-in-the state Z; we can check that it is asympto-
tically stable if D is positive (or more precisely, as long as D does not remain equal to
zero for too long. The stability condition is indeed written as follows: if there exist
positive constants p and 8 such that n+ 8 D( r )dr > f3 for all t). Therefore Eq. (5.49)
can be used to compute Z on-line from the knowledge of the feed rates F, the gas-
eous flow rates Q, and the knowledge of some combinations of the yield coefficients.
The variables Z can then be used to estimate on-line the unmeasured variables from
the measured variables by implementing an asymptotic observer which is indepen-
dent of the reaction kinetics (see [2]):
dZ A

dt = -DZ + Ao(Fa - Qa) +Fb - Qb (5.51)

(5.52)
where Z and €b are the on-line estimates of Z and Sb, respectively. Let us apply the
above transformation to our example. Here, n=5 and r=m=2. A particular form of
the asymptotic observer is when Sa contains the measured process components and
Sb the unmeasured ones. However, as it is illustrated here, we have chosen another
state partition. Let us choose, e.g., Sb= [S, C, L] T and Sa= [K, G] T and define z as follows:

z= [~~l = [S-k;,~~~k4Xl (5.53)

Z3 L - ksX +~G

The dynamical equations of Z are readily derived from Eqs. (5.49), (5.50), and
(5.53):

(5.54)

dZ z kz
dt = - DZz + Qin - k3 QI (5.55)

(5.56)
5.4 Adaptive Linearizing Control of Bioprocesses 163

Let us see how it specializes in the animal cell culture example in the particular
situation where we are only interested to have an on-line software measurement of
the biomass X. The asymptotic observer is then written as follows by using
Eqs. (5.53) and (5.56); Eqs. (5.57) and (5.58) represent Adaptive Linearizing Control
(Part I): On-line Estimation of x:
dZ J k4 - kJ
dt = (5.57)
A

- DZJ + DSin + -k-3- QJ

(5.58)

where X represents the on-line estimate of X provided by the aymptotic observer.

Note that in the second equation of the above software sensor (Eqs.5.57 and 5.58),
we have considered the low solubility assumption for the CO 2 (G=O). Note that the
asymptotic observer requires the knowledge of the ratio (k4-kdlk3 and the yield
coefficient k4 •

5.4.3.2
The Monitoring Tool 2: The Parameter Estimation

The parameter estimation has a double objective here. The first one is to give on-line
estimates of uncertain parameters (the reaction kinetics in our example, see assump-
tion A3). As already mentioned in Sect. 5.3.7, a second important objective of the
parameter estimation is to incorporate integral action in the control scheme (to
eliminate offsets). In adaptive control, there are typically two different approaches
corresponding to the way the parameter adaptation is introduced in the controller:
the direct and the indirect methods. Here we have considered the latter (this choice
is indeed somewhat arbitrary). In Sect. 5.3.7, we have considered an observer-based
estimator; here we consider an alternative approach: the unknown parameter(s) will
be estimated on-line via a recursive least-square (RLS) estimation algorithm with a
forgetting factor. The usual form of the RLS form with forgetting factor for a system
described in discrete-time by the equation
(5.59)
(where e is the vector of parameters to be estimated, and t the time index) is written
as follows:
(5.60)

Gt=Gt-J(I_ 4VP~Gt ),O</' ::;1, Go>O (5.61)

/' /' + ¢t Gt¢t
where Gt is the RLS gain matrix and y the forgetting factor. Let us derive the RLS
estimation algorithm in our animal cell culture example. Here we assume that kJ1k3
is (sufficiently well) known and to report all the uncertainty on the kinetics para-
meter /Lp. We further assume (in line with basic kinetics laws) that there is no growth
in absence of the limiting substrate S (/LF (S=O)=O): the simplest way to express it is
to write /LF as follows:
(5.62)
164 5 Bioprocess Control

where B is an unknown, possibly time varying, parameter. The next step in the deri-
vation of the estimation algorithm is to rewrite the glucose equation (Eq. 5.44) in the
format at Eq. (5.59) by discretizing the time derivative dSldt via a first order Euler
approximation (dSldt~St+l-St)ITwith T the sampling period). This gives:

(5.63 )

i.e., Xt = St+1 - St + T~QI,t - TF~;t (Sin,t - St) and ¢t=-Tk4 StX t. T~e RLS estima-
tion algorithm with a forgetting factor specializes as follows (with X t given by the
asymptotic observer): Eqs.5.64 and 5.65 represent Adaptive Linearizing Control
(Part II); On-line Estimation of B:

(5.64)

(5.65)

5.4.3.3
The Control Tool: The Adaptive Linearizing Controller

Recall that the control problem under study here is the control of some process
component concentration(s) (denoted by y) by acting on the flow rate(s) or influent
substrate concentration(s) (denoted by u). Assume that we wish to have the follow-
ing linear stable closed-loop dynamics for y:

dy = A(y* _ y) (5.66)
dt
with Y* the desired values for y, and A an (arbitrary) stable matrix (see Sect. 5.2.2.2).
By combining Eqs. (5.42) and (5.66) and eliminating dyldt, we obtain the following
linearizing control law:
u =g(~, Brl[A(/- y) - f(~, B)] (5.67)
By considering that some process parameters are estimated on-line, the above con-
trol algorithm (Eq.5.67) is written in its discrete-time version as follows:
u=g(~,erl[A(y*-Y)-f(~,e)] (5.68)
In the animal cell culture example, if we consider the following closed-loop desired
dynamics (with S* the desired ethanol concentration):

~! = '>'(S* - S),.>. > 0 (5.69)

the adaptive linearizing control, Eq. (5.68), specializes as follows: Eq. (70) represents
Adaptive Linearizing Control (Part III); The Controller:
Vt * kl
[.>.(S - St) + -k QI,t + k4 BtStX tl
A A

Fin,t = (5.70)
Sin,t - St 3

The above equation (Eq.5.70), together with Eqs. (5.57), (5.58), (5.64), and (5.65)
constitutes the adaptive linearizing controller of animal cell culture process. Note
5.4 Adaptive Linearizing Control of Bioprocesses 165

that the controller structure is similar to the one presented in the preceding section
(Eq.5.22) (with the addition of the CO2 gaseous outflow rate term), and the com-
ments mentioned then obviously readily apply here also.

5.4.4
Experimental Results

The adaptive linearizing controller (Eqs.5.57, 5.5S, 5.64, 5.65 and 5.70) has been
implemented on a 22-1 pilot-scale bioreactor (see [16]). Figure 5.6 presents one set
of experimental results. The glucose has been measured via an FIA biosensor device.
The temperature and pH were maintained at constant values of 37 DC and 7.2, respec-
tively. The initial conditions were equal to

Vo = 19L, Xo = O.lSx 10 6 eel/simi, So = 1.3 mmolll, Sin = 3.3 molll

In this experiment, CO2 was not available for on-line measurement. Therefore the
adaptive linearizing controller without the carbon dioxide term has been implemen-
ted. The values of the initial conditions and design parameters have been chosen via
numerical simulation and simple computation:
Bo = 0.25 h- 1 mM-t,X=Xo, A=O.4 h-t, y=0.9, S*= 1 mmolll
The initial value of Bis given by iLp,olSo where iLp is given by the simulation model
validated on batch experimental data. The controller gain has been chosen so as to
have closed time constant shorter than in open loop: the highest open loop time
constant was around 5 h, the chosen A corresponds to 2.5 h time constant. The value
of the forgetting factor A (0.9) allows one to handle time variations of the parameter
e. Figure 5.6 presents the controlled variable, i.e., the glucose concentration S

5
b
1.5
- 3 S· .lmM Z
I .oS
::J
f
"'2
0.5

0
0 50 100 150 200

2.5...,- - - -- - - ------.
2
1.5
c
-~ 1
, 0 .5
;- 0 ~,~---------
-0.5
-I

o 50 100 150 200

t {l1)

Fig. S.6a-d. Adaptive linearizing controller of an animal cell culture: experimental results
166 5 Bioprocess Control

(Fig.5.6a), the control variable, i.e., Fin (Fig.5.6b), and the estimated values of e
(Fig.5.6c), and X compared to off-line measurements (Fig. 5.6d), respectively.
Acknowledgments. This paper presents research results of the Belgian Programme on Inter-
University Poles of Attraction initiated by the Belgian State, Prime Minister's Office for
Science, Technology and Culture. The scientific responsibility rests with its authors.

References
1. Axelsson JP (1989) Modelling and control of fermentation processes. PhD thesis, Lund
Inst Technology
2. Bastin G, Dochain D (1990) On-line estimation and adaptive control of bioreactors. Else-
vier, Amsterdam
3. Chen L, Bastin G, Van Breusegem V (1995) A case study of adaptive nonlinear regulation
of fed-batch biological reactors. Automatica 31:55-65
4. Dochain D, Perrier M (1993) Control design for nonlinear wastewater treatment pro-
cesses. Water Sci Technol 28:283-293
5. Dochain D, Perrier M (1997) Dynamical modelling, analysis, monitoring and control de-
sign for nonlinear bioprocesses. Advances in Biochemical Engineering/Biotechnology
56:147-197
6. Feyo De Azevedo F, Dahm B, Oliviera FR (1997) Hybrid modelling of biochemical pro-
cesses: a comparison with the conventional approach. Computers Chern Eng
21 [SuppljS751-S756
7. Holmberg A (1982) On the practical identifiability of microbial growth models incorpor-
ating Michaelis-Menten type nonlinearities. Math Biosci 62:23-43
8. Ljung L (1987) System identification - theory for the user. Prentice-Hall, Englewood
Cliffs, NJ
9. Montague G, Morris J (1994) Neural network contribution in biotechnology. TIBTECH
12:312-323
10. Mosey FE (1983) Mathematical modelling of the anaerobic digestion process: regulatory
mechanisms for the formation of short-chain volatile acids from glucose. Water Sci Tech-
nol 15:209-232
11. Perrier M, Dochain D (1993) Evaluation of control strategies for anaerobic digestion pro-
cesses. Int J Adaptive Cont Signal Proc 7:309-321
12. Pomerleau Y, Perrier M, Bourque D (1995) Dynamics and control of the fed-batch pro-
duction of poly-~-hydroxybutyrique by Methylobacterium extorquens. Proc 6th Int Conf
Computer Appl in Biotechnol, pp 107-112
13. Pomerleau Y, Vie! G (1992) Industrial application of adaptive nonlinear control for baker's
yeast production. In: Karim NM, Stephanopoulos G (eds) Proc 5th Int Conf Computer
Appl in Biotechnol, Pergamon, Oxford, pp 315-318
14. Renard P, Dochain D, Bastin G, Naveau H, Nyns EJ (1988) Adaptive control of anaerobic
digestion processes. A pilot-scale application. Biotechnol Bioeng 31:287-294
15. Seborg DE, Edgar TF, Mellichamp DA (1989) Process dynamics and control. Wiley, New
York
16. Siegwart P, Male K, Cote J, Luong JHT, Perrier M, Kamen A (1998) Adaptive control at
low-level glucose concentration to study HEK-293 cell metabolism in serum pure cultures.
Biotechnol Progress 28:927-937
17. Willems JL (1970) Stability theory of dynamical systems. Nelson, London
6 On-Line Simulation Techniques for Bioreactor Control
Development
Reiner Luttmann, Klaus- Uwe Gollmer

6.1
Introduction

During the last decade, bioprocess development automation tasks have become more
significant for overall process performance. Quality requirements on the one hand
and growing process knowledge on the other have resulted in applied control strate-
gies of increasing complexity. At the same time modern information technology fa-
cilitates higher flexibility of the technical systems used for practical realization of
control tasks [1, 2]. Nevertheless, one of the most important goals in bioprocess
automation is the manipulation of the process to meet desired performance criteria.
Examples are the control of pOl to a prescribed limit or the realization of an expo-
nential substrate feeding strategy. In practice, the process and the controller form a
closed loop, often realized in a feedback fashion (Fig.6.1a). For design and stability
investigations of these control-loops both process and controller require treatment
using a theoretical mathematical model (Fig.6.1b). Obviously the time-dependent
changes of the relevant process variables are of major interest for these applications.
Therefore the underlying model is highly dynamic in nature and typically a set of
differential equations based on mass and energy balances is used for simulation of
the process behavior.
The main advantage of this model-based design strategy is that the optimization
procedure is almost independent of application. Therefore well established methods
from control theory can be applied to bioprocess automation tasks. Finally, the theo-

-_._---,
a
on-line simulation

b
simulation

Fig. 6.1 a-c. Different types of process simulation

168 6 On-Line Simulation Techniques for Bioreactor Control Development

retically optimized strategy has to be transferred to the real-world application by

setting the parameters of commercial PID-controllers or using the high level lan-
guage of a process control system.
Because of the inherent complexity of the underlying software program a number
of test fermentations has to be performed to ensure proper functionality and inter-
action of all control tasks. Due to economic reasons, the verification of a new strate-
gy by on-line simulation is gaining influence in the design cycle of bioprocess auto-
mation methods.
The essential idea behind on-line simulation is to use a model (virtual bioreactor)
to emulate the process behavior and connect this artificial system to the real-world
automation hardware (Fig.6.1c). Therefore, the performance of a new implementa-
tion, for example, can be validated without running a real-world process. In contrast
to the above-mentioned model-based design methods, on-line simulation has to be
done in real-time interacting with the existing automation hardware. Nevertheless,
the computational power of today's personal computers enables on-line simulation
even for complex models.
On-line simulations have been applied to a wide field of technological systems and
engineering problems. One of the most popular examples is the flight simulator,
which runs on almost any personal computer. In this case, a mathematical model
of the airplane is manipulated in real-time by the human operator to ensure a proper
flight. "Virtual reality" is one of the strongest arguments for using on-line simula-
tion tools for training and instruction of operators.

6.2
Application
The fields of application may be classified into two groups, industrial and educa-
tional. Whereas industry is predominantly interested in maximizing profit, aca-
demics are looking for efficient teaching tools.

6.2.1
Application in the Biochemical Industry

Recently the chemical industry has been directing its attention to the development
and application of on-line simulation methods. Remarkable benefits are expected in
the following areas.

6.2.1.1
Plant Set Up

Today, modern software packages allow us to simulate the combination of discrete

and continuous units that are involved in the production and purification of bio-
chemical products [3]. A typical plant model for batch processes consists of unit
operations such as product dilution, filter press, solvent extraction chromatography
operation, ultrafiltration, and pasteurization. Furthermore, cleaning in place (CIP)
equipment is required for various processing areas. Plant-wide simulation challenges
for computer aided process design and the identification of bottle-necks in the whole
production process [4]. From the automation point of view, checking a new control
6.2 Application 169

system in a simulated environment is a typical procedure for the chemical and

petrochemical industries. The results of preliminary tests have significant influence
on the design of production plants. Operators are able to familiarize themselves with
the behavior of the new units months before the real plant is ready. Based on simula-
tion, the reliability of the planning process and the time for putting the plant into
operation could be reduced significantly.

6.2.1.2
Economy

Yield improvement caused by efficient control strategies has an evident influence on

the overall production costs. Application of model-based control algorithms (e.g.,
OLFO or adaptive control) guarantees an optimal performance even under small
disturbances of actual process conditions. On-line prediction gives deeper insight
into future process behavior and enables, for example, the determination of an opti-
mal scheduling strategy for harvest or transfer operations. Deviations from the ex-
pected optimal trajectory could be detected and faults could be identified easily
based on such predictions. As a consequence the operators are able to reduce the
economic loss due to interruption.

6.2.1.3
Quality

Reproducibility is a prerequisite for success of biotechnology in research and indus-

trial production. Producing a product of constant high quality is one of the main
goals in process automation. As a consequence of biological uncertainties, all process
conditions as far as they can be influenced must be brought under control. As al-
ready mentioned, on-line simulation is a powerful tool to optimize and verify the
characteristics of a particular control algorithm even in extracurricular situations.
With it, there is considerable evidence that the outcoming product has the expected
properties.

6.2.1.4
Validation

Reproducibility of product quality is strongly coupled with the question of valida-

tion. Validation plays an important role, especially in pharmaceutical industries,
where the production process has to be in compliance with the requirements of
GMP rules [5]. Integration of on-line simulation into the life cycle for automated
equipment provides the chance to detect failures (e.g., electrical cabling, software
bugs, calibration errors) at a very early stage. As a result, the number of validation
runs could be reduced significantly.

6.2.1.5
Complexity

In the past, a lot of supervision and control tasks in industry has been done by the
human operators. Today, increasing quality requirements and economic reasons have
led to more complex automation structures. The current situation is characterized by
170 6 On-Line Simulation Techniques for Bioreactor Control Development

a high degree of automation and the application of sophisticated control [6]. That
applies to the whole automation hierarchy, beginning from the sensor-actuator system
with built-in fault detection capacity [7] to the plant control system using complex
fuzzy rules. The present state of efficient biotechnology processes, for example high
cell density cultivation techniques, requires fully instrumented bioreactors and com-
plex fermentation strategies. Today's plants for production of pharmaceuticals are
more or less fully automated. A good example is the production of Erythropoietin
using robots to assist the human workers [8]. Particularly in research and develop-
ment applications with its highly flexible experimental set up, there is a need for rapid
realization and verification of efficient automation methods. To get the novel structure
under control, on-line simulation is inevitable. After running a "virtual fermentation"
the probability of correct functional behavior is greatly increased.

6.2.1.6
Training

A fundamental field of application is in professional training of industrial staff. Due

to the fact that nearly every process situation could be simulated, operators are able
to grasp the behavior and influence of control actions by the technique of learning
by doing. In contrast to real plants, fault conditions and accidental situations could
be simulated without danger. This experience and the challenge to evolve and check
new ideas for process improvement leads to a higher motivation. Moreover, the fi-
nancial expenditure for development and maintenance of simulation equipment is
considerably lower than for a real training plant.

6.2.2
Application in Education

The restricted financial situation is one of the strongest arguments for using on-line
simulation at schools and universities. From the educational point of view, teaching
modern methods using "virtual reality" is much more meaningful than obsolete
hardware. Combining real-time and speed-up simulations allows for the skipping of
uninteresting phases, whereas important situations could be replayed in slow-mo-
tion. Obviously, in educational concepts the emphasis is more on models for unit
operations than for simulation of complete plants. Building their own automation
program and verifying the strategy with on-line simulation increase students'
knowledge about interaction in complex technological systems and helps to bring
the theoretical solution into practice.

6.3
General Architecture of On-Line Simulation Systems

The automation infrastructure on the one hand and the user-interface on the other
has formative influence on the general architecture of a typical on-line simulation
system. Maximum reality could be achieved by using an independent device for bior-
eactor imitation and the existing automation infrastructure for process visualization
and control. As shown in Fig.6.2a, the on-line simulation is coupled to the control
system by using A/D- and D/A-converters claiming the electrical plugs in the field.
6.3 General Architecture of On-Line Simulation Systems 171

a real
process
simulation-
operator
•
[SJ
~

p===j
,~

on-line simulation
reaction
system
L exhaust gas

~~~~::;=~ [ijL ::~: trans~r

I aeration
system

agitation
system

b process
real-time simulation
control

[Mil
~
•
process- and
r
programmable
logic controller

controller
'-_ _ _ _ _ _---1

,S~~I~~~~. i:
rocess contr~-I-
syslem
-----'
------agitation
system

Fig.6.2a,b. Integration of process control equipment and simulation systems

One advantage of such an architecture is the complete review of all automation re-
levant devices including the established cabling and the existing user-interface for
operation and visualization. Although quite expensive, such strategy is well suited
for validation purposes in the pharmaceutical industries. After plant set-up, a sec-
ond set of automation devices is needed for running a training-simulation and the
real-world plant in paralleL
172 6 On-Line Simulation Techniques for Bioreactor Control Development

However, an integrated approach shown in Fig. 6.2b emulates the control facilities
and simulate the process simultaneously using only one computer. The main draw-
back of this very economic solution is the lower performance in terms of reality due
to software-imitation of control and visualization devices. It is up to the user to
adjust the expenditure for reproducing the exact behavior of the emulated process
control components. Nevertheless, a high functionality combined with low-cost
hardware makes this architecture convenient for educational or small plant applica-
tions. Some of the commercial simulation systems incorporate Microsoft OLE auto-
mation technology or DDE support and offer the capabilities to build powerful hy-
brid programs by integrating the visualization facilities of SCADA systems like Lab-
View or InTouch.

6.3.1
Components of Simulation Systems

In spite of conceptional differences, three components are common to all on-line

simulation systems. Namely the type of model, the build in mathematical methods,
and the user interface characterize a particular solution.

6.3.1.1
Models

Depending on the intended use, the mathematical description could be more or less
complicated. Precise models for almost all chemical unit operations are available
from the literature or application of mass- and energy conservation laws. After ad-
justing the parameter of such mechanistic formalism, the models coincide with the
real-world application in a wide range of experimental conditions. In contrast to
modeling and simulation for process optimization, the range of validity is one of
the key issues in on-line applications. Especially in training and educational use the
model has to be able to describe process behavior under normal and exceptional
conditions. Moreover the complete process including all instrumentation and auxili-
ary facilities (full scope model) has to be covered by the set of equations. Fortu-
nately, it is often sufficient to describe the dynamic behavior of peripherical devices
in tendency rather than to be in exact quantitative coincidence with the reality. That
is the reason for simple approximation of typical sensor devices by using linear dif-
ferential equations (low-pass filter). Most often, the biological activity is included by
using formal kinetic models.

6.3.1.2
Numerical Methods

From the mathematical point of view, the simulation of typical biochemical engi-
neering models is equivalent to the problem of finding a solution for a set of ordin-
ary differential equations (odes). Generally there is no analytic expression available,
so it is necessary to approximate the solution by numerical means. Beginning at
initial time and with initial conditions, the well known solver (e.g. Runge-Kutta for-
mula) steps through the time interval, computing a solution at each time step. If the
solution for a time step satisfies the solver's error tolerance criteria, it is a successful
step. Otherwise, it is a failed attempt; the solver shrinks the step size and tries again
6.4 Full Scope Model of the Fermentation Process 173

[9]. To meet the real-time requirements in on-line applications, first make sure that
there is enough computing power to finish this iterative process in a defined time
interval. By observation of process- and simulation-time a synchronization task is
responsible for real-time, slow-motion or speed-up simulation mode. At the end of
each predefined time step there has to be an interaction with the external interfaces.
This could be the export of output variables (measurements, controlled variables)
and the import of manipulated variables from the automation devices or operator
action. In addition to ode-solver and synchronization, some optimization proce-
dures should be implemented for identification of unknown parameters or on-line
optimization application.

6.3.1.3
User Interface

To achieve maximum reality, the operation and visualization of the process under
investigation should be done by using the user interface of the existing process con-
trol system. By means of this system, the user is able to redefine set points and
change parameters for the control-loops. At the same time, the simulation itself has
to be operated. Mainly the mode (real-time, speed-up, slow-motion, pause) or model
events (sensor/actuator fault, calibration errors etc.) should be manipulated on-line.
In any case, the user has be able to change or extend the build-in model by using a
high level modeling language or a block oriented diagram.

6.4
Full Scope Model of the Fermentation Process

The principle of virtual processing requires an extended model going far beyond the
classical description of the gas and fluid phase reaction space of the reactor.
Figure 6.3 shows a photograph of a 7.5-1 bioreactor BIOSTAT ED5, for which a
complete dynamic model is to be developed later.
The model, in the form of a virtual process in conjunction with the real fermentor
process control system, is a useful tool in education and research.
This principle, termed on-line simulation, is used in the research field to develop
and test complex processing strategies, which can then be implemented in the real
reactor.
The methodology to be developed requires a description of the complete technical
and biological bioreactor system. This includes the final control elements and local
controllers, the processing, physicochemical and biological behavior, and the indivi-
dual in-line and on-line measurement systems.
Figure 6.4 shows the required subsystems of a model for on-line simulation.
The model covers not only the engineering aspects (temperature control, mixing,
aeration and gas removal, and the mass transport behavior), but also the biological
reaction process (growth, substrate and oxygen uptake, COz- and product forma-
tion) and their interactions (pH, viscosity, foam formation, etc.).
When modeling the actual reaction process the major problems are found in the
exact description of mass transport as a function of agitation speed, aeration rate
and rheology, notation of the interaction of the relevant cell-specific reaction rates,
and the determination of the parameters influencing pH.
174 6 On-Line Simulation Techniques for Bioreactor Control Development

Fig. 6.3. Highly instrumented 7.5-1 bioreactor with integrated on-line simulation. Photo: E.
Stagat, University of Applied Sciences Hamburg

Fig. 6.4. Submodels of the bioreactor system

6.5 Submodels of the Bioreactor Process 175

In addition, feed, harvest and pH titration systems, and the reactors measurement
and control systems must be described.
The individual sub models are introduced in detail below and then integrated to
form the aggregate model.
In contrast to classical modeling of biotechnological processes, a complete de-
scription of the process behavior is required in this case.
Thus, the general function is assured and may be checked when the real process
control system is connected to the simulator.
Furthermore, the real-time and control behavior of the modeled parameters, e.g.,
temperature, agitation speed, aeration rate, or pH may be studied on-line.

6.5
Submodels of the Bioreactor Process

6.5.1
Engineering Components

6.5.1.1
Temperature Control System

The bioreactor temperature control system comprises the reactor liquid phase, the
surrounding jacket, and the supply unit connected by piping. This in turn includes
the heat exchangers for cooling (water) and heating (steam or electrical).
Figure 6.5 shows a schematic diagram of the temperature control system.
Exact modeling requires partial differential equations as both space- and time-
dependent heat transfer processes take place.
The complete model will be treated with a package for the solution of odes within
the scope of this contribution. Thus, a number of simplifications are assumed: ide-

HEATING

~'---+-"""7
230 V

"Ie ¢---=-----, 9. 9,
9, A ••
v,
""
A,.. p,

m,." 9~ m",

m,

24V

Fig. 6.5. Schematic diagram of the temperature control system

176 6 On-Line Simulation :rechniques for Bioreactor Control Development

ally mixed subsystems in the temperature control circulation and no heat loss from
the pipework.
The temperature control system thus comprises the sub model's thermostat heating
(Th), cooling water (C), thermostat cooling (Tc), double jacket (D), and liquid phase
reaction system (L). The heat capacities of the reactor walls (W) are distributed
proportionately to the systems D and L.
Temperature control takes place in a pressurized closed circuit in which water is
pumped with a mass flow mT through the subsystem K, K=Th, Tc, and D. The reci-
procal mean residence time DK in the exchanger system K,

(6.1 )

where
mT = thermostat circulation mass flow rate
mK = exchange mass in subsystem K,
and the equivalent for the cooling water system Dc,
mc(t)
Dc ()
t =-- (6.2)
me
where
me = temperature control manipulated cooling water flow rate
me = exchange mass in primary cooling circuit,
are defined here.
The heat transmission coefficient k JK ,
k _ 1
(6.3)
JK - .1. + OJK +..l.
0'1 AJK Q'K

where
OCr = heat transfer coefficient in subsystem I; I=J,K
DJK = wall thickness between J and K; J,K E {C,Tc,D,L,E}
AJK = thermal conductivity of the wall between J and K,

is used to describe classical heat transfer through a plane wall, which in simplifica-
tion may be assumed for the reactor geometry in this case.
Thus both, the overall heat transmission resistance RJK ,

(6.4)

where
AJK = heat transfer area between J and K,
and the thermal transport time constant TJK,
TJK = RJK · CJ (6.S)
where
CJ = volumetric heat capacity in system J,
are defined.
In subsystems C, Th, and Tc the volumetric thermal capacity CJ,
6.5 Submodels of the Bioreactor Process 177

CJ=mJ· CH20 (6.6)

where
CH20 = specific heat capacity of water
is determined by the mass of water present therein.
The heat capacity of the reactor wall is taken into consideration by fractions al-
lotted to the double jacket D and the liquid phase L of the reactor
Co = mo . CH20 + mwo . cw (6.7)
and
CL(t) = PL(t) . VL(t) . CH20 + mWL· Cw (6.8)

where
mWI = mass of reactor wall fraction in subsystem I; I=D,L
Cw = specific heat capacity of the wall
PL = reactors liquid phase density
VL = liquid volume in reactor.
Thus, the odes of the temperatures in the thermostat system are as follows.
The thermostat heating system Th,
·
OTh(t) = -DTh . OTh(t)
~W
+ DTh . Oo(t) +-- (6.9)
CTh
where
OTh = temperature in thermostat heating system
PH = required electrical heating power
0 0 = temperature in double jacket,
contains only convective heat flux and has the pulse-modulated heating power PH of
the temperature controller as manipulating variable.
The cooling heat exchanger has on the primary side the cooling water system C,

·
Oc(t) [
= - Dc(t) + - 1 ] . Oc(t) + Dc(t) . OCin(t) +--
OTe(t) (6.10)
TCTe TCTe

where
Oc = temperature of the cooling water system
OCin = temperature of the cooling water feed,
and on the secondary side the thermostat cooling system Tc,

·
OTe(t) [
= - DTe + - 1 ] . OTe(t) + DTe . OTh(t) +--
Oc(t) (6.11 )
TTeC TTeC

where
OTe = temperature of thermostat cooling system,
which is connected convectively to the heating system.
The double jacket system D,

Jo(t) = - [Do + _1 + _1] . Oo(t) + Do. OTe(t) + Odt) + OE(t) (6.12 )

TOL TOE TOL TOE
178 6 On-Line Simulation Techniques for Bioreactor Control Development

where
{)L = temperature of liquid phase
{)E = temperature of reactor environment,

is connected convectively to the cooling system and is involved in heat exchange

with the liquid phase L and the reactor environment E.
The liquid phase of the reactor, the reaction volume L,

(6.13)

where
QSt thermal power of the stirrer
QM thermal power of the microorganisms,

is involved in heat exchange with the double jacket D and surroundings E via the
reactor lid and base.
The liquid phase receives heat generated by the stirrer QSl>
(6.14)

where
KHSt = proportional gain of stirrer heat generation
Nst = agitation speed,

and heat produced by the microorganisms QM,

(6.15)

where
KHM = proportional gain of microbial heat generation
OUR = microbial oxygen uptake rate.
The electrical circuit representation in Fig. 6.6 shows the connection of the reactor
thermostat model to the control hardware of the (laboratory) system of Fig. 6.3.
The cooling water entry heat resistance Rc,
1
Rc(t) = . , (6.16)
mc(t) . CH20
describes the convective primary cooling water flow, whereas the thermostat circula-
tion heat resistance RT,

RT =-:-.--- (6.17)
mT . CH20

accounts for the convective heat flux in the thermostat system.

There is no feedback from the convective streams which are thus depicted as a
trap amplifier.
The jacket entry temperature {)Tc and the liquid phase temperature {)L are fed as
actual values to a cascade control system. The slave controller output pulses the
signal COOLING to the cooling water valve, described by the connection of resis-
tance Rc, or the signal HEATING to switch on the electric heating power.
6.5 Submodels of the Bioreactor Process 179

HEATING COOLING

Fig. 6.6. Electrical circuit representing the temperature control system

6.5.1.2
Pressure Behavior

The pressure in the liquid phase PL and gas phase PG are equal for the reactor de-
scribed here.
During cultivation ('l9 L<100°C) they correspond to the set point PGw of an ideal
pressure controller according to

(6.18)

During sterilization ('l9 L>100°C) both air inlet and outlet are closed.
In this case, fermentor steam pressure, given by

(6.19)

with interpolation constants according to Dupres-Rankine,

KpLJ = 9.807 N/m2
KpL2 = 10.9
KpL3 = 2461 K
KpL4 = 2.065,
is controlled using the liquid phase temperature TL,

(6.20)

where
TnG = 273.15 K, (normalized conditions).
180 6 On-Line Simulation Techniques for Bioreactor Control Development

6.S.1.3
Aeration Behavior

A well-equipped bioreactor has a gas-mixing station with four massflow controllers

for air, O2, N2 and CO 2.
These are controlled via the corresponding set points Fn1wo Assuming ideal control
the corresponding gas throughput at normalized conditions is obtained with
FnI(t) = Fn1w(t) (6.21)
for I = AIR, 02, N2, and C02.
Hence, total aeration rate FnG is obtained,
(6.22)
again referring to normalized conditions.
Together with the O2 mass fraction XOGim
. ( ) _ Fn02(t) + XOAIR . FnAIR(t) (6.23)
XOGm t - FnG(t)

where
XOAIR = 0.2094 O2 mass fraction of air,
and the CO 2 mass fraction XCGim
. ( ) _ FnC02(t) + XCAIR· FnAIR(t) (6.24)
XCGm t - ()
FnG t
where
XCAIR = 0.0003, CO 2 mass fraction of air,
all conditions being defined at the gas entry point.

6.6
Mass Balances of the Complete Aerobic Growth Process

The model to be developed refers to the description of a growth process of

Escherichia coli (cell mass X) on the substrates glucose (S1) and glycerol (S2). The
cells consume oxygen (0) and generate carbon dioxide (C).
The cell mass is initially viewed as a dissolved component.
In addition, the production of acetate (P1) and its reuse (S3) is taken into account.
In order to simulate the whole process the pH system with the components buffer
acid (B1), buffer base (B2), added acid (Ac), and added base (AI) is required. This
requires not only mass balancing of the liquid phase reaction space (L) but also of
the acid (Tl) and base (T2) feed vessels. The substrate feed vessel (R) must also be
included in the model, as substrate is fed during the process.
Since O2 supply and CO 2 removal is realized via a flowing gas phase, both these
components must be balanced over this phase.
The general mass balances of the process form the starting point of modeling [10 J.
This is done over the macroscopic volume element of an aerated biphasic system
shown in Fig. 6.7.
6.6 Mass Balances of the Complete Aerobic Growth Process 181

off gas

mOGin ~OGout
meGin mCGout

F
R
.{CS1A
COR

feed
harvest and
samples

Fig. 6.7. Macroscopic volume element used to develop the mathematical model of an aerobic
growth process

It will be assumed that the reactor may be described as an ideally mixed vessel,
i.e., exit stream component fractions correspond to the values in the related subsys-
tems.

6.6.1
Gas Phase Balances

The oxygen and carbon dioxide mass balances are the most common on-line tools
for the observation of cell activity of aerobic processes.
By introducing an inert balance for N2 and assuming ideal gas cooling we have the
basis for a global gas phase mass balance.
The oxygen supply rate Q02,

Q (t) = mOGin(t) - mOGout(t) =

02 VL(t)
(6.25)
FnG(t) . MOl XOGin(t)· [1 - XCGout(t)]- XOGout(t)· [1 - XCGin(t)]
VnM . VL(t) . 1 - XOGout(t) - XCGout(t)
where
mOGin/out = O2 mass flow at gas phase entry and exit, respectively
FnG = aeration rate at normalized conditions
M02 = mole mass of O2
VnM = gas mole volume
VL = liquid volume
XIGin/out = mass fraction of component I; I=0(02)'C(C0 2),
is calculated with respect to the reaction volume VL, as is the CO 2 removal rate QC02'

( t) _ mCGout(t) - mCGin(t) _
QC02 - VL(t) -
(6.26)
FnG(t) . Mco2 XCGout(t)· [1 - XOGin(t)]- XCGin(t)· [1 - XOGout(t)]
VnM . VL(t) . 1 - XOGout(t) - XCGout(t)
where
mCGin/out = CO 2 mass flow at gas phase entry and exit, respectively
MC02 = mole mass of CO 2,
182 6 On-Line Simulation Techniques for Bioreactor Control Development

By introducing the molar respiration quotient RQ,

RQ(t) = QcOl(t) . MOl , (6.27)

Mcoz . QOl(t)
and a rating for the gas supply with the maximum oxygen supply rate QOlmax which
may be achieved at equal aeration rate with pure oxygen,

(6.28)

one obtains, in conjunction with Eq. (6.25), a clear representation of the O2 supply
rate Q02,

() () XOGin(t) - XOGout(t)
QOl t = Q02max t . 1 _ [1 _ RQ(t)] . XOGout(t) . (6.29)

In the special case of RQ=l the O2 and CO 2 balances become decoupled with
(6.30)
In this case the difference in mass fractions,
(6.31)
is the same for O2 and COl> and both Eq. (6.25) and Eq. (6.26) decoupled.
Balancing the gas volume in contact with the liquid phase shown in Fig. 6.7 leads
to the gas phase O2 mass balance,
d
dt (mOG(t)) = VL(t) . [Q02(t) - OTR(t)] (6.32)

where
mOG = O2 mass in the reactor gas phase
OTR = O2 transfer rate between gas and liquid phases,
and to the gas phase CO 2 mass balance,
d
dt (mCG(t)) = Vdt) . [-QC02(t) - CTR(t)] (6.33)

where
mCG = CO 2 mass in the reactor gas phase
CTR = CO 2 transfer rate between gas and liquid phases.
An exact solution of the odes (Eqs.6.32 and 6.33) requires knowledge of the gas
volume VG. Since the accumulation terms may generally be neglected, only the qua-
si -steady state terms,
OTR(t) = Qoz(t) (6.34)
and
CTR(t) = -QC02(t) (6.35)
are used.
Thus, both transfer rates may be measured via the exit gas balances.
6.6 Mass Balances of the Complete Aerobic Growth Process 183

6.6.2
The O2 - and CO 2-Transfer Equations

The descriptions of the Oz supply to and COz removal from the liquid phase are
significantly different.
The oxygen transfer rate OTR,

(6.36)

where
kLa = volumetric Oz transfer coefficient
C~L = Oz equilibrium concentration with the gas phase
COL = Oz concentration in liquid bulk

is coupled to the gas phase via the equilibrium concentration c~L'

c* (t) = POG(t) = PG(t) . x (t) (6.37)

OL H02(t) H02 (t) OG

where
POG = oxygen gas phase partial pressure
Hoz = Oz Henry coefficient
PG = total gas phase pressure
XOG = Oz mass fraction in the gas phase.

By defining two Oz transfer par~meters, the maximum Oz saturation concentration,

PG(t)
COLmax(t) = Hoz (t) , ( 6.38)

and the theoretical maximum Oz transfer rate,

(6.39)

and substituting Eq. (6.29) and Eq. (6.36) in Eq. (6.34) with COL ---+0 the only useful Oz
transfer scale-up criterion is obtained, the oxygen transfer capacity OTC,

OTC(t) = QOZmax(t) . OTRmax(t) . XOGin(t) .

Q02max (t) + OTRmax (t)
2 (6.40 )
1 _ 4·[1-RQ(t)]·Q02mox(t)·OTRmox (t)·XOGin(t)]
1+ (Q02mox (t)+OTRm,,(t))2

The same kLa value is used to describe the carbon dioxide transfer rate CTR,

(6.41)

where
DC/o = ratio of COz/O z transfer coefficients
ceL = COz equilibrium concentration with the gas phase
CeL = non-dissociated fraction of the CO 2 concentration in the liquid phase,

as used for O2 transfer (Eq.6.36).

184 6 On-Line Simulation Techniques for Bioreactor Control Development

The gas phase driving force CeL>

(6.42 )

where
HcOl = Henry coefficient for CO2
XCG = mass fraction of CO 2 in the gas phase,
is derived from Henry's law.
The driving CO 2 concentration in the liquid phase CCL>

CCL (t) -- (C~L(t))2 . Mco2 .C (t)

(6.43 )
+ 2 + CLlot
(CHL(t)) + KCl . CHL(t) + KCl . KC2

where
C~L = molar H+ concentration in the liquid phase
CCLlot = molar CO 2 total concentration in the liquid phase
KCj = dissociation constant of CO 2; j=1,2,

is obtained solely from the non-dissociated fraction of the molar CO 2 total concen-
tration CCLlot which is composed of CO 2, carbonate, and bicarbonate.
The molar hydrogen ion concentration C~L is calculated from the pH model de-
scribed later.

6.6.3
The kLa Correlation

The literature reports numerous kLa correlations based on aeration rate, power input
from the stirrer, and the viscosity of the fermentation broth [11],

(6.44)

where
CkLa = fitted constant
PSt = stirrer power input
UG = superficial gas velocity
Tleff = effective viscosity
(X,(3,"( = correlation parameters.

For the on-line simulation the kLa value must in some way be linked to the agitation
speed, aeration rate, and viscosity.
In the following the stirrer power input PSt,

(6.45)

where
Ne = Newton number
d 2 = stirrer diameter
Nst = stirrer speed,
6.6 Mass Balances of the Complete Aerobic Growth Process 185

will be substituted by the stirrer speed Nst, and the gas superficial velocity UG,

(6.46)

where
TnG,PnG = state of the gas phase at normalized conditions
TGimPGin = state of the gas phase at reactor entry
Av = vessel cross-sectional area,
by the aeration rate FnG .
Assuming the simplification of constant Ne, Pi> 11e[6 a simple kLa correlation may
be derived,

(6.47)

where
NStmax = maximum stirrer speed
FnGmax = maximum aeration rate
VLmin = minimum working volume
kLamax = maximum kLa value.
The minimum working volume VLmin is that volume which, under maximum aera-
tion rate FnGmax and maximum stirrer speed NStmax, does not result in stirrer flood-
ing. In this case the maximum kLa value is obtained.

6.6.4
The Liquid Phase Balances

In order to describe the reaction process with the three modes of operation - batch,
fed batch, and continuous - a mass balance on the liquid volume VL is required,
(6.48)
in which it is presumed that the material properties in the entry streams and in the
reactor are the same.
The volumetric flow into the reactor Fin,
(6.49)
where
FR = substrate feed rate
FTi = acid titration rate during pH control
FTZ = base titration rate during pH control
Fv = water evaporation rate due to aeration,
takes into account medium components such as feed and the loss due to evaporation.
The volumetric flow from the reactor Fout,
(6.50)
where
FH = harvest or transfer rate
Fs = sampling rate,
186 6 On-Line Simulation Techniques for Bioreactor Control Development

describes the loss of medium from the reactor.

Hence, the general mass balance of component I in the liquid phase L may be
given with respect to the component's (mass) concentration CIL,

(6.51)

where
CIR = concentration of I in the substrate vessel
rlL = volumetric reaction rate of component I
ITR = transfer rate of component I from the gas phase.
The volumetric reaction rates rlL,
rIL(t) = ± qI/x(t) . CXL(t) (6.52)
where
qI/X = cell-specific reaction rate of component I
CXL = dry biomass concentration in the liquid phase,
will be replaced in the following by the cell-specific activities and the cell concentra-
tion, whereby the cell-specific reaction rate M is termed qX/x,
The following balance equations are stated for dissociable components with molar
concentrations CI.
The cell mass balance (dry biomass X),
. Fill(t)
CXL(t) + -(-)
V t
. cxdt) = qx/x(t) . CXL(t), (6.53)
L

the glucose balance (substrate S1),

. Fill(t) FR(t)
CSlL(t) + -(-)
VL t
. CSlL(t) = -(-) . CSIR(t) - qSI/x(t) . CXL(t),
VL t
(6.54)

the glycerol balance (substrate S2),

CS2L(t) + Fill((t)) . cS2dt) = -qS2/X(t) . cxdt) , (6.55)

VL t
the acetate balance (product PI and substrate S3),

CPILlo! (t) + Fin(t)

Vdt)
.C
PILlo!
(t) - [qPI/X(t) - qS3/X(t)].
- M pi
(t)
CXL , (6.56)

the dissolved oxygen balance (0),

. () Fin(t). ( ) _ FR(t) . COR + FT!(t) . COT! + FT2 (t) . COTZ
COLt + VL ()t COLt - VL ()
t
+ (6.57)
+ OTR(t) - qo/x(t) . cxdt),
the dissolved carbon dioxide balance (C),
C () Fin(t). C ( ) _ FR(t) . CORIot + FT!(t) . CCTlto! + FT2(t) . CCT2!o!
CLIo! t + Vdt) CLIo! t - Vdt) +
+ CTR(t) + qC/x(t)· CXL(t) , (6.58)
MC02
6.6 Mass Balances of the Complete Aerobic Growth Process 187

and the ammonia balance (pH base Al and nitrogen source),

· () Fin(t) () FT2 (t) AlTR(t) - qAljx(t)· cxdt)
CAILtot t + Vdt) . CAILtot t = VL(t) . CAITZtot + MAl ) (6.59)

comprise the reaction model.

The pH is controlled by addition of alkali component ammonia at the titration rate
FTZ , fed from vessel T2.
The loss of ammonia by evaporation,

(6.60)

where
KAlvol = evaporation constant of ammonia
Kw = ion product of water
KAI = dissociation constant of ammonia
is given as analogue to Raoult's Law for the non-dissociated fraction.
A model of the pH behavior requires the description of the pH buffer and added
acid in addition to the correction by the alkali component ammonia.
In the case example the pH will be corrected by phosphoric acid H zP0 4 •
The phosphoric acid balance (pH acid Ac),
· Fin(t) FT! (t)
CAcLtot(t) + Vdt) . CAcLtot(t) = VL(t) . CAcTltot) (6.61)

contains only the pH controller correction with the acid titration rate FT!'
The buffer chosen comprised potassium dihydrogen phosphate KH zP0 4 (B1) and
potassium hydrogen phosphate K zHP0 4 (B2).
Both components are diluted during feeding.
Thus, two purely hydrodynamic balances for the buffer acid 0=1) and base 0=2)
are required,
· Fin(t)
CBjLtot(t) + -(-)
VL t
. CBjLtot(t) = o. (6.62)

6.6.5
The Feed and Titration Vessels System

Measurement of the weights of the substrate, acid, and base feed vessels can be used
to obtain information on the integral activity of the cells.
A description of the whole bioreactor thus requires a mass balance of the substrate
feed vessel,
(6.63)
the acid feed vessel,
mTl(t) = -PTl' FTl(t) (6.64)
and the base feed vessel,
mTz(t) = -PTZ . FTZ(t) (6.65)
188 6 On-Line Simulation Techniques for Bioreactor Control Development

where in each case

PK = density of the medium in vessel K.

6.7
The pH Model

The mass balances mentioned form the prerequisite basis for modeling the pH value
which is influenced by the solution of Eqs. (6.56), (6.58), (6.59), (6.61), and (6.62).
By definition the pH value is a dimensionless measure of the number of hydroxo-
nium hydrogen ions (H30+) present in the electrolyte solution,
(6.66)

which may be found in various hydration forms and are given as a mass concentra-
tion C~L in molrl.
The starting point of a pH model is defined by the ionic product of water,
C~L(t) . COHL (t) =Kw = 10- 14 moe 1-2, (6.67)

and the electroneutrality condition which states that the sum of the product of con-
centration and ionic valency must be equal for positive and negative ions [12, 13].
From this one derives the basic pH equation,

COHL(t) + LPi' Cft(t) = C~L(t) + Lqj' C~+(t) (6.68)

j

in which
= an arbitrary anion (except OH-)
= an arbitrary cation (except H30+)
Pi, % = valency of the anion i or cation j
CiL ' CjL = molar concentration of anion i or cation j.
Substituting Eq. (6.66) in Eq. (6.68) and setting
x(t) = C~L (t) (6.69)

one obtains the solution polynomial,

(6.70)

for the unknown hydrogen ion concentration C~L'

The modeling objective is to find all ions in the feed medium, to determine the
degree of dissociation for them, and to describe the changes to processes due to cell
activity or actions taken externally.
In the present example the pH is influenced by the following:
l. By the pH buffer mixture of potassium dihydrogen phosphate (KH2P04), CBILtot,
and potassium hydrogen phosphate (K2HP04), CB2Ltot.
These two salts of phosphoric acid dissociate completely, resulting in three posi-
tive potassium ions,
(6.71)
6.8 The Reaction Model 189

The liberated phosphoric acid,

CBLtot(t) =CB1Ltot(t) + CB2Ltot(t) (6.72)

dissociates in three stages with constants KB1 , KB2 , and KB3 , resulting in three ionic
fractions.
2. By acetate (CH 3COOH), a metabolic product, CP1Ltot.
Acetic acid protolyses to a single ionic charge with a dissociation constant Kp1 .
3. By carbon dioxide (C0 2), a metabolic product, which may enter the system via the
aeration system or exit via the exhaust, CCLtot'
Carbon dioxide can give up two protons in aqueous solution as carbonic acid.
4. By the ammonia (NH 3) added for pH control and as a nitrogen source, CAlLtot '
Ammonia forms ammonium cations (NHt).
5. By the phosphoric acid (H 2P0 4 ) added for pH control, CAcLtot.
In the same way as the buffer, three ionic fractions are obtained.

After substituting the variables involved for all dissociation stages in Eq. (6.70) a
solution is obtained for the protons C~L determining pH,

f(x(t)) = x2(t) + { [CB1Ltot(t) + 2· CB2Ltot (t) + KA1~~(~~~)Kw . CAILtot(t)]

Kpl KC1·X(t)+2·Kcl·Kc2
[
- x(t) + Kp1 . CPlLtot(t) + x2(t) + KCl . x(t) + KCl . KC2 . CCLtot(t)
KBl . x2(t) + 2· KBI . KB2 . x(t) + 3· KBI . KB2 . KB3
+ .
x3(t) + KBI . x2(t) + KBI . KB2 . x(t) + KBI . KB2 . KB3

. (CBILtot(t) + CB2Ltot(t) + CACLtot(t))]} . x(t) - Kw = O. (6.73)

6.8
The Reaction Model

The reaction model chosen is based on the bottle-neck principle in which a single
enzymatic catalysis determines the reaction [14].
The cell-specific growth rate, qx/x,
(6.74)

where
/L = cell-specific growth rate observed
/Llmax = maximum possible cell-specific growth rate on glucose
/LSgr = substrate-determined cell-specific growth rate
/LOgr = Or determined cell-specific growth rate
qX/Xm = cell-specific (death) maintenance rate

is controlled by the rate-determining step in the substrate or O2 supply to the cells

and is limited by the maximum specific growth rate on glucose, /Llmax'
In order to avoid zero order reactions in the substrate mass balances, the uptake
for maintenance purposes is substituted by qX/Xm in the cell mass balance equation.
Thus, the cell obtains its maintenance energy from storage substances.
190 6 On-Line Simulation Techniques for Bioreactor Control Development

acetate

q S11X

1A<]- - - -I - --YPI1$2

[do,, .,
YXIO

~XlS3) YXlS2 YXISI

limitation
).IJ
+112 ).II
- +
~lsw ).10
)
~IS
.,

LJ
).lOw ?

substrate
limitation

~
nOrTllally closed
(lim.itation) (;
normally open
0: (il\hibitl0n)
q)(J)(

Fig.6.S. Bottle-neck reaction scheme of a multi-substrate model

Figure 6.8 shows the multi substrate reaction scheme described in a simplified
form.
With adequate O2 supply the cell-specific uptake rate of the preferred substrate 1,
glucose,

eSlL (t)
qSI/xopt(t) = qSI/Xmax(t) . k + CSIL (t) (6_75)
SI

where
qSI/Xopt = optimum cell-specific glucose uptake rate
qSI/Xmax = maximum cell-specific glucose uptake rate
kSI = glucose limitation constant
is only limited by the associated concentration, CSIL' This is depicted in the bottle-
neck as a spring-loaded valve (normally closed).
6.8 The Reaction Model 191

Glucose inhibits glycerol uptake,

CS2L(t) kI21
qS2/Xapt (t) = qS2/xmaA t) . kS2 + CS2L (t) kI21 + Cm (t) (6.76)

where
qS2/Xapt = optimum cell-specific glycerol uptake rate
QS2/Xmax = maximum cell-specific glycerol uptake rate
kS2 = glycerol limitation constant
kl21 = inhibition constant for glycerol uptake (S2) due to glucose (S1),

and both these inhibit the uptake of acetate,

CS3L (t ) rr2 kl3j

(6.77)
QS3/xapt(t) = QS3/Xmax(t) . k
S3
+ CS3L (t) . j=! k I3j + CSjL (t)

where
QS3/Xapt = optimum acetate uptake rate
QS3/Xmax = maximum cell-specific acetate uptake rate
kS3 = acetate limitation constant
k13j = inhibition constant for acetate uptake due to glucose (j=1) and glycerol
(j=2).

The inhibition mechanism is symbolized by the characteristic curve of a spring-

loaded (normally open) valve in Fig. 6.8.
The respective maximum cell-specific uptake rate, QSi/Xmax,

l1imax(t)
QSi/Xmax ()
t = - - - + QSi/Xm (6.78)
YX/Sigr
where
YX/Sigr = cell mass growth yield for substrate i
QSi/Xm = cell-specific maintenance rate for substrate i,

contains the pH and temperature dependent maximum cell-specific growth rate on

substrate i [15],

l1imaAt) = l1iapt' KT(t) . KpH(t) (6.79)

where
l1iapt = optimum specific growth rate on substrate i
K1 = environmental growth control function of parameter 1, 1='!9u pH
In the temperature range '!9 L E ['!9 Lmi n> '!9Lmax] the growth rate is controlled by

(6.80)
where
'!9Lmin = minimum temperature boundary of growth
'!9 Lmax = maximum temperature boundary of growth
'!9Lopt = temperature of optimal growth.
192 6 On-Line Simulation Techniques for Bioreactor Control Development

Similar behavior of growth is observed in a pH-range of [pH mim pHmax],

(pH(t) - pHmin) . (pH(t) - pHmax)
KpH (t ) = 2 (6.81)
(pH(t) - pHmin) . (pH(t) - pHmax) - (pH(t) - pH opt )

where
pHmin = minimum pH boundary of growth
pHmax = maximum pH boundary of growth
pH opt = pH-value of optimal growth.

Both the temperature control function KT and the pH control function KpH are zero
outside their defined ranges.
Due to considerations of consistency, the cell-specific maintenance rate, qX/Xm,
must be of equal magnitude for all substrates,

qX/Xm =YXiSigr' qSi/Xm, i = 1,2,3. (6.82)

In order to maximize its metabolic efficiency, the cell will take up all three carbon
sources simultaneously (mathematically) when there is adequate O2 supply,
3

J'LSgr(t) = LYX/Sigr' qSi/Xopt(t) ::::; J'Llmax(t) + qX/Xm' (6.83)

i=l
In reality, due to catabolite repression, the substrates are taken up in sequence.
This is brought about by the inhibition terms.
The supply of oxygen is the second bottle-neck in Eq. (6.74).
The cell-specific O2 uptake rate, qo/x,

qO/x(t) =qO/Xgr(t) + qO/Xm, (6.84)

consists of a growth fraction, qO/Xgr> i.e., the energy for substrate uptake, and a frac-
tion for conversion of maintenance storage substances, qO/Xm'
In the case of an O2 bottle-neck the possible growth fraction, J'LOgr>

J'LOgr(t) = YX/Ogr . qO/xgr(t)

COL(t) (6.85)
= (J'Llmax(t) + qX/Xm) . ko + COL(t)
where
YXlOgr = O2 growth yield coefficient
ko = O2 limitation constant,
is controlled by the dissolved oxygen concentration, cOL'
After evaluation of the rate-determining step in Eq. (6.74),

qx/xgr(t) = min {J'LSgr(t), J'LOgr(t)} (6.86)

where
qX/Xgr = resulting cell-specific growth fraction,
the associated cell-specific reaction rates may be calculated.
The cell-specific growth rate, qx/x,

qXlx(t) = qX/Xgr(t) - qX/Xm, (6.87)

6.8 The Reaction Model 193

the cell-specific O 2 uptake rate, qo/x,

qx/xgr(t)
qo/x (t ) = + qO/Xm, (6.88)
YX/Ogr

the cell-specific glucose uptake rate, qSI/X,

qx/xgr( t)
qSI/x ()
t = , (6.89)
YX/Slgr

and the cell-specific ammonia (N 2) uptake rate, qAIJX,

_ qX/Xgr (t)
qAl/X (t) - (6.90)
YX/Algr

where
YX/Algr = cell mass yield for ammonia,

may be calculated directly from the growth fraction, qX/Xgr-

The cell-specific glycerol uptake rate, qS2/X,

_ qX/Xgr(t) - YX/Slgr . qSI/x(t)

qS2/X (t) - , (6.91)
YX/S2gr

and the cell-specific acetate uptake rate, qS3/X,

2
qx/xgr(t) - LYX/Sigr· qSi/X(t)
qS3/X(t) = i=1 (6.92)
YX/S3gr

are calculated recursively, resulting in the residual fraction for the optimum growth
rate.
Finally, the cell-specific acetate production rate, qPI/X,

qPl/X = [6~
i=1
( )]
YPI/Si . qSi/X t . k
kIP
10
+ (t)
IPIO COL
(6.93)

where
YPI/Si = acetate yield from substrate i, i=1, 2
k IPlo = inhibition constant for acetate production due to oxygen,

completes the reaction model.

Acetate is formed by uptake of both glucose and glycerol. Product formation is
inhibited by the oxygen content of the medium.
The simulation of a complete process in a bioreactor is possible using the model
developed in Sects 6.3-6.6.
For a discussion of further details of the model described, e.g., foam development
and control, the reader is referred to [16].
In the next section the model is used for the development of processing strategies.
194 6 On-Line Simulation Techniques for Bioreactor Control Development

6.9
Application Examples of On-Line Simulation Techniques

Figure 6.9 shows the Digital Control Unit DCU of the bioreactor BIOSTAT ED5,
shown in Fig. 6.3. The DCU is linked to the on-line simulation system BIOSIM and
the bioreactor host process control system UBICON.
The UBI CON system on the right is the master control system of the actual fer-
mentation process [17]. It is developed for biotechnological research and based on a
VMEbus system. This master computer is connected to its peripheral I/O units via a
CAN-bus. The I/O station shown, CTERM, is responsible among other things for
communication with the DCU and other serial peripherals by translation of serial
protocols to the CAN data transfer.
The VMEbus system on the left of Fig. 6.9 is the on-line simulator BIOSIM which
includes the I/O structure of the basic bioreactor, the reactor periphery, and the
measurement amplifier outputs. The simulator is part of another UBI CON system
and forms the virtual bioreactor process.

6.9.1
Training with Virtual Reaction Processes

It is frequently impossible to carry out complex cultivations in the course of student

practicals. The extensive preparation and the slow course of a fermentation stand in
stark contrast to the tight scheduling of such practicals. Furthermore, most students
lack the practical abilities necessary to guarantee reproducible and thus useful re-
sults. On-line simulation is a possible remedy. The students run the virtual process
with the same automation technology which they use for other real but simpler ex-
periments in the bioreactor. The advantage is that the students simultaneously ac-
custom themselves to the system but can make mistakes without grave conse-
quences.

Fig. 6.9. On-line simulation setup. Photo: E. Stagat, University of Applied Sciences Hamburg
6.9 Application Examples of On-Line Simulation Techniques 195

Cp1ltct cS2L cS1L p02 pH

(mmoll~'J (g t'J [9 t'] [%1 1-]
10 5.5 5.5 100 8.0

._._._-_._._.-.-._._._._._._._._._._._-_._._._.- ......

,
8 44 44 80 7.6 \. c
\ , S2L
\.

3.3 3.3 60 7.2

2.2 2.2 40 6.8

1.1 1.1 20 6.4

0.0 0.0 a 6.0

0.0 1.1 5.5

I, 12 1 [hi

Fig. 6.10. Course of experiment during students' training

By alteration of the reaction equations and definition of the associated parameters

it is thus possible for the supervisor to specify well-defined processes. In addition,
checking the experimental evaluation is facilitated.
Figure 6.lO shows an example of a practical experiment. The process corresponds
to the reaction behavior of Escherichia coli K12 introduced in Sect. 6.6, with an in-
oculation concentration of 1 g 1-1. The cells grow on the two substrates, glucose and
glycerol, initially present at concentrations of 5 g 1-1. The experiment shows the shift
from oxygen to substrate limitations, the diauxic behavior, the formation and reuse
of acetate, and its influence on the pH. Addition of glucose during the glycerol
growth phase at t=t2 clearly shows the preference for glucose uptake. By alteration
of the settings for the various process phases, e.g., the agitation speed at t=t 1, the
students can run virtual experiments to determine unknown reaction parameters.
This is done either off-line after the experiment is finished, e.g., with Excel, or on-
line using a MAT LAB subsystem running simultaneously.

6.9.2
Development of a High Cell Density Cultivation

6.9.2.1
The wStat Problem

Figure 6.11 shows in principle the reaction behavior commonly encountered with
recombinant production.
The cell-specific product formation rate qP2/X exhibits a marked dependency on
the specific growth rate /-l. At low and high growth rates the cell ceases to produce.
196 6 On-Line Simulation Techniques for Bioreactor Control Development

-~­
~m""

normalized specific normaJized specific

production rate growth rat.

0,6 0.4 0.2 2 4 6 8 10

~
-{).2 ks,
Fig. 6.11. Dependency of cell-specific production rate on substrate concentration and growth
rate

dissolved oxygen

drive
).I-controller unit

-----'OTR

FnG t aeration

Fig. 6.12. The control concept of the j-l-stat

6.9 Application Examples of On-Line Simulation Techniques 197

Between these extremes a production maximum is observed under strong substrate

limitation.
The processing objective is to hold the cell-specific growth rate at an optimum lLopt
using a dynamic fed batch scheme. In order to obtain a high product yield the pro-
cess is to be executed as a high cell density cultivation with a cell dry weight up to
120g1- 1 [18j.
It is not possible to control the specific growth rate using the substrate concentra-
tion as reliable on-line measurements of glucose in the kSI range of E. coli «5 mg I-I)
are not available [19j.
The objective of the strategy which will be developed below is to allow a freely
definable cell-specific growth rate qX/x during production of recombinant proteins
[20j. This mechanism, termed a IL-stat, can be forced by running under substrate or
oxygen limitation. The latter, however, leads to the formation of undesirable bypro-
ducts. It is especially important to suppress acetate production as it leads to disturb-
ing growth and production inhibition. Hence, the measured relative dissolved oxy-
gen partial pressure p02 is kept by addition of substrate in an uncritical O 2 range, as
shown in Fig. 6.12.
This control mechanism simultaneously fulfills the requirement of substrate lim-
itation. However, the process is highly sensitive to falsely set control parameters.
There is a resultant over-supply of substrate for a short period and a shift in the
limitation type.
Constant stirrer speed N st and aeration rate FnG result in an almost constant feed
rate FR. The cell mass in the reactor, mXL, grows linearly in this case and the specific
growth rate IL falls hyperbolically.
A second control mechanism is therefore required to increase the feed rate in or-
der to hold IL constant.
This is done with the IL/agitation controller. IL is determined using reactor mass
balances or a Kalman filter and kept constant by a cascaded stirrer speed controller.
A reduction in IL results in a stirrer speed increase which in turn leads to a higher
p02. The p02/feed controller increases the feed rate FR and thus IL.
From a control viewpoint a number of problems must be solved:
1. The two controllers for p02 and IL are coupled within the cell by unknown regula-
tion mechanisms.
2. The processing objective requires that for a given set point ILw exponential growth
of cell mass mXL>

(6.94)

takes place after feeding has started at t l • This is only possible by coupling the two
controllers.
3. The control variable IL is not directly measurable. Hence, observers must be im-
plemented in the automation scheme.
4. Badly set controller parameters lead to a shift in the reaction-limiting component
and thus to a structural switch in the controlled process.
5. The variability in the dynamical process behavior due to the exponential increase
in cell mass requires adjustment of the controller parameters.

These problems will be solved below using appropriate process control technology
and with the aid of on-line simulation.
198 6 On-Line Simulation Techniques for Bioreactor Control Development

6.9.2.2
Observation of Ce//-Specific Growth Rate

The observation of the cell-specific growth rate is based, in the case of the fL-stat, on
the cell mass balance of a substrate limited cultivation,
(6.95)
where
YX/I = effective cell yield of component I.
The O2 transfer mass flux, mOT,
(6.96)
is obtained form the gas balance Eq. (6.25) without knowledge of the reaction vo-
lume V L •
Formal integration of Eq. (6.95) results in the solution for qX/x,
Yx/o(t) . mOT(t)
qx/x(t) = fL(t) = --'-t-'----'---'-------'-'---

mXLl + IYx/o(T)' mOT (T)dT

tl
(6.97)
YX/SI (t) . FR(t) . CSIR
t
mXLl + I YX/SI (T) . FR(T)' cslRdT
tl

where
tl = time-point begin of fL estimation
mXLl = cell mass at time t l '
Equation (6.97) can unfortunately not be used on-line because both yield coefficients
are themselves functions of the unknown specific growth rate,

(t) _ fL( t) . YX/Ogr

(6.98)
YX/O - ()
fL t + qo/Xm . YX/Ogr + qSI/Xm . YX/SIgr
and

(t) _ fL(t) . YX/SIgr

(6.99)
YX/SI - ()
fL t + qSl/Xm . YX/SIgr
However, for the application in question fL is controlled to the set point fLW'
Assuming perfect control, both yield coefficients in Eqs. (6.98) and (6.99) are con-
stant.
As a result, the fL-observer is derived from the O2 mass balance (Eq.6.96) and its
integration,
mOT(t)
Jl,o (t) = ---t--'-'---- (6.100)
Kow + I mOT (T)dT
tl

where
Jl,o = estimated fL from O2 mass balance.
6.9 Application Examples of On-Line Simulation Techniques 199

The integration constant Kow at time t l ,

K Ow -_ mXL! . (f-kw + qO!Xm . YX!Ogr + qSl!Xm . YX!Slgr) (6.101)

f.1w . YX!Ogr
must be estimated before the f.1 observation is activated. Equation (6.100) is conver-
gent even in the case of incorrect estimation of Kow> since the integral in the denomi-
nator increases.
Equations (6.100) and (6.101) may also be applied for non-limited batch phase
growth on glucose (f.1w=f.11max)'
The second equation for determining f.1,
FR(t)
ils(t) = I (6.102)
KS1w + JFR(T)dT
II

where
ils = estimated f.1 from the glucose feed mass balance,
is applicable only for the substrate limited fed batch phase and is dependent on the
glucose integration constant KSlw>

K Slw _
-
mXL! . (f.1w + qSl!Xm . YX!Slgr) , (6.103)
f.1w . YX!Slgr . CSlR

which must itself be estimated at time t I'

The integration of the feed rate FR in Eq. (6.102),

(6.104)

may be substituted by on-line measurement of the feed vessel mass.

6.9.2.3
Course and Testing of Processing Strategies

Figure 6.13 shows an on-line simulation of a high cell density cultivation (HCDC) for
the production of recombinant DNA products with Escherichia coli in a 75-1 bioreac-
tor. One aim of the process was to cultivate the cells with air aeration and without
oxygen transfer limitation up to 100g1- 1 dry weight [21].
The process starts in a batch mode, where cells grow on the main substrate glu-
cose. In order to avoid Oz-transfer limited growth conditions, at t=4.9 h, a classical
pOz/agitation-control is switched on. During this process stage the pOz-controller is
not allowed to reduce the agitation speed by more than 25% of the previous max-
imum value. Therefore, the end of the batch phase is observable by a steep increase
in the pOz-levei. This is followed by a double diauxic growth phase. First the cells
change to a by-substrate, glycerol, and then to their own product, acetate.
At t=12.85 h all substrates are consumed and the f.1-stat procedure is switched on
automatically, where the process changes into a fed batch-cultivation. During this
stage the cells grow limited on glucose. A set point f-kw=0.11 h- 1 avoids any oxygen-
transfer limitation. Figure 6.13 demonstrates the powerful control method. At the
200 6 On-Line Simulation Techniques for Bioreactor Control Development

P02 [%] vL [I] ~ [h- t ] NSt [rpm]

100 50.0 0.5 1200

47.5
0.4 1000
80
45.0

0.3 800
42.5
60

40.0 0.2 600

40
37.5
0.1 400

35.0
20
0.0 200
32.5

0 30.0 -0.1 0
0 5 10 15 20 25 30 35
t [hI

Fig. 6.13. Course of a simulated high cell density cultivation with f.L-stat strategy; pOz=dis-
solved oxygen tension, VL=liquid volume, f.L=specific growth rate, Nst=agitation speed

pO. [%] FR [I h- t ] ~ (h-t, N., [rpm]

100 0.7 0.5 500-,----------------------,

0.6 0.4 - -.-.-.L~.:..::.z-......,~.,

80 400 ---- " ,I
0.5 0.3
tL~~--A __ .;__ ,
60 300 I;: \
0.4 0.2

j(?\(\ \
40
0.3 0.1
200 !~
~ ·:;t
'.f.
.:" i
~
\
,
0.2 0.0
..... _.
20 100
0.1 -0.1

0.0 -0.2 0
12.0 12.2 12.4 12.6 12.8 13.0 13.2 13.4
t(h]

Fig. 6.14. Transient behavior from batch to fed batch during a high cell density cultivation;
pOz=dissolved oxygen tension, FR=substrate feed rate, f.L=specific growth rate, Nst=stirrer
speed

end of batch, the cells grow up to 8 g r 1 and at the end of the process, 110 g dry mass
per I can be achieved.
The problem of HCDC is an automatic switch from batch to fed batch processing
and the tuning of control parameters during the fL-stat procedure. The advantage of
on-line simulation lies in a simple reproduction of complex process behavior with-
out any preparation of bioreactor and biological material.
6.9 Application Examples of On-Line Simulation Techniques 201

The process discussed in Fig. 6.13 was simulated in real time conditions [20]. As
an example of on-line control, interactions between the process control system UBI-
CON and the simulator BIOSIM is shown in Fig. 6.14. The transient region between
batch and fed batch starts with a step decrease in the specific growth rate, J-L, due to
the loss of glucose. The pOl/agitation control is switched off after the agitation
speed, Nst, is reduced by about 25% of the previous maximum value.
The change of growth from substrate 1, glucose, to substrate 2, glycerol, is now
observable with the first pOl peak. After consumption of glycerol, pOl increases
once more. The second pOl peak indicates the change to substrate 3, the product
acetate. After consumption of the last carbon source, pOl increases again.
The specific growth rate, J-L, is less than zero, due to the substrate maintenance
(death) rate (Eq.6.87), whereas a small amount of oxygen is consumed further on,
due to the Oz-maintenance rate (Eq.6.88).
The last pOz-peak is identified by the control system in order to start the fed batch
process. The agitation speed, Nst, is reduced to a low level and the simultaneous con-
trol of pOl by substrate addition and J-L by agitation speed alteration is switched on.
A high p02 control deviation leads to a strong substrate addition after the begin-
ning of control. The pOl decreases drastically and the substrate-limited process
changes to an oxygen-transfer-limited one. Now the p02 controller reduces the sub-
strate feed rate, FR , slowly to zero. The last steep increase of pOl is due to the glucose
consumption and the next change to substrate limitation. The restart of substrate
feeding ends in a stable pOz-control at the set point of 10% with substrate-limited
conditions and a slowly increase of feed. The time course of J-L indicates the dis-
cussed process behavior. The J-L-control transient behavior is more or less smooth
and ends at the set point of 0.11 h- 1 with a slightly increase of agitation speed.
Figure 6.15 shows an application of high cell density cultivation based on the de-
veloped J-L-stat strategy [22].

pO, 1%1 vL PI ~ [It'l N., {rpml

100 55,0 o.so 1200

!Ml 52.5 0.45 l OBO

80 so,o 0,40 960

70 "7 ~ 0 .35 840

P
60 "li.o 0,30 720

so '2.5 0,25 600

40 40.0 0.20 480

30 37.5 0 .15 380

20 35.0 0 ,10 240

10 32.5 O.OS 120

a 30,0 0.00
10 15 20 25 30 35
l(h]

Fig.6.15. High cell density cultivation with /L-stat strategy; pOl=dissolved oxygen tension,
V L =liquid volume p=observed specific growth rate, Nst=agitation speed
202 6 On-Line Simulation Techniques for Bioreactor Control DeveiopmentReferences

Similar control hardware is used as in the previous on-line simulation in Fig. 6.14.
During batch phase p02/agitation control and jl-calculation are switched on at
t=6.6h.
The end of batch is observed by the double diauxic growth phase.
The consumption of all carbon sources is followed by the J1-stat procedure.
Experiment in Fig. 6.15 and simulation in Fig. 6.13 show similar behavior.
At t=30 h, the end of the controlled range, the cells grow up to 95 g 1-1 and at the
end of the process, 110 g r l dry biomass can be achieved.

6.10
Summary
The integration of on-line simulation methods in the process control system of bio-
reactors is an attractive addition both for training of students as well as for operators
of industrial plants and for the development of advanced automation tools.
The method of virtual processing introduced here relies on complete modeling of
the bioreactor system and its replacement by the real-time simulator with duplicate
I/O connections.
Furthermore, the real process control system of the bioreactor serves to control
the simulation.
This offers an inexpensive method for validating the automation hardware and
software and enables the development of complex processing strategies on the basis
of advanced functions.
The fully automated high cell density cultivation with controlled specific growth
rate (J1-stat procedure) described demonstrates the power of the methods.
Even beginners in fermentation can successfully carry out real high cell density
cultivations with concentrations of over 100 g I-I dry cell weight after training on
the simulator.

References

1. Bailey JE (1998) Mathematical modeling and analysis in biochemical engineering: past

accomplishments and future opportunities. Biotechnol Prog 14:8-20
2. Gollmer K, Posten C (1995) Fieldbus application in the hierachical automation structure
of a biotechnological pilot plant. J Biotechnol 40:99-109
3. Zhou YH, Holwill ILJ, Titchener-Hooker NJ (1997) A study of the use of computer simu-
lation for the design of integrated downstream processes. Bioprocess Engineering 16:367-
374
4. Hwang F (1997) Batch pharmaceutical process design via simulation. Pharmaceutical En-
gineering, Jan/Feb, 26-43
5. GAMP, Good Automated Manufacturing Practice. International Society for Pharmaceuti-
cal Engineering (ISPE). http://www.activa.co.uklgamp/
6. Sonnleitner B (1997) Bioprocess automation and bioprocess design. J Biotechnol 52:175-
179
7. Clarke DW, Fraher PMA (1995) Model-based validation of DOx Sensor. Preprint of IFAC
Workshop on on-line fault detection and supervision in chemical process industries, pp
216-223
8. Kunitake R, Suzuki A, Ichihashi H, Matsuda S, Hirai 0, Morimoto K (1997) Fullyauto-
mated roller bottle handling system for large scale culture of mammalian cells. J Biotech-
nol 52:289-294
9. Mathworks (1997) Using Matlab. The Mathworks Inc
References 203

10. Dunn IJ, Heinzle E, Ingham J, Prenosil JE (1992) Biological reaction engineering. VCH,
Weinheim
11. Reuss M (1993) Oxygen transfer and mixing: Scale up implications. In: Rehm HJ, Reed G
(eds) Biotechnology vol 3. VCH, Weinheim, pp 188-217
12. Fomferra N (1993) Mathematical modelling and real-time simulation of biotechnological
processes. Diploma thesis, University of Applied Sciences, Hamburg
13. Gustafsson T (1982) Calculation of the pH value of a mixture of solutions. Chern Eng Sci
37:1419-1421
14. Bellgardt KH (1991) Cell models. In: Rehm HJ, Reed G (eds) Biotechnology, vol 4. VCH,
Weinheim, pp 267-298
15. Rosso L, Lobry JR, Bajard S, Flandrois JP (1995) Convenient model to describe the com-
bined effect of temperature and pH on microbial growth. Am Society Microbiol 61:6lO-
616
16. Luttmann R, Gollmer K (1998) BIOSIM, an on-line simulation system for education and
research. In UBI CON - universal bioprocess control system, user's manual
17. Gollmer K, Gabel T, Nothnagel J, Posten C (1992) UBICON - an universal bioprocess
control system. DECHEMA biotechnology conferences 5. VCH, Weinheim
18. Luttmann R, Hartkopf J, RoB A (1994) Development of control strategies for high cell
density cultivations. Mathematics and computers in simulation 37:153-164
19. Luttmann R, Slamal H, Evtimova V, Berens M, Scheffler U, Elzholz 0 (1997) On-line de-
termination of reaction kinetics. Proceedings of the 8th European congress on biotech-
nology, Budapest, Hungary
20. Luttmann R, Bitzer G, Miiller D, Scheffler U, Friedriszik U (1995) Development of a my-
stat with online simulation methods. In: Schiigerl K, Munack A (eds) Postprint volume of
CAB 6 - International conference on computer applications in biotechnology, Garmisch-
Partenkirchen, Germany
21. Lee SY (1996) High cell density culture of Escherichia coli. TIBTECH 14:98-104
22. Riesenberg D, Schulz V, Knorre WA, Pohl HD, Korz D, Sanders EA, RoB A, Deckwer W-D
(1991) High cell density cultivation of Escherichia coli at controlled specific growth rate. J
Biotechnol 20: 17 - 28
 Part B
Application of General Principles for Reactor Models
7 Application of Computational Fluiddynamics (CFD)
to Modeling Stirred Tank Bioreactors
Matthias Reuss, Sven Schmalzriedt, Marc Jenne

7.1
Introduction
The stirred and aerated tank fermentor is still the most important type of bioreactor
for industrial production processes. The agitator or agitators are required to perform
a wide range of functions: adequate momentum, heat and mass transfer, and mixing
as well as gas-dispersion and homogenization of suspensions. As these various op-
erations make different demands, optimization of the individual tasks would result
in different design of the impellers. Therefore, agitators used in practice always re-
flect compromises. Conventional impellers used in fermentation are typically classi-
fied into axial and radial flow impellers. Examples of the two groups are illustrated
in Fig. 7.1. Of the many impeller geometries, the six-blade disk impeller (Rushton
turbine) with gas sparging below the impeller is most often used in standard confi-
gurations. However, an interest in using high-flow, low-power-number agitators such
as Intermig (Fig.7.2a), Lightnin A315 (Fig.7.2b), Prochem Maxflow T (Fig.7.2c),
and Scaba 6SRGT (Fig.7.2d) has been developed and improved performance of fer-
mentation processes has been reported with such alternative designs [1-3].
The physiological state of microorganisms and their related behavior (growth and
product formation) is determined by the immediate environment. During design of
a bioreactor, and in particularly during scale-up, an attempt is made to recreate this
environment in the large scale reactor as similar as possible to that established in
bench scale and/or pilot plant vessels. Engineering solutions to this problem include
maintaining geometric similarities whenever possible, and also criteria such as con-
stant power per unit volume, volumetric mass transfer coefficient, circulation time,
terminal mixing time, shear rate, tip speed, etc. (Table 7.1). The logic behind the
different criteria is that preserving each of these singly represents maintaining the
constancy of the corresponding characteristic of the extracellular environment at
different scales. And, if this environment property is the one that most critically
influences the desired microbial productivity, a successful scale-up might result. In-
deed, a number of microbial systems are in accord with these techniques and permit
a production scale operation reasonably in agreement with that established in the
laboratory. However, it is easy to see that these criteria are mutually exclusive and,
therefore, do not allow an exact replication of environmental similarity at any two
different scales. Under such circumstances, the behavior of microorganisms in dif-
ferent fermentors remains uncertain. As a matter of fact, the use of volumetric pro-
perties as scale-up criteria demands a guarantee of uniformity of properties
throughout the system, something that may be impossible even in a small vessel.
208 7 Application of Computational Fluiddynamics (CFD)

Radial Flow Impellers

DiSk Styte Flat Blade Turbine Swept back or Curved Blade Turbine
Commonly Referred to as (a Spiral Turblnel
the Rushton Impeller

Axial Flow Impellers

Propeller 45° Pitched Blade Turbine

Fig. 7.1. Impellers used in stirred tank bioreactors [58]

~ . .--- "
c5i€J "" , .... ~~

_ h ____ . )

Fig.7.2a-(. Low-power number agitators: a Scaba; b Lightning A 315; c Prochem Maxflow T

[3]
7.1 Introduction 209

Table 7.1. Common criteria for scale-up of stirred tank bioreactors

Volumetric oxygen transfer coefficient kLa

Volumetric power input P/V
Volumetric gas flow QdV
Impeller tip speed nd;
Agitation speed n
Terminal mixing time Boo

Feed
,
.. .....
;. . :: :J
I. ' ••

c:~ --+
,
mass & momentum
transfer

...,: : .
. . : : ..
Intracellular

Fig. 7.3. Illustration of interactions between extracellular environment and metabolism

Complex interactions between transport phenomena and reaction kinetics charac-

terize bioreactors and determine their performance. A quantitative description of
these phenomena should consequently rest upon the two interwoven aspects of
structured modeling (Fig. 7.3). The first aspect concerns the complex interaction of
the functional units of the cells, including the mathematical formulation of reaction
rates and the key regulation of these networks in response to changes in the envi-
ronment. The second aspect has to do with the structure of the abiotic phases of the
bioreactor in order to analyze the quality of mixing and other transport phenomena
such as mass transfer between the phases causing gradients in the concentrations of
various substrates and products.
These problems are particularly important for those processes in which nutrients
are continuously introduced into the broth. For specific nutrients such as oxygen
and sometimes other nutrients such as carbon source, the time constant for their
distribution (mixing-time) may be of the same magnitude as those of their con-
sumption in any reasonable sized reactor beyond the bench-scale. If we accept that
spatial variations exist, we are faced with the problem of dynamically changing en-
vironment conditions. This in turn may result in drastic changes in metabolism and
final outcome of the process. The long-term mathematical description of these phe-
2lO 7 Application of Computational Fluiddynamics (CFD)

nomena requires flexible tools to be adapted to different systems and able to inte-
grate the process and reactor.
Despite the strong interdependence, this chapter will concentrate on the detailed
modeling of the abiotic phases of the reactor. Various tools are available to tackle
problems of incomplete mixing. The models suggested for stirred tank reactors
may be classified into two groups - reactor flow models and computational fluiddy-
namics (CFD) including turbulence models.
The most important tools for modeling situations of incomplete mixing based on
reactor flow models are compartment models and recirculation models.
The application of the various approaches to bioreactor modeling has been exten-
sively described by Reuss and Bajpai [4] and Reuss [5]. As a result of these critical
reviews, several serious limitations of these modeling strategies have been outlined.
One of these limitations is related to the omission of backmixing of the gas phase
and, therefore, the disregard of the material balance equation for oxygen in the gas
phase. Second, in many of the models the number of compartments is connected to
the intensity of mixing and, thus, the structures are virtual in space. In other words,
the compartments are without identity because of missing coordinates. It is easy to
see that the two mentioned problems are closely related if attempts are made to
couple mixing of gas and liquid phase including local mass transfer across the inter-
face.
In contrast to these classical approaches of reactor flow models, this chapter sum-
marizes the state of the art in the application of computational fluiddynamics (CFD),
making use of the numerical solution of the state equations for mass, momentum,
kinetic energy, and energy dissipation. The application of this approach is illustrated
for single-phase and two-phase flow. Single as well as multiple impellers will be trea-
ted.
With the aid of some examples of industrial importance the use of the fluiddy-
namic simulations will be demonstrated. The first examples deal with the three-di-
mensional transient concentration field in mixing experiments at different scales of
operation. The process examples chosen are treating the problem of substrate distri-
bution during fed batch operation as well as oxygen distribution during the oxygen
sensitive microaerobic production of acetoin and butanediol with Bacillus subtilis.

7.2
Modeling and Simulation of Gas/liquid Flow in Stirred Tank Reactors

It is generally accepted now that Reynolds-averaging the turbulent Navier-Stokes

equations and modeling the Reynolds-stresses with an appropriate turbulence model
is a promising way of modeling the flow behavior. Ongoing development of commer-
cial computational fluiddynamics software (CFD) and increasing computer power
are continuously improving the conditions for the simulation of the three-dimen-
sional and turbulent flow structure in stirred tanks. Among the variety of impellers
the Rushton turbine is well established for many tasks, mainly due to good gas dis-
persion and mixing of liquids with low viscosities. The Rushton turbine generates a
flow leaving the impeller in radial and tangential direction. This radial-tangential jet
flow divides at the vessel wall, and the flow then recirculates back into the impeller
region. Besides turbulent dispersion, recirculation of the flow is the main reason for
the mixing capability of stirred tanks.
7.2 Modeling and Simulation of Gas/Liquid Flow in Stirred Tank Reactors 211

In spite of the improved hardware and software, which have greatly expanded the
tools available for simulating fluid flow in stirred tank reactors, a number of un-
solved problems and open questions still exist.
A critical analysis of the many publications concerning the simulation of the li-
quid flow in baffled stirred tank reactors equipped with a Rushton turbine shows
several discrepancies. The most important differences between the simulations con-
cern the dimensionality of the simulations (three-dimensional or axisymmetric),
turbulence modeling, the modeling approaches for the Rushton turbine as well as
the accuracy of the numerical predictions, which depends on the grid size.
In what follows, different modeling approaches for the Rushton turbine are exam-
ined and critically reviewed. For more than a decade (see, e.g., [6-8]) it has been an
established method to specify boundary conditions for the impeller with the aid of
experimental data. One advantage of this method is the reliable description of the
outflow region of the impeller, which can essentially be considered as a circumfer-
entially and time averaged radial-tangential jet. A resulting additional advantage is
the reduced computational expense of stationary simulations compared to transient
simulations. The resolution of the vortex system behind the stirrer blades (see, e.g.,
[9]) in applying this method, however, is not possible. In order to specify boundary
conditions for other types of impellers, one has to perform time-consuming experi-
ments in advance. To remove the two last disadvantages mentioned, recent attempts
have been made to simulate the unsteady flow within and outside the impeller swept
region in applying the so-called sliding-mesh technique (see, e.g., [10, 11]). A critical
comparison of the results from the sliding mesh technique and simulations with
measured data in the impeller region has been presented by Brucato et al. [12]. How-
ever, the sliding mesh technique requires excessive computational resources and for
most engineering applications knowledge of the full time varying and periodic flow
field may not be necessary. Another possibility to simulate flow details between the
impeller blades is the so-called snapshot approach (see, e.g., [26]). This is often also
called a multiple reference frame method (see [13]). Experimental data to specify
boundary conditions are not necessary. An advantage compared with the sliding-
mesh technique is that the full time dependent transport equations need not be
solved. This offers an interesting and promising approach. However, the essential
comparisons with experimental observations are lacking.
Besides the modeling approaches for the Rushton turbine, the dimensionality of
the simulation is another point of discussion. Table 7.2 shows an overview of publi-
cations concerning simulations of the liquid flow in stirred tank reactors using ex-
perimental data as impeller boundary condition for the Rushton turbine. The fourth
column of Table 7.2 informs about the dimensionality of the simulations. In axisym-
metric simulations the stationary baffles fixed at the vessel wall are not resolved as
geometrical bodies, but modeled with the drag resistance force of the baffles to
match the experimentally observed velocity fields (see, e.g., [15]). The coefficient
required in the formulation of the drag resistance force lacks physical meaning and
needs to be adjusted. Axisymmetric simulations are then able to describe the experi-
mentally observed bulk circulation in radial-axial direction as the most important
flow characteristic in the stirred tank (see, e.g., [15, 17]). Increasing computer power
has continuously improved the conditions to resolve the flow field in three dimen-
sions. Three-dimensional simulations show flow details in the vessel (e.g., vortex
formation behind the baffles), which cannot be observed in axisymmetric simula-
tions.
212 7 Application of Computational Fluiddynamics (CFD)

Table 7.2. Summary of fluiddynamic simulations of stirred tank vessels

Author Year turbulence model dimensionality

Single-phase flow
Platzer [14] 1981 standard k-£O axisymmetric
Harvey and Greaves [IS] 1982 standard k-£O axisymmetric
Placek et al. [16, 17] 1986 three-equation kp-k re axisymmetric
Middleton et al. [18] 1986 standard k-£O 3D
Ju [6] 1987 modified k-£O 3D
Joshi and Ranade [7] 1990 standard k-£O 3D
Kresta and Wood [8] 1991 standard k-£O 3D
Bakker and Van den Akker [19] 1994 algebraic stress 3D
Togatorop et al. [20] 1994 standard k-£O axisymmetric
Brucato et al. [12] 1998 standard k-£O 3D
Two-phase flow
Issa and Gosman [21] 1981 standard k-£O axisymmetric
Tragardh [22] 1988 standard k-£O axisymmetric
Politis et al. [23] 1992 standard k-£O 3D
Morud and Hjertager [24] 1993 standard k-£O axisymmetric
Bakker and Van den Akker [25] 1994 algebraic stress 3D
Ranade and Van den Akker [26] 1994 standard k-£O 3D

The third column of Table 7.2 concerns the choice of the turbulence model. Most
of the authors use the standard form of the so-called k-f, model, although it is well
known that this approach of turbulence modeling may fail in flow regimes with
strong streamline curvature and vortex generation. Some authors use modifications
of the k-f, model or other turbulence models like the three-equation kp-ky-f, model or
the algebraic stress model.
An important feature in modeling the two-phase flow is to distinguish between
Eulerian and Langrangian approaches. In the Langrangian approach, the continuous
phase is treated as a continuum while the dispersed gas bubbles are modeled as
single particles. In the Eulerian approach the dispersed phase is also considered as
a continuum resulting in the so-called two fluid model. Only the Eulerian approach
has been considered for aerated stirred tank reactors so far. If only gravitation, pres-
sure, and drag force are taken into account in the momentum equation for the gas
phase, the relative velocities of the gas phase are calculated from algebraic equations.
This is the so-called algebraic slip model. The disadvantage of this simple approach
is the fact, that additional interface forces are neglected. Issa and Gosman [21] cal-
culated the flow in a gassed and stirred vessel equipped with a Rushton turbine using
the algebraic slip model. Furthermore, they used very coarse grids because of limited
computing power. Experimental verification of their simulations was not shown.
Tragardh [22] reported about 2D-simulations with the algebraic slip model for a
stirred vessel equipped with three impellers. Politis et al. [23] performed three di-
mensional simulations with the two fluid and k-f, model. They considered different
interfacial forces and critically examined their influence. These authors were able to
show that, in addition to the drag force, the virtual mass force particularly needs to
be considered. For boundary conditions in the impeller region, values for averaged
tangential velocities as well as k and f, from measured data were used.
7.3 Single Phase Flow 213

Morud and Hjertager [24] followed an axisymmetrical approach based on the two
fluid and k-£ model. The virtual mass force was neglected. These authors observed a
considerable deviation between measured and simulated data.
Bakker and Van den Akker [25] proposed a single phase model in which a reduc-
tion of pumping capacity due to aeration was taken into account. The turbulence was
modeled with the aid of an algebraic stress model. A more detailed model for the
flow in the impeller region was presented by Ranade and Van den Akker [26]. These
authors used the two-fluid approach and modeled the two-phase flow in the impeller
region with the aid of a snapshot method.

7.3
Single Phase Flow

7.3.1
Transport Equations

The transport equations describing the instantaneous behavior of turbulent liquid

flow are three Navier-Stokes equations (transport of momentum corresponding to
the three spatial coordinates r, z, cp in a cylindrical polar coordinate system) and a
continuity equation. The instantaneous velocity components and the pressure can be
replaced by the sum of a time-averaged mean component and a root-mean-square
fluctuation component according to Reynolds. The resulting Reynolds equations and
the continuity equation are summarized below:

O(pUi) O(pUiUj) __ ~ ( .. ' ') _ op . (7.1 )

'"
ut
+ '"
UXi - UXi
'" T'J + puiuj UXi'" + pg,

op a (7.2)
~+-(pUi) = 0
at OXi
A reasonable compromise for model accuracy and computational expense are
eddy viscosity models relating the individual Reynolds stresses to mean flow gradi-
ents:
, , (OUi OUj) 2 (7.3)
puiuj = -PVturb OXj + OXi +"3 pOijk

where Vturb is the turbulent eddy viscosity. The transport of momentum, which is
related to turbulence, is thought of as turbulent eddies, which, like molecules, collide
and exchange momentum.
The family of two-equation k-c models is the most widely used of the eddy vis-
cosity models. A k-c model consists of two transport equations, one for the turbulent
kinetic energy k and one for the energy dissipation rate c. The turbulent eddy vis-
cosity is calculated from

(7.4)

where CJ.l is a parameter which depends on the specific k-c model.

The standard k-c model, as presented by Launder and Spalding [27], is by far the
most widely-used two-equation eddy viscosity model, also for modeling turbulence
214 7 Application of Computational Fluiddynamics (CFD)

Table 7.3. Parameter values in the k-£ model

Parameter Value

C"m,s 0.09
C1,s 1.44
C2,s 1.92
ch,s 1.00
(J'E,S 1.314

in stirred tank reactors (see Table 7.2). The popularity of the model and its wide use
and testing has thrown light on both its capabilities and its shortcomings, which are
well-documented in the literature [27-33). For high turbulent Reynolds numbers, the
model may be summarized as follows:

o(pk) 0
--+-(puik) 0 ( veff Ok)
= - p - - +p(Pk- E) (7.5)
ot OXi OXi IJk,S OXi

O(pE) 0
--+-(pUiE)
ot OXi
= -
0 (p
OXi
veff
--
IJE,s OXi
8c) + p (CI,S-Pk
E
k
E )
- CZ,s-E
k
(7.6)

The model parameters of the standard k-E model are listed in Table 7.3.
The production of turbulent kinetic energy Pk is modeled with the aid of the eddy
viscosity hypothesis:

(7.7)

The dissipation rate E can be regarded as the rate at which energy is being trans-
ferred across the energy spectrum from large to small eddies. The standard k-E mod-
el assumes spect~al equilibrium, which implies that once turbulent kinetic energy is
generated at the low-wave-number end of the spectrum (large eddies), it is dissi-
pated immediately at the same location at the high-wave-number end (small eddies).
In other words, the standard k-E model assumes that Pk is near to E. As far as the
stirred vessel is concerned, this is a very restrictive assumption, because there is a
vast size disparity between those eddies, in which turbulence production takes place
(mainly at the stirrer), and the eddies, in which turbulence dissipation occurs.
The standard k-E model employs a single time scale Td=klE called dissipation
range time scale in the equation to characterize the dynamic processes occurring
in the energy spectrum. Thus, Eq. (7.6) can be rewritten as:

O(pE)
ot
+ ~ (pUi E) = ~ (p veff
oX i OXi
8c) +
IJE,s OXi
p (CI,S Pk - CZ,S~)
Td Td
(7.8)

The energy spectrum, however, comprises fluctuating motions with a spectrum of

time scales, and a single time scale approach is unlikely to be adequate under all
circumstances. Consequently, the model has been found to perform less satisfactorily
in a number of flow situations, including separated flows, streamline curvature,
swirl, rotation, compressibility, axisymmetrical jets etc.
7.3 Single Phase Flow 215

Because of its wide use, variants and ad-hoc modifications aimed at improving its
performance abound in the literature. The most well-known modifications are the
Chen-Kim and the RNG variant of the k-E model.
In order to ameliorate the previously mentioned deficiencies in the standard k-E
model, Chen and Kim [34] proposed a modification which improves the dynamic
response of the 10 equation by introducing an additional time scale
k
Tp=- (7.9)
Pk
which is called the production-range time scale. The final expression of the trans-
port equation for the dissipation rate is given as
1st part Znd part )
~~
a(pE) a a Veff & Pk Pk 10
- - + - (pUi E) = - (p----) + P ( CI,CK - + C3,CK - -CZ,CK- (7.10)
at aXi aXi CYE,CK aXi Td Tp Td
" v J

production term

The parameters of the Chen-Kim model are summarized in Table 7.4.

The first part of the production term corresponds with the production term of the
standard k-E model. Notice that the second production term is related to the time
scale Tp. The introduction of this additional term enables the energy transfer to
respond more efficiently to the mean strain than does the standard k-E model. Thus,
T p enables the development ofa field of 10 suppressing the well-known overshoot
phenomenon of the turbulent kinetic energy k. This overshoot appears when the
standard k-E model is applied to flow conditions with large values of mean strain
(see [29,32,33]).
The modification may be summarized as follows. Production of 10 appears in two
energy fluxes divided by two different time scales T d and T p. The multiplying coeffi-
cients might be seen as weighting factors for these two energy fluxes. One may ex-
pect that this feature offers advantages in separated flows and also in other flows
where turbulence is far from local equilibrium (P k is far from E). If Pk is near to 10
(local equilibrium) the Chen-Kim modified k-E model is almost identical to the stan-
dard k-E model. Then, T p equals T d and summing up the two 10 production terms
leads to the 10 production term of the standard k-E model. The resulting coefficient
CI,CK+C3,CK=1.4 is only slightly lower than CI,S' This is the reason why for simple
boundary type flows the Chen-Kim modified k-E model gives results similar to those
predicted by the standard k-E model. However, for complex elliptic turbulent flow
problems (internal turbulent recirculating flows) involving rapid changes of turbu-

Table 7.4. Parameter values in the Chen-Kim model

Parameter Value

Cp"CK 0.09
Cl,CK 1.15
C 2,CK 1.90
C 3,CK 0.25
(h,CK 0.75
(JE,CK 1.15
216 7 Application of Computational Fluiddynamics (CFD)

lent kinetic energy production and dissipation rates the Chen-Kim modified k-E
model has been shown to give much better results than the standard k-E model [34].
To improve further the agreement between simulations and experimental observa-
tions the ratio of two parameter Cj,CK/C3,CK in the Chen-Kim model was slightly
modified. For modification of these parameters the ratio of the Eulerian macro
length scale Lj to the impeller blade height w have been employed. The property
can be compared in geometric similar vessels.
According to Batchelor's [35] energy cascade theory the energy dissipation rate E
is given by

(7.11 )

Introducing the kinetic energy k, Eq. (7.11) reads

(7.12 )

Wu and Patterson [36] normalized their own observations as well as results from
other observations with the blade height w:

(7.13 )

If the macro length scale for anisotropic turbulence L res = vI: L; is replaced by
the length scale for isotropic turbulence the following equation holds:

L A k 3/ 2
(7.14)
W V3 E

For the estimation of the improved parameters in the turbulence model, the simu-
lated values of k and E are used to predict the local values of Lw-I, which are then
compared with the experimental observations reported by Wu and Patterson [36] in
the impeller region. For more details, the interested reader is referred to the original
paper of Jenne and Reuss [37]. At conditions of spectral equilibrium the dynamic
response of the optimized E transport equation should always be identical to that of
the standard k-E model. Thus, the sum of the two parameters of the Chen-Kim mod-
el is constrained by

(7.15)

Only one parameter has been varied. The values obtained for the two parameters
at minimal deviations between computed and measured length scale in the impeller
region are

and

(C3,CK)Opt =0.04
7.3 Single Phase Flow 217

7.3.2
Simulations and Comparison with Experimental Observations

The Reynolds equations, the continuity equation, which is turned into an equation
for pressure correction (see [38]), and the transport equations for the turbulence
quantities k and £ are integrated over the respective finite volume elements resulting
from the discretization of the stirred tank domain. The convection and diffusion
terms in the transport equations are approximated using the hybrid-scheme of Pa-
tankar [38]. The resulting algebraic equations are then solved with the aid of the
commercial CFD software Phoenics (Version 2.1). So-called false-time-step relaxa-
tion is used in order to achieve stationarity. The semi-implicit method, which con-
siders the pressure-link of the pressure correction equation and the Reynolds equa-
tions, is the Simplest algorithm. The sets of algebraic equations for each variable are
solved iteratively by means of the AD! technique.
With the aid of a systematic comparison of simulations and experimental data
obtained by Costes and Couderc [39] the performance of the following turbulence
models has been analyzed:
- standard k-£ model
- optimized Chen-Kim k-£ model
- Chen-Kim k-£ model
- RNG k-£ model (renormalized group method, Yakhot and Smith [40])

Figure 7.4 demonstrates that in some positions within the tank it is easy to discrimi-
nate between the various k-£ models on the basis of a comparison with measured
mean velocities. Thus, the profiles of axial mean velocities belonging to the baffle
plane ('P=4S0) in Fig 7.4, show significant deviations for the velocities obtained with
the aid of the Chen-Kim and the RNG k-£ model. The standard and the optimized
Chen-Kim k-£ model fit the experimental data well.
However, the differences in performance of the four k-£ models are even more
pronounced by inspecting Figs. 7.5 and 7.6. Figure 7.5 concentrates on the outflow
of the impeller (radial-tangential jet flow) exactly between two baffles ('P=O 0). Fig-
ure 7.Sa shows radial profiles of radial mean velocities at the impeller centerline and
at the height of the lower impeller blade edge. The optimized Chen-Kin k-£ model
fits the experimental data well, whereas the other k-£ models deviate significantly.
The values obtained with the Chen-Kim and RNG k-£ model at the impeller center-
line are higher than the measured data. The standard k-£ model overestimates the
energy loss in the outflow of the impeller. Consequently, the values of the predicted
radial mean velocities are too low. The values of radial mean velocities obtained with
the Chen-Kim and the RNG k-£ model at the height of the lower impeller blade edge
are lower than the measured ones, because the predicted turbulent shear stress does
not drag the entrainment flow sufficiently. The standard k-£ model overestimates the
turbulent shear stress and consequently, the values of the radial mean velocities are
too high. Figure 7.sb shows radial profiles of tangential mean velocities; again at the
impeller centerline and at the height of the radial velocity components equally apply
to the tangential components.
Figure 7.6 shows radial profiles of tangential mean velocities at different heights in
the bulk of the vessel. Again, the optimized Chen-Kim k-£ model fits the experimen-
tal data well, whereas the other k-£ models deviate significantly. Figure 7.7a illus-
trates the flow field between two baffles at 0°. One can see the flow leaving the im-
218 7 Application of Computational Fluiddynamics (CFD)

0.'

'.1

.0 ~,; ..• 0.0 ]

....
,;
...,
.. ~

lLEJ 0.3

0.25
.,
O~

0.' ... 0.1

§"
,; O~
0.0

".1
]
,;
,1m) ~.

0.25

.,
] 0.0
0;
§
,;'

....
0.2
,;'
, 1m) ".1

Optimized Chen-Kim k-e

Standard k-e
Chen-Kim k-e
RNG k-e

a b
Fig. 7.4a,b. Radial profiles of axial mean velocities in the bulk of the vessel. Simulated profiles
obtained with four different k-£ models. Measurements from [39]: a between two baffles
(47=0°); b baffle plane (47=45°)

peller in a radial direction and the subsequent entrainment flow into the outflow of
the impeller. Thus, the flow divides at the vessel wall and recirculates back either
directly into the impeller or again as entrainment flow into the outflow of the im-
peller. The aim of Fig. 7.8 is to demonstrate that the flow in the T-tp direction is in
fact three-dimensional. Figure 7.8a shows the flow field at a short distance below the
7.3 Single Phase Flow 219

..75
75

...
...S

"
]
;i
0.0
M I:!
, ImJ ,ImJ

0.0 0.(15

'Iml , Iml

Optimized Chen-Kim !toe

S\andard k-E
Chen-Kim k~

RNG k-e

a b
Fig.7.5a,b. Radial profiles of: a radial; b tangential mean velocities in the outflow of the im-
peller between two baffles (<p=OO). Simulated profiles obtained with four different k-E models.
Measurements from [39]

impeller. The flow is directed to the vessel wall and has a positive tangential velocity
component like the radial-tangential jet originating from the impeller. The radial-
tangential jet is able to drag the flow in its vicinity due to turbulent shear stress.
Again, the impeller tip velocity Utip is shown as reference vector and indicates high
velocities. Behind the baffle, vortex formation can be observed.
Figure 7.8b illustrates the flow field at the height of the lower bulk circulation cen-
ter. The reference vector indicates lower values of velocity. The flow, originally direc-
ted to the vessel wall above the lower bulk circulation center (see Fig. 7.7a), is turned
to the bottom of the vessel. This is the effect of the baffles. And the flow, originally
directed to the vessel axis below the lower bulk circulation center, is turned to the
impeller region. Consequently, the larger amount of momentum is associated to the
flow directed in negative and positive z direction, respectively, and not in r-cp direc-
tion. The vortex behind the baffle still exists. The vortex occupying the vessel axis
220 7 Application of Computational Fluiddynamics (CFD)

Q.l

0.05
u;
.....
o.Q §
0.2
=;
-0.05

0.1

u;
....
0.0 §
"
.
r [m! -0.1

u;
....
0,0 §
,,'
-0.1

-Q.2

Optimized Chen-Kim k-£

Standard k-.
Chen-Kim k-.
RNG k-e

Fig. 7.6. Radial profiles of tangential mean velocities in the bulk of the vessel between two
baffles ('P=OO). Simulated profiles obtained with four different k-£ models. Measurements from
[39]
7.3 Single Phase Flow 221

U 11 110 :: 1,28 m/s

~

,117'1 1

Fig. 7.7. Flow field between two baffles ('P=OO)

with a slight angular motion in opposite direction to the impeller rotation (see [39,
41] can also be observed.
Figure7.8c shows the flow field near the bottom of the vessel. The flow accumu-
lates in the high pressure region in front of the baffle. The main part of the flow has
to leave the baffle with a negative tangential velocity component. A smaller part of
the flow evades in a low pressure region with a positive tangential velocity compo-
nent. The smaller part meets the main part of the flow leaving the baffle, which is
shifted 90° in positive cp direction, in a sharp angle (see [76]). The two parts of the
flow join each other and form the basis of the vortex occupying the vessel axis. Due
to the prevailing negative tangential momentum in the united flow the vortex obtains
its slight angular motion in the opposite direction to the impeller rotation. For more
details (distribution of pressure and turbulence properties) the interested reader is
referred to the original publication ofJenne and Reuss [37] as well as Jenne [42].
222 7 Application of Computational Fluiddynamics (CFD)

U [III' = 1.28 m/s

---+

a no 210
'"

U tI ,. - 1.28 m/s

'00 .0
"

b 'HoG 270 :210

7.4 Multiple Impellers 223

~ ' .28 m/s

•
U li lll

,1)0 aD ID

c 2SO 710 210

Fig.7.8a-c. Flow field in r-!.p direction: a at short distance below the impeller; b at the height of
the lower bulk circulation center; c near the bottom of the vessel

7.4
Multiple Impellers

In industry, reactors are usually equipped with two or more impellers. Very little
data are found concerning the details of flow patterns, particularly quantitative in-
formation about velocity fields and distribution of turbulence intensities. However,
many workers have investigated the effect of different impeller types and configura-
tions on mass transfer gas liquid and mixing. Important and quite useful results have
been summarized by Bouaifi et al. [43] and John et al. [44]. Improved reactor per-
formance has been observed when incorporating mixed flow systems (e.g., with low-
er impeller acting radially and the upper impeller axially) in a baffled system [45-
47]. In particular for large scale fed batch fermentations these configurations should
offer advantages because of improved axial mixing. A few CFD simulations together
with simulations of mixing behavior presented in Sect. 7.6.2 will serve to elucidate
these phenomena.
To reduce complexity and computation time, two-dimensional simulations have
been performed for this comparison. In these simulations the baffles are modeled
as a momentum sink in the Reynolds equation for tangential direction [15]. The
assumption that these simplified simulations are able to reasonably approximate
the radial-axial flow behavior seems to be justified because of the large height-to-
diameter ratio.
224 7 Application of Computational Fluiddynamics (CFD)

U II, " 5.13 m/s U .. , "'"' 5.13 m/s

."",
a c
Fig.7.9a-c. Flow field for different multiple impeller systems: a 4 Rushton turbines; b 2 Rush-
ton turbines and 2 pitched blade impellers; c 4 pitched blade impellers

Fig. 7.10. Pitched blade axial-flow impeller

In Fig. 7.9a velocity fields are shown for a system offour Rushton turbines. In addi-
tion to the velocity vector field, large arrows are used to illustrate the flow behavior.
Each impeller creates a more or less independent symmetrical flow field. The multiple
impeller system therefore shows very poor axial convection. The transport between
the individual cells is performed mainly with the aid of axial turbulent dispersion.
The results from similar simulations with two Rushton and two pitched blade im-
pellers as well as four pitched blade impeller are shown in Fig.7.9b and c, respec-
7.5 Gas-Liquid Flow 225

tively. The pitched blade axial-flow impeller is shown in Fig. 7.10. In both cases an
improved convection in axial direction can be observed. The results of simulated
mixing experiments presented in Sect. 7.6.2 will confirm more rapid mixing for both
systems.

7.S
Gas-Liquid Flow

The simulations of the gas-liquid flow are based on the Eulerian two-fluid model
originally derived by Ishii [50]. In this approach, each phase is treated as a conti-
nuum. After averaging the general transport equations, we get the following set of
multi-phase conservation equations [23, 26]:
CONTINUITY:

a a ( Vturb aCk)
-(Pkcd +- PkCkUk,j - Pk---- =0 k=L,G (7.16)
at aXi SCturb aXi

A dispersive transport of gas bubbles and liquid has been considered in both con-
tinuity equations. SCturb is the Schmidt number for turbulent transport which is as-
sumed to be 1 [26]. The global mass conservation is given by:
(7.17)
MOMENTUM:
Liquid phase:

a(PLCLUL,i) a(PLCLUL,iULj) _ _ ~ (.. I ,)_

!C\
ut
+ !C\
uXj
- cL!C\
uXj
TL,t) + PLUL,jU Lj
(7.18)
ap
- cL",
uXj
+ PLcLgi + Si
with the laminar shear stress 'tL,ij and turbulent Reynolds-stresses given by the Bous-
sinesq approximation:

I I
-PLUL,iUL;j = PLVturb
(a UL,i
a Xj + aaULj)
Xi - '23 PLoijk (7.19)

Gas phase:

a(PGcGuG,i)
---.0..----,_-'--'- + a(PGcGuG, iUGj) = -CG -
ap
+ PGcGgi - Sj (7.20)
at aXj OXj

Reynolds-stresses in the gas phase can be neglected.

7.S.1
Interfacial Forces

The interfacial coupling term Si in Eqs. (7.18) and (7.20) is a linear combination of
several forces. Politis et al. [23] have compared the order of magnitude of the various
forces and concluded that only the drag force and the virtual mass force need to be
considered.
226 7 Application of Computational Fluiddynamics (CFD)

7.S.1.1
Drag Force

From the definition of the drag coefficient CD of a single bubble the following expres-
sion for the drag force can be derived:
PL
PD,i = 2CdAbl~ul~Ui (7.21)

Ab is the sectional area of the bubble=(O.251t)d~, !:..Ui is the relative velocity be-
tween the bubble and the liquid in the direction i, l!:..ul is the absolute value of the
relative velocity vector. The momentum equation (Eq.7.18) is related to the total
volume dV which contains gas and liquid. The volumetric force Si is therefore
Pi Pi
S'=-=cG-- (7.22)
I dV dVG
and with Eq. (7.21):
3 PL
SD,i = CG 4cD db l~ul~Ui (7.23)

The following correlations for air bubbles rising in distilled and tap water have
been proposed by Kuo and Wallis [49). For distilled water the equations for the drag
coefficent read:
Re<0.49 CD =24 Re- 1
0.49<Re<33 CD =20.68 Re- O.643
33<Re<661 CD =72 Re- 1
661<Re<1237 and We~4 CD =O.02083 Re4 Mo
Re> 1237 and We<8 CD =O.125 We
For tap water Kuo and Wallis [49) proposed the following equations:
Re<0.49 CD =24 Re- 1
0.49<Re<100 CD =20.68 Re- O.643
100<Re<717 CD =6.3 Re- O. 385
Re» 717 and We<8 CD =WeI3

The dimensionless numbers in these equations are defined by:

Reynolds number Re=PLI,:;uldb
I"L
Weber number We=PLI,:;ul'd b (J

Morton number Mo (PL-PG)gl"i

PE(J"3

Making use of the correlation proposed by Ishii and Zuber [48) for correction of the
drag force in a bubble swarm, the drag coefficient CD is multiplied by the correction
factor!k which is given by:

fi
k
= (1 + 17.67
18.67/
10) 2
(7.24)

with
~j),L
/ = y1-cG- (7.25)
j),m
7.5 Gas-Liquid Flow 227

and

(7.26)

for db> 1.8 mm. For bubble diameter smaller than 1.8 mm the Reynolds number is
calculated with {trn from Eq. (7.26).

7.5.1.2
Virtual Mass Force

The virtual mass force represents the force required to accelerate the apparent mass
of the surrounding continuous phase in the immediate vicinity of the gas bubble.
Drew and Lahey [51] have proposed the following formulation:
(7.27)

(7.28)

The virtual volume coefficient Cvrn for potential flow around a sphere is 0.5. For
ellipsoidal bubbles with a ratio of semiaxes 1:2 Cvrn is 1.12. For ellipsoidal bubbles
with random wobbling motions Lopez de Bertodano et al. [52] calculated Cvm to be
about 2. In addition, Cvm is a function of the specific gas hold up [77-79]:
(7.29)

with
Cvm =Cvma (1- 2.78 min [0.20, £G]) (7.30)

and 0.5<Cvma <:;2.0.

In context with the discussion on the effects of the virtual mass force it is worth-
while to insert a short comment on the so-called algebraic slip model. The momen-
tum equation of the gas phase can be simplified if the inertial forces are neglected
and only drag is considered as interfacial force. Equation (7.20) then becomes
op
o= -CG ~
UXi
+ PGcGgi - SD,i (7.31)

The relative velocities between gas and liquid phases are then calculated from
simple algebraic equations. The computing effort is therefore similar to one phase
simulations which makes the algebraic slip model very attractive. However, it is not
possible to include the virtual mass force, because the algebraic characteristic disap-
pears if derivatives for gas and liquid velocities, Eq. (7.28), are introduced.

7.5.2
Turbulence Model

For applications in two-phase flow the k-£ models have been modified in different
ways. One possibility is to insert additional sources into the transport equations for
k and £ [53-56]. An alternative is to consider an increase of the turbulent viscosity
228 7 Application of Computational Fluiddynamics (CFD)

in the liquid phase caused by the bubbles. According to Sato et al. [57] and Lopez de
Bertodano et al. [52] this effect can be described by

k2
Vturb = c", ~ + C""bdbEGIb.ul (7.32 )

For the parameter C""b Lopez de Bertodano et al. [52] suggested a value of 0.6.
Assuming that the optimized version of the Chen-Kim model is still valid, the trans-
port equations for the turbulence quantities k and £ for the two phase system are
given by

O(PLELk) a a ( veff Ok)

---"-=:-::"""':"+-(pLELUL,ik) = - PLEL-- +
at OXi OXi O"k,s OXi
(7.33)

and

(7.34)

with

(7.35)

7.5.3
Impeller Model

An important effect of the aeration is the reduction of the pumping capacity (volu-
metric rate ofliquid leaving the impeller in radial direction). As already discussed in
context with the single phase simulations, again the boundary conditions at the im-
peller are predicted from measured data of the averaged velocities. The first step
towards estimation of these velocities is a reasonable estimate of the pumping capa-
city at non-aerated conditions. Taking into account an averaged value from the large
variations in pumping capacities for Rushton turbines and the strong influence of
the impeller off-bottom clearance [58], a pumping capacity of QL,0=1.53nd3 i was es-
timated for the impeller configuration used by Bombac et al. [59,60]. The data from
these investigations will be used for the following comparison between measured
and simulated values of the specific gas hold up (see Sect. 7.5.4).
The quantitative estimation of the effect of aeration is even more difficult. Rousar
and Van den Akker [61] reported the following correlation between power consump-
tion and pumping capacity for aerated impellers:

( OL,G) = (PG) 0.341

(7.36)
QL,o 1 Po
7.5 Gas-Liquid Flow 229

which is close to the effect on the circulation time for low values of the aeration
number reported by Reuss and Bajpai [4]. On the other side, Joshi et al. [62] recom-
mended

(OL)
QL',:
(P)
P: 2=
1.0
(7.37)

Due to these uncertainties, it was decided to fit this parameter from the compar-
ison between measured and simulated values of the specific gas hold up with the
constraint

(OL,c)
QL,O
< (QL,C)
QL,O
1 estimated
< (OL,c)
QL,O 2
(7.38)

7.5.4
Simulation Results

The simulations discussed in the following have been performed for the vessel and
impeller geometries used by Bombac et al. [59, 60] in their systematic investigations
of the distribution of specific gas hold-up at different speed of agitation. These mea-
surements were performed by using conductivity sensors. Table 7.5 summarizes the
geometrical properties of the system as well as operation conditions (aeration and
agitation).
For the prediction of the interfacial forces (Sect. 7.5.1) it is necessary to estimate a
representative bubble diameter. From measurements of bubble size distributions in
stirred vessels reported by Barigou [63, 64] for coalescing and non-coalescing sys-
tems, an average value of the bubble diameter was calculated with

L qidb,i
nb

dlO = (7.39)
i=1

where qi=relative number fraction and db,i=bubble diameter in class i. Additionally,

the Sauter mean diameter was predicted according to

L qid~,;/ L qid~,i
nb nb

d32 = (7.40)
i=1 i=1

For the bubble size distribution measured by Barigou [63, 64J a value of

dlO = 0.67 (7.41)

d32

Table 7.5. Geometrical and operational parameters in the investigations of Bombac [59,60]

Nr. N Q PdPo QL,dOL,o

min-I m3 S-I Eq. (7.36) Eq. (7.37) Estimated
1 376 5.56x 10-4 0.77 0.91 0.77 0.91
2 376 1.67 x 10- 3 0.43 0.75 0.43 0.64
3 266 1.67x 10-3 0.48 0.78 0.48 0.62
Dy=0.44 m, H Dy-'=l, d; Dy-'=0.33, off-bottom clearance/liquid height C 11'=0.25
230 7 Application of Computational Fluiddynamics (CFD)

Table 7.6. Predicted bubble diameters for the three operation conditions of Bombac [59, 60]

Nr. N Q Po Pc/Po PG Pc/VL cG,Ca d 32 diO

[min-I] [m3 S-I] [W] [-] [W] [Wm- 3 ] [%] mm mm

367 5.56x 10-3 99 0.77 76 1058 2.7 3.1 2.1

2 376 1.67 x 10-3 99 0.43 42 591 6.0 5.0 3.4
3 266 1.67x 10-3 35 0.48 17 235 5.4 6.5 4.4

was calculated. To estimate bubble diameters for the operation conditions in the
work of Bombac [59, 60], in the first step the Sauter mean diameter (in cm) was
predicted from the well known Calderbank equation [65]:

(J0.6 )
d32 =4.15 ( 04 cGCa°.5+0.09cm (7.42)
(PG/Vr) . p2· 2 '

with

cG Ca =
,
( 0-
UG cG,Ca )0.5 +0.0216 (
Ut
06 )-1(0)0.5
(P /vd.4 p2·
uG (J .
2 Ut
(7.43 )
G

For the rise velocity Calderbank assumed a constant value of 0.26 m S-I.
The information necessary to apply Eq. (7.42) for the three operation conditions
from the work of Bombac [59,60] are summarized in Table 7.6.
Results from the simulations performed with the software package Phoenics are
shown in Figs.7.11-7.14 [42]. Figure7.11a,b shows exemplary simulated flow velo-

u " ... z..e5, JOI.

---+

...
Fig. 7.11. Simulated flow field of the gas and liquid phase for operation condition 2 (Table 7.5)
7.5 Gas-Liquid Flow 231

EC [%]

v.

0,Q5 0.. D.. D.

rlml

Fig. 7.12. Comparison between simulated (left) and measured (right) local value of the gas-
hold up at operation condition 1 of Bombac [59, 60)

Table 7.7. Comparison of integral values for the specific gas hold up

Operation lOG,c. Eq. (7.43) [%) lOG Bombac [59,60) EG, Simulations [%)
condition No. experimentally observed [%)

1 2.7 2.2 1.8

2 6.0 4.2 4.0
3 5.4 3.3 4.3

cities for the gas and liquid phase. Figures. 7.12-7.14 summarize comparisons be-
tween simulated and predicted local values of the specific gas hold up. Figure 7.15
shows contour plots of volume fractions at three different angles between the baf-
fles as well as at two different levels of the tank. In total, the simulated values of
the local gas hold up show a satisfactorily agreement with experimental data for
the three different operation conditions. Table 7.7 shows the comparison of the
232 7 Application of Computational Fluiddynamics (CFD)

fG [%]

0.0 O.OS 0. 1 0. 15 0.2

, 1m.

Fig. 7.13. Comparison between simulated (left) and measured (right) local value of the gas-
hold up at operation condition 2 of Bombac [59, 60]

integrated value of the specific gas hold up. With the exception of operation con-
dition 3 again, a reasonable agreement is observed. A careful analysis of these
deviations illustrates that the intensity of gas recirculation is slightly overpredicted
in the simulations. This is probably caused by the uncertainties in the prediction
of the average bubble diameter.

7.6
Application of CFD to Simulations of Mixing and Biotechnical Processes

7.6.1
Methodology

The approach used for the application of CFD is illustrated in Fig. 7.16. It is based on
the assumption that the stationary flow field is not affected by mass transfer and
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 233

fG [%1

.,. 0,0.5 . ,1 0.'5

r Iml

Fig. 7.14. Comparison between simulated (left) and measured (right) local value of the gas-
hold up at operation condition 3 of Bombac [59,60]

reactions. This is a reasonable assumption for many biotechnical processes with

Newtonian flow behavior. Also the depletion of oxygen from the gas phase is rather
low and usually compensated by the desorption of carbon dioxide. The methodology
is attractive because it permits a separation of fluiddynamics (momentum balances,
continuity equations, and turbulence model) from material balance equations for the
state variables of interest. Figure 7.16 illustrates how results from the fluiddynamic
simulations (mean velocities u~~~'P (r, Z, 'P), turbulent dispersion coefficient Deff
(r,z,'P), and local gas hold up EG (r,z,'P)) can be used as parameters in the material
balance equations.
The main advantage of the separation is the reduction of the computational
effort. Another aspect is the fact, that the two sets of equations can be solved
with different numerical methods and on different numerical grids. Due to the
nature of the nonlinearities in material and momentum equations they usually
require different grid refinements in different areas of the computational domain.
Additionally, if the assumption of a stationary flow field is valid, the simulation
234 7 Application of Computational Fluiddynamics (CFD)

ITU -
"*, ,,
\.00010 ' %
QJ)I)o,o· 'IC. •
• INlI . '..1)0-'0 ' ...
"*' _ 0..00-10' %
•
•
,...... -
"*, ..
, ,(10010 ' ~
0,,00.'0' '%
••
0,"
"'"
0" ."
0," 0'"
ellS II \..~

I
0'0
I .'"
OJ) n ~j

.oo "OO

." ""
0,. I'JoIO

.os DOS

000 ~J III)

_ o~ _ 0 I, _ 0 10 a llO'i ,,""
1 / 111'11 f 1"11
a
rna.. • 1.00"0 ' '"
.,... .. 0.00-10· % •• rna.- -
"*' -
1.00·10 ' %
0.00' 10' '%
••
..: 1II: 0. :1 m

... :ttO HI

b
Fig. 7.15. a Contour plots of specific gas hold up at three different angles between the baffles.
b Contour plots of specific gas hold up at two different levels of the tank

of the coupled set of equations would be unnecessarily slowed down by solving

the momentum equations. The material balance to be solved for each of the re-
acting components reads

Bc
"ut Bc
+ Ui ~
UXi
= - ~
B (
UXi
8c) +
Deff ~
uXj
Sc (7.44)
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 235

D.f/(r,z,¢) !
Simulation of one or two phase l Simulation of
'G(r,z,¢)
momentum balances -----tl material balances
u~.,.(r, z, ¢)
together with ) including sources and sinks
U~.1:,IP(r, z, ¢)
turbulence model l corresponding to kinetic model

Fig. 7.16. Separation of momentum and material balance equations

Sc = reaction+mass transfer gas/liquid+inlets/outlets

Deffr,z,<p) = turbulent dispersion coefficient=v effr,z,<p )/SCturb
with SCturb=turbulent Schmidt number, assumed to be 1, and veft<r,z,<p) from turbu-
lence modeling. Discretization of material balance equations is made using finite
volume elements. They are solved either with the differential-algebra solver Limex
(two-dimensional simulations) or ug (unstructured grids, development of the Insti-
tute for Computer Applications, University of Stuttgart, Prof. G. Wittum) for three-
dimensional simulations.

7.6.2
Simulation of Tracer Experiments

Tracer experiments are used to determine mixing characteristics. The most impor-
tant tracer substances are acids or bases, electrolytes, colored agents and heated li-
quid. Pulse experiments can be simulated by solving the material balance equation of
the tracer. Figure 7.17 illustrates an example of the simulated dynamics of a tracer
distribution for a single, symmetrically positioned Rushton turbine. The terminal
mixing time (), used as a quantitative measure, can be easily predicted from the
simulated response within the finite volume element which corresponds with the
sensor position.
From systematic simulations of tracer experiments in a stirred vessel equipped
with a Rushton turbine, a height/tank diameter ratio HDT -1=1, an impeller/tank dia-
meter ratio=0.3125, and an impeller clearance/height of liquid ratio=0.31, the follow-
ing correlation was obtained:
n(}9S = 27.5 (7.45)
with (}9s=time for 95% homogeneity. This result is in reasonable agreement with
numbers reported in the literature (n(}=35 [66] and =32 [67]).
Despite this good agreement, the terminal mixing time itself does not provide a
very informative measure of the mixing process. Because transient concentrations in
all volume elements are available from the simulations, it is also possible to calculate
the time course of inhomogenity defined by Landau and Prochazka [68]:
1 n
I(t) = V(c oo _ cO) ~ Vil(Ci(t) - COO) I (7.46)

Figure 7.18 summarizes the transient inhomogenity for three simulated tracer ex-
periments differing in the position of the tracer input.
The following example serves to illustrate that tracer experiments also help to dis-
criminate between different turbulence models. For this purpose, the classical tracer
236 7 Application of Computational Fluiddynamics (CFD)

•
9
! ~
••
iI·
~~
~!
! ·It ,

!I
g•
~ ~
••
o·
•·
••
§
,
I iI

~
g•
~ :
••
••S;
! '
It
.

experiments suggested by Khang and Levenspiel [69] have been employed. In this
experiment (also described by Tatterson [58] and Reuss and Bajpai [4]), the pulse
injection of the tracer is made with the aid of a concentric ring near the stirrer tips
and the response measured with a concentric ring electrode nearby. As a simple
representation of the recirculation flow for an impeller symmetrically placed in a
tank with HD T - 1=:1 Khang and Levenspiel [69] suggested a tank-in-series model
with recirculation loop. The approximated pulse response for such a recycle system
takes the form

y( t) = (
1 + 2 exp - -2~)
nttc
(27r + -27r)
t cos -
tc nt
(7.47)
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 237

t"", t [sl t",.2 t",.J

Fig. 7.18. Time course of inhomogenity for three different tracer pulses, 1: circular pulse into
stirrer zone, 2: pulse onto liquid surface, 3: pulse into stirrer zone

a b
2.0

I.S

i C 1.0
'-'
>.

I O.S

- A(t)
0.0
20 T1mt • .lie •• 0 10 IS

t[s]
Fig.7.19a,b. Response to a circular pulse into the stirrer zone: a measured (Khang and Leven-
spiel [69]); b simulated

The two required parameters - circulation time tc and number of tanks in the
cascade n t - can be predicted from frequency and decrease of the amplitude of the
measured or simulated response, respectively. Figure 7.19 shows a measured [69] and
a simulated (material balance Eq.7.44 and CFD) response. Simulated and measured
responses agree qualitatively well. The more interesting results of this exercise are
illustrated in Fig.7.20. The simulated response, obviously, is very sensitive to the
turbulence model used in the simulations. As a consequence of the overestimation
of the turbulent viscosity by the standard k-£ model, the corresponding simulations
show a response in which, in contrast to the experimental observations, the oscilla-
tions are completely damped down.
Large scale fermentation equipment usually contains more impellers. Based on the
CFD simulations shown in Sect. 7.4, it is then also possible to investigate the mixing
behavior of these systems. Figure 7.21 depicts the flow field and distribution of a
tracer 60 s after introduction of the pulse from the top in a 22 m 3 fermentor (geome-
trical properties: see [70]). Figure 7.22 shows a comparison of measured [70] and
simulated responses for the same reactor.
238 7 Application of Computational Fluiddynamics (CFD)

t [s]

Fig. 7.20. System response to a pulse into the stirrer zone: graph (a) flow field calculated with
the optimized Chen-Kim model, graph (b) flow field calculated with standard k-E model

tracer pulse at t=o

.
1

Fig. 7.21. Flow field (left) and tracer distribution (right) 60 s after the pulse (tank volume
22m 3 )
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 239

nurbattom

n • 1151/min

-o.,+--~~-~~-~~-l
o 50 100 150 200 250 300 350

t [s1 t[s1

Fig. 7.22. Comparison of measured [70] and simulated responses for the 22-m 3 reactor

A careful analysis of these comparisons illustrated that the measured signals al-
ways show a slightly faster response. With the aid of further investigations it was
possible to indicate that these differences are caused by the insufficiency of the
two-dimensional approximations. The two-dimensional approximations show a very
poor axial mixing between the impeller regions of the individual Rushton turbines
resulting in a more or less independent symmetrical flow field. The three-dimen-
sional simulations of the multiple impeller systems (not shown) indicate an addi-
tional tangential mixing effect primarily caused by the flow around the baffles. De-
spite these insufficiencies of the two-dimensional simulations, a reasonable agree-
ment between simulated and measured terminal mixing times could be observed.
The mixing time calculated from the equation for multiple impeller systems pro-
posed by Groen [71] which was experimentally verified in vessels between 0.015 m 3
and 130m3

(7.48)

where ni=number of impellers, H=liquid height, w=height of impeller blade, and

DT=tank diameter, is (B 9s )measured=257s. This value is in reasonable agreement with
the mixing time estimated from the simulations, (B 9S )simulated=288s.

7.6.3
Simulations of Substrate Distribution in Fed Batch Fermentations

The following example shows how simulations can be applied for optimization of fed
batch processes which are very common for a variety of biotechnical processes to
avoid oxygen limitations, heat transfer problems, over-flow metabolism, or catabo-
lite regulation. The concentration of the carbon and energy source in the feed are as
high as possible (in the range of 500 kg m- 3 ). The concentration inside the tank is
very often in the range of the saturation constant, e.g., in the concentration range of
1-100 mg 1-1. The challenge in the scale up of these processes is then to prevent con-
centration gradients resulting in further limitations (cS<CS,crit) as well as unwanted
byproduct formation or inhibition of production rates (cS>CS,crit). Examples of over-
flow metabolism are the growth of Saccharomyces cerevisiae (production of ethanol)
and Escherischia coli (production of acetate).
240 7 Application of Computational Fluiddynamics (CFD)

a b c
Fig. 7.23a-(. Substrate distribution for different positions of the inlet of the concentrated feed
solution: a liquid surface; b wall; c stirrer zone

-- ... .
"... ...... 10 1
...... ,
... .

J J J

a .... b .... c ....

0.0 0 0. 2 0.04 0.06

c. [gil)
Fig.7.24a-(. Substrate distribution in a stirred tank bioreactor (volume 22 m 3 ) equipped with
different impellers: a 4 Rushton turbines; b 2 Rushton turbines and 2 pitched blade impellers; c
4 pitched blade impellers
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 241

Assuming that oxygen supply is sufficient to avoid local oxygen limitations, the ki-
netic model required for the simulation includes only the material balance equation
for the substrate. As suggested in earlier simulations based on recirculation models
(micro-macromixer) by Bajpai and Reuss [80] the uptake kinetics is only considered
in the vicinity of the so-called critical sugar concentration. Thus, a rather simple
unstructured empirical model was chosen for the purpose of this study. It involves
a Monod type of kinetics for substrate uptake

rs = rsmax - -
Cs
- Cx = - Sc (7.49)
Ks + Cs

which holds at a certain time of the process for the corresponding biomass concen-
tration cx. If substrate concentration Cs locally exceeds the critical concentration
CS,crit> an ethanol production rate is superimposed which is given by

max Cs - CS,crit
rp = rp ( ) Cx (7.50 )
Ks + CS,crit - Cs

Equations (7.49) and (7.50) are then used as source terms in the material balance
equation, Eq. (7.44). Additionally, the feeding rate is considered as a source term in
the volume element corresponding to the feeding point. Figure 7.23 shows results of
simulations of the substrate distribution at three different positions for the substrate
inlet for a vessel with a volume of 681 [81]. As expected, feeding of the concentrated
sugar solution into the impeller region leads to the best equidistribution of substrate.
Again simulations have been performed for large scale vessels with multiple im-
pellers. Figure 7.24 summarizes the distribution of sugar in a vessel of 22 m 3
equipped with different combinations of Rushton turbines and axial impellers. As
expected, the axial impellers generally lead to a better distribution of the substrate.

7.6.4
Production of Acetoin/Butanediol with Bacillus subtilis

Bacillus subtilis has been used several times as a model for an oxygen sensitive cul-
ture for the characterization of the effects of inhomogeneities on the intensity of
oxygen transfer gas liquid in stirred tank reactors [73]. At dissolved oxygen concen-
trations below 1% saturation, the ratio of the two production rates of acetoin and
butanediol strongly depends on the dissolved oxygen concentration. In other words,
the selectivity of the process is very sensitive to changes in dissolved oxygen under
micro aerobic conditions. A detailed simulation of these effects requires a model for
the oxygen gradients as well as local production rates for the two products as a
function of dissolved oxygen concentration.
For the simulations presented in the following, a kinetic model proposed by Moes
et al. [73, 74] has been used. This model takes into account the formation of biomass
and the two products acetoin and butanediol as well as substrate and oxygen con-
242 7 Application of Computational Fluiddynamics (CFD)

sumption. The source terms in the material balance equations (Eq.7.44) for the six
state variables then read

Sx = rATP~X YXjATPCX
rS~ATP
SS = ----CX
Y pjS
SAe = (rS-->Ae - r Ae-->Bu + rBu-->Ae)CX (7.51)
SBu = (rAe-->Bu - rBu~Ae)CX
Sd, = -rNADH Yo,fNADCX + qo,
1- EG
SoG = qo,---
2 EG

Mass transfer gaslliquid is modeled as local mass transfer rate:

qo,(r,z,cp) = kL(r,z,cp)a(r,z,cp)(c~,(r,z,cp) - ~,(r,z,cp)) (7.52)

The kinetic expressions are given by:

r ATP~X = [( YATP,aer (Y :jS - Y:ejJ + YATP,anaer) rS~AC +

YATP,aerYXjATP)
+ YATPjBu(rAe~Bu - rBu~Ae )Rs1/ ( 1 + YXjS

rS-->E = ( 1 1)
- - - - - rS-->Ae -
YPjS YAejS
YXjATP
---rATP~X
YXjS
rNADH = YNAD,resprS~E + YNADjAerS~Ae + rBu-->Ae - rAe~Bu + rATP~XYXjATPYNADHjX
rS-->Ae = klRs
rAe~Bu = k2RAeRS + k2,eqRAe(1 - Rs)
rBu-->Ae = k3RBuRS + k3,eqRBu(1 - Rs)
Cs cAe CBu CO,
Rs = --- RAe = ---- RBu = Ro, = --"------
Ks + Cs KAe + CAe KBu + CBu Ko, + CO,
kl = kl,o + mlRo, k2 = k2,o + m2RO, k3 = k3,o + m3 RO,

YATP,aer = YATPjPYR + YNAD,resp YATPjNADH,

YATP,anaer = YATPjPYR + YNAD,AeYATP,NADH,
YATP,Bu = YNAD,BUYATPjNADH, (7.53)

The values of the model parameters are given in the original work of Moes [73].
The concentration at the gas liquid interface in Eq. (7.52) is calculated from
Henry's law,

H= po,(r,zcp)
(7,54)
xo, (r, z, cp)

with xo, (r, zcp) = co, (r, z, cp)/ Li Ci and Po, (r, z, cp) = Yo, (r, z, cp )P(r, z, cp)
7.6 Application of CFD to Simulations of Mixing and Biotechnical Processes 243

Additionally, it is possible to estimate local values for the volumetric mass transfer
coefficient. For this purpose the mass transfer coefficient is estimated with the aid of
the equation suggested by Kawase and Moo Young [75]

(7.55)

and the local value of the specific surface area from

_ 6cG(r, z, <p)
a (r,z,<p ) - d (7.56)
32

Due to the reduced coalescence in the fermentation broth, a constant Sauter mean
diameter has been assumed in these simulations .

.... '.
mil)! - 2..25o'tO"' II"IOUm' •
~-o"oo.1011ftClU1't'I1 •

I I

a ... b ....
Fig. 7.25. a Dissolved oxygen concentration after 8 h of batch fermentation of Bacillus subtilis.
b Production rate of butanediol after 8 h of batch fermentation of Bacillus subtilis

a b
15.0 15.0
12.5 12.5

J
J
10.0
7.S
S.O
2.5

1.0 2.0 3.0 4.0

P/( P Y) [Wlkg]
5.0
J
u
~
10.0
7.S
5.0
2.5
0.0
0.0
/-
1.0 2.0 3 .0
PI( PV) {Wlkg]
4.0 5.0

Fig. 7.26. Ratio of acetoin to butanediol as a function of specific power input: a measured (data
from [72]); b simulated
244 7 Application of Computational Fluiddynamics (CFD)

Figure 7.25 shows typical results from the simulation of the system of material
balance equations which are parameterized from the CFD simulations as described
before. In this figure, distributions of dissolved oxygen concentration and the pro-
duction rate of butanediol after 8 h of simulated batch fermentation are shown. Fig-
ure 7.26 illustrates a comparison of measured [72] and simulated ratios of the two
products acetoin and butanediol at the end of the fermentation as a function of spe-
cific power input. The simulations show a behavior qualitatively similar to that ob-
served by Griot [72]. No attempt was made to obtain a quantitative fit due to the
diversity of sources of parameter values.

References

1. Balmer GJ, Moore IPT, Nienow, AW (1987) Aerated and anaerated power and mass trans-
fer characteristics. In: Ho ES, Oldshue JY (eds) Biotechnology, scale-up and mixing.
AIChE, New York, p 116
2. Gbewongyo K, Dimasi D, Buckland BC (1987) Characterization of oxygen transfer and
power absorption of hydrofoil impellers in viscous mycelial fermentations. In: Ho ES,
Oldshue JY (eds) Biotechology, scale-up and mixing. AIChE, New York, p 128
3. Nienow AW (1990) Trends Biotechnol 8:224
4. Reuss M, Bajpai R (1991) Stirred tank models. In: Schiigerl K (ed) Biotechnology, A mul-
ti-volume comprehensive treatise, vol. 4: Measuring, modelling and control. VCH, Wein-
heim, New York, Basel, Cambridge, p 299
5. Reuss M (1995) Stirred tank bioreactors. In: Asenjo JA, Merchuk JC (eds) Bioreactor sys-
tem design. Marcel Dekker, New York Basel Hong Kong, p 207
6. Ju SY (1987) Three-dimensional turbulent flow-field in a turbine stirred tank. Ph. D. The-
sis, Louisiana State University
7. Joshi, JB, Ranade VV (1990) Trans I Chern Eng 68:19
8. Kresta SM, Wood PE (1991) AIChE J 37:448
9. Van't Riet K, Smith JM (1975) Chern Eng Sci 30:1093
10. Perng CY, Murthy JY (1993) AIChE Symp Ser 89:37
11. Takeda H, Narasaki K, Kitajima H, Sudoh S, Onofusa M, Iguchi S (1993) Computers and
Fluids 22:223
12. Brucato A, Ciofalo M, Grisafi F, Micale G (1998) Chern Eng Sci 53:3653
l3. Gosman AD (1998) Trans I Chern Eng 76:153
14. Platzer B (1981) Chern Tech 33:16
15. Harvey PS, Greaves M (1982) Trans I Chern Eng 60:201
16. Placek J, Tavlarides LL (1985) AIChE J 31:11l3
17. Placek J, Tavlarides LL, Smith GW, Fort 1(1986) AIChE J 32:1771
18. Middleton JC, Pierce F, Lynch PM (1986) Chern Eng Res Des 64:18
19. Bakker A, Van den Akker HEA (1994) Trans I Chern Eng 72:583
20. Togatorop A, Mann R, Schofield DF (1994) AIChE Symp Ser 299:19
21. Issa RI, Gosman AD (1981) The computation of three-dimensional turbulent two phase
flows in mixer vessels. In: Numerial methods in laminar and turbulent flow. Pineridge
Press, Swansea, p 829
22. Tragardh C (1988) A hydrodynamic model for the simulation of an aerated agitated fed-
batch fermentation. In: Bioreactor fluid dynamics. Elsevier Applied Science, p 117
23. Politis S, Issa RI, Gosman AD, Lekakon C, Looney MK (1992) AIChE J 38:1946
24. Morud K, Hjertager BH (1993) Computational fluid danymics simulations of bioreactors.
In: Mortensen U, Noorman H (eds) Bioreactor performance. IDEON, Lund, Sweden, p 47
25. Bakker A, Van den Akker HEA (1994) Trans I Chern Eng 72:594
26. Ranade Vv, Van den Akker HEA (1994) Chern Eng Sci 49:5175
27. Launder BE, Spalding DB (1972) Mathematical models of tubulence. Academic Press
28. Launder BE, Spalding DB (1974) Comp Meth Appl Mech Eng 3:269
References 245

29. Kim JJ (1978) Three dimensional turbulent flow-field in a turbine stirred tank. PhD the-
sis, Louisiana State University
30. Pope SB (1978) AIAA J 16:279
31. Hanjalick K, Launder BE (1980) Trans ASME 102:34
32. Kline SJ, Cantwell BJ, Lilley GM (1981) The 1980-1981 HFOSR-HTMM-Stanford Confer-
ence on Complex Turbulent Flow, Standfort University, I, II, III
33. Roback R, Johnson BV (1983) NASA CR-168,252
34. Chen YS, Kim SW (1987) NASA CR-179,204
35. Batchelor GK (1953) The theory of homogeneous turbulence. Cambridge University Press,
Cambridge
36. Wu H, Patterson G (1989) Chern Eng Sci 44:2207
37. Jenne M, Reuss M (1999) Chern Eng Sci (in press)
38. Patankar SV (1980) Numerical heat transfer and fluid flow. Mc Graw-Hill
39. Costes J, Couderc JP (1988) Chern Eng Sci 43:2765
40. Yakhot V, Smith LM (1992) J Sci Comput 7(1)
41. Nagata S (1975) Mixing: principles and application. Wiley
42. Jenne, M (1999) Modellierung und Simulation der Stromungsverhliltnisse in begasten
Ruhrkesselreaktoren. PhD thesis, Universitat Stuttgart
43. Bouafi M, Roustan M, Djebbar R (1997) Mixing IX, multiphase systems. Recents progres
en genie des pro cedes 11(52):137
44. John AH, Bujalski W, Nienov AW (1997) Mixing IX, multiphase systems. Recents progres
en genie des pro cedes 11: 169
45. Nienow AW, Elson TP (1988) Chern Eng Res Des 66:5
46. Cooke M, Middleton JC, Bush JR (1988) Proc 2nd Int Conf Bioreactor Fluid Dynamics.
BHRA/Elsevier, p 37
47. Abradi V, Rovera G, Baldi G, Sicardi S, Conti R (1990) Trans I Chern E 68:516
48. Ishii M, Zuber N (1979) AIChE J 25:843
49. Kuo JT, Wallis GB (1988) Int J Multiphase Flow 14:547
50. Ishii M (1975) Thermo-fluid dynamic theory of two-phase-flow. Eyrolles
51. Drew DA, Lahey TJ (1987) Int J Multiphase Flow 13:113
52. Lopez de Bertodano M, Lahey TJ, Jones OCC (1994) Trans SME 116:128
53. Lopez de Bertodano M, Lee SJ, Lahey RT, Drew DA (1990) ASME J Fluids Enging 112:107
54. Svendsen HF, Jakobsen HA, Torvik R (1992) Chern Eng Sci 47:3297
55. Johansen ST, Boysan F (1988) Metall Trans B 19B:755
56. Lahey RT, Lopez de Bertodano M, Jones OC (1993) Nuclear Enging Des 141:177
57. Sato Y, Adatomi M, Sekoguchi K (1981) Int J Multiphase Flow 7:167
58. Tatterson GB (1991) Fluid mixing and gas dispersion in agitated tanks. McGraw Hill
59. Bombac A (1994) PhD thesis, University of Ljubljana
60. Bombac A, Zun I, Filipic B, Zumer M (1997) AIChE J 43:2921
61. Rousar I, Van den Akker HEA (1994) Proc 8th Europ Conf on Mixing, Cambridge, UK,
p89
62. Joshi JB, Pandit AB, Sharma MM (1982) Chern Eng Sci 37:813
63. Barigou M (1987) PhD thesis, University of Bath
64. Barigou M, Greaves M (1992) Chern Eng Sci 47:2009
65. Calderbank PH (1958) Trans I Chern Eng 36:443
66. Voncken RM (1966) Circumlatie stromingen en menjing in geroerde vaten. PhD thesis,
Delft University of Technology
67. Hoogendoorn CT, Hartog AP (1967) Chern Eng Sci 22:1689
68. Landau J, Prochazka J (1961) Coli Czechoslov Chern Commun 26:1976
69. Khang SJ, Levenspiel 0 (1976) Chern Eng Sci 31:569
70. Cui YQ, van der Lans RGJM, Noorman HJ, Luyben KCAM (1966) TransI ChernE 74
(A):261
71. Groen DJ (1994) Macromixing in bioreactors. PhD thesis, Delft University of Technology
72. Griot M (1987) MaBstabsvergroBerung von Bioreaktoren mit einer sauerstoffempfindli-
chen Testkultur. PhD thesis, ETH Zurich
246 7 Application of Computational Fluiddynamics (CFD)

73. Moes J (1985) Untersuchung von Mischphiinomenen mit Hilfe von Bacillus subtilis. PhD
thesis, ETH Ziirich
74. Moes J, Griot M, Keller J, Heinzle E, Dunn LJ, Bourne JR (1985) Biotech Bioeng 27:482
75. Kawase Y, Moo-Young M (1990) Chern Eng J 43:B19
76. Yianneskis M, Popiolek Z, Whitelaw JH (1987) Fluid Mech 175:537
77. Watanabe T, Hirano M, Tanabe F, Kamo H (1990) Nuclear Engng Design 120:181
78. Huang B (1989) Modelisation numerique d'ecoulements disphasiques it bulles dans des
reacteurs chimiques. PhD thesis, Lyon
79. Kowe R, Hunt JCR, Hunt A, Couet B, Bradbury LJS (1988) lnt J Multiphase Flow 14:587
80. Bajpai R, Reuss M (1982) Can J Chern Eng 60:384
81. Schmalzriedt S, Reuss M (1997) Mixing IX, Recent advances in mixing, Recent Progres en
Genie des Procedes 11 (51): 171
8 Bubble Column Bioreactors
Andreas Liibbert

8.1
Introduction

From the various types of bioreactors discussed in the literature, only a few are ac-
tually used in industry. By far the majority of processes are performed in stirred
tank bioreactors. Besides these standard reactors, only bubble columns or, more gen-
erally, airlift reactors reached a noticeable number of applications.
In bubble column bioreactors, agitation is performed by density driven fluid mo-
tions induced by sparging air into the continuous liquid phase at the bottom of the
reactor. Air sparging leads to density inversion in the culture medium and, thus, to
buoyancy motions in the dispersion. At sufficiently high gas flow rates, these con-
vective flows become turbulent, which has the practical advantage that the fluid be-
comes intensively mixed. Since in bioreactors for aerobic cultures, air dispersion is
necessary anyway in order to provide sufficient oxygen mass transfer rates, gas spar-
ging serves two purposes in these reactors: agitation and mass transfer. In many
practical systems, mixing by these buoyancy-induced flow effects makes additional
mechanical agitation needless. This is the domain where airlift reactors are applied
in industrial practice.
Bubble column bioreactors are usually constructed as cylindrical vessels with an
aspect ratio (reactor height/diameter) greater than two. The air can be dispersed at
the bottom of the vessel by any gas sparger. For example, one can use the conven-
tional ring sparger most often used in stirred tank bioreactors. Thus, these simple
bubble column reactors do not contain any moving part and are thus not only in-
expensive in terms of investment cost, but also easy to maintain, e.g., easy to clean
and sterilize.
Often, these simple bubble column reactors are divided into two segments by
means of additional wall elements. This may most easily be accomplished by insert-
ing a tube concentrically into the cylindrical reactor. When then the liquid is aerated
in one segment only, e.g., in the annulus, a substantial density difference is generated
between both segments. Then a strong circulatory fluid motion results. The net ef-
fect is that dispersion rises or is lifted in the aerated part, named the riser section,
and, by continuity, the liquid flows back in the other section referred to as the down-
comer. Because of this circulatory motion, we then speak about an airlift loop reac-
tor. There are many different ways to construct such segmented airlift loop reactors.
Since a bioreactor, as we will understand it in this chapter, is a container in which
a production process is being performed, there are some obvious requirements to be
met. The criteria are almost the same for all of them. One most obvious primary
248 8 Bubble Column Bioreactors

task of the bioreactor is mixing. The central objective in most bubble column bior-
eactors is to obtain a high degree of homogeneity with respect to the species dis-
solved in the continuous liquid phase, in order to provide nearly the same chemical
and physical environmental conditions for all cells, if possible close to the optimum
with respect to a maximal rate of the desired biochemical conversion. Additionally,
the reactor must be able to provide sufficiently high oxygen mass transfer rates in
order to supply the cells in aerobic cultures with sufficient oxygen, and, finally, it
must be able to keep the culture temperature close to its optimal value.
When obtaining a high mass transfer rate is the main objective in a particular
process, a stirred tank reactor might be preferable as compared to a bubble column
since these reactors allow one to apply higher agitation power inputs and thus higher
mass transfer rates. However, when bubble columns are able to provide sufficiently
high oxygen transfer rates, they do this at a considerably higher energetic efficiency.
This is often a decisive argument for using bubble column reactors, e.g., in the yeast
manufacturing industry. Another essential advantage as compared to stirred tank
reactors is that there is virtually no serious restriction with respect to the size of
bubble column bioreactors. Thus, bioreactors larger than 300 m 3 are nearly exclu-
sively constructed as airlift reactors.
In order to design a bubble column, it is necessary to understand the multiphase
fluid motion in the reactor. Intensive bubble column research over many decades has
shown that simple assumptions about the flow patterns do not allow one to make the
quantitative predictions necessary for serious reactor design. Hence, physically
based detailed fluid mechanical models are required for such a purpose. From the
design point of view, models represent that part of our current knowledge about
these reactors that can be exploited quantitatively. This chapter deals with bubble
column models. Only recently has it become possible to perform detailed computa-
tional fluid dynamical simulation of the complex flow in these reactors. The focus in
this chapter is on mechanistic insight into the complex two-phase gas-liquid flows in
bubble column reactors.

8.2
Phenomenology

As the driving mechanism for the fluid flow in bubble column reactors, the density
differences of the fluid elements in this system were identified. These density differ-
ences are mainly caused by the dispersed gas not being homogeneously distributed
in the column. The gas hold-up E, defined as the fraction of the gas-phase of the total
volume VD of the dispersion considered, is thus of main importance for airlift reac-
tors:

(8.1 )

VG is the total gas volume within YD' The gas hold-up is thus a volume-related
quantity.
When the entire culture within a bioreactor is considered, we more precisely refer
to the corresponding hold-up as the total or global gas hold-up. We may, however,
particularly in systems where the gas phase is not quasi-homogeneously distributed,
8.2 Phenomenology 249

also speak about spatial distributions of the gas phase. Then we prefer to speak
about local gas holdups lOb which are defined by the differential quotient

(8.2)

In this case we assume that we are allowed to speak about the dispersion as a quasi
homogeneous fluid. This is a reasonable assumption when we speak about technical
reactor sizes where the bubbles are usually very small as compared to the reactor
dimensions. However, when we discuss effects on small scales, we should keep in
mind that in gas-liquid dispersions the gas-phase is composed of individual bubbles
of volume Vj. Then this assumption might be violated, e.g., it does not make sense to
speak about continuous gas hold-ups when we are looking for effects on scales of the
order of the mean bubble size. On such small scales, the gas-liquid two-phase flow
system must be treated in a discontinuous way.
When we consider the individual bubbles of volume Vj' we may rewrite the total
gas hold-up as

(8.3)

where the sum is drawn over all bubbles in the entire volume V D of the dispersion.
The volume Vj characterizes the bubble size. However, it is often more convenient to
use its diameter instead, since the diameter can most easily be estimated by visual
inspection of the dispersion. Unfortunately, the bubble diameter is not well defined
since different bubbles of the same volume might have different forms. This problem
has been circumvented by defining the so-called equivalent diameter, d j , which is the
diameter of a sphere with the bubble's volume Vj. Then we can write the gas volume
VG as

(8.4)

Bubbles have several functions in bioreactors, their most important one being the
transport of a gaseous reactant from the dispersed gas phase across the bubble sur-
faces into the liquid phase. In this respect, the transport cross section A available for
this mass transfer is of primary importance. A is the total interfacial area between
the gas and the liquid phase. This is, apart from the top surface of the dispersion, the
sum of all bubble surfaces

(8.5)

where Aj is the surface area of the j-th bubble. In biochemical reaction engineering,
one prefers to discuss the transport cross section as a specific, in this particular case
a volume related quantity. This specific interfacial area a is defined as

(8.6)
250 8 Bubble Column Bioreactors

With these definitions we are able to relate the gas hold-up with the quantities of
primary interest in biochemical reaction engineering. Starting from the definition at
Eq. (8.1), we get
VG AV G A ~Ldj ~Ldj
E=-=--=----=a--- (8.7)
VD VDA VD7rLdJ 7rLdJ
Since the bubbles in biotechnologically relevant dispersions are not of uniform
size, such expressions are of little practical use. One is more interested in character-
izing the bubbles by some mean bubble diameter. In the light of the preceding dis-
cussion it is convenient, as proposed by Sauter, to take the quantity
Ldj
(8.8)
ds = LdJ
as such a mean bubble diameter. The sums are drawn over all bubbles j. In the lit-
erature, d s is referred to as the Sauter diameter.
With the Sauter diameter d., we obtain a simple, practically very important rela-
tion between the relevant gas-phase characteristics of a bubble dispersion:

E= a ds or a = 6E (8.9)
6 ds
which essentially says that the important specific interfacial area a is proportional to
the gas hold-up £, the proportionality constant being dependent in a most simple
way from the Sauter mean bubble diameter d s • In other words, we can increase the
specific interfacial area a by increasing the gas hold-up £ and we are more effective
when we are able to keep the mean bubble size smaller.
Assuming that the volume VL of the liquid-phase of a culture within a bioreactor is
given, then the higher the gas hold-up £, the larger the total volume VD of the dis-
persion will be. Usually, in discussions about the gas-phase behavior within bioreac-
tors, the solid phase, in particular the cells, can be attributed to the liquid phase. In
reactors where the reactor horizontal cross sectional area ARCS is uniform with re-
spect to the height HD of the dispersion, we can write the gas hold-up as a function
of the height HD and the height HL the liquid will assume within the reactor when
there are no gas bubbles would be present.
VG VD - VL (HD - Ht)ARCS (HD - Ht)
E=-=--- (8.10)
VD VD HDARCS HD
This relative enhancement of the culture level in the reactor by aeration or simply
the presence of the gas bubbles is of practical importance, since one must take care
of this level during the operation of the bioreactor.
In a global view, a higher gas hold-up of the dispersion leads to a lower mean
density PD of the dispersion. When the gas phases can be assumed to be quasi-
homogeneously distributed, the density PD can be represented as a linear function
of the gas hold-up £:
(8.11)
since the gas density PG is very much smaller than the density PL of the liquid phase.
This has an immediate consequence for the static pressure at a particular height H
within the reactor. Since the static pressure drop ~p over a height interval ~H is
8.3 Basic Equations of Motion 251

dependent on the density of the dispersion, it is possible to relate directly the static
pressure drop and the gas hold-up

(8.12)

(8.13)

a relationship that is often used to measure the gas hold-up in multiphase reactors. g
is the gravitational acceleration.
The gas hold-up can also be viewed from the point of view of the aeration rate.
The aeration rate is often represented by the gas throughput QG relative to the cross
sectional area ARCS of the reactor. This quantity is referred to as the superficial gas
velocity wsg ' since it formally has the dimension of a velocity:
QG
W sg = - - (8.14)
ARCS

When the mean bubble velocity in the laboratory coordinate system is Wbb then
with

(8.15)

where 'tB is the mean residence time of the bubbles in the reactor, and

(8.16)

we get

(8.17)

In other words, the gas hold-up is proportional to the superficial gas velocity, the
proportionality constant being the reciprocal value of the mean bubble rise velocity
in the laboratory coordinate system. The superficial gas velocity is the most impor-
tant manipulatable variable. The effective bubble velocity, Wbb however, depends on
the fluid-dynamic properties within the reactor as well of the broth rheology.

8.3
Basic Equations of Motion

8.3.1
Fundamental Laws of Fluid Motion

The state of a fluid flow system is basically determined by the quantitative descrip-
tion of dynamics of the flow field, i.e., the spatial (r) and time (t) dependent fluid
velocity u(r,t). The basic equations of motion for fluid dynamical systems are de-
rived from the elementary conservation laws for mass and momentum (e.g., [1]).
252 8 Bubble Column Bioreactors

8.3.1.1
Mass Conservation

If we consider some finite fluid volume, the mass conservation law says that the
amount of mass m within this volume can only be changed by transport of mass
across its boundary surface. If we consider the density p, i.e., the mass per volume
instead of the mass, the mathematical formulation of this statement is

fJp = -div (pu) (8.18)

fJt

8.3.1.2
Conservation of Momentum

The momentum equation is Newton's second law applied to fluid motion, saying that
a substantial change of the momentum mu requires the action of a force F. In a
volume related representation, the substantial derivative of pu is thus equal to the
sum f of all volume related forces acting on the fluid element under consideration:
D(pu) = f (8.19)
DT
The volume-related forces acting on a fluid element are usually distinguished by
body forces fb' pressure forces fp, and surface forces fs.
From the forces acting on the fluid volume, the field or body forces, the gravitation
force is the most important in our case as long as we are not dealing with special,
e.g., magnetic particles which would be influenced by magnetic fields. The gravita-
tion leads to
(8.20)
where g is the gravitational acceleration.
The pressure term is the pressure gradient across the fluid element considered:
(8.21)
It can also be interpreted as a force resulting from normal stresses on the fluid
element. Finally, there are forces acting tangentially on the surface of the fluid ele-
ment. These are the viscous shear forces, which can be described by
fs = VT (8.22)
where T is the viscous stress tensor. In a viscous fluid flow, the two terms fp and fs
describe the immediate momentum transfer between adjacent fluid elements in a
viscous fluid flow.

8.3.1.3
Navier-Stokes Equation System

In fluid dynamics, the fluid elements considered as control elements are usually re-
presented by the cells of the numerical grid used during the solution of the equa-
tions of motion. The flow velocities u are thus considered at discrete, spatially fixed
points in the numerical grid only. This representation is referred to the Eulerian
8.3 Basic Equations of Motion 253

representation. In Cartesian coordinates, the substantial (or total or convectional)

derivative operator becomes
D 0 0 Oxl 0 OX2 0 Ox3 0
-=-+--+--+--=-+\7u (8.23)
Dt at Oxl ot OX2 ot OX3 ot ot
where Xi denote the components of the position vector r. The stress tensor T is
then defined by its components in the following form:
oui OUj 2 fJu n )
Tij = Meff ( - +---Dij- (8.24)
OXj OXi 3 oXn

and Meff the effective dynamic viscosity of the dispersion. Oij is the Kronecker sym-
bol.
Hence, the momentum balance reads
Opu
at + \7(puu) = - \7p + \7T (8.25)

and the equation of mass conservation reads in a corresponding representation

Op
at + \7(pu) = 0 (8.26)

It is known as the continuity equation.

The combination of both equations is usually referred to as the Navier-Stokes
equation system. The Navier-Stokes equation system, known for about 150 years, is
a most general equation system and encompasses all effects of fluid motions.
When we are dealing with two-phase gas-liquid flows, as we are forced to do in
bioreactors, the fluid density p becomes a further quantity that dynamically changes
with space and time. Hence, it is necessary to provide independent information
about the density variations. The appropriate approaches will be discussed at a later
point. First it is necessary to recognize the main practical problems that appear dur-
ing current attempts to solve the Navier-Stokes equation system, since this will shed
some light on our possibilities to extend this equation system.

8.3.1.4
Problems with Solving the Equations of Motion

The Navier-Stokes equation system is a quite general differential equation system. Its
adjustment to the flow field in a particular technical apparatus, for example to the
flow in a cylindrical vessel used as a biochemical reactor, is made by definition of the
boundary conditions, which must include the precise shape of the reactor. In the
case that we are dealing with, dynamic flow situations, we additionally need initial
conditions to the flow.
The initial conditions are usually very simple. We can, for example, assume that
initially there is no flow within the reactor. The boundary conditions thus define the
technical problem under consideration. In the particular case of a fluid flow in a
cylindrical vessel, the boundary conditions are as follows: the fluid velocity at the
walls of the container are zero. At the top surface (excluding the effects of the vessel
wall) only the axial velocity is restricted to zero because the horizontal components
are free.
254 8 Bubble Column Bioreactors

To a novice, the Navier-Stokes equation system looks quite simple and the bound-
ary conditions for the fluid motions in the cylindrical vessel are very simple as well.
And from this point of view, one would expect that the flow problem is easy to solve.
Unfortunately, however, things are not as simple as they would seem.
The first problem is the spatial resolution which would be necessary to cover all
relevant fluid motions in these reactors. Turbulent fluid motions are of primary im-
portance, as they are induced in order to obtain good mixing properties. According
to Kolmogoroff's turbulence theory, the order of magnitude of the diameter of the
smallest eddies, which are of importance, is roughly 10-4 m. It was estimated that one
needs at least 10 grid points to resolve the motion of an individual eddy [2]. If we
consider a biochemical reactor of a characteristic length scale of 1 m, a simple esti-
mation shows that we would need 105 elements in each direction of the numerical
grid, or 10 15 elements in a three-dimensional flow. With our current computers we
are able to handle numbers of grid points of the order of magnitude of 10 6 only. This
immediately shows that a straightforward approach covering all scales, which might
be of importance, will be impossible in the near future.
The second severe problem is that the pressure p cannot be determined separately.
Both the continuity and the momentum equation influence the pressure. Hence, so-
lutions of both must be determined in such a way that both lead to the same pressure
distribution p(r) at the same time t. In practice this must be obtained by a time-
consuming iterative approach.
The third major problem is that there are significant nonlinear terms in the Na-
vier-Stokes equation system. The solution of the differential equation system thus
requires implicit strategies, since explicit ones are known to lead to considerable
instabilities. Implicit schemes, however, require time-consuming iterations as well.
These three problems significantly reduce the number of computational grid ele-
ments that can be tackled within an acceptable computation time. Unfortunately, one
must deal with a rather low spatial resolution and, consequently, only large scale
motions can be resolved. Thus, these calculations of fluid motions in technical scale
reactors are often called Large-Eddy-Simulations. Flow effects expected to result
from motions on smaller length scales must then be lumped together in the material
parameters of the equation of motion. The main material parameters in the equa-
tions are the fluid density and the fluid viscosity. In the two-phase flow of bubble
column reactors both are dependent on the resolution of the numerical grid used in
the calculations. The effective viscosity depends on the fluid motions with character-
istic length scales significantly smaller than the fluid elements that can be resolved in
the calculations. The density depends on the local positions of the bubbles. Since the
bubbles are not homogeneously distributed, the density of the dispersion in a parti-
cular computational block is also dependent on its size. We need so-called subgrid-
models to account for these effects. We will go into more detail later on.
Generally the physical flow is independent of its representation. The representa-
tion must be chosen for convenience reasons. Instead of using cylindrical coordi-
nates to solve the equations of motion discussed, as would be the natural first choice,
it proved to be computationally more efficient to use Cartesian coordinates in order
to reach a predefined accuracy. The reason is that the matrices which result from
these representations during the solution algorithm get a form that can be solved
much faster. The disadvantage is that the spherical cross section of the column is
more difficult to describe in this representation. However, this is overcompensated
by the computational advantage.
8.3 Basic Equations of Motion 255

One might argue that the problem can be treated in a two-dimensional way when
dealing with cylindrical column reactors, since this would allow one to increase con-
siderably the numerical resolution. Such a representation, however, is not possible,
since the crucial ingredients of turbulent flows are vortex interactions [3], and inter-
action between eddies are three-dimensional effects. Comparisons of computations
in so-called waver columns, vessels of rectangular cross section with one dimension
of the cross section being much smaller than the other, showed a significant differ-
ence between both results from two- and three-dimensional calculations, although it
was assumed that in this geometry the two dimensional approach would be possible
without problems, as the degree of freedom in the third direction is much restricted
in these devices.
One often made further approaches to reduce the computational load such as time
averaging of the Navier-Stokes equation system. In such a way the random fluctua-
tions of higher frequency in the fluid flow, associated with the turbulent motion, are
smoothed out, while the slow variations of the flow field remain. The idea behind
this averaging/smoothing is to reduce drastically the number of grid points neces-
sary to simulate the flow in this way. The most often performed average procedure is
based on the decomposition, referred to as the Reynolds decomposition, of the flow
velocity u into two parts:

u=u+u' (8.27)

where u is a slow varying average velocity and u' is the fast and randomly fluctu-
ating component that is averaged away by this particular averaging procedure, called
Reynolds averaging. The averaged equations are called Reynolds equations (e.g. [2]).
Evidently a physical problem cannot be solved by averaging alone [3]. The advan-
tage of the smaller number of grid elements is obtained at the cost of additional
assumptions about the influences of the essential physics averaged or separated
away, namely the turbulent fluctuations. In other words, the Reynolds-averaged
equations need additional information about the influence of the turbulence in the
form of apparent turbulent stresses on the fluid motion. The problem of modeling
these relationships is referred to as the closure problem. Thus in order to solve the
Reynolds equations, an additional turbulence model is required that is not necessary
when the original Navier-Stokes equation system is solved directly. The best known
of such models is the so-called k-£ model.
The k-£ model was often discussed in turbulence theory (e.g., [[2]). According
Boussinsq's approximation, the Reynolds' stress, i.e., the additional term appearing
with Reynolds' averaging, is proportional to the changes of the mean velocity com-
ponents Ui in a perpendicular direction, the proportionality constant being the tur-
bulent viscosity fJ,T:

-PUiU; = fJ,T (~~; + ~~:) (8.28)

Thus, the problem is to estimate the turbulent viscosity fJ,T' General kinetic theory
shows that

fJ,T= pvl (8.29)

where v and I are the velocity scale and the corresponding length scale of the turbu-
lent motion. Thus, these quantities have to be estimated. Kolmogoroff and Prandtl
256 8 Bubble Column Bioreactors

showed in the 1940s that v can be estimated by the root mean value of the mass
related kinetic energy k [m 2 /s 2 ] of the turbulent fluid motions (k=u'iu';I2). It was
also shown that the corresponding length scale 1 can be estimated to be proportional
to k 3/2 /1':, where I': [m2 /s 3 ] is the mass related energy dissipated per unit time. Then
we get

(8.30)

where C and c are constants. Thus the turbulent viscosity is essentially estimated by
the kinetic energy k in the turbulent fluid oscillations and the energy dissipation
density f, i.e., the power loss of the turbulently fluctuating fluid. This relationship
is often called the k-I': model of turbulent viscosity (e.g., [2]). Sometimes people try
to use this expression to estimate the effective viscosity in the large eddy simulation
procedure as well.
In this chapter we do not consider averaged equations for the entire flow system
but concentrate on the Large-Eddy-solution of the Navier-Stokes equation system.

8.3.1.5
Numerical Aspects

There are well established solution procedures that can be used to solve the Navier-
Stokes equations system (Eqs.8.25 and 8.26) provided sufficient computation power
is available, e.g., the SIMPLER technique [4].
Here only a few main problems will be mentioned qualitatively in order to show
the practical problems which may appear in practice. As shown before, only a very
limited number of computational grid elements is possible, and the accuracy in the
solution, which can be obtained at a ftxed number of grid elements, very much de-
pends on the discretization procedure of the equations of motion, i.e., the particular
algorithm used to solve the equations numerically.
There are two types of errors due to the discretization method, which are of con-
siderable importance to bubble column simulation. The ftrst is that sharp gradients
in a process variable, e.g. a concentration of a species dissolved in a fluid, are
smoothed out by the numerical procedure. Since this effect is similar to a physical
diffusion process or a process with an increased momentum transport due to an
enhanced viscosity, this error is referred to as numerical dispersion or artiftcial visc-
osity. Sokolichin et al. [5] showed that the most often used UPWIND discretization
may lead to numerical diffusion effects which are much larger than the real physical
ones so that the final results are qualitatively different (damped fluctuations) to the
observed patterns of the quantity investigated. The consequence is that one is forced
to use so-called higher order solving procedures that can maintain the gradient.
However, these higher order methods lead to another error.
This second typical error also appears at gradients and leads to overshooting ef-
fects on one or both sides of the position of the gradient. This error is referred to a
the dispersion error, since it significantly influences the phase relationships of wave
propagation in the fluid. One of the most problematic results of this error to bior-
eactor simulations is that it might result in negative concentrations.
The numerical schemes are intensively being developed in order to avoid these
errors as much as possible. One currently much investigated method is the implicit
total variation diminishing (TVD) method that considerably reduces the artificial
8.3 Basic Equations of Motion 257

dispersion effects [5]. This method tries to reduce both errors at the cost of addi-
tional iteration steps which are, of course, time consuming.

8.3.2
Two-Fluid Model

In two-phase gas-liquid flows, the fluid density varies with time and spatial position.
Hence we need an additional model describing the density variations. One can relate
the density P=PD of the dispersion to the hold-up or volume fraction E of the gas
phase by Eq. (8.11).
The volume-related mass EPG of gas in the control volume element may be ba-
lanced [5]. Assuming that the gas-holdup can be regarded as a continuous quantity,
we can then formulate a continuity equation for EPG, which essentially means that we
make a balance across the mass of the gas content per unit volume. Under the as-
sumption that there is no significant mass transfer or chemical reaction, we obtain

(8.31)

where the term on the right hand slide of accounts for some generalized effective gas
phase losses due to random fluctuations of the bubble paths, which is described by a
diffusion term. When gas absorption, desorption, or chemical reaction effects are
significant, a further source term must be added on the right hand side of the equa-
tion. Thus, since PG can be considered constant and known, we obtained a dynamic
equation for the local gas holdup E.
A problem might be to determine the velocity u' G of the gas phase. However, this
can be estimated from the liquid phase velocity u by assuming a slip velocity U'slip
between the gas phase and the liquid phase:

UG =U + U'slip (8.32)

There are several proposals for the slip force in the literature, which account for
the slip between individual bubbles of different sizes and a stagnant continuous li-
quid phase or a bubble swarm and the liquid in which it rises (e.g., [6]). As experi-
ments showed, a fix slip velocity of 20 cmls in a vertical direction is not too bad an
approximation for UG' The reason for this roughly constant bubble velocity as a
function of their size is that larger bubbles more easily change their form and they
do so in such a way that their flow resistance coefficient Cw is roughly constant for
the relevant bubble Reynolds numbers range between 500 and 5000. In water we
obtain for such a range

cw = 50 [---;-] (8.33)
cm s
The assumption of the local gas-holdup E to be a continuous quantity is equivalent
to the assumption that the gas phase is considered a continuous, i.e., a second space-
filling fluid. The corresponding equation of motion can then also be formulated in a
Eulerian representation. Hence, the simulation of the two-phase gas-liquid flow re-
quires a simultaneous solving of two fluid flow systems - that for the dispersion and
that for the gas phase. This approach is thus termed the Euler-Euler representation
or simply the two-fluid-model.
258 8 Bubble Column Bioreactors

8.3.3
Euler-Lagrange Approach

The number of grid points that can currently be handled in simulation software for
two-phase gas-liquid flows on the available workstations is about 33x33x200. When
we are dealing with a reactor diameter of 1 m, the computational grid elements have
a volume of roughly 30 cm 3 , while a bubble with an equivalent diameter of 4 mm has
a volume of only 0.03 cm3 • Hence, the typical bubble is small as compared to the grid
element. This fact that the bubble motion cannot be resolved on such grids seems to
justify the assumption of considering the gas phase in numerical simulation as a
continuous phase. However, small dispersed particles are often considered part of
the continuous liquid phase, just as the microorganisms dissolved in the continuous
liquid phase must be considered part of the liquid.
Landau and Lifschitz [7] showed that the Navier-Stokes equation system can be
applied to two-phase flows when the dispersed phase elements are small, do not
drastically change the overall fluid density, and if the momentum of the particles or
bubbles can be neglected. Then, obviously, the corresponding material properties
appearing in the Navier-Stokes equation system must be properly corrected: The
density must be chosen as the effective density of the dispersion and, similarly, the
usual viscosity f-L must be replaced by an effective viscosity f-Leff.
In bioreactors, these conditions are widely fulfilled. The sizes of the dispersed
bubbles are small as compared with the characteristic reactor scale, and hence we
are dealing with small particles. The density changes of the dispersion within the
computational grid elements due to the appearance of bubbles, however, is not neg-
ligible. It is inhomogeneously distributed and changing with time, since the bubbles
move relative to the continuous liquid phase.
The natural way of considering the gas-phase motion, and hence to estimate the
fluid density, is to follow the self-evident physical picture and regard the bubbles as
individuals that rise in their turbulently agitated liquid surrounding. As already
mentioned, this cannot be done to a sufficient accuracy on the same grid as the fluid
flow simulation. The bubble motion must be determined with a much higher spatial
resolution, within the numerical grid of the Navier-Stokes calculation. At each time
step of the solution of the Navier-Stokes equations, the positions and the sizes of the
individual bubbles then allow one to determine the average density within the indi-
vidual grid elements simply by averaging. This determination of the apparent den-
sity in the numerical grid elements can thus be termed a subgrid-model for the
density.
When the dispersion is treated in the Eulerian representation and the gas phase
motion in a Lagrangean one, the entire two-phase flow representation is referred to
as a Euler-Lagrange representation.

8.3.3.1
Dynamics of the Dispersed Gas-Phose

The Lagrangean treatment of the motion of individual gas bubbles has the advantage
that many effects observed in gas-liquid flow investigations can be treated in a direct
way. However, a large number of bubbles must be followed on their path through the
bubble column in order to get accurate results. Lapin and Liibbert [8] and Lapin et
8.3 Basic Equations of Motion 259

al. [12] were able to follow simultaneously up to roughly 200,000 bubbles. Hence, one
needs simply to apply equations of motion to determine the bubble trajectories.
The slip velocities Uslip of the bubbles relative to the velocity Ul of the surrounding
liquid-phase of the dispersion is dominated by pressure gradients as well as their
flow resistance. These effects can be taken into account by the force law:
dUslip
CvmPL ~ = -\7p - CUslip (8.34)

The bubble velocity Ub in the laboratory coordinate system, which is needed to

determine the bubble position, is simply related to these two velocity values by
(8.35)
The positions Sb of the bubbles after one time step of length Lit, can then directly
be determined by simply applying the equation
(8.36)
assuming that their positions at the previous time step in the larger-scale calculation
was SQ'

8.3.3.2
Effective Viscosity

For the effective viscosity appearing in the Navier-Stokes equation one also needs a
model. Here, it is necessary to remember that the effective viscosity in a turbulent
flow is essentially determined by the energy dissipation in the eddy motion. As is
well known, the random fluid motions on the smaller scale do dissipate more kinetic
energy than the motions on the larger scale.
In the large-eddy simulation of bubble column reactors, the mesh width is, as al-
ready estimated, of the order of magnitude of some centimeters. With respect to the
statistical turbulence theory this scale falls into the range of inertia. This is the range
of scales in which inertia and buoyancy forces are significantly larger than the vis-
cous shear forces. Consequently, the influence of the viscosity on the fluid flow pat-
terns is expected to be small. This can be confirmed by simple numerical experi-
ments, which show that a change in the numerical value of the viscosity by 100%
leads to an integral change of the velocity patterns in the range lower than 5%.
Hence, the influence of the viscosity is rather low for the large scale fluid motion.
The effective viscosity can thus be taken as constant for simulations of the fluid flow.
The concrete value can either be taken from experiments or from turbulence models.
As such models, use of the k£ model (Eq. 8.30) was tried. Experimental information
is expected to provide more reliable results.

8.3.3.3
Mass Transfer and Chemical Reaction

In order to consider correctly the concentrations of the gaseous component, for ex-
ample oxygen, dissolved in the liquid phase, the simulation must also contain the
gas-liquid mass transfer. As usual in chemical and biochemical engineering, the
mass transfer resistance can be assumed to be the liquid-side boundary layer around
260 8 Bubble Column Bioreactors

the bubbles. In this picture, the mass m of gas transferred per unit time into the
continuous liquid phase is

dm
-=kLA (*
0 -0 ) (8.37)
dt
where kL is the mass transfer constant, A the interfacial area, and (0*-0) the driving
concentration difference, i.e., the difference between the saturation concentration 0*
of the dissolved oxygen assumed at the physical interface and the local concentration
o of oxygen dissolve in the liquid bulk. Once the gas has been transferred into the
liquid phase, it is mixed across the liquid phase by fluid motions. This mixing pro-
cess is described by a combination of convection and diffusion:
80
at + (uV')O = Deff~O - Rc + Rmt (8.38)

Besides the diffusive and convectional transport of the dissolved gas within the
liquid, a source (Rmt ) and a consumption term (Rcg) must be considered in this rate
equation to account for gas-liquid mass transfer and consumption by chemical reac-
tion.
The gas consumption by means of chemical reaction can be described by a kinetic
expression. A typical example is a Monod-type rate expression of the form
o
Rc ~--f(T) (8.39)
Ko+O
where Ko is a Monod constant. The conversion rate is often a function f of the tem-
perature T as well, which may be approximated by a set of two Arrhenius-terms as
shown in standard text books (e.g., [10]). The source term corresponds to the mass
transfer rate already mentioned in Eq. (8.37):
(8.40)
The biochemical conversion process usually produces heat which must be re-
moved via the heat exchangers, which are most often welted as jackets onto the re-
actor wall. A heat balance can than be used to model the temperature behavior of the
dispersion:
aT
at + (uV')T = Keff~T + RT (8.41 )

where Keff is the effective temperature conductivity. The source term, the heat pro-
duction rate, may be considered proportional to the chemical reaction rate:
(8.42)
This heat transfer as well as the subsequent temperature increase within the cool-
ing jackets can easily be considered in such an extended bubble column reactor
model.

8.3.3.4
Mixing Due to the Bubble Rise

Bubbles are known to carry liquid in their wakes relative to the continuous bulk
liquid phase. They continuously pick up liquid and, on average they release this
8.3 Basic Equations of Motion 261

liquid after some residence time from their wakes. Thus, the bubbles actively mix
liquid [ll] at least on a scale of some bubble diameters. This is a scale which is still
smaller than the typical resolution of the large eddy simulation described so far. The
question is whether or not this wake mixing effects have some influence on the glo-
bal mixing in the bubble column. This question was addressed by Lapin et al. [12].
These authors developed a model that starts by describing the exchange of liquid
elements between wake and bulk. For rp, denoting the local wake volume per unit
volume (wake holdup), and <I> the local volumetric flow rate of the liquid through
the wakes, the following equations were established:

(8.43)

fJew fJew) = -<I> (Cw - Cb )

rp) ( Tt+ubfu (8.44)

Ub is the bubble or wake velocity relative to the bulk. Index w refers to wake proper-
ties while index b refers to bubble properties.
The analytical solution of these equations for a typical tracer dispersion by the
bubble wakes with the initial conditions:
Cb(X,t =0) =b(x), cw(x,t =0) =0 (8.45)
looks a little bit simpler when we take ",=(1-rp)/rp, i.e., the relative volume of bulk
and wake liquid.
cw(x, t) = "'7](x)e-l<x-t+xIo(2J",x(t - x)) (8.46)

Cb(X, t) = 8(x)e- t + 7](X)~(t - x)e-l<x-t+xI 1 (2J",x(t - x)) (8.47)

where T\(x) is a unit step function where 10 and 11 are modified Bessel functions of
the first kind. Both expressions are only defined for O<x<t. Elsewhere, the solutions
are identically equal to zero. Physically, these solutions show how a tracer spot be-
comes dispersed with time over the coordinate x. The property that most clearly
characterizes the dispersion effect is the variance cr 2 of such a tracer distribution,
which is obviously a function of time. The expression, Lapin et al. [12] found, is
rather complex, however, for with large times a rather simple expression was ob-
tained for the variance:
(8.48)
This variance is proportional to time t and T.r=rp/<I>, the mean residence time of the
liquid in the bubbles' wake. For the often used axial dispersion model, the standard
deviation is
(8.49)
where Dax is the axial dispersion coefficient. By comparison of the coefficients, we
thus obtain for the axial dispersion coefficient with respect to the bubble wake dis-
persion
_ u~(1- rp)2rp2
Daxw - ----"----=---'-- (8.50)
<I>
262 8 Bubble Column Bioreactors

The wake additionally leads to a convective transport of liquid with a mean velo-
city U, which immediately turns out to be
(8.51)
The velocity U of the center of mass of the tracer cloud can easily be interpreted in
physical terms. It is equal to the velocity Ub of the bubble or wake relative to the bulk
liquid times a factor between 0 and 1, namely the wake hold-up. The physical expla-
nation is simple: a given fluid element is either transported within or outsides the
wake. If it is within the wake, it is being convected with the wake's velocity for a
while. The mean time it is transported with the wake is equal to the mean residence
time of the liquid in the wake. The probability of a liquid fluid element of being
within the wake is equal to the wake hold-up <p.
An order of magnitude estimations shows that this bubble wake contribution to
axial mixing accounts for roughly 20% of the total mixing in terms of the axial dis-
persion coefficient.

8.3.3.S
Problem of Bubble Coalescence and Redispersion

Coalescence and redispersion of bubbles change the bubble number and size distri-
butions. The influence to be expected on the large-scale motion of the dispersion is
not dramatic as long as one is dealing with low viscous liquids where the mean
bubble size is not to be expected to become very large. Small bubbles easily follow
the larger-scale liquid motions. Preliminary calculations showed that the character-
istic quantity of the bubble size distribution.- which is by far the most important to
the flow pattern of the dispersion - is its first moment, i.e., the mean bubble dia-
meter. The second, i.e., the width, is of much less importance to the bubble column
fluid dynamics. The main influence of the bubble size distribution on the perfor-
mance of a bubble column bioreactor is to be expected to be due to its influence on
the mass transfer and the micro-mixing properties of the reactor.
Both effects, coalescence and redispersion, can easily be incorporated into the Eu-
ler-Lagrange simulation software, as the paths of individual bubbles are followed up
in this approach. The problem is to provide probability distributions for the coales-
cence for the case that two or more bubbles approach each other to a distance where
coalescence may occur. Well accepted physically based theories are not yet available.
For redispersion, a corresponding probability distribution must be provided, which
basically depicts the probability of a bubble decay as a function of the energy dis-
sipation density in the fluid flow around the bubble. Once again, expressions that
reliably map such processes are still not available.

8.3.3.6
Rating of the Euler-Lagrange Representation

Since all the effects, the fluid flow, the mass, the heat transfer, as well as the chemical
reaction appear simultaneously, all these equations must be solved at the same time.
Lapin and Lubbert [13] performed such dynamical simulation directly in three di-
mensions on a Cartesian frame of computational nodes. The grid size used was
33 x 33 x 200. A typical example of a velocity pattern in a yeast fermenter is shown
in Fig. 8.1. The reactor diameter in this test case was 1 m.
8.3 Basic Equations of Motion 263

..
.'.'
"
, ,

"
~ . .. .. ..
I • • - .. :~ ~ ": ~

,<L~~--_
Fig.8.l. Typical instantaneous flow structure of the dispersion in the bubble column. The scale
at the lower left marks the velocity of 10 cm/s. Since the arrow length is not easy to recognize,
the velocity is additionally represented by the size of the triangle which marks the direction of
the arrow. Blue arrow color indicates that the vectors have a component coming out of the
plane, while red arrows are directed into the plane, black ones fairly remain within the plane

In simulations of the flow in such large bubble columns we are not able to increase
the number of elements in the numerical grid, and thus the sizes of the grid elements
become larger. At the same gas holdup we thus have more bubbles in the individual
computational grid elements. At the same time the total number of bubbles in the
reactor is larger as well, and it may easily become larger than the number that can be
dealt with individually. In this case it is of advantage to remember that bubbles are
moving in swarms or bubble clusters through the reactor. The motion of these
swarms can be treated in roughly the same way as the motion of single bubbles. In
this case the center of a bubbles swarm can be followed up on its way through the
bubble column. For more details see [5).
Finally, as also shown in [5), both representations, the Euler-Euler and the Euler-
Lagrange techniques, lead to practically the same results where they could be applied
to the same fluid flow simulations. Hence, in many cases one can choose one of these
two alternatives. However, some aspects can be represented more directly with the
Euler-Lagrange representation. Some applications of the Euler-Lagrange approach
will be discussed in the next section.
264 8 Bubble Column Bioreactors

8.4
Modeling of Particular Aspects of Bubble Column Reactors

8.4.1
Velocity Patterns in Bubble Column Reactors

The most general characteristics of the flow in a bubble column bioreactor is the
pattern of the velocity vectors across the volume of the dispersion. A three-dimen-
sional velocity field is difficult to represent graphically. Figure 8.1 depicts a typical
pattern of velocity vectors computed for points on a plane through the column axis.
Only the projections of the velocities on this surface can be shown. However, in
order to provide a little more information about the three-dimensional velocity vec-
tors, the arrows were colored. The red arrows indicate that the corresponding vector
has another component showing in the projection area, the blue ones represent com-
ponents coming out of the area. Those which do not have a significant component
perpendicular to the projections are shown in black. The velocity scale is shown as
an arrow at the lower left side of the left graph.
The velocity pattern depicted in Fig. 8.1 represents the flow computed for a parti-
cular instant in time. It shows a typical situation which appears after the system had
sufficient time to reach a quasi stationary state. The structure of the flow pattern
changes continuously with time in a chaotic way, although the assumed superficial
gas velocity, wsg=2 cmls, is not too high. About 200,000 bubbles clusters were as-
sumed to rise in the column in that situation, leading to a gas holdup of about 10%.
The outer appearance of this transient fluid flow shows the pattern expected for
turbulent fluid flow in a biochemical reactor. It is dominated by an eddy motion with
eddies of various sizes that are being generated and are decaying. From the videos,
which were constructed from many successive such velocity patterns, one can ob-
serve that the eddies are generated randomly, they move for some distance within
the reactor, and then they decay to form smaller eddies.
Since the eddy motion is chaotic, it is practically impossible to compare images
calculated for different times. Quantitative comparisons of the flow patterns are only
possible on a statistical basis. The easiest-to-obtain of such representations is time
averaging of the transient flow patterns. They must be obtained by averaging over
times spans that are long compared to the characteristic time-constants of the mo-
tion of the characteristic randomly appearing flow structures, i.e., over periods
which are long enough to average out all random flow structures.
Such an averaged velocity pattern is shown in Fig. 8.2 in the same representation
as before. A fairly symmetric flow profile appears after averaging. This profile is the
type of flow patterns usually discussed in bubble column literature. It was obtained
after averaging over a period of 30 min.
As can be seen from the figure, long-time averaging of the transient flow velocities
leads to the well known gulf stream patterns (a term proposed by Freedman and
Davidson [14]). Obviously all the random motions, which are induced for mixing
purposes, are averaged away. The long-time-averaged velocity profile shows the pro-
files that have been measured many times over the past 30 years. Comparisons of the
results of the simulation software and such velocity profiles have been performed on
several different scales (e.g., [159 or 12] to validate the computational code.
8.4 Modeling of Particular Aspects of Bubble Column Reactors 265

1111"1111' Irlll

Fig. 8.2. Long-time averaged version of the velocity pattern depicted in Fig. 8.1. The averaging
time is 30 min. The scale, 10 cmls, is shown in form of the arrow at the lower left side of the
graph

8.4.2
Fate of Individual Cells in the Bubble Column Bioreactor

Cells submersed in the liquid phase of a culture medium respond in their metabo-
lism to various concentrations in their environment. In larger reactors we expect, as
we will see later on in some detail, inhomogeneities in the concentrations of sub-
strate components. The flow patterns in bioreactors show that the cells move
through the reactor and will thus experience different concentrations in their envir-
onment. The transient eddy motion suggests that the environment of an individual
cell changes randomly, while the averaged profiles suggest a periodical change on the
average. Hence, the question appears as to what concentrations an individual cell
really sees.
For questions like these, the Euler-Lagrange representation of the fluid-mechanical
bubble column reactors is optimal. A single cell can be simulated by a fluid particle
of neutral buoyancy just as a small bubble with the gas density equal to the cell's
density, which is practically equal to the liquid density, and followed up on its path
through the column. Figure8.3 [16] presents a typical trajectory calculated for a
bubble column of 21 em diameter operated at a superficial gas velocity of 2 cmls.
The path appears to be nearly totally random. There is no similarity to the well-
ordered flow field that appears after long-time averaging of the local velocities. The
particular model reactor, for which the computations were performed, was chosen
266 8 Bubble Column Bioreactors

Fig. 8.3. Simulated trajectory of a single fluid particle in the bubble column reactor. The tra-
jectory appears to be a random path

since measurement values of the path of radioactive flow-follower particles were

available for a column of this size, so that the simulations could be validated experi-
mentally [17]. The radioactive particle could be tracked for hours on its path
through this bubble column, which was operated under the conditions assumed in
the numerical simulations.
The question as to whether or not the dispersed particle will see some periodicity
in the flow was answered by the question how long an intelligent cell would need to
recognize a ordered average gulf stream flow when it would be able to record and
analyze all the velocities it saw on its trip through the reactor. This problem could
also be investigated by the particle tracking method, since the velocities of the par-
ticle are available for the entire path. The answer to the question is that the particle
would need times of the order of hours to recognize the gulf stream. This time is
well above any time constant relevant to the biochemical conversion processes.
Hence, in practice the particle does not recognize periodic flows. From the point of
view of a single particle, the flow is entirely chaotic.
As the measurement of the particle trajectory and its numerical counterpart show,
the individual particles do not notice the deterministic orderly motion which ap-
peared after averaging the local velocities over a sufficiently long time. Thus, from
the point of view of the cells suspended in the liquid, the averaged gulf-stream mo-
tion can be regarded as an artifact.
That means that model calculations assuming the particles to pass regions of high
and low oxygen concentration periodically do not seem to be very realistic. Scale-
down experiments, which try to mimic periodic changes in the environment condi-
8.4 Modeling of Particular Aspects of Bubble Column Reactors 267

tions, may provide information on the dynamics by which the cells react to changes,
but do not approximate true conditions within bubble column bioreactors. For stir-
red tank reactors the situation is expected to be much similar.

8.4.3
Influence of Tilted Columns

The Lagrangean representation of the gas-phase motion is perfectly suited to deter-

mining gas residence time distributions, simply by recording the time for each bub-
ble from its entrance into the reactor at the sparger to the time at which the bubbles
leave the dispersion at its top surface. Figure 8.4 depicts two typical examples calcu-
lated for a bubble column operated at different vertical alignments [15].
In the figure, the behavior of a truly vertical column is compared with a column
slightly tilted by only OS away from the vertical. As already found experimentally,
the influence of a column inclination is considerable [18-20].
In Fig. 8.5, the corresponding velocity profiles are recorded.
The considerable changes in the mean residence time effect cannot immediately be
recognized in the corresponding transient flow patterns. These remain practically
unchanged. However, the long-time averaged patterns disclose the fundamental dif-
ference between both flow situations: While the gulf stream pattern characterizes the
averaged flow in the vertical column, the mean flow pattern in the tilted column
changed to a different flow mode. This is characterized by a single large asymmetric
circulation mode (Fig.8.4). Such a mode depicts larger mean flow velocities, and
hence the mean residence times become smaller.

8.4.4
Oxygen Distribution in a Yeast Fermenter

Sufficient oxygen supply for the culture is one of the key issues in bioreactor engi-
neering. A main objective is to keep the dissolved oxygen concentration in the med-
ium above some critical value throughout the entire reactor. From this point of view
the dissolved oxygen concentration pattern in bubble column fermenters during a
cultivation process is of primary interest. Here we discuss the results of such a pat-

00 6.-----------------------~

l: 9=0.5"
2: 9=0.0"

Fig. 8.4. Typical gas residence time distributions for a bubble column, perfectly aligned to the
vertical and one tilted by os away from the vertical
268 8 Bubble Column Bioreactors

a b c

Fig. 8.5. Left and middle: transient velocity patterns in a bubble column tilted by os from the
vertical. Both represents velocity patterns on planes through the axis of the reactor. They are
perpendicular to each other. The right graph shows the long-time-average of transient patterns

tern simulated with the Euler-Lagrange representation. The bioreactor considered is

a cylindrical pilot-scale bubble column bioreactor with 1 m diameter, aerated homo-
geneously across its entire flat bottom. Since the time scale at which the yeast grows
is much larger than the time scale at which the fluid flow structures in the reactor
are changing, we assume that the yeast concentration is not changing with time
during the situation simulated.
The dissolved oxygen concentration pattern O(r), compiled for one cross sectional
area through the axis of the fermenter, is shown on the left hand side of Fig. 8.6. The
corresponding fluid flow situation is the chaotic velocity pattern shown in Fig. 8.1;
i.e., both these plots depict the same momentary situation for different variables. The
most striking result which can be notices at first glance is that, due to the oxygen
consumption by the cells, the dissolved oxygen concentration does not reach a state
of saturation in the reactor and the oxygen does not become homogeneously distrib-
uted across the reactor. The patterns remain chaotic with respect to space and time.
Nevertheless, globally, a quasi-equilibrium is obtained in the sense that the distribu-
tion density p(O) of all the dissolved oxygen concentration 0 values at a given time
in the reactor does not change with time, as long as the yeast concentration can be
considered constant.
The long-time averaged dissolved oxygen concentration pattern, however, which is
depicted on the right side of Fig. 8.6, exhibits a clear structure, indicating that the
concentrations near the walls are considerably smaller than in the center of the col-
umn.
8.4 Modeling of Particular Aspects of Bubble Column Reactors 269

Fig. 8.6. Patterns of the dissolved oxygen concentration on a plane through the axis of a bub-
ble column bioreactor during a yeast cultivation: On the left, a transient situation is shown,
while on the right, a profile is shown that was obtained by averaging the transient results over
30 min real time

This has an immediate consequence for dissolved oxygen concentration measure-

ments, which are most often performed by means of electrochemical electrodes. In
practice, such electrodes reach only a few centimeters into the reactor and are usual-
ly mounted near the bottom of the vessel, where they are easily accessible to the
personnel. Each point in the graph of the right hand side of Fig. 8.6 can be inter-
preted as a time-averaged dissolved oxygen concentration signal. It is immediately
clear from the figure that the values to be expected at typical measurement positions
are far away from the true representative mean concentration for the entire reactor.
In other words the true mean oxygen concentration in a yeast production fermenter
is extremely difficult to measure.
In indirect measurements of key parameters of the fermentation process, e.g., the
specific growth rate Jl, one tries to make use of values of simple-to-measure vari-
ables. Usually only a few measurements are performed at large pilot or production
reactors that could be used for that purpose. Oxygen partial pressure in the vent line
of the reactor and the dissolved oxygen concentration 0 are proper candidates,
which are closely related to growth. In cases where the growth or product develop-
ment rate is oxygen limited, we need kinetic data about the oxygen dependency of
these parameters. Assume that the oxygen limitation effect can be described by a
Monod-type kinetic limitation term. In such cases, an estimation of the growth rate
would require knowledge of the corresponding Monod constant K. This value must
be determined from process data. For that reason, dissolved oxygen concentration
270 8 Bubble Column Bioreactors

measurements must be performed. In oxygen-limited process states one can assume

that the oxygen transferred into the reactor is immediately consumed by the cells, so
that no oxygen accumulation occurs. That means that the oxygen consumption rate
OCR is equal to the oxygen transfer rate OTR:

con st. _0_ = kLa(O* - 0). (8.52)

K+O
Let us assume that the kLa value can be estimated from off-gas measurements and
0*, the solubility, is available from the literature. Then, when the constant const. is
also available, K can be determined with a measurement value of 0, provided 0 is
homogeneously distributed.
When 0 is not homogeneously distributed, and that is why this problem has been
addressed at this point, one needs the probability density function p(O) in order to
determine the correct average of OCR. Since its oxygen dependency is nonlinear, one
cannot simply take the expectation value 0 of 0 and use Eq. (B.52). For the oxygen
consumption rate the correct averaging is

J
Omax A

0 0
OCR ex: p(O)--dO # - - A (B. 53)
Ko+O Ko+O
o

Since it is linear, the averaging procedure is not necessary for the OTR relation-
ship:

OTR =kLa(O*- 0) = E{kLa(O*- O)} (B.54)

From Fig. B.6 we just learned that the homogeneity assumption cannot be justified.
Hence, the average of the expressions in Eq. (8.53) must be determined using the
dissolved oxygen distribution p(O). Since this probability density cannot be properly
provided in practice, one must estimate the mean 0 and determine a rough estimate
Ka from Eq. (8.52).
Qualitatively, there is obviously an error when one uses the average value 0 in
Eq. (8.52). The question, however, is how big this effect is, quantitatively in typical
biotechnical applications, and whether or not it worth considering. This answer can
be given by numerical simulation, since in this case we have access to p(O).
As numerical experiments have shown, the answer largely depends on the actual
dissolved oxygen concentration probability density p(O). Different situations con-
cerning the oxygen supply in the reactor must be distinguished. When the oxygen
concentrations in the reactor are rather low and the corresponding probability dis-
tribution p(O) is sufficiently narrow, only the nearly linear "small-O-interval" in the
Monod-like oxygen uptake kinetics is affected. Hence, the influence of the nonlinear-
ity is small. The same applies for very high oxygen concentrations. The latter case,
however, does not often appear in industrial practice, since the corresponding oxy-
gen transfer rates would be too costly. In between these two extreme cases, the ef-
fects of the nonlinearities are expected to be significantly higher.
An easy way to characterize the error is to determine the ratio Ka by the K ob-
tained via a correct averaging procedure, which can be determined without specify-
ing the constant const.. When we consider the effects in the turbulent flows in bubble
column fermenters as depicted in Figs. B.l-B.5, then the quotient is shown in Fig. B.7.
8.4 Modeling of Particular Aspects of Bubble Column Reactors 271

12.

Kf~
0
~ 1.20

ii
·u 1.15

~
U
.§ 1.10
i'ii
~ 1.05
0.05 0,10 0.15 D.2<I 0.25 D.30
Oxygen Input Rate

Fig. 8.7. Ratio of K obtained with a correct averaging by Ko obtained using an estimate for the
average dissolved oxygen concentration and the value as a function of the aeration rates

As expected from the preceding discussion, the deviations are maximal for a med-
iate oxygen concentration interval and smaller at small and high oxygen concentra-
tions. At its maximum, at mediate aeration rates, the effect exceeds 20%.
At this point it is interesting to recall that the mean oxygen concentration 0 used
in the preceding discussion, which is the first moment of the distribution density
p(O), is nearly impossible to measure as the long-time averaged results in Fig. 8.6
indicate. With measured dissolved oxygen concentration values the effect is expected
to be significant larger. As the long-term average shows, there are considerable oxy-
gen concentration profiles to be expected in bioreactors, a fact that was measured
several times by different groups even in stirred tank bioreactors (e.g., [21].
This discussion essentially gives some idea of the errors in determining kinetic
parameters in inhomogeneous reactors under the wrong assumption that the reactor
would be well mixed.
This study clearly shows that the turbulent gas-liquid two-phase flow in biochem-
ical bubble column production reactors lead to chaotic patterns of the dissolved oxy-
gen concentration. An equilibrium in the sense of a homogeneous distribution of the
dissolved oxygen concentration cannot be obtained. In order to characterize the dis-
solved oxygen concentration pattern, two statistical measures can be used. The first
is the time averaged oxygen pattern as shown in Fig. 8.2. This pattern does not have
much relevance to the production process as such, but serves as a medium for com-
parison with measurements, which are usually time averages as well. The second
statistical measure is the dissolved oxygen concentration probability density, charac-
terizing the probability by which a yeast cell traveling through the reactor will see an
oxygen concentration 0 in its environment.
The probability density p(O) is highly dependent on the flow conditions, the local
mass transfer rate, as well as the oxygen consumption rate by the biomass. The ef-
fects of the oxygen concentration probability density p(O) on the integral oxygen
uptake by the culture is considerable at medium oxygen concentration where the
kinetics depicts its nonlinearities. Interestingly the influence of the bioreactor con-
struction shows that more regular global flow structures increase the effect. This was
demonstrated at flat bubble columns with rectangular cross section. Similar large-
scale global flow structures are to be expected in stirred tank reactors. Hence, it is
recommended to investigate this effect at stirred tanks as well.
What can be learned from the aspects considered in this section is that measure-
ments cannot provide accurate mean oxygen concentrations in larger bioreactors. The
272 8 Bubble Column Bioreactors

determination of kinetic parameters from measurements made in large reactors is

usually biased. Since the error is due to the fluid flow properties of the particular scales,
this bias is obviously scale-dependent. This means that values of kinetic parameters,
which are generally based on measured values in fInite reactors, are in most cases scale-
dependent values and cannot be used to an arbitrary accuracy for scale-up studies.

8.S
Conclusions
Modeling is understood as a quantitatively exploitable formulation of our current
knowledge about the processes under consideration. It takes a lot of effort and is
thus justifIed only when it does help to solve open questions of signifIcant impor-
tance. Different objectives require different models, since the models must consider
those mechanisms which are relevant to the problem to be solved.
The current problem with two-phase gas-liquid flows in bubble column bioreac-
tors is that we do have the model but cannot solve it to an arbitrarily chosen spatial
resolution. That means the problem lies in the possibility of us exploiting the model.
The decisive limitation is, and will be in the near future, the available computing
power to solve the Navier-Stokes equation system. Currently we can tackle detailed
fluid motions on a relatively rough scale only. However, as was shown in this chapter,
several basic questions about fluid flow could be answered with the currently avail-
able software.
The random nature of the flow can be shown quite directly. It has immediate con-
sequences for all effects which are usually discussed on the single particle basis. The
long-time averaged flow patterns do not have the importance that might be assumed
from their frequent appearance in fluid dynamical discussions in literature. It is of
importance only with respect to those questions which directly consider the long-
time averaged behavior of the fluid flow. An example of .practical importance is the
gas residence time distribution. Another more technical aspect is that comparisons -
may they be between computational results or between computational results and
measurement data - can only be performed on averages. All other problems in par-
ticular mass transfer and mixing, which are the key objectives of inducing flows in
bioreactors, are governed by the chaotic flow components.
The main activity to cope with coarse computational grids is compensating for the
fluid flow effects originating from motions on the small scale by means of appropri-
ate subgrid models. Since we cannot resolve the details of these fluid flow effects, we
must lump them together and project them into the material properties of the fluid,
namely the effective viscosity and the fluid density.
Although the current CFD results are qualitatively good enough to serve as basic
reactor models, they unfortunately cannot yet be incorporated into process optimi-
zation studies, simply because their evaluation takes too much time. With the com-
puters available, full three-dimensional dynamical simulations take several days and
thus cannot be used in the iterative procedures by which we optimize the reactor
performance. Such optimization procedures would require us to recompute the flow
patterns many times during a single optimization study. However, if the computing
power of our workstations continues to rise at the rates we have seen in the last
decade, the time at which we will be able to incorporate such models will not be
too far ahead.
References 273

Acknowledgement. Deutsche Forschungsgemeinschaft DFG generously supported my work on

fluid dynamics with several different projects over the previous year.

References
1. Feynman RP, Leighton RB, Sands M (1963) The Feynman lectures on physics. Addison
Wesley, Reading, Ma
2. Anderson DA, Tannehill JC, Pletcher RH (1984) Computational fluid mechanics and heat
transfer. Hemisphere/McGraw-Hill, Washington New-York
3. Liepmann HW (1979) The rise and fall of ideas in turbulence. Am Scientist 67:221-228
4. Patankar SV (1980) Numerical heat transfer and fluid flow. McGraw-Hill, New York
5. Sokolichin A, Eigenberger G, Lapin A, Liibbert A (1997) Dynamic numerical simulation
of gas-liquid two-phase flows: Euler-Euler versus Euler-Lagrange. Chern Eng Sci 52:611-
626
6. Abou-El-Hassan ME (1986) Correlations for bubble rise in gas-liquid systems. In: Cher-
emisinoff NP (ed) Encyclopedia of fluid mechanics, vol 3: Gas-liquid flows, chap 6. Gulf,
Houston, pp 110-120
7. Landau LD, Lifschitz EM (1971) Lehrbuch der Theoretischen Physik, vol VI, Hydrodyna-
mik, 2. Aufl., Akademie, Berlin
8. Lapin A, Liibbert A (1994) Numerical simulation of the dynamics of two-phase gas-liquid
flow in bubble columns. Chern Eng Sci 49:3661-3674
9. Lapin A, Maul C, Junghans K, Liibbert A (2000) Industrial-scale bubble column reactors,
gas-liquid flow and chemical reaction (to be published)
10. Nielsen J, Villadsen J (1994) Bioreaction engineering principles. Plenum, New York
11. Fan LS, Tsuchiya T (1990) Bubble wake dynamics in liquids and liquid-solid suspensions,
Butterworth-Heinemann, Boston
12. Lapin A, Paaschen T, Liibbert A (2000) Local mixing in bubble column reactors: Bubbles
as local mixing elements (to be published)
13. Lapin A, Liibbert A (1999) Problems with measurements of kinetic constants in bioreac-
tors (to be published)
14. Freedman W, Davidson JF (1969) Hold-up and liquid circulation in bubble columns.
Trans Inst Chern Eng 47:T251-T262
15. Liibbert A, Lapin A (1996) Dynamic numerical experiments with 3D-bubble columns.
Revue de nnstitut Francais du Petrole 51:269-277
16. Liibbert A, Paaschen T, Lapin A (1996) Fluid dynamics in bubble column bioreactors:
experiments and numerical simulations. Biotechnol Bioeng 52:248-258
17. Devanathan N, Dudukovic M, Lapin A, Liibbert A (1995) Chaotic flow in bubble column
reactors. Chern Eng Sci 50:2661-2667
18. Konig B, Buchholz R, Liicke J, Schiigerl K (1978) Longitudinal mixing of the liquid phase
in bubble columns. Ger Chern Eng 1:199-205
19. Rice RG, Littlefield MA (1987) Dispersion Coefficients for ideal bubbly flow in truly ver-
tical columns. Chern Eng Sci 42:2043-2051
20. Rice RG, Barbe DT, Geary NW (1990) Correlation of nonvertically and entrance effects in
bubble columns, AIChE J 36:1421-1424
21. Oosterhuis NMG, Kossen NWF (1984) Dissolved oxygen concentration profiles in a pro-
duction scale bioreactor. Biotechnol Bioeng 26:546
 Part C
Application of General Principles for
Process Models Including Control
9 Baker's Yeast Production
Karl-Heinz Bellgardt

9.1
Introduction
Yeast production is an important industry in several fields: besides baker's yeast,
these are animal feed yeast, yeast for hydrolyzates used in biotechnology and the
food industry, and in biotechnology as host organism for foreign genes. Thus there
is a wide application area of models as presented in this chapter, which is structured
as follows. An introduction to important phenomena of yeast growth is given first.
Only the important subjects for modeling and control will be summarized. A de-
tailed survey of yeast metabolism and the baker's yeast process can be found, e.g.,
in [1].
The next two sections are concerned with modeling of growth kinetics in ideally
mixed stirred-tank reactors (Sect. 9.2) and airlift tower-loop reactors (Sect. 9.3). The
baker's yeast production is an example for a process that cannot be described in all
its important aspects by simple growth models. Yeast metabolism is very flexible
with respect to usage of different substrates and oxygen, and several regulatory phe-
nomena have to be taken into account for modeling of growth. The proper approach
are structured models based on a representation of the metabolic reaction network
and a cybernetic criterion for its regulation.
The product quality of the final yeast product is influenced by the state of the cells
within their propagation cycle. This requires the extension of the growth models by
a population balance model to cover the distribution of cells in the cell cycle. Popu-
lation balances are introduced in Sect. 9.4. At first both models, the growth and po-
pulation models, are combined to provide information on the development of the
fraction of budding cells during fed-batch cultivations. Then emphasis is put on the
analysis of self-synchronized growth in chemostat cultivations.
The fraction of budding cells is then used in Sect. 9.5 as a product quality index to
evaluate optimal fed-batch control policies of the yeast production process with re-
spect to product quality and profit. The economic optimization requires a global
view of interconnected processing steps of the whole process and a detailed look at
the fermentation steps. Possibilities for automatic feed-back control of the substrate
supply by means of measurements in the liquid phase and exhaust gas are then dis-
cussed in Sect. 9.6.
278 9 Baker's Yeast Production

9.1.1
Metabolic Types of Yeast Growth and Regulatory Effects

Yeast can react on the availability of substrates such as carbon and nitrogen sources
or oxygen by a flexible choice of different metabolic pathways. A simplified map of
the metabolic network is shown in Fig. 9.1. Oxygen supply is one of the main deter-
mining factors of yeast metabolism and the limiting parameter for the production
process. In excess of oxygen, at low sugar concentrations, yeast can achieve an oxi-
dative metabolism. For energy production, the carbon source is completely oxidized
to carbon dioxide and water via glycolysis and tricarboxylic-acid cycle (TCC), with
almost no by-products. The reduction equivalents (NADH z) are transferred mostly
into the respiratory chain. These pathways can produce up to 36 moles of ATP per
mole glucose, resulting in a very high yield for cell mass of up to 0.5 g dry cell mass
per gram of sugar. Therefore, oxidative growth is the preferred metabolic type for
baker's yeast production, although the growth rate is limited to about 1L<0.25 h -1.
Ethanol can also serve as a sole substrate for oxidative metabolism. The related en-
zymes of gluconeogenesis are inducible, which usually results in a lag phase of
growth after a switch from growth on sugars to ethanol. Although glucose is the
preferred substrate, parallel ethanol uptake is also possible up to a maximum rate
that is determined by the availability of glucose and oxygen.
On sugar substrates, a second metabolic type is fermentative metabolism. This can
be due to complete absence of oxygen under anaerobic conditions, due to partial
oxygen limitation under aerobic conditions, or by a saturation in the rates of oxida-
tive pathways (oxidoreductive or respiro-fermentative metabolism). Then, ethanol is
produced as a final proton acceptor for NADH z. Glycerol, acetaldehyde, and organic
acids can be found in small amounts as by-products. The chemical energy then
comes mainly from glycolysis with a yield of 2 moles ATP per mole of sugar, plus
2 moles of ethanol and CO 2 as by-products. By this low efficiency compared to oxi-
dative growth and due to the loss of much carbon in the products ethanol and COl>
the cell yield is also very low. Under anaerobic conditions, the yield reaches only

Fig. 9.1. Simplified map of yeast metabolism

9.1 Introduction 279

Yxs=O.l and the growth rate stays below {l=O.3h-1, compared to {l<0.5h- 1 for oxi-
doreductive growth. The production of a large volume of carbon dioxide is used in
baking to leaven the dough. The fermentative activity - the ability to produce a
certain amount of CO 2 in a given time - is, beside storage properties and resulting
flavor of the dough, an important quality index for the yeast product.
The respiratory quotient can be used to estimate the metabolic type of yeast
growth. For RQ> 1 growth is fermentative, RQ= 1 indicates respiratory growth on
sugar substrate, and RQ< 1 can be found during ethanol consumption.
For the directed coordination of its metabolism in dependence on substrate and
oxygen supply the yeast disposes of a complex regulatory system on the epigenetic
and reaction levels. This regulation affects uptake of carbon sources, glycolysis, glu-
coneogenesis, respiration, and product synthesis.
The Pasteur-effect - suppression of fermentative activity by respiration - is the
earliest regulatory effect discovered in yeast. It is connected to the repression of en-
zymes of the respiratory chain and TCC under fermentative growth. Since the induc-
tion level under aerobic conditions depends on the available oxygen, a sudden in-
crease in oxygen supply usually leads to short lag phase until the enzyme level is
adjusted to the new condition and the metabolism is directed towards more effective
energy production.
The Crabtree-effect is referred to as oxidoreductive metabolism under excess of
oxygen and sugar, caused by a saturated capacity of the respiratory system of the
cell. Metabolic fluxes are directed towards fermentative pathways with ethanol pro-
duction. This is very important for the control of baker's yeast production because
an over-supply of sugar in relation to a limited capacity of oxidative metabolism -
either by internal saturation or by external oxygen supply - leads to fermentative
type of growth with low yield of cell mass.
The Glucose-effect is the repression of uptake systems or pathways needed for
other substrates at high concentrations of glucose, the substrate that can most easily
be used by the cells. As a result of this repression, glucose is preferred and catabo-
lized first, even when other substrates are present at high concentrations. For baker's
yeast production, the glucose-effect with respect to ethanol - as a possible substrate
for oxidative growth - is most important, because ethanol can be either produced or
metabolized during the cultivation. Ethanol and sugar can even be taken up in par-
allel, then both compete for the oxidative pathways.

9.1.2
The Asymmetric Propagation of Yeast

The baker's yeast Saccharomyces cerevisiae is a single-cell micro-fungus. Cells are

spherical to ellipsoid with a diameter ranging from 2.5 {lm to 10.5 {lm and a length
from 4.5 {lm to 21 {lm. The vegetative multiplication by budding is most important for
its propagation in technical processes. The budding cycle of yeast is a dynamic pro-
cess of a series of ordered steps as schematically shown in Fig. 9.2: growth in mass
and/or volume of single cells in the intermediate phase Gl> DNA-synthesis phase (S),
a phase of bud formation and bud growth including mitotic phase (G 2 , M), and a
further intermediate phase G1* before separation of parent cell and bud, which be-
comes the new daughter cell. The period for the completion of the cycle is asymmetric
for parent and daughter cells. These are smaller than the parent cell and must increase
in size before initiation of chromosome duplication [2, 3]. For this reason, a newly
280 9 Baker's Yeast Production

D B

,if Daughler cell

Bud

B
r
AppeBrance d Ihe bud
Separation of tho ood

Fig. 9.2. Schematic diagram of the asymmetric budding cycle of yeast with single cell phases
of daughter cells (D) and parent cells (P), and double cell budded phase (B)

formed daughter cell needs more time in the Gl-phase for the formation of its first bud
than for the subsequent buds. Therefore, yeast populations are heterogeneous mixed
populations with a certain distribution of distinguishable daughter cells, parent cells,
and budding cells. The duration of the several cycling phases determines the genealo-
gical age distribution of the yeast population. The genealogical age of the cell can be
easily determined by counting the number of bud scars on the cell surface. For expo-
nentially growing cultures these data was used to calculate the age distribution of a
population and the duration of the cell cycle phases [4, 5]. The duration of the Gl -
phases of parent and daughter cells is positively correlated to the cell number doubling
time. In contrast, it is found that the division process - the duration of the budded
phase (G2 +M+G l*) - is almost constant and independent of the external growth con-
ditions. Only for very slow growth does the length of this phase increase slightly.
Hartwell and Unger [3] found that the length of the budding interval and the genera-
tion time of parent cells do not vary with the genealogical age, at least not for a few
generations. In addition, the budding intervals of daughters and parents are practi-
cally identical in length and undergo only minor changes for variations in the growth
rate. The control mechanism for the cell cycle of Saccharomyces cerevisiae in batch
culture can be smoothly explained by the concept of critical mass for the initiation
of division. In this concept, the generation time of daughter cells is longer than that
of parent cells, because they must first accumulate more cell mass. Since the size of the
bud at separation becomes smaller at low growth rates, the enforced increase in
daughter generation time with average doubling time is evident.
Based on these findings, the whole cycling process of yeast cells can be divided by
a simplified view into three cycling phases (see Fig.9.2): unbudded daughter cell
phase D with length T[» unbudded parent cell phase P (length Tp), and budding
phase B (TB ), which has equal length for both budding daughter cells and budding
parent cells. The experimental results demonstrated [3,4] that T[» T~ and TB can be
linearly correlated with the average cell number doubling time T, even for a variety
of strains, different temperatures, and substrates. The lengths of the cell cycle phases
can be described by the following simple correlations, the cycling phase equations:

TD = KDl T + Km
Tp = Kpl T + KP2 (9.1 )
TB = KBl T + KB2
9.1 Introduction 281

~,-------------------------,

To
_N
E.~

Tp

2 4 6 8 10 12
T [h]

Fig. 9.3. Graphical representation of the cycling phase equations [5]: lengths of the cell cycle
phases (TIft Tp, TB ) over averaged doubling time T

Table 9.1. Experimentally determined parameters in the cycling phase equations for Sacchar-
omyces cerevisiae S288c/1

KDJ KD2 Kp1 KP2 KBJ KB2 Remark

1.302 -1.30h 0.440 -0.367h 0.180 0.767h batch culture [4]

0.330 0.05h 0.250 -0.233h 0.710 0.100h T<3.17h
1.600 -3.93h 0.390 -0.433 h O.llO 1.800 h T> 3.17 h, continuous
culture [5]

Model parameters from the literature are given in Table 9.1. For an extended range
of growth rates in chemostat cultivation, Thompson and Wheals [5] found biphasic
correlations for the generation times of parent and daughter cells and the length of
the budding phase as shown in Fig. 9.3: for T<3.2 h, growth is practically symmetric
with TD=Tp< 1 h, and TB<2.5 h, the length of the cycling phases are nearly propor-
tional to T. For T>3.2h, growth becomes asymmetric and TD is more and more
prolonged at the cost of TH while TB is almost constant at about 2.5 h. The para-
meters for both regions of doubling times are also given in Table 9.1.
There were a lot of investigations to clarify the interconnection of metabolism and
cell cycle, which are summarized in [6]. It was found that most of the proteins, cell
wall components, and RNA are synthesized with a constant rate during the cell cycle.
The activity of most enzymes is also not strongly modulated or their concentrations
are much higher than needed for the catalytic activity. So, while most cellular pro-
cesses may be considered as continuous, the periodic events originate from the re-
plication and partitioning of the DNA and the separation of the bud. The rate-limit-
ing step of the cell cycle is obviously the growth process, i.e., the rates for cell mass
or volume increase, which are directly connected to the continuous processes of me-
tabolism. The cells need to reach a certain mass or volume to initiate the budding.
This also explains why the G1-phase of the first generation is longer than that of
parent cells. The genetic program for cell division can normally be executed faster
than the accumulation step. Growth and division are coordinated by monitoring the
level of key metabolites, which then trigger the start of the next cell cycle phase.
Alberghina et al. [68] and Mariani et al. [69] investigated the protein distribution
in the cell cycle by means of extended population balance models. They proposed
that the protein content of the cells controls bud initiation and the critical protein
282 9 Baker's Yeast Production

level increases at each generation of parent cells. The dominant role of the mass
accumulation facilitates modeling since the continuous increase of mass can be de-
scribed by structured growth models averaging over the population. The population
dynamics are then influenced mainly by the average growth rate.
A very special feature of baker's yeast is self-synchronized growth in continuous
culture, where oscillations appear on population variables such as the fraction of
budding cells, or on process variables like the rates of oxygen uptake and carbon
dioxide evolution, even when the cultivation conditions are kept constant. It has
been subject to several investigations that were reviewed in [32, 33]. Self-sustained
oscillations that are directly related to the asymmetric budding cycle of yeast are
mostly found in the region of oxidative growth. These can appear spontaneously or
be induced by shifts in the dilution rate or by small pulses of substrates such as
glucose, ethanol, acetic acid, and pyruvic acid. The oscillation period varies in gen-
eral with growth rate, but eventually multiple oscillation frequencies at a single
growth rate could be observed. During synchronized growth, the variation of the
metabolism during the cell cycle becomes visible on macroscopic variables. This
facilitates experimental investigations of the related intracellular biochemical pro-
cesses and physiology [7]. At the beginning of the budding phase, the metabolism
is oxidoreductive with ethanol production. The ethanol is reconsumed in the later
stages of the cell cycle. Also a periodic built up and degradation of storage material
and proteins is found during the cycle. For sustained oscillations, a feed-back me-
chanism must be assumed, which controls the cell cycle by variations in the concen-
trations of substrates or other growth factors. Many works lead to the conclusion
that ethanol together with glucose is responsible for the stabilization of synchronous

800 50
=111,""" pmoo of
4.5 i - - - - -_ _ _ _---.!rrrde~~II'_n_--,,---6!L.----l
700
4.0
" "",Ie 112
3.5 1>". _" _" _____ . ___ " _. " ___ __ _______ ______ . __

8
40

20

Fig. 9.4. Experimental data for self-sustained oscillations in continuous culture of Saccharo-
myces cerevisiae at D=O.097h- 1 during mode change from 12 to 11: (- - - - - -) agitation speed,
C·····) dissolved oxygen, (--) CO 2 in exhaust gas, (.1.) experimental period length in compar-
ison to the expected values. Reprinted by permission of Wiley-Liss. Inc., a subsidiary of John
Wiley & Sons, Inc. from [34] "Oxygen, pH and carbon source induced changes of the mode of
oscillation in synchronous continuous culture of Saccharomyces cerevisiae". Copyright © 1999.
9.2 Growth Modeling 283

oscillations. The oscillations can only exist when a periodic switch from ethanol
uptake to ethanol formation is possible during the cell cycle. Under reduced dis-
solved oxygen concentration, the oscillations disappeared or oscillation periods are
greatly prolonged and rather irregular, because ethanol uptake is hindered. An ex-
ample for self sustained oscillation during continuous cultivation of yeast is given in
Fig. 9.4.
The quality of the yeast product is influenced by fermentative activity, storage
material, and protein content. Since all these vary over the cell cycle, controlled syn-
chronization and harvest at a certain point in the cell cycle may lead to increased
product quality. This property can be used for determination of optimum fed-batch
control policies as shown in Sect. 9.5.

9.2
Growth Modeling

A model of growth kinetics in bioprocesses includes the two sub-models of the re-
actor, i.e., of the abiotic phase, and of the population of microorganisms, the biotic
phase. This section is concerned with growth of baker's yeast in stirred tank reactors
under batch, fed-batch, and continuous operation. The general reactor model for
stirred tank reactors is given in Chap. 1 in Table 1.3. For the yeast process, it includes
the main components of the liquid phase: cell mass, substrate (e.g., glucose or mo-
lasses), ethanol, dissolved oxygen, and dissolved carbon dioxide. The vector of con-
centrations and specific reaction rates are then

(9.2)

The substrate flow rate F is the main manipulating variable of the reactor during
fed-batch operation. It determines the increase of the volume of the liquid phase,
Eq. (1.9), and the dilution of dissolved substances. For an accurate description of
oxygen limited growth, the model of the gas phase (Table 1.3, f and g) must also be
included.
A review of models for baker's yeast is given by Nielsen and Villadsen [8]. The
models range from very simple compartment models up to complex structured ones
[9, 10]. Cybernetic models (see Sects. 2.3.3 and 2.3.4 in Chap. 2) are well suited for
baker's yeast because they simplify greatly the modeling of the complex growth phe-
nomena and related regulatory effects. The cybernetic model of the biotic phase
presented here [45, 47, 85] is built of three parts: the stoichiometric model for a
representation of the metabolic reaction network, a cybernetic criterion for determi-
nation of the optimal flux distribution in the network under consideration of the rate
limiting steps of metabolism, and a dynamic model for slow regulation by induc-
tionlrepression of key enzymes for some of the rate-limiting steps.
284 9 Baker's Yeast Production

9.2.1
Stoichiometric Model

The stoichiometric model is based on the simplified map of the metabolic network
of yeast as shown in Fig.9.1. It includes the following building blocks: glycolysis,
tricarboxylic-acid cycle - approximated by a linear chain of net-throughput of pyr-
uvate to CO 2 , respiratory chain, gluconeogenesis, and product formation for ethanol,
CO 2 , glycerol, and acetaldehyde. Synthesis of macromolecular cell constituents is
assumed to start only from fructose-6-phosphate and TCC, by this representing
quite a number of reactions that were analyzed in more detail in [11, 12]. The very
simplified modeling of the TCC also requires a simplified view of the metabolism
during growth on ethanol. As accurate investigations have shown [l3, 14] a substrate
shift from 100% sugar to 100% ethanol proceeds via at least four different metabolic
flux regimes that also influence several reactions in the TCe. Nevertheless, the sim-
plified models are also quite well in agreement with experimental data. The pentose-
phosphate way was neglected for the model since it is of greater importance for
other substrates such as glucose. These simplifications are compensated for during
parameter identification by slight deviations in some model parameters. The meta-
bolic map can be represented by a system of reactions as given in Table 9.2. From
that scheme of reactions the stoichiometric model can be derived by establishing
balances for each considered metabolite. According to the principle of optimality
(see Sect.2.1.2.2 in Chap. 2), no surplus intracellular metabolites are accumulated
during the growth process, so that for a modeling of cultivation processes with a
relatively long time horizon the reaction model can be introduced as quasi-station-
ary. For investigation of short-term effects and intracellular kinetics, dynamic mod-
eling of cellular reactions may be of interest [15-17].
Before establishing the stoichiometric model, further simplifications are intro-
duced. The biomass is assumed to have constant composition and consist only of
structural compounds. Then the rates of both biosynthetic pathways must be pro-
portional to the specific growth rate,

rMal = KMa1Ji-
(9.3)
rMaZ = KMaZJi-

Ideally, the parameters K Ma1 and K Ma2 are not independent of each other but
should satisfy the condition

(9.4)
Since the model reflects a simplified view of the metabolism, the fitting of experi-
mental data can be improved by taking both as independent parameters. The energy
consumption for growth and maintenance requirements, rAm is modeled in analogy
to Pirt's concept, Eq. (2.24) in Chap. 2, as

(9.5)

The rates of glycolysis, rpl, and gluconeogenesis, rGy> are joined to only one rate,
and the differences in ATP-yield are considered by a switch function R(rp)
rp=rFl-rGy>
in the ATP-balance.
9.2 Growth Modeling 285

Table 9.2. Metabolic reactions included in the stoichiometric model

Step Reaction Remarks

a Glucose uptake, 1st CLU +ATP ~ F6P+ADP F6P and G6P are as-
phosphorylation and sumed to be in equili-
isomerization brium

b 2nd phosphorylation F6P+ATP ~ FDP + ADP

c Biosynthetic pathway 1 F6P + QATP . ATP representative reaction

~ QMa! . MAC! + QATP . ADP
TGy
d Glycerol synthesis !FDP + NADH2 -+ CLY + NAD+ TCy can be modeled as
starting from F6P with
respective modification
in the ATP balance
e Glycolysis FDP + 4ADP + 2NAD+

~ 2PYR + 4A TP + 2NADH2

f Acetaldehyde synthesis PYR ~ ACD + CO2

g Acetaldehyde excretion ACD ~ACDex

h Ethanol synthesis ACD + NADH2 ~ EtOH + NAD+

Ethanol uptake EtOH + NAD+ ~ ACY + NADH2 simplified to only one

step
Acetyl-coA synthesis PYR + NAD+ ~ ACY + NADH2 + CO2

k Biosynthetic pathway 2 ACY ~ QMa2MAC2 representative reaction

Tricarboxylic-acid cycle ACY + ADP + 4NAD+ Tee is approximated by

linear net-reaction, FAD
~ 2C02 + ATP + 4NADH2 is considered as NAD

m Respiratory chain O2 + 2P /0 . ADP + 2NADH2 FAD is considered as

NAD
~ 2H20 + 2P/0· ATP + 2NAD+

n Gluconeogenesis 4ACY +4ATP via malate, oxalacetic

acid, phosphoenol-pyru-
.':3!... F6P + 4ADP + 2C02 + Pi vate, and FDP
286 9 Baker's Yeast Production

The formation of by-products acetaldehyde and glycerol is according to experi-

mental findings correlated to the ethanol production,

fAdE = KAdTEI
(9.6)
TGy = 2KEGTEI

Both rates were not independent if the simplified map of the metabolism would be
exact. Then, since NADH must be balanced, the following relation can be derived for
anaerobic growth

(9.7)

One reason for deviations from Eq. (9.7) is a possible excretion of other by-pro-
ducts.
For modeling of technical cultivations, the knowledge of the intracellular rates TG,
TAd, and TTCC is of less interest. So they can be eliminated from the system of equa-
tions of the stoichiometric model for simplicity. By further use of the conditions at
Eqs. (9.3), (9.5), and (9.6), the stoichiometric model can be written with molar spe-
cific reaction rates in the notation of Eq. (2.81) in Chap. 2 as

Yr=m (9.8)

including the following matrix and vectors:

K Mal 0 o 0
KEG )
-4KMaZ 5 -2 1 -1 - 2KEG - K Ma3
-KMZ-_l-
a YATP
2P/O 0 -2KEG
o -1 o -1 - KAD

Tp

r= (9.9)

The net reaction rates q, defined by Eq. (2.1) together with Eq. (9.2), appearing in
the reactor model, are then given by

(9.10)

Although the stoichiometric model may cover most of the complex growth phe-
nomena of yeast, it contains - when considering the conditions of Eqs. (9.4) and
(9.7) - only five independent stoichiometric parameters, which all have a direct bio-
logical interpretation.
9.2 Growth Modeling 287

The sugar uptake system is assumed to be constitutive. Therefore, the sugar uptake
rate rs is included into the vector m of growth independent reaction rates and mod-
eled according to Monod-kinetics as
Cs
rs = rs max - - - (9.11)
Ks + Cs
although an accurate analysis shows that it is also controlled by ATP [15, 18], espe-
cially at high glucose concentrations. There is also evidence in the literature of dy-
namic regulation of sugar uptake in medium and long time scales above 1 h that may
be relevant for cultivation processes under excess of substrate [19, 20]. This is also
neglected here for the sake of simplicity. For growth on substrates of natural sources
such as molasses, the uptake of other sugars beside glucose is important. A respec-
tive model including fructose and sucrose was published by Barford et al. [21,22]. In
batch cultivation the uptake of these sugars proceeded in parallel at different rates,
while for maltose a sequential uptake was found [23]. A cybernetic model for adap-
tation during yeast growth on galactose and melibiose was published by Gadgil et al.
[24]. Under sugar limited growth in fed-batch cultivation all sugar components are
normally used immediately and it is then not necessary to discriminate between
them.

9.2.2
Cybernetic Modeling of Metabolic Regulation

In the cybernetic modeling, the growth kinetics are determined by an optimum

strategy for the metabolic regulation. When applying the metabolic regulator ap-
proach (see Sect. 2.3.4 of Chap.2) for the modeling of yeast growth kinetics, the
optimum strategy of maximization of the specific growth rate, Eq. (2.83), was found
to work quite well:
r(t)
p,(t) ) maximum (9.12)
Eq.(2.84)

The solution of Eq. (9.8) for r(t) by the above optimization procedure, also called
the metabolic coordinator, has to consider constraints by additional restrictions for
possible rate limiting reactions as given by Eq. (2.84). The inherently rate limiting
steps in the model are on one hand the uptake steps for ethanol rE2' and oxygen ro
For simplicity, they are introduced as first-order kinetics. In conjunction with the
other constraints this leads to Blackman-kinetics for these uptake steps. For high
ethanol concentration inhibition effects should be considered in the uptake kinetics
[25]. A careful investigation of oxygen-limited growth also reveals a more compli-
cated nature of the oxygen uptake being dependent on the intracellular ATP levels
[26]. This can even cause hysteresis effects with respect to variations in the dissolved
oxygen concentration. The intracellular rate-limiting steps on the other hand are the
limited capacity in the oxidative pathway, rAcmax' the dynamic regulation of respira-
tion rOmax(t), and of gluconeogenesis, rPmaAt). For each of the latter two reactions,
metabolic regulators are introduced into the model as given below. Beside the TCC-
start-reaction, rAcmax, as assumed here, alternative mechanisms for causing oxidore-
ductive growth are also possible, e.g., saturation of the respiration. This slightly dif-
ferent view of the Crabtree-effect is adopted in the model of Sonnleitner and Kappeli
[28]. Such a mechanism could also easily be used instead in the constraints given
288 9 Baker's Yeast Production

below. But rate-limiting steps may also be located within the TCC [27]. With these
assumptions, the minimum and maximum reaction rates in Eq. (2.84) are as follows:

-rPmax(t) regulation of gluconeogenesis

-00 specific growth rate, not restricted
o entry of TCC, irreversible
(9.13)
rmin =
o oxygen uptake only
o ethanol uptake only
o ethanol production only

00 glycolysis, not restricted

00 specific growth rate, not restricted
rAcmax Crabtree effect
rmax =
min(KoCo , romax(t)) 1st order oxygen uptake kinetics and regulation
KECE 1st order ethanol uptake kinetics
00 ethanol production, not restricted
(9.14)

Since the oxidative metabolism is most effective with respect to energy yield
(ATP), the coordination strategy for the metabolism, Eq. (9.12), directs as much as
possible of the substrate fluxes into the tricarboxylic-acid cycle, as long as this is
below saturation (rAcmax) or the respiratory chain can regenerate all of the produced
NADH z. This also gives an explanation of the Pasteur-effect. The rate of the respira-
tory chain is restricted by the oxygen supply or by the actual induction state of its
key enzymes. When the sugar flux cannot saturate the capacity of these pathways,
ethanol is used as substrate, when available, up to that limit. Here, the glucose effect
is explained by the lower energy yield from ethanol compared to hexoses. In con-
trast, when the sugar flux exceeds the capacity of oxidative metabolism, the surplus
amount is directed to ethanol production and growth becomes fermentative. Then
energy production becomes less effective, but nevertheless the growth rate can be
increased further by using the additional ATP produced during glycolysis. The
switch from ethanol production to uptake into the TCC proceeds instantaneously; a
lag-phase is only observed when gluconeogenesis is being involved [29].
The above metabolic model is a static one for the instantaneous adaptation of the
cells to the actual supply of substrate and oxygen. By the action of the metabolic
coordination, Eq. (9.12), it covers the fast metabolic regulation on the reaction level
by activation/inhibition. The model can already be used for simulation of steady
states in chemostat cultivations, as presented in the next section. But the respiratory
activity and the pathways for gluconeogenesis are subject to long-term regulation by
enzyme induction and repression. The resulting lag-phases of growth during phases
of regulatory adaptation must be described by corresponding dynamic models: two
metabolic regulators for the oxidative capacity and induction of gluconeogenesis
during growth on ethanol. The latter one is modeled according to Eq. (2.85) as

drPmax(t)
dt = (KIP - f.k)rPmax(t) + Kzp(R( -rp(t)) + rpmin) (9.15)

In extension of the presentation in Sect. 2.3.4 of Chap. 2, here the metabolic regu-
lator of respiratory activity is established as a dynamic system of third order, which
9.2 Growth Modeling 289

Table 9.3. Model parameters used for simulations in Figs. 9.5-9.7

Parameter Fig. 9.5 Fig. 9.6 Fig. 9.7 Units

rSmax 0.019 0.016 0.018 mole (ghr 1

Ks 0.8 0.45 0.45 gr 1
KE 0.5 1.1 1.1 I (gh)-l
Ko 12.8 I (mgh)-l
KMal 0.0075 0.0053 0.002 mole (ghr 1
KMa2 0.0 0.0044 0.007 mole (ghr 1
KMa3 0.0002 0.39 0.016
KEG 0.036 0.05 0.05
KAd 0.033 0.08 0.07
PIO 0.82-1.5 0.93-2.7 1-2.5
YATP 10.2 12 12.9 g mole- 1
mATP 0.002 0.0025 0.002 mole (ghr 1
rAcmax 0.0028 0.0034 0.016 mole (ghr 1

on
""! on
.... 0
-
N

'" '"
qo
.,.v \J
'" 0

'"
0
N
'"
0 .,.
C r,J rS
on
0 ~ ~
I

9 0.11 0.19 0.27 0.35 0.43

D[h- I ]

Fig. 9.5. Simulation (lines) of steady-states of chemostat cultivation of Saccharomyces cerevi-

siae H1022 on a mixture of 1.5% glucose and 1.5% ethanol in comparison to experimental data
(symbols) of Rieger et aI. [48]: concentrations in kgm- 3 of glucose Cs, ethanol CE (e), cell
mass Cx (6), specific rates in g(g hrl of oxygen uptake qo (... ) and carbon dioxide produc-
tion qc. Reprinted from (45) with permission by VCH Weinheim

shows a damped oscillatory behavior, as can be seen in Fig. 9.6. The model is in state
space representation

dy(t)
-----;[t = F(J.-t)y(t) + f(J.-t) (ro(t) + rOmin) (9.16)

where the elements of the state vector

omax
y = ( r CEl ) (9.17)
cEl
290 9 Baker's Yeast Production

'" ~
'" ""
tJ
0
~
'"
rJ
r::;

'" I ~

r:J

., 0
0 11
0
0
b

I
'-
~

i2•
<0

~ '" i2

0 0
0
I 11 22
t [hI

Fig.9.6a,b. Simulation (lines) of diauxic growth in batch cultivation of Saccharomyces cerevi-

siae HI022 on glucose in comparison to experimental data (symbols): a concentrations in
kgm- 3 of glucose Cs (e), ethanol CE (0), cell mass Cx (8), and dissolved oxygen Co [gm-3]
(.); b specific oxygen uptake rate To and output of the metabolic regulator rOmax in mo-
le(g hf 1, specific rate of gluconeogenesis -rp and output of the metabolic regulator rPm ax in
mmole (ghfl. Reprinted from [45] with permission of VCH Weinheim

are the maximum specific oxygen uptake rate, TO max, and intrinsic concentrations of
two fictitious enzymes, CEl and CE2' Due to the ft-proportional dilution terms in in-
trinsic balances, Eq. (2.64), the elements of F and f are non-linear functions of the
specific growth rate ft [85], and can be written as follows:

~ 0~) f
o
= 0.27 ( 2ft
ft2
! 0.36)
+ 10.9
(9.18)

where ft is in h- 1 and TO max , TO in mmole(ghrl. The output variables of the regula-

tion models, Eqs. (9.15) and (9.16) are as possible rate-limiting steps of the metabo-
lism included into the constraints, Eqs. (9.13) and (9.14).
The assumption of a minimum respiratory capacity as described by the parameter
rOmin in Eq. (9.16) is supported by the investigation of Sonnleitner and Hahnemann
9.2 Growth Modeling 291

~ ...,
0 0
on

\I") ~

c
.2 .2 ~
..
0
.
.....,
....
......
.,
....0
....c
.,
0

c en
g

(.)
"
"0 F
15'"
(.)
c c
'"
.S g .... 00
a;
-0
.,\1") .,
(.)-

~o '0c en
0
0
(.) ~
....
.t=
~ c:;"

20 25 30
t[h]
Fig. 9.7. Simulation (lines) of a batch/fed-batch cultivation of Saccharomyces cerevisiae H620
on molasses and experimental data (symbols): concentrations in kg m- 3 of molasses Cs (*),
ethanol CE (~), cell mass Cx (0), and substrate flow rate F [I h- I ]. Reprinted from [49]

[30]. It was also found that ethanol can have a very long lasting negative effect over
several days on the maximum oxygen uptake rate. This was not included in the above
model. After prolonged adaptation the respiratory capacity may even be increased
over the value corresponding to saturation of the respiratory system [31]. The above
model is much simpler than the more detailed structured models including a num-
ber of intrinsic balances presented by Steinmeyer and Shuler [9] and Coppella and
Dhurjati [10]. But an accurate description of a wide range of process conditions that
are important for baker's yeast production is provided.

9.2.3
Application of the Model for Simulation of Batch, Fed-Batch, and Continuous Cultivations

In this section, simulation results with the above model for continuous, batch, and
fed-batch cultivations of baker's yeast in laboratory-scale stirred-tank reactors are
presented. The model parameters used in the simulations are summarized in Ta-
ble 9.3. The simulation results for growth of yeast in chemos tat culture on a mixed
substrate of l.5% glucose and l.5% ethanol and experimental data of [48] are given
in Fig. 9.5. At low growth rates, controlled by low dilution rates, both substrates are
used up completely for growth by oxidative metabolism. For D>0.23 h -1 some of the
ethanol is left in the medium because of the limited capacity of oxidative metabolism
by saturation of the TCC (rAcmax)' In this range of dilution rates, glucose is the pre-
ferred substrate due to the higher efficiency in energy production. The increase in
ethanol is accompanied by a decrease in cell concentration. For D>O.3 h -1 even etha-
nol is produced since the flux from glucose alone already exceeds the oxidative ca-
pacity of the cells. All these phenomena are predicted by the metabolic coordinator
that maximizes the growth rate with respect to the actual inherently rate-limiting
steps. With usual structured models for growth on multiple sugars the effect cannot
be described so smoothly [46].
Figure9.6 shows simulation results from [45] for an aerobic batch cultivation in
glucose medium in comparison to experimental data of Kuhlmann (cited in [45]). In
292 9 Baker's Yeast Production

Head

. ,,
,
24 m

11

Riser Down·
comer

,
,
,,, ,,
,, ,,
(.I

Fig.9.S. Schematic drawing of the pilot-scale 5_m 2 airlift-tower-loop reactor for fed-batch cul-
tivation of baker's yeast [52]

the first growth phase up to 5.5 h, the sugar concentration is high and ethanol is
produced due to the Crabtree-effect. The yeast is growing on the preferred substrate
glucose and the uptake of ethanol is suppressed by the glucose-effect until all sugar
is used up. This behavior leads to the well-known growth-diauxy of yeast. There are
two mechanisms for suppression of ethanol uptake: inhibition by the metabolic co-
ordinator and repression of the key-enzymes for gluconeogenesis. By the latter effect,
rpm ax stays at low values. In the transient phase when the metabolism is switching
from glucose to ethanol growth, both regulatory systems for rp and ro become rate
limiting for a short period, and therefore their activities are induced. But growth is
mainly limited by low oxygen concentrations during the second growth phase with
few exceptions between t=14 hand 16 h, where the actual capacity of the respiratory
system is at the limit. At t=12, 13, 14, and 15 h the agitation speed was increased to
follow the higher oxygen demand by growth of cell mass. During growth, the energy
requirements for maintenance are covered by the substrate catabolism. In the starva-
tion period, beginning from t= 18 h, the cell mass concentration decreases. Here the
model predicts a degradation of cell material to produce energy as required for
maintenance by the rate mATP'
Figure 9.7 presents a combined batch and fed-batch cultivation of Saccharomyces
cerevisiae strain H620. After the end of the diauxic batch phase at t=15 h, substrate
flow is switched on and later increased further according to the profile shown. The
aim of a fed-batch control is to ensure purely oxidative growth without ethanol pro-
duction by limited feeding of the sugar. During oxidative growth, yeast has high
requirements for oxygen of about 1 g O2 per g of dry cell mass. The profile of the
molasses feed in these stages has to be adjusted to meet the oxidative capacity of the
cells or the maximum oxygen transfer of the reactor. If one tries to keep the growth
rate as high as possible, it follows more or less a two-phase feeding scheme: in the
first exponential growth phase under excess of oxygen a constant specific growth
rate of about 0.25 h -I - just at the critical value for the Crabtree-effect - is set by an
exponentially increasing substrate flow. In the second phase the growth rate has to
9.3 Growth in Airlift Tower-Loop Reactors 293

Table 9.4, Correlations for fluid-dynamic parameters in the reactor model, Eqs. {9.63}-{9.70}

Parameter Correlation Section Reference

Dispers.ion. coefficient
of the hqUld phase
.~
DLJ ~ 1.23 em
2
5
-1 (dem
j ) 1.5 ( UGj
em 5- 1
) 0.5 j=r,d,k
all sections
Towel and
Ackerman a

Volumetric mass
transfer coefficient (kLa)r= 2ULr ( UGr ) 0.87 ( 1 + Ad )-1 riser Bello et al. [58]
Hd ULr Ar

downcomer StrauB [52]

head Kastaneka

Gas holdup UGr (z) riser Hills [59]

EGr(z) = 093
0.24 + 1.35 uG,(z)
ms-
+ (UL'(Z))
ms-
l i
.

downcomer Bello et al. [58]

UGk(Z)) 0.775
EGk(Z) = 0.63 ( m 5- 1 head Weiland [60]

"Recommended in [57]

be limited by an almost constant feed-rate according to the maximum oxygen trans-

fer capacity into the reactor to avoid oxygen limitation and ethanol production. An
optimum production with maximum yield and good productivity can only be
achieved by automatically controlling the substrate feed just at the critical value for
the onset of the Crabtree-effect. This will be further discussed in Sect. 9.6.

9.3
Growth in Airlift Tower-Loop Reactors

In this section the model is applied for baker's yeast fed-batch production in a pilot-
scale airlift tower-loop reactor. In this type of reactor aeration is very energy effi-
cient, but inhomogeneities must be expected in the liquid phase, especially for the
concentrations of substrate and oxygen, that can influence the growth and be very
critical for baker's yeast production. The reactor is schematically shown in Fig. 9.8.
Among the three sub-systems - riser, downcomer, and head - only the head space
has a variable working volume due to the substrate feed. To cover extended fed-batch
periods, its height and diameter are unusually large for airlift reactors. Substrate
feeding is distributed along the height of the riser. The model is used to analyze
the process and study possible directions for its optimization.
With respect to high productivity, the process has to be operated at the critical
substrate feed rate for switch of metabolism to oxidoreductive growth with ethanol
294 9 Baker's Yeast Production

.8

10 I~
"
I [hI

I [hI

Fig.9.9a,b. Simulation (lines) and experimental data (symbols) of a fed-hatch cultivation of

Saccharomyces cerevisiae in the airlift tower-loop reactor [52] : a spatially averaged concentra-
tions of glucose, ethanol, cell mass (0), and dissolved oxygen; scaling: Co = 0.5 mmole r l ,
Cs =2mmolel- l , CE =0.5mmoler 1, Cx = 100gl- 1; b total oxygen transfer rate OTR
[kg m- 3 h- I] (0)

·5
1 0
1 0

10

• I)

,E

Fig. 9.10. Simulation results for spatial and time dependence of the specific ethanol production
rate rEI (left side) and uptake rate rE2 (right side) for the fed-hatch cultivation in Fig. 9.9 [52];
top: riser; middle: downcomer; bottom: head; The normalized spatial coordinate in all three
sections is z=O for bottom and z= 1 for top, the scaling unit for rEI and rE2 is 5 xl 0- 7 mmo-
le(gs(
9.3 Growth in Airlift Tower-Loop Reactors 295

production. To avoid loss of productivity in an inhomogeneous environment, the

flow rate must not be decreased so much that in every location of the reactor the
metabolism is purely oxidative. Therefore, the cells will pass through various regions
where by different substrate concentrations the substrate uptake rate is either above
or below the critical value. This value may also vary due to different levels of dis-
solved oxygen. Abel et al. [53] concluded from experiments with enforced periodic
Orvariation that as long as the residence time in the oxygen-limited region is below
1 min and the volume of the downcomer less than one-third of the riser, the yield
should not be negatively influenced. The yeast can switch very fast between oxidative
growth with ethanol uptake and oxidoreductive growth with ethanol production, as
long as ethanol is used in parallel to sugar and only catabolized in the TCC for en-
ergy production. On the other hand, comparative investigations by George et al. [51]
on a 215-m3 bubble column production reactor and a scaled-down reactor point to a
more critical role of inhomogeneous substrate distribution by localized feeding.
There, with a recirculation interval of 60 s, the yield decreased by about 6%, while
at the same time the gassing power of the yeast product increased. But in ATL reac-
tors with a higher recirculation flow the inhomogeneities are less pronounced.
For an investigation of the influence of concentration profiles on the yeast growth,
the model of the previous section was extended by a distributed reactor model to
describe the spatial dependencies of process variables [52]. The reactor model, as
presented in Chap. 1, Eqs. (1.63)-(1.70), has to describe the time and space-depen-
dent concentrations in the gas and liquid phase as a function of initial conditions,
manipulating variables, and biological reactions of the yeast cells. Beside reaction
and mass transfer, convection and dispersion are considered for gas and liquid
phase. In airlift reactors, the fluid-dynamics and flow characteristics are controlled
by the aeration rate which also influences the mass transfer and mixing properties.
The correlations given in Table 9.4 were used for the dispersion coefficients, trans-
port parameters, and gas holdup. The parameters were estimated from experimental
data based on measurements by Liibbert et al. [54, 55]. Since it is ensured in normal
fed-batch operation that sufficient amounts of sugar are always available, gluconeo-
genesis is avoided and the respiratory system is fully induced. Therefore, the dy-
namic regulation models as part of the metabolic model can be omitted for this
study. This also avoids the problem with structured segregated models as was dis-
cussed in Sect. 2.3 in Chap. 2. .
After verification of the model with experimental data, simulation studies for the
pilot plant were carried out to investigate the influence of non-ideal mixing and
concentration gradients. Some results are shown in Fig. 9.9 for averaged concentra-
tions of the liquid phase, and in Fig. 9.10 for specific ethanol uptake and production
rates. The spatial concentration profiles (not shown) are relatively flat and the cell
mass could be considered as a lumped variable. Dissolved oxygen is limiting during
the entire cultivation. Therefore, the cell mass increases almost linearly. To meet the
increased oxygen need of the culture, the aeration rate was increased stepwise sev-
eral times as can be seen from the oxygen transfer rate. In the average, there is etha-
nol accumulation in the first half of the cultivation which is later on reconsumed.
As expected, the yeast metabolism switches from fermentative with ethanol pro-
duction to oxidative with ethanol uptake during each circulation. Under the given
operating conditions in the head space of the reactor, only ethanol production is
found. After an initial phase of about 3 h with overfeed of sugar, there are both etha-
nol production and uptake in downcomer and riser (Fig. 9.10); ethanol is produced
296 9 Baker's Yeast Production

in the upper region of the two sections, where the oxygen transfer is lowest. At these
locations there is also a slightly higher sugar concentration due to distributed feed-
ing over the height of the reactor. In the bottom regions of the system ethanol uptake
can be found due to lower sugar and higher dissolved oxygen concentration. Produc-
tivity and yield are not greatly influenced as expected from [53], since ethanol serves
as a buffer that equalizes the differences in growth conditions for the growth-deter-
mining intracellular energy production. When sufficient sugar is available and etha-
nol flux is only directed to the TCC, there is practically no loss in energy efficiency.
The analysis reveals that for the given system substrate, feeding in the riser is not
optimal because then the substrate concentration is highest just in those parts of the
reactor where the oxygen supply is lowest. This limits the productivity because the
substrate flow rate has to be kept low to avoid too much ethanol production and loss
of yield. It can be concluded from the simulation studies that in this special reactor a
distributed substrate feeding in the downcomer would be favorable in order to flat-
ten the profiles for oxygen in the entire reactor. The reason is the very high recircu-
lation flow that transports enough gas bubbles and oxygen into the downcomer
down to its bottom, where oxygen transfer is increased by pressure effects. This at
first unexpected effect was also found by Pollard et al. [56].

9.4
Population Balance Models for the Asymmetric Cell Cycle of Yeast

In this section, age distribution models for growing yeast populations are presented,
aimed at two aspects: first, to provide a quality parameter for the yeast production
that can be used for optimal dynamic control of fed-batch processes, and second,
analysis of stable synchronous oscillations in continuous processes. There were dif-
ferent attempts to develop such models for balanced and non-balanced growth,
mostly on the basis of the general population models presented in Chap. 2,
Sect. 2.4.6, Eq. (2.123), or the age distribution model, Eq. (2.126). The variables used
to characterize the population depend on the desired application of the model and
the kind of analysis that should be done. A model for studying biochemical pro-
cesses during the cell cycle will naturally use other quantities than a model that is
aimed at a description of some population parameters during industrial cultivations.
Of course one would always like to establish a model that is as close as possible to
the biochemical mechanisms of cell growth and proliferation. But this is presently
far out of reach. Therefore, usually some representative variables without direct cor-
respondence in the metabolism are used, such as total mass, age, or volume of a cell,
or mass of a representative intrinsic constituent, e.g., protein content or storage ma-
terial. The next question for population models is whether the length of the cell cycle
phases is assumed as stochastic with a continuous variation around an average value
- so individual cells may behave differently - or assumed as deterministic, where all
cells finish their phase at the same discrete value of the characteristic variable. The
latter type of model has a higher degree of simplification, but has the advantage of a
facilitated solution, either numerically or even analytically for some special cases.
This property makes it much easier to check the model against a great number of
experimental data and to apply it for data interpretation, e.g., for balanced asynchro-
nous growth of yeast [3]. As any model, unstructured or structured, a population
model also adopts a very simplified view of reality, and one must always be aware of
9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 297

it [62]; but one should not succumb to the illusion that less simplified models must
give a more exact representation. They generally have a higher degree of freedom
and include more parameters and mathematical functions, so the experimental ver-
ification becomes much more difficult - often due to the limited availability of ex-
perimental data, or difficulties in the model solution - or even impossible when the
model can be arbitrarily fitted to almost any data set. So the general guideline is also
imperative for population models: a model is never reality, but a model should be as
simple as possible as long as it is in agreement with experimental data. Conceptual
models that are compared to no or limited experimental data are of less use for
practical application [63-66].

9.4.1
Age Distribution Model of Yeast for Batch and Fed-Batch Processes

Many microorganisms such as bacteria propagate by symmetrical division. The par-

ent cell "dies" as it is reborn in two identical daughter cells. This is reflected in the
model, Eqs. (2.123) and (2.128) in Chap. 2, by the factor 2 that is also called the birth
rate. The propagation of yeast by budding follows an asymmetric mechanism as
shown schematically in Fig. 9.2: the bud grows out of the parent cell. After bud se-
paration, the remaining bud scar is closed and the parent cell keeps growing and can
give birth to further buds. A new-born bud behaves differently from its parent cell
since it is usually much smaller. One of its characteristics is a longer generation time
compared to parent cells. Only after its first budding does the former daughter cell,
now in the first generation of parents, behave quite similar to the elder generations
of parent cells. Therefore, a population model for yeast has at least to consider two
clearly distinguishable sub-populations of parent and daughter cells. A short review
of population models for baker's yeast is given in [32,33].
The early investigations of the population dynamics [3-5], where the cycling
phase equations, Eq. (9.1), were established and verified, can be represented by an
age distribution model, as given by Eq. (2.126) in Chap. 2, with discrete deterministic
division ages and a loss function g that equals the dilution rate D of the reactor. But
the clearly distinguishable sub-populations of daughter and parent cells have to be
described by at least two coupled population balances. The respective cells number
densities in dependence of age 't and time t, !D(t, 't) and !p(t, 't), are used as charac-
terizing variables of the population. The M'Kendrick-von Foerster equations accord-
ing to the simplified scheme of the asymmetric cell cycle as shown in Fig. 9.2 then
become

afp(t,T) afp(t,T) __ .f( )

at + aT - D JP t, T (9.19)

a!D(t, T) a!D(t, T) _ _ . f ( )
at + aT - D JD t, T (9.20)

The processes of bud separation and formation of new unbudded daughter and
parent cells, each with a birth rate equal to one, is then introduced into the model
by the boundary conditions

(9.21 )
298 9 Baker's Yeast Production

For an the initial condition at t=O with exponential age distribution

fD(O, T) = fp(O, T) = /oe- w (9.22)
growth is balanced and the solutions of the population balances for the number
densities become
(9.23)
where D=O for batch cultivation. The number of cell in the cell cycle intervals and in
total are

J J
TD Tp

ND(t) = fD(t, T)dT Np(t) = fp(t, T)dT

o o
(9.24)

J J
TD+TB Tp+TB

NB(t) = fD(t, T)dT + fp(t, T)dT

TD

Ntot(t) = ND(t) + Np(t) + NB(t) (9.25)

and the fraction of budding cells is then

FBC(t) = NB(t) (9.26)

Ntot(t)
Hartwell and Unger [3] found from the above solutions the following condition -
which can be called stationary condition - for the length of the cell cycle intervals of
an exponential asynchronously growing yeast population,
(9.27)
where the specific growth rate and average doubling time are connected by
T = In(2)
(9.28)
JL

r;;
d(l)
\T
•
d(/) d(no)
II II II IJ,
To
unbudded daughter phase

p(j) p(np) f
II II

un budded parent phase

Fig.9.11. Structure of the discrete cell cycling model

9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 299

Tyson et al. [106] derived from the previous model the relation between the length
of the budding phase and the fraction of budding cells as

(9.29)

that is in good agreement with experimental data. Adams et al. [61] presented a
population model for asymmetric division with discrete age classes. The transitions
from one age class to the next or to the next cycle are determined by probabilities.
For steady state the model is equivalent to that of [3]. The simulation showed
strongly damped oscillations during transients of growth. The dynamics of the cell
cycle are of interest for fed-batch cultivations when it is desired to predict the pro-
duct quality which depends on the fermentative activity and storage properties. Ta-
kamatsu et al. [71] have found that the fermentative activity is correlated to the
fraction of budding cells (FBC) in the yeast population. Since the storage properties
are also related to FBC [50], this single parameter can be used in a process model to
describe several quality aspects. A low FBC at the end of the final fermentation stage
gives a good product quality. In an industrial-scale baker's yeast production process,
the cell growth rate generally changes with time. Moreover, the whole fermentation
period is only about 10 h. A true steady exponential growth state is not obtained
under these operating conditions.
A time-discrete cell cycling model (CCM) - combined with a growth model for
Saccharomyces cerevisiae - to describe this asymmetric budding process under dy-
namic conditions was developed and verified with experimental data by Yuan et al.
[47, 49, 50]. The growth model, given in Sect. 9.2, represents the continuous reac-
tions of growth. The oxidative pathways and the gluconeogenesis are subject to
long-term regulation by enzyme induction and repression. The respective dynamic
regulation models were omitted here because in the final stage of fed-batch produc-
tion the respiratory system is fully adapted and gluconeogenesis is avoided. The
age distribution model covers the discontinuous processes of the asymmetric cell
multiplication process and the age distribution of the population. As was discussed
in Sect. 9.1, the cell metabolism at large does not vary greatly during the cycling
process and the increase of cell mass is the rate limiting step of cell multiplication,
i.e., the cell division cycle is controlled to a great extent by the specific growth rate.
Since in the range of interest the mean cell volume and mean cell density vary only
slightly but in opposite direction, the cell number doubling time was assumed to be
equal to the cell mass doubling time, and Eq. (9.28) was applied under dynamic
conditions to connect the growth model with the age distribution model. This
makes it possible to simulate the control of the cell cycling process by manipulat-
ing the common controlling variables, e.g., substrate feeding rate. For a limited
degree of synchrony, a feed back from the cell cycling model to the metabolic
model need not be considered. For simplification, parent cells and daughter cells
were assumed to have the same metabolic activity as described by a unique specific
growth rate.
To facilitate simulation, the continuous balances of the population model,
Eqs. (9.19) and (9.20), are approximated by discrete shift-registers with nD> n~ and
nB storage elements, as shown in Fig.9.11. The cycling phases D, P, and B (see
300 9 Baker's Yeast Production

JIJ
/[hj

Fig.9.12a,b. Experimental data (symbols) and simulation results (lines) by the process model
for a combined batch and fed-batch cultivation of baker's yeast Saccharomyces cerevisiae
H1022 on glucose: a sugar feed rate F [lh- 1 ] and concentrations in kgm- 3 of cell mass Cx
(0),. substrate Cs \*) and eth~nol CE (~);.b specific growth rate J.L [h-1 ], cell number concen-
tratIon X N [10-5 1- ] and fractIOn of buddmg cells FEC [%] ('\7). Repnnted from [49]

~ ~:.-----------------------~

10 15 2.1
I [hI

Fig.9.13a,b. Experimental data (symbols) and simulation results (lines) for a pulsed-feeding
fed-batch cultivation of baker's yeast Saccharomyces cerevisiae H1022 on molasses: a sugar feed
rate F [lh- 1 ] and concentrations in kgm- 3 of cell mass Cx (0), substrate Cs (*), and ethanol
CE (~); b specific growth rate J.L [h- 1], cell number concentration X N [10-5 1- 1] (*) and fraction
of budding cells FEC [%] ('\7). Reprinted from [49]
9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 301

Table 9.S. Identified parameters in the growth model for Saccharomyces cerevisiae H1022

Parameter Fig. 9.12 Fig. 9.13 Units

'Smax 0.014 0.019 mole (ghr 1

K Mal 0.003 0.007 mole (gh)-l
K Ma2 0.012 0.005 mole (ghr 1
K Ma3 0.006 0.0001
KEG 0.05 0.05
K Ad 0.02 0.031
PlO 1.2-3.9 1.5-2.4
YATP 10.3 10.5 g mole- 1
mATP 0.001 0.003 mole (ghr 1
rAcmax 0.019 0.003 mole (ghr 1
c/ 351 325 kg m- 3

Table 9.6. Identified parameters in the cycling phase equations for Saccharomyces cerevisiae
H1022

Experi- KDJ KD2 Kp1 KP2 KBJ KB2 Substrate Tempera-

ment ture

Fig.9.12 1.84 -1.09 h 0.46 -0.1 h 0.29 l.39h glucose 303K

ethanol
3.32 -0.26h 0.36 -0.1 h 0.35 0.08h
Fig. 9.13 1.88 -l.l1 h 0.59 -0.1 h 0.26 l.2h molasses 307K

Fig. 9.2) are divided into small cycling age intervals ~'t. The number of these inter-
vals in each cycling phase are given by
TD Tp TB
nD = !:::.T np = !:::.T nB = !:::.T (9.30)

where the lengths of the cell cycle phases for single daughter, single parent and
budded cells, T[» T~ and TB, are determined by Eq. (9.1). The continuous cell num-
ber densities are replaced by discrete ones in the following manner:
fD(1 ... TD) --+ d(1 ... nD)
fp(1 ... Tp) --+ p(I ... np) (9.31)
fD(TD + 1 ... TD + TB) + fp(Tp + 1. .. Tp + TB) --+ b(I ... nB)
In each simulation time step with a length ~'t, the following substitutions are per-
formed
d(i - 1) =? d(i) b(nB) =? d(l)
p(i - 1) =? p(i) b(nB) =? p(l) (9.32)
b(i - 1) =? b(i) d(nD) +p(np) =? b(l)
For varying T, the cells are redistributed over more or less cycling intervals by
keeping the old distribution pattern and the relative positions of the cells in the
cycling phases. When T decreases, only cells in D are redistributed by an exponential
302 9 Baker's Yeast Production

non-uniform pattern, so as to adapt to the dynamic dividing behavior during the

cycling process [49]. The averaged population variables can be calculated by simple
summations as for the total number of cells, the cell number concentration and the
fraction of budding cells,
nD np nB

Ntot = L d(i)+ Lp(j)+ L b(k) (9.33)

i=! j=! k=!

(9.34)

nB b(k)
FBC=L- (9.35)
k=! N tot

The model was tested with a number of experiments. The results of two fed-batch
experiments are shown in Figs. 9.12 and 9.13. The identified model parameters are
given in Tables 9.5 and 9.6. Although the latter parameter values show some varia-
tion, the calculated lengths of the cycling intervals are close to the literature values
for balanced growth. In the first experiment, the time series of the feeding rate was
eventually equivalent to an industrial incremental feeding pattern. At the beginning,
the model simulation for FBC differs somewhat from the measurement. This is due
to an initial lag-phase that was not taken into account in the model: there the simu-
lated specific growth rate is too high, resulting in the sharp peak in the simulated
FBC, as illustrated in Fig. 9.12. There were actually three different growth phases in
the experiment according to the available substrates: oxidoreductive on glucose only,
oxidative on ethanol only, and oxidative on glucose during the fed-batch. In Fig. 9.12.
the cycling model is compared with experimental data for FBe. Besides the tendency
of increased FBC with higher growth rates, some oscillations in FBC can be ob-
served. The final value of FBC is about 20%. In the ethanol growth phase, the dura-
tion of the daughter cell phase was much longer than on sugar at comparable f.L. This
is probably caused by the changes of cell morphology due to poor nutrition [67].
Therefore, the parameters of the cycling phase equations were identified again for
ethanol growth as given in Table 9.6.
In the second experiment shown in Fig. 9.13, a pulsed feeding pattern was used,
based on the CO 2 content in the exhaust gas. If CO 2 was below a preset lower thresh-
old, the flow was turned on, whereas when it had reached a preset upper limit, the
pump was shut down. The feeding pulse sequence is shown in Fig. 9.13a. The width
of the pulses was about 1-1.5 min and the pulse period 0.6-0.9 h. At the beginning,
cells grow on ethanol that remains from a previous batch culture and FBC decreases.
The first feeding pulse at about t=3 h leads to a short period of oxidoreductive
growth with ethanol production and high growth rate so that FBC increases again.
During the remaining part of the cultivation the average growth rate decreases mod-
erately, accompanied by a similar trend on FBC with slight oscillations.
In the experiments, the simulated cell number concentration, XN , was in good
agreement with the measurements, though the specific growth rate varied over a
wide range. Also the FBC simulated by the CCM follows the measurements well.
From these results, one may conclude that the CCM with its simple structure is able
to describe the oscillatory transients by partial synchronization of the cell popula-
tion of baker's yeast under dynamic conditions, and can be applied in process opti-
9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 303

mization, for example minimization of FBC. For minimization of FBC at the harvest-
ing time in fed-batch culture, the feeding strategies should be in accordance with
such an intrinsic oscillation and reinforce it in a way as to make the FBC approach
its minimum at a desired harvesting time. This will be further outlined in Sect. 9.5

9.4.2
Age Distribution Model for Data Analysis of Stable Synchronous Oscillations
in a (hemostat

Population synchrony is an important factor for autonomous and forced oscillations

in continuous culture of yeast. For a rather mechanistic modeling of self-synchroniza-
tion and sustained oscillations, the population balances must be extended by a struc-
tured growth model that introduces a non-linear feed-back mechanism for the stabi-
lization of the oscillation due to differences in the metabolism of budded and un-
budded cells. In the model of Strassle [35-37) a bottleneck in the respiratory pathway
and description of the metabolism of storage carbohydrates was included. The range
of dilution rates and the frequency of sustained oscillations could roughly be repro-
duced. Cazzador et al. [38) established a model for a size control mechanism that is
modulated by the substrate concentration. The model was in qualitative agreement
with observations of frequency and amplitude of oscillations at different dilution
rates. Unfortunately, there are no analytical solutions of the balance equations of such
models, and one must turn to numerical methods, which require an enormous expen-
diture in computer time. This limits the simulation studies to only a few points of the
parameter space and makes it difficult to draw some general conclusions. For a broad
evaluation of experimental data these models are too complicated.
Sustained synchronous oscillations of baker's yeast were also analyzed by means
of an age distribution model without consideration of cell metabolism [32,33). Many
characteristics of the oscillations - with the exception of the actual shape of the age
distribution - are already determined to a great extent by the asymmetric cell cy-
cling behavior of yeast in conjunction with the mathematical properties of the solu-
tion of the population balance equations. The assumption of discrete division ages
without variation in the lengths of the cell cycle intervals of individual cells of the
population seems to be a very natural approximation, because such variations
should be reduced by the feed-back, which is responsible for self-stabilization of
the oscillation. The accompanying oscillations on the concentrations of substrates
actively force the cells into a narrow distribution and reduce the dispersion in cy-

Table 9.7. Oscillation periods for a few mode numbers

Mode TlJT- 1 TIl (T= 2.5 h) alJ Remark

I J
1 2.5 0.5 Symmetric division
2 0.694 1.736 0.6180 Golden mean
3 0.551 1.379 0.6823
2 2 0.5 1.25 0.7071 Symmetric division
2 3 0.406 1.014 0.7549
2 4 0.347 0.868 0.7861
304 9 Baker's Yeast Production

5
..
4
."
.....•
- - - Mode 11 =T
3
T,J[h] ........ Mode 14
- -Mode 22 =T/2
2 ·········Mode 23
·············Mode 24
- - - Mode 33
.•.•.•.. Mode 34

o
2 3 4 5 6 7 8 9 10
T [h]

Fig. 9.14. Calculated oscillation periods (lines) of self-synchronized growth over doubling time
in comparison to experimental data (symbols) from the literature: .&. [35), • [39), 0 [40),
• [43), ~ [44)

cling events. The model considers incomplete synchronization and an age distribu-
tion of cells within the cell cycle, but only due to a phase shift in the growth of
individual cells. The assumption of discrete division age is actually not a great re-
striction for analyzing stable synchronous oscillations, because under such condi-
tions also in a model with division age distribution, the statistical fluctuation must
be exactly leveled out during the oscillations with an average value identical to the
discrete division age. For a distribution being conserved between subsequent oscilla-
tion periods it makes no difference whether all cells exactly keep their order during a
cycle or some of them exchange their positions by stochastic processes. Further-
more, when the cell metabolism during one oscillation period is determined by the
average doubling time, analytical solutions for the model can be obtained. That pre-
condition is supported by the results on yeast physiology reported in Sect.9.1 and
also by the work of Munch et al. [39] who showed that very slight disturbances are
sufficient to establish or alter the oscillations.
For extending the above model to stable synchronous oscillations the initial con-
dition, Eq. (9.22), can be rewritten as

(9.36)

and the solutions of the population balances become for chemostat processes, with
D=jL,

(9.37)

Here, an additional shape function, 1,

of the age distribution appears, which is
responsible for the oscillatory behavior of the solution,

(9.38)
9.4 Population Balance Models for the Asymmetric Cell Cycle of Yeast 305

Cell number (relative units) a

N..
0.5
ND

Np

Ns
0
100

G,

50

0
0 2 3 4 5
I[h]

Fig.9.1Sa,b. Simulation (lines) of synchronous growth in comparison to experimental data

(symbols) from Strassle [35]: Mode 12, T=4.4h, T B =1.6 h, Ts=0.5 h: arelative numbers of total
unbudded daughter (ND ) and unbudded parent cells
(N,o')' (Np ); b fraction of cells in the bud-
ding (_), S- (.... ), and GI-phase (+)

Equation (9.37) describes a sustained oscillation with constant amplitude over

time, but exponentially decreasing amplitude over cell age due to the dilution. It also
matches the stationary condition at Eq. (9.27). The possible oscillation periods are
TB+Tp TB+ TD
Tlf= I J for 1=1 , 2, 3 .. . , J = I , 2,3 .. . ,1 (9.39)

During sustained oscillations in population variables the lengths of the parent and
daughter cycles must be multiples of the oscillation period as defined by the mode
numbers I and]. By introducing the decrement of the oscillation,
(9.40)
Eq. (9.27) changes to
(/+aJ =1 (9.41)

which can be solved for a when the mode numbers are given, yielding a value aI/'
The oscillation period of a certain mode then becomes
In( a IJ) In( a lf )
Tlf = - - f l , - = -T In(2) (9.42)

Some values of aIJ and TlJ for low mode numbers are given in Table 9.7. The
solution for sustained oscillations in chemos tat processes, Eqs. (9.37)-(9.42), has
some very interesting properties due to the asymmetric population dynamics.
Although the main determining factor for the frequency of the oscillation is the
average doubling time T, the oscillation frequency is not unique for a certain T,
but may change with the mode of oscillation; fixed discrete but multiple oscilla-
306 9 Baker's Yeast Production

tion frequencies are possible. The actual mode present in the culture depends on
initial conditions, cultivation conditions, and growth kinetics. Therefore, it can be
possible that under identical cultivation conditions different oscillation frequen-
cies are observed as shown in Fig.9.4. Oscillation periods from the experimental
work of several authors and calculated from the model for different modes of
oscillation are presented in Fig. 9.14 over doubling time. The predicted oscillation
frequencies agree well with experimental data. The observed variability of the fre-
quency at a fixed doubling time can be explained to a great extent by the multi-
ple modes of oscillations. The prevailing number of measurements close to mode
12 fall into the region T<7 h, and for higher modes into the region T<4.5 h. This
coincides with situations where the daughter cycle is shortened by the oscillation
and the parent cycle is prolonged compared to asynchronous growth. For T<3.2 h,
the asynchronous growth follows a nearly symmetric division pattern, which is
very close to the oscillation modes 11, 22, and 33. It can be expected that under
such conditions one of these modes is preferred. In a very wide region for med-
ium doubling times, the modes 12, 24, and 36 - where the length of the daughter
cycle is twice that of the parent cycle - come very close to the asynchronous case.
Hence, these modes may be called the natural modes of oscillation of a synchro-
nous growing yeast population.
Further evaluation of experimental data by the model revealed that sustained os-
cillations can only exist in a region where the length of the parent cycle is increased,
and the length of the daughter cycle is decreased, compared to asynchronous growth
[33]. The preferred region of oscillations of a certain mode seems to be determined
by the length of the budding phase which also controls the oscillation amplitude.
The predicted tendency of the oscillation amplitude in dependency on the oscillation
frequency was also in good agreement with experimental data for the CO 2 evolution.
The theoretical analysis showed that two different types of synchronous oscillations
exist under excess of oxygen and oxygen limitation. The latter one, with oscillation
periods greater than the doubling time, could not be explained by the model.
While the oscillation period of a certain mode is exactly known from the above
theory, the direct determination of the actual mode of oscillation in the experiments
still remains difficult since exact age distribution data is mostly not available. One
method is the careful analysis and simulation of the time course of the oscillation on
average population parameters, such as the total number of cells in a certain phase of
the cell cycle. In Fig. 9.15 a sinusoidal shape of the age distribution

f(t, T) = 1+ sin (27r

TI]
(t - T)) (9.43)

was used to compare the model to experimental data of Strassle [35]. The culture is
growing at T=4.4 h under the natural mode 12. The budding phase was estimated at
TB =1.6h and the length of the S-phase for the DNA-synthesis to Ts=O.5h. These
estimates are in good agreement with the data of Munch [41] and Beuse et al. [42].
The number of cells in the P- and B-phases does not return to zero in this experi-
ment, although the age distribution does. The simulated phase difference between
the oscillating cell number in the S-phase and the other phases of the cell cycle is
consistent with the assumption that the S-phase is located at the end of the single cell
phase just before the bud appears.
9.5 Considerations for Process Optimization 307

The important property of the model is that the general characteristics of the os-
cillation are already fIxed by the mode numbers and the doubling time, while culti-
vation conditions, growth kinetics, and cell metabolism mainly influence the shape
of the age distribution. For all other parameters of the oscillating population vari-
ables, such as relative amplitude, average value, or phase shift, there is no remaining
degree of freedom in the model; they are all uniquely determined by the mode num-
bers. Bya different mode the model calculations as given in Fig. 9.15 would clearly
disagree with the experimental data. The quite good agreement with the experimen-
tal data for a wide range of experimental conditions supports the simplifying as-
sumption of the model. Nevertheless, synchronous oscillations may be more compli-
cated in nature than can be reflected by the simple pure age distribution model.
Beuse et al. [42] found additional modes of oscillations that can be explained by a
further daughter cell class and extended the model accordingly. There were also
hints that the length of the budding phase may be different for parent and daughter
cells. Hjortso and Nielsen [70] extended a population model by a substrate balance
and could show that in such a model multiple modes of oscillations could also exist.
But for a full understanding of all the details there is still a long way to go.

9.S
Considerations for Process Optimization
In this section the dynamic process model is used to investigate optimum control
strategies in the baker's yeast process and to determine optimal operating parameters.

10
.-------------------~~

10
t[hl

Fig.9.16a,b. Comparison of simulations (lines) and experimental data (symbols) for a fed-
batch optimum quality control strategy: a sugar feed rate F [lh-'] and concentrations in
kg m- 3 of cell mass Cx (0), substrate Cs\*), and ethanol CE (~); b specific growth rate /1
[h- 1], cell number concentration X N [10- rl] (*) and fraction of budding cells FBC [%]
(\7). Reprinted from [49]
308 9 Baker's Yeast Production

Table 9.8.Identified parameters in the model for Saccharomyces cerevisiae H1022 used in the
simulation of Fig. 9.16

Parameter Value Units

rSmax 0.018 mole (gh)-l

K Mal 0.003 mole (ghr 1
K Ma2 0.01 mole (ghr 1
K Ma3 0.034
KEG 0.05
KAd 0.024
PIO 1.8-2.1
YATP 11 g mole- 1
mATP 0.001 mole (gh)-l
rAcmax 0.018 mole (gh)-l
KDJ 1.23
KD2 -2.04 h
K p1 0.435
KP2 -0.43 h
KBI 0.023
KB2 1.49 h

The baker's yeast production process is mostly optimized with respect to yield be-
cause the substrate costs account for the greatest cost factor of production. By fed-
batch operation, the substrate concentration and uptake rate are kept relatively low
at the critical level just before oxygen limitation or Crabtree-effect. This strategy
avoids ethanol production with loss of yield, but also limits productivity. An optimi-
zation should also take into account, in addition to yield, productivity and economic
profit, as well as the quality of the yeast product. The latter aspect is often neglected in
model-based optimization because it is difficult to relate quality to the usual model
variables. In practice the proper control of the relevant manipulating variables with
the aim of getting the end-product with desired quality, Le., good storage stability and
leavening power for baking, is still more an art than science [72,73].

9.S.1
Optimization of Product Quality

For baker's yeast, several aspects of product quality are connected to the cell cycling
process and the fraction of budding cells, as was mentioned above. The FBC at harvest
is the result of the entire dynamic development of the cell cycling process during the
cultivation, which is difficult to optimize without a mathematical model. Therefore, a
combined growth and age-distribution model should be applied for such an extended
optimization. For simplicity, only the final production tank is considered here in de-
tail, because it has a major influence on yield, productivity, and product quality, and it
uses a great percentage of the resources. A complete model for the entire series of
cultivations in the multi-stage production process could be built by multiplication of
the elementary model. Quality parameters for yeast that can be influenced by cultiva-
tion conditions are the fermentative activity in terms of the ability to produce CO 2 for
leavening the dough, and the storage stability. For active dry yeast the loss of activity
9.5 Considerations for Process Optimization 309

during the dehydration-rehydration cycle is also of importance. The storage stability

of compressed baker's yeast depends on several factors besides storage temperature
[74, 75]: trehalose content, protein content, and maturity. It has been shown that sto-
rage carbohydrates such as trehalose and the average protein content [68] of yeast cells
are strongly influenced by the fraction of budding cells and the age distribution of the
population. Dairaku et al. [82] and Takamatsu et al. [71] pointed out that a minimized
FBC at harvesting time of the cultivation corresponded to a high leavening activity.
Some investigations revealed a dual effect of FBC on the storage properties [50]: if the
yeast is used within a period of a few days, samples with lower FBC values have less
leavening activity than those with higher FBC values. However, if the compressed
yeast cells are stored for more than ten days before being used, the samples with high-
er FBC values lose their leavening activity very rapidly. Therefore, minimization of the
FBC is favorable for enhancing the storage stability of compressed baker's yeast. Con-
trol of product quality may also be of interest for the production of active dry yeast
(ADY). It was found that the leavening activity of ADY samples decreases, and the loss
ofleavening activity during the dehydration-rehydration cycle increases strongly with
FBC. This suggests that minimization of the final FBC is also important during the
ADY -production.
A symmetric division model for optimization of the IL-profile was applied in [71,
82] for fed-batch cultivation. But there the cycling pattern of the budding yeast will
be far from symmetric division [4,5]. The optimal control of the fraction of budding
cells, or equivalently, the maturity of baker' yeast, by dynamic optimization of the
substrate feeding rate for fed-batch cultivation was therefore studied by means of the
asymmetric CCM [47], as described in Sect. 9.4.1. The growth model takes the real
control variables, e.g., substrate feeding rate, as the input and delivers an accurate
estimation of IL to the cell-cycling model, where it is translated into the proper
lengths of the cycling phases in the asymmetric age distribution model. An optimal
feeding strategy was designed to minimize the FBC at harvesting time, while keeping
yield and productivity at a relatively high level [49, 50]. A stepwise flow rate, F(t i ), as
shown in Fig. 9.16, was determined by minimizing the following objective function:

J( Tp ) = [Cs2 + CE2 + FRC 2 + (45 - Cx )2] t=T P(ti),i=I,2, .. 9

: min (9.44)
F Fmin <P(ti)<Pmax

where Tp is the cultivation period. The flow rate is restricted by the installed equip-
ment in the range Fmin to Fmax. The sugar concentration in the feed was 250 g I-I. The
model parameters used for the optimization were obtained from the experiment gi-
ven in Fig. 9.12 including the ethanol growth phase. The optimal control policy was
applied in an open loop without feedback from process variables. Figure 9.16 shows
the simulation results of the experiment in comparison with measured data. The
model parameters were slightly readjusted after the experiment was carried out to
eliminate the influence of ethanol-growth. Table 9.8 gives the results of the parameter
identification.
The principal strategy for quality optimization by minimization of FBC as quality
index is as follows. At constant conditions, low FBC could be achieved only with long
doubling times and, therefore, low productivity. But for unsteady operation and step
changes in the growth rate the cell cycling process of the yeast population tends to
synchronization, and oscillations in FBC can be observed. This property can be used
for maximization of fermentative activity of the final product under high productiv-
ity by suitable dynamic control of the specific growth rate. The cells are then har-
310 9 Baker's Yeast Production

Table 9.9. Consideration of the upstream and downstream processing steps in the model

Type of cost Model Remark

Fixed operation, capital of invest-

ment and staff
Cultivation medium Ps=PMe+PSa+PSp Total
PMe=(yZCJ+Ks1(Vf- yZ)Cf)PMe Substrate
Psa=(l-Ksd(Vf - yZ)PSa Salts and other media com-
ponents
Preparation and storage,
including waste water treat-
ment
Aeration Aeration rate is assumed
constant
Downstream processing PD=PDC+PDD Total
PDc=KDlVf Centrifugation
PDD=VfCfpDD Drying
Proceeds for the sold PE=Vf(Cf-~)(l-KQFBC')pE Considering the quality of
product minus pitching the yeast product by the
yeast fraction of budding cells with
an empirical correlation

vested in the dynamic minimum of FBC at the end of the fermentation, being much
lower than the stationary value for the averaged specific growth rate. In this way the
productivity and quality can be decoupled to some extent to optimize both of them.
The minimum FBC value in the experiment is as low as 1.7%, which is very close to
complete synchronization. This is normally unattainable with the conventional feed-
ing strategies. Compared with conventional incremental feeding, the flow profile is
characterized by pulsing. During the periods of high flow rate, ethanol is formed
due to the Crabtree-effect at high growth rates. However, since ethanol is recon-
sumed completely afterwards - but in parallel to sugar - the yield is not lowered
significantly in comparison to conventional control strategies, where usually also
some ethanol is accumulated. In the period of ethanol uptake the growth rate is re-
duced. By this alternation in the specific growth rate, the pulsed feeding course en-
forces to some extent synchronization of the population growth to prepare a trough
of the FBC at the desired harvesting point, and final FBC is minimized. The steeply
decreased feeding during the last portion of the cultivation (about 3 h) supports the
down-swing of the FBC further and allows the uptake of ethanol. In contrast to some
industrial practice, the feeding is not stopped completely before harvest. This is ben-
eficial to the cells for maintaining a higher trehalose content and, hence, a better
storage stability. The studies in a number of experiments demonstrated that the
quality index of the storage stability of compressed baker's yeast can be controlled
and, indeed, be improved by the model-based open-loop feeding strategy in fed-
batch cultivations. The low feeding at the end does not harm the productivity of
the entire process, because there is an unusual high peak flow during the earlier
parts.
9.5 Considerations for Process Optimization 311

9.5.2
Economic Optimization

For more accurate optimization, the performance index should be uniquely deter-
mined by an economic input-output balance of the process. All economic indices
for a process such as quality, productivity, and efficiency are interconnected. An
optimization of one of them also influences the others in a complicated manner. It
is therefore preferable to take the total economic profit in dependence on the other
indices as criterion for multi-objective optimization. Such a procedure was applied
in [76] to an industrial scale airlift reactor with a volume of VR =60 m 3. The total
economic profit per unit of time, J, is calculated by

J -~ PE - Ps - PAe - PD - Pp
(9.45)
Tr
where
(9.46)
is the total duration for the completion of the process cycle, TF the fermentation
period, and Tp the time for preparation and downstream processing. The specific
costs for an industrial process were about 2% for inoculum, 80% for raw materials,
9% for energy, and 9% for fixed costs (capital, staff, and so on). For the economic
optimization, additional model equations are introduced to consider in a simplified
form the costs and time requirements of other processing steps besides cultivation
as included in Eq. (9.45): medium preparation, cleaning, inoculum propagation,
aeration, operation, and downstream processing. The additional model equations
are summarized in Table 9.9, where pMe and PSa are the specific prizes for 1 kg of
pure molasses and media components, pAe the specific cost for 1 m 3 of compressed
air, and PDD the specific costs for drying of 1 kg of yeast product. For consideration
of the product quality, its exact influence on the proceeds should be known; here, an
empirical factor was introduced, where KQ is a quality index of the selling prize, and
PE the specific maximum selling prize for 1 kg of best quality dry yeast.
Based on the extended process model the total economic profit of the baker's yeast
production was optimized,
F(t),h,~,~,C:
J ) max (9.47)
Subject to the optimization were, as constant operating parameters, the substrate
concentration in the feed and inoculum, the aeration rate, the mass of pitching yeast,
and the fermentation period, T F • For the molasses feed F(t) an optimum control
profile is determined with an hourly change of flow rate. For technical reasons, sev-
eral constraints for the optimization have to be taken into account: the maximum
working volume, the maximum and minimum aeration rate, the maximum molasses
flow rate, and the maximum substrate concentration in the feed that is given either
by the sugar content of the molasses or the maximum viscosity that can be handled.
For the given economic indices the point of optimal operation of a 60-m3 airlift
reactor was obtained for a cultivation period TF=lOh, aeration rate FG =90m- 3
min-I, and a substrate concentration of Cl=1l4kg m- 3.
The time course of flow rate and fraction of budding cells resembles in general the
pulsed feeding pattern of the pure quality optimization, as already shown in
312 9 Baker's Yeast Production

Fig. 9.16. But the optimization result is strongly influenced by the relative weights of
the economic indices in Eq. (9.44). For example, if the energy costs for aeration in-
crease, lower aeration rates would be preferred, which on the other hand reduces
oxygen supply to the reactor. This can be compensated for by slower feeding and
an extended cultivation period. Generally, if the yield is the most important aspect
of yeast production, the exponential-linear fed-profile is optimal because it avoids
oxygen limitation and any ethanol production. The exponential increase at the be-
ginning becomes more distinct with slightly higher weight on productivity. Then
there is moderate overfeed of sugar, accompanied by some ethanol accumulation.
In the final cultivation period, the flow rate is decreased to convert the remaining
ethanol. By a proper balance of sugar feeding and ethanol uptake, the decrease in
yield is very limited, because ethanol is only used for oxidative energy metabolism
via TCe. If much emphasis is put on high productivity the optimization results in a
different two-phase strategy: in the first phase there is a strong over-feeding with
ethanol production and sugar accumulation, the growth conditions corresponding
to batch operation. In the second phase the feed rate is lowered to small values so
that the ethanol can be used up completely. The loss of yield can be limited to about
16%. This strategy is very similar to the combined batch/fed-batch-scheme of De-
Loffre [76].
Where the product quality is considered, the optimization tends towards higher
productivity with slight loss of yield since the substrate pulses for growth synchro-
nization lead to over-feeding with subsequent ethanol uptake. But since this effect is
moderate, it was concluded that the quality can be improved with low additional
effort and small cut-off on the other economic indices.

Flow rate
Reactor

~-t/~
Substrate Measurements
reservoir
c,' CX,cE,C S1 e.g., OU R, CPR ,RQ,
Co, V L eEl Co

Feedforward-
+~ control
~

':' Estimator
I Estimation
Adaptation
Tuning I e.g. ex,
(dynamic/static) ~
I
W

+--;;;;rtrolled variable
- Feedback-
controller
_ _ Setpoint

Fig. 9.17. Schematic diagram of a control system for baker's yeast fed-batch production
9.6 Automatic Control of Fed-Batch Processes 313

9.6
Automatic Control of Fed-Batch Processes

9.6.1
General Remarks

In the production of baker's yeast, the costs for substrates contribute the biggest part
of the total production cost. Therefore, the main goal for process control is to obtain
the maximum yield at the highest possible productivity during the cultivation. This
means the growth rate has to be kept at its highest possible value that just avoids
fermentative growth by either Crabtree-effect or oxygen limitation in the reactor.
The simplest method from a control-point of view would be continuous cultivation
where the dilution rate can be fixed at the optimum point. But by the well-known
problems of operational procedures and of strain stability in continuous cultivations,
fed-batch operation is preferred for industrial production, although its control is
much more difficult.
In an ideal fed-batch process the specific growth rate would be fixed at the opti-
mum value resulting in exponential growth. For a number of reasons such a simple
strategy cannot be realized in practice. Usually the oxygen transfer capacity of the
reactor is limited so that the growth rate has to be reduced at high cell mass concen-
trations when oxygen becomes limiting, by keeping the substrate flow rate constant.
This results in the so-called exponential-linear feeding pattern. As discussed in the
previous section this basic strategy is modified to some extent when other criteria
than yield, e.g., productivity, product quality, or economic profit, are used for deter-
mination of the optimum control strategy. Further modifications have to consider
the true growth kinetics of the yeast including adaptation phenomena. Therefore,
the pattern of the specific growth rate is more complicated, but can in principle be
translated into a suitable dosage scheme for the substrate feed rate when all operat-
ing parameters are known. For many years, fed-batch processes were carried out
using such predetermined fixed feeding profiles that were at most manually adjusted
during the running process. As an advantage this simple open-loop control does not
require measurement equipment, but it is also difficult to guarantee reproducible
cultivations.
The open-loop strategy cannot ensure that the cultivation runs near the desired
optimum trajectory since it does not consider disturbances during the process, or
variations in operational parameters: e.g., power supply, temperature of the cooling
medium, composition of the substrate, initial biomass density, operational problems
with equipment. In addition, the growth kinetics of the yeast may vary in response
to such disturbances or variations in the precultures. When measuring information
from the running process is available, the actual flow rate can be corrected for these
deviations in the cultivation conditions by automatic feedback control, to keep the
process near to the desired optimum. A number of different concepts for automatic
control have been established by control engineers (see, e.g., [77, 78] for an over-
view), which all have their specific advantages and draw-backs when applied to bio-
technical processes, and baker's yeast production in particular. Several of these con-
cepts have also been adopted and experimentally tested for control of baker's yeast
fed-batch production.
314 9 Baker's Yeast Production

0.8
100

o
20
Stirrer speed

10 15
t [h]

Fig. 9.18. Experimental results for a fed-batch process controlled by the ethanol concentration
in the medium; ethanol concentration CE [kgm- 3 ], percentage of dissolved oxygen p02 [%],
stirrer speed [s-'], and flow rate F [1O-9 m-3 s-']. Reprinted from [87]

Another general problem for control of fed-batch processes is the increase of both
volume and cell mass. Under the desired constant growth conditions the latter re-
sults in an exponentially increasing disturbance. Since all reaction rates are propor-
tional to the cell mass due to the autocatalytic nature of growth, the process has -
from the control engineering point of view - variable gain and time constants. Con-
ventional controllers are designed and then tuned to have best performance at a
certain operating point. But this is steadily changing in the fed-batch process. There-
fore, the controller response may become slow or oscillatory in phases of the cultiva-
tion where the controller was not tuned for. In the worst case the control can become
unstable and then drives the process far away from the desired state. A stabilization
can be made only of the cost of controller performance, which then may become
unsatisfactory for the entire process. The methodology of adaptive control targets
this problem by changing the controller parameters according to the actual state
and parameters of the process, which then, in addition to the controlled variable,
must be measured, calculated from other variables, or estimated by means of a
mathematical model (self-tuning control). Summarizing the above remarks, a gener-
al control-scheme can be drawn as shown in Fig. 9.17, although in particular not all
blocks may be present and not all variables used. The control itself consists of the
feedforward part for reducing known or measurable disturbances, e.g., by the in-
creasing cell mass, and of the feedback controller (regulator) for fine regulation
around the rough value of the flow rate determined by the feedforward block. Usual-
ly the two blocks will not use all the same measured variables. The feedback con-
troller corrects the flow rate to make the controlled variable as close as possible to
the setpoint. This may be constant or varied according to a given profile that meets a
specific, possibly optimum strategy for the cultivation. There is a wide choice of
controller types, ranging from simple switching control or PID control to advanced
model based approaches. The third block for estimation of cell mass or other vari-
ables can be added to improve the disturbance compensation, adapt the employed
control law, or control non-measured variables. The estimation method can range
from simple stoichiometric calculations up to model-based dynamic observers for
the state variables of the process.
In a conventional control system, the controlled variable is measured, compared to
a given set point value, and the difference is looped back to the process via controller
9.6 Automatic Control of Fed-Batch Processes 315

and controlling device. In baker's yeast production there are several possibilities for
measuring a characteristic variable that is suitable for control. But to begin with
negative examples, control of substrate concentration in the medium is not practical,
not only due to lack of reliable and accurate sensors, but also because there is no
unique relation of substrate concentration or substrate flux - as input of the meta-
bolic system - to the critical point for onset of fermentative growth. As already men-
tioned, the latter also depends on the intracellular state of the respiratory system and
the oxygen supply. The same argument can be used for control schemes that track a
predefined profile of cell mass or specific growth rate, as investigated, e.g., by [B1-
B3]' or only one of the rates of gaseous metabolism, CPR and OUR. These quantities
per se give no information on the metabolic state of the cells, fermentative or oxida-
tive, and therefore cannot be used as single control variable if the yield or ethanol
concentration has to be kept in a narrow range in the presence of disturbances.
Nevertheless, profile control schemes can greatly reduce the variability among the
production runs. This can also be an important control objective, because it pro-
motes a constant product quality and facilitates scheduling of the entire plant in
fixed time intervals.

9.6.2
Examples for Applied Control Systems

Direct information on the metabolic state of growth can be obtained from ethanol
concentration, respiratory quotient, or equivalently, the difference between CPR and
OUR. In 1961 the principle of automatic control of the molasses feed rate by the
ethanol concentration in the gas phase was proposed by Rungeldier and Braun [79].
The ethanol content in the medium can be calculated from mass transfer equili-
brium. This control was realized only years later because of the lack of cheap and
reliable sensors. Nowadays, cheap in-situ membrane probes based on semiconduc-
tors are available [BO]. The ethanol concentration is usually kept constant below
0.1 vol.%. This means that during fed-batch with increasing volume there is always
moderate ethanol production, but strong over- or underfeeding can be avoided. Ac-
tually, some limited ethanol production during baker's yeast production is even de-
sirable to ensure high growth activity of the cells and reduce bacterial contamina-
tion. Furthermore, in reactors with non-ideal mixing, ethanol production could only
be avoided with great loss in productivity, which is not acceptable. But the loss of
yield can be limited by reducing the feed rate shortly before the end of the cultiva-
tion to allow for reuse of ethanol. When applied in the liquid phase, the response
time of the ethanol sensor is very short and control becomes much easier compared
to gas-phase measurements. The feasibility of the control concept was tested in sev-
eral papers with simple PI control up to non-linear adaptive control [B2, 92, 97, 102].
Axelson et al. [B7] showed that with this type of sensor during the entire cultivation
a stable control can be obtained by a simple PID rule. An example is given in
Fig.9.1B. The stability problems reported by other authors [B4, 92], mainly during
oxygen limited growth at high cell concentration, could be avoided. This relatively
robust and reliable type of control is also used in industry. Nevertheless, an adapta-
tion scheme covering the increasing cell mass can improve the controller perfor-
mance and disturbance rejection. Such a system, including a Luenberger observer
for the exponential disturbance by biomass growth, was presented in [103].
316 9 Baker's Yeast Production

An observer-based approach that allowed a very accurate estimation of cell mass

and specific growth rate was published by Pomerleau and Perrier [104, 105]. The
algorithm included as state variables the concentration of glucose, ethanol, cell mass,
and dissolved Oz and COz. Glucose feed, OUR, and CPR were used as measured
variables. Another means for model-based state and parameter estimation is the Kal-
man-filter. Although its theoretical concept - founded on a stochastic approach - is
quite smooth, it can be difficult in practice to ensure stability of the estimation. This
technique was also applied for baker's yeast adaptive state and parameter estimation,
including cell mass and growth rate [85, 86], but seemingly there is no on-line ap-
plication within a closed loop control system.
The ethanol control concept by set-point control was slightly extended by Schubert
et al. [107]. A predetermined optimum profile for the ethanol concentration was
followed during the fed-batch. An internal-model-control including an artificial
neural network was developed to improve the performance compared to PID-con-
trol. There were also reports in the literature about fuzzy state recognition [108]
and fuzzy control of baker's yeast production [110, 111]. Although this rule-based
approach offers a new view on the problem of control and seems to work reasonable
well on the laboratory scale, it is objected to by many control engineers because the
performance and stability is theoretically difficult to evaluate. It can be seen as an
advantage that backtracking of the inference rules for determination of the control-
ling variable's values from the measurements gives some explanation of the control
action.
Another common method for automatic control of the molasses feed rate uses the
respiratory quotient, RQ. Compared to control by ethanol concentration this method
not only requires higher effort for measuring devices, i.e., Oz- and COz-analyzers,
but also has inherent stability problems. As explained earlier, RQ equals one for oxi-
dative growth on sugars with a sugar uptake rate below the critical threshold for
Crabtree-effect. At fermentative growth above the critical value RQ is greater than
one, and during underfeeding of sugar with ethanol uptake, RQ is less than one.
When no ethanol is present in the medium, the respiratory quotient gives no infor-
mation about the degree of underfeeding to allow for a proper adjustment of the flow
rate by the controller. The usual way to reduce this problem is to choose a set point
in the range RQ=1.1-1.2, which also means moderate ethanol production. A very
accurate gas phase balancing is then necessary to prevent stability problems by a
bias in the RQ-ca1culation. An over-estimation of RQ can lead to complete ethanol
uptake and drive the process to relatively low growth rates in the purely oxidative
region, while under-estimation gives rise to pronounced ethanol accumulation and
loss of yield. Further problems in this control are the additional delay by the gas
analyzers and a strong cross-coupling of the RQ-ca1culation to pH in the cultivation
medium by dissolved carbon dioxide. An early investigation on the RQ-control was
by Aiba et al. [88]. A switching control was used that increased stepwise the sub-
strate flow rate for RQ< 1, and decreased it for RQ> 1. The control action occurred
in intervals of about 0.5 h. Based on a mathematical model the authors discussed
possible improvements of the control. For presetting the flow rate, two formulas
were derived, which are still used for feed-forward compensation of the growing cell
mass. The first uses the estimated cell mass,

F_k CxVL (9.48)

- YxsCf
9.6 Automatic Control of Fed-Batch Processes 317

and the second the measured oxygen uptake rate,

F = k OUR . VL (9.49)
C~
In the first equation, k should equal a desired specific growth rate, Ji-set> and in the
second the ratio of yield coefficients for oxygen and substrate. The parameter k is in
general not constant during the entire fed-batch, but can in principle be estimated by
an observer as mentioned above.
Control by RQ and other variables obtained from exhaust gas analysis was studied
further in a number of papers with different types of controllers including adaptive
ones [89-97]. The same basic idea was worked out further by O'Connor et al. [101].
They investigated the performance and failure conditions of different control
schemes for feed-forward control of the flow rate, including the calculated cell mass
and oxygen uptake rate. A parallel operating inferential feedback controller, as
shown in principle in Fig. 9.17, corrects the actual flow rate for disturbance and er-
rors in the cell mass. Controlled variable is the calculated ethanol reaction rate. A
satisfactory control of a lab-scale reactor was obtained by the proportional-integral
control law
Feed-forward PI-control
~, ,
F = Ji-setCX VL _ K V R _ K C '"' VLR E (9.50)
YC R PLE [X~C
D S X

The cell mass was calculated from either oxygen uptake rate, ammonium uptake
rate, or by balancing that also considered the total oxygen uptake and carbon dioxide
production. The ethanol reaction rate is inferred from the following linear relation
(OUR, CPR in molar quantities):
RE = 1.04 (CPR - OUR)
(9.51)
= 1.04 OUR (RQ - 1)

According to the last line, the above control law may also be considered as RQ-
control.
A quite different approach is the LlA-control proposed by Lakrori and Cheruy
[98]. It can also be formulated as adaptive control. The simple structured controller
uses the ratio of the controlled variable to the set-point as basic non-linear element,
and its power is used as a tuning or adaptation parameter. Such a control system for
baker's yeast fed-batch production was developed by Dantigny and Lakrori [99] in
simulations, then modified and experimentally tested [100]. The control law used
was

F= Ji-VL(Y~~~r (9.52)

where Cx and Ji- were estimated from ammonia uptake. The adaptation parameter of
the control law was determined by measurements of the respiratory quotient and
ethanol concentration as

k= ~5 ( gL-
CEI + RQ + 3.5) (9.53)
318 9 Baker's Yeast Production

In practical applications by this control, ethanol could be kept below 2 g 1-1 and RQ
close to one. The control law, Eq. (9.52), can also be interpreted as gain adjustment
of some kind of feed-forward compensation, similar to Eq. (9,48), by a feedback ac-
tion due to Eq. (9.53).
A control by directly using Eq. (9,48) with fine-tuning of the gain k was presented
in [109]. Based on process information obtained from state observation by a mathe-
matical model, k was stepwise up/down-adjusted to obtain high growth rate without
by-product formation in baker's yeast fed-batch cultivation.

References

1. Rose AH, Harrison JS (eds) (1987) The yeasts. Academic Press, London
2. Slater ML (1981) In: Arnold WN (ed) Yeast cell envelops: biochemistry, biophysics, and
ultrastructure, vol 2. CRC Press:65
3. Hartwell LH, Unger MW (1977) J Cell Bioi 75:422
4. Lord PG, Wheals AE (1980) J Bacteriol 142:808
5. Thompson PW, Wheals AE (1980) J Gen Microbiol 1221:401
6. Wheals AE (1987) In: Rose AH, Harrison JS (eds) The yeasts, vol 1. Academic Press,
London, p 283
7. Duboc P, Marison I, von Stockar U (1996) J Biotechnol 51:57
8. Nielsen J, Villadsen J (1992) Chern Eng Sci 47:4225
9. Steinmeyer DE, Shuler ML (1989) Chern Eng Sci 44:2017
10. Coppella SJ, Dhurjati P (1990) Biotechnol Bioeng 35:356
11. Cortassa S, Aon JC, Aon MA (1995) Biotechnol Bioeng 47:193
12. Jin S, Ye K, Shimizu K (1997) J Biotechnol 54:161
l3. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
14. Vanrolleghem PA, de Jong-Gubbels P, van Gulik WM, Pronk JT, van Dijken JP, Heinjen S
(1996) Biotechnol Prog 12:434
15. Rizzi M, Theobald U, Querfurth E, Rohrhirsch T, Baltes M, Reuss M (1996) Biotechnol
Bioeng 49:316
16. Theobald U, Mailinger W, Baltes M, Rizzi M, Reuss M (1997) Biotechnol Bioeng 55:305
17. Rizzi M, Baltes M, Theobald U, Reuss M (1997) Biotechnol Bioeng 55:592
18. Cortassa S, Aon MA (1994) Enzyme Microb Technol 16:761
19. Auling G, Bellgardt KH, Diekmann G, Thoma M (1984) Appl Microbiol Biotechnol
19:3533
20. Sonnleitner B, Rothen SA, Kuriyama H (1997) Biotechnol Prog l3:8
21. Barford JP, Phillips PJ, Orlowski JH (1992) Bioproc Eng 7:297
22. Barford JP, Phillips PJ, Orlowski JH (1992) Bioproc Eng 7:303
23. Lee YS, Lee WG, Chang YK, Chang HN (1995) Biotechnol Letters 17:791
24. Gadgil CJ, Bhat PJ, Venkatesh KV (1996) Biotechnol Prog 12:744
25. Wasungu KM, Simard RE (1982) Biotechnol Bioeng 24:1125
26. Grosz R, Stephanopoulos G (1990) Biotechnol Bioeng 36:1030
27. van Urk H, Mak PR, Scheffers WA, van Dijken JP (1988) Yeast 4:283
28. Sonnleitner B, Kappeli 0 (1986) Biotechnol Bioeng 28:927
29. Enfors SO, Hedenberg J, Olsson K (1990) Bioproc Eng 5:191
30. Sonnleitner B, Hahnemann U (1994) J Biotechnol 38:63
31. Barford JP (1990) Biotechnol Bioeng 35:921
32. Bellgardt KH (1994) J Biotechnol 35:19
33. Bellgardt KH (1994) J Biotechnol 35:35
34. Beuse M, Kopmann A, Diekmann H, Thoma M (1999) Biotechnol Bioeng 63:410
35. Strassle C (1988) PhD thesis No 8598, ETH Zurich
36. Strassle C, Sonnleitner B, Fiechter A (1988) J Biotechnol 7:299
37. Strassle C, Sonnleitner B, Fiechter A (1989) J Biotechnol 9:191
38. Cazzador L, Mariani L, Martegani E, Alberghina L (1990) Bioproc Eng 5:175
References 319

39. Munch T, Sonleitner B, Fiechter A (1992) J Biotechnol 24:299

40. von Meyenburg K (1969) PhD thesis, ETH Zurich
41. Munch T (1992) PhD thesis No 9772, ETH Zurich
42. Beuse M, Bartling R, Kopmann A, Diekmann H, Thoma M (1998) J Biotechnol 61:15
43. Heinzle E, Furukawa K, Tanner RD, Dunn IJ (1983) In: Halme A (ed) Modelling and
control of biotechnical processes. Pergamon Press, New York, p 57
44. Parulekar SJ, Semones GB, Rolf MJ, Lievense JC, Lim HC (1986) Biotechnol Bioeng
28:700
45. Bellgardt KH (1991) In: Rehm HJ, Reed G (eds) Biotechnology, vol 4, p 267
46. Dantigny P (1995) J Biotechnol 43:213
47. Deckwer WD, Yuan JQ, Bellgardt KH, Jiang WS (1991) Bioproc Eng 6:265
48. Rieger M, Kappeli 0, Fiechter A (1983) J Gen Microbiol 129:653
49. Yuan JQ, Bellgardt KH, Deckwer WD, Jiang WS (1993) Bioproc Eng 9:173
50. Yuan JQ, Bellgardt KH (1994) J Biotechnol 32:261
51. George S, Larsson G, Olsson K, Enfors SO (1998) Bioproc Eng 18:316
52. StrauB G (1990) PhD thesis, Universitat Hannover
53. Abel C, Hubner U, Schiigerl K (1994) J Biotechnol 32:45
54. Lubbert A, Korte T, Larson B (1987) App Biochem Biotech 14:207
55. Lubbert A, Larson B (1987) Chern Eng Tech 10:27
56. Pollard DJ, Ayazi Shamlo P, Lilly MD, Ison AP (1996) Bioproc Eng 15:279
57. Deckwer WD (1985) Reaktionstechnik in Blasensaulen, Verlag Salle/Sauerlander
58. Bello RA, Robinson CW, Moo-Young M (1985) Biotechnol Bioeng 27:369
59. Hills JH (1976) Chern Eng J 12:89
60. Weiland P (1978) PhD thesis, Universitat Dortmund
61. Adams J, Rothmann ED, Beran K (1981) Math Biosci 53:249
62. Hjortso M (1995) J Biotechnol 42:271
63. Hjortso M (1987) Biotechnol Bioeng 30:825
64. Hjortso MA, Nielsen J (1994) Chern Eng Sci 49:1083
65. Hjortso, MA, Bailey JE (1983) Math Biosci 63:121
66. Hjortso, MA, Bailey JE (1984) Biotechnol Bioeng 26:528
67. Porro D, Martegani E, Ranzi B, Alberghina L (1988) Biotechnol Bioeng 32:411
68. Alberghina L, Mariani L, Martegani E, Vanoni M (1983) Biotechnol Bioeng 25:1295
69. Mariani L, Martegani E, Alberghina L (1986) lEE Proceedings 133D:21O
70. Hjortso MA, Nielsen J (1995) J Biotechnol42:255
71. Takamatsu T, Shioya S, Chikatani H, Dairaku K (1985) Chern Eng Sci 40:499
72. Becker MJ, Rapoport AI (1987) Adv Biochem Eng/Biotechnol 35:127
73. Suomalainen H (1975) Eur J Appl Microbioll:l
74. Ginterova A, Mitterhauszerova L, Janotkova 0 (1966) Branntweinwirtschaft 106:57
75. Hautera P, Lovgren T (1975) Baker's Digest 49:36
76. Bellgardt KH, Yuan JQ (1991) In: Rehm HJ, Reed G (eds) Biotechnology, vol 4, p 383
77. Fisher DG (1991) Canad J Chern Eng 69:5-26
78. Johnson A (1987) Automatica 23:691
79. Rungeldier K, Braun E (1961) US Pat 3,002,894
80. Schugerl K (1998) Bioreaction engineering III. Wiley, Chichester, Amsterdam
81. Wu WT, Chen KC, Chiou HW (1985) Biotechnol Bioeng 27:756
82. Dairaku K, Izumoto E, Morikawa H, Shioya S, Takamatzu T (1982) Biotechnol Bioeng
24:2661
83. Takamatsu T, Shioya S, Okada Y, Kanda M (1985) Biotechnol Bioeng 27:1675
84. Dairaku K, Izumoto E, Morikawa H, Shioya S, Takamatzu T (1983) J Ferment Technol
61:189
85. Bellgardt KH, Kuhlmann W, Meyer HD, Schiigerl K, Thoma M (1986) lEE Proceedings
133D:226
86. Shi Y, Yuan WK (1988) Chern Eng Sci 43:1915
87. Axelson JP, Mandenius CF, Holst 0, Hagander P, Mattiasson B (1988) Bioproc Eng 3:1
88. Aiba S, Nagai S, Nishizawa Y (1976) Biotechnol Bioeng 18:1001
89. Cooney CL, Wang HY, Wang DIC (1977) Biotechnol Bioeng 19:55
320 9 Baker's Yeast Production

90. Wang HY, Cooney CL, Wang DIC (1977) Biotechnol Bioeng 19:69
91. Peringer P, Blachere HT (1979) Biotechnol Bioeng Symp 9:205
92. Woehrer W, Hampel W, Roehr M (1981) Proc Ferm Symp 6:419
93. Verbruggen HB, Eelderinck G, Van den Broeck PM (1985) In: Johnson A (ed) Modelling
and control of biotechnical processes. Pergamon, Oxford, p 91
94. Verbruggen HB, Eelderinck G, Van den Broeck PM (1987) BTF 4:185
95. Deckers R, Voetter M (1985) In: Johnson A (ed) Modelling and control of biotechnical
processes. Pergamon, Oxford, p 73
96. Yamane T, Matsuda M, Sada E (1981) Biotechnol Bioeng 23:2509
97. Williams D, Yousefpour P, Wellington EMH (1986) Biotechnol Bioeng 28:631
98. Lakrori M, Cheruy A (1987) Bio-Sci 6:154
99. Dantigny P, Lakrori M (1992) Biotechnol Bioeng 39:246
100. Dantigny P, Lakrori M (1991) Appl Microbiol Biotechnol 36:352
101. O'Connor GM, Sanchez-Riera F, Cooney CL (1992) Biotechnol Bioeng 39:293
102. Chen L (1992) PhD thesis, Universite Catholique de Louvan
103. Axelson JP (1989) PhD thesis, Lund Institute of Technology
104. Pomerleau Y, Perrier M (1990) AIChE Journal 36:207
105. Pomerleau Y, Perrier M (1992) AIChE Journal 38:1751
106. Tyson CB, Lord PG, Wheals AE (1979) J Bacteriol 138:92
107. Schubert J, Simutis R, Dors M, Havlik I, Lubbert A (1994) J Biotechnol 35:51
108. Shimizu H, Miura K, Shioya S, Suga KI (1995) Biotechnol Bioeng 47:165
109. Ringbom K, Rothberg A, Saxen B (1996) J Biotechnol 51:73
llO. Siimes T, Linko P, von Numers C, Nakajima M, Endo I (1995) Biotechnol Bioeng 45:135
111. Zhang XC, Visala A, Halme A, Linko P (1994) J Biotechnol 37:1
10 Modeling of the Beer Fermentation Process
Andreas Lubbert

10.1
Introduction

Over the last decades, huge, centrally located breweries with factory-style produc-
tion and considerable brewing capacities were built. At the same time the beer con-
sumption has been decreasing in central Europe. Thus, the breweries fight over mar-
ket shares. Consequently, optimization of the benefit/cost ratio becomes decisive for
breweries to keep themselves at the competitive edge. As the beer quality is not ex-
tremely different between breweries in a given area, the beer price and hence the
production cost becomes an overwriting criterion. Cost must be cut wherever pos-
sible, but labor and energy costs are the primary targets of process optimization.
Central in the creation of beer is fermentation. Here we focus on this fermentation
step, which can be regarded as the heart of the entire production, since the condi-
tions of this fermentation process have a marked effect on the flavor and taste of the
beer (e.g., [1, 2]). The beer fermentation process is a batch process where a Sacchar-
omyces yeast (Saccharomyces cerevisiae, Saccharomyces carlsbergensis, Saccharomyces
uvarum, etc.) is used to produce ethanol from barley malt. This anaerobic process
typically takes 7 days. Here we discuss optimization procedures for the fermentation
process with respect to the operational procedure and do not consider constructive
aspects of the equipment. In other words, we are dealing with everyday work in
breweries: optimizing a running fermentation system for the production of a parti-
cular brand of beer.
Beer fermentation is one of the large-volume industrial processes that are essen-
tially controlled manually. There are several reasons that excluded automation from
this process so far. One is that the essential process state variables are difficult to
measure on-line. Another is that the fermentation process appears to be a complex
system, which had not been modeled precisely enough so far. On the other hand,
however, the brewmasters make excellent beers with their current technology and
any automatic process supervision and control system must compete with these
highly skilled human controllers. Automatic systems would only be of interest to
breweries if they would help the brewmasters to produce the same quality of beer
in a much less expensive way. There are some obvious entry points for further auto-
mation in beer fermentation. In beer breweries, many tasks presently being per-
formed by men, such as manual cleaning, filling, pitching, measuring, temperature
control, yeast harvesting, pumping into lager vessels, etc., could be automated [3].
A main issue in automating beer fermentation is sensing the actual state of the
process and making the necessary decisions concerning eventually necessary correc- .
322 10 Modeling of the Beer Fermentation Process

tions of the process based on this information. The key variable by which the
brewers characterize the state of their fermentation is the extract (substrate) concen-
tration. It is measured off-line, and thus the measurement values become available
with a considerable time delay. The decisions the brewmasters make upon deviations
from the data they expect are essentially based on their experience with the time
profiles of the extract concentration. At least in Germany (Reinheitsgebot), they
can influence a running fermentation by changing the process temperature only.
Adding yeast extract or water for control purposes is seldom performed in real prac-
tice.
Automatic process supervision and control would require one to formulate the
relevant knowledge about the fermentation process in such a way that it can be
exploited numerically with computers. The software, representing the process in
such a control system, is the way by which we implement the process model. Ob-
viously it does not suffice to know only the profiles of a perfectly running fermenta-
tion process. Control actions are required whenever the process significantly devi-
ates from those paths, and hence the dynamics of the process around the typical
trajectories must be known as well. Thus, dynamic process models are required.
Here we discuss the problem of how to formulate and process the knowledge re-
quired to perform process optimization, supervision, and control tasks. The accom-
panying examples used throughout are taken from industrial-scale beer fermenters
(Fig. 10.1) operated in one of the breweries of the Gilde Brauerei AG, Hannover, Ger-
many. The brand brewed in the fermenters is based on a wort with 110 gil extract

5
TIlermornt:[tr. Room, T.
TI'u:mlQr'l\c:[t=r. HC':.d Sp:;u:e. T I).,
15 Thermometer. Reactor Wall . T..
Vent Lim:
Measure-men! Probe

'\
ThennOlTIctl::T. Beer. Head Spac:c:.
6 T..
7 Cooling Surf""",
I 8 Thermometer. Vent Line
9 Pteuurt: Sensor
7 10 Swing Whirl
11 fR-De\'ice Polytron IR -
12 Collc::ction Line for CO,
16 13 Sampling l'limp
14 Thtnl'lomelcT. Beer, Ta
IS r-o.am Layer
16 Beer
17 Sampling Line

14

Fig. 10.1. Production -scale beer fermenter operated in one of the breweries of Gilde Brauerei
AG, Hannover
10.2 Process Optimization 323

(barley malt extract) concentration. The initial yeast concentration is of the order of
1 g(dry weight)/L. The fermenters are modern cylindro-conical fermenters with a
total volume of 300 m 3 and a diameter of 4.2 m. They are cooled by means of three
cooling elements each, welded as jackets onto the reactor walls around their cylind-
rical parts. The cooling is performed by evaporation of ammonia within the closed
cooling system. All data were taken during their usual operation [4, Sa, 6, 7].

10.2
Process Optimization
10.2.1
Different Knowledge Representation Techniques

In order to optimize a process, one first of all needs sufficient knowledge about its
dynamics. Beer fermentation is a classical biochemical production process, which
has been performed by humans for many centuries. Hence, there is a lot of heuristic
knowledge available. From the process engineering point of view, however, systema-
tic optimization requires quantitative models that can be exploited numerically.
Since the fermentation process is basically a biochemical process, namely the con-
version of the raw material barley malt (sugar) into ethanol and carbon dioxide, the
overall biochemical reaction equation is the first approach to a process model. In the
beer community this is referred to as the Balling formula. It is usually formulated on
a mass base (component changes measured in kg) in the following form:
1 kg Sugar --+ 0.484 kg EtOH + 0.463 kg CO 2 + 0.053 kg Biomass (10.1)
As the process is exothermic, the mass balance can be complemented by
~H = 568 kJ/kg (10.2)
leading to the heat generated per kg sugar consumed.
This basic balance does not take into account that a fraction of the fermentable
carbohydrates, here summarized as sugar, is utilized for the production of further
metabolites such as glycerol, organic acids, acetaldehyde, higher alcohols, and esters.
The reason is that, although these other fermentation products are crucial from the
point of view of the beer flavor, they do account for only 0.27% of the sugar utilized
(e.g., [8]).
Of practical interest is to keep the biomass production rate relatively low in order
to get a high beer/substrate yield. In practice this is widely achieved, and thus the
first rough models of the beer fermentation process neglect biomass. That means
that over a wide part of the fermentation time the net conversion is essentially sugar
to ethanol and carbon dioxide. Such a reaction equation is a static relationship be-
tween the amount of sugar consumed and the amounts of products built. It tells what
the decisive quantities are and how they are interrelated. However, it does not tell us
about the process dynamics, i.e., the time course of the individual concentrations in
the fermenter.
324 10 Modeling of the Beer Fermentation Process

10.2.1.1
Classical Approach

The classical approach to dynamic process models is to start with mass balance
equations for those components, which are consumed or being built to a significant
extent during the conversion process. These are the components discussed in the
stoichiometric model (Eq.10.1). Since beer fermentation is operated in the batch
mode, the first approach to a dynamic model is a rather simple component balance
equation. When the balance is drawn over the entire mass of young beer, we obtain
dc
-=R (10.3)
dt
where R is the biochemical rate of change of the components considered, the con-
centrations of which are combined in a vector c= [biomass; sugar; ethanol; carbon-
dioxide].
The essential additional aspect in this balance equations is the kinetics, which is
coming into play via the conversion rate vector R. The most obvious way to specify it
is first to try the simple kinetic expressions formulated by Monod. However, one
immediately finds that this does not work sufficiently well. It is necessary to note
that in a batch fermentation the initial substrate concentration is too large for the
yeast cells to work at full speed so that substrate inhibition must be taken into ac-
count. At the end of the fermentation the concentration of the product ethanol be-
comes inhibiting. So product inhibition must be taken into account as well.
For both effects simple extensions of the Monod expressions are known from the
literature. 0", the specific substrate consumption rate, essentially determines the dy-
namics. It may be described as a function of the concentration S of the substrate
(extract) and of the product ethanol (E) being built:
, S Ksi Kei
IJ = IJmax - - - - - - (10.4)
Km + S Ksi + S Kei + E
The second factor is the pure Monod expression; the third considers the substrate
inhibition and the last the product inhibition. The corresponding parameters Km,
Ksi, and Kei must be determined from measurement data.
The rate vector R might be formulated by:
(10.5)
where X is the biomass concentration, and Yxs , Yes' and Yes the yields, the values of
which can be estimated from the stoichiometric relationship at Eq. (10.1).
Since beer fermentations are controlled by the temperature only, the temperature
influence on the yeast activity must be quantified. 0" is affected in two ways, being
increased in essentially the same way as any chemical reaction. This can be de-
scribed by an Arrhenius relationship. On the other hand, at higher temperatures a
decrease in the activity is observed, which can also be described by an Arrhenius-
like expression. The entire temperature dependency of 0" then has the form
IJ = IJ exp( -Eakt/RT) (10.6)
(1 + KdexP( - Edeak/RT)
The corresponding activation energies Eakt and Edeak as well as the intensity factor
Kd must be determined from experimental data. R is the gas constant.
10.2 Process Optimization 325

unfortunately, these extensions of the pure Monodequation are not sufficient

either. The reason is that, as known from literature, the concentration of the yeast
and its activity is not readily described by simple stoichiometric or kinetic relation-
ships.
Sufficient knowledge about the concentration of biomass in production-scale beer
fermenters is usually not readily available. Values published in the literature depict
much scatter. They also much depend on the particular control strategy of the brew-
ery. Reliable data about the active part of the biomass during the beer fermentation
is missing in the literature. Only some measurements performed with hemacyt-
ometers after staining the yeast cells with methylene blue to determine viability
(e.g., [9]) are available.
Hence, the fermentation must be describe in an alternative way.

10.2.1.2
Heuristic Approach

Heuristically it is straightforward to divide the entire process into phases, which can
separately be described in the classical way. Such a partition makes sense whenever
the process behavior will be dominated by different mechanisms in the different
phases. For instance, in the batch at the beginning of beer fermentation, substrate
inhibition takes place which is completely unimportant at the end of the fermenta-
tion. Hence, this type of inhibition only needs to be taken into account in a model
for an early phase of the fermentation.
This simplification of the models for the individual phases has the advantage that
the models can be kept smaller, i.e., they contain a smaller number of free para-
meters. Consequently, the accuracy by which the model parameters can be identified
from given data sets is much higher than in a comprehensive "world model."
The price one has to pay for this advantage is that one must decide when to switch
over from one phase to the next in a practical simulation of the entire process. For
that decision, the heuristic knowledge about the process accumulated by brewers can
be used. Since this knowledge can often be formulated in terms of simple "if ... then
... - rules", it is straightforward to used small rule-based systems in order to make
use of that knowledge. Havlik et al. [4] showed that such rule systems can easily be
formulated by fuzzy rules which can be processed in computers using fuzzy logic. In
this way, missing mechanistic knowledge is partly compensated for by heuristic pro-
cess knowledge together with the basic mechanistic knowledge contained in the var-
ious classical models for the individual phases. The simultaneous utilization of such
different representations of a priori knowledge in a process model is referred to as a
hybrid model approach. It could be shown that such hybrid models can be used for
process state estimation or process supervision purposes. Utilization of heuristic
knowledge in addition to the classical models leads to an added value in the formu-
lation of process knowledge and thus to an enhanced performance of model sup-
ported techniques where the models are used.
In practice one proceeds as follows. One starts with the model for the first phase of
the fermentation and, with the passage of time, particularly during the time interval
where a switch to the next phase is expected, the next model is also exploited. If
more than a single model is considered, an appropriate weighting of the individual
models becomes necessary. It is straightforward to do this by fuzzy reasoning. This
326 10 Modeling of the Beer Fermentation Process

allows a smooth switch over from one model to the next, thus avoiding unrealistic
jumps in the signals for the key process variables.
It must be clearly stated that the reason for formulating such hybrid models is not
to replace classical models, but instead to extend the latter by making available addi-
tional knowledge in order to increase the performance of the model. Where there is
no further knowledge to extend the existing models on the basic mechanistic level,
one must make use of the relevant heuristic knowledge, and the fuzzy rule proces-
sing provides a powerful tool to formulate these heuristics and to add this knowl-
edge to a classical model.

10.2.1.3
Alternative Methods to Describe the Kinetics

When even additional heuristic knowledge about the part of the process under con-
sideration is missing, one is forced to look for further information about the process
dynamics. For production processes measurement data records are usually available
and it is straightforward to proceed on the traditional path of engineering, namely to
develop so-called engineering correlations. As is well known, such correlations are
black-box models relating the desired key quantities of the process, i.e., those deter-
mining the bioreactor performance, to quantities which are adjustable indepen-
dently. For instance, the Monod formula is such an expression which formally de-
scribes the dependency of the key variable specific substrate consumption rate (j
from the quantity substrate concentration which can be adjusted in a fermentation.
Recently, artificial neural networks were shown to depict a much higher capacity
for representing complex multidimensional nonlinear relationships between process
variables. Their flexibility and accuracy in describing complex nonlinear relation-
ships in biotechnical processes was demonstrated by many researchers, e.g., Simutis
et al. [Sb 1, who showed in several examples that the kinetics in biotechnological
production systems can be accurately described by artificial neural networks. When
the kinetics can be represented by an artificial neural network this can easily be used
during the solution of the basic dynamic mass balance equation system just by call-
ing the network evaluation routine as a subroutine for determining the actual
rates R.
The advantage of using this most flexible representation for kinetic expressions in
hybrid models must be paid for by dispensing with the conventional network train-
ing procedures, since the relevant rates cannot usually be measured directly. For
such situations, Schubert et al. [10] developed a special training technique: the sen-
sitivity approach that uses the data usually available during the cultivation. Evolu-
tionary algorithms can also be used to train such a hybrid model on the available
process data.
When it is possible to describe the kinetics alternatively by a network or the clas-
sical Monod-like approach, there are two reasons for using both. The first is that
artificial neural networks cannot be used for extrapolation, i.e., they cannot be used
out of the area in the state space from which data were taken to train them. In cases
where an extrapolation becomes necessary, it is believed that the classical models are
much more reliable. However, one should generally be careful in using models (may
they be classical ones or those based on networks) outside of the area which was
experimentally explored before. Serious predictions can only be based on models
that have been validated, and where there are no data no validation can be per-
10.2 Process Optimization 327

formed. The second reason is that the Monod-type kinetic models are known to have
excellent global description properties in contrast to the networks which perform
better when locally describing the process around the actual point in the state space.
This suggests the use of both representations simultaneously just as if considering
two votes given from different points of view.
Obviously, the weighting of the votes must be based on the amount of data avail-
able in the relevant area of the state space. By means of cluster analysis it is possible
to evaluate the relevant area in the state space with respect to the sufficiency of data,
and to express this by some kind of reliability or evidence measure E, with values
between 0 and 1. This evidence measure can be used to weight the neural network
component of the model, and the classical representation then gets the complemen-
tary weight (complement to 1) in the superposition of the model outputs.
Before one can make use of the models described for process optimization it must
be ensured that they are reliable enough for long-term process state prediction, i.e.,
for predicting the course of the fermentation process over long time horizons up to
the scale of the entire fermentation time. Brewers will only make decisions on the
outcome of such state predictions after the underlying models have be sufficiently
validated on measurement data. This can only be done by comparison with experi-
mental data from the particular fermentation process under consideration. Of
course, such a comparison must be made with independent data, i.e., data that have
not been used during the development of the process model. This test procedure is
referred to as the cross validation procedure in the literature.

10.2.2
State Prediction for Process Optimization

State predictions are of immediate interest to the industrial beer brewery. They allow
one to estimate the end of the current fermentation runs more accurately than was
possible before. Hence, scheduling of the tank capacity in the brewery can be made
more precisely. This is a cost argument. On the other hand, one is interested in mon-
itoring the quality of the young beer.
In the beer brewing industry, the characteristic buttery taste of vicinal diketones
(VDK) has long been known to be a major problem. Human taste thresholds of the
two vicinal diketones diacetyl and pentanedione, which may appear at significant
concentrations in lager beers, are rather low: 50-100 I-lg/l for diacetyl and 100-
500 I-lg/l for pentanedione [11]. Because of the considerably higher taste threshold
of pentanedione, which appears in lager beers at roughly the same concentration as
diacetyl, breweries mostly concentrate on keeping the diacetyl concentration low. At
the end of the fermentation, the diacetyl concentration should be below the thresh-
old value in order to avoid extended lager times or blending. Thus, it is of interest to
predict the final diacetyl concentration at the earliest possible time in order to be
able to reduce it by changing the fermenter control, i.e., the temperature profile,
appropriately. Since diacetyl is not considered in the basic model discussed above,
the latter must be extended appropriately.
There is enough experimental evidence showing that the diacetyl concentration in
a normal beer fermentation rises in the first few days, crosses a maximal value [12,
13] about the third day, and then decays monotonously with time. The formation of
vicinal diketones result from the oxidative decarboxylation of excess a-acetolactate
hydroxyacid that leaked from the isoleucine-valine biosynthetic pathways (e.g., [14]).
328 10 Modeling of the Beer Fermentation Process

dDA
dynamo
ANN

(4,10,\)

Fig. 10.2. Artificial neural network (ANN) structure used by Manikowski et aL [7] to represent
the kinetics of the diacetyl (DA) degradation, The ANN is complemented by an integrator
(INT) as explained in the text

When the yeast cell produces too much a-acetolactate, it exports it into its culture
medium, The biochemical explanation for the degradation of this diketone is that
the yeast cells take up and degrade it enzymatically.
The formation of diacetyl appears to be a very complex process with many degrees
of freedom, while the degradation of diacetyl seems to be significantly more simple
as it depends mainly on the metabolic activity of the yeast cells. Since both the
knowledge about the diacetyl degradation and, in particular, its formation are not
so clear that a mechanistic model could be built, a first approach for a quantitative
representation was made by means of a black box approach based on the available
experimental data only. As already mentioned, the most universal and flexible way of
black-box modeling is using artificial neural networks.
Fig. 10.2 shows the network structure used by Manikowski et al. [7]. It is a simple
feed-forward network with a single hidden layer containing 10 nodes. As input
signals, the fermentation time t, the young beer temperature T, the substrate con-
sumption rate q" and, finally, the diacetyl concentration DA, as interpolated from
the measured values, were taken. The only output value is the net rate of change
d(DA)/dt of the diacetyl concentration. In order to get the diacetyl concentration,
an integrator must follow the network output. This construction was chosen since
the rate depends more sensitively on the input variables. The a priori information
used in the network representation is restricted to the assumption that the tem-
perature T is influencing the process as well as the metabolic activity of the yeast
as represented by the extract degradation rate qs' Furthermore, the heuristic
knowledge about the sensitivity of the diacetyl conversion rate on the input vari-
ables is used.
In order to allow for a dynamical determination of this rate of change, the outputs
were smoothed and fed back to the entrance of the network. In this way it is possible
to determine the time development of the output quantity when only a single start
value is provided. For the network training, however, complete data sets for the dia-
cetyl concentration as a function of time are required.
This network must be trained with data records from several fermentations. Such
a training is stopped at the minimal deviation between the network's output and the
corresponding measurement data. As already mentioned, further independent data
sets from additional production runs must then be used to validate the network.
The graphs in Fig. 10.3 provide an impression of the quality of the neural network
representation of the diacetyl profile. In this figure the trained network is compared
with the measurement data from one of the training data records. It provides a good
representation of the process.
However, when it is then applied to the validation data set that was not known to
the network before, much less agreement is observed. Figure 10.4 shows that the data
lO.2 Process Optimization 329

0,7,------------------,
Fennentation M3186 A Measurement Val ue
0,16
0,6
------- Estimation/Neural Net 0,14
Deviation Estimation I - 0,12
Measurement
- 0,10 ;...
- 0,08 Oil

- 0,06
oS
=
~...
- 0,04
- 0,02

0,1
- 0,00 ~
- -0,02
-0,04
168
Fermentation Time [hI

Fig. 10.3. Modeling result of the total diacetyl profile (training data) together with experimen-
tal data. The curve in the lower part of the figure represents the deviation of both results

0,7
Fermentation M3204 A Measurement Value
0,16
0,6 ------ Estimation I Neural Net
I> 0,14

::-" 0,5 0,12

;... ;...
1
I> 0,10
Oil
0 ,4 O,OS
1>1> !
~O,3 ~'"
1>1> "

I>
0,06
~
=
0,04

S'<U"
"'"
.~
0,2 0,02
~
...
0,00
0,1 - - Deviation Estimation
I> Measurement -0,02

0,0 -0,04
24 48 72 96 120 144
° Fermentation Time [hI

Fig. 10.4. Modeling of the total diacetyl profile using the test data

used for the training were not sufficient to train the neural network in such a way
that it becomes possible to describe the diacetyl profile to a sufficient accuracy.
This is an excellent example to demonstrate that modeling without a solid valida-
tion can provide misleading results. After a closer investigation of the problem it
became clear that the entire process, i.e., both diacetyl formation and degradation,
cannot be modeled on the available data. However, it was found that at least the
simpler diacetyl degradation phase can be modeled with the available measurement
values.
The degradation process was again modeled by means of a dynamic artificial
neural network of the same structure as before, where the input variables were the
actual fermentation time, the substrate degradation rate qs as determined from the
extract data, the temperature of the young beer, and the measured diacetyl concen-
tration.
Alternatively, for comparison, a simple mathematical model can be developed
based on the assumption that the diacetyl degradation follows a first order reaction
as suggested by Dellweg [15] and Inoue [16]. The corresponding temperature depen-
330 10 Modeling of the Beer Fermentation Process

dency of the rate was chosen according to Rice and Helbert [17] and Yamauchi et al.
[18]. As with the model of Garcia et al. [19] one can formulate as follows:

dCDA/dt =kl x qs + k z X CDA x exp(-B/T) (10.7)

with
kl = l.S88e-3 [-]
k z = 4.694e-2 [h- 1 ]
B = 18.26 [0C]

The three parameters can be obtained by fitting this model to the experimental data.
The results are depicted in Fig. 10.5
The full lines in Figs. 10.5 and 10.6 show typical results of the two diacetyl estima-
tion techniques. The simple mathematical model (dashed curve) and the artificial
neural network (full curve), both not used in MM and ANN training procedures,
agree quite well with the measurement data. Such tests were repeated with several
other test data records. Thus, by cross validation, it was ensured that this model can be

Fermentation M3189 1::. Measurement 0,18

0,50
0,45
---.- Math. Model 0,16
_Neural Net
0,14
::-" 0,40 0,12 .:::-"

~ -~_>- ~
!::
~
0,35

0,20
------~--"~
1::.
0,10

~:: ~
0,04 :8
~ 0,15 0,02 .~
0,10 0,00 Q
0,05 -0,02
0,00 -"---'--~--'-~_-T--_--~--_--_Marth_.~Mod_e,I_~NeTu_ral~N_et-,-.1 -0,04
84 96 108 120 132 144 156

Fermentation Time [h]

Fig. 10.5. Estimation of the diacetyl degradation in a typical fermentation supported by means
of a simple mathematical model and an artificial neural network

Fermentation M4187 t:. Measurement 0,18

0,50 -----Math. Model 0,16
--------Neural Net
0,45 0,14
P--I Q,4Q
......

~ 0,35 AM
AAAAltA~ !::. -, --__ "'_
0,12
O,10~
rO,3Q -- O,OS CJl
- A 0,06 .§.
:&0,25
1:l 0,20
t:."',& 0,04 §
~ 0,15 \ ~"--_ r:::: 0,02 -.;
0,00 .;;:
0,10 \j=zJ' -0,02 Q
0,05 Deviation of Estimation from Measurement
-0,04
---- Math. Model -Neural Net
0,00 -I--~---,--.-,...,r-o~--,---.----,~-,--..--,~---,-~,--1
24 48 72 96 120 144 168 192 216

Fermentation Time [h]

Fig. 10.6. Model-supported estimation of the diacetyl degradation in a typical fermentation

using a simple mathematical kinetic model
10.3 Process Supervision 331

used throughout the fermentations in the fermenters from which the data were taken
for training.

10.2.3
Remarks on Hybrid Models

The beer fermentation process can be described quite accurately by means of hybrid
process models that combine mechanistic information about the process in the form
of general mass balance expressions with heuristic information about the various
process phases and detailed local information extracted from the extended data re-
cords, usually available from production fermenters, to describe kinetic details by
artificial neural networks. Since, compared to classical mathematical models, the
hybrid models make use of additional knowledge and measurement information,
they perform better for state prediction.
It must be stressed however, that hybrid models cannot be validated when the data
base is insufficient. The important problem of diacetyl generation cannot be predicted
on the data usually acquired during an industrial beer fermentation process. When its
development is to be predicted as well, additional measurement information is re-
quired. On the other hand, the diacetyl degradation can be described; however, this
requires a start value which expresses the amount of diacetyl to be degraded. Thus, as
shown before, at least one measurement value of the diacetyl concentration must be
measured at a time where the diacetyl degradation dominates over the development
process to be able to predict its path for the rest of the fermentation.
One important remark which should be made here is that all the models discussed
so far assume that the fermenter is a well mixed reactor. This is at least questionable,
since it is well known that big fermenters often involve considerable inhomogeneity.
The incorporation of fluid dynamics into these models to an accuracy comparable
with the modeling of the conversion process is inhibitively difficult and time con-
suming at the moment. Experience showed that detailed fluid dynamics is not indis-
pensable for this type of model. The only consequence that must be drawn from
recognizing this problem is that the kinetic parameters used in the models must be
taken from experiments in the fermenters of the original size, since they contain all
the information about the transport problems not resolved in the models.
In this section it was described how to formulate the knowledge about the process
that is quantitatively available in order to predict the behavior of the process under
different operational conditions and to optimize it with respect to a predefined ob-
jective function. When this task has been performed the real process must be driven
along that control path. In reality one must usually cope with distortions so that an
open loop control must be augmented by a closed loop feedback control in order to
counteract deviations from the desired path. In the next section we first discuss on-
line state estimation and in a later section control. Since it is only possible to manip-
ulate temperature, we will concentrate on temperature control.

10.3
Process Supervision
In order to supervise a running production process, information about its actual
state is required. The definition of the state is problem dependent. With the state
332 10 Modeling of the Beer Fermentation Process

variables, those aspects must be described that are relevant to the task to be per-
formed. Here, this task is producing a particular beer at minimum cost. What
brewers found by experience over centuries is that the main state variable is the
extract concentration, and if one would like to go into more details ethanol concen-
tration and the diacetyl concentration are the next candidates. For state estimation it
is indispensable to make on-line measurements.

10.3.1
On-Line Measurements are Difficult to Perform

Obviously, one first tries to measure the state variables discussed above. The most
important problem appearing in this context is that none of the concentrations of
the key components can easily be measured on-line. Thus, indirect measurements
are required which make use of measurement data that can easily be obtained during
the process and which are uniquely related to the desired state variables, i.e., the
concentrations of the components in the basic reaction equation (Eq. 10.1).
As already mentioned, beer brewers traditionally take the extract concentration as
their key process state variable. They assume that the extract is uniquely related to
the other quantities that significantly change during the fermentation by means of
the stoichiometric equation (Eq.1O.1). Since even direct off-line measurements of
the extract concentration are not reasonably possible, they perform indirect off-line
measurements. From every point of view, extract concentration measurements via
manual density measurements proved to be the most economical technique in the
past. Even these measurements require a careful pretreatment of the samples drawn
from the fermenter in order to avoid distortions by COr gas bubbles developing
from beer as known to everybody. Hence, these measurements are time-consuming
and thus cannot be performed very often in a commercial beer fermentation. The
measurements required to keep a production-scale brewery under control are a sig-
nificant cost factor and large breweries try to reduce the amount of measurement
wherever possible. In practice, they are most often measured only once per day and
fermenter.
Density measurements are not easy to perform online during a fermentation.
Hence, much work has been devoted to finding alternatives. On-line measurability
is not the only criterion; the measurement technique must also be robust enough to
be applicable in the harsh everyday working environment in breweries at acceptable
expenses. Here expenses essentially not only means investment costs; more relevant
is the expense of manpower for operation and maintenance, which must be clearly
below that necessary for the currently used technique.
The alternatives to extract concentration measurements can be found in the basic
reaction equation (Eq. 10.1), which at the same time shows how they are related to
each other. Apart from biomass, which did not prove to be a good candidate, the
others - CO 2 development rate, ethanol production, and heat development rate -
are possible alternatives and have been discussed as such in the literature.
CO 2 development rates can be measured on-line by volumetric or mass flow mea-
surements in the vent line of the fermenters. The problem is that appropriate mea-
surement devices for the relevant conditions are rare on the measuring instrument
market. A more significant problem, however, is that CO 2 dissolves in the young beer
to a large extent, leading to supersaturation effects. The maximum level of super-
saturation of CO 2 reached during active fermentation increases with the fermenta-
10.3 Process Supervision 333

tion vessel capacity. The amount of CO 2 that remains dissolved at the end of brewery
fermentation is essentially constant (at a degree of supersaturation of approximately
1.5, equivalent to 11 % of the total CO 2 production) and could be calculated. The
profile of change in dissolved CO 2 concentration during fermentation followed the
same pattern as that of the rate of CO 2 evolution. The maxima for both profiles were
found to coincide [20).
Although the major portion of the dissolved carbon dioxide exists as the aqueous
species CO 2 (aqu), in typical ethanol fermentations, the solubility of the CO 2 is a func-
tion of the concentration of other non-polar and ionic species, in particular ethanol.
Thus, one has to consider the other molecular species that are in equilibrium with
CO 2 :
(10.8)
Because CO 2 (aqu) is able to react with different components dissolved in the culture
medium, e.g., with free amino groups of proteins, and H2 C0 3 to associate with posi-
tively charged groups on proteins via a dipole-protein interaction, the total amount
of CO 2 in the solution can increase significantly, causing high supersaturation levels.
The gas exchange CO 2 (aqu)=C0 2 (gas) further depends on the rate of nucleation.
This rate could be a rate-limiting effect for the liquid-gas mass transfer. The result-
ing supersaturation leads to a non-ideal behavior of CO 2 in solution and the pre-
sence of increasing amounts of HC0 3 and CO 2 (aqu) [21). This discussion shows that
it is not simple to determine the progress of the fermentation with simple point
measurements of the specific weight or the carbon dioxide evolution rate.
Since the COr transfer into a bubbly gas phase is also dependent on the local tem-
perature and the local mixing conditions, the actual gas flow rate depicts heavy fluc-
tuations of more than 100% during the main fermentation, where the measurement
values would be needed to supervise the process. From that point of view, CO 2 is not
a preferred measurement variable.
The situation is much better with ethanol, since the ethanol concentration does
not fluctuate very much during the fermentation. However, direct measurements of
the dissolved ethanol is also difficult to obtain. An alternative is to make use of the
fact that there is an equilibrium ethanol concentration in the gas phase of the reac-
tor. Hence, some ethanol leaves the reactor via the vent line. It was shown that the
ethanol concentration in the off gas nicely reflects the dissolved ethanol concentra-
tion in the young beer. The partial pressure of ethanol within the gas-phase can be
measured with several devices. For on-line measurements in a brewery, infrared sen-
sors are best. Measurement devices for supervising the ethanol concentration in
room air that can be used to monitor the ethanol in the vent line are commercially
available. The same applies to devices used to detect ethanol in the breath of hu-
mans. Both work in the relevant concentration interval and can be adapted to the
needs in a fermenter.
Finally, heat development measurements can be performed by measuring the heat
transferred into the cooling system of the beer fermenters. In new fermenters the
appropriate measurement devices can easily be installed. Already producing fermen-
tation halls, however, cannot be retrofitted easily.
The reason why biomass is not an appropriate measure for the state of the fermen-
tation is that it is not developed in a stoichiometrically simple and stationary way
from the sugar being consumed during the fermentation. Thus, the basic stoichio-
metric equation is changing with time. Hence, we need dynamic process models
334 10 Modeling of the Beer Fermentation Process

even for process supervision. This discussion shows that process supervision is not
simple and must be performed indirectly, using so-called software sensors for the
most important process variables.

10.3.2
Estimation of the Extract Degradation

As the state variables cannot be measured on-line, they must be measured indirectly.
Hence we need a relationship between the measurement quantities and the desired
state variable. It would be ideal if there were a simple correlation between the mea-
sured quantities and the extract degradation. Unfortunately things are more complex
so that one needs a more detailed description. An entirely mechanistic model of the
brewery fermentation is not yet possible. Furthermore, the exact description of the
extract degradation suffers from the fact that many parameters of the natural raw
material influencing the fermentation cannot be described quantitatively to a suffi-
cient accuracy. Also, the parameter values fluctuate from charge to charge. Mani-
kowski et al. [7] proposed a first approach in the form of a simple first-order math-
ematical model relating the extract concentration to the ethanol measurement data.

10.3.2.1
Simple Mathematical Model

The steady state ethanol concentration CE,g in the off gas generally depends on the
ethanol concentration CE,\ in the beer and the temperature T. Raoult's law allows one
to determine the partial pressure PE of the ethanol in the liquid phase (e.g., [22]:
(10.9)
where XE is the relative amount of ethanol in the liquid, PE,O the vapor pressure of
pure ethanol, and YE the activity coefficient of ethanol. By Dalton's law one gets

PE =YE X Pges (10.10)

where YE is the relative amount of ethanol in the gas phase, and Pges the total pres-
sure of the gas mixture. By combination of both formulae, we obtain the Raoult-
Dalton law for non ideal solutions of ideal gases:
(10.11)
The activity coefficient YE is a function of the ethanol concentration in the beer.
The relative amount of ethanol in the beer usually does not rise to values beyond a
value of 0.016. In that region of low EtOH concentrations, the activity coefficient YE
can be considered constant and the transformation of relative amounts to concentra-
tions can be linearly approximated. In a fermentation performed at normal atmo-
spheric pressure in the head space of the fermenter we then get
YE = const., Pges. = const.
XE = alxCE,\ al = const.
It follows:
kz = al xyE/Pges.
YE = kZxCE,\XPE,O
10.3 Process Supervision 335

and, since YE=PM-kl> the ethanol concentration CE,\ in the beer is:

CEI -_ YE - kl (10.12)
, k2 X PE,o
Since the coefficients kl and k2 depend on the properties of the concrete fermenter,
they must be determined separately. Fits to the available data led to values
kl = -222.9845 x 10-6 H
k2 = 13.5006 x 10-4 [lg-Ibar- I]
According to Balling, the relationship between the fermentable sugar consumed and
the ethanol concentration CE,\ formed in the beer is
S=STW-CE,Ji0.484
where S [g I-I] is the total concentration of the fermentable sugars and STW the
concentration of the original wort [g tl]. With the sugar table, the true extract Ew
and by Balling's formula the virtual extract Es can be determined [23].
By Antoine's equation, the vapor pressure PE,O over a pure ethanol phase depen-
dents on the temperature T:
(10.13)
The parameters A, B, and C are material dependent. They are tabulated by Gmehl-
ing et al. [24].
Since the temperature Tg at the gas-liquid interfacial area cannot be measured,
Manikowski et al. [7] tried to find out which of the available temperature values
can best be used to approximate its true value. Two main alternatives were tested -
the gas temperature TKR in the head space of the fermenter as measured with the
sampling station and the beer temperature TB' The results indicated that the head
space temperature should be used rather than the beer temperature.
However, there is usually no thermometer in the head space of a production fer-
menter. Since it is too cost intensive to install additional thermometers in an existing
production fermenter, a simple alternative was looked for. The result of this investi-
gation showed that it is possible to obtain fairly reliable estimates for the relevant
temperature from a contact-thermometer attached to the outer surface of the wall of
the fermenter directly at the upper edge of the cylindrical part below the insulation
layer (Fig. 10.1).

10.3.2.2
Estimation of the Extract Degradation by Artificial Neural Networks

As an alternative to the simple mathematical model described in the previous sec-

tion, Manikowski et al. [7] used a black box model: An artificial neural network was
taken to make direct use of the original measurement data from the plant for repre-
senting the relationship between the extract concentration and the available mea-
surement values without referring to potentially imprecisely known physical details.
Apart from the data, the only a priori information they used concerned which of the
available measurement variables will influence the substrate degradation.
Different possibilities for the construction of artificial neural networks were tried.
Two static feed-forward three-layer sigmoidal networks with three input nodes, ten
nodes in the hidden layer, and a single output node were trained. Both use the fer-
mentation time t and the temperature T at the reactor wall as input variables. They
336 10 Modeling of the Beer Fermentation Process

differ by the third input variable which, in the first network is the EtOH concentra-
tion in the off-gas line, and the corresponding EtOH value determined by the simple
mathematical model described above in the other one. The networks were trained on
the available process data using the conventional error backpropagation technique
[25].
In order to allow a cross validation of the artificial neural networks, the available
data were divided into a set of training data and a set of test data. A set of 20 fer-
mentation records, which were obtained under the same operational conditions,
were used as the validation data set. The results of the extract estimations are de-

11 Fermentation M4 194

_._ lnterpolabQt'l 01 SaOCharorreler Valoos

- - - . Extract EshmlbOn on Math. Mo(feI
- Extract Esllma'boo on ANN 1

24 36 48 60 72 84 96

Fermentation Time [hI

Fig. 10.7. Estimated extract values for a typical fermentation

FermentatIOns M3186· M7222

.
~6
~
.
.
..
..
. Means

. ".
.2] 5 ,/
" .. "
..
W ....'h, ModeI

" . .. • ... "I ·"

\..
(/)
"I" ....

... . "
::i! ANN1
II: 3 .. " ANN2
\"
: .
6 10 12 14 16 18 20 22

Fermentation Number

Fig. 10.8. Comparison of the rms error values of the individual fermentations

Table 1. Summary of the mean errors of the three models for the extract estimation

[grlj Mean deviation Mean deviation Mean deviation RMSE RMSE RMSE
(math. model) (ANNI) (ANN2) (math. model) (ANNl) (ANN2)

Mean 3.67 2.69 2.83 4.42 3.17 2.96

CiE 1.28 1.33 1.50 1.37 1.53 1.70
lO.3 Process Supervision 337

picted in Fig. 10.7 for a typical example. Neural networks proved to deliver better
results than the mathematical model approach. Both networks depict to about a si-
milar quality. Quantitative results obtained with this method from the fermentations
investigated are summarized in Table 1O.l.
The mean values of the deviations were taken over all fermentations investigated.
The corresponding standard deviations O"E characterize the consistency of the ex-
tract estimations. The rms errors show that the neural network models depict a sig-
nificantly lower error as compared to the simple mathematical model.
In order to provide an overview over the individual measurement data, the means
and the rms errors of the predictions of all three estimations were plotted in
Fig. 10.8.

10.3.2.3
Hybrid Modeling

Hybrid models are used either when there are different models available to describe
the behavior of a process or when different parts of the process are described by
different models. The first alternative is of advantage when the different models can
be viewed as descriptions of the same process from different points of view. The
second alternative is proposed for the case where the knowledge about different as-
pects of the process is available at different levels of sophistication. For example, the
basic mass balance equations are known to a high accuracy, while our knowledge of
kinetic details of the biochemical conversion and growth processes involved are at a
much lower level.
The latter case led to the hybrid model, schematically depicted in Fig. 10.9. In ad-
dition to the basic mass balance, which is formulated by means of a set of ordinary
differential equations, the kinetics may be described by means of an artificial neural
network. This type of hybrid model (with respect to the process representation tech-
nique) can be extended by using an alternative representation of the kinetics, e.g., by
classical Monod expressions.
The training of the hybrid model containing the artificial neural network and a
differential equation system was performed by the method of Schubert et al. [10],
since the conventional training methods cannot be used for hybrid models. The data
sets used for training and the subsequent cross validation were the same as used
before for the simple artificial neural networks.

Conversion
1,,0 (S=9Og/1)
tr= t - tr,o
ANN s
Math, Model
log(pE,o/p·) = A - B/(C+Tw~d) CE,1
CE,g ----II- YE

Fig. 10.9. Scheme of the hybrid model for estimating the extract degradation
338 10 Modeling of the Beer Fermentation Process

110,---------------------,
Fennentation M4216 45
- Ethanol in the young beer
100 .. -- Interpol. Saccharometer Values 40
- - - Extract Estimation Model
35
-:
30,!!!
~
25 g
"0
~

20~
W
15

Tirre [h]

Fig. 10.10. Extract estimation with the hybrid model as example of a beer fermentation. In
order to keep the graph simple, the offline measured values were depicted by a line obtained
through a spline interpolation

The rms error obtained with the hybrid model approach was between 1.26 grand
2.52 g 1- for the fermentations investigated. This is significantly smaller than the er-
rors obtained with the other models. where the rms errors were between 2.64 g 1- and
3.17 g r. Thus, the extract concentration can be estimated more consistently with a
hybrid model as demonstrated in Fig. 10.10.

10.3.3
Kalman Filters, and an Advanced Method for State Estimation

The different techniques used for state estimation are distinguished by the quality
and the amount of knowledge they use about the measurement data and their rela-
tionships to the extract degradation. While the model-based indirect measurement
used so far only use very general assumptions about the data and assume that the
relationships are essentially correct, more detailed assumptions about the measure-
ment noise and the modeling error are taken into account in Kalman filters.
Kalman filters assume the dynamic process model to be given by a set of differen-
tial equations in the state variables c, which can be influenced by some control vari-
ables u. The state variables in the case of beer fermentation are the biochemical state
variables, i.e., the concentrations of the species appearing in the basic reaction equa-
tion. As already discussed, the basic control variable in a beer fermentation is the
temperature T. The parameters of the model might be combined to a parameter vec-
tor p. The uncertainties of the measurement values and in the model are provided by
covariance matrices. The Kalman filter is a one-step-ahead prediction method,
which estimates the state vector at the current time instant from the current mea-
surement values and the value computed by the underlying process model. Since the
latter must be based on the estimate made for the last step, one needs a total of three
covariance matrices, since the last value estimated was obviously also not completely
correct.
The Kalman filter is considered the optimal state estimator for a linear system with
predetermined parameters. Unfortunately, it is not directly applicable to fermenta-
tion systems due to their nonlinearity and difficulties in determining model para-
10.4. Process Control 339

meters. To cope with nonlinear models as well, Extended Kalman Filter (EKF) were
developed.
Originally the Extended Kalman Filter was designed to estimate the state of a non-
linear process for which a model is already available (including its parameter va-
lues). However, Jazwinski [26], and later on many other authors, showed that it can
also be applied to estimate the parameters of the process model even when these are
time-varying or ill-defined model parameters are used. Simutis et al. [27] applied
Extended Kalman Filters to estimate the extract degradation in production-scale
beer fermenters. They extended the process model needed as one base of the Kalman
filter by a heuristic component, namely the division of the entire process into several
phases, which were described by different models. The different model components
they handled with a small fuzzy expert system, just as shown for the modeling of the
beer fermentation with fuzzy supported artificial neural networks described above.
The detailed formulae were often presented in the literature [28, 29] and will not
be repeated here, particularly since the application of Kalman filters in industrial
breweries currently seems to be quite unrealistic: Although conceptionally excellent,
extended Kalman filters are known to be difficult to implement, the main problem
being the tuning of the covariance matrices, which are the decisive components of
the underlying concept. In particular, it was shown that the accuracy of process state
estimation by means of Kalman filters is seriously degraded, when the uncertainties
in the process model are not properly known. Thus, its use in control may be limited
to companies which own control specialists that are able to deal with these advanced
techniques.

10.4.
Process Control

Since the temperature is the only variable that can be manipulated during a beer
fermentation process, it is of primary interest to keep it close to a predefined set
point. In order to keep the temperature fluctuations safely within a limited interval
around the temperature set point, it is controlled in all breweries. In practice, simple
temperature controllers, often two-point controllers, are used. With this technique it
is possible to keep the temperature in the production fermenters during the main
fermentation phase within an interval of ±1 K around the set point, which is usually
in the range of 9-13 °C. Here we are speaking about the temperature fluctuations as
measured with a single temperature sensor installed at the fermenter.
One essential point to note is that experimental (e.g., [30]) and model-based in-
vestigations [31] showed that the temperature inhomogeneity, i.e., the spatial varia-
tion of the measured temperatures, is of the same size as the temporal temperature
fluctuation. It was observed that the mean temperatures over a fermenter of several
hundred cubic meters differ by about 1 K only. Hence, with respect to temperature,
beer production fermenters are quite well mixed reactors, a fact which, a posteriori,
justifies the modeling approach discussed before.
Tightening the temperature variances by improved control, however, is a nontri-
vial problem, since production-scale brewery fermenter vessels are very big and thus
not able to follow immediately changes in the temperature set point profiles. In order
to reduce the variance of the temperature fluctuations, a model supported control is
required which can make use of information about the fermenter fluid dynamics,
340 10 Modeling of the Beer Fermentation Process

particularly with respect to the heat transfer into the cooling system and the model
for the heat generation by the biochemical reaction (Eq. 10.2).
Fluid dynamic models of beer fermenters are now coming to the stage where de-
tails about the temperature field can be calculated [31, 32]. However, such calcula-
tions currently need weeks of computing time so that an easier-to-use technique
must be chosen, when dealing with control aspects.
The detailed response in the young beer temperature upon changes in the tank's
cooling systems, i.e., switching on or off the coolant flow through the cooling jack-
ets, is complicated. The stirring action of the COz bubbles rising in the liquid phase
is, as already mentioned, by no means continuous, since considerable eruptions
characterize the carbon dioxide gas flow through the vent line which influence the
heat transfer. Thus supersaturation levels cannot be predicted. The flow in the fer-
menter is chaotic and since it depends on the COz transfer from the liquid into the
gas phase it is not really stationary. The flow of COz through the vent line of the
fermenter and the temperature T at one point in the reactor are the measured quan-
tities in this example.
Since most of the parameters determining the heat transfer coefficients are un-
known, the cooling behavior of the tank must be determined experimentally by ana-
lyzing the temperature response to changes in the cooling control signals. In order to
keep the dynamical representation simple, a linear difference equation was used by
Gvazdaitis et al. [33]. In this equation, the temperature sampling values Tk> where k
is the time index, are related to the registered control signals Sik of the i-th cooling
element:

Tk = ao + al Tk-I + az Tk-z + a3 Tk-3 + a4 Tk_4 + as Tk-s

+ b l SI,k-1 + bz SI,k-Z + b3 SI,k-3 + b4 SI,k-4 + bs SI,k-S
+ b l SZ,k-1 + bz SZ,k-Z + b3 S2,k-3 + b4 SZ,k-4 + bs SZ,k-S (10.14)
+ b l S3,k-1 + b z S3,k-Z + b3 S3,k-3 + b4 S3,k-4 + bs S3,k-S
+ CI Kk
+ Cz Ck

The authors determined the parameters aj' bj, Cj by appropriate identification pro-
cedures. The variables Kk characterize the heat production, while the Ck are charac-
teristic for the COr off-gas mass flow. As a characteristic COz mass flow through the
off-gas line, a smoothed signal (typical average over 2-3 h) of the COz mass flow data
was taken. The characteristic heat production parameter Kk can be determined in
two alternative ways, either from the long term COz development, which is directly
proportional to the heat production, or by directly incorporating the output of the
neural net describing the kinetics of the substrate degradation, as described above.
According to the general reaction(Eqs. 10.1 and 10.2), the heat production was as-
sumed to be proportional to the substrate conversion rate. The identification of the
other parameter was performed with standard optimization procedures.
In order to get an impression of the performance of this model, the temperature
was predicted for periods of 2 h in advance using the available measuring data
sampled up to the actual point in time. These simulated temperatures are depicted
in Fig. 10.11 together with the temperature measurement values obtained 2 h later.
IDA. Process Control 341

.. , -------,-------,-------r-------, tS

., r---~--+-_,~--+-------+-------;

...
Fermentation Time [hI
.'"

Fig. 10.11. Comparison of the simulated temperatures together with the temperatures mea-
sured 2 h later. Both curves cannot really be distinguished. Also, the control profile of the
cooling system is shown in the lower part of the graph (label on the right). Its values 1, 2,
and 3 denote that only the upper cooling section, the upper two only, and all three cooling
sections are switched on

10.4.1
Controllers that Consider the Dynamics of the Fermenters

Temperature predictions over a time horizon of 2 h can be used in a model sup-

ported feedback controller. For this purpose it is straightforward to use a model
predictive controller that compares the predictions with the desired temperature
profiles and changes the cooling control signal in such a way that the deviations
between the expected and the desired temperature profiles become minimal.
Gvazdaitis et al. [33] computed the required control vector for the three cooling
elements of the real fermenter every 8 sampling time steps (12 min each) only, in
order to reduce the controller actions. As already shown in the previous section,
the simple model was able to predict the temperature profile over such time hori-
zons. They used the time between two such predictions to adapt the model para-
meters to the process. Hence, after each 8 time steps, the next control profile segment
was computed with an updated model. The controller used in this example was an
adaptive controller. It is quite easy to determine the optimal temperature control
vectors for the 8 time steps ahead. They can simply be chosen from all possible com-
binations of activating the three cooling sections in such a way that the mean
squared deviation from the set point profile in that interval becomes minimal. In
the case where the actual temperature is below the set point temperature, nothing
is done because it is not possible to heat the fermenter by means of the cooling
system. In this case, one must wait until the internal heat production enhances the
temperature to a value close to the set point.
The most important parameter in the model-supported controller is the adapta-
tion gain by which the deviations .:1T of the actual temperature value from the tem-
perature predicted by the model influence the model parameters used in the control-
ler. If the gain is high, i.e., the influence of the temperature deviations on the model
adaptation is high, the controller might become unstable. If it is too low, the con-
troller might not follow changes in the system quickly enough. Since the response of
342 10 Modeling of the Beer Fermentation Process

the reactor on the actions in the cooling system is rapidly changing during the fer-
mentation, a high gain is desirable.
In order to minimize the risk of obtaining an unstable controller behavior it is
recommended to make use of a second controller with the task of compensating for
larger deviations of the actual temperature from the desired level. This can be done
by a robust fuzzy expert controller based on the available knowledge on the cooling
process. Gvazdaitis et al. used an additional watch-dog module in the control soft-
ware in order to supervise the controller. If this adaptive controller is not able to
control the temperature, the watch-dog switches over to the more robust but not as
accurate fuzzy controller. Figure 10.12 depicts a schematic view of the controller
structure.
An example of the improvement of the temperature signal by means of the con-
troller is shown in Fig. 10.13. It can be seen that there is a significant improvement
over the previous situation.

T.

Fig. 10.12. Schematic view of the temperature controller consisting of the adaptive controller,
the fuzzy controller, and the watch-dog which works as supervisor

'3r----------------------------------,

'2
.,
Q.... ; , \ • .,'''.

.3" 11

"
E
&
~ '0
- - modol prc(lictive control
- - convcnliOl1al control

60 70 80 90 '00

Tim [11]

Fig. 10.13. Comparison of typical temperature signals obtained with conventional temperature
control (upper and lower signals) and with model predictive control (middle)
IDA. Process Control 343

One might argue that the controller controls the temperature at a single point only,
namely at the measurement point. This might be true; however, by the model sup-
ported controller one does not change the mixing behavior of the fermenter, and
thus one does not change the general behavior of the tank. The operational tempera-
ture in a production fermenter is most often measured at one point only and we
know from measurements and from detailed simulations that the inhomogeneity of
the mean temperature values (mean with respect to the time) is less than 1 K during
the main fermentation phase. We can assume that the temperature inhomogeneity
does not become larger with an improved control.
From the experience with beer fermentations we know that temperature fluctua-
tions of ±l °C around the set point do not influence the beer quality. When this
variance is reduced by means of an improved control, we do not change the beer
quality, although the predictability of the extract degradation and its reproducibility
will be better. This might not be a strong enough argument to justify investments for
an improved temperature control. The benefit of an improved controller becomes
more obvious, if one considers its integration into an optimized feedback control
strategy where, as well as the quality arguments, the cooling costs are taken into
account.

10.4.2
Reduction of Energy Costs by Temperature Profile Optimization
and Control in a Production-Scale Brewery

As already mentioned at the beginning, cost reduction is a major aim in process

optimization. Besides manpower, energy consumption is a main target in this re-
spect. Prerequisite to all changes in the process operation and control is that the
product quality goals are assured. This being assured, breweries make significant
efforts to reduce the cost of the energy consumption. For instance, one entry point
to cost reduction is to benefit from the reduced "off-peak" electricity price during
the night. Since electric energy is not conservable in sufficient amounts, some brew-
eries, e.g., Ashai in Japan, installed so-called ice-banks, which are cooled down over-
night and deliver a cooling medium at a sufficiently low temperature level during the
day time (e.g., [34]). From the Guinness brewery in Ireland a similar approach was
published [35).
In order to avoid the significant expenses in investment, maintenance, and ground
space for such an ice-bank, one can control the fermentation process in such a way
that the cooling expenses area mainly required overnight [33). The obvious con-
straint to such an approach is that the temperature must not go outside the known
temperature variation observed in a fermenter that is conventionally controlled,
since otherwise the beer quality would be affected. The basic idea is to first improve
the temperature control so that the temperature fluctuations become significantly
smaller than before. Then the temperature set point can be guided through the al-
lowed temperature interval in such a way that the temperature decreases overnight
and reaches the lower bound of the interval in the morning, when the electric power
becomes more expensive again. Then the temperature is allowed to rise until the
upper level of the interval is reached. From there until the time instant where power
becomes cheaper again the temperature is kept just below the higher bound.
Since the metabolic heat production changes over the course of the fermentation,
a combination of the heat generation and the heat removal processes must be incor-
344 10 Modeling of the Beer Fermentation Process

porated into the model. In order to find an optimal temperature profile, one first of
all needs a measure of optimality.
The following cost functional J was used by Gvazdaitis et al. [33]:

J = kJ(Sf - Sg)2+ (Deviation from extract set point)

k2D2+ (Deviation from diacety set point)
~UkP~t (Cooling cost)

This is a weighted sum of three major contributions to the optimal solution. Its first
component describes the influence of a deviation of the actual extract concentration
from the model value. It says that the process must not deviate too much from the
predefined optimal way. The coefficient kJ is typically chosen to be 0.05. The second
term describes the influence of the diacetyl concentration D. It is assumed that con-
centration D influences the quality loss quadratically, expressing the cost of the lager
process which becomes necessary to reduce the diketone concentration below its
acceptable limits. Parameter k2 is zero as long as the concentration is below some
threshold value (usually 0.2 mg/l). Above that concentration it might be chosen to be
100. The third term sums up the direct cooling cost up to the actual point of time. ~ t
is the time increment. Uk, with k=l, 2, 3, respectively, is a binary control signal in-
dicating whether the individual cooling elements are switched on or off. P=P(t), a
function of time t, is the price of electric power for the cooling aggregate. The mini-
mization of the cost functional can be obtained with classical optimization techni-
ques.
Figure 10.14 shows the structure of the feedback controller schematically. The con-
troller contains an on-line part and an off-line component. In the off-line part, the
optimal control profile is calculated every 24 h in advance. Then the on-line part
controls the temperature along this profile. During this closed loop control, the de-

ON-LINE (a)
T ........ STATE X(t)

Fig. 10.14. Schematic of the optimal feedback controller used to minimize the fermentation
cooling cost in a running production [33]
10.5 Conclusion 345

14 ,-------------------------------------------,

Control Signal for :

the Cooling
6 .... J~ctl

o 50 100 150 200

Fennentation Time [h]

Fig. 10.15. Typical example of the optimized temperature signal in a production fermenter. By
means of a feedback controller the temperature is kept close to the optimal temperature profile
beforehand obtained in an optimization run [32]

viations of the temperature from the predefined profile are used to correct the para-
meters of the model within the controller. Every 24 h, after new extract values be-
came available, the kinetic representation by the neural networks was improved. So,
for the next day the profile can be calculated using an updated model.
Figure 10.15 presents a typical example of a temperature profile measured during
such a controlled experiment. The temperature set point due to the fermentation
scheme used before by the brewmaster was constant at 12 DC. It is shown together
with the ±O.S DC margins within which the temperature was hopefully kept with the
simple conventional controller. The optimized set point profile is a saw-tooth-like
profile where the extreme values were kept within the margins given by the classical
control approach. In the figure the measured temperature is shown which was kept
very close to the set point profile by model supported feedback control. The opti-
mized controller fully exploits the temperature interval allowed during the extract
degradation. As far as possible, cooling actions are performed overnight.
For a single fermenter, the cost reductions may come up to more than 20% in this
way. This is in the same order of magnitude as the energy reduction claimed by
Asahi and Kinoshita [34], who used an unspecified AI computer system in combina-
tion with their cold storage system to maintain an optimized economical balance.

10.5
Conclusion
There is enough knowledge and quantitative data that can be used to optimize the
process operation and to supervise and control the beer fermentation process. The
knowledge about this process is now at a state where the fermentation plants can be
automated in order to reduce the cost of manpower and other operational costs like
electric energy.
346 10 Modeling of the Beer Fermentation Process

Two main problems influence the success. The first touches the question of how to
solve the measurement problem. There are several possible solutions discussed in
the literature, the most promising seeming to be heat removal and ethanol concen-
tration measurements. The second main problem is how to keep the process models
on the actual state. This requires an educational change in the brewers' community
and an adaptation to the developments in other process industries like the pharma-
ceutical industry (e.g., [14]).
In this environment, hybrid modeling provides a chance to make direct use of
model-supported optimization and control in beer fermentation processes, since it
allows one to make use of the wide range of relevant heuristic knowledge of the
process dynamics as well as the extended data records. The alternative would be to
dispense with improvements possible by modern control technology and to make
long term investigations into the biological and transport processes ruling this pro-
cess before starting with improved optimization and automatic control. There is no
doubt that the cost of beer production can be significantly reduced when fermenta-
tion control is automated.

10.5.1
Summary of the Application of the Techniques to Beer Fermentation

Modeling only makes sense when it leads to some obvious advantages. In beer fer-
mentation such advantages would be to simplify achieving improvements of the pro-
duct quality or cost reduction in terms of consumables, energy, or manpower.
The approach taken here is process improvement by optimization of the opera-
tional procedure and making sure through feedback control that the optimal proce-
dure is maintained during the production. This requires tight process supervision.
Systematic optimization of the process operational procedure first of all requires a
clear-cut criterion of optimality. The objective of the process must be defined quan-
titatively in all necessary detail. In beer fermentation the first issue is to produce
beer with the desired quality. Second, this beer must be produced at a minimum of
cost. Traditionally the state of the fermentation process is observed by means of the
actual extract concentration, the operational procedure for a particular beer being
determined by the initial conditions of substrate and yeast concentration as well as
the temperature and the extract degradation profile.
Here we have discussed models designed to provide a sufficiently accurate quanti-
tative description of the beer fermentation process such that they can be applied in
numerical optimization procedures. They are built to represent the knowledge of the
process needed for optimizing the operational procedure (initial conditions as well
as temperature and substrate degradation profiles). Furthermore, models are re-
quired for improved process supervision and feedback control. In order to guarantee
a benefit over trial and error based design methods, the models must meet some
accuracy requirements.
Prediction of the state over a larger time horizon poses the biggest challenge to the
models. Such predictions are required for process optimization in the sense of deter-
mining the optimal process control path. Sufficiently accurate predictions of the path
of the beer fermentation over the entire fermentation time of roughly 5-7 days were
not possible with simple classical models that are often used in biochemical engi-
neering. Hence, extensions had to be made. With hybrid models, e.g., fuzzy sup-
ported artificial neural networks, additional knowledge or a priori information
References 347

about the process was exploited. This allowed one to predict the extract degradation
profile accurately enough. One essential heuristic extension was to divide the process
virtually in several phases, which are separately described by individual models.
This procedure reflects the fact that one speaks of different phases of a process when
its dynamics is dominated by different mechanisms.
Process supervision poses a different requirement to the model. It is not so much
the long term prediction capability of the model that is important, as the detailed
local description. The main problem with process supervision is state estimation.
Obviously supervision requires one to know the actual state of the process as accu-
rately as necessary for the general task to be fulfilled. Such a task could be closed
loop process controlled.
Process control is a measure to make sure that the process runs strictly along the
predetermined control path. Closed loop or feedback controllers are designed to
correct automatically for deviations of the actual process state from the desired one.

References
1. NarziE L (1990) Dber einige Faktoren, die den Biergeschmack beeinflussen. Brauwelt
130:1554-1556, 1558-1562
2. Mitsui S, Shimazu T, Abe I, Kishi S (1991) Simulation and optimization of aeration in beer
fermentation. MBAA Techn Quart 28: 119-122
3. Dymond G (1991) The brewer in control: modern brewery automation. Brewers Guardian
120/3:17-21
4. Havlik I, Dors M, Beil S, Simutis R, Liibbert A (1992) State prediction in production scale
beer fermentation using fuzzy aided Extended Kalman Filters. Dechema Conferences
5b:623-626
5a. Simutis R, Havlik I, Liibbert A (1993) Fuzzy aided neural network for real time state
estimation and process prediction in a production scale beer fermentation. J Biotechnol
27:203-215
5b. Simutis R, Havlik I, Schneider F, Dors M, Liibbert A (1995) Artificial neural networks of
improved reliability for industrial process supervision. In: Munack A, Schiigerl K (eds)
Proceedings of the 6th International Conference Computer Appl in Biotechnology. Else-
vier, Oxford
6. Manikowski M (1996) Modellgestiitzte Dberwachung und Steuerung des Garprozesses in
einer Bierbrauerei, Doctoral Dissertation, University of Hannover
7. Manikowski M, Havlik 1, Simutis R, Liibbert A (1999) Model-supported estimation of the
diacetyl concentration during a production-scale beer fermentation. J Inst Brew (sub-
mitted)
8. Daoud IS, Searle BA (1990) On-line monitoring of brewery fermentation by measurement
of C02 evolution rate. J Inst Brew 96:297-302
9. Austin GD, Watson RWJ, Nordstrom PA, DAmore T (1994) An on-line capacitance bio-
mass monitor and its correlation with viable biomass. MBAA Techn Quarterly 31:85-89
10. Schubert J, Simutis R, Dors M, Havlik I, Liibbert A (1994) Bioprocess optimization and
control: Application of hybrid modelling. J Biotechn 35:51-68
11. Landaud S, Lieben P, Picque D (1998) Quantitative analysis of diacetyl, pentanedione, and
their precursors during beer fermentation by an accurate GC/MS method. J Inst Brew
104:93-99
12. Nakatani K, Takahasho T, Nagami K, Kumada J (1984) Kinetic study of vicinal diketones
in brewing (I): Formation of total vicinal diketones. MBAA Technical Quarterly 21:73-78
13. Nakatani K, Takahasho T, Nagami K, Kumada J (1984) Kinetic study of vicinal diketones
in brewing (II): Theoretical aspect for the formation of total vicinal diketones. MBAA
Technical Quarterly 21:175-183
14. Masschelein CA (1997) A realistic view on the role of research in the brewing industry
today. J Inst Brew 103:103-113
348 10 Modeling of the Beer Fermentation Process

15. Dellweg H (1985) Diacetyl und Bierreifung, Monatsschr Brauwiss 6:262-266

16. Inoue T (1977) Forecasting the vicinal diketone of finished beer. Rept Res Lab Kirin Brew-
ery 20:19-24
17. Rice JF, Helbert JR (1973) The kinetics of diacetyl formation and assimilation during
fermentation. Proc Annual Meeting ASBC, PP 11-17
18. Yamauchi Y, Okamoto T, Murayama H, Kajino K, Amikura T, Hiratsu H, Nagara A, Ka-
miya T, Inoue T (1995) Rapid maturation of beer using an immobilized yeast bioreactor.
1. Heat conversion of a-acetolactate. J Biotechnol 38: 10 1-108
19. Garcia AI, Garcia LA, Diaz,M (1994) Modeling of diacetyl during beer fermentation. J Inst
Brew 100:179-183
20. Pandiella SS, Garda LA, Diaz M, Daoud IS (1995) Monitoring the production of carbon
dioxide during beer fermentation. MBAA Technical Quarterly 32:126-131
21. Doelle HW, Kirk L, Crittenden R, Toh H (1993) Zymomonas mobilis - science and indus-
trial applications. Crit Rev Biotechnol13:57-98
22. Kempe E, Schallenberger W (1983)Measuring and control offermentation processes, part
1. Proc Biochem 18:7-12
23. DeClerk J (1965) Lehrbuch der Brauerei, Bd II Analysenmethoden und Betriebskontrolle.
Versuchs und Lehranstalt fiir Brauerei, Berlin
24. Gmehling J, Onken U, Arlt W (1977) Vapor-liquid equilibrium data collection, vol 1, pt la,
pp 116-155
25. Werbos P (1990) Backpropagation through time: what it does and how to do it? Proc IEEE
78:1550-1560
26. Jazwinski AH (1970) Stochastic process and filtering theory. Academic Press, New York
27. Simutis R, Havlik I, Liibbert A (1992) A fuzzy-supported extended Kalman filter: a new
approach to state estimation and prediction exemplified by alcohol formation in beer
brewing. J Biotechnol 24:211-234
28. Stephanopoulos G, Park S (1992) Bioreactor state estimation. In: K Schiigerl (ed) Biotech-
nology, measuring, modelling, and control, vol 4. VCH, Weinheim, pp 225-249
29. Shioya S (1992) Optimization and control in fed-batch bioreactors. Adv Biochem Eng
Biotechnol 46: 111-142
30. Schuch C (1996) Temperaturverteilung in zylindrokonischen Tanks, Brauwelt 13:594-597
31. Lapin A, Liibbert A (1994) Dynamic simulation of the gas-liquid flows in bubble columns,
demonstrated at the example of industrial-scale beer fermenters; paper presented at the
13th International Symposium on Chemical Reaction Engineering, Baltimore
32. Lapin A, Liibbert A (1999) Fluid dynamics in industrial-scale cylindroconical beer fer-
menters (to be published)
33. Gvazdaitis G, Beil S, Kreibaum U, Simutis R, Havlik I, Dors M, Schneider F, Liibbert A
(1994) Temperature control in fermenters: application of neural nets and feedback control
in breweries. J Inst Brew UK, 100:99-104
34. Asahi K, Kinoshita M (1992) Asahi's new high-tech and energy efficient "Ibaraki" brew-
ery, Technical Quarterly of the Master Brewer's Association of the Americas 29:48-52
35. O'Shea JA (1990) Ice-banks at Guinness Dublin. The Brewer 76:478-480
11 Lactic Acid Production
John Villadsen

11.1
Introduction

Lactic acid bacteria (LAB) comprise a diverse group of microorganisms with many
common traits, but also with a great variation in phylogenetic characteristics. They
are all Gram positive bacteria, and they all produce lactic acid as part of their meta-
bolism, although different genera - even closely related LAB species - have different
yields Ysp oflactic acid (p=HLac) on sugar substrate (s), and widely different rates of
HLac production from the same sugar. One particular species of LAB may even
change its metabolic product pattern quite dramatically when the environment
changes from one sugar concentration to another.
LAB are very well adapted to life in a nutrient which is rich in fermentable sugars
and also rich in nitrogen-containing compounds which can be taken up directly or
after proteolysis and thereafter used as building blocks for biosynthesis. The auxo-
trophy of LAB, i.e., the lack of machinery for de novo synthesis of a majority of the
essential amino acids, is an evolutionary draw-back, and in the nutrient rich media
in which they are forced to survive a great many other organisms would also thrive.
The production of HLac is the main defense action of LAB, and their ability to
acidify the surroundings to a pH of 4 or below where few other microorganisms
survive will give LAB a competitive advantage which outweighs their poor biosyn-
thetic capability. LAB lack the ability to synthesize the heme group and are conse-
quently unable to gain energy by respiration. Escherichia coli has a fully operational
oxidative phosphorylation system with a PIO ratio of 2 and is able to grow with a
maximum specific growth rate of 1.8-2 h -I. Still LAB grow remarkably well, with
specific growth rates in the range 0.7-1 h- I on preferred sugars. This is explained
by their astounding capacity for energy production by substrate level phosphoryla-
tion of the sugar. Not only is the uptake of the preferred sugars high, but the glyco-
lytic machinery is fast and well modulated to handle different environmental situa-
tions.
It is the ability to acidify the environment that has, since the dawn of history,
made LAB such valuable fermenting agents for mankind, inhibiting the growth of
many spoilage and pathogenic bacteria. We use LAB to make a wide variety of fer-
mented dairy products from milk, to protect and to develop desirable sensory qua-
lities of meat products such as salami and sausages, and to produce pickled vegeta-
bles and fruit (olives), soy sauce, sour dough breads, and silage. Detailed accounts of
the use of LAB in the food industry are available in a large number of text books and
review articles of which [1-3] are recent examples.
350 11 Lactic Acid Production 11.3 Sugar Metabolism of LAB

The present study is, however, focused on the production of lactic acid as a bulk
chemical rather than on the use of lactic acid fermentation in the development of
specialized food products. Here the available literature is much more sparse since -
at present - the potential of lactic acid production from agricultural waste or from
byproducts in the food industry has not yet been realized. There are some reports
from academia - some are to be cited in Sects. 11.6 and 11.7 but almost no data from
industrial scale production.
Still, the successful design of a large-scale, fermentation-based process for lactic
acid is critically dependent on the choice of the right LAB species, the right nutri-
ents, and the right operating conditions. Consequently considerable attention must
be paid to these items in order to minimize both the capital costs and the cost of
nutrients.
The following Sects. 11.2-11.5 will therefore give a brief outline of the genomic
organization of LAB, the energy production from sugar - leading to synthesis of
HLac, and the additional nutritional requirements of the bacteria. Finally the at-
tempts to derive expressions which relate the rate of HLac production to the concen-
trations of key substrates will be reviewed.

11.2
Classification of Lactic Bacteria

The majority of organisms used in industry are members of the genera Lactococcus,
Lactobacillus, Streptococcus, and Leuconostoc. The genera Pediococcus and Carnobac-
terium also supply important LAB species with specialized use in the meat industry
and for the production of bacteriocins. Lactococcus lactis has been the most closely
analyzed LAB species in the upsurge of molecular biology since the early 1980s, and
the chromosome and the plasmid content of this group of cocci is well described,
i.e., for the common laboratory strain L. lactis MG1363 and other genetic relatives of
L. lactis subsp. cremoris. The plasmids which serve a wide variety of metabolic pur-
poses [4] are normal components of L. lactis strains - many plasmids are conjec-
tured to be recent additions to the genome of L. lactis which enables these bacteria to
grow well on milk.
The genus Lactobacillus (rod shaped bacteria in contrast to the cocci) contains a
number of important species (Lb. delbrilckii (e.g., subsp. bulgaricus), Lb. helveticus,
Lb. plantarum and Lb. casei) which are used to acidify milk. In general lactobacilli
are more acid tolerant than other LAB [5] and a number of interesting strains can be
chosen from this genus of LAB as potential candidates for bulk lactic acid produc-
tion. The genetics of lactobacilli and their plasmids is described in a number of
recent reviews of which [6] appears to be particularly complete.

11.3
Sugar Metabolism of LAB

As mentioned in the Introduction, the most noteworthy feature of LAB metabolism

is the enormous capacity for degradation of carbohydrates, thereby producing ATP
for biosynthetic purposes. The predominant metabolic product is lactic acid and the
yield Ysp of lactic acid on the sugar can be as high as 95-97%.
11.3 Sugar Metabolism of LAB 351

Most sugars are taken up by lactic acid bacteria either by a PEP (phospho enol
pyruvate) dependent phosphotransferase system (PTS) or via a permease. In the first
case the sugar enters the cell in an already energized form, having been phosphory-
lated at the expense of a PEP molecule during the transmembrane transport process.
In the latter case a kinase mediated phosphorylation takes place in the cytoplasm.

11.3.1
An Example Showing the Functioning of PTS Systems

The best described PTS system is the plasmid encoded, constitutively expressed
mannose (man) -PTS which exists in a number of LAB, including Lactococcus lactis
subsp. cremoris FDI which was studied by Benthin et al. [7-10]. The following ex-
perimental results are compiled from the work of Benthin et al.
The man-PTS has a much higher affinity for glucose than for mannose or in gen-
eral for any other of the sugars or sugar analogues that can be translocated via this
transmembrane protein. In fact it has received its somewhat inappropriate name in
order to distinguish it from a glucose (glc) -PTS which was first studied in E. coli
[11].
Figure ILl shows the metabolism of a pulse of sugar (0.87 g r l fructose+0.87 g I-I
glucose, both in mutarotational equilibrium) added to a glucose limited chemos tat
fed with 7 g r l glucose+ 10 g I-I yeast extract, casein peptone (YECP) at D=O.11 h -I.
Before t=O the glucose concentration in the chemostat was less than 5 mg I-I. At t>O
the flow to the chemos tat was stopped and therefore the time profiles on Fig. ILl
show batch degradation of the pulse. Throughout the experiment the only measur-
able metabolic product was lactic acid.
The figure contains much information concerning the functioning of the man-
PTS:

0.9~~;;;----"""---------1

0.8

0.7
::::-
S 0.6
.,
II)

~ 0.5
:::J
u:- 0.4
:l:
o
g 0.3
a
0.2

0.1

20 40 60 80 100 160 180

Time (min)

Fig.11.1. Metabolism of a pulse offructose (0.87 g I-I) and glucose (0.87 g rl) added to a stea-
dy state glucose limited chemostat (D=O.ll h- I) Feed: 7 gl-I glucose+ 10 g I-I YECP before t=O.
Feed was stopped at t=O. Microorganism L. lactis subsp. cremoris FD1. Simulation of the fruc-
tose profile by Eq. (ILl)
352 11 Lactic Acid Production

- A rapid glucose metabolism immediately starts when the glucose concentration

increases above 10-20mgl- l . Before t=O the rate of glucose conversion is
7xO.ll=0.77 g I-I h- I. For t >0 the rate is more than doubled.
- Assuming a constant biomass concentration of 1.4 g I-I during the pulse, glucose is
metabolized with an almost constant specific rate r s =1.24h- 1 during the pulse. The
corresponding specific lactic acid production is an impressive 1.24 gHLac g bio-
mass-I h- I.
- Fructose must be taken up by the same transport system since no fructose is con-
sumed before the glucose is gone. Thereafter fructose is consumed according to

0.65 (h- I ) Cjru

g fructose g biomass- I h- I (11.1)
rjru = cfru + 160 (mgL-I)
- The corresponding saturation constant for glucose is of the order of 1 mg I-I. The
lines on Fig. l1.1.show simulated results corresponding to the two kinetic expres-
sions.
- If the fructose pulse is given without the corresponding quota of glucose the fruc-
tose consumption after t=30 min on Fig. 11.1 is exactly followed. This confirms
that the two sugars pass through the same transport system - which is blocked
for transport of fructose as long as there is a measurable glucose concentration
in the medium.

In Fig. 11.2 the consumption of a fructose pulse of 0.5 g r I added to a fructose lim-
ited chemostat at D=O.lOh- 1 (feed: 7g1- 1 fructose, lOgl-1 YECP) is shown. For a
biomass concentration of 0.4 g r l one obtains approximately the same very high
specific sugar consumption rate of -1.4h- 1 (and saturation constant of less than
3 mg 1-1) as obtained for glucose in Fig. 11.1. Again the only metabolic product found
is lactic acid. Clearly a high capacity fructose -PTS system has been induced after
prolonged fermentation of fructose without glucose present in the medium.

0.5,....--------------------,

0.4

S
c:::-
0.3
..
lIj
U
2 0.2
IL

0.1

O.O+---,..---,---r--..----r--r--,.---,--,-~_._.....".
o 5 10 15 20 25 30 35 40 45 50 55
Time (min)

Fig. 11.2. Metabolism of a pulse offructose (0.50 g rl) added to a steady state fructose limited
chemostat (D=0.lOh- 1 ). Feed: 7gr 1 fructose, 10gr 1 YECP before t=O. Feed was stopped at
t=O. Microorganism as in Fig. 11.1
11.3 Sugar Metabolism of LAB 353

24

20

--:c
16 ::::ii
.§.
CII

Total Galactose 12 !
c
0
1;
8 :J
"tJ
E
Co
4 "tJ
(;
c(

-4
20 40 60 80 100 120
Time (min)

Fig. 11.3. Metabolism of a lactose pulse (1 g I-I) added to a steady state glucose limited chemo-
stat (D=O.ll h- I ). Feed and microorganism: as in Fig. 11.2. Apr=volumetric HLac production
rate in mmolrlh- I

From Figs. ILl and 11.2 combined we see that

- A LAB (L. lactis subsp. cremoris FD1) can have a constitutively expressed PTS with
a preference order for a range of different sugars. A specific PTS with much higher
affinity for a particular sugar can be induced.
- If the same pulse of fructose is added to a chemos tat operating at the much higher
dilution rate D=0.46 h -I and with the feed of Fig. 11.2 the pulse is consumed much
faster - with a constant specific rate of 2.9 h -I. This shows that the catabolic ma-
chinery - in this case the quantity of or the activity of the fructose transporter and
of the glycolysis pathway enzymes - is much more efficient when the organism has
been grown at high specific growth rate f.L=D.

L. lactis subsp. cremoris also has a constitutively expressed, plasmid encoded PTS for
lactose uptake. Figure 11.3 shows the metabolism of a 1 g I-I lactose pulse added to a
steady state glucose limited chemostat at D=O.l1 h -I (feed 7 g r l glucose, 10 g I-I
YECP). The consumption of lactose starts immediately, showing that the transport
protein is not repressed by glucose nor inhibited by the low (less than 5 mg rl) glu-
cose concentration present before t=O.
The rate of lactose consumption is constant and high - in 27 min the lactose pulse
is consumed. The production oflactic acid (Apr) (20-24 mmoll- I h -I) does not quite
correspond to the rate oflactose consumption (4x6.17=24.7 mmoll- I h- I ). The rea-
son is that a considerable fraction of the galactose moiety of the lactose is excreted to
the medium and is only taken up, and very slowly metabolized long after the glucose
moiety has been metabolized. The rapid drop in acid production rate followed by a
long tail closely mirrors the slow drain from the medium of almost half of the origi-
nal galactose formed. This is the phenomenon of a "secondary acidification" which
is highly undesirable in cheese manufacture.
354 11 Lactic Acid Production

Physiologically the phenomenon is explained by the following uptake mechanism

[10, 12, 13]:

(Lac)ex + (PEP)in -t (Lac - 6P)in + (PYR)in (11.2)

(Lac - 6P)in -t (Gal- 6P)in + (Glc)in (11.3)

(Gal- 6P)in +-+ (Gal)ex + P* (11.4 )

(Gal- 6P)in -t-t (Tag - 1, 6P) +-+ (GAP)in (U.S)

(GAP) in +-+ (Fru - 1, 6P)in +-+ (Fru - 6P) ( 11.6)

(GAP) in -t HLac + byproducts + 2 ATP (11.7)

The glucose moiety formed in Eq. (11.3) is rapidly phosphorylated by glucokinase
and enters the Glc-6P, Fru-6P pool to be used for HLac synthesis and for synthesis of
biomass precursors in the Pentose Phosphate pathway. Gal-6P is metabolised in the
Tagatose-6 phosphate pathway of Eq. (11.5). The resulting glyceraldehyde-3 phos-
phate (GAP) can either be used as a gluconeogenic substrate, entering the Fru-6P,
Glc-6P pool via fructose 1,6 diphosphate, Eq. (11.6), or it can be metabolized to lac-
tic acid. Apparently the enzymes of the Tagatose-6 phosphate pathway are inefficient
in L. lactis subsp. cremoris FD1, and Gal-6P accumulates to unacceptably high levels
and is secreted out of the cell, donating the activated phosphate group to glucose.
The inefficiency of the Tagatose-6 P pathway is also known for other starter cul-
tures, especially for thermophilic LAB [14, 15], while the inability to dephosphorylate
Fru 1,6 P to Fru-6 P in Eq. (11.6) (due to lack of an FDP-ase) is a specific trait of the
FD1 strain used in [10].

11.3.2
Sugar Uptake by LAB in General

The experimental results shown for L. lactis subsp. cremoris FD1 can certainly not be
generalized to all LAB - not even to strains which are closely related to it. But the
description in some detail of the sugar uptake by one particular strain does show the
great variation in uptake pattern for different sugars.
Most industrial starter cultures (at least from the L. lactis genus) have plasmid
encoded constitutive PTS systems for glucose, lactose, mannose, and sometimes also
for sucrose and galactose, although these PTS systems are likely to be inducible and
severely glucose repressed. This is what makes the starter cultures such efficient cat-
alysts for converting sugar to lactic acid.
However, an equally common way to take up lactose in LAB is by means of a
permease followed by cleavage of the non-phosphorylated (Lac)in by ~-galactosidase
(~-gal as opposed to the enzyme ~-gal P used to cleave Lac-6P in Eq. 11.3). Glucose
and galactose can also be transported to the cell via permeases. Glucose is phos-
phorylated to Glc-6P, while galactose is phosphorylated to Gal-1P which is readily
converted to Glc-6P in the Leloir pathway (in contrast to Gal-6 P which is formed
when galactose enters via a specific PTS system or as a result of cleavage of Lac-6 P).
11.3 Sugar Metabolism of LAB 355

Some strains have both transport systems which then serve as a high affinity (the
PTS) and a low affinity (permease) system. Other strains have only one transport
system for a given sugar.
Thus L. lactis subsp. cremoris FDI studied in [7-10] has no lactose permease (no
activity of ~-gal is detectable) and no galactose permease, but has PTS systems for
both sugars. Consequently it cannot grow on galactose since it lacks an FDP-ase, but
if a small amount of lactose is fed together with galactose, sufficient precursors for
biosynthesis are formed from the glucose moiety of lactose to support growth at the
same specific rate as if fed by lactose alone - an indication that energy generation is
more critical than precursor synthesis.
In contrast L. lactis subsp. cremoris MG 1363, a laboratory strain used for research
purposes by a large number of research groups, lacks a lactose PTS and therefore it
grows much more slowly on lactose than the related strain FD1.
The two transport systems for fructose found in L. lactis subsp. cremoris FDI gives
rise to an interesting phenomenon [9]. The mannose-PTS has very low affinity for
fructose, but translocates fructose to Fru-6P. The dedicated fructose-PTS yields Fru-
IP which can be phosphorylated to Fru 1,6 diphosphate, and thereafter it enters
glycolysis, but the absence of an FDP-ase makes Fru-lP useless for synthesis of bio-
mass precursors. The flow through the fructose transporter furthermore disturbs the
functioning of the mannose PTS. The combined action of the two transporters is [9]
that the biomass concentration in a steady state chemostat increases with increasing
dilution rate, a very atypical feature. It also gives rise to instabilities, and apparently
stable oscillations with periods of 5-20 h are observed in both HLac production rate
and in biomass concentration.

11.3.3
Homolactic vs Heterolactic Fermentation

LAB from the Leuconostoc genus and some Lactobacilli metabolise Glc-6P through a
variant of the Pentose phosphate pathway. At the level of Xylulose-5 phosphate the
enzyme phosphoketolase splits the pentose into a C3 (glyceraldehyde-3P) and a C2
(acetylphosphate) compound. The end result is generation of 1 mol each of HLac and
ethanol, together with 1 ATP from one mol of glucose. If these LAB have an NADH
oxidase activity they may under aerobic conditions produce acetic acid (HAc) in-
stead of ethanol. In either case these strictly heterofermentative LAB, which own
their special fermentation pattern to the lack of an FDP-aldolase to split Fru-l,6 P
into two C3 compounds, will be of no interest in the production of lactic acid since
half the carbon is lost to CO 2 and to a C2 compound.
A vast majority of LAB including Lactococci and most Lactobacilli ferment the
sugars through the EMP pathway. These are the homofermentative LAB, and they
generate 2 ATP and a maximum of 2 HLac per metabolized hexose molecule.
The main features of the EMP pathway from glucose to pyruvate, and further to
HLac and the most common byproducts, are shown in Fig. 11.4.
The linear pathway from glucose or any of the other fermentable sugars to pyru-
vate is common for all the homolactic LAB and is traversed in the same fashion for
low and for high sugar flux. For a high sugar flux (and this corresponds to sugar
levels higher than about 20-50 mg tl in the exterior medium) the phosphoenol-pyr-
uvate (PEP) pool is very small while the fructose 1,6 diphosphate (FDP) pool is
356 11 Lactic Acid Production

PTS
system

Fig. 11.4. Schematic of sugar metabolism in LAB: 1: PK=Pyruvate kinase, 2: LDH=Lactate de-
hydrogenase, 3: PFL=Pyruvate formate lyase, 4: PDH=Pyruvate dehydrogenase

large. The pool of triosephosphates (GAP and DHAP) is also large, but at all condi-
tions much smaller than the FDP pool.
At low sugar flux through the pathway the PEP pool increases in size while the
FDP pool, and in particular the pool of triosephosphates, diminishes. The change
between "high" and "low" flux can occur quite abruptly, e.g., somewhere between
20 mg r l and 5 mg rl glucose in the medium. For sugars which are taken up by per-
meases rather than by the highly efficient PTS systems the condition of "low" flux is,
however, obtained at much higher levels of sugar in the exterior medium.
The metabolism of pyruvate is highly dependent on whether the sugar flux
through the EMP is "high" or "low:' At "high" flux only lactic acid is formed; at
"low" flux a considerable amount of formic acid (anaerobic conditions) or CO 2
(aerobic conditions) is formed in one of the two branches leading from pyruvate to
Acetyl coenzyme A (AcCoA). When the PDH catalysed pathway is inactive (under
anaerobic conditions) and the sugar flux is low, perhaps 50-60% of the sugar is
converted to formic acid, ethanol, and acetic acid in the ratio 2:1:1:
(11.8)
When the flux through the EMP pathway increases there is an immediate reloca-
tion of fluxes at the pyruvate branchpoint. This is clearly seen in Fig. 11.5 which
shows the development of the concentration of lactic acid and formic acid after the
dilution rate in an anaerobic chemostat fed with 7 g rl glucose, 10 g r l YECP has
been changed from 0.075 h -I to 0.409 h -I. The strain is L. lactis subsp. cremoris
FDl. At the low dilution rate the glucose concentration (measured by FIA [16]) is
between 1 mg 1-1 and 5 mg r l , after the transient the steady state glucose concentra-
tion at D=0.409 h -I is 8 mg rl. The remarkable change in product formation pattern
is of course not directly a result of this small change in exterior glucose concentra-
tion, but rather of the considerable change in glucose flux through the EMP pathway
which results when the dilution rate is increased.
11.3 Sugar Metabolism of LAB 357

6..-------------------.2.0

1.5
<>
g 4
{t Formic acid
OJ

..
-ri 3
U
<>
1.0 gj
8
..
U
..J
2
lactic acid
iii

or 0.5
co
0
<>
a" Glucose
01.j.......-'!""''"'!'''-~_r__:r:___r__::2II!'_~~~!i!"qO.O
0 2 3 4 5 6 7 8 9 10 11 12 13
Time (h)

Fig. 11.S. Profiles of metabolic products after a change of dilution rate (from D=O.075 h- I to
D=OA09h- I ). Feed and microorganism as in Fig. ILl

Before the change of dilution rate, 2.1 g HLac 1-1 and 3.2 g HCOOH 1-1 is produced
from the feed. Since the production of formic acid as measured by its concentration
is 20% higher than what would be calculated by Eq. (11.8) from the glucose feed,
7 g t 1 corrected for the produced lactic acid some part of the YECP must also be
catabolized, but this is not important in the present context. The key observation in
Fig. ll.s is that the formic acid production immediately stops when the flux through
the EMP pathway increases, and that production of HLac immediately increases. The
two lines are simulations of this situation. The increasing flux cannot be seen di-
rectly, but at the end of the transient more than five times as much glucose is meta-
bolized by approximately the same amount of biomass (1.6 gt 1 ) and the hump of
glucose extending for about 5 h and with more than 500 mg t 1 glucose shows that
the fermentation is in no way glucose limited until the excess glucose has been me-
tabolized.
Figure 11.6 gives additional information related to the shift in metabolism. The
specific growth rate JL is of course equal to D both before the transient and after
the transient. In between, JL is first seen to be smaller than D=OA09 h -1 but for al-
most 3h JL is larger than OA09h- 1 • The first part of the JL (t) curve is explained by a
higher catabolism, i.e., a larger generation of ATP than corresponding to the ATP
need for anabolism. The overshoot of JL is an effect of the large enzymatic, including
biosynthetic, machinery produced during the long exposure of the culture to a rich
glucose medium.
Of particular interest is the jump in r p' the specific rate of lactic acid production
occurring as soon as the flux increases. This is another sign that the machinery for
HLac production, i.e., the high activity of the LDH is immediately available as soon
as the glucose flux through the EMP pathway increases.
Another interesting feature of Fig. 11.6 is the slow increase of the catabolic ma-
chinery which gives rise to an almost doubling of rp over 5 h. Also shown in the
figure.is the RNA content of the cell as measured by a very precise method [17].
The increase of rp faithfully follows the increase of RNA, i.e., of cell activity in gen-
eral as the cells become acclimatized to the high sugar concentration during the
358 11 Lactic Acid Production

1.5~-------------------r 12

{
6 ~
<
Z
a:

Glucose
o.otO;;:;;:~2=~3;::~4~5"--""'6""'-""'7-"'8~:~~10;;;11;;1"'2J130
Time (h)

Fig. 11.6. RNA, glucose profile and specific rates of production of HLac (rp) and biomass (fh)
after the shift in dilution rate as Fig. 11.5

transient. When the sugar concentration again drops below 20-30 mg 1-1 rp abruptly
decreases and the RNA level slowly settles on the value corresponding to
D=00409h- l . Again it is remarkable that rp is three times as high at the end of the
transient than at its start - and this in spite of the seemingly small difference in
extracellular glucose level from 1-5 mg I-I to 8 mg I-I.
For a long time the dramatic change in product formation pattern was ascribed to
the allosteric regulations shown in Fig. 1104.
In vitro experiments demonstrated that FDP acts as an allosteric activator of LDH,
and consequently it was argued that lactic acid production was enhanced by the high
FDP level which was found at high glucose flux through the EMP pathway. Recently,
however, it has been shown that the FDP level can have only a minor effect on the
product formation pattern, since with the in vivo concentrations found by more ac-
curate assays [18] the FDP concentration is probably always high enough to saturate
the LDH.
It is more likely that the inhibition of PFL activity by the pool of triose phosphates
GAP and DHAP is responsible for the mixed acid fermentation at low sugar flux
through the EMP pathway since the level of this metabolite pool is shown to decrease
appreciably under these conditions. With the focus on GAPIDHAP rather than on
FDP, a dominating influence of the NADH/NAD+ ratio has recently been postulated
[19, 20] and strongly supported by measurements of the ratio [20] at different
growth conditions. First, the activity of LDH is very sensitive to the NADH/NAD+
ratio, and by in vitro experiments the enzyme is found to be totally inhibited when
the ratio has a value lower than 0.03. Also an almost linear relation was found be-
tween the value of NADH/NAD+ and the specific sugar (glucose, galactose, or lac-
tose) consumption rate. For r s =19mmolg- I h- 1 the ratio was 0.08 and the LDH was
fully activated. Second, a low value of NADH/NAD+ will speed up the oxidation of
GAP since this reaction has NAD+ as cofactor. This will lead to a small pool of trio-
sephosphates and the negative allosteric regulation of PFL is alleviated.
The details of the regulatory system which determines whether almost 100% of the
sugar is fermented to HLac or whether a substantial portion is directed towards by-
1l.4 Nitrogen Uptake and Metabolism 359

products may not yet have been settled, but as outlined above the main mechanisms
are known. Qualitatively speaking this knowledge may be used to plan an industrial
production of lactic acid by fermentation.
It is desirable to have a high sugar flux through the central metabolism of the LAB.
Transport of sugars to the cell by PTS systems gives a higher flux than transport via
permease systems, and one might consider importing a lac-PTS to strains (mostly of
plantal origin) which normally lack this transport system, typical of starter cultures
from the dairy industry. One may also consider an overexpression of genes encoding
for the enzymes in the EMP pathway and for LDH. This is less likely to give a notice-
able advantage due to the very tight regulatory control of the enzymes in the central
metabolism.
In any case very low sugar concentrations in the medium must be avoided since
this will necessarily lead to activation of the PFL or the PDH enzymes - a very useful
option for the organism which will gain an extra ATP (i.e., a 50% increase in the ATP
produced by metabolism of sugar) in the energy limited situation at low specific
growth rate, but hardly an advantage for the production of lactic acid. In this context
it should be remembered that a "low" sugar concentration in the medium is 20 mg 1-1
or lower for sugars transported by a PTS. These transport systems are incredibly
efficient, and they will furthermore spring into action as soon as the sugar concen-
tration in the environment increases - at low flux through the glycolysis the PEP
level is high due to inactivation of PK, and the uptake of sugar through the PTS
can start immediately.

11.4
Nitrogen Uptake and Metabolism

In contrast to yeast and to E. coli, LAB lack the ability to synthesize most amino
acids from ammonia or another inorganic N-source. The requirements for amino
acids varies between species and it is difficult to predict which amino acids must
be added. Thus some strains of L. lactis subsp. lactis are prototrophic for most ami-
no acids whereas most L. lactis subsp. cremoris strains have requirements for about
15 amino acids, and almost all strains grow slowly on a chemically defined medium
compared to growth on a complex, ill defined nitrogen source like YECP from which
the proteolytic systems of the LAB can secure an ample supply of amino acids.
Even when an apparently rich N-source has been used, the LAB may experience a
deficit in one or more amino acids during a batch fermentation. Incorrectly inter-
preted this may easily lead to the erroneous conclusion that the culture is sugar (en-
ergy) limited.
Figure 11.7 shows the progress of a batch fermentation with initially 18 g r l glu-
cose and 20 g 1-1 YECP inoculated with 1 mg 1-1 L. lactis subsp. cremoris FD1. At first
the culture grows with a specific growth rate of 1-"=0.82 h -\ but after 8 h and at a
biomass concentration of 0.3 g 1-1 the growth rate abruptly decreases, probably due
to depletion of some component in the N-source.
In Fig. 11.8 the change in specific growth rate is seen to be closely correlated with a
change in the specific acid production rate rp. No trace of byproducts can be de-
tected in either growth period and the lactic acid produced corresponds closely to
the glucose metabolized (Fig. 11.7, accounting for the volume change due to addition
of 2 mol 1-1 NaOH to keep pH constant at 6.2). It appears that after a while both I-"
360 11 Lactic Acid Production

10.00
Lactic acid
a
15 a
Biomass
~ 12 1.00
"C
U
.. c
S
.
u co
'fi 9 co
+ co
-' ; E
.,- ./
0
iii
g
::0
6
........
0.10
(5
3 .or
....~
~~ •
• Glucose
•
0.JaOIllm"""=WIJ!III!!!!!~~~--:-:----:>:--';·~~-~0.01
2 4 6 8 10 12 14 16 1B
TIme (h)

Fig. 11.7. Batch fermentation with a rich N-source (lOgl-1 glucose and 20gl- 1 YECP) L. lactis
subsp. cremoris FDI

7.-----------------------~r_----_r 21

....
....
'"
....
i-
6+-----------------=-----__+
....
....
....
st---------------------.r--------+
18

15 0
:c
.,
.
..
Apr/3.1

.
....
:c ~.... ~'" ....
C 4+-----~~-----~~~~~~-__+ 12 "":
E
.§.
.. .
,. .!3
t" Co
RNA
~ 3+------~~~--~--~~~__+ 9 ca:
e-
2+---------~----~~----~~----_+ 6
rp I?;.
ca:
zII:
3

0
2 4 6 8 10 12 14 16 18
Time (h)

Fig. 11.8. Specific (rp) and volumetric (Apr) production rate of HLac during the batch of
Fig. 11.7. Also shown: specific growth rate J-l and RNA profile

and r p pick up again and J-l reaches an approximately constant value of 0.4 h -I until
the sugar is depleted and both rp and J-l abruptly drop to zero.
In Fig. 11.9 rp is plotted as a function of J-l. The two specific production rates are
clearly linearly dependent throughout the batch fermentation:
rp = 1.75 + 2.24J-l (ggbiomass-1h- 1) (11.9)
or
rp = kATP + YxATP . J-l = 19 + 25 J-l (m mole g biomass- 1h- 1) (11.10)
Since in homolactic fermentation the rate of lactic acid production corresponds
exactly to the rate of ATP generation (inmmolh- 1 ) the relation at Eq.(ll.lO) also
11.4 Nitrogen Uptake and Metabolism 361

4,---------------------------------,

O+---~--~--~--r_--~--~--r_--.-~
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
u (1/h)

Fig. 11.9. Cross-plot of fL and rp from Fig. 11.8

expresses the specific rate of ATP formation needed to support growth. In repeated
fermentations (results not shown) with the same feed, somewhat lower values of the
parameters of Eq. (ILl 0) are found. An average for three fermentations gives
YxATP =17 mmol g biomass- I and kATP=21 mmolg- I h- I with an estimated RSD of 6%
and 11 % respectively.
An ATP consumption of 17 mmol g biomass -I is far below the theoretical value 37-
40mmolgbiomass- 1 for biosynthesis of bacteria [21] from a minimal medium.
Considering that LAB have a low protein content (45 wt %) compared to that of
many other bacteria (60 wt %), and that some of the building blocks can be taken up
directly from the yeast extract-peptone medium, one may calculate a smaller lower
bound of 26 mmolg-I for the theoretical ATP requirement YxATP , but this is still
much smaller than 17mmolg- l .
The reason for the ATP requirement being apparently below the theoretical limit
must be that the cell gains energy concomitant with the uptake ofN-containing com-
pounds from the medium.
In Figs. 11.1 0 and 11.11 this assumption is verified experimentally using less N-
source than in the previous batch experiment. In Figs. 11.10 and 11.11 the feed com-
position is 20 g I-I glucose and 7 g tl YECP.
For f.L>0.3 h- I the energetic parameters YxATP =15 mmolg- I andkATP =18 mmolg- I h- I
correspond closely to the parameters determined from Fig. 11.9. For f.L<O.3h-I, YATP
is 50 mmol g-I and the non-growth related maintenance consumption kATP is only
7 mmolgh- I.
In Fig. 11.11 the net consumption of primary amino groups (AN) is shown to-
gether with the total consumed N and the specific growth rate - all as functions of
the biomass concentration during the batch fermentation.
The net rate of AN consumption is the difference between the AN groups con-
sumed for synthetic purposes (or metabolized) and the AN groups formed by hydro-
lysis of peptides.
In the first part of the fermentation until f.L first reaches the value 0.3 h- I the pro-
duction and consumption of AN balance. There is a constant but small net consump-
362 11 Lactic Acid Production

3T--------------------------r------------~

O+----r----.---~--_,----~--~----r_--~
M ~ U ~ U U MUM
u (1/h)

Fig. 11.10. Cross-plot of Jt and rp for a batch fermentation with less N-source than in Fig. Il.7:
20gl- 1 glucose and 7gr 1 YECP at t=O

0.18 5

0.15 4

~
!!! 0.12 3 ~
"'- .§.
z
"""
0;
;; 0.09 2 ..
<I:
."

..
."
E
0.06
..'"
E
c
'"
til
C
0
CJ
0
CJ
0.03 0

0.00 -1
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2
Biomass (g/~

Fig. 11.11. Consumed total nitrogen (N, gl-l) and primary amino groups (AN, mmoll- 1 ) dur-
ing the batch fermentation of Fig. ILl o. A cross-plot of Jt and the biomass concentration is
also shown

tion of AN during the next part of the fermentation where JL increases to 0.45 h -I and
again decreases to 0.3 h -I. Throughout the first two periods there is a constant yield
coefficient YxN of O.l13ggbiomass- l • In the last part of the fermentation
YxN =0.018ggbiomass- 1 is so small that normal growth with cell division is severely
inhibited. Rather the cells grow in mass, but with decreasing N content. The high
rate of AN consumption together with the small rate of total N consumption is best
explained by a considerable cell lysis which takes place at the same time as cells with
a different composition are formed at an extra expense in ATP derived by glycolysis
of glucose to lactic acid.
11.5 Growth Kinetics and Product Formation Kinetics 363

In [22] the growth energetics of Lactococcus cremoris FD1 is analyzed for condi-
tions of energy-, carbon-, and N-limitation respectively. Amongst other hypotheses
tested, it appears that import of Cr C4 pep tides from the YECP medium and excre-
tion to the medium of some of the amino acids released by hydrolysis of the pep tides
in the cytoplasm during the first phase of a batch cultivation is the explanation for
the very low measured YxATP Import of a peptide costs only 1 ATP while export of
each of the amino acids from the hydrolyzed peptide gives rise to the production of
1 ATP.
The unknown amounts of trace compounds of possible stimulating effect and the
general variability of any complex nitrogen source such as the yeast extract, casein
peptone medium used in [22] makes it attractive to study the nitrogen metabolism
with a feed which contains known quantities of each of the 20 "essential" amino
acids. The composition of one particular "defined" medium is given in [23]. When
the concentration of each of the added amino acids is followed during a batch fer-
mentation or at different steady states in a chemos tat, a very detailed picture of the
uptake and metabolism of amino acids by a given LAB strain is obtained. Thus in a
study of L. lactis subsp. cremoris MG l363 it was found [24] that the organism de-
graded arginine almost quantitatively to ornithine, gaining ATP in the process. If the
arginine present in the feed was lower (but still much higher than required for bio-
synthesis) the organism became heterofermentative. An increasing serine concentra-
tion (a non-essential amino acid for L. lactis) led to a drastic change in product
distribution. With 11.6 mmoll- 1 serine in the feed the catabolism of sugar changed
from strictly homofermentative (with no serine in the feed) to almost entirely het-
erofermentative with YSHLac=0.2gg-t, the remainder being the 2:1:1 mixture of by-
products shown in Eq. (11.8). Many of the phenomena observed up to now have not
been adequately explained by reference to well defined metabolic reactions, but ob-
viously the study of LAB fermentation with a defined medium can give much insight
into the formation of a desirable flavor spectrum of the final product.
It must, however, be realized that no defined medium is likely to contain all the
"growth factors:' "stimulating species:' etc., which in general make LAB grow better
on a complex medium than on a defined medium.

11.S
Growth Kinetics and Product Formation Kinetics

The purpose of setting up a kinetic model for any chemical reaction is to predict the
selectivity for different products and to find the rate of consumption of the main
reactant at reaction conditions outside the experimental range used to construct
the kinetic model. This type of research has proved highly successful in the petro-
chemical and related industries, and the rate expressions derived for typical gas
phase catalytic reactions can often be used at much higher pressures and far from
the temperature range of the original experiments.
Similar successes are not to be expected in fermentation studies. In contrast to the
simple reaction scheme which describes, e.g., the formation of NH3 from N2 and H2
with only one final product and a limited number of plausible mechanistic models
for the surface reactions on the metal oxides used as catalysts, the process of con-
verting sugar and other substrates to biomass and metabolic products is far too
complex to be described by a single overall kinetic model.
364 11 Lactic Acid Production

For all of the industrially important microbial systems - yeast, E. coli, LAB, Bacil-
lus subtilis, Penicillium chrysogenum, etc.- many attempts have been made to set up
kinetic models which could be used in the design of fermentors, as the basis for
process control of fermentation processes or to gain insight into the metabolism of
the microorganism. Kinetic models may be of some help in a crude fermentor de-
sign, and if they are sufficiently equipped with results from different fermentation
conditions they can be used in a computer run expert system for control purposes.
Where they fail is in scientific studies of metabolic behavior.
The preceding sections have been illustrated with a number of experimental results of
relevance for lactic acid fermentations. When considered together these results should
make it abundantly clear why any overall kinetic model for LAB fermentation must fail
if used to extrapolate beyond the range of the original experiments. This pessimistic
forecast is not a peculiar feature of LAB which in comparison with other microorgan-
isms have a relatively simple metabolism. The prediction of the behavior of eukaryotic
microorganisms - e.g., in an attempt to calculate by a general model the production of
proteins from yeast or mammalian cells at conditions somewhat different from those of
the data basis used to construct the model- would probably fail even more miserably.
An overall kinetic model used to predict the specific rate of lactic acid production
rHLac (g g biomass- I h- I) or volumetric production rate qHLac (g 1medium- I h- I)
would fail for the following reasons:
- Different species of LAB have different substrate requirements and different meta-
bolic machinery. Therefore the influence of environmental parameters may be very
different, both concerning the metabolic rates and, in particular concerning the
product formation pattern.
- Experiments made with the same strain of LAB give different results depending on
the time scale of the experiment. Chemostat experiments such as those shown in
Fig. 11.6 can be used to construct relations between specific growth rate j1, the
specific acid production rate rp or the RNA content, and the concentration of the
limiting substrate (glucose) in the steady state. These rate expressions cannot be
used to predict what happens during a transient. The immediate jump in rp=rHLac
in Fig. 11.6 followed by a slow increase over 5 h are features which are not even
mentioned in the steady state model, nor is the abrupt shift (Fig. 11.5) in the rate
of formic acid production from 3.2 x 0.075 g I-I h -I to zero as soon as the dilution
rate is changed to D=OA09 h -I.
- Also in experiments with the same strain the influence of one sugar on the uptake
of another sugar, or the combined effect of uptake of one sugar through several
membrane transport systems (Figs. lLl-l1.3) is as yet difficult to predict. The
same holds for the influence of the N-source on the batch kinetics, Fig. 11.7. The
specific growth rate decreases by a factor of 2 at a biomass concentration which
will depend not only on the ratio of carbon to nitrogen source, but also on the
composition of the nitrogen-containing substrate. This composition is very depen-
dent on the origin of this substrate (yeast extract, casein peptone, hydrolysed whey
protein, etc.). Hence the yield coefficients Ysx (gbiomass g sugar-I) and Ysp
(g HLac g sugar-I) as well as the non-growth related maintenance requirement k ATP
are heavily dependent on the nitrogen source as seen in Figs. 11.9 and ILl O. They
are also dependent on the sugar: in the experiment of Fig. 11.2 Ysx is practically
zero since the strain cannot grow on fructose admitted through the fru-PTS,
whereas Ysx can be high if a little glucose is used in the feed and the transport of
both sugars occurs through the man-PTS.
ll.S Growth Kinetics and Product Formation Kinetics 365

The value of kinetic models as data fitters for a large amount of experimental data
should, however, not be underrated. In this sense the mathematical model can cer-
tainly be used for design purposes when the desired operating conditions are with-
in the parameter space spanned by the experimental data. Each detail of the phy-
siological map of the organism can of course also be modeled quite faithfully, and
mechanistic features of the metabolism can be discovered by mathematical model-
ing of carefully designed experiments. Thus the fructose uptake in Fig. ILl is very
well described by Eq. (1Ll) as seen from the simulation (full line on figure), the
decrease of formic acid concentration on Fig. 11.5 is adequately modeled by the
assumption that the formic acid production immediately stops after the shift in D,
and the stepwise uptake of the N-source is understood by the simulations in
Fig. lLl1. An intuitively appealing explanation for the existence of a double
stranded man-PTS in contrast to the single trans-membrane strand observed for
other PTS is the result of a study [7, 8] of the transport of a-d- and ~-d-glucose
to L. lactis subsp. cremoris FD1.
An overall model for growth and product formation should be simple and should
account only for the most basic features of the growth process. In this case it can
express in a compact form our empirical knowledge of the process without unwar-
ranted claims of exactness.
Practically all LAB will in some sense be inhibited when the lactate concentration
increases. In non-pH-controlled cultures an inhibiting effect of the acidification of
the medium during the process is also observed.
Consequently a model of the form

qx = JL x = -JLmax
-- s ( 1 - -P- ) x (ILll)
s + K, Pmax

for the volumetric rate qx of biomass production has some value. s is the concentra-
tion of the limiting substrate and p is the concentration of HLac. At P>Pmax~50 g r l
the fermentation stops and the cells lyse.
The trouble with Eq. (1Ll1) is that the growth limiting substrate may be different
in different phases of the fermentation. Consequently the value of JLmax may change
during a batch fermentation as seen in Fig. 11.7. If the limiting substrate is assumed
to be the sugar it is evident from the results both of pulse experiments and of steady
state experiments that Ks is extremely low, of the order of 20 mg I-lor less, and the
first factor of Eq. (1Ll1) is independent of s throughout the batch experiment. The
difficulty of assigning a reasonably constant value to JLmax for growth on different
sugars in N-rich medium has also been touched upon: if galactose or fructose is
not able to enter he Fru-6P/Glc-6P pool after uptake JLmax will be practically zero
for these sugars due to a failure to produce biomass precursors.
The value of Pmax varies considerably for different species of LAB which in
itself makes Eq. (ILl 1) difficult to use in general. Even the concept of a reduced
specific growth rate at higher lactic acid concentration has been questioned. Thus
in [25] it is concluded that the inhibition of HLac on growth of Lb. delbrukii
subsp. bulgaricus at constant pH is primarily seen as an extremely prolonged lag
phase increasing from 2 h to 39 h when the concentration of HLac added at the
beginning of the batch increases from 0 g rl to 25 g I-I and a reduced maximum
cell yield. There is virtually no effect of added HLac on the specific growth rate
once the batch fermentation takes off. The effect of I-lactic acid is also much
366 11 Lactic Acid Production

more pronounced than the effect of d-lactic acid, the isomer naturally produced
by the organism. Clearly none of these effects are even suggested by the rate
equation (Eq.11.11).
In a classical study of lactic fermentation by Lb. delbriikii, Luedeking and Piret
[26] modeled the product formation kinetics as the sum of a growth associated and
a non-growth associated (maintenance) term

(1l.l2)

For any fermentation where product formation is a result of the catabolic me-
tabolism leading to ATP formation it is expected that the product formation
kinetics has the structure indicated in Eq. (11.12). In Fig.11.9 and Eq. (11.10)
we have seen that the linear relation between rp and JL holds true throughout a
batch fermentation with excess nitrogen source. Unfortunately the two para-
meters of Eq. (11.12) are heavily dependent on the fermentation conditions. In
the original paper on batch fermentations at constant pH Luedeking and Pi ret
[26] had already noted that Yxp increases by a factor of 2 in the narrow pH
range from 5.2 to 4.5 whereas the yield coefficient was constant between pH 6
and 5.4. In the same pH range the maintenance term decreases by a factor of
2.5 - one may assume that the microorganism down-regulates its ATP expendi-
ture in futile cycles etc. when at low pH the growth becomes more energy de-
manding.
The results in Fig. 11.10 for a somewhat N-limited batch fermentation show that
the nitrogen source also has a significant influence on the parameters of Eq. (11.12),
and the great variability of the two parameters when using different sugars or mix-
tures of sugars [22] gives further evidence to prove that Eq. (11.12) can only be used
with extreme caution for predictive purposes.
In recent years a number of structured kinetic models have been proposed. The
aim of these models has been to predict on the basis of steady state experiments
certain features such as the length of the lag phase in a batch fermentation, the
course of a wash-out experiment, and the diauxie observed in a mixed sugar fer-
mentation. The concept of these models is that the biomass is divided into an
"active" part XA and a "structural part" Xc which is synthesized by the "active"
part and at the expense of XA • The "limiting" substrate is consumed in one or in
both reactions. It is evident that the structure of these models where the major
part of the biomass Xc is synthesized from a small, active portion of the cell XA
in a consecutive fashion will make it possible to predict a lag phase. Also the slow
increase of RNA coupled to a parallel increase of rp (Fig. 11.6) and a delayed in-
crease in the rate of total biomass production (Fig. 11.5) lend much credibility to
simple structured models.
In the context of lactic acid fermentation the structured models of Nielsen et al.
[27] are based on a very large database of steady state continuous fermentations.
They are therefore quite robust in their predictions of batch fermentations and tran-
sients with time constants of several hours. Processes with time constants in the
range of minutes, e.g., the alleviation of enzyme inhibition by metabolites in the
catabolic pathways when the dilution rate is changed are not accounted for in the
models of [27]. Hence the jumps in rp in Fig. 11.6 are not envisaged.
The structure of the models in [27] is as follows.
ll.S Growth Kinetics and Product Formation Kinetics 367

Lactic acid formation by catabolism:

s
sugar(s) --> HLac; rl = k l --- XA (11.13)
S + KI
Formation of active compartment XA : YXA 5 S + YXA N SN --> XA
(11.14)
Formation of structural compartment XG: YXG XAXA --> XG
One may assume that the formation of macromolecules is rate limiting in both
reactions of Eq. (11.14) and that the kinetics of the two reactions is similar:
( 11.15)

where the rate constant as well as the parameters of fl and f2 may be different for the
two reactions whereas the autocatalytic nature of all three reactions in Eq. (11.13)
and Eq. (11.14) is acknowledged through the factor XA , the fraction of the total bio-
mass (X A +Xc=l) which is "active." In [27] simple saturation type expressions were
used for both fl and f2 in Eq. (11.15). The two parameters of Eq. (11.13) and the six
parameters of Eq. (11.15) can be determined from experiments with either sugar or
nitrogen limitations.
It is not expected that a further compartmentalization of the cell will give much
more generality to the model - except of course if the production of a particular
enzyme is highlighted. Also the structured models of [27] must be conceived as
purely empirical although more biochemical knowledge is included in order to
explain phenomena which are foreign to the unstructured models of Eqs. (11.11)
and (11.12). These unstructured models are incapable of explaining a change in
the cell activity and consequently in the capacity for producing more biomass
and more product when the environment, e.g., the substrate concentrations s or
SN changes.
Very recently a family of models [28] which are capable of predicting both short
term and long term changes in cell metabolism have been advocated. Such changes
result after an abrupt change in the environment of a chemostat, e.g., following ad-
dition of a substrate pulse or a step change in the dilution rate. The experimental
basis for the models is built on anaerobic and aerobic yeast fermentations, but the
concepts of the models can certainly be applied to explain the events following a step
change in D (Figs. 11.5 and 11.6) in a continuous lactic acid fermentation.
The almost dormant cells at the low dilution rate before the transient are assumed
to possess a hidden capacity for energy production in the catabolic pathways. As
soon as the sugar concentration in the medium increases this capacity is activated
and rp jumps to a higher value. Biomass synthesis is considered to start at the rate
corresponding to the low dilution rate, and more biosynthetic machinery (e.g., RNA,
PTS proteins, proteins in the glycolysis, and finally proteins in the anabolic path-
ways) is produced by first order processes starting from the original level. All the
features of Figs. 11.5 and 11.6 are qualitatively explained by this oral model: the peak
in glucose concentration which is consumed once the active machinery (RNA) has
increased with the corresponding overproduction of biomass which compensates for
the initial wash out of biomass (Fig. 11.5). The accumulation of PEP at the low dilu-
tion rate could be the "hidden" capacity of the catabolic pathways which is mobilized
as soon as the PTS system for glucose senses an increasing glucose concentration in
the medium.
368 11 Lactic Acid Production

11.6
Lactic Acid Production on the Industrial Scale
Lactic acid is a commodity chemical with many different applications. The current
global production is approximately 60,000 tons year- 1 in different qualities, e.g., food
grade quality 88% by weight with a sale price of 1.9-2.4 US$ kg- 1 (fob works), [29].
The market price has been sliding since 1990 (2.5-2.75 US$kg- 1), probably related to
an increasing production volume, especially by the fermentation route.
Besides being an inhibitor of bacterial spoilage of food, lactic acid is used as an
acidulant, flavoring, pH buffering agent in soft drinks, jams and jellies, candy and
bakery products. Specialized dairy products and pickled vegetables such as sauerk-
raut and olives rely on lactic acid both for taste and for preservation. Approximately
50% of food-related lactic acid applications is found in conditioners and emulsifiers,
particularly in the bakery. The emulsifying agents are esters of lactic acid with long-
chain fatty acids.
Other, but in terms of volume much smaller, outlets for lactic acid are in drugs
(e.g., as the antiacne preparation ethyl lactate), in cosmetics, and as biodegradable
polymers for controlled release of drugs and in surgical sutures.
An exciting new and potentially enormous outlet for lactic acid has been created
by the recent development of technologies for commercial production of transparent
polymers for food packaging purposes. The properties of lactic copolymers ap-
proach those of petroleum derived polymers such as polystyrene and flexible Pvc.
The polylactide polymers have a good shelf life, degrading slowly by hydrolysis, but
they are easily compostable when deposited in dumps after use. Other recent appli-
cations of lactic acid is in food safe, biodegradable esters with low molecular weight
alcohols. These can be used as solvents and plasticizers. Together with the biode-
gradable and (from a carcinogenic point of view) absolutely safe polymer films for
food packaging, lactic acid derivatives are almost certain to win a solid position as
regulations and consumer preferences rapidly change our attitude towards conven-
tional chemical products.
It is therefore not strange that many patents appear which relate to either produc-
tion processes oflactic acid (of a purity and d-, I-composition suitable for polymer-
ization purposes) or relate to new applications oflactic acid in the food industry and
in production of bulk chemicals.
Chemically based production routes to lactic acid from acetaldehyde via lactoni-
trile or from propane via nitrolactic acid [30] are bound to become uncompetitive as
larger supplies of otherwise almost useless agricultural byproducts such as corn
steep liquor, a byproduct of the corn wet-milling industry, and with a lactic acid
content of 20-25 wt %, become available. Another route to lactic acid is of course
the fermentation of carbohydrate sources, e.g., molasses, corn syrup, or whey.
Consider production of lactic acid from whey as an example. A large, modern
dairy (Golden Cheese in Coronado, California) has access to milk from 60-
70,000 head of cattle. From its raw material, 13-1500 tons of milk per day, the factory
produces approximately 100-115 tons of cheese, 1-1.5 tons of whey protein, and
1000-1100 tons of water with 45 g r 1 lactose together with a number of minerals. It
is easy to calculate that if this lactose was dedicated to lactic acid production the
factory could produce 15-18,000 tons per year of lactic acid, thus doubling the pro-
duction of lactic acid in the US. The waste water with 4.5% lactose is of little value -
11.7 Process Technology in Lactic Acid Fermentation 369

and it could even be considered as a liability, incurring considerable costs for biode-
gradation. The lactic acid could have a sales value of more than 1.5 US$ kg- 1 in food
grade quality. It could immediately be used as a substitute for citric acid in beverages
and in other food products. The mild taste of lactic acid would be preferred by many
customers and the huge citric acid market (550,000 tons per year in 1990) could
easily be scavenged, although the price of citric acid is "soft" (-1.9 US$ kg- 1 [29])
due to competition from new producers in, e.g., China. In the longer term the lactic
acid might find applications for polymer production or elsewhere in the bulk che-
mical industry.
Considering the importance of having a cheap raw material available in sufficient
quantities, the dominating role of Cargill Inc. in development of large scale inte-
grated lactic acid fermentation processes is easily understood. This Mid-West based
company has access to raw materials from the corn-belt of the US and is involved in
dairies throughout the continent. The role of Du Pont and its associated companies
in polymer production is also obvious. Thus both Cargill and Du Pont have a num-
ber of patents, e.g., [31] for production of high purity lactide polymers and both
companies are developing their particular processes to a commercial scale, intending
to manufacture biodegradable polymers [32].
The lactic acid market is, however, not an easy one in which to operate. In late
1994 the Ecochem (Ecological Chemical Company) of Wisconsin, USA dropped out
of lactic acid production after 3 years of operation as a Du Pont-Con Agra partner-
ship [33]. The reason for the failure was stated to be that production oflactic acid for
the commodity market was not successful due to production difficulties, while the
parallel research in pilot scale to produce polylactides was claimed to be successful.
Both the Ecochem plant and a recent entrant in the market, Chronopol Inc. of Color-
ado, US [34] have produced lactic acid by purification of agricultural waste (corn
steep liquor). The fermentation route from whey is possibly a safer one; a closing
remark in [33] concerning the possible re-entry of Ecochem in the market might
suggest this to be the case.

11.7
Process Technology in Lactic Acid Fermentation

Considering that the raw material for a lactic acid fermentation is always going to
have a very modest price, the design of an industrial plant is critically dependent on:
- A low capital cost for the total plant, including the down-stream operation
- Low energy costs, both in the fermentation and in the down-stream processes
The low cost of the substrate does not imply that the yield of lactic acid on sugar is
of little importance. On the contrary, an almost complete conversion of the sugar to
lactate, and to lactate alone, is crucial for the cost of the down-stream processes, and
consequently for the total process economics. The investment costs, both in the fer-
mentation unit and in the down-stream operations, will be small if the cell density in
the fermentor is high. This means that the inhibitory effect of the product on the
process kinetics must be minimized.
Further design considerations are centered on the cost of the non-sugar substrates.
The LAB need a rich nitrogen source to grow, and yeast-extract and other commer-
cially available nitrogen sources may well constitute a substantial portion of the pro-
duction costs.
370 11 Lactic Acid Production

Finally one needs to consider the alternatives of batch and continuous fermenta-
tion. Several patented lactic acid production processes are built around a batch fer-
mentation process, but neither batch nor fed batch operation can be the preferred
design solution when the product is as cheap as lactic acid and has to be produced in
such large quantities as indicated by the potential scale of the Golden Cheese exam-
ple (1000 tons per day of feed). The LAB are genetically quite stable organisms and
the desired product is formed as part of the central metabolism which is not going to
change over time. Hence the usual argument for batch production of chemicals or
proteins from the secondary metabolic pathways will not hold for lactic acid fermen-
tation. Finally, the process control, e.g., separation of lactic acid from the broth is
much simpler in a continuous fermentation than in a batch where the level of the
sugar substrate decreases and the level of the product increases during the fermenta-
tion.
Different strains of LAB have been used by different research groups working with
pilot plant lactic fermentation. Lb. helveticus was used in [35, 36], Lb. casei by [37-
41], Lb. delbrilckii by [42, 43], and Lb. plantarum by [44].
Industry has picked one or several of the above-mentioned organisms (unpub-
lished information). It is interesting that they are all chosen from the genera of Lac-
tobacillus. Perhaps this is due to their higher acid tolerance, but it appears that sev-
eral fast acid producers from the L. lactis genera might offer a good alternative to the
Lb. strains since in any case the lactic acid has to be removed continuously in all
industrial productions with high cell density.
Although any type of cheap sugar source can be used it appears that all academic
research groups and (as far as can be conjectured) all industrial producers prefer to
use whey permeate - the 45 g t 1 lactose aqueous solution which results when the
whey proteins have been filtered from the tail-end of the cheese production lines.
Yeast extract, which can of course be used in academic research, is too expensive as
a nitrogen source, and hydrolyzed whey protein is also recommended in academic
research reports focused on the economics of the lactic acid process [36, 45]. When
the fermentation process is located at the site of a large dairy the use of whey protein
rather than any other nitrogen source must be the only rational choice. To make the
nitrogen from whey protein more accessible, enzymatic degradation of the protein
may be advisable as recommended in the Lactoscan process [46].
The two major considerations in the fermentation step are (1) to have a high bio-
mass concentration and (2) to remove the lactic acid to avoid slowing down or stop-
ping the fermentation.
Cell recirculation and continuous removal of lactic acid is included in most of the
pilot-scale studies cited above. With recycling of cells one may in principle obtain an
infinitely high capacity of a given reactor volume.
Let the cell density of the effluent from the plant be XE which is lower than the
biomass concentration x in the fermentor due to a separation step (filtration, cell
sedimentation, hydrocyclone) between the reactor outlet and the plant-outlet. A re-
circulation stream of cell density XR and flow rate Rv (R=recirculation rate and v net
volumetric flow through the plant) is returned to the reactor inlet where it is mixed
with fresh feed.
Define the net dilution rate D=v!V where V is the reactor volume.
Further define f=xE/x and the cell separation factor ~=XR/X.
11.7 Process Technology in Lactic Acid Fermentation 371

A steady state continuous operation can now be maintained for any given value of
D and at any given specific growth rate J-l by selecting f such that J-l=fxD. The corre-
sponding recirculation rate R is given by
f=I-R(~-l) (11.16)
Simultaneous removal of lactic acid and recirculation of biomass can be achieved
in a membrane type reactor - ultrafiltration with a cut-off of 20 Da has been pro-
posed. In a continuous LAB fermentation with a reasonably small dilution rate the
residual sugar concentration is very low as seen in Sects. 11.2-11.4. Consequently the
loss of sugar is negligible. If the lactic acid content of the retentate is considered to
be a nuisance, one may run the retentate through an ion-exchanger to trap the acid
before the retentate is returned to the fermentor. Occasional back flushing of the ion
exchange column with NaOH adds more lactic acid to the permeate from the UF
unit. Various types of ion exchange resins (e.g., Amberlite RA-400 from Rohm and
Haas [41]) have been suggested.
Concentration of the lactate stream from the UF-unit is most economically done
by electrodialysis [39, 45] which is easy to scale up and can be run at a low electric
power requirement (1 kwh kg HLac- 1 [30]). The process is continuous and lactate
will pass through a (positively charged) anion electrodialysis membrane. It is advi-
sable to remove calcium and magnesium ions from the solution by inserting a ca-
tion-chelating resin column before the electrodialysis unit. Ca and Mg salts of lactic
acid are very sparsely soluble and they will invariably foul the electrodialysis mem-
branes and decrease the current efficiency. After removal of divalent cations and
electrodialysis the process stream is free of sugar and protein residues and the lactate
is usually in the form of NH4Lac. In a final separation step bipolar electrodialysis is
recommended [46] to give a pure lactic acid which can easily be concentrated to
80 wt% lactic acid free of other organic acid residues by evaporation of water in a
low pressure continuous distillation column.
Each of the down-stream purification process steps must be carefully considered
and as far as possible optimized since otherwise a substantial loss of lactic acid and
prohibitive energy costs may be incurred. The know-how of these process steps and
the best integration offermentor and down-stream operations (e.g., best cell concen-
tration, optimal bleed of biomass from the reactor) is of course proprietary informa-
tion in the companies engaged in commercialization of the process, but the major
stages of both the fermentor design and of the design of the purification process are
probably always those described above.
It is encouraging to follow the intense research effort both in academia and in
industrial companies to prove the economic viability of a really large scale fermenta-
tion process with cheap raw materials and a product which, after suitable further
process steps, can become quite valuable. Lactic acid fermentation is in some sense
deeply rooted in the current efforts to develop "green" products from "sustainable"
raw materials. The integration of all process steps is certainly required to make the
product competitive with other cheap products like citric acid. But many of the sec-
ondary processes derived from the main production process - such as the produc-
tion of electrical energy by anaerobic digestion of the process effluent in the Lactos-
can process [46] - are indicative of the ecological awareness of entrepreneurs in the
modern biotechnology based industry.
372 11 Lactic Acid Production

References

1. Salminen S, Wright A von (ed) (1998) Lactic acid bacteria, microbiology and functional
aspects, 2nd edn. Marcel Dekker, New York
2. Roissart H de, Luquet FM (ed) (1994) Bacteries lactiques. Aspects fondamentaux et tech-
nologiques, vols 1 and 2. Lorica, Uriage (France)
3. Venema G, Huis in't Veld JHJ, Hugenholtz J (eds) (1996) Lactic acid bacteria, genetics,
metabolism and applications. Kluwer Academic Publishers, Dordrecht Holland for
FEMS-reprinted from Antonie van Leeuwenhoek (1996) 70:2-4
4. McKay LL (1983) Antonie van Leeuwenhoek 49:259
5. Kashket ER (1987) FEMS Microbiol Rev 46:233
6. Pouwels PH, Leer RJ (1993) Antonie van Leeuwenhoek 64:85
7. Benthin S, Nielsen J, Villadsen J (1992) Biotechnol Bioeng 40:l37
8. Benthin S, Nielsen J, Villadsen J (1993) Biotechnol Bioeng 42:440
9. Benthin S, Nielsen J, Villadsen J (1993) Appl Environ Microbiol 59:3206
10. Benthin S, Nielsen J, Villadsen J (1994) Appl Environ Microbiol 60:1254
11. Postma PW, Lengeler JW, Jacobsen GR (1993) Microbiol Rev 57:543
12. Thompson J (1979) J Bacteriol140:774
l3. Thompson J (1987) FEMS Microbiol Rev 46:221
14. Hickey MW, Hillier AJ, Jago GR (1986) Appl Environ Microbiol 51:825
15. Hutkins RW, Morris HA (1987) J Food Prot 50:876
16. Benthin S, Nielsen J, Villadsen J (1991) Anal Chim Acta 247:45
17. Benthin S, Nielsen J, Villadsen J (1991) Biotechnol Tech 5:39
18. Jensen NBS, Jokumsen K, Villadsen J (1998) Biotechnol Bioeng (submitted)
19. Snoep JL, Teixeira MJ, Neijssel OM (1991) FEMS Microbiol Lett 81:63
20. Garrigues C, Loubiere P, Lindley ND, Cocaign-Bousquet M (1997) J Bacterio1179:5282
21. Nielsen J, Villadsen J (1994) Bioreaction engineering principles. Plenum Press, New York
22. Benthin S, Schulze U, Nielsen J, Villadsen J (1994) Chern Eng Sci 49:589
23. Jensen PR, Hammer K (1993) Appl Environ Microbiol 59:4363
24. Melchiorsen CR (1997) MSc thesis, Techn Univ, Denmark
25. Benthin S, Villadsen J (1995) J Appl Biotechnol 78:647
26. Luedeking R, Piret EL (1959) J Biochem Microbiol Technol Eng (Biotechnol Bioeng) 1:393
27. Nielsen J, Nikolajsen K, Villadsen J (1991) Biotechnol Bioeng 38:1, 11
28. Duboc P, Stockar U von, Villadsen J (1998) Biotechnol Bioeng 60:680-689
29. Chern Mktg Rep 251:(10) 7 March 1997
30. Kirk-Otmer (ed) (1995) Encyclopedia of chemical technology, vol l3 (4th edn). Wiley,
New York
31. Gruber PR (1992) US. Pat 5,142,023 (Aug 25, 1992) (to Cargill Inc.); Bhatia KK (1989) US
Pat 4,835,293 (May 30, 1989) (to EI du Pont de Nemours & Co. Inc.)+about 15 follow-up
patents
32. Chern Eng News 6 (Aug 6, 1992)
33. Chern Mktg Rep 246: (22) 28 November 1994
34. Chern Mktg Rep 249:(15) April 8 1996
35. Aschlimann A, Stockar U von (1990) Appl Microbiol Biotechnol 32:398
36. Boyaval P, Goulet J (1988) Enzyme Microb Technol10:725
37. Vaccari G, Gonzales-Vara RA, Campi AL, Dosi E, Brigidi P, Matteuzzi D (1993) Appl Mi-
crobiol Biotechnol 40:23
38. Hjorleifsdottir S, Seevaratnam S, Holst 0, Mattiasson B (1990) Curr Microbiol 20:287
39. Krischke W, Schroder M, Trosch W (1991) Appl Microbiol Biotechnol 34:573
40. Tuli A, Sehthi RP, Khanna PK, Marwaha SS (1985) Enzyme Microb Technol 7:164
41. Senthuran A, Senthuran V, Mattiasson R, Kaul R (1997) Biotechnol Bioeng 55:841
42. Stenroos S-L, Linko Y-Y, Linko P (1982) Biotechnol Lett 4:159
43. Major NC, Bull AT (1989) Biotechnol Bioeng 34:592
44. Shamata TR, Sreekantiah KR (1988) J. Ind Microbiol 3:175
45. Tejayadi S, Cheryan M (1995) Appl Microbiol Biotechnol 43:242
References 373

46. Final report on lactic acid fermentation and purification (1998) (confidential report pub-
lished by BIOSCAN A/S, Odense 5230 Denmark from whom further information may be
obtained)
12 Control Strategies for High-Cell Density Cultivation
of Escherichia coli

Clemens Posten, Ursula Rinas

12.1
Introduction

Growth of microorganisms under conditions of substrate saturation in batch pro-

cesses is often accompanied by substrate inhibition, by-product formation, or endur-
ing adaptation. These problems can be overcome, in principle, by fed-batch or con-
tinuous cultivation, where the concentration of one substrate (in general the carbon
source) is kept at a limiting level and, furthermore, where the growth rate can be
kept at a desired value. Production systems with E. coli in batch culture suffer mainly
from the formation of acetate, which is produced in response to oxygen limitation or
excess carbon. The prevention of the accumulation of toxic levels of acetic acid is the
main task for the achievement of high cell and high product concentrations in the
bioreactor. The formation of acetate can be circumvented when cell growth occurs at
reduced growth rates, thus preventing carbon overflow metabolism. The critical
growth rate above for which acetate formation occurs is influenced by the growth
medium, temperature, and, more general, by the overall physiological status of the
cells. In addition, it is important, to allow cell growth at a constant growth rate to
permit the synthesis of a product (e.g., a recombinant protein) of reproducible and
defined quality. In continuous cultivation, a constant and non-critical growth rate
with respect to acetate formation can be achieved, but problems with plasmid stabi-
lity and, therefore, decreasing productivity may occur. So in many research and in-
dustrial applications, fed-batch processes for the production of recombinant proteins
are employed. With an accurate choice of the process and control parameters, very
high cell concentrations in the bioreactor can be achieved. The following will give a
concise mathematical description of the related microbial reactions, of the reactor
dynamics, and of feasible control strategies, which finally can be called a basic model
for high-cell density cultivation (HCDC).

12.2
Basic Modeling of a Fed-Batch Strategy

A desired improvement in process performance by means of feedforward or feed-

back control can only be achieved by employing a process model. Such a model
should contain as much available knowledge as possible on the one hand but on the
other should be easy to obtain, robust, and feasible under process conditions. Many
process models in bioprocess engineering consist of a physiological sub-model re-
presenting the metabolic stoichiometry and kinetics and a reactor model to describe
12.2 Basic Modeling of a Fed-Batch Strategy 375

the relation between the process parameters, kinetic parameters, and material con-
centrations inside the reactor. Since the main goal of control strategies in HCDC is
the estimation and control of the specific growth rate, the model has to relate this
non-measurable physiological variable with measurable process variables, a task of
physiological control.

12.2.1
The Physiological Model

For modeling and control of high-cell density cultivation (HCDC), a simple struc-
tured physiological model for a typical aerobic microorganism is sufficient. The sim-
plified structure of the metabolic pathways of such a micoorganism, namely Escher-
ichia coli, is given in Fig. 12.l.
With the assumption that the time constants of accumulation of intracellular me-
tabolites are very small compared to reactor time constants, a set of linear equations
for the process stoichiometry can be obtained by applying material balances and
stoichiometric relations to the reaction network. According to network theory, a sui-
table choice of balance boundaries has to guarantee that:
- Each network knot lies inside at least one balance boundary
- Each pathway is cut by at least one balance boundary
- For each boundary with n pathways only n-l linearly independant equations are
possible

Details are given in the chapter Metabolic Flux Analysis.

s
0,
rs
r
0,

,/ r ATI'
_ _ ..-OC

0" --;'
,
" energy
• balance
r Tr.m

co,
x

Fig. 12.1. Simplified metabolic structure of an aerobic microorganism for set-up of a partially
structured process model; the circles represent the balance boundaries
376 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

For the indicated case, one obtains a linear set of equations

E'r=O (12.1)

with the vector of specific turnover rates

(12.2)

and the system matrix

1 0 0 -1 -1 0 0 0 0 0
0 0 1 0 -1 0 0 0 0
0 eN,NH, 0 0 0 -eN,X 0 0 0 0
E= 0 0 0 0 -1 -1 0 0 (12.3)
0 0 1 0 eo,s 0 -eo,co, -eO,H20 0 0
0 0 0 0 ec,S 0 -eC,C02 0 0 0
0 0 YATP,02 0 0 YX,ATP 0 0 0 -1

Here, the rows represent the mass balance of the substrate utilization, the anabolic
mass balance, the anabolic nitrogen balance (assuming a constant elemental compo-
sition of the biomass), the catabolic mass balance, the catabolic oxygen balance, the
catabolic carbon balance (the latter three items representing the substrate oxidation
stoichiometry), and, finally, the energy (ATP) balance, which couples the anabolism
and the catabolism, thus substituting the missing stoichiometry of the substrate uti-
lization balance.
From these considerations, seven linear equations are obtained for the eight un-
known metabolic fluxes. Since only the ratio between energy generation and energy
consumption is required for the modeling purposes here, rATP is not calculated as an
absolute value, while rATP,m is taken as an unknown parameter. Finally, the system is
completely described by an additional kinetic equation for the specific substrate up-
take
cs
rs = rS,max . (12.4 )
ks + cs
which is, from a modeling point of view, a link between the physiological and the
reactor model. Oxygen and nitrogen limiting conditions are not considered, since
they are usually avoided during HCDC. These conditions would lead to additional
kinetic uptake equations for the respective medium components.
To describe the performance of an HCDC strategy, which is basically a fed-batch
process strategy [1] leading to a constant specific growth rate fL, the kinetic
Eq. (12.4) and a subset of Eq. (12.1), namely at least one yield equation for biomass
production from nutrient uptake, giving, e.g.,

fL =Yx,s' rs - fLm' (12.5)

are sufficient. The parameters can be derived from Eq. (12.1) by linear combination
of the mass balance of substrate utilization and the energy balance. However, in the
representation according to Eq. (12.5), the parameters can be determined experi-
mentally.
12.2 Basic Modeling of a Fed-Batch Strategy 377

12.2.2
The Reactor Model

In case of a continuously stirred tank reactor (CSTR), no spatial distribution of con-

centrations have to be considered. The reactor equations can be obtained by concise
application of the material balances for balance boundaries around the reactor.
The material balance for biomass, including accumulation and reaction term, gives
dmx (t)
-d-t- = fL(t) . mx(t) (12.6)

while the carbon substrate balance with an additional transport term for the carbon
substrate pumped into the reactor can be written as

dms(t)
~= fs,f(t) . CS,f - rs(t) . mx (t) (12.7)

The corresponding differential equations for the concentrations can be derived

from the total differentials

dmx(t) = d(cx ((t) . VR(t)) = dcx(t) . V(t) + dVR(t) . cx(t)

dt dt dt dt (12.7a)
= fL(t) . mx(t)

and

dms(t) = d(cs(t) . VR(t)) = dCs(t) . V(t) + dVR(t) . cs(t)

dt dt dt dt (12.8)
= fs(t) . CS,f - rs(t) . mx(t)

Relating to the reactor volume VR(t) and rearranging finally yields

dcx(t) fs,f (t)
-- = fL(t) . cx(t) - - - . cx(t) (12.9)
dt VR(t)

and
dCs(t) fS,f(t)
-- = -
dt
rs(t) . cx(t) +- - . (CSf, -
VR(t)
cs(t)) (12.10)

Since these equations include a dilution term ~:i~;, the total biomass but not the
biomass concentration will increase exponentially assuming constant fL.
The volume VR(t) of the fluid phase in the reactor itself can be regarded as

dVR(t) = £ (t) (12.11)

dt S,f

while differences of the density of the feed medium and the reactor broth are ig-
nored as well as evaporation, sampling, and water production from the oxidation
reactions of the microorganisms. A robust process strategy should not rely on a
precise knowledge of the reactor volume on- or off-line. However, Eq. (12.11) can
be used for simulations and to compare modeling results with measurements.
378 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

For the gas transfer into the reactor, separate balances for the fluid and for the gas
phase are necessary. The oxygen balance in the fluid phase reads

dC02(t) = OTR(t) - OUR(t) (12.12)

dt
where the oxygen transfer rate OTR(t) into the reactor follows a linear material
transport law, namely Fick's first law related to the volumetric gaslfluid exchange
interface a,

OTR(t) = kLa . (C~2 (t) - C02 (t) ) (12.13)

while the volumetric oxygen uptake rate OUR(t) is given as

OUR(t) = r02 (t) . cx(t). (12.14)

Quasi-steady-state conditions with respect to dissolved oxygen

d~~2 = OTR(t) - OUR(t) :::::: 0 =? OTR(t) = OUR(t) (12.15)

can be assumed unless dynamics of pOz-control or sudden substrate depletion are

considered.
From the gas phase balance equations for oxygen, carbon dioxide and inert gases,
mainly nitrogen, one obtains

MW C02 . (OTR - Qtmol

. . MW 02 . X~2)
~m_
(12.16)
02 OTR. MW C02 - MW 02 ' (Q~mol . MW c02 + CPR)

for the oxygen molar fraction in the off-gas, and

t -MW02 ' (Qtmol·MWc02 'X~02 + CPR)

X~'b2 = '. (12.17)
OTR· MWC02 - MW02 . (Q~,mol . MWC02 + CPR)

for the carbon dioxide molar fraction in the off-gas. Details of the gas phase balance
can be found in [2].

12.3
Growth Rate Control via Substrate Feeding

For maintaining a constant specific growth rate,

(12.18)

a limiting and, in case of constant ks, a constant substrate concentration Cs is re-

quired.
Equation (12.6) can be solved directly yielding

(12.19)

for the biomass time profile that will be obtained.

12.3 Growth Rate Control via Substrate Feeding 379

Quasi-stationary conditions dcs/dt=O for the substrate concentration Cs can be

derived from Eq. (12.10) by substitution of rs=(I1+l1m)/yx,s (Eq. 12.5) and rearran-
ging to
dcs (t) I1set + I 1 m t
- - = fsf(t)· (CSf-CS)- ·mxo·elL =0 (12.20)
dt " yx,s'
This is only achieved, as can be shown by comparison of the coefficients, by feed-
ing with an exponentially increasing feeding rate

C .
IS f
(t) = (l1set + 11m) . mx,o . e~"
"t· t
(12.21)
. yx,s . (Cs,f - Cs (t))
As usually cs«CS,f> the feeding rate is often referred to as

fs,f(t) = (l1set + 11m) . mx,o . eIL",·t (12.22)

Yx, S . CS,f
which can be interpreted as the feeding rate for the amount of substrate that the
microorganisms use for maintaining the desired growth rate, while the neglected
part (difference between Eqs.12.21 and 12.22) is the amount that is necessary to
keep the substrate concentration constant despite the increasing culture volume.
Simulation results for this basic strategy are given in Fig. 12.2. Values for process
variables, model parameters, and starting values of the state variables are obtained
from [3].
As can be seen from a comparison of the exponentially increasing feeding rate, the
biomass concentration also increases in the first part of the fed-batch phase nearly
exponentially as long as the additional feeding volume is low compared to the start-
ing volume in the bioreactor. Later, the increase slows down to a theoretical limit of
CX,final=CS,fXYX,S' This value can hardly be reached because of the limited volume of
the reactor. A smaller starting volume would result in a higher final biomass con-
centration, but also causes problems such as longer cultivation time, low volumetric

- Volume [11 - - Biomass [gill Feeging rate [Uh)

100 - - - Substrate [gn) 0,030

0,025
80 3
O,020~
..
."

",60 ~.
.2!
O,015~
'"coE'" 40
0 ~
iii 0,010

20 0,005

0 ~=-_~~_r-_-' __~__~O 0,000

0 5 10 15 20 25
Cultivation Time [hI

Fig. 12.2. Simulation of an HCDC using glucose as carbon source and, with the equations gi-
ven above (Eqs.12.1-12.5, 12.9-12.11, and 12.21), the parameters have been chosen from [3] as
cs,£=795 gil, VR,o=2.5}j Cs,o,ba,ch=25J/I, cx,O'~f'~h=O.05 gil, J.imax=0.45 h-l, YX,s=O.5, m=O.025 gl
(gxh)--+J.im=O.0125h , J.ise,=O.12h (O.14h III [3])
380 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

Model for High·Density-Cultivation

~ qf~
f-----=-'-,Sf. m my,rS,rAct,rProt.I02,rC02

Feed ~ ~~Q
RLea-c-to-r-m-o-,Jdel Off·gas balmce
x02, xC02,RO,CI.C02.QC2
cX,CS.cACE!,CPlot

mY.rS.tAce.IPfot,,02.rC02

Fig. 12.3. Basic model structure of a fed-batch process, here in the graphical representation of
SIMULINK, the program used for the described simulation studies

productivity, and the need for pumps working in several orders of magnitude. An
optimization approach can be found in [4]. For production purposes, the initial vo-
lume is in most cases not smaller than one-third of the final volume, which means a
maximum of three to four doubling times of the biomass after the start of the fed-
batch phase of the HCDC. However, the feed concentration should be as high as
possible and is often chosen to be at the limit of solubility. Therefore, even feed of
crystalline substrate has been proposed in the literature [5].
After depletion of the initial carbon source present in the batch phase of the culti-
vation, the carbon substrate concentration remains of the order of magnitude of the
limitation constant ks of the E. coli cells as long as /Lset</Lmax. This self-stabilizing
property of fed-batch processes without external feedback control can be understood
by the process internal feed-back loop as shown in Fig. 12.3. An infinitesimal pertur-
bation of the glucose concentration, e.g., an increase, will result in a higher glucose
uptake rate and a higher specific growth rate. These reactions will then reduce the
glucose concentration in the next time instant.
This previously described strategy of achieving a constant growth rate by expo-
nential feeding depends strongly on the assumptions of absence of substrate accu-
mulation and constant yield of biomass from substrate consumption YX,s' Short dis-
tortions of feeding can be compensated by estimation of the actual biomass using
the same yield equations as shown above. However, in practical applications, the
substrate yield may not be known exactly or may change during the cultivation.
Therefore, the question arises what the effect of an error in the assumed substrate
yield with regard to the true specific growth rate can be. The fully relative sensitivity
functions

ocx(t) y~,s
scx,yxs = -",-- . O-(t) (12.23)
uYx,s C x
12.4 Growth Rate Control via Oxygen Supply 381

--Scx,yxs
1,0
, \ - - - S/l,yxs [7h)
\ -S/l ]
\
0,8

:;;;: 0,6
z. ,,
.,
'S;

'~ 0,4
en

0,2

0,0
0 5 10 15 20 25
Cultivation Time [hi

Fig. 12.4. Fully relative sensitivity functions of biomass and specific growth rate with respect
to changes in the yield coefficient of the bacterial metabolism at different time instances; the
nominal parameter values are the same as given in Fig. 12.2

and
O/-l (t) y~,s
s!',yxs = - - .- (12.24)
OYx,s /-lset

can be evaluated. These functions give information about the time course of relative
changes in biomass concentration and specific growth rate in case a relative error of
the yield occurs which is not compensated in the feeding profile. Figure 12.4 shows
simulation results where the sensitivity functions are computed via numerical differ-
entiation. Here, a sudden change of Yx,s at the beginning of the fed-batch-phase is
assumed.
As expected, if the assumed yield is too low, lower biomass concentrations are
obtained. The sensitivity function exhibits an asymptotic behavior with a limiting
value 1. This means that an error in yield causes the same error in biomass.
A change in the yield parameter results in a corresponding change of the specific
growth rate. Fortunately, this error is damped and vanishes slowly. This is due to the
fact that, e.g., a higher yield results in a higher growth rate but also, after some time, in
a higher biomass. The amount of substrate per cell fed to the reactor then falls slowly
under the level calculated beforehand thus compensating the yield effect. Similar
sensitivity functions are obtained for errors in the initial feeding rate, e.g., by an
error in the measurement of the initial biomass. So the strategy of direct substrate
feeding is to some extent robust to changes in yield or other parameters used for
calculating the initial feeding rate (compare Eq. 12.21 for t=O).

12.4
Growth Rate Control via Oxygen Supply

As shown earlier, a constant growth rate is achieved only in case of constant sub-
strate yield and in case of a preset growth rate lower than the critical growth rate
382 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

which results in the formation of acetic acid. These parameters may also change
during cultivation. So additional control strategies are advisable. From the simple
linear model (Eq. 12.1) different control approaches can be deduced and have been
proposed in the literature. However, all of them assume at least one constant yield
coefficient, i.e., one constant element in the system matrix. Since the oxygen uptake
is closely related to the energy metabolism of aerobic microorganisms and molecular
oxygen is not incorporated into the biomass, some authors based their control strat-
egy on the specific oxygen uptake rate, which is related to the specific growth rate by

/J = YX,02 . r02 - /Jrn (12.25)

as shown above for the specific substrate uptake rate as an analogous case. Exponen-
tial growth with a constant specific growth rate is obtained by an exponential in-
crease of the oxygen transfer rate via increasing stirrer speed, aeration rate, oxygen
fraction, or pressure. This strategy, proposed and applied by Riesenberg et al. [6]
dates back to Mori et al. [7] and Cutayar and Poillon [8]. It is up to this point analo-
gous to the strategy using an exponential substrate feeding profile. However, there
are two basic differences. First, an exact relation between the oxygen mass transfer
and process parameters is hardly available. Second, since oxygen is not a limiting
substrate, special attempts have to be made to force the cells to use the exponentially
increasing oxygen supply. These two problems define the task of establishing two
control loops as will be pointed out below.
For both control loops, it is necessary to estimate the oxygen uptake and the oxy-
gen mass transfer. The oxygen transfer rate can be calculated from the off-gas ana-
lysis from application of balances such as

(t) = fG,inMWO,. [x~(1-X~(i2(t)) -Xg~t(t)(I-X~o2)] (12.26)

q02 Vmol 1 - x out ( ) out ( )
02 t - x C02 t

which is a rearrangement of the Eqs. (12.16) and (12.17), here in a representation

for on-line application without considering the reactor volume. Now q02(t) is ob-
servable and controllable as long as there is a gradient in the oxygen partial pres-
sure between fluid and gas phase, and can be chosen to serve for the demand of
the E. coli cells as

q02(t) = r02(t) . mx(t) = /Jset + /J m . mx(t) (12.27)

YX,02

where for constant /J=/Jset and an exponential increase of mx(t), an exponential

increase in oxygen supply is required in analogy to the substrate feeding strategy
described above. Using the approach via oxygen supply, potential overfeeding can
easily be detected by a measurable increase of the dissolved oxygen concentration.
In practical applications, disturbances may occur causing a remarkable deviation of
the biomass profile from the precalculated value. It is also possible that a non-con-
stant growth profile is desired. In such cases, an estimation of the actual biomass
in the reactor has to be provided in order to calculate the oxygen demand accord-
ing to Eq. (12.27). The biological model information for such an estimator of the
biomass can be based on the assumption of constant yield coefficients. Employing
12.4 Growth Rate Control via Oxygen Supply 383

again Eqs. (12.6) and (12.25) as well as r02=q02(t)/mX(t) (Eq.12.14a), the model
equation

(12.28)

can be used as an observer model. This differential equation can be either computed
numerically on-line or can be solved (MAPLE V) yielding

J
t

rnx(t) = YX,02' e- llm . t ellmT . q02(T)dT+ mx,o' e- Ilm · t (12.29 )

o
where integration is only necessary during the process. Moreover, with the knowl-
edge of the biomass, the specific growth rate can be estimated as
q02
jl = YX,02' rnx(t) - f-Lm (12.30)

For negligible maintenance or constant growth rate, the integration of Eq. (12.29)
can be simplified, yielding

jl= t
(12.30a)
k1(Yx,02) cx,o) + I q02(T)dT
to

as given in the literature [2, 6]. So all prerequisites are given to supply the culture
with an oxygen mass flow suitable for the current biomass and the chosen growth
rate. Finally, the task of this first control loop is to control the estimated specific
growth rate at the desired value f-Lset via control of the oxygen supply rate q02' There-
fore, according to Eq. (12.13), either kLa or C02 * can be influenced by the process
parameters mentioned above as control variables.
Since oxygen is not the limiting substrate, a second control loop has to be estab-
lished in order to feed enough substrate to allow the cells to utilize the offered
oxygen for respiration. At first glance, the stoichiometry between glucose and oxy-
gen consumption could be considered which would end up in a glucose feeding
profile proportional to the given oxygen transfer profile. However, even small
changes in the related yield coefficients could lead to substrate accumulation on
one hand or loss of oxygen controllability on the other hand. At this point, another
difference between the glucose and the oxygen feeding profile approach comes into
action, namely the fact that, in contrast to the substrate concentration, the oxygen
partial pressure in the reactor can be measured on-line. As long as dp02/dt=0, the
oxygen transfer rate OrR equals the oxygen uptake rate OUR which is - via meta-
bolic stoichiometry - a function of the substrate consumption rate. The basic idea
of this second control loop is to control the oxygen partial pressure at a constant
value by substrate feeding, which guarantees a substrate consumption suitable to
reach finally OUR=OTR. A special point here is that no knowledge of Yx,s or Y02,S
is required.
Simulation results of this strategy are shown in Fig. 12.5 with the assumption of
ideal controller performance. Values for process variables, model parameters, and
starting values of the state variables are obtained from [6]. While the aeration rate
was constant, a linear relation between the kLa-value and the stirrer speed is as-
sumed. However, this is not essential for the process performance.
384 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

--Biomass [gil] 100 2000

100
- - - Substrate [gil]
----- p02[%]
80 1600
80 ----- n-stir [min-']

60 1200
~60 ::J

'"
() ~ ~
i< /
40 800 3'
() 40 /

.,. ~
/

,~ j:
,I: ~ / 20 400
20 , /
/

;:. ,. t1:--.-:".-.-'.
/---
--------~
0 0
0
0 5 25 30
Cull;~ation 1i~e [hI 20

0,2
-°2 ---- OTR [gll/h]
- - - - - CTR [g/l/h]
- - - CO2
12
0,2
10
Sl
0,1 / 8..::0
-'-
~
./ 6-~
~- 0,1
J
0,0
2

0,0
5 10 15 20 25 30
° Cultivation Time [hI

Fig. 12.Sa,b. Simulation of an HCDC with the equations given above (Eqs.12.1-12.17, 12.21,
12.26, and 12.28); the parameters have been chosen from [6] as cS,f=700 gil, VR,o=371,
cs,o=25 gIl, cX,o=0.1 gil (0.07 gIl in sim.), flmax=0.45 h-l,(rs,max=0.9 h- 1), YX,s=0.5 gig, m=O.O gl
(gh)--tflm=O.O h -1, flset=O.ll h -I, fG ,in=4440 llh, P02,set= 10%, kLa=f(nstir) assumed: a measur-
able concentrations and process parameters in the fluid phase; b measurable concentrations
in the off-gas and calculated volumetric mass transfer rates

Of course, the trajectories of glucose and biomass concentrations are comparable

to the case of growth rate control via substrate feeding. The volumetric oxygen
transfer coefficient kLa is adjusted at the beginning of the batch phase so that the
partial oxygen pressure can drop toa small but not limiting value p02,set. Then, p02
is usually controlled by control of the oxygen transfer via stirrer speed. After deple-
tion of the glucose, no more carbon source is available for respiration and, conse-
quently, the oxygen uptake stops, while the continuing oxygen transfer leads to a
sudden increase in p02' In some cases other peaks occur in response to glycerol
and acetate depletion, which can be taken as a starting signal for glucose feeding
[9]. In the further course of the cultivation, OTR and eTR increase exponentially as
expected. When p02 is indeed constant, these quantities can be calculated from off-
gas analysis.
Similar to the question of accuracy in the case of the substrate feeding strategy, the
robustness of this strategy to inaccurate assumptions is investigated in the following
paragraph. As already mentioned, the sensitivities to unknown substrate yield Yx,s
namely scx,yxs=sjl,yxs=O vanish. This is indeed one of the main advantages of the oxy-
12.4 Growth Rate Control via Oxygen Supply 385

1,0 -- e~.)'X02
;,
e~.)'X02
e X,XQ
~ 0,5 - - - - - eJi.XO
o
.,<:
c:
o
~ 0,0
.,
13o

.
~-0 ,5
~
,
-1.0 +---,.-~----r-..;..
" -.----""-~----r---'
o 5 10 15 20 25 30
Cultivation Time IhJ

- - e.cx.yxo2
1,0 ------ e.my.yxo2
e.cx.cxhut

0,5
."'< -... -e.my.C"xhut
Q)

~ 0 ,0
lii
1\!
l05
~,

·1.0 +-~-...--~--r---="
" '--r--~-r--~-.-~-'
o 5 10 15 20 25 30
Cultivation Time [hi

Fig. 12.6.a,b. Relative observation errors of growth and biomass in response to erroneous as-
sumptions of biomass yield with respect to oxygen and starting values of the observer for
biomass rnx(t=O) as a result of an erroneous measurement of mx,a. b Relative process errors
of growth and biomass in response to erroneous assumptions of biomass yield with respect to
oxygen and starting values of the observer during application of the control strategy via oxy-
gen supply

gen feeding strategy. However, the oxygen yield YX,02 may also change from one
cultivation to another or during an experiment. In such a case, the observation of
biomass and of the specific growth rate deviates from the real values. Simulations of
this observation errors (Fig. 12.6) reveal that an error in the assumed oxygen yield
causes an error in the estimated biomass and the estimated growth rate as well.
These errors do not completely vanish during the entire cultivation. In a similar
way, differences between the assumed and the real biomass concentration at the be-
ginning of the fed-batch phase (caused, e.g., by an inaccurate off-line measurement)
will cause an observation error which only slowly vanishes.
This situation continues considering a controlled cultivation as pointed out above.
Even if it is possible to control the estimated specific growth rate at its setpoint as
{L=flset an error in YX,02 will cause a process error, e.g., if the biomass yield with
386 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

respect to oxygen is overestimated, then the real growth rate is too low. It should be
noted that the biomass and the growth rate observer do not have independent mea-
surement information which could be taken for convergence of the estimation re-
sults. However, this strategy with the measurement of p02 gives at least the oppor-
tunity to detect and avoid carbon substrate overfeeding. This can be considered as
the second advantage of the described strategy. Altogether, the main merit of this
approach is to prove that it is possible to combine the features of the self-stabilizing
property of the limiting substrate and the observability of a co-substrate like oxygen
or eventually a product like carbon dioxide.

12.S
Considerations for Improved Observation and Control

Altogether, the robustness of fed-batch strategies in general depends on the reliabil-

ity of the considered yield coefficients. Since the oxygen yield is coupled to the dis-
similatory carbon flux by stoichiometry, it is expected that a low substrate yield is
accompanied by a low oxygen yield, e.g., in cases of high maintenance, decoupled
growth, or other factors which affect the energy flux of the cells. While the energy
generating processes may be more easy to calculate, the energy consuming processes
are subjected to unknown influences [10], e.g., unbalanced metabolism during re-
combinant protein synthesis. From the viewpoint of process control, YX,s and YX,02
do not contribute independent information regarding the actual growth rate since
both rely on a constant energy efficiency of the cells.
More reliable methods for estimation of growth processes could be based on spe-
cific uptake rates which are more closely linked to the anabolism of the cells. The
uptake of ammonia could be one possible candidate, because the nitrogen demand is
directly coupled to the biomass (e.g., protein, nucleic acids) synthesis reactions.
While no maintenance has to be considered, YX,NH3 depends on the cellular composi-
tion. The macromolecular composition may change during fed-batch cultivation [11]
and, especially, during the production phase to some extent, but, fortunately, this
does not effect the elemental composition very strongly. So for the purpose of bio-
mass estimation, the elemental balance

cx(t) = _ 1 ._MW-N .
eN,X MWNH3
1t0
QNH3,f(r)dr (12.31)

can be applied with sufficient accuracy. However, while the amount of ammonia fed
for pH-control can certainly be measured, it is neither the limiting substrate nor as
easily measurable on-line as the dissolved oxygen concentration. So a reliable esti-
mation based on ammonia consumption has to make sure that no acid formation
leads to ammonia accumulation at constant pH. This drawback can be circumvented
by employing estimators for acetic acid formation. So modern concepts of observa-
tion of HCDC should include, apart from the feeding strategies pointed out above,
an additional consideration of other yield coefficients e.g. YX,NH3 which will permit a
more reliable biomass estimation which is valid at different physiological conditions
[12] and, furthermore, the estimation of accumulation, e.g., of acetate [13], ammo-
nium, or carbon substrate, which will also give the possibilities for on-line optimiza-
tion of the feeding strategy.
12.6 A Case Study: Kinetics of Acetate Formation 387

The control task of complex fed-batch strategies has been discussed previously
(e.g., [14]; for a more recent review see also [15]). In most cases linear control with
constant parameters has been applied. This is in contrast to the exponentially chan-
ging system time constants and gains. Linearization of the process equations along
the exponential trajectory, e.g., by MAPLE, delivers a linear system, where the linear
system parameters depend on the actual biomass concentration. This offers the op-
portunity to use the biomass estimation for an adjusted linear controller, where the
control parameters are optimally set on-line according the process state employing
simple tuning rules. Furthermore, the actual kLa-value, which represents the main
time constant of the gas phase, can be calculated from the off-gas analysis and the
pOz, thus allowing for an adjusted linear controller as well.

12.6
A Case Study: Kinetics of Acetate Formation
and Recombinant Protein Synthesis in HCDC

Even during a well-designed high-cell density-cultivation, small amounts of acetate

may be produced (e.g., during the initial batch-phase or at the end of the recombi-
nant protein production phase). Therefore, it is of some interest to understand this
kind of by-product formation. The basic idea for modeling is the assumption of an
overflow metabolism which has been proposed, e.g., [16] and has been supported by
data from, e.g., [17].
If more glucose is taken up by the cell than can be processed in the oxidative path-
way, the glucose in excess is transformed to acetate to yield additional metabolic
energy, even if this is much less efficient compared to complete oxidation. This con-
cept has been already successfully applied to understand and model the Crabtree
effect in yeast [18].
The stoichiometric model (Eq. 12.1) can be modified by the inequality

r 02 < r 02,crit (12.32)

where r02,crit represents the maximum capacity of the respiratory chain. Conse-
quently, a critical catabolic substrate flux

r S,cat,crit < r S,cat (12.32a)

can be introduced according to stoichio metry. The specific acetate formation rate
rAce is then obtained from the catabolic mass balance as

rAce = YAce,S' (r s - r S,cat,crit)' (12.33)

which is valid in case of rs>rs,cat,crit and therefore r02=r02,crit.

The yield yAce,S,ferrn for acetate from glucose in excess is given by the stoichiometry
of acetate formation

1 mol glucose ---+ 2 mol acetate + 2 mol CO 2 as

YAce,S,ferrn = 2 MAce/MGlu ~ 0.666 gIg. (12.34)

Accordingly, an additional term YATP,Ace' rAce has to be added to the energy bal-
ance. Acetate inhibition [19] and possible inhibitions by small molecules at high
biomass concentrations [20] are not considered here.
388 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

From the viewpoint of downstream processing, a high intracellular protein con-

centration is at least as important as the achievement of high cell concentrations.
After induction, the protein is produced with an initial rate which depends on the
actual growth rate and, of course, on the expression system itself. In many cases, a
decrease of the production rate with time is observed. This may be due to intracel-
lular regulation or by other inhibition mechanisms leading to a decrease of gene
expression with a concomitant decrease of recombinant protein synthesis. In the si-
mulation study given below, an exponentially decreasing specific protein production
rate

rpr ot = fpr ot,ini . eTp·(t-t In

d)
(12.35)

is considered. In efficient expression systems, the amount of recombinant protein

produced is so high that a major part of cell energy is channeled into the synthesis
of this protein. This has been formulated in the concept of metabolic burden [21],
where it has been estimated that the amount of energy needed for the production of
protein is nearly twice as high as for the other cellular compounds (e.g., DNA, li-
pids). In the model presented here, this is accounted for by introducing another term
YProt,ATP . rprot into the energy balance, which finally reads

J.l =YX,s . rs + YX,Ace' rAce - YX,Prot' rprot - J.lm (12.36)

in the condensed parameter form. With these assumptions (Eqs. 12.32-12.36) in ad-
dition to the model shown above, simulation studies are presented in Fig. 12.7. using
data from [22], where exponential carbon substrate feeding has been employed with-
out further control and assuming constant yield and maintenance coefficients. The
model parameters related to acetate and product formation were estimated directly
from the data.
With these simple assumptions described above, the data are represented reason-
ably well (Fig. 12.7). During the batch and the first fed-batch phase (both at 30°C),

x c_Ace_Mess
80 • cSMess + c~ProCMess 8
c:::S:::Sim .c_Ace_Sim
.. c_X_Sim + c_Prot_Sim ... c_Prot_Sim
.............
.
,
....
,. •
. .-
,e
.' +
- .-• • -.-...... !II
" ,,- .. ;+
',.~.
'. ..... I 2
\, xl
.\ xl
o~~~~~~·~··~·~·~·~~~~~~~~~~~~o
o 6 12 18 24 30 36
Cultivation Time [hI

Fig. 12.7. Simulation of HCDC with temperature-induced production of an insulin B-chain fu-
sion protein; data are from [22] with process and model parameters given as cs,0=30 g!1,
VR,0=2S I, cs,£=0.6 gIg, J.lse,=0.12 h-1(bef. ind.), J.lset=0.08 h-1(after ind.); estimated parameters
are YX,s=0.48 gIg (bef. ind.), Yx,s=lJ.23 gIg )after ind.), Y3,Ace=0.OS g!p (assumed),
YX,Pro'= 1.80 gIg, YAce,S,ferm=0.666 gIg, rs,cri,=0.22 h , rprot,ini=0.OS4 h , 'tp=0.6 h
References 389

the system behaves as expected. Less good agreement of experimental data and mod-
el prediction after induction may result from the deleterious effects associated with
the overexpression of the heterologous gene. Beside the influence of the protein for-
mation on the energy flux of the cells - as considered by the production kinetics and
the energy balance - the accumulated protein itself may have an additional feed-back
influence on the metabolic activity, namely reduction of growth and protein synth-
esis itself [23]. Also, in addition to heterologous protein production, heat shock pro-
teins are being produced which can be considered as additional metabolic burden
[24]. However, these effects cannot be considered on the level of lumped parameter
models since the non-uniform distribution of recombinant protein in the population
makes the concept of a uniform biomass dubious.
In the cultivation shown above, the feeding rate suitable to allow for a specific
growth rate of ILset=O.12 h- I under non-inducing conditions was reduced simulta-
neously with the induction of recombinant protein synthesis according to
ILset=O.08 h -I assuming no change in the biomass yield with respect to the carbon
substrate [22]. However, the true growth rate and the true biomass yield were lower,
therefore leading to a lower biomass profile than expected. So more and more glu-
cose was fed per cell, and the specific substrate uptake rate increased and, finally,
reached the critical value which results in progressively increasing acetate excretion.
So, the experimental results here can indeed be modeled in accordance with the
concept of carbon overflow metabolism. Nevertheless, the critical specific substrate
uptake rate rS,crit during the induction phase is lower compared to those of the batch
and the first fed-batch phase, where no acetate production was observed at even
higher specific substrate turnover rates. Again, this should not be overinterpreted.
At least it indicates that an even better process performance can be obtained by
adjusting the feeding rate during the induction phase to the real energy demand of
the cells. This can be done, e.g., by employing on-line state estimation to eventually
occurring acetate formation. Other observations such as the increase of the oxygen
uptake rate after induction of recombinant protein synthesis are not considered here.
Yet it supports the assumptions of a high energy demand of the cells during the
protein production period. Furthermore, a decrease of the yield coefficient Yx,s and
an increase of maintenance ILm [25] cannot be distinguished because of the piecewise
constant growth rate not allowing for a linearly independent estimation. Altogether,
it has to be stated that much more work has to be carried out to understand the
influence of induction and protein production with respect to energy turnover and
intracellular regulation, thus allowing for an even improved coordination of gene
expression and process strategy.

References
l. Yamane T, Shimizu S (1984) Fed-batch techniques in microbial processes. In: Fiechter A
(ed) Adv Biochem Eng 30. Springer, Berlin Heidelberg New York
2. Luttmann R, Bitzer G, Hartkopf J (1994) Development of control strategies for high den-
sity cultivations. Mathematics and Computers in Simulation 37:153-164
3. Korz DJ, Rinas U, Hellmuth K, Sanders EA, Deckwer W-D (1995) Simple fed-batch tech-
nique for high cell density cultivation of Escherichia coli. J Biotechnol 39:59-65
4. Waldraff W, King R, Gilles ED (1997) Optimal feeding strategies by adaptive mesh selec-
tion for fed-batch bioprocesses. Bioprocess Eng 17:221-227
5. Matsui T, Yokota H, Sato S, Mukataka S, Takahashi J (1989) Pressurized culture of Escher-
ichia coli for a high concentration. Agric BioI Chern 53:2115-2120
390 12 Control Strategies for High-Cell Density Cultivation of Escherichia coli

6. Riesenberg D, Schulz v, Knorre WA, Pohl H-D, Korz D, Sanders EA, RoG A, Deckwer W-D
(1991) High density cultivation of Escherichia coli at controlled specific growth rate. J
Biotechnol 20: 17 - 28
7. Mori H, Yano T, Kobayashi T, Shimizu T (1979) High density cultivation of biomass in
fed-batch system with DO-stat. J Chern Eng Jpn 12:3l3-319
8. Cutayar J, Poillon D (1989) High cell density of E. coli in a fed-batch system with dissolved
oxygen as substrate feed indicator. Biotechnol Lett 11:155-160
9. Gollmer K, Posten C (1996) Supervision of bioprocesses using dynamic time warping
algorithm. Control Eng Practice 4:1287-1295
10. Russell JB, Cook GM (1995) Energetics of bacterial growth: balance of anabolic and cata-
bolic reactions. Microbiol Rev 59:48-62
11. Pramanik J, Keasling JD (1997) Stoichiometric model of Escherichia coli metabolism: in-
corporation of growth-rate dependent biomass composition and mechanistic energy re-
quirements. Biotechnol Bioeng 56:398-42l
12. Schmidt M, Viaplana E, Hoffmann F, Marten S, Villaverde A, Rinas U (2000) Secretion-
dependent proteolysis of heterologous protein by recombinant Escherichia coli is con-
nected to an increased activity of the energy-generating dissimilatory pathway. Biotechnol
Bioeng (in press)
l3. Akesson M, Nordberg Karlsson E, Hagander P, Axelsson JP, Tocaj A (1999) On-line detec-
tion of acetate formation in Escherichia coli cultures using dissolved oxygen responses to
feed transients. Biotechnol Bioeng 64:590-598
14. O'Connor GM, Sanchez-Riera F, Cooney CL (1992) Design and evaluation of control stra-
tegies for high cell density fermentations. Biotechnol Bioeng 39:293-304
IS. Riesenberg D, Guthke R (1999) High-cell-density cultivation of microorganisms. Appl
Microbiol Biotechnol 51:422-430
16. Majewski RA, Domach MM (1990) Simple constrained-optimization view of acetate over-
flow in E. coli. Biotechnol Bioeng 35:732-738
17. Han K, Lim HC, Hong J (1992) Acetic acid formation in Escherichia coli fermentation.
Biotechnol Bioeng 39:663-671
18. Sonnleitner B, Kappeli D (1986) Growth of Saccharomyces cerevisiae is controlled by its
limited respiratory capacity: formulation and verification of a hypothesis. Biotechnol
Bioeng 28:927-937
19. Nakano K, Rischke M, Sato S, Markl H (1997) Influence of acetic acid on the growth of
Escherichia coli K12 during high-cell-density cultivation in a dialysis reactor. Appl Microb
BiotechnoI48:597-601
20. Markl H, Zenneck C, Dubach AC, Ogbonna JC (1993) Cultivation of Escherichia coli to
high cell densities in a dialysis reactor. Appl Microb Biotechnol 39:48-52
21. Da Silva NA, Bailey JE (1986) Theoretical growth yield estimates for recombinant cells.
Biotechnol Bioeng 28:741-746
22. Schmidt M, Babu KR, Khanna N, Marten S, Rinas U (1999) Temperature-induced produc-
tion of recombinant human insulin in high-cell density cultures of recombinant Escheri-
chia coli. J Biotechnol 68:7l-83
23. Kurland CG, Dong H (1996) Bacterial growth inhibition by overproduction of protein.
Mol Microbiol 21:1-4
24. Rinas U (1996) Synthesis rate of cellular proteins involved in translation and protein fold-
ing are strongly altered in response to overproduction of basic fibroblast growth factor by
recombinant Escherichia coli. Biotechnol Prog 12:196-200
25. Bhattacharya SK, Dubey AK (1995) Metabolic burden as reflected by maintenance coeffi-
cient of recombinant Escherichia coli overexpression target gene. Biotechnol Lett 17:1155-
1160
13 ~-Lactam Antibiotics Production with Penicillium
chrysogenum and Acremonium chrysogenum
Karl-Heinz Bellgardt

13.1
Introduction

The history of antibiotics production began in 1941 following the discovery of Peni-
cillin by Fleming in 1928/29 and the first therapeutic applications in 1940 by Florey
and Chain. Then the productivity and product concentration was increased by sev-
eral orders of magnitudes by development of strains and improvements in cultiva-
tion media and process control. Surface culture was soon replaced by batch suspen-
sion culture in stirred tank reactors on complex media, followed by further improve-
ment with a two phase fed-batch strategy. This extended the length of the production
phase and avoided repression during high substrate levels. The industrial production
capacity is in the range of several tens of thousands of tons, a market of billions of
US$. In the production of secondary metabolites by filamentous fungi, such as the ~­
lac tam antibiotics, Penicillin by Penicillium chrysogenum, or Cephalosporin C by
Acremonium chrysogenum (Cephalosporium acremonium), many physiological, che-
mical, and technical factors influence the growth of the cells and product formation.
Small alterations in preculture, media, and operating conditions change the produc-
tivity in a seldom predictable way. On the basis of new and improved analytical
techniques, mathematical models can help to provide new insight into details, as well
as an integral view of the entire process, by putting together the knowledge of the
different levels from biology and cultivation to process control.
The morphology of the cells, which is a property of the strains but is also influ-
enced by the cultivation conditions, is an important parameter for antibiotics pro-
duction that is closely related to the fungal life-cycle. Antibiotics are often formed in
connection with hindered growth, aging and maturation of cells, or metamorphosis.
It is a general view that mainly non-growing morphological states contribute to anti-
biotics synthesis. Mycelial growth can range from filamentous to pelleted with any
combination in between [13-16]. Important processes during the growth of filamen-
tous fungi are shown in Figs. 13.1 and 13.2. After germination of the spores, the
hyphae grow by linear extension in the apical region of the active tips, and by
branching, into a complex network, the mycelial flocs. Their size is restricted by
breakage or fragmentation due to mechanical shear forces. But under proper condi-
tions dense pellets that can be rather stable may develop from hyphal filaments by
increase in spatial hyphal cell density (see Fig. 13.2), up to sizes in the millimeter
range. Agglomeration of spores before outgrow of filaments and attachment to solid
particles supports pellet formation. The pellet size is strongly influenced by shear
forces during preculture and in the production reactor [17, 18, 127]. High energy
392 13 ~- Lactam Antibiotics Production

Spores Mycelia
~~
Fragmentation

o Germination ~ ~
..........

Subapical cells
! Hyphal cells
Apical cells

Fig. 13.1. Processes at hyphal growth and morphological development of Penicillium chryso-
genum (septa are indicated)

Mycelium
Pellet

q
•
lormation

Growth D. Erosion
\}.
Chip-oft

c)

Breakage D Lysis in the center

Hollow pellet

Fig. 13.2. Processes at the development of Penicillium chrysogenum pellets

input supports small, smooth, and compact pellets by an erosion process, while un-
der mild shear stress the pellets can grow larger and develop a broad outer hairy
region. Small pellets often have a decreasing biomass density over the radius [10,
11], but generally the biomass density can pass through a maximum [7]. Large old
pellets are often hollow in the center as result of autolysis and can then easily break
apart. This is caused by transport limitation of substrate and mainly oxygen into the
inner zones. Measured pH and oxygen profiles by microprobes clearly support the
autolysis hypothesis [137]. The kinetics of autolysis were investigated under excess
and limitation of oxygen in [114, 115]. Although some autolysis was also observed in
the growth phase, it was clearly induced by limitation of substrate or nitrogen. Peni-
l3.1 Introduction 393

cillin degradation was then accelerated. Under additional oxygen limitation the onset
was more sudden but hyphal disintegration was less. After restoring proper growth
conditions, some regrowth of cells could be observed.
Morphology also indirectly influences production kinetics by determining the
rheology of the broth. Free filaments increase viscosity and hinder gas-liquid oxygen
transfer, which then limits cell density and productivity. Pellets result in low viscos-
ity compared to free filaments and thus facilitate oxygen transfer to the liquid phase,
but an additional transport resistance for substrates and oxygen from bulk liquid to
the inner parts of the pellets can become important. Experimental results show that
pellets up to a size of 200-400 JLm are optimal and have no negative effect on the
productivity, because the viscosity still remains low and there is no serious transport
limitation into the pellets as is observed for larger pellets. The mechanisms of the
inner-pellet transport are still under discussion. In [7, 16] it was found that with
molecular diffusion, turbulent and directed convective flows have to be expected in
the same order of magnitude. Diffusion seems to be the main mechanism in the
center of the pellets while turbulence penetrates only the outer parts. Generally, the
transport-reaction phenomena can be described by the dispersion model given in
Chap. 2. The rheology of pellet suspensions was also examined in a number of pa-
pers, e.g., [19, 83, 111]. The respective correlations should be included in a complete
process model to permit calculation of power supply and maximum oxygen transfer
as parameters for optimization.
The production strains are far from their natural ancestors, but the reason for
antibiotics synthesis might have been originally to provide some protection of the
cells against competing species in critical growth situations [37]. Secondary metabo-
lites such as many antibiotics are not part of the central metabolic pathways or of
immediate use for growth. Synthesis of p-Iactam-antibiotics is in a complex reaction
network coupled with catabolism and anabolism and subject to various kinds of
metabolic regulation. The biochemistry, metabolism and its regulation are reviewed,
e.g., in [1-4,54]. Maximum productivity is only possible in the presence of sufficient
amounts of carbon source and additional precursors in the medium. Sugars such as
glucose or lactate, dextrin, starch hydrolyzates, fatty oils, and other complex sub-
strates may serve as carbon sources for the cultivation. The disadvantages of oil are
increased oxygen demand and hindrance of the oxygen transfer by bubble coales-
cence.
Sugar substrates are catabolized via FDP- or pentose-phosphate-pathway, which
are both used in a similar order of magnitude. The metabolic flux distribution was
studied by J0rgensen et al. [116] during fed-batch cultivation for production of Pe-
nicillin V with a detailed stoichiometric model. The pentose-phosphate pathway was
used at 20% during rapid growth but up to 40% in the production phase. By addition
of the three amino acids precursors the productivity could be increased by 20%. In
another investigation of chemostat cultures, the pentose-phosphate-pathway was
even expected to be used up to 60% at high producing conditions [126]. The differ-
ence was explained by the higher specific production rate compared to fed-batch
cultivation. The theoretical maximum Penicillin V yield was calculated as 0.43 mole
per mole glucose. The antibiotics synthesis is subject to catabolite repression or in-
hibition during growth phases with high concentrations of easily convertible sugars,
or at high rates of carbon source uptake [4, 80].
Precursors for the antibiotics come from the metabolism of amino acids (see also
Figs. 13.5 and 13.10). Therefore, the kind of C- and N-sources and their amounts
394 l3 ~-Lactam Antibiotics Production

present in the medium have a strong influence on growth and productivity [5-7].
Soybean extract, yeast extract, peanut flour, or corn steep liquor (CSL) serve as or-
ganic nitrogen sources, beside ammonium. Addition of protein to the substrates im-
proves the productivity by freeing resources for product formation in the amino-
acid pathways.
The metabolic properties and kinetics determine the generally employed two-
phased control strategy for antibiotics production: in the first phase of the process
high growth rates are supported by an easily usable carbon source in batch mode or
under fed-batch with high feeding rate. The repression in the following production
phase can be either avoided by using carbon sources with lower maximum growth
rate compared to glucose, such as lactose or oil, or by limited feed of the carbon
source. In the first case, it is not necessary to establish a fed-batch strategy for the
production phase. But this is preferred because it extends the possibilities for control
of the process as further discussed in Sect. 13.4. Another and often the limiting fac-
tor for productivity is the oxygen supply to the culture because it sets bounds on the
maximum attainable growth rate and cell density. Furthermore, the cyclization reac-
tion of the tripeptide and the last steps of Cephalosporin C synthesis are directly
oxygen-dependent [2, 8]. Even short oxygen limitations can have lasting negative
effects on the product synthesis.
Summing up, the antibiotics production processes can be characterized by a diver-
sity of influencing factors originating from medium, growth conditions, and biolo-
gical properties of the strains. Thus the kinetics of production are much more com-
plicated than for primary metabolites as was already outlined in Chap. 2. The diffi-
culties in quantifying these effects and relating them to the morphological hetero-
geneity of the culture on one side, and to resulting variations in the productivity on
the other side, are the cause for the so-called bad reproducibility of the process.
Mathematical models can help to bring some light into the kinetics and mechanisms
on the cellular level and connect them to the observed macro-kinetics. Mathematical
modeling of antibiotics production covers all process levels beginning with the in-
tracellular reaction network, growth of hyphae and pellets in space and time, over
kinetics of substrate uptake and product formation, up to population dynamics for
mycelial flocs and pellets. Since production is closely related to morphological
changes, morphologically structured or segregated models are the proper approach
for modeling of production processes for 13-lactam-antibiotics. In the following, sev-
eral models for Penicillium chrysogenum and Acremonium chrysogenum are pre-
sented as examples on the different modeling levels. In a first step, these models
can improve our understanding of the observed macro-kinetic phenomena during
the cultivations, and of the interrelation of preculture, operating conditions, growth,
and production. In a second step, the models can be used to evaluate control strate-
gies and optimize the production process. The biochemical and morphological as-
pects of cultivations of filamentous fungi and their modeling are discussed in several
papers and reviews, e.g., [9-12,76,79,93,99,109,110,112]. The aim of this chapter
is to give by examples an overview of the modeling methods without looking at all
details of the models. For more information the reader is referred to the more spe-
cialized reviews on biological aspects and modeling, or to the original papers.
13.2 Modeling of Penicillin Production 395

13.2
Modeling of Penicillin Production

13.2.1
Unstructured and Simple Segregated Models

The complexity of models for Penicillin production has increased significantly over
the years. Most of the early models were purely unstructured and tried only to give a
formal description of the distinct growth and production phases. A simple model for
growth as logistic law was chosen by Constantinides et al. [20, 70]. The production
was assumed to be growth-independent but starting only after some dead-time due
to induction of product synthesis. Temperature dependence of growth and produc-
tion were modeled as polynomials or Arrhenius-type functions and used for optimi-
zation of batch cultivations. In a similar way, Calam and Russel [22] used a model
with extremely long lag-phase to explain the delayed product formation. In some
models an age dependence of the product synthesis is proposed to explain the de-
layed onset of production and its stagnation at the end of prolonged cultivations [33,
38].
Heijnen et al. [31] developed a model based on elemental and enthalpy balances,
extended by kinetic expressions for substrate uptake, maintenance, product forma-
tion, and product hydrolysis. The slowdown of the production at the end of the cul-
tivation is simply referred to a dilution effect due to the substrate feeding, and to
product hydrolysis. In the model, a direct coupling of growth and production up to
a critical growth rate of 0.01 h- 1 is proposed. Balancing methods were also employed
for computer control of cultivations. For this task, correlations of penicillin produc-
tion, growth, and maintenance metabolism to the COz-production is of special inter-
est [34-36].
A mechanistic model based on the Contois-kinetics for carbon source and oxygen
was developed by Bajpai and ReuB [28]. Yield variations are considered by a main-
tenance term in the balance equations. The unstructured model (see Table 13.1 for
model equations and Table 13.3 for parameter values) included a substrate inhibition
term in the rate of product synthesis to cover the low production during the growth
phase. The model was adapted by Menezes et al. [129] to industrial Penicillin G
pilot-plant fed-batch cultivations. Product synthesis was found not to be repressed
at high growth rate and described by Monod-kinetics, and an autolysis-term for the
cell mass was added. The uptake of components form complex substrates (CLS) was
dealt with by combining them with glucose to a representative single substrate. Due
to its simplicity, the model of Bajpai and ReuB served as a basis for a number of
further investigations, not only to improve the validity of the model, but also to use
it for extensive process optimization studies. Montague et al. [29] extended the mod-
el for carbon dioxide production and then applied it for on-line state estimation and
parameter-adaptive control purposes. Further extensions and improved parameter
estimation were given by Nicolai et al. [30] and van Impe et al. [125]. To overcome
model inconsistencies during batch and fed-batch simulations, the authors intro-
duced a variation in maintenance metabolism, endogenous metabolism, and yield
in relation to the substrate concentration. Hegewald et al. [27] incorporated into
their model not only the dependence of the growth and production kinetics on car-
bon source and dissolved oxygen, but also on nitrogen sources.
396 13 ~- Lactam Antibiotics Production

Table 13.1. Unstructured model of Bajpai and ReuB for Penicillin production [28]

Description Model equations

a Specialized vectors, r
considers maintenance
metabolism, see Eqs. (1.12) c=
and (2.2)

y=

b Intrinsic rate expression Growth

based on Contois-kinetics, Cs Co
substrate inhibition kinetics f-L = f-Lsmax KsCx + Cs KoCx + Co
for product
Production

rp = rpmax
Ksp
Cs
+ Cs
()
1+ * Co
KopCx + Co

c Explicit model equations dCx(t)

for the liquid phase, the -;It = -DCx(t) + f-LCx(t)
product balance considers dCp(t)
first order decay, ----;It = -DCp(t) + rpCx(t) - kDCp(t)
see Table 1.3
dCs(t)
dt
= D(t) (Cf(t) - Cs(t)) - (~+ rpCx(t)
Yxs Yps
+rns) Cx(t)
dCo(t)
dt
= OTR(t) _ (~+ rpCx(t)
Yxo Ypo
+rn o) Cx(t)

The morphological development of the hyphae cannot be described by simple un-

structured models. As shown in Fig. 13.1, the hyphae are only growing at their tip
that forms the extending end of the growth-active apical cells. It is believed that the
antibiotics are mainly synthesized in the non-growing subapical region, while the
elder parts of the hyphae are more or less inactive. This heterogeneity of the biomass
is more adequately described by segregated models or morphologically structured
models. Simple segregated models for the antibiotics production can be taken as a
first approximation to the inhomogeneities in the population, or the effects of mor-
phological differentiation and aging of cells. For cultivations in the presence of pel-
lets, the macrokinetics are also influenced by inhomogeneous biomass distribution
in the pellets, transport limitation of substrates, and alterations in the size and age
distribution of the pellet population. Growing cells are located mainly in the outer
layers of the pellets, while production is assumed to take place preferably by the non-
growing hyphae in the inner regions of the pellets. Also these effects can be de-
scribed in a first attempt by simple segregated models.
13.2 Modeling of Penicillin Production 397

In the morphologically structured model of Nestaas and Wang [39] the cell mass is
distinguished by three types: growing tips, non-growing but producing hyphae, and
degenerated non-producing inactive cells. Different sub-models were proposed for
growth and production phases. Product synthesis proceeds via a postulated inter-
mediate and is by this indirectly coupled to the specific growth rate; penicillin pro-
duction starts above a minimum growth rate. Beside endogenous metabolism of the
fraction of active biomass, a cell lysis mechanism is also used to explain variations in
the yield. In the model, the specific growth rate, estimated from experimental data, is
used as input to the balance equations instead of kinetic expressions in dependence
on the substrate concentration. Although the model is, therefore, not predictive, it
was used by Guthke and Knorre [40] to investigate the optimal control of repeated
fed-batch cultivations. The tendency of the predicted optimal control policy was in
agreement with experimental data. Kluge et al. [42] developed a segregated model as
extension of the ideas of Heijnen et al. [31] that considers lactose and lysed biomass
as additional substrates. For modeling of a lag-phase in product synthesis, a first-
order delay system for its induction at glucose limitation is introduced. The model
considers active and inactive biomass, as well as a temperature dependency of some
model parameters, and thus could be used for simulation of experiments with tem-
perature shifts or for temperature profile optimization.
Tiller et al. [47] presented a segregated model that was aimed for cultivations of a
highly-producing strain of Penicillium chrysogenum with pronounced pellet forma-
tion. In a first batch growth phase, a mixture of pharma-medium and glucose was
used as a substrate. During the fed-batch production phase, glucose was continu-
ously added to the culture to avoid catabolite repression of the production. In the
growth phase up to 90% of the total biomass were found as pellets with a diameter of
>250/Lm. Pronounced fragmentation and lysis of bigger pellets was observed in the
later phase of the cultivation and explained by aging effects and changes in the
macroscopic morphology. The model, summarized in Table 13.2, subdivides the en-
tire biomass into two fractions, as shown in Fig. 13.3, to take this morphological
change into account - growing and producing, Xl> and a non-growing but produ-
cing, X2 • The specific growth rates on glucose and pharma-medium are taken as
Monod-kinetics. It is assumed that cell lysis contributes to pharma-medium with a

Pharma -----I~~I
medium

Lysis

Fig. 13.3. Structure of the segregated model of Tiller et al. [47]

398 13 ~-Lactam Antibiotics Production

Table 13.2. Excerpts of the segregated model of Tiller et al. [47]

Description Model equations

a Specialized model vectors

(see Chap. 2.4.1 for CX2
explanation) c=
C')
Cs
CPM
Cp
Cx = CX! +CX2

(,;;; )
1 0

~l)
~1

Y1 = (-i-,0
XS
0
0
~YXftM
0
~YpSI
0
0
0
fl =
r12
0 0 0 ms
0

Y2 = (J-, ~l )
0
PS
~1

0
1
f2 = (i)
1 0

b Intrinsic growth rate Cs CPM

expressions /-Ls = /-LSmax Ks + Cs /-LPM = /-LPMmax KpM + CPM
c Metamorphosis reaction k12 = f12A
d Lysis of non-growing k1y =aly + blyA
hyphae
e Specific rate of sugar m=bm+amA
uptake for maintenance

J
t

f Mean culture age A(t) = Cxl(t) Cx(r)dr

0

g Specific product formation rp =

{ ~;;'Tpmn
/Ln
rpmax "
(2/-LPI ~ /-L)
for
/-L :::; /-LPI
/-LPI < /-L :::; /-LP2
/-L > /-Ln
0 /-L> 2/-Ln

yield of one. The amino acid and protein constituents of pharma-medium are an
important source of nitrogen during the cultivation. The rates of morphological dif-
ferentiation, r12' of lysis, rlY' and of the maintenance requirements, ms, are intro-
duced as age-dependent. Both types of cells are assumed to produce penicillin by a
specific rate rp. The balance also considers a product decay by a first order mechan-
ism with rate constant k[)o Product synthesis (see Table 13.2, row g) is maximum for a
region of moderate growth rates, while at strict limitation or above a critical level for
total catabolite inhibition, p,>2p,p2> no product is synthesized. The model was success-
fully applied for simulation of fed-batch cultures. Selected model parameters are gi-
ven in Table 13.3. The simulation results for substrate, cell mass and product concen-
tration are reported in Fig. 13.4. The growth phase during the experiment lasts for
13.2 Modeling of Penicillin Production 399

Table 13.3. Selected parameters of the models of Tiller et al. [47] and Bajpai and Reu6 [28]

Parameter Tiller et al. (Table 13.2) Bajpai and Reu6 (Table 13.1) Unit

!-lSmax 0.06 0.09 h- '

rPmax 0.0046 0.005 h- 1
Ks 0.07 0.15 g 1- 1/_
/-lPMmax 0.03 h- '
KpM 2.0 g I-I
0.47 0.45 -I
Yxs gg
YXPM 0.51 g g-I
Yps 1.2 0.9 g g-I
kD 0.0006 0.04 h- '
j.lPI 0.003 h- '
j.lP2 0.014 h- '
k12 0.00046 h- 2
am 0.001 0.014 (ms) h- 1
bm 0.00015 0 h- 2
aly -0.0008 h- '
bly 3 10-6 h- 2

~ e! ~.,-----------------------------------~

:G
:;:
l<l

l<l

:e
'"
Ci ~ J c.i . ./
!O! ~/
,£
0/
0 b,
~

40 60 120 100 200

t [hI

Fig. 13.4. Experimental data (symbols) and simulation results of a fed-batch experiment of
Penicillium chrysogenum: total biomass (ex), active biomass (Cxd, product concentration
(Cp), all in kgm- 3 ; glucose concentration Cs in gm- 3 • Reprinted from [47] "Segregated math-
ematical model for the fed-batch cultivation of a high producing strain of P. chrysogenum",
Copyright (1999), with permission of Elsevier Science

about 50 h. Glucose becomes already limiting at 30 h. After that point, the growth is
supported by the glucose feed. The transition to production coincides with the point,
where pharma-medium becomes limiting, because then the specific growth rate falls
below the upper critical value 2/-l P2 . The pattern of active biomass (eX!) showed a
striking correspondence to the pellet sieve fraction with diameter >250 /-lm, as
shown in Fig. 13.8a.
400 13 ~-Lactam Antibiotics Production

13.2.2
Biosynthesis Model of Penicillin V

For optimization of the Penicillin production a detailed analysis of the biosynthesis

of the antibiotics is necessary to understand its kinetics in dependence on the pre-
cursors, dissolved oxygen, and activities of the involved enzymes, or the accumula-
tion of intermediates and by-products. A mathematical model can help to identify
rate-limiting steps of the synthesis, which then may be attacked by metabolic engi-
neering. The biosynthesis pathway is shown in Fig. 13.5. Penicillin is synthesized in
several steps from a metabolite of the lysine pathway, L-a-aminoadipic acid; further
amino acid precursors are valine and cysteine. The cyclization step of the tripeptide
ACV (L-a-aminoadipyl-L-cysteinyl-D-valine) to Isopenicillin N is directly oxygen-
dependent. The following reactions may proceed by a one-step or two-step mechan-
isms via 6-APA. The precursors phenylacetic acid (for Penicillin G) or phenoxyacetic
acid (for Penicillin V) are incorporated into the penicillin molecule in the last synth-
esis step.
J0rgensen et al. [117] and Nielsen and J0rgensen [118] investigated the biosynth-
esis of Penicillin V. The Penicillin V biosynthetic pathway as shown in Fig. 13.5 was
described by the kinetics in Table 13.4 with model parameters given in Table 13.5.
Conversion of IPN to Penicillin V was assumed to take place either in the one-step
reaction, where the alpha-aminoadipate side chain is exchanged directly with the
precursor without release of 6-APA from the enzyme, or in a two step reaction via
free 6-APA. The kinetics of carboxylation of 6-APA to 8-HPA were assumed to follow
first order, since COr concentration was considered as constant. They found that in a
production phase of a fed-batch cultivation Isopenicillin N synthetase (IPNS) is flux
limiting, and afterwards the ACV synthetase [93]. In a culture without feeding of

lYAVC
L-a Aminoadipic acid L- Cysteine L-Valine

~ S",tretae

Tripeptide ACV (ACV)

02~l I, 13openi::illh N Syntletase

Cycl1se

Isopenicillin N (IPN)
J30pentillh N
Amidohyd:JO::ase

6-aminopenicillanic
acid (6-APA)

~
Penrillh
18 am ±lare

Phenylacetic Phenoxyacetic
acid (PAA) Ii; acid (POA)
CO2 Carbo,>,
atClll

S-hydroxy-
penicillic acid
Penicillin G (PEN G) (S-HPA) Penicillin V (PEN V)

Fig. 13.S. Penicillin biosynthetic pathway

13.2 Modeling of Penicillin Production 401

Table 13.4. Model for Penicillin V biosynthesis [117, 118]

Step Kinetics'

ACV formation by ACV synthetase

Isopenicillin N formation by IPN synthetase

Formation of 6-APA from Isopenicillin N by

Isopenicillin N Amidohydrolase (IAH)

Formation of Penicillin V from activated

side-chain precursor and 6-APA r - kX 1
4 - 4 AT 1 + K6APA -pOA +~
by Acyl-CoA: 6-APA Acyltransferase (AAT) C6APA CPOA-·CoA

One-step conversion of IPN to Penicillin V , - kX 1

5 - 5 AT 1 + KIPN-POA + K pOA
C1PN CPOA-CoA

Carboxylation of 6-APA to 8-HPA

Cleaving of penicillin to 6-APA and CPenV

k 8 X AT--c:...::.:.:-'----
'8 =
phenoxyacetic acid by Penicillin amidase CPenV + K penv
ax is the activity of the enzymes

Table 13.S. Parameters of the Penicillin V biosynthesis model

Parameter Value Unit

KIPN-POA 0.023 mmolr!

K pOA 0.006 mmolr!
K 1PN 4.0 mmoll-!
K6APA-POA 0.0093 mmoll-!
Kl 8.9 mmoll-!
Ko 0.13 mmoll-!
KAAA 0.63 mmoll-!
Kyal 0.3 mmoll-!
Keys 0.12 mmoll-!
K ACV 12.5 mmoll-!
KPenv 2.0 mmoll-!
kI 17.77 mmoll-! h-!
k2 16.85 mmolr!h-!
k3 4.03 mmoll-1h-!
k4 1.95 mmolr!h-!
ks 13.74 mmolr! h-!
k6 0.065 mmoll-! h-!
k8 0.4 mmolr! h-!
XACVs:PXI 615 mmoll-!h-!
XIPNS:PX2 330 mmoll-! h-!
X AT:PX3 420 mmolr!h-!
CPOA-CoA 0.05 mmolr!
402 13 /3- Lactam Antibiotics Production

precursors, amplification of both enzymes by genetic engineering would be advanta-

geous. It was also expected from simulation that oxygen concentrations above 45% of
saturation could increase productivity. This effect was further evaluated in a subse-
quent paper [128] with an extended model that was verified with experimental data.
The model indeed predicted a significant increase in specific penicillin production
rate. Above 50% saturation, ACV synthetase became the only rate-limiting step.

13.2.3
Morphologically Structured Models for Growth of Hyphae

Based on the early experimental results [13,24]' Righelato [25], Bull and Trinci [26],
and van Suijdam and Metz [10] proposed models for the development of mycelial
flocs and average pellet populations. Hyphae are only growing in length at the free
tips, as shown in Fig. 13.1. New free tips are formed by a branching mechanisms so
that their number also increases during growth. Synthesis of growth metabolites
mainly takes place in the apical region before the first septum. The subapical cells,
separated by septa, have a cellular composition very similar to the apical cells and
are considered as being metabolically active. The elder cells in the hyphal compart-
ment often have large vacuoles and are assumed to have lower metabolic activity.
The average hyphal length, the number of free tips, and the hyphal extension rate
were found to be proportional to the specific growth rate. An important parameter
for characterization of the filamentous morphology is the hyphal growth unit length
(lhgu) as the ratio of total hyphal length to number offree tips [100],

It total hyphal length

Ihgu = nt total number of tips (13.1)

Similar characteristic parameters can be defined by using the mass, Le., the hyphal
growth unit mass,

(13.2)

or volume of hyphae. Since their diameter and density are relatively constant, these
definitions are practically equivalent. Ihgu and mhgu are almost independent of the
growth rate, and therefore the linear extension rate and the branching frequency of
the hyphae are proportional to J1 [10].
A morphologically structured model for an Aspergillus strain was presented by
Megee et al. [23] in 1970. They discriminated the actively growing tips, active
hyphae (subapical), and further morphological states. Growth and differentiation
were described by kinetic expressions in dependence on the substrate concentra-
tion. Formation of growth and non-growth associated products was also modeled
in connection with the morphological states. Using the basic ideas of Megee et al.
[23], Paul and Thomas [46] presented, as extension of an earlier model [130], an
interesting structured-segregated model for submerged growth of Penicillium
chrysogenum filaments with far-reaching experimental verification by data from
image-analysis. In this model the vacuole formation is considered as an important
physiological process during growth and aging of hyphae. The model divides the
biomass into basically three distinct regions according to the activities and struc-
ture of hyphal compartments: these are actively growing tips, non-growing Peni-
cillin-producing regions, and degenerated or metabolic inactive regions that can
13.2 Modeling of Penicillin Production 403

be subject to fragmentation and autolysis. The non-growing region is considered

to consist of cytoplasm and vacuoles. Growing tips are transformed by septation
into non-growing cells that initially contain no vacuoles. Under substrate-limited
conditions vacuole formation is initiated; these grow in size until they fill the
entire cell volume and finally lead to the degenerated form. The vacuole size dis-
tribution in the hyphae is described by a population balance equation. Penicillin
production is assumed to take place in the cytoplasm at low substrate concentra-
tions. The differentiation phenomena and product formation are represented by
kinetic equations in dependence on the substrate concentration. Unknown model
parameters were estimated from experimental data for the development of cell-
dry-mass equivalents of the four morphological types during fed-batch cultiva-
tions. This quantitative information was obtained by methods of digital image
analysis. The model was in well agreement with the experimental data and had
remarkable predictive capabilities for a number of fed-batch cultivations with
varying sugar-feeding strategies. The highest product concentration was obtained
by a feeding profile with gradually decreasing glucose-feed during the production
phase.
A detailed hyphae model was proposed by Aynsley et al. [43] as an extension of the
ideas of Prosser and Trinci [44]. The hyphae are considered as self extending tubular
reactors, where substrate diffusion into the cells take place all over the hyphae. The
substrate is converted into growth-precursor containing vesicles, which are trans-
ported to the tip. The flow of vesicles determines the tip extension rate, and their
level the branching frequency. Fragmentation was assumed to occur mainly at sub-
strate limitation. The model could be fitted to experimental data of total biomass and
hyphal growth unit.
A morphological model was combined with a population balance model by Niel-
sen [45] to analyze morphological data from different batch and continuous cultiva-
tions of filamentous microorganisms in more detail, with respect to the growth me-
chanisms tip extension and branching, as well as fragmentation. The model consid-
ers three compartments, growing apical cells, subapical cells which still participate
directly in the tip extension process, and hyphal cells that are inactive with respect to
the growth process but provide material for the growth of apical cells by contribut-
ing to the stream of vesicles. Branching points are assumed to be located at the sub-
apical region. It is an important point for the description of submerged cultures that
the model discriminates between active and inactive tips since fragmentation is a
frequent event under agitated conditions. If one assumes that practically every my-
celia floc was subject to fragmentation, it must contain exactly two inactive tips
which add to the measurable total number of tips. On average, the total number of
tips and the number of active tips are related by

(13.3)

The morphological model is complemented by balance equations for the average

number of hyphal elements and actively growing tips as well as the hyphal mass.
These equations are derived from a distribution model that is summarized below.
The dynamic balance for the hyphal growth unit mass was found to be

(13.4)
404 13 ~- Lactam Antibiotics Production

where <I> is the specific branching frequency and '¥ the specific rate of fragmenta-
tion. For most species, the hyphal growth unit mass is constant when no fragmenta-
tion occurs. It can then be calculated from the average specific growth rate as

(13.5)

and the rate of tip extension is then proportional to the average specific growth rate,
(13.6)
The model was in good agreement with experimental data of continuous cultures
of Penicillium chrysogenum. It was also found that the specific rate of fragmentation
is linearly correlated with the energy input, regardless of batch or continuous opera-
tion. The morphological model was extended by kinetics for growth and product
synthesis by Zangirolami et al. [98] under the assumption of three hyphal sections
(see Chap. 2, Table 2.14): growing apical tips, subapical sections with glucose inhib-
ited product formation, and inactive parts. A complex substrate (CSL) was consid-
ered by an equivalent amount of glucose. The model was successfully compared to
experimental data of fed-batch and repeated fed-batch cultivations by simulation.
The macroscopic models for hyphal growth on average can easily be compared to
experimental data. But for in-depth theoretical studies and more accurate investiga-
tion of the microscopic morphology and related mechanisms, population balance
models are a proper approach. Takamatsu et al. [136] developed a discrete distribu-
tion model for growth of Aspergillus niger in the form of mycelial flocs. The popula-
tion is characterized by a number density function for flows with a certain number
of cells and branches. The model included the mechanisms of branching and break-
age and was used to establish a simplified segregated model of the population. An-
other morphological model for growth of hyphae based on the population balances
of Ramkrishna (see Chap. 2, Sect. 2.4.6) was published by Nielsen [95, 113] and re-
cently worked out further [97]. Hyphal elements are characterized by their total
length It and number of tips n. Their number density f is given by the population
balance
8f(lt, n, t) 8(qtip(lt, n,z)nf(It, n, t)) 8(qbra(Zt, n, z)f(lt, n, t))
--'---'-:---'---'- + + --'-''---'-----.,.-'..:...-'---'--'-
8t 8lt 8n (13.7)
= hg(lt, n, z, t) + hjra(lt, n, z, t) - Df(lt, n, t)
where qtip is the average tip extension rate, and qbra the branching frequency. Net
formation of hyphal elements is by spore germination with rate hg and fragmenta-
tion with rate hfra> both also depending on the conditions in the bioreactor denoted
by the general vector z. It is assumed that by spore germination newly formed hy-
phal elements have fixed properties ng and ltg, and fragmentation results in binary
fission, then
(13.8)

JJ
CXJ CX)

hf(lr, n, z, t) = 2 qjra (l; , n', z)p(l;, n', It, n)f(l;, n', t)dndl t - qjra(lr, n, z)f(lt, n, t)
n It

where qfra is the breakage function, p the partitioning function, and g spore germi-
nation frequency. The latter was fitted to a B-distribution to estimate the spore via-
13.2 Modeling of Penicillin Production 405

bility and germination time interval from experimental data. From the above model
balance equations for macroscopic population parameters, the total number of hy-
phal elements, e, average number of tips, naJ<' and average hyphal length, laJ<' as de-
fined by

JJ
00 00

e= Itf(lt, n, t)dltdn
ng itg

~ JJ
00 00

Itav = Itf(lt, n, t)dltdn (13.9)

ng lrg

~ JJ
00 00

nay = nf(lt, n, t)dltdn

ng 11g

were derived for the situation when all spores have germinated as follows:

de
dt = (qjra(ltav, z) - D)e
dl
tav
dt = navqtip(ltav, z) - qjra(ltav, z)ltav (13.10)
dn
av
dt = qbra(ltav, z) - qjra(ltaVo z)nav

The following kinetic expressions were in well agreement with experimental data:

() Cs Itav
qtip Itav , z = ktipmax Cs + Ks Itav + K/ (13.11)

0 for Itav < 44 fLm

qbra(ltav, z) = { k
bra
()I
Z tav
fior Itav ~ 44 fLm with fL = Jktipkbra
0 for leav < leeq nay-+-1 ( I -I )
. leav = -
qfra(ltav,Z)= { kfra ()(12
Z eav -[2)
eeq fior leav ~ leeq
wIth
2nav
tav tip

where kjra was correlated to the specific power input of the reactor and geometric
parameters. The average effective length of hyphal elements is leaY' Growth and frag-
mentation of single hyphae were described by a modified population model that was
solved for steady states by Monte Carlo simulation [96]. It was shown that it is pos-
sible, from measurements of the number of tips and total hyphallength, to discrimi-
nate between four sets of models consisting of two different fragmentation kinetics
and two partitioning functions.
Turbulent breakage of hyphae in mechanically stirred bioreactors was also inves-
tigated by Shamlou et al. [132] and found to be approximately of first order, i.e., the
mean main hyphallength could be described by

(13.12 )
406 l3 [3- Lactam Antibiotics Production

where leq is the equilibrium length. The rate constant k was found to depend on the
product of mean energy dissipation rate and reciprocal of the impeller circulation
time.

13.2.4
Models for Growth of Fungal Pellets

Similar to the models for growth of hyphae, the early models for pellet populations
considered the development of pellets only on average assuming uniform size and
density. The growth of a pellet in mass can in a very simplified way be described by
the cube-root law under assumption of constant radius growth:

mp(t) = ( m!(o) + kt) 3 (13.13)

The kinetics correspond to the following volumetric and specific rate equations for
growth of cell mass:
2 1
Qx = 3kcI f-t = 3kC? (13.14)
The growth of fungal pellets on average with consideration of oxygen transport
limitation can be described by simple unstructured models in connection with the
diffusion equation, Eq. (2.9) in Chap. 2 [10, 12, 14,41, 131], assuming constant bio-
mass density within the pellet. This allows one to estimate the penetration depth of
oxygen, which determines the active portion of a pellet, and its total oxygen and
substrate turnover. For very dense pellets, the dispersion model has to consider the
maximum possible cell density that is limited by two mechanisms: available space
for hyphae and reduced free outer-cell volume for transport of substrates. Both ef-
fects can be modeled in analogy to the dusty-gas model by a cell mass dependent
diffusion coefficient, Eq. (2.13) in Chap. 2.
Pellet breakage and disruption due to mechanical sheer stress as a consequence of
high energy input for agitation of the cultivation medium was studied by van Suij-
dam and Metz [12) in stirred tank reactors. They derived the following theoretical
correlation for the mean pellet diameter under shear stress that was in good agree-
ment with experimental data:
dp = kN- L2 d;?;8 (13.15)
where dp is the pellet diameter, N the stirrer speed, and dStir the stirrer diameter.
In order to proceed toward a better understanding of influencing factors for
growth and product formation of Penicillium chrysogenum under the presence of
pellets, Meyerhoff et al. [49) developed a detailed model for the growth of single
hyphae up to mycelia flocs and pellets. The model adopted the basic concepts of
Yang et al. [50, 51] for the growth of Streptomyces tendae. This is a combined me-
chanistic-stochastic approach for the description of the three-dimensional develop-
ment of mycelia by linear extension of hyphae through apical growth, septation, and
branching. But in addition, kinetics for growth limitation by space, substrate or oxy-
gen, cell inactivation, or lysis under starvation condition, e.g., in the center of pel-
lets), and mechanisms for erosion of hyphae at the surface of pellets were intro-
duced. Growth of hyphae is determined by a key-compound that is synthesized from
substrate transported into the cells through the outer wall along the branches of the
13.2 Modeling of Penicillin Production 407

hyphae. The event of septation and location of the septum are also controlled by the
concentration of the key-compound. The linear extension rates of the tips are deter-
mined by diffusion of a key-compound along the branches to the tips, where it is
used to build new cell material. The diffusion equation can be solved analytically
and gives for the quasi-stationary case the maximum linear extension rate of a com-
partment with one tip:
f3e 2d1 + 'Y (13.16)
al,max = a max f3 e21<L 'Y
1_

with known auxiliary variables ~, y, and K that depend on the geometry of the
branches [49]. The events of branching and the spatial direction for growth of new
tip-sectors are given by probability functions in the stochastic part of the model.
Transport processes of substrate and oxygen from the bulk medium to the center of
the pellet are described by the diffusion equation, Eq. (2.9) in Chap. 2, with cell mass
dependent effective diffusion coefficient. The growth limitation mechanism for the
linear extension rates of individual tips,
Co Cs
ai = ai,max Ko + CoKs + Cs (13.17)

is assumed to follow Monod-kinetics for each limiting compound, oxygen, and sub-
strate. Cell inactivation with subsequent lysis is assumed to take place when the sub-
strate concentration falls below a critical value. Shear forces are an important factor
for the development of the pellets that limits the growth in diameter by the erosion
process. It was assumed that only hyphae growing out from the denser parts of the
pellet are subject to breakage according to the probability function

(13.18)

where rthr(Pthr) is the critical radius above which breakage can occur, rtip is the
radial position of the tip, and AShear is the characteristic parameter of the probability
distribution that can be varied to adjust the model to experimental conditions. The
critical radius is determined by the local hyphal density Pth,. Simulations of a single
pellet have been carried out with the model that were quite realistic by visual inspec-
tion. Under mild shear conditions, simulated pellets developed a broad and more
hairy outer region than the compact and dense pellets resulting under enforced shear
stress; these can have a very smooth surface. In the simulation studies, the cell den-
sity within the pellets approaches a maximum volume fraction of about 0.4. After the
cell density has reached this maximum, the active growth zone keeps a thickness of
about 200/-Lm since no oxygen reaches the inner parts due to transport limitation,
and the density profile then moves more or less only outside without altering the
shape or maximum value.
The detailed model above is not well suited for more far reaching simulation stu-
dies of bigger pellets or pellet ensembles, because the demand for computer re-
sources is tremendous. Therefore, the model was simplified by neglecting individual
hyphae and looking at average morphological quantities, while keeping the morpho-
logical basis and information on the microscopic structure of the pellets [52]. In the
modified model the pellet is considered as having spherical symmetry and averaging
is done in small radial layers. The simulation proceeds in discrete time steps. The
simplified model considers only the total length of hyphae within a layer as repre-
408 13 [3- Lactam Antibiotics Production

sentative variable. The hyphal growth unit length is then reconstructed using a cor-
relation derived from the detailed model as a function of the radius by

1 360p,m - 2rR
hgu = 70p,m + ~ (13.19)
3 L: n tR
R

where ntR is the total number of tips including inactive ones. The number of tips nR
in the radial layer is then evaluated from Eq. (13.1). Similarly, the linear extension is
only calculated on the average. The total hyphal length increment in the simulated
pellet layer by growth is

f::,l+ {D:maxnRr(CO)r(Cs) p < 0.6, Cs > CS,Crit, Co > CO,Crit

~= D:maxnRr(CO)r(CS) (0.79 - p) where p ~ 0.6, C s > CS,Crit, Co > CO,Crit
f::,t
-D:lynR Cs :::; CS,Crit or Co:::; CO,Crit
(13.20)
The first condition is for weakly limited growth while the second also considers
spatial limitation; the most dense packing of cylinders of equal diameter is 0.79. The
third condition describes reduction of the active hyphal length by inactivation at
growth limitation. The inactive parts could then be subject to lysis. Length reduction
by pellet erosion is given by
f::,lffi.
f::,t = -lShearnShearR (13.21)

where the number of broken tips, nShearR is calculated as in the detailed model, and
the average mycelia length loss per break event, lShean is a model parameter.
Simulation results of the simplified model were compared to microprobe measure-
ments for glucose and oxygen profiles as described in [137], and to cell density pro-
files obtained by digital image analysis of microtome preparations of a pellet. Fig-
ure 13.6 depicts the simulated concentration profile of a small pellet by using the cell

100 2.09 0.5

0--
y--
p-
OD §0.4 0--
2.04 '.;l 0-
g 0-
<l:l "
60 0,-
00. ]0.3
rJ 1.99 0 0
- CJ--Ll---G---~-- 0
0, 40
0 >
'i'J 0.2
u o --
0
- .cJ-- d
_0--

0--'/--- __ -i:( 9-
1.94
20
I 0.1 ---_ ... -- 0
0 0

1.89 O~rr~TT~"rr~TT~"'-~~~~~
o 0.05 O.lO 0.l5 0.20 0.25 0.30
radius [nun]

Fig. 13.6. Comparison of radial oxygen (p02 in %) and glucose profiles (Cs in kg m-3) within a
small pellet as obtained by microprobe measurements (symbols) to simulation results (lines)
by using the measured cell volume fraction, 216 h after germination; parameters:
u max =5jLmh-I, IShear=200jLm, Uly =50jLmh-I, Cocri,=8 lO-Sgl-l. Reprinted from [52]
13.2 Modeling of Penicillin Production 409

100 0.5
,yell volmne fraction
80 0.4 ' ," ,,,
,,
......

60 ·1 0.3 ,
00.. ¢:'" ,

40
"] 0.2
0
> !
'5
u 0.1
20 --

a
0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.4 0
radius [nunI

Fig. 13.7. Simulation results (dashed lines) for radial profiles of dissolved oxygen (p02) in %)
and cell volume fraction, the latter in comparison to experimental data (symbols with interpo-
lating line) as obtained from digital image processing; de novo simulated pellet after 115 h of
shake flask cultivation; parameters: a max = 18 j.Lm h-I, IShear=214 j.Lm, aly =50 j.Lm h-I, COcri'= 10-5 g
I-I. Reprinted from [52]

volume fraction from image analysis. The simulation can reproduce the measured
profiles quite well. In a second step it was tested whether the model yields realistic
cell density profiles. Shake flasks were inoculated with spores and cultivated up to
pellets under mild shear forces. Pellets were taken out at different time and the cell
volume fraction determined. Simulations over an identical time interval yielded very
similar cell density profiles, as shown in Fig. 13.7 by one example.

13.2.S
Models for Growth of Pellet Populations

In a bioprocess the pellet population is never homogeneous but shows a wide dis-
tribution in shape, size, density, and growth activity that all influence the macroki-
netics of the process. This is due to the stochastic nature of the important processes
during the pellet development as shown schematically in Fig. 13.2: germination, ero-
sion, and breakage events follow probability functions instead of being exactly de-
termined in every single occurrence. The unequal development can be further pro-
nounced by transport limitation of substrates into the inner regions of the differently
developing pellets. The inhomogeneity can be described by population models with
continuous variation in characteristic variables of a pellet.
A purely theoretical model for growth of pellet populations was developed by
Edelstein and Hadar [134], later modified by Tough et al. [135] in the growth equa-
tion, and translated into a finite elements description that generated more stable
solutions in the pellet size distribution. Pellet fragments after breakage were as-
sumed to have all the same size; chip-off of hyphae at the outer region of the pellet
was not considered. Nielsen [95] presented a population balance model in analogy to
the model of hyphal development given in Sect. 13.2.3. The pellets are characterized
410 13 13- Lactam Antibiotics Production

by their diameter dp' The number density of pellets fp with this diameter is described
by the population balance
afi,( dp , t) a( (qgro (dp, z) - qpfra (dp , z))fi, (dp, t))
at + adp (13.22)

where qgro is the rate of increase of pellet diameter, qpfra the rate of hyphal fragmen-
tation (erosion) at the pellet surface, hfor the net formation rate of pellets, and h bre
the loss rate due to breakage. Functional expressions were not suggested. In analogy
to hyphal growth, it was proposed to describe the growth in average pellet diameter
as
d(dpav )
-----;It = qpgro ( dpav , z) - qpfra ( dpav , z) (13.23)

with the kinetic expressions

qpbre (1tav, Z ) = k bre ()d

3.2 N6.65d8.72
Z pay Stir
for dpav < dpeq · (13.24)
qpjra(ltav, Z) = { kO ()(d2 _ d 2 ) for
pjra Z pay peq dpav ;::: dpeq

where kpjra depends on the specific power input of the reactor and geometric para-
meters. The equilibrium diameter, d peq , was correlated with the specific power input
[12],

dpav = djrag (Vp)-k (13.25)

The agitation induced fragmentation of pellets was successfully described by Jiis-

ten et al. [94] by a population balance model founded on detailed experimental data
from image analysis [133]. Correlations to agitation parameters were also developed
on this basis. The projected area Ap of a pellet was used as characteristic variable in
the population balance for the number density function f(Ap,t),
af(Ap, t) a(j(Ap, t) (-k) (Ap - A~))
(13.26)
at aAp
The breakage mechanism is of first order in some driving force with fragmenta-
tion rate constant k. Driving force is the size difference in the actual and the mini-
mum projected area AP' of a pellet, below which no breakage occurs. The rate con-
stant was found to depend only on the agitation conditions and could be correlated
to impeller tip speed, specific power input, and the "energy dissipation/circulation"
function [132]. No direct significant effect of fermentation conditions was found.
The simulated distributions were in very nice agreement with the experimental data.
In an investigation on pellet morphology with Aspergillus awamori, Cui et al.
[138] found that pellet breakage is not important. The effect of agitation influenced
mainly the rate of hyphal chip-off in the outer hairy zone of the pellet. This mechan-
ism restricts the size of pellets and increases the density of pellets as well as the mass
of free hyphae. Nevertheless, when hollow pellets are formed due to transport limita-
tion, pellet breakage seems to be more likely.
13.2 Modeling of Penicillin Production 411

The simplified pellet model of Meyerhoff and Bellgardt [52] presented in

Sect. 13.2.4 also provides a tool for an integrated, morphology-based simulation of
fed-batch processes in the presence of pellets with non-uniform property distribu-
tion. In [53] the distribution function was approximated by simulation of up to 100
different pellets that represent all of the pellets in the cultivation. The simulation
started from the very beginning of the process, i.e., inoculation of spores in the pre-
culture. This opens the additional chance of studying the influence of variations in
the pre culture conditions on the main cultivation. For this task, the model of [52]
was extended by a description of the germination time distribution of spores. A
modified truncated Gauss-distribution

<I>(lnx) = vz;r
1 J
Inx

e-,-dlnx
-Inx'
(13.27)
o

was found to be in agreement with experimental data, where the actual germination
time tg is

t = (,61nx + 1') tg max, x > 0, tg 20 (13.28)

where ~ and yare model parameters. Since the observed pellet size distribution
could only be explained by assuming breakage events, this mechanism was also in-
cluded in the model. Breakage events are influenced by the actual state of the pellet
with respect to cell density profile, size, and age. Although the model provides this
information, a simple correlation for the breakage probability of a pellet, Pbre, in a
time interval M was introduced for the beginning:

(13.29)

The model was applied to simulations of pellet size distributions in shake-flask

cultivations in the presence of glass beads under varying conditions and for inter-
pretation of the data. Also complete fed-batch cultivations were simulated. For the
process shown in Fig. 13.4, the simulations of the usual process variables based on
this pellet population model were quite close to the segregated model. The simulated
pellet size distribution is compared to experimental data in Fig. 13.8. At the beginning
there is an exponential growth phase with significant formation of pellets with dia-
meter >250 Mm. During the production phase with only slightly increasing total bio-
mass, the formation of bigger pellets slows down at first. Later on, after about 100 h,
the pellets begin to disappear due to breakage and autolysis. After 140 h, there re-
main only filaments and pellet fragments with small diameter. The simulation
showed that the pellets are most susceptible to oxygen limitation in the early phases
of the cultivation although the dissolved oxygen concentration remains high. The
reasons for this effect are the high growth rate and large mean pellet diameter at
the beginning of the cultivation. Afterwards oxygen transport limitation into the
pellets plays no role because of the reduced growth rates and decreasing mean dia-
meter.
412 13 13- Lactam Antibiotics Production

30
b
:5
:0
:5
.,
:0
':0 -250)-:':1
5

0
100 ::0 :'::0
IIIlJ
Fig. 13.S.a Experimental data for pellet sieve fractions. b Simulations results of the pellet size
distribution for the cultivation in Fig. 13.4. Parameters: u max=20 J-lmh-l, IShear=214 J-lm, Uly=
20J-lmh-1, AShear=O.1. Reprinted from [53] "A morphology-based model for fed-hatch cultiva-
tions of P. chrysogenum growing in pellet form" Copyright (1999), with permission of Elsevier
Science

13.3
Modeling of Cephalosporin C Production

The morphology and metabolism of Acremonium chrysogenum deviates somewhat

from Penicillium chrysogenum which has consequences for the modeling of the pro-
cess. In submerged cultures A. chrysogenum can be found in three main morpholo-
gical types as shown in Fig. 13.9: slim thin-walled filamentous hyphae (mycelia),
swollen thick-walled hyphae or hyphal fragments, and arthrospores. The organism
has no clear tendency to pellet formation. At high growth rate mainly slim hyphae
and a low percentage of swollen hyphae are formed. Gradual limitation of the carbon
source accelerates the transition from the slim filamentous to the second thicker
form. Under growth limitation this develops septa and can subsequently break down
to arthrospores. Cephalosporin C (CPC) production is usually maximum in process
13.3 Modeling of Cephalosporin C Production 413

Hyphae Swollen hyphal

Arthrospores
fragments

~ 00

J~
y y 000
00
0 0

Fig. 13.9. Morphological development of Acremonium chrysogenum

phases with a high rate of differentiation from swollen hyphae to arthrospores [2,
56-58, 77]. Thus it correlates with a high number of swollen hyphae as well. The
arthrospores can germinate and grow out to new hyphae, but seem to be non-produ-
cing.
Similar to Penicillin production, high productivity can be obtained by a two-phase
control strategy with clearly distinguishable growth and production phase. Oxygen
limitation has to be avoided and product formation must be supported by sufficient
feed of substrate. The CPC-production is stimulated by addition of methionine dur-
ing the growth phase, but inhibitory high concentrations or overfeed is contra-pro-
ductive [57,68]. Methionine is converted to cysteine [1] and can supply sulfur to the
CPC-pathway, inducing its enzymes as well as stimulating morphological differentia-
tion from thin hyphae to produce swollen hyphae. During later phases of the culti-
vation, the production slows down and it must be stopped before the product degra-
dation surmounts production. In all, 25-40% of the product can undergo such hy-
drolysis, which follows first order kinetics [2]. Another problem of the cultivation is
the accumulation of undesired precursors of the Cephalosporin C synthesis, e.g.,
deacetylcephalosporin C, which is rather stable.

13.3.1
Biosynthesis of Cephalosporin

The synthesis steps of Penicillin and Cephalosporin C up to Isopenicillin N are

equivalent, as is shown in Fig. 13.10. This is then converted to Penicillin N without
precursor, and in a further oxygen-dependent reaction sequence to Cephalosporin C.
Under oxygen concentrations below 20% of saturation, these further steps are inhib-
ited and Penicillin N is accumulated [106]. Glucose and other rapidly utilized carbon
sources repress the ring-expansion enzyme deacetoxycephalosporin-C-synthetase
and inhibit the first enzyme of the CPC-pathway, the ACV synthetase. The ring ex-
pansion enzyme is not inhibited after induction [55], but may be rate limiting [107,
108].
Malmberg and Hu [59] presented a detailed kinetic model for analysis of rate lim-
iting steps in the CPC biosynthetic pathway under conditions of constant dissolved
oxygen tension. The model, given in Table 13.6, includes the six reaction steps to
Cephalosporin C, starting with the amino acids L-a-aminoadipic acid, L-cysteine,
and L-valine. Cell growth was simulated according to a formal balance with fixed
specific growth rate, and loss of Penicillin N into the medium was considered. In
vivo data of enzyme activities were converted to the intracellular conditions required
for the model simulations. From the simulation results the authors conclude that the
414 13 ~- Lactam Antibiotics Production

L-aAminoadipic ac~cI~ L-Valine

Ave Synthet.s!e

Tripeptide ACV (ACV)

O~:rcYCl Cyclise

Isopenicillin N (IPN)

o
l
Penicillin N (PEN N)
Epm e:ase

K ~ ~
a- etog utarate ~ tI DAOCsyntlu;t<llie
Ring expansion

Deacetoxycephalosporin C (DAOC)

a-Ket~~utar~
.... ~....
IlHYd
'" ...
Expandase;
HydlOxyhse

Deacetylcephalosporin C (DAC)
Acetyl-CoA ________ I
.. ... Acet}l:rnsfera=

Cephalosporin C (CPC)

Fig. 13.10. Cephalosporin C biosynthetic pathway

Table 13.6. The Cephalosporin biosynthesis model of Malmberg and Hu [59)

Description Model equation

Intrinsic tripeptide ACV dCACV _ k CAAA ceys Cval

- - - ACV
dt KAAA CAAA Keys +
Ceys Kval Cval + +
CACV
- k cycl K
ACV
+ CACV - 11,cACV

dCPenN _ 1 (k CACV k CPenN

Intrinsic Penicillin N -- - - Cycl - Exp
dt 2 K ACV + CACV K penN + CPenN

- kSE (CPenN - kCpenN») - J.LCPenN

Intrinsic deacetoxy- dCDADC _ k CPenN k CDADC

-d-- - Exp - Hyd - J.LCDADC
cephalosporin C (DAOC) t K penN + CPenN K DADC + CDADC

Cephalosporin C dCcpc _ k CDADC C

----;It - P Hyd K DADC CDADC+ X

External Penicillin N dC PenN

~ = (
pkSE CPenN - kCPenN
)
Cx

ACV -tripeptide formation catalyzed by the ACV -synthetase is the rate limiting step
of CPC-synthesis. Increase of the intracellular level by genetic engineering methods
should enhance the production.
13.3 Modeling of Cephalosporin C Production 415

13.3.2
Simple Cybernetic Model for Growth and Production on Sugar and Soy-Oil

An expert system for modeling and control design of biotechnical processes [104]
was used for modeling of Cephalosporin C production on sugar syrup (main com-
ponent maltose) and soy-oil [105]. The process is divided into a growth phase on
sugar with parallel uptake of soy-oil and moderate production, and a second produc-
tion phase on soy-oil only with continuous feeding. According to the data analysis
and knowledge acquisition by the expert system, the cybernetic metabolic regulator
model (see Chap. 2.3.4) was established as summarized in Table 13.7.
The initial lag-phase on sugar and the diauxic lag-phase for adaptation from
sugar substrate (C s) to soy-oil (COil) are modeled by metabolic regulators, Eq. (2.85)
in Chap.2, for rlmax(t) and r2max(t); see Table 13.7, row d). Limitation terms for
soy-oil or oxygen were not considered because both concentrations were kept
above the critical values. The production of Cephalosporin C with specific rate
rp was assumed to proceed more or less with constant specific rate during growth
on soy-oil, but slight inhibited by sugar and product. The product inhibition
terms in Table 13.7, rows d and e, were suggested by the expert system to de-
scribe the reduced product formation and growth at the end of the cultivation,
because the data did not allow for clear discrimination of other possible mechan-
isms. The effect could also be caused by degradation of the product, variation in
the synthetic activity of the cells, or in the active portion of the cell mass. Simu-
lation results for a fed-batch cultivation are given in Fig. 13.11 with the model
parameters in Table 13.8. The structured unsegregated model is in good agree-
ment with the experimental data when considering the measurements errors. In

Table 13.7. Cybernetic model for Cephalosporin C production in a fed-batch process [102,
103]

Description Model equation

a Stoichiometric model q =Yr+m
0 0

(~,) (l yt)UD U)
0
y-I 1
51
0 0 I
y-I y-I ycp mc
qC02 CI C2

b Specific growth rate fL = YX51 r51 + YX52 r52

c Metabolic coordination J = fL (r51, r52) -t max

d Constraints for metabolic OS rSI S min (rl max(t), rSlmax--C_S_)

coordination KSI + Cs
OS rS2 S min (r2max (t), rS2max~ (1- ~))
K 1S + Cs KpI

e Specific rate of
Cephalosporin C synthesis
416 13 ~-Lactam Antibiotics Production

3 100
a
75
2 Cx
CO2 50
Cp
25

0 0
0 20 40 60 80 100 120 140 160
50
b
Cs 40
++
30 + +
COil <>
20 <> <> <>
<> <>
10
0
0 20 40 60 80 100 120 140 160
t [hI
Fig. B.lla,b. Simulation results with the cybernetic model (lines) in comparison to experi-
mental data (symbols) for a fed-batch cultivation of Acremonium chrysogenum [102): a concen-
trations of cell mass (Cx ), Cephalosporin C (Cp ), and mole fraction of CO 2 in the exhaust gas;
b concentrations of sugar (Cs ) and soy oil (COil); all liquid phase concentrations are in kg m- 3

Table B.S. Parameters in the metabolic regulator model, Fig. 13.11 and Table 13.7

Parameter Value Unit

rSl max 0.04

rS2max 0.024
rPmax 0.013
Y XS1 0.75
YXS2 1.0
YS1 5.5
YP1 0.9
YC1 0.9
YC2 0.7
Ycp 2.0
KSl 1.5 kgm- 3
K[s 0.5 kgm- 3
KIP 22.0 kgm- 3
K p1 35.0 kgm- 3
KP2 35.0 kgm- 3
mc 0.008 h- 1

the simulation, the initial and diauxic lag-phase can be seen on the exhaust gas
CO 2 -concentration.

13.3.3
Segregated Models Describing Morphological Differentiation

Matsumura et al. [56] presented a structured-segregated model describing the mor-

phological differentiation and regulation of the Cephalosporin C synthesis by the
13.3 Modeling of Cephalosporin C Production 417

carbon source and methionine. They discriminate the three cell types slim hyphae
(Xl), swollen hyphae (Xz), and arthrospores (X3 ) as shown in Fig. 13.9. The funda-
mental equations of the segregated model that emphasizes the importance of an in-
tracellular methionine pool are given in Chap. 2, Table 2.13. It is assumed that
methionine stimulates product synthesis and metamorphosis of slim hyphae to
swollen, thick-walled hyphae. The intrinsic methionine concentrations in slim hy-
phae, C521, swollen hyphae, C522, and arthrospores, C523, are calculated by
CS2l CS22 C S23
CS21 = -C CS22 = -C CS23 = -C (13.30)
Xl X2 X3

and the totally measured intrinsic methionine concentration is then

CS2! C X ! + CS22 C X2 + CS23 C X3 (13.31)
CS2 =
Cx
The intrinsic balance of methionine in slim hyphae considers uptake from the
medium with rate qS21, endogenous de novo synthesis from inorganic sulfate with
rate rNS21' utilization for protein synthesis with rate rG521' first order decay, and dilu-
tion by growth. Since metamorphosis of hyphal cells to thick-walled cells does not
change the intrinsic concentration of hyphal cells, the balance becomes
Uptake Synthesis Degradation Utilization Dilution by growth
~ ~ ~
at
dCS21 "....,--.
= QS21 + rNS21 k Dl cS21 - rGS21 -
~
/-t12 CS21 (13.32)

The following methionine balance in swollen hyphae has to consider by the last
summand the inflow of methionine from the newly formed cells by metamorphosis,
Inflow by metamorphosis
Uptake Synthesis Degradation Utilization, " ,
dCS22 "....,--. ~ ~ ~ CXI
- d - = QS22
t
+ rNS22 - k D2 CS22 - rGS22 + /-t12 -
CX2
(CS21 - CS22) (13.33)

In arthrospores there is no growth activity. Therefore, the balance simplifies to

Inflow by morphogenesis
Degradations A

dCS23 ~' C X2 '

-d- = - kD3 CS23 + /-t23 - (CS22 - CS23) (13.34)
t CX3
A fictive rate limiting enzyme, e.g., the ring-expansion enzyme, in the swollen
hyphae that is repressed by glucose is assumed to be responsible for the CPC synth-
esis. Formation of this enzyme in the swollen hyphae is assumed to follow a delayed
activation by methionine with lag-time t1 and immediate repression by glucose, and
is described by the intrinsic balance
r max ECS22(t - td
. 1 + KnCsI(t) . CX2(t - td
KE + CS22(t - td 1 + K I2 Cs I (t) C X2 (t)
(13.35)
- (kDE+/-t12~:~~:DCE(t)
The dilution by newly formed Xz-cells being deficient of enzyme is considered in
the last term. The above balance is actually inconsistent in regard to structured
models, because the intrinsic concentration CS22 - being averaged over the popula-
tion - is used in the non-linear Michaelis-Menten kinetics. The averaged intrinsic
418 l3 ~- Lactam Antibiotics Production

rate can only be calculated from averaged intrinsic concentrations when the model is
linear, i.e., the first factor must be either constant or proportional to C522. Please refer
to Chap. 2 for further discussion of structured models. The model was used to simu-
late qualitatively a process with linear-increasing substrate feed-rate over time. By
the feeding of glucose and methionine the process duration could be extended and
the product concentration was increased.
A model considering only thin hyphae (Xl) and swollen hyphae (Xl), which are
formed at impaired growth conditions, was developed by Chu and Constantinides
[55] for batch cultivations on glucose and sucrose. The basic segregated model is
described in Table 2.12, Chap. 2. Both morphological forms catabolize substrate, can
grow, and form new slim hyphae. In the model, production of CPC, formed by swol-
len hyphae, is repressed by glucose. The diauxic lag-phase for the switch to growing
on sucrose is modeled by a formal delay function in a non-mechanistic way,

(13.36)

where tl is the time of glucose limitation. Functional expressions for tlag in depen-
dence on the total cell mass and sucrose concentration were afterwards proposed by
Basak et al. [101]. Similar to the previous model, delayed synthesis of a fictive rate
limiting enzyme is assumed to be responsible for a lag-phase before start of the CPC-
production after depletion of glucose. Product synthesis is proportional to the con-
centration of the enzyme in the thick-walled cells that is synthesized during meta-
morphosis. The intrinsic balance for the rate-limiting enzyme given in the paper
violates the conservation law of mass and is not repeated here. A modified expres-
sion with a synthesis rate /1 J2 kCX1 and a decay constant kD could read

dCE CXl
- = /112-(k - CE) - kDcE (13.37)
dt CX2

This means that when the percentage of thick-walled cells is high, the rate of
change in enzyme concentration by the inflowing enzyme produced during meta-
morphosis is slowed down, and vice versa. The concentration will not change if the
newly formed cells have the same state as the existing ones, i.e., k=CE. The model was
supplemented by introducing a dependency of kinetic parameters on temperature
and pH. The extended model was then used to generate optimal control profiles for
these control variables, which increased the productivity by more than 50%.
A segregated-structured model aimed at industrial fed-batch cultivations on com-
plex substrate mixtures was developed for Cephalosporin C production by Meyerhoff
[60]. In the experiments, dry glucose syrup, which contains mainly maltose and oli-
gomers of this sugar, and soy-oil were used as carbon source [61]. There is no clear
separation of the growth phases on sugars and oil, both substrates being used in
parallel. Nevertheless, after limitation of the sugars a lag-phase can be observed,
where the cells adapt further to growth on oil as sole carbon source. The high-pro-
ducing strain used in the experiments also showed no clear separation of growth and
production phase. Production starts very early when still significant amounts of su-
gars are present in the medium. In this model, no precursors were considered, since
during the experiments limitation of methionine, ammonia, or phosphate were
avoided.
13.3 Modeling of Cephalosporin C Production 419

Table B.9. Segregated model for a high-producing strain of Acremonium chrysogenum [60]

Description Model equation

a Growth of active hyphae

Cs
{ls = {lSmax Ks +Cs

COil Cs
{lOil = {lOilmax KOil + COil c E + {lOilmaxS Ks +Cs

b Total biomass

c Metamorphosis reactions
of inactive cells to active
hyphae, active hyphae to
swollen hyphae, and
swollen hyphae to
{l34 -
-k~
34 Kl + Ks
arthrospores
d Product formation COil
by swollen hyphae qP = kP
K;';il + COil
e Sugar uptake

f Soy oil uptake QOil = -

{lOil ( C X2 + C X3 ) + qpCX3)
- C X2 + (mail * COil
YXOil YPOil KOil + COil
g Intrinsic balance of oil
catabolizing enzyme

Fig.B.12. Structure of the segregated model of Meyerhoff [60]

420 13 ~-Lactam Antibiotics Production

The model given in Table 13.9 considers initially inactive (Xl), growing but non-
producing (Xz), and swollen producing hyphae (X3 ), as well as inactive arthrospores
(X4 ) as shown in Fig. 13.12. By the inactive form an initial lag phase after inocula-
tion of the reactor is modeled. These cells are transformed into the active growing
form, which then can differentiate to swollen hyphae, and, after limitation of the
sugars, to arthrospores. Growing hyphae use the sugar and oil substrates with spe-
cific growth rates 115 and 1l0il. Arthrospores are formed from swollen hyphae at low
sugar concentrations. The sugar uptake, mainly maltose, is determined by growth
and product formation in the respective morphological states. Since in the measured
sugar concentration some sucrose is contained, which cannot be metabolized by the
fungus, the total sugar concentration is modeled as the sum of maltose and a con-
stant value of sucrose. The balance equation for soy oil considers the effects of
growth, maintenance, production, and feeding. The experimental results suggest that
the oil uptake in parallel to sugar follows a different mechanism than for oil as a sole

*
40 80 120 180
t [h]

Fig. 13.13. Experimental data (symbols) and simulation results (lines) by the segregated model
[601 for a fed-batch cultivation of Acremonium chrysogenum. Concentrations of total biomass
(Cx ), total sugar (Cs ), Cephalosporin (Cp ), and soy oil (COil), all in kg m- 3

Table 13.10. Model parameters for simulation of the fed-batch cultivation of Acremonium chry-
sogenum in Fig. 13.13 by the segregated model in Table 13.9

Parameter Value Unit

{iSmax 0.028
{iOilmax 0.032
kmax 0.115
Yxs 0.53
YXOi1 0.8
Yps 1.5
YPOi1 1.2
Ks 3.3 gt l
K] 0.3 g I-I
k12, k23 0.027 h- I
k34 0.032 h- I
kp 0.023 h- I
mail 0.0027 h- I
13.4 Process Control and Optimization 421

carbon source. The first one is assumed to be controlled by the sugar uptake rate,
and the second one by Monod-kinetics dependent on the oil concentration. The
total specific growth rate on soy oil then includes both terms as a sum. The key
enzyme for pure growth on oil is inducible and described by an intrinsic balance,
Table 13.9, row g).
The model was able to describe experimental data for several cultivations with
only small variations in the parameters. The simulation results of one cultivation
are given in Fig. 13.13 with the model parameters in Table 13.10. Maltose is the pre-
ferred substrate for this organism. Therefore, the oil uptake rate in the first growth
phase is much lower than after the sugar limitation and adaptation of the oil cata-
bolizing system. For this process, the unique discrimination in growth and produc-
tion phase is difficult. The presence of sugar has no strong repression effect on the
product synthesis. Therefore, Cephalosporin C is synthesized just from the begin-
ning. The lag-phase in growth and oil uptake after 70 h, when maltose became limit-
ing, can be deduced from the temporary increasing oil concentration, and more
clearly from exhaust gas CO 2 that is not shown.

13.4
Process Control and Optimization

13.4.1
Problems and Possibilities

The field for process control can be divided into four areas: determination of the
optimum mode of operation of the bioreactor, including optimum conditions for
inoculation under consideration of the preculture; dynamic optimization of process
variables; feedback control of operating parameters and process variables to con-
stant setpoints or optimum dynamic profiles; and finally optimum scheduling of
the entire plant. Process variables of interest for control are temperature, pH, sub-
strate feed, oxygen supply, precursor feed, and nitrogen supply.
On the plant level, the optimization can be done by dynamic scheduling of the
process steps and determination of a suitable harvest time for the cultivation. Using
the on-line estimated profit and the natural variance in the productivity, Yuan et al.
[75] suggest a strategy for deciding when a cultivation should be stopped. In com-
parison to a set of historic data, the actual cultivation run is classified as bad, nor-
mal, or good. By stopping bad runs earlier and prolonging the cultivations classified
as good ones, the profit can be increased without changing the control scheme for
the cultivations. The advantage of this method is that it does not directly rely on a
mathematical model. This is only used for estimation of the product concentration
when direct measurements are not available. Further improvements can be expected
by combining the dynamic scheduling with optimal dynamic control of each run.
The frame for process control is determined by the physiology of the fungi. The
catabolite repression of antibiotics synthesis requires roughly a two-phase process
control strategy with a first phase of optimum growth and a second phase with low
growth rate but maximum production. This procedure was extensively studied ex-
perimentally and is considered as optimal. It can be realized in two ways. In a batch
process, a mixture is provided of a rapidly catabolizable first substrate, e.g., glucose,
and a slowly degradable second substrate, e.g., lactose. When the first substrate is
422 13 13-Lactam Antibiotics Production

used up, the process switches from growth to production phase by itself. But the
growth rate is then fully determined by the choice of the substrate and cannot be
further influenced directly, so it may deviate from optimum conditions for produc-
tion. Therefore, additional precautions such as nitrogen limitation must then be em-
ployed. These disadvantages are avoided by fed-batch cultivation, where the growth
rate in the production phase can in principle be controlled to any value below the
maximum rate. In the first growth phase the process is usually still run as batch
cultivation. The two-phase strategy may be supported by a proper control of tem-
perature and pH levels, because these variables also have different optimum values
for growth and production; pH strongly influences morphological differentiation
and degradation of the product. The biomass should grow as much as possible in
the first growth phase to ensure high productivity afterwards, but oxygen limitation
must be avoided [19, 62]. The optimal initial amount of substrate and the maximum
attainable biomass depends on the oxygen transfer capacity of the reactor [22, 27, 28,
48]. The production phase is usually continued into the stationary phase without net
growth. But the process has to be stopped, e.g., by cooling below 10 °C, before the
onset of autolysis induces a strong increase in pH that leads to degradation of the
product. Other stopping criteria with respect to downstream processing are the
rheology or filtration properties of the broth.
The optimum mode of operation of a general bioprocess, where substrate concen-
tration determines the growth and production kinetics, was investigated by Modak
and Lim [124]: batch, fed-batch, repeated batchlfed-batch, or continuous. According
to their kinetics, the processes were classified into four types: for maximization of
yield, batch operation is optimal when the intrinsic yield (the ratio of specific pro-
duct formation to substrate uptake) increases with substrate concentration; contin-
uous operation is optimal when the specific yield decreases, fed-batch when it goes
through a maximum, and repeated operation for constant specific yield. For maxi-
mization of productivity, continuous operation is always optimal from the theoreti-
cal point of view, but in practice this is avoided by reasons of operational and strain
stability. The process can be prolonged and the productivity also increased by re-
peated batch/fed-batch operation mode, which reduces the required time for reactor
preparation and growth phases. This could be shown generally in theoretical inves-
tigations for production of antibiotics and non-growth associated products [32, 85-
87]. Hasegawa et al. [84] used a model for batch penicillin production [70] to inves-
tigate repeated batch strategies for enhancing productivity. The improvement is due
to increased production rate, prolonged cultivation, and higher final concentration.
Guthke and Knorre [40] investigated the optimal control of repeated fed-batch culti-
vations by the model of Nestaas and Wang [39], and found an optimal cycle period
of 48 h and a draw-off ratio of about 80%. This is lower than the result in [32] but
still unusually high. The time for addition of fresh media can be very critical, and
the production may even stop under adverse conditions [33]. The maximum number
of repeated cycles is bounded by the capacity of the plant for substrate preparation,
accumulation of toxic or inhibitory substances, or degeneration of the strain. Never-
theless, repeated fed-batch cultivation is of great interest for industrial production.
Several authors have dealt with optimal control of fed-batch processes in general,
and antibiotics production in particular [21, 63-68]. Generally unstructured kinetic
models are used and comparison to experimental data is rare. Mostly the substrate
feed or concentration is optimized. The optimum control profile consists of a se-
quence of maximum feeding, batch, singular feeding during production phase, and
13.4 Process Control and Optimization 423

a short final batch phase. Depending on the boundary conditions for the optimiza-
tion, not all phases of the sequence may be present. The optimum initial substrate
concentration was found to be higher when productivity is maximized instead of
yield. Numerical algorithms for practical computation of the control were given in
Lim et al. [78] for penicillin production. The results were later improved with Itera-
tive Dynamic Programming by Luus [120]. A penalty function was used to handle
constraints of the state variables and the global optimum was supposed to be found.
Diener and Goldschmidt [122] published a suboptimal control strategy for maximiz-
ing product concentration, yield, or productivity in fed-batch cultivations using
phase partitioning. In this scheme, the global optimization problem defines the se-
quence of phases being computed off line, while during each phase only local criteria
have to be fulfilled which can easily be translated into feedback control laws. The
method was applied for penicillin production using the model of Bajpai and Reug
[28]. The results of the suboptimal strategy were quite close to the global optimum
that requires a much higher effort for its determination. Guthke and Knorre [69]
calculated optimum concentration profiles for substrate during antibiotics produc-
tion by using a general model where the specific product formation is maximum
below the maximum growth rate. Suboptimal strategies were proposed that can be
implemented more easily. Bajpai and Reug [48] evaluated different feeding strategies
on the basis of simulations with their simple unstructured model. The effect of oxy-
gen limitation is discussed. It was found that the differences in productivity among
the feeding strategies are less than expected, so the robustness in practical applica-
tion should be decisive for their selection. San and Stephanopoulos [121] investi-
gated the optimal control of the substrate inlet concentration at fixed flow rate under
constraints in maximum cell mass and substrate concentration by the model of Baj-
pai and Reug [28]. Rodrigues and Filho [123] optimized the productivity numeri-
cally by the above model using a simplex method. Temperature and pH optimization
of Cephalosporin production was investigated by Chu and Constantinides [55]. Sui-
table temperature profiles were determined by Constantinides et al. [20,70] for batch
penicillin production by means of a simple unstructured model and application of
the continuous maximum principle.
Automatic feedback control of the feeding rate of carbon sources for antibiotics
production was not so extensively studied as for other biotechnical processes, e.g.,
baker's yeast production. This has a number of reasons. The setpoint - or better time
profile - of operation for maximum productivity is not so clearly defined due to the
still somewhat limited knowledge of the process dynamics, although the mathemati-
cal models may delude a clearer image. Furthermore, there is no direct sensitive and
unique indication for the metabolic state of the culture on easily accessible measure-
ments, such as RQ for baker's yeast. Due to the morphological heterogeneity of the
culture the optimum rate of substrate supply depends on the biomass fraction in the
producing morphological state. So even for identical total biomass the optimum sub-
strate feed is different when the morphological composition of the population
changes, and up to now this is difficult to determine on-line. Another problem is
the use of ill-defined substrates from natural resources that may only be catabolized
partially during the process. Application of a single-loop control for substrate flow
seems to be generally problematic in antibiotics production. When there is indeed a
maximum for the specific production over growth rate as assumed by a number of
models, it is impossible to decide whether the substrate supply should be reduced or
increased without additional knowledge. So further information from other process
424 13 ~-Lactam Antibiotics Production

variables has to be taken into account. Therefore, practical feedback control schemes
are mostly realized as tracking control according to a predetermined profile that was
found experimentally or with assistance of a model. The preset profile can be for cell
mass; but because this is difficult to measure, estimation schemes must be estab-
lished by balancing or mathematical models. These can use, for example, exhaust
gas measurements of O2 and CO 2. Variants of this control take the specific growth
rate as controlled variable instead of cell mass. Nevertheless, these control schemes
are at most suboptimal since they do not consider the true actual morphological or
physiological state of the culture. But often the productivity and reproducibility can
be improved even by simple control, because critical situations during the process
can be avoided or reduced.
Mou and Cooney studied the computer control of penicillin production by means
of balancing principles [35] with the aim of increasing reproducibility. Carbon diox-
ide production was used to estimate cell mass and specific growth rate, and to con-
trol both according to predefined schemes. The concept was extended for cultiva-
tions containing corn-steep-liquor [119] as complex substrate by using an approxi-
mate carbon balance. Fed-batch control by different measured variables was the sub-
ject of a number of papers: p02 [62, 88]; pH, [89, 90]; carbon dioxide production
[91]; specific growth rate [22, 34, 36, 92]. Montague et al. [29] used an extended
Bajpai-ReuB-model for estimation of cell mass by a Kalman-filter by means of the
carbon dioxide production rate. By adaptive control of substrate feed, the cell mass
was kept on a predetermined trajectory. Carbon dioxide production and RQ were
used by Nelligan and Calam [36] for estimation of f1 and desired oxygen uptake rate.
On the basis of different predetermined strategies the process was then controlled by
both variables while avoiding oxygen limitation. Nestaas and Wang [39] used the
morphologically structured model for open-loop and feedback control of fed-batch
penicillin processes according to the desired cell mass and specific growth rate.
Feedback control was achieved by measuring the hyphal density with a filtration
probe. Suboptimal control strategies were investigated by Van Impe et al. [125]. A
heuristic control law for the feed-rate during production phase by the substrate con-
centration as control variable gave good performance in simulations and was quite
close to the global optimum. Adaptive extensions, either based on an a simple ob-
server for the substrate concentration or on the carbon dioxide production rate, were
also proposed.

13.4.2
Example for Dynamic Optimal Control of Fed-Batch Antibiotics Production

In the work of Meyerhoff [60], summarized in [99], the model given in Sect. 13.3.3,
Table 13.9, was applied for dynamic optimization of the sugar and oil feeding for
cultivations of Acremonium chrysogenum by using the Iterative Dynamic Program-
ming (IDP) algorithm [71-74]. This is easy to use and has good convergence proper-
ties compared to methods based on Pontryagin's Maximum Principle. Several per-
formance criteria were tested: total mass of product, product yield per supplied sub-
strate, and economic profit. During the simulation studies, a maximum filling vo-
lume of the reactor and a maximum oil concentration of 8 g r 1 were introduced as
additional restrictions. Since the oil promotes coalescence of the air bubbles, the
sense of the latter restriction is to maintain a sufficient oxygen supply to the reactor.
13.4 Process Control and Optimization 425

10
45
8 Cx 40
,
~COil
, 35
6 J=- __/ 30
Cs ,,/ -. , ,
/
25
, Cp -
4 , /- 20
,- IS
/'
2
, 10
5
0 - 0
0 25 SO 75 100 125 ISO
t [h]
0,125
12
10 FOil
0,100
8
6 0,075
4
2
0
-2
-4
0 25 50 75 100 125 150
t [h]

Fig. 13.14a,b. Simulation of the optimal control of sugar and soy-oil feeding for a fed-batch
cultivation of Acremonium chrysogenum by a segregated model [60]: a concentrations of total
biomass (Cx ), sugar (Cs), Cephalosporin C (Cp ), and soy oil (COil), all in kg m- 3 ; b flow rates of
sugar (Fs in lh-'), and soy oil (FOil in gh-'). Reprinted from [99]

For the optimization, the total process duration (tf =150 h) and maximum feeding
rates were fixed (Fs <0.05 I h-t, F Oil <12gh- I ).
Figure 13.14 shows the optimization results for both of the control variables, sugar
(Fs) and soy oil feed (Fail), for a performance index of final total mass of the pro-
duct. Compared to the original process, the growth phase is prolonged to about
105 h by a higher sugar feed under near-batch conditions. The sugar flow rate also
supports sufficient growth in the production phase when the oil concentration de-
creases. Soy oil is used up to the maximum amount that can be afforded under the
given restrictions on states and feeding rate. From the simulation, the optimized
process shows a clearly increased product concentration and productivity by en-
forced growth. This is due to the seemingly weak catabolite repression of product
synthesis that allows for relative high product formation even under growth favoring
conditions.

13.4.3
Economic Optimization for Mycelia Fed-Batch Cultivation

Optimization of a bioprocess is mostly a multi-objective problem tackling several

criteria such as productivity, yield, and quality of the product. These are usually
not independent of each other, and the manipulating variables influence them simul-
taneously. A more general way is to optimize the economic profit directly. The profit
function is an objective description of a process under the aspect of economy. Global
426 13 p-Lactam Antibiotics Production

economic optimization of a general antibiotics plant was investigated by Okabe and

Aiba [66] based on data of an industrial process. They considered the sections culti-
vation, filtration, extraction, and final treatment of the product, e.g., by crystalliza-
tion. A major part of the costs is for cultivation, i.e., raw material and energy, and
therefore its optimization is of greatest interest. The influence of operational proce-
dures, variation in down-stream processing steps, disposal of solid waste, and selec-
tion of cultivation medium could be successfully evaluated with respect to the eco-
nomic effectiveness. Genon and Saracco [32] extended the Bajpai-ReuB model [28]
by economic balances to evaluate the profitability of fed-batch, repeated fed-batch,
and continuous cultivations in one- and two-stage reactor systems. It was found that
with this simple unstructured model the last system performed best when neglecting
the known operational difficulties for continuous cultivations. Repeated batch with a
cycle period of 96 h and a draw-off ratio of >95% of the reactor operating volume
was the second choice.
The on-line profit estimation and optimization of industrial multi-reactor plants
for mycelia fed-batch cultivations is further discussed by Yuan et al. [75]. For a sin-
gle batch, the gross profit is the revenue from product sales minus the total input for
raw materials, energy, capital, personnel, and other costs, which were expended for
winning the product. The goal of optimization is to get the highest gross profit with-
in a minimal operation period. The profit function is then given by
revenue - total costs
J()
tf = ------- (13.38)
tf + tpr
where tf is the process duration and tpr the time interval between two successive
batches in the same reactor in which emptying, cleaning, batch medium preparing,
sterilization, and inoculation are done. The total costs as summarized in Table 13.11
were divided into two parts - the direct production-related costs for raw material
and energy consumption from pretreatment to product package, and the remaining
indirect costs. The material consumption in downstream processing is product con-
centration-dependent. So the related terms, rows e-g in Table 13.11, are multiplied
by a correction factor, tp, being the ratio of actual product concentration to desired
product concentration under design conditions. The operating factor, representing
the fraction of time where the plant is in full-scale production, was assumed as 90%
in row h). The profit function is then in detail

VL(tf )Cp(tf )0.91Psa le - Pini - Pin - Popttf - VL(tf) (P extr + Peryst + Pdry)
J(tf) = t +t
f pr
(13.39)

where PSaie is the sales price of product, and the separation yield was estimated to
0.9l.
The profit function can be used either for off line optimization similar to the ex-
amples above, or for real-time control and scheduling of the process. In this case its
development in time is of great interest. Figure 13.15 shows the typical time course
of the profit function of a batch. At the beginning, the profit is negative because of
the cost for inoculum and initial batch medium preparation. During the following
cultivation, the profit increases in connection with the amount of product. The be-
ginning of the profitable operation is at time tj; the stationary phase of the cultiva-
13.4 Process Control and Optimization 427

Table B.ll. Costs per m 3 of Penicillin production for a 50-m3 production reactor [75,81,82]

Cost Calculation formula

Direct
a Seed inoculum, approximated as direct production costs:
PINo=O.23Plmp+ 1730PAir+ 1.0Pw+251PcWJs+ 1.01PEle+O.OI2PSteam+30PGlu+
10PLac+O.5PKH2P+ 5Pcaco3 +2PNH4S+ 15PStarch
b Initial input, includes expenses for sterilization, medium preparation and inoculum:
Pini=33Pw+6PSteam+6PINo+ 1.8PDesmo+ 108PGlu+45PStarch+ 14PKH2P+ 180Pcaco3
+ 180PNH4s+90PK2so4+ 360PPharma
C Feed solution: nitrogen source, carbon source and precursor, the factor of 1.15 con-
siders the costs for pretreatment:

J
Tf

Pin = 1.15 [Fs(t)Cf(t)Pclu + FpAA(t)PPAA +FNH3(t)PNH3

o

d Energy for aeration, agitation and cooling, including electric power and cooling
water: Popt =28PcW4 + 132PEle+4000PAir
e Product extraction from broth: consumption of extract solvent, cooling water and
electric power: Pextr=2.24PBuAc 'P +O.2PNet 'P +O.08PH2S04+PCWlS+2PcW4+97PEle
f Product crystallization: salts, cooling water and electric power:
Pcryst=1O.6PKAc 'P +O.6Pcw4+50PEle

g Drying: consumption of steam, nitrogen, wash solution, and so on:

Pdry=2P1sop 'P + 146PN2 'P + 1.5Pcw4+5PEle+O.OlPSteam
Indirect
h Total indirect costs: credit repayment, labor expense, maintenance and repairs,
research and development, insurance, storage and transportation, sales expenses:
Pindir=Total indirect costs/(360x24xO.9xNumber of reactors x V R )

Profit J

J(t)
--------------------------------------------7-..,.--~_

time

Fig.B.15. Typical time course of a profit function during fed-batch antibiotics production
428 l3 I3-Lactam Antibiotics Production

Table 13.12. Prices of raw materials and product for penicillin production [75]

Price Symbol Value Unit

compressed air PAir 0.004 DM m- 3

butyl acetate P BuAc 2.37 DM kg- 1
CaC0 3 P Cacm 0.45 DM kg- 1
cooling water (4 DC) P CWI5 0.15 DM m- 3
cooling water (15 DC) P CW4 0.3 DM m- 3
anti-foam solution PDesmo 5.0 DMr 1
electricity P Ele 0.15 DM (kWhr 1
glucose P G1u 0.53 DM kg- 1
H2 S0 4 (98%) PH2S04 0.46 DM kg- 1
inoculum liquid with a volume of 100 ml PImp 10.0 DM per flask
approximate indirect costs Pindir 1.86 DM h- 1m- 3
isopropanol P rsop 1.0 DM kg- 1
K2 S0 4 PK2S04 0.99 DM kg- 1
KH 2 P0 4 P KH2P 4.09 DM kg-I
lactose PLac 20.0 DM kg-I
nitrogen P N2 0.05 DM m- 3
net promote P Net 2.0 DM kg-I
ammonia solution (16.5%) P NH3 0.15 DM kg-I
(NH 4 h S0 4 P NH4S 1.18 DM kg-I
phenylacetic acid PPAA 3.25 DM I-I
pharma-medium PPharma 0.25 DM kg-I
sales price of penicillin G PSaie 80.0 DM kg-I
corn starch PStarch 0.85 DM kg-I
steam PSteam O.oz8 DM kg-I
process water Pw 1.6 DMm- 3

1(tj )
300
... Pen 1
..........
..-..-...... .........
250
• Pen 2 ...
......
...
...~ ",.

.....
200
• Pen 3

....
150
100

50
.. :....
0

·50 ...."'.-
•
...•
·100 ~

-150
-200
o 20 40 60 80 100 120 140 160 180 200 220
t [h]

Fig. 13.16. Profit functions of three industrial Penicillin G fed-batch cultivations. Reprinted
from [75] "Profit optimization for mycelia fed-batch fermentation" Copyright (1999), with
permission of Elsevier Science
References 429

tion is more or less reached at time tb where the profit approaches the maximum
value; later on the profit function may drastically decrease, which can be caused by
autolysis of mycelia, hydrolysis of product, or contamination. The cultivation must
be stopped and downstream processing started before this point at t3' The estimated
profit function, converted to equivalents of a 50-m3 reactor, for three industrial Pe-
nicillin production runs is presented in Fig. 13.16. In Table 13.12 approximate prices
of raw materials and product for the profit function estimation of Penicillin G pro-
duction are summarized. In the stationary phase, the difference of the profit func-
tion between runs Pen 1 and Pen 2 is about 50 DM h -1. When these two batches are
stopped at 186 hand tpr is set to 14 h, the gross profit difference is
50x(186+14)=10,000DM. From Fig. 13.16 the potential of dynamic scheduling by
using the natural variance in the profit function is also obvious. Compared to fixed
harvesting time, the profit can be increased by stopping those cultivations in ad-
vance that have a lower profit or reach their maximum profit earlier, while processes
with high maximum profit may be prolonged even when the increase in profit func-
tion is delayed. The possible gain in profit by such procedure was estimated at about
5% without changing the control scheme of the cultivation.

References

1. Vandamme EJ (ed) (1984) Biotechnology of industrial antibiotics. Dekker, New York

2. Moo-Young M (ed) (1984) Comprehensive biotechnology, vol 3: the practice of biotech-
nology, current commodity products. Pergamon, Oxford
3. Queener S, Neuss N (1982) The biosynthesis of ~-lactam antibiotics. In: The chemistry
and biology of ~-lactam antibiotics, vol 3. Academic Press, New York, p 141
4. Martin JF, Revilla G, Zanca DM, Lopez-Nieto MJ (1983) Carbon catabolite repression of
penicillin and cephalosporin biosynthesis. In: Trends in antibiotics research. Antibiotics
Research Association, Japan, p 258
5. Court JR, Pirt SJ (1981) J Chern Technol Biotechnol31:235
6. NiehoffJ (1987) PhD thesis, Universitat Hannover
7. Wittler R, Schiigerl K (1985) Appl Microbiol Biotechnol 21:834
8. Hilgendorf P, Heiser V, Diekmann H, Thoma M (1987) Appl Microbiol Biotechnol27:247
9. Nielsen J (1992) Adv Biochem Eng Biotechnol46:187
10. van Suijdam JC, Metz B (1981) Biotechnol Bioeng 23:111
11. Metz B, de Bruiijn EW, van Suijdam JC (1981) Biotechnol Bioeng 23:149
12. van Suijdam JC, Metz B (1981) J Ferment Technol 59:329
13. Trinci APJ (1970) Arch Microbiol 73:353
14. Metz B, Kossen NWF (1977) Biotechnol Bioeng 19:781
15. Whitaker A (1987) Int Ind Biotechnol 2:85
16. Wittler R, Baumgartl H, Liibbers DW, Schiigerl K (1986) Biotechnol Bioeng 28:1036
17. Hotop S, Willer J, NiehoffJ, Schiigerl K (1993) Proc Biochem 28:99
18. Edwards N, Beeton S, Bull AT, Merchuck JC (1989) Appl Microbiol Biotechnol 30:190
19. Wittler R (1983) PhD thesis, Universitat Hannover
20. Constantinides A, Spencer JL, Gaden EL (1970) Biotechnol Bioeng 12:803
21. Fishman VM, Biryukov VV (1974) Biotechnol Bioeng Symp 4. Wiley Interscience, New
York, p 647
22. Calam CT, Russel DW (1973) J Appl Chern Biotechnol 23:225
23. Megee RD, Kinoshita S, Fredrickson AG, Tsuchiya HM (1970) Biotechnol Bioeng 12:771
24. Trinci APJ (1974) J Gen Microbiol 81:225
25. Righelato RC (1979) The kinetics of mycelial growth: fungal walls and hyphal growth.
Cambridge University Press, London, p 385
26. Bull AT, Trinci APJ (1977) Adv Microbiol Physiol 15:1
430 l3 f3-Lactam Antibiotics Production

27. Hegewald E, Wolleschensky B, Guthke R, Neubert M, Knorre WA (1981) Biotechnol

Bioeng 23:1133
28. Bajpai RK, ReuB M (1980) J Chern Technol Biotechnol 30:332
29. Montague A, Morris AJ, Wright M, Aynsley M, Ward A (1986) Can J Chern Eng 64:567
30. Nicolai BM, Van Impe JF, Vanrolleghem PA, Vandewalle J (1991) Biotechnol Letters
13:489
31. Heijnen JJ, Roels JA, Stouthamer AH (1979) Biotechnol Bioeng 21:2175
32. Genon G, Saracco G (1992) Chern Biochem Eng Q 6:75
33. Calam CT, Ismail BAK (1980) J Chern Technol Biotechnol 30:249
34. Cooney CL, Mou DG (1982)' Application of computer monitoring and control to the
penicillin fermentation. In: 3rd International Conference on Computer Applications in
Fermentation Technology, Manchester, England, 1981, p 217
35. Mou DG, Cooney CL (1983) Biotechnol Bioeng 25:225
36. Nelligan I, Calam CT (1983) Biotechnol Letters 5:561
37. Calam CT (1979) Folia Microbiol 24:276
38. Germain B, Mele M, Kremser M (1980) Biotechnol Bioeng 22:255
39. Nestaas E, Wang DIC (1983) Biotechnol Bioeng 25:781
40. Guthke R, Knorre WA (1987) Bioproc Eng 2:169
41. van Suijdam JC, Hols H, Kossen NWF (1982) Biotechnol Bioeng 24:177
42. Kluge M, Siegmund D, Diekmann H, Thoma M (1992) Appl Microbiol Biotechnol 36:446
43. Aynsley M, Ward AC, Wright AR (1990) Biotechnol Bioeng 35:820
44. Prosser JI, Trinci APJ (1979) J Gen Microbiol111:153
45. Nielsen J (1993) Biotechnol Bioeng 41:715
46. Paul GC, Thomas CR (1996) Biotechnol Bioeng 51:558
47. Tiller V, Meyerhoff}, Sziele D, Schiigerl K, Bellgardt KH (1994) J Biotechnol 34:119
48. Bajpai RK, ReuB M (1981) Biotechnol Bioeng 23:717
49. Meyerhoff J, Tiller V, Bellgardt KH (1995) Bioproc Eng 12:305
50. Yang H, Reichl U, King R, Gilles ED (1991) Biotechnol Bioeng 39:44
51. Yang H, King R, Reichl U, Gilles ED (1991) Biotechnol Bioeng 39:49
52. Meyerhoff J, Bellgardt KH (1995) Bioproc Eng 12:315
53. Meyerhoff}, Bellgardt KH (1995) J Biotechnol 38:201
54. Demain AL, Ahronowitz Y, Martin JF (1983) Metabolic control of secondary biosyn-
thetic pathways. In: Vining LC (ed) Biochemistry and genetic regulation of commer-
cially important antibiotics. Addison-Wesley, Reading, MA
55. Chu WBZ, Constantinides A (1988) Biotechnol Bioeng 32:277
56. Matsumura M, Imanaka T, Yoshida T, Tagichi H (1981) J Ferment Technol 59:115
57. Matsumura M, Imanaka T, Yoshida T, Tagichi H (1980) J Ferment Technol 58:197
58. Matsumura M, Imanaka T, Yoshida T, Tagichi H (1980) J Ferment Technol 58:205
59. Malmberg LH, Hu WS (1992) Appl Microbiol Biotechnol 38:122
60. Meyerhoff J (1995) PhD thesis, Universitat Hannover
61. Tollnik C (1993) PhD thesis, Universitat Hannover
62. Giona AR, De Santis R, Marelli L, Toro L (1976) Biotechnol Bioeng 18:493
63. Modak JM, Lim HC (1987) Biotechnol Bioeng 30:528
64. Modak JM, Lim HC, Tayeb YJ (1986) Biotechnol Bioeng 28:1396
65. Modak JM, Lim HC (1989) Biotechnol Bioeng 33:11
66. Okabe M, Aiba S (1975) J Ferment Technol 53:730
67. Hong J (1986) Biotechnol Bioeng 28:1421
68. Vicik SM, Fedor AJ, Swartz RW (1990) Biotechnol Prog 6:333
69. Guthke R, Knorre WA (1981) Biotechnol Bioeng 23:2771
70. Constantinides A, Spencer JL, Gaden EL (1970) Biotechnol Bioeng 12:1081
71. Luus R (1989) Hung J Ind Chern 17:523
72. Luus R (1990) Int J Control 52:239
73. Bojkov B, Luus R (1994) Chern Eng Res Des 72:72
74. Bojkov B, Luus R (1992) Ind Eng Chern Res 31:1308
75. Yuan JQ, Guo SR, Schiigerl K, Bellgardt KH (1997) J Biotechnol 54:175
76. Paul GC, Kent CA, Thomas CR (1993) Biotechnol Bioeng 42:11
References 431

77. Bartoshevic YE, van den Heuvel JC (1990) Chern Eng J 44:Bl
78. Lim HC, Tayeb YJ, Modak JM, Bonte P (1986) Biotechnol Bioeng 28:1408
79. Prosser JI, Tough AJ (1991) Crit Rev Biotech 10:253
80. Keirn J, Shen YQ, Wolfe S, Demain AI (1984) Europ J Appl Microbiol Biotechnol 19:232
8l. Grant EL, Ireson WG, Leavenworth RS (1982) Principles of engineering economy II, 7th
edn. Wiley, New York
82. Ulrich GD (1984) A guide to chemical engineering process design and economics. Wi-
ley, New York
83. Wittler R, Matthes R, Schugerl K (1983) Eur J Appl Microbiol Biotechnol 18:17
84. Hasegawa S, Shimizu K, Kobayashi T, Matsubara M (1985) J Chern Tech Biotechnol
35B:33
85. Shimizu K, Kobayashi T, Nagara A, Matsubara M (1985) Biotechnol Bioeng 27:743
86. Abulesz EM, Lyberatos G (1987) Biotechnol Bioeng 24:1059
87. Guthke R, Knorre WA (1982) Biotechnol Bioeng 24:2129
88. Giona AR, Marelli L, Toro L, De Santis R (1976) Biotechnol Bioeng 18:473
89. Pan CH, Hepler L, Elander RP (1972) Dev Ind Microbiol13:103
90. Jensen FG, Nielsen R, Emborg C (1981) Eur J Appl Microbiol Biotechnol 13:29
9l. Itsygin SB, Biryukov VV, Lurie LM, Berezovskaya AI (1977) Pharm Chern J 1O:ll9
92. Ryu DDY, Hospodka J (1980) Biotechnol Bioeng 22:289
93. Schugerl K, Seidel G (1998) Chern Ing Tech 70:1596
94. Justen P, Paul GC, Nienow AW, Thomas CR (1998) Bioproc Eng 18:7
95. Nielsen J (1995) Physiological engineering aspects of Penicillium chrysogenum. Polyte-
knisk Forlag, Lyngby
96. Krabben P, Nielsen J, Michelsen ML (1997) Chern Eng Sci 52:2641
97. Krabben P, Nielsen J (1998) Adv Biochem Eng Biotechnol 60:125
98. Zangirolami TC, Johansen CL, Nielsen J, Jorgensen SB (1997) Biotechnol Bioeng 56:593
99. Bellgardt KH (1998) Adv Biochem Eng Biotechnol 60:153
100. Caldwell IY, Trinci APJ (1973) Arch Microbiol 88:1
101. Basak S, Velayudhan A, Ladisch MR (1995) Biotechnol Prog ll:626
102. Ludewig D (1999) Expertensystem zur Entwicklung von Proze6modellen fUr biotech-
nische Prozesse, Fortschritt-Berichte 20 Nr 289, VOl Verlag, Dusseldorf
103. Ludewig D, Bellgardt KH (1997) BioTec 4:31
104. Gomersall R, Grumann S, Ludwig D, Seeger M, Hitzmann B, Bellgardt KH, Munack A
(1998) Automatisierungstechnik 46:368
105. Tollnick C (1996) PhD thesis, Universitat Hannover
106. Zhou W, Holzhauer-Rieger K, Dors M, Schugerl K (1992) Enzyme Microb Technol
14:848
107. Weil J, Miramonty J, Ladisch MR (1995) Enzyme Microb Technol 17:88
108. Weil J, Miramonty J, Ladisch MR (1995) Enzyme Microb Technol 17:85
109. Martin JF, Liras P (1989) Adv Biochem Eng/Biotechnol 39:153
llO. Vining LC, Sharpiro S, Madduri K, Stuttard C (1990) Biotech Adv 8:159
lll. Pedersen AG, Bungaard-Nielsen M, Nielsen J, Villadsen J, Hassager 0 (1993) Biotechnol
Bioeng 41:162
ll2. Packer HL, Thomas CR (1990) Biotechnol Bioeng 35:870
113. Nielsen J, Krabben P (1995) Biotechnol Bioeng 46:588
ll4. McNeil B, Berry DR, Harvey LM, Grant A, White (1998) Biotechnol Bioeng 57:297
ll5. Harvey LM, McNeil B, Berry DR, White S (1998) Enzyme Microbiol Technol 22:446
ll6. J0rgensen H, Nielsen J, Villadsen J, M011gardt H (1995) Biotechnol Bioeng 46:ll7
ll7. J0rgensen H, Nielsen J, Villadsen J, M011gardt H (1995) Appl Microbiol Biotechnol
43:123
ll8. Nielsen J, J0rgensen H (1995) Biotechnol Prog ll:299
ll9. Mou DG, Cooney CL (1983) Biotechnol Bioeng 25:257
120. Luus R (1993) Biotechnol Bioeng 41:599
121. San KY, Stephanopoulos G (1989) Biotechnol Bioeng 34:72
122. Diener A, Goldschmidt B (1994) J Biotechnol 33:71
123. Rodrigues JAD, Filho RM (1996) Chern Eng Sci 51:2859
432 13 13-Lactam Antibiotics Production

124. Modak JM, Lim HC (1992) Chern Eng Sci 47:3869

125. Van Impe JF, Nicolai BM, De Moor B, Vandewalle J (1993) Chern Biochem Eng Q 7:13
126. Henriksen CM, Christensen LH, Nielsen J, Villadsen J (1996) J Biotechnol 45:149
127. Moller J, Niehoff J, Hotop S, Dors M, Schiigerl K (1992) Appl Microbiol Biotechnol
37:157
128. de Noronha Pis sara P, Nielsen J, Bazin MJ (1996) Biotechnol Bioeng 51:168
129. Menezes JC, Alves SS, Lemos JM, de Azevedo SF (1994) J Chern Tech Biotechnol 61:123
130. Thomas CR, Paul GC, Kent CA (1994) Alberghina L, Frontali L, Sensi P (eds) Proceed-
ings of the 6th European Congress on Biotechnology, p 849
131. Michel FC, Grulke EA, Reddy CA (1992) AIChE J 38:1449
132. Shamlou PA, Makagiansar HY, Ison AP, Lilly MD, Thomas CR (1994) Chern Eng Sci
49:2621
133. Jiisten P, Paul GC, Nienow AW, Thomas CR (1996) Biotechnol Bioeng 52:672
134. Edelstein L, Hadar Y (1983) J Theor BioI 105:427
135. Tough AJ, Pulham J, Prosser JI (1995) Biotechnol Bioeng 46:561
136. Takamatsu T, Shioya S, Furuya T (1981) J Chern Tech Biotechnol 31:697
137. Cronenberg CCH, Ottengraf SPP, van den Heuvel JC, Pottel F, Sziele D, Schiigerl K, Bell-
gardt KH (1994) Bioproc Eng 10:209
138. Cui YQ, van der Lans RGJM, Luyben KCAM (1997) Biotechnol Bioeng 55:715
 Part D
Metabolite Flux Analysis, Metabolic Design
14 Quantitative Analysis of Metabolic and Signaling
Pathways in Saccharomyces cerevisiae
Klaus Mauch, Sam Vaseghi, Matthias Reuss

14.1
Introduction

The yeast S. cerevisiae has been used in the oldest biotechnical workshops and in-
dustries, baking, brewing, and wine making, from the earliest days of recorded his-
tories. In our times it is in particular the application as bakers yeast which has re-
sulted in production processes with very impressive scales of operation. With the
advent of recombinant DNA technology, the yeast has also proved to be an excellent
host for the production of a number of different proteins with exciting commercial
applications. With increasing information on metabolic processes and physiology it
can be expected that this organism becomes a very suitable cell factory for the pro-
duction of high-added-value products.
In 1996, the complete genome sequence of S. cerevisiae was deposited in the public
data libraries. However, very soon after this important event, attention has shifted to
the so-called functional analysis which is characterized by different approaches
grouped in the domains genome, transcriptome, proteome, and metabolome. The
purpose of this concerted action is to derive a better knowledge and to enable us to
understand, predict, and eventually control the metabolism of the yeast to mold it to
the purpose of old and new production processes. It is very clear that the rather
recent field of metabolic engineering, defined as the purposeful and rational design
of metabolic networks, will play an important role in this field. The challenge for the
new discipline in this difficult and ambitious endeavor is to bridge the gap between
the fingerprints of molecular assemblies and the physiological behavior of the sys-
tem. This requires an integrative biological systems analysis as the quantitative de-
scription at the hierarchical levels of molecular, cellular, and phenotypic functions
including their interaction with the environment is extremely complex. Indeed, there
is a growing awareness that this complex interplay between the genome and physio-
logical functions of the cell needs a new holistic and fully integrative view by placing
the organism, rather than the genome, at the center of investigations.
This ambitious goal, first of all, points to a pivotal problem: more specific require-
ments for the design of experiments necessary to elucidate the functions of un-
known genes as well as the complex interplay of metabolic and signaling networks.
The demand of better definition of the physiological state of the cells and the need of
new methods for sampling to analyze gene products (transcripts or proteins) as well
as metabolites has been recognized in the scientific community engaged in the func-
tional analysis of the yeast genome [1]. Experimental methods described here ad-
dress this issue of investigations at defined physiological conditions. Also measure-
436 14 Quantitative Analysis of Metabolic and Signaling Pathways

ments and analysis in ways that accurately reflect the status within the living cell are
discussed. A systematic and consequent application of these experimental strategies
for quantitative analysis of the transcriptome (DNA arrays) and proteome (two-di-
mensional-gel electrophoresis) should enable a deeper understanding at these levels
of protein biosynthesis and its regulation.
The concepts regarding the quantitative analysis of the metabolome are less clear.
Indeed, in context with the holistic and integrative view of cell metabolism there is
growing recognition that the quantitative description of metabolic and signaling
networks is strongly limited by the insufficiency of the experimental methods as well
as quantitative analysis of the data. The implications of these missing links are quite
striking in context with attempts to predict structural and functional aspects of me-
tabolic networks from the genome [2-6]. It must be stressed that mathematical ana-
lysis of reconstructed metabolic systems to predict control mechanisms of functional
blocks or even dynamic behavior of the systems are based on information about
metabolic models available in data banks. The practical feasibility of these ap-
proaches including the application of the tool boxes of Metabolic Control Analysis
(MCA) obviously depend on the quality of the data available for the kinetics of in-
tracellular reactions.
The biochemical research in the last few decades has been concerned with purify-
ing the individual enzymes and studying in isolation the chemistry of the reactions
they bring about. Most of the present knowledge about the kinetics has been derived
from these investigations in which the enzymes are isolated and thus free from in-
tracellular interference. There are two pivotal questions regarding the application of
these kinetic for studying metabolic networks: (1) to what extent does the multitude
of interacting processes inside the cell lead to kinetic behavior which differs from in
vitro conditions and (2) what is the influence of the functioning of the entire ensem-
ble; in other words - can we simply sum up every enzyme reaction to understand the
system quantitatively?
The answers to these questions depend on the possibilities of our determining the
enzyme kinetics under in vivo conditions of the living and growing cell and analyzing
these kinetics by taking into account the open character of the system. The essential
tools to tackle these ambitious tasks are steady state metabolic flux analysis, measure-
ment of intracellular metabolites at dynamic conditions, and strategies for deriving
the structure of in vivo enzyme kinetics and their parameters from these data.

14.2
Metabolic Flux Analysis
Accurate quantification of intracellular fluxes under stationary conditions (Meta-
bolic Flux Analysis) has become an indispensable tool in metabolic engineering [7,
8]. Basic requirement for metabolic flux analysis is the experimental determination
of an adequate number of metabolic fluxes as well as sufficient knowledge of the
metabolic reaction network. Particularly with regard to the validation of stoichio-
metric models, labeling experiments provide a major contribution [9, 10]. The flux
distribution observed is always an immediate consequence of regulation phenomena
on the genetic and metabolic level. Hence, metabolic flux distributions in no way
own a predictive character.
14.2 Metabolic Flux Analysis 437

Spatial separation of enzymes in cellular compartments is a characteristic property

of eukaryotic organisms to which the yeast S. cerevisiae belongs. For estimating meta-
bolic fluxes in eukaryotes, compartmentation into cytosol and mitochondria is highly
relevant because in those compartments the predominant part of the reactions within
the central metabolic pathways proceed. With regard to the control of metabolism, the
observation of transport fluxes between cell compartments is of high interest. More-
over, by using compartmented metabolic networks, the diversity in the usage of me-
tabolites in individual compartments can be represented.
Most of the work published on flux distributions in S. cerevisiae [11-14], however,
is based upon non-compartmented or partly compartmented stoichiometric models.
This seems mainly due to the incomplete knowledge on transport systems, but tech-
nical difficulties in treating the usually large compartmented models may have cir-
cumvented a broader application as well.
In this section, metabolic flux distributions in S. cerevisiae CBS 7336 are observed
at three physiological states by means of a completely compartmented stoichiometric
model. The main emphasis is placed on the estimation of fluxes crossing the inner
mitochondrial membrane as well as on energetic aspects.

14.2.1
Metabolite Balancing in Compartmented Systems

Under stationary conditions, a compartmented metabolic network can be described

by the superposition of reaction and transport according to

rr+Tt=O (14.1)

where matrix nm x n) contains the stoichiometric coefficients Yi.j of the n bio-

chemical reactions. m denotes the sum of metabolites present in each compartment.
The membership of metabolites with respect to a compartment is indicated by a
superscript. Matrix T (m x t) comprises the stoichiometric coefficients ti,j of t trans-
port equations. r denotes the vector of reaction rates, t the vector of transport rates.
A transport equation can formally be treated similar to a biochemical equation. If,
for example, metabolite A is transported from compartment ex to compartment ~ by
facilitated diffusion, the transport equation is written as

(14.2)

whereas, when A is transported by the symport of one proton H, the transport equa-
tion is formulated according to

(14.3)

The stoichiometric coefficient ti,j of transport matrix T is defined to be negative

when metabolite mj is consumed and positive at the metabolites production. When r
and t in Eq. (14.1) are used on a volumetric basis, the coefficient tj,j has to be
weighted by the ratio of volumes va, vI3 in between which metabolite mi is ex-
438 14 Quantitative Analysis of Metabolic and Signaling Pathways

changed. The transport equations, Eqs. (14.2) and (14.3), for instance, are then re-
formulated according to Eq. (14.4) and Eq (14.5), respectively:
va
_An = AI' (14.4)
Vi!

(14.5)

For a metabolic network located solely in a single compartment, multiplication of

transport matrix T with the transport vector t immediately leads to vector Q com-
prising the systems net-conversion rates. Hence, Eq. (14.1) can be simplified to
rr=Q (14.6)
Matrices rand T can be combined to matrix N according to

[r I T] m= Nv = 0 (14.7)

where vector v now contains both the reaction vector r and the transport vector t.
Methods for estimating metabolic fluxes [8, 12], data reconciliation, and error diag-
nosis [16-18] can directly be applied to Eq. (14.7).

14.2.2
Stoichiometric Model

In the following, characteristic properties of a compartmented stoichiometric model

for the yeast S. cerevisiae CBS 7336 are described. All biochemical reaction equa-
tions, abbreviations for metabolite names, and the C-mol composition of macromo-
lecules are listed in the appendix. The organism for which the model is proposed,
has been grown in continuous culture with glucose as sole carbon and energy source
and ammonia as nitrogen source at dilution rates of D=O.lOh-1, D=O.24h-1, and
D=0.33 h -I. Details on fermentation conditions, media, sampling methods and meth-
ods for the experimental determination of extracellular metabolites can be extracted
from Theobald et al. [19].
The stoichiometric model is partitioned into the compartments cytosol and mito-
chondria. EMP pathway, PPP, synthesis of nucleotides, amino acids, polysaccharides,
and reactions of the C1 - transfer are all located in the cytosol. Pyruvatecarboxylase is
assumed to be present in the cytosol as well, just as the pathways for the formation
of ethanol, acetate and glycerol. Cytosolic acetyl CoA synthetase produces acetyl
CoA which is further used as precursor for the synthesis of fatty acids and amino
acids in the cytosol. Malate dehydrogenase and citrate synthase are present in both
compartments [20]. All other enzymes of the citric acid cycle and pyruvate dehydro-
genase are located in mitochondria.
No transport systems are formulated for NADP, NADPH, NAD, NADH [21], acetyl
CoA, and oxaloacetate. As a consequence, those metabolites are blocked by the inner
mitochondrial membrane. For the transfer of inorganic phosphate [22], pyruvate
[23], and succinate, transport equations for proton symport are formulated. For the
remaining metabolites simultaneously occurring in the cytosol and mitochondria,
equations for facilitated diffusion/passive transport enable transport across the inner
mitochondrial membrane.
14.2 Metabolic Flux Analysis 439

Biochemical reactions of the respiratory chain as well as ATP regeneration by ATP

synthase are located in the mitochondria [20]. It is assumed that cytosolic NADH is
oxidized by ubiquinone and not coupled to ATP synthesis [20]. Moreover, shuttle
systems for cytosolic NADH (for an overview see [24]) are assumed to be inactive.
The efficiency of oxidative phosphorylation 11 is introduced to take into account the
unknown numbers of protons pumped from mitochondria to the outside and in turn
used to regenerate ATP. The symbolic stoichiometric coefficient 11 is normalized
such that for 11=1 a PlO-ratio of 2 results for the oxidation of cytosolic NADH and
mitochondrial FADH2 whereas a PIO-ratio of 3 results for the oxidation of mito-
chondrial NADH. 11 can be considered as a measure of the degree at which the re-
spiratory chain is coupled with ATP-regeneration driven by ATP synthase. Thus, for
11=0 reduction equivalents are oxidized, but no ATP is produced, indicating a com-
plete uncoupling.
A reaction for ATP hydrolysis is included in the stoichiometric model to make
allowance for ATP consumption through growth dependent and growth indepen-
dent maintenance (e.g., turnover of macromolecules, "futile cycles:' repair pro-
cesses).
Within the present stoichiometric model, the PPP is assumed to be the sole
source of NADPH production. In turn, NADPH serves as reduction equivalent in
anabolic reactions. Activity of transhydrogenases is assumed to be negligible. Data
on the amino acids composition of cellular protein was taken from [13], whereas
the nucleic composition of RNA has been adopted from [25]. Lipid composition
(saturated and non-saturated fatty acids, glycerol) has been analyzed by Maser
[26] and was found to be virtually constant at different dilution rates. Therefore,
the stoichiometry of lipid formation is kept at a constant value. Similarly, for the
composition of protein with respect to amino acids as well as the composition of
RNA with respect to nucleic acids, almost constant values are reported for differ-
ent growth rates in the literature [12, 25]. For each prolongation step, a demand
of four ATP for translation and a demand of one ATP for transcription is as-
sumed, and one ATP is used for the extension of a polysaccharide with one car-
bohydrate unit [27].
As the fraction of macromolecules constituting biomass is a distinct function of
the specific growth rate, the weight fraction of protein, lipids, RNA, and polysacchar-
ides has been determined experimentally for S. cerevisiae CBS 7336 [28]. Results are
depicted in Fig. 14.1. The data is in good agreement with data published on other
strains of S. cerevisiae [29,30].
Most remarkable, by increasing the dilution rate, the protein fraction increases at
the expense of polysaccharides. For the derivation of the stoichiometry for biomass
formation, an ash content of 3.8% (31) has been considered for each physiological
state.

14.2.3
Computational Aspects

Weight fractions of monomers within macromolecules as well as the weight fractions

of macromolecules within biomass are converted to integers and fractions of integers
and subsequently translated into biochemical equations for polymerization by the
programming system METAFLUX [32]. Based upon biochemical reactions edited
textually in a form similar to equations found in textbooks, the matrices rand T
440 14 Quantitative Analysis of Metabolic and Signaling Pathways

60

50

~ 40 Polysaccharides
0~

~ 30
.!.
20
RNA
11)
Ii
~.
0.09 0.12 0.15 0.18 0.21 0.24 0.27 0.30 0.33
D (h'l)

Fig. 14.1. Weight fraction of protein, lipids, RNA and polysaccharides in S. cerevisiae CBS 7336
as a function of the dilution rate

are generated automatically. To prove the consistency of the biochemical equations,

matrix E comprising the elemental composition and the charge of the metabolites
included in the stoichiometric model is created from a metabolite composition data-
base. Since matrix r multiplied by matrix E equals the zero matrix, a comprehensive
topological analysis (for an overview see [33]) is carried out revealing, for instance,
degrees of freedom, optimal and suboptimal yields, conserved moieties, unbranched
subsystems, and dead ends of the metabolic network, By analyzing its topological
structure, unexpected or incorrect properties of the metabolic network can be de-
tected in an early phase of the modeling process. Application of matrix multiplica-
tion and matrix inversion by using integers and fractions of integers allows a precise
computation of flux distributions independent of the network's size. Hence, comput-
ing the condition number of matrix N becomes dispensable. Due to the built in
symbolic computation facilities of MAPLE [34], METAFLUX allows the treatment of
unknown stoichiometric factors. Those unknowns (i.e., the P/O-ratio) may be incor-
porated in the stoichiometric matrix symbolically.

14.2.4
Results

The degree of freedom of the stoichiometric model proposed in this contribution is

found to be five. As seven substrate uptake and product production rates have been
determined experimentally at every physiological state, the degree of redundancy is
two. This redundancy has been used to reconcile the data according to a method
published by van der Heijden et al. [16, 17]. Absolute values per liter of reactor vo-
lume for the measured and reconciled net conversion rates are given in Table 14.1.
The error criteria £ computed from the differences in measured and reconciled
data [16] has been compared to the X2-distribution by using a confidence level of
95%. For each data set the ratio X2 95 11- 1 is below 0.03. From there, a significant error
resulting, for example, from excreted compounds omitted in the measurement vector
cannot be diagnosed. According to Table 14.1, metabolic flux distributions within the
central metabolic pathways are summarized in Fig. 14.2. All numbers shown are nor-
malized to the glucose uptake rate ("100").
 14.2 Metabolic Flux Analysis 441

Table 14.1. Measured Ost column) and reconciled (2nd column) net conversion rates

D QGlucose QBiomass ~thanol QGlycerol ~cetate Qco z Q02

(h~l) (mmol rlh~l) (C-mmol rl h~l)(C-mmol rl h~l)(C-mmol rl h~l)(C-mmol rl h~l)(mmoll~lh~l) (mmoll~lh~l)

0.10 16.7 16.4 64.8 65.4 0.15 0.15 0.03 0.03 <0.1 0 32 32.6 30 29.6
0.24 39.9 39.9 150.7 150.5 1.6 1.6 0.13 0.13 <0.1 0 83 85.4 79 76.9
0.33 53.1 51.9 71.6 72.4 71.6 70.5 0.18 0.18 0.9 0.9 93 93.6 19 19

At a dilution rate of D=0.10h- 1 a normalized flux of 50 is branched off into the

PPP whereas a flow of 23 is used for the formation of polysaccharides. The remain~
ing flow of 27 at the G6P branch point is funneled into the EMP pathway. The flux
into the PPP appears high compared to values obtained by radioactive labeling ex~
periments [35]. However, the flux into the PPP largely depends upon the linkage of
GDH either to NADH or NADPH which varies according to the nitrogen source [20].
A linkage of GDH exclusively to NADH results in a reduced flux of 20 into the PPP
(instead of 50 for an NADPH linked GDH).
Around a half of the pyruvate produced in the cytosol is consumed for the forma~
tion of amino acids (22), acetaldehyde (21), and cytosolic oxaloacetate (23). The
other half of pyruvate formed in the cytosol is directly transported to mitochondria
by means of proton symport. The flux to acetaldehyde is further transferred to cyto~
solie acetyl CoA where the flux is split into the flux forming fatty acids (13) and
amino acids (8). The bigger part of oxaloacetate produced in the cytosol (15) is used
for the production of aspartate whilst a flux of 8 leads to the formation of cytosolic
malate. The latter is then transported to the mitochondria where it is used to replen~
ish the citric acid cycle. Fumarate also serves as anaplerotic flux even though to a
lower extent (4).
As malate and fumarate are transported into mitochondria, a~ketoglutarate
crosses the inner mitochondrial membrane from mitochondria to the cytosol (12)
where it is transformed at the high rate of 58 to glutamate. The relatively low net flux
of a~ketoglutarate transferred to the cytosol (12) is mainly due the high rate at which
a~ketoglutarate is recovered from amino acids synthesis (46) in the cytosol.
Pyruvate dehydrogenase transforms mitochondrial pyruvate to acetyl CoA (48),
which further fuels the citric acid cycle. Carbon dioxide formed at a rate of 131 by
pyruvate dehydrogenase (48), malate dehydrogenase (48), and citrate synthase (48)
in the mitochondria clearly outweighs the carbon dioxide produced in the cytosol
(67). Likewise, the amount of reduction equivalents produced in mitochondria
(NADH 181; FADH2 36) is large compared to NADH (142) formed in the cytosol.
Between mitochondria and cytosol a flow of 328 ATP is observed. ATP transported
to the cytosol supplies the anabolic ATP demand which exceeds glycolytic ATP pro~
duction. At the same rate (328), ADP and inorganic phosphate is transported from
the cytosol to mitochondria. The high value of the ATP/ADP counter~exchange is
noteworthy as enzymes transforming metabolites at a high rate usually underlie
tight control. Otherwise pool concentrations of metabolites formed or consumed
quickly overshoot or deplete. The consequences for cell metabolism through en-
zymes operating at a high rate are all the more distinct when globally potent meta-
bolites (i.e., ATPI ADP, inorganic phosphate) are involved.
Although the absolute value of the glucose influx significantly increases (240%),
only minor changes in the distribution of metabolic fluxes are observed when shifting
 r-'-'-'-'-'-'-'-'-'-'-'-'-'-'-'~ (--'-----'-'-'-'-'-------'-----'1 I'-'-'---'-'-'-'-'-'-'-'-'-'-'-'~
polysacch • ::::
i 0 . Cytosol I i Cytosol ! N
i ""'7J: Cytosol ! o.h
@-(lLUCOSE-G6P~~~ ! ~GLUCOSE - G6~ ---{g>--1l'U5P 1 @-GLUCOSE-- Gap --@>-~U5P !
I ~ 3 ~ ! I !
I IQ ~ ~ ! . 17"'1 ~ 3 ~ jP~
i PPPDF6P}{XYLSP RIBSp6PrpP! I PPPC>F6P}{XYLSP RIB5PC> Prpp l 1 RIBsp6prppl
pppbF6PMXY~L5P
I. 0.' $ 13 17
~ I• i 03~ 14 ~ ! 03~ 5 6 •
), ), t"p~'Y' I
tryp~'Y' I t"p~'Y ~
!I
~GLYCEROL-GAP E4P F6P I ~GLYCEROL-GAP E4P F6P i ~GLYCEROL-GAP E4P Fap i
$ 0' 0 i 8
I1t;I
? 0'
eho EMP
0 i
..
'*'
3"1
0'
cho EMP
0 i•
serd 3PG cho EMP i ""<:I3PG I
...'<:I3PG I
, 1$1 CO,!$> 11301 CO,[E1> ,~ CO,~
1'Y . :;
cho<:l PEP I cho<:lPEP I cho<:l PEP I
okl,ileu,ala,v,1 1 ~ 22 akl,lleu,ala.vaI i
ok;';'au,ala,v.' ~
$<J 1 , $b O"~ ~.
om met,trypbpYR~ALD--ETOH~ met,Uypc>PYR ~ALD - ETOH-c!> om met,t"pC>PYR)--0§r-~ - ETOH 137

) I y,
~-
~
~:, ) $oom,cy& i :om g Jom,Of'
? $ ~, Yoom,c d> ~.
~s: AC E9 ~: A~C <1 ~~~ A~C'
ileu ileu 24 25 • ileu 10 8 I'
ala ala 4
leu 15
23 $1 13 •
~ I ~
pl~n arg a OAe ACCOA D pal, slear i pl~~n arg .0 OAe ACCOA C> pat, stear i
phen 0", <:I OAC ACCOAg:,,; ~y.' ! tyr aic:ar asp c::> om. Iys. cy&•• t I asp C> orn, Iys. cys, •
a.p cy&,
""
his glut
t"""
amp
I e leu
', ! his glut amp
I 11 leu I tfs glut ~::
I 4 leu I -[
..(I 0.. 0 NADH ! "0 0" 0 NADH i• "0 0,. 0 i o
'".....,
AKG FUM MAL ADP
!WI I AKG FUM MAL ADP 11521 ~ AKG FUM MAL ADP ~~ i
48 .131.Ql.-:,~,:,,-·Y!-4ijo, t":"'.@":"'~":"'· 51 .141 9 ":"''':'~'':'":''',~-:~, l-:-..6). :...,:@-:-· 7 ~-:-..:.C0..:.-:-.~"7~O, s::
(1)
f-:,~-:,~· PYR CO, AlP - r PYR CO ATP (@ rij .
PYR CO, AlP g ~ i ~~ , ' I g.
AC~OA NADH FADH, • I NADHFADHJ ACCOA NADHFADH.. g.
ACCOA
n'
§
0..
Mitochondria Mitochondria Mitochondria C/J
Qq'
:::
~
S'
CJQ
"'1:1
l
L
_____ ._._._. ___
~ • _____ . D=O,10h-
_._._._._

~
~
Fig. 14.2. Metabolic flux distributions in 5_ cerevisiae CBS 7336 as a function of the dilution rate D
'"
14.2 Metabolic Flux Analysis 443

from a dilution rate of D=0.lOh- 1 to D=0.24h- 1 • Due to an increased proportion of

protein in biomass, the flux branched into the PPP and the ATP/ADP counter-ex-
change is slightly higher. Furthermore, a rising efflux of ethanol (4) can be noticed.
Compared to flux distributions discussed previously, the situation is completely
different at a dilution rate of D=0.33 h- 1 • The drastic increase of an ethanol efflux
coincides with a decrease of the biomass yield on glucose from Y s,x=0.51
(D=0.24 h -1) to Y s,x=0.17 (D=0.33 h -1). In the literature, this phenomenon has been
described as long-term Crabtree effect [15]. Consequently, an increased flux through
the EMP is observed. Associated with the lower biomass production rate, the flux
into the PPP decreases from 50 (D=0.24 h -1) to 20 (D=0.33 h -1). At the cytosolic
pyruvate branch point, the flow towards acetaldehyde increases drastically (137).
Solely a flux of 7 reaches mitochondria as pyruvate. Accordingly, the rate at which
reduction equivalents are produced in mitochondria is significantly reduced. More-
over, at a dilution rate of D=0.33 h- 1 a disproportionate efflux of acetate (2) is ob-
served.
Most remarkable, however, is the collapse of the ATP/ADP counter-exchange. In
contrast to a value of 331 estimated at a dilution rate of D=0.24 h -\ a value of 2 is
observed for the counter-exchange at a dilution rate of D=0.33 h -1. At the corre-
sponding physiological state, the need of ATP in anabolic reactions is almost com-
pletely covered by substrate chain phosphorylation in the cytosol. Despite the low
net flow of ATP to the cytosol, a significant amount of reduction equivalents is still
oxidized in the respiratory chain. From this finding it can be concluded that either
(a) the respiratory chain is uncoupled from oxidative ATP regeneration to a high
degree or (b) the use of ATP for maintenance increases drastically. To elucidate this
result more closely, the linear relation between the consumption of ATP by mainte-
nance rm,ATP and the efficiency of oxidative phosphorylation 11 have been derived
symbolically. By weighting 11 with the proportion at which mitochondrial FADH 2 ,
NADH, and cytosolic NADH are oxidized in the respiratory chain, the effective over-
all P/O- ratio is obtained and we write
(14.8)
By relating the amount of ATP consumed for maintenance to the biomass pro-
duced (in C-mol), we obtain the maintenance coefficient k introduced by Verduyn
et al. [36]. Hence, at the three physiological states under investigation, the energetic
relations can be expressed as
D = 0.10 h- 1: 11 = 0.44 k + 0.74 or P/O = 1.04 k + 1.73 (14.9)

D = 0.24h- 1: 11 = 0.44 k + 0.74 or P/O = 0.91 k + 1.64 (14.10)

D = 0.33 h- 1: 11 = 0.44 k + 0.74 or P/O = 1.77 k + 0.03 (14.11)

From Eqs. (14.10) and(14.1l), minimal values for 11 and P10 are easily obtained for
k=O. The minimal values are given in Table 14.2.
Minimal P/O values of 0.74 and 0.70 at the dilution rates D=O.lO h- 1 and D=0.24h-\
respectively, differ only to a minor extent. Importantly, the minimal P/O values re-
ported here are already well above the P/O ratios of 1.09 [13] and 1.2 [14] reported
in the literature for S. cerevisiae at a dilution rate D=0.lOh- 1• The reason can be
specified by a detailed implementation of active transport across the inner mito-
chondrial membrane and the oxidative ATP-regeneration in this contribution. Both
444 14 Quantitative Analysis of Metabolic and Signaling Pathways

Table 14.2. Minimal values for 11 and PIO

0.10 0.74 1.73

0.24 0.70 1.64
0.33 0.01 0.03

30
D·O.10h '
§o 25 D·O.241T'

E
- 2.0 D·033h '
Q..
t-
« 1.5
(5
.§.. 1.0
o
il: 05

0.0 02 0.' 0 .6 0 .8 10
k (mol ATP I C- mol Biomass)

Fig. 14.3. PIO-ratio as a function of the maintenance coefficient k

processes are driven by the same proton gradient. Since ATP-generation and active
transport essentially have to be lumped together when a non-compartmented ap-
proach is used, the effective PIO ratio is apparently lower. Or, in other words, the
apparent loss in the efficiency of oxidative phosphorylation is partly caused by active
transport.
While 11 is a measure of the degree by which ATP regeneration is coupled to the
respiratory chain, 1-11 can be regarded as a measure for the degree of the uncou-
pling. Accordingly, the maximal degree of uncoupling at the lower dilution rates
D=O.lOh- 1 and D=0.24h- 1 is comparatively small whereas at dilution rate of
D=0.33 h- I the degree of uncoupling can almost be as high as 100%.
Figure 14.3 shows the linear relations between the P/O-ratio and the maintenance
coefficient k. A constant P/O-ratio of 1.73 for all three dilution rates would result in
k=o for D=O.lOh-I, k=0 .12 for D=0.24h- I, and k=0.94 for D=0.33h- l • In contrast,
assuming a constant k of 0.6 reported for at an average for aerobic [13, 14) and
anaerobic [30) fermentations, a vanishing degree of uncoupling results for
D=0.10h-I, 6% for D=0.24h- I, and 49% for D=0.33h- l .
The quest for the actual values for k and PIO cannot be resolved from the experi-
ments analyzed in this section. However, the breakdown of transport fluxes across
the inner mitochondrial membrane observed at a dilution rate of D=0.33 h -I sup-
ports the thesis of a distinct uncoupling of the respiratory chain with oxidative phos-
phorylation during the long-term Crabtree effect as both findings are strongly re-
lated to the properties of the mitochondrial membrane.
14.3 Measurements of Intracellular Compounds 445

14.3
Measurements of Intracellular Compounds

Beside measurement of fluxes [9, 10], it is the quantitative determination of various

biotic state variables which occupies a central place in metabolic engineering. The
ambitious goals in the quantitative description of the interplay between the hierarch-
icallevels of the genome, transcriptome, proteome, and metabolome require in vivo
measurements which can be grouped into:
- Macromolecular pools (specific compounds and integral values)
- Enzyme activities
- Metabolites and signals

For a quantitative functional analysis it is an essential prerequisite to define the

physiological state of the cells used for these measurements. Of course, this impera-
tive requires experimental conditions and related process operations which are de-
fined and reproducible. We also need to devise methods for sampling, quenching,
and extraction that ensure that the results of the subsequent analysis accurately re-
flect the status of the living cell. The concrete design of the appropriate tools and
operations in the sequence:
- Process operation (steady state, fed batch, transient)
- Sampling (time span after disturbances and frequency)
- Quenching
- Extraction
- Analysis

depend on (a) the biotic variables to be measured and (b) the information to be
derived from these observations. Attributes related to (a) include chemical and bio-
logical stability of the compound, turn-over times, analytical methods applied, etc.
The focus of the second point is the purpose for these measurements. The interest in
such measurements may be related to dynamic responses of metabolite pools to ex-
tracellular disturbances for identification of in vivo kinetics including modulation at
the metabolic level. Another focal point of these measurements could be regulation
phenomena involved in transcription, translation, or posttranslational processing.
In the following, examples for this sequence of operations are provided from work
conducted in the laboratory authors'. It is important to emphasize, that this is by no
means a complete review of the activities in this area. In particular, the interesting
field of measurements of macromolecular pools including the more specific infor-
mation about the transcriptome and proteome are left out.

14.3.1
Measurements of Intracellular Metabolites and Signals - General Tools

Now as far as we have formulated the general framework for the measurements of
intracellular compounds, let us address some of the more specific problems asso-
ciated with the observations of concentrations of intracellular metabolites. In the
preceding section it was mentioned that design of the experimental conditions is
an essential prerequisite. This can be pursued by an analogy between measurement
and modeling. Bailey [37] correctly complained that sometimes modelers have not
clearly defined the purpose of modeling and then referred to Casti [38, 39] who
446 14 Quantitative Analysis of Metabolic and Signaling Pathways

asserts an inextricable coupling between a model and its intended application. Be-
cause measurements are an integral part of the modeling cycle, the same statement
holds for the experimental design. It is impossible to evaluate the success or failure
of experimental observations without specification of the demand or use for the
data.
With the measurements of intracellular metabolites presented in the following, the
goal is clear. This data should be used for identification of in vivo enzyme kinetics
which afterwards could be applied for the purpose of Metabolic Control Analysis
(MCA).
The strategy for this in vivo diagnosis of intracellular reactions uses experimental
observations of intracellular metabolites under transient conditions. For this pur-
pose, a continuous culture (or fed-batch process at a controlled specific growth rate)
is disturbed by a pulse of glucose and changes in metabolite concentrations which
are measured within milliseconds, seconds, or minutes after the disturbance. There
are two reasons for choosing this time scale:
- Regulation at the metabolic level occurs within seconds or even faster
- Within this time scale, changes are caused only by metabolic regulation. Biosyn-
thetic reactions can be regarded as staying in a "frozen" state.

As will be shown later, there are, however, some special reactions related to fast en-
zyme modifications (such as phosphorylation) which need special consideration.
Precise measuring of intracellular concentrations in the time window of seconds re-
quires appropriate techniques for rapid sampling, inactivation of metabolic enzymes
(quenching), and extraction of metabolites, taking into account the high metabolic
turnover rates of the compounds of interest.
Figure 14.4 illustrates two different techniques developed for the aforementioned
rapid sampling and quenching experiments. Both sampling devices are connected
with a stirred tank bioreactor operating in a continuous mode. In the first approach

a r ucose pulse

.... time window: seconds

Quenching at ·20 · C

3O<N<50 " - 180 sec

\ 2 I N·l N

b Glucose injection
steam time window: milliseconds
Quenching at -196 · C
u
~~~~~

11-1

Fig. 14.4a,b. Two different sampling techniques for measurement of intracellular metabolites at
transient conditions: a manual sampling after a pulse of glucose in continuous culture; b
stopped flow technique with glucose shift with the sampling valve [40]
14.3 Measurements of Intracellular Compounds 447

[ Intracellular

~
~ Metha",,1 ·70· C ~
PCA .20· C_ -, Quenching Filtration O' C
Simultaneous
Quenching
Extraction & [ peA ·20·C
• rBoiling
I
'
Extr~ Ethanol ,

_ 1_
Measurement

Fig. 14.5. Overview on quenching and extraction procedure [40,44]

a pulse of glucose is injected into the bioreactor with a syringe to give an initial
glucose concentration of, for example, 1 g 1-1 (steady state concentration of glucose
before the pulse is ca. 20 mg t 1 ). Samples are then rapidly taken aseptically with
vacuum-sealed, precooled glass tubes through a special sampling device [19, 40].
The frequency of sampling is indicated in Fig. 14.5. The sample tubes contain an
appropriate quenching fluid depending on the metabolite to be measured (perchloric
acid solution: -20°C; Methanol: -70 °C, liquid nitrogen: -196°C). Systematic inves-
tigations have indicated that the most important quenching effect is due to the low
temperature [43].
This sampling technique can easily be automated to increase the frequency of sam-
pling [41]. However, as far as the very fast and initial response of intracellular meta-
bolites in the millisecond range is concerned, this method shows an inherent limita-
tion. The time span for the first sample after the disturbance is determined by the
mixing time of the glucose pushed into the bioreactor. Even in small laboratory re-
actors mixing times are of the order of 2-3 s.
Figure 14.4b shows a new sampling technique designed to overcome this problem.
It is based on the well-known stopped-flow-method used for fast measurements of
enzymatic reactions. As shown in Fig. 14.4b in its application to sampling from bior-
eactors, a continuous stream of biosuspension leaving the bioreactor is being mixed
with a glucose solution in a turbulent mixing chamber (mixing time a few millise-
conds). The position of the valves in the cascade illustrated in the figure then deter-
mines the residence time of the biosuspension before being quenched in the corre-
sponding sampling tube.
The main features of this sampling device may be characterized as follows. (1) The
culture remains at steady state because the organisms are stimulated by the glucose
in the mixing chamber within the valve. (2) The sampling time and reaction time are
decoupled. The volume of the individual sample can be chosen independently. (3)
The time span between glucose stimulus and first sample can be less than 100 ms.
The only limitation of the technique is the problem of oxygen limitation at aerobic
growth. Thus, the longest reaction time is determined by the oxygen consumption
rate in the sampling tube, which then depends on the organism, growth rate, and cell
concentration in the reaction. For studying the complete response we therefore re-
commend use of both sampling devices, the stopped-flow-technique for the first sec-
onds and the pulse technique with manual (or automated) sampling for longer time
periods.
Figure 14.5 summarizes some aspects of the various techniques used for the simul-
taneous or sequential operations of quenching and extraction for extra- and intra-
448 14 Quantitative Analysis of Metabolic and Signaling Pathways

cellular concentrations of metabolites. In utilizing any of these methods, attention

should be paid to ensure that:
- Metabolic activity is rapidly and irreversibly inactivated
- All the metabolites to be measured are completely extracted
- The quenching/extraction solutions used do not interfere with the assay used for
the analysis

The left hand side of Fig. 14.5 shows the well-known and often applied method of a
simultaneous quenching and extraction with the aid of perchloric acid solutions
[42]. There are some critical points with this technique which should draw attention.
Often it is claimed that the sample is quenched at -20°C if 25% solution of PCA is
used as quenching/extraction fluid. However, careful measurements with thermoele-
ments [43] indicate that, depending on the volume ratio, the temperature of the mix-
ture (sample+quenching/extraction fluid) might be as high as 10 °C for several sec-
onds immediately after sampling. Thus, in general, attention should be paid to the
choice of the volume ratio. Because of the additional neutralization, PCA -extractions
always result in high salt concentrations, which may create problems in HPLC mea-
surements. For optimal extraction results with PCA solutions, it is important to per-
form at least three freeze-thaw cycles [43].
As an alternative to the quenching with methanol - which is very efficient as far as
the time course of the temperature during sampling is concerned [43] - and addi-
tional extraction with PCA, it is also possible to combine this quenching method
with other extraction procedures. An interesting procedure has been suggested and
systematically applied by Gonzales et al. [Sl]. After quenching in methanol at -70°C
and centrifugation at -20 °C the yeast pellet is incubated in boiling ethanol (SO DC).
After evaporation of ethanol, resuspension in water, and centrifugation at 4°C, the
metabolites are measured in the supernatant.
For measurements of metabolites, a variety of analytical techniques are available.
In the laboratory authors' mainly enzymatic methods and HPLC have been used. For
more detailed information about these analytical techniques the reader is referred to
the original papers of Theobald et al. [40,44] and Mailinger et al. [SO].

14.3.2
Dynamic Response of Metabolite Pools of Glycolysis

Figure 14.6a-c show illustrative examples of some metabolite concentrations during

steady state and after a glucose pulse introduced into the fermentor. The application
of two different sampling techniques allows a resolution either in the range of milli-
seconds or seconds.
The extracellular measurements show the consumption of substrate glucose after
the pulse (marked by the arrows) and excretion of glycerol, ethanol, and acetic acid.
The changes in intracellular metabolites show the feedback regulation of glucose
uptake. Immediately after the pulse glucose appears in the cell as G6P due to the high
activity of hexokinase and a low activity of the next enzyme, PFKl. PFK1 must be
activated by the decreasing ATP and the increasing F6P, AMP, and ADP concentra-
tions. Next, glucose uptake stops the decrease in ATP level. One can imagine that a
further decrease in ATP could limit PFKl. After 30-40 s, however, the PFK1 is acti-
vated to work at a level higher than the glucose uptake and the G6P-concentration
14.3 Measurements of Intracellular Compounds 449

6 ~
;f5
~ 4
E
o£3
ill 2
8=> 1

;f
'" 0
0.15

~
E 0.10
o£
~ 0.05
2>
'" 0.00
c
!L
'j5
E
o£2
(;
c
1 -
~
;;
D
;f 0.5 -
~ 0.4
E
o£
~ 0.2
0.3
•
! 0.1
0.0
-30 30 60 90 120 150 180

a time [sec]

5"6
J A
tl 5
S
~ 4
o£3
D..
2
'"
LL

• •
B

D..
III 1
LL

~ 0.04
E
o£ 0.03
D.. 0.02
UJ
D.. 0.01

0.00
5" 0.15
tl
~ 0.10
E
o£
D.. 0.05
c:;
0.00

b
450 14 Quantitative Analysis of Metabolic and Signaling Pathways

10 ~ 3.0
~

•
A I B
9 I
2.5
8 I

~
0
7
o~ 2.0
6 Cl)
~
0
E 5 ""~ 1.5
2- 2-
0.. a. 1.0
f- \l 0
« «
2
0.5

0.0
0 30 60 90 120 ·30 0 30 60 90 120
time [sec) time [sec)

3.0
~ ~
C 1.0 0
2.5 whole cell
0.9
o~
0
2.0 W
whole cell
~ 0.8
.e>
(; ---
E 1.5
'"
.t::. ./

2-
a.
::<
«
1.0 '" , u 0 .7
.,
;.,
0>

.,
<: 0.6
I

:\./
I
/
/

cytoplasm
I
0.5 0.5 I
I
I
Y I
0.0 0.0/ ! /
-30 0 30 60 90 120 -30 0 30 60 90 120
C time [sec) time [sec)

injection or
gluco",

G6PmM
l

d time sec

Fig. 14.6.a. Changes in the levels of glucose (A), glycerol (B), ethanol (C), and acetic acid (0).
b Changes in the levels of G6P, F6P (A), Fl,6bP (B), PEP (C), and GAP/3PG (0) during steady
state conditions and after a glucose pulse. c Changes in the levels of ATP (A), AOP (B), AMP
(C), and energy charge in cytoplasm and the whole cell (0) during steady state conditions and
after a glucose pulse. d Time course of G6P: Comparison of injection- and pulse-experiments
(ll) and injection experiment (AI.,) long term pulse-experiment
14.3 Measurements of Intracellular Compounds 451

decreases. Note that phosphoglucose-isomerase maintains near-equilibrium between

G6P and F6P even under dynamic conditions.
Whereas the first two intracellular metabolites of the glycolysis have a concentra-
tion maximum shortly after the glucose pulse, Fl,6bP, and GAP remain at a high level
for a long time. This behavior may result from the interact on between glycolysis and
the pentose-phosphate-cycle. The initial increase of G6P concentration is accompa-
nied by a delayed increase of 6PG, the first metabolite in the PPP. This will cause an
increasing flux through the PPP, which most likely will serve to keep Fl,6bP, and GAP
concentrations at higher levels. The most striking result, however, concerns the oppo-
site trend of the next metabolites 3-GP and PEP. This dynamic behavior may be ex-
plained by the well known activation of the pyruvate kinase by Fl,6bP. Thus, the in-
creasing Fl,6bP-level activates the conversion of PEP on pyruvate and leads to a sharp
decrease of the PEP-concentration. If we assume that the reactions catalyzed by eno-
lase and mutase are very close to equilibrium, then 3-PG may show the same trend.
The forward activation ofPK also helps to keep the 1,3-DPG concentration low. At first
glance it seems that GAP-DH would act as a bottleneck because of the increased level
of cytoplasmatic NADH. Alcohol dehydrogenase cannot use this NADH because it is
still substrate limited (pyruvate decarboxylase shows a co-operative effect with re-
spect to its substrate, pyruvate). However, three effects counteract the increase of
NADH level: (1) the cytoplasmic artha-phosphate level which is also a substrate of
GAP-DH increases, (2) the activated PK withdraws the product (1,3-DPG) of the re-
versible reaction of GAP-DH, and (3) glycerol excretion reduces the level of cytoplas-
mic NADH. Moreover, direct oxidation of cytosolic NADH by ubiquinone [20] or
shuttle systems [24] capable of transferring cytosolic NADH to mitochondria may
also influence strongly the dynamics of the NADH concentration in the cytosol. The
intracellular adenine nucleotides respond extremely rapidly. Only 2 s after the glucose
pulse, the cytoplasmic ATP decreases to 42% of the steady state value. The first steps

Glee

Fig. 14.7. The pentose-phosphate pathway

452 14 Quantitative Analysis of Metabolic and Signaling Pathways

6 ,_-----------------------------------,
5
S 4
E
oS 3
•
II..
~ 2

o +------------------+ 3.0
2.5 S
2.0 ~
1.5 ;--
II..
1.0 !;(
0.5 ~
I--r~~r_r~~r_r,_~~~,..,.__,_~,...,..~~~+ 0.0
-30 o 30 60 90 120
lime [secl

1.1
1.0
0.9
0.8 •
I 0.7
0.6
Cl 0.5
II..
II) 0.4

0.3
0.2
0.1 +--,--.--r-.--7~~/---r------~
·30 0 3D 60 90 120 180 200
time [sec]

0.24
j''\
I
SO 0.22 NADPH
~ 0.20 I ~~. ___ .. _
~
II..
0.18 .... -.-.~ ..
~ 0.16
z. 0.14
+
&
c(
0.12
z 0.10
<='
0.08 0.5 ""g
0.4 ....
0_ E
c:::z:
0.3,gil.
02 l:C
c::c(
SZ
0.1 c::]
8 Q.to
----T------,--------,------,-------+ 0.0
-:30 0 30 60 90 120 ~
Z
time [sec)

Fig. 14.8. Time course of G6P, NADP, NADPH and ATP after a glucose pulse in a glucose lim-
ited continuous culture at D=O.l h- 1 [46]
14.3 Measurements of Intracellular Compounds 453

of glycolysis (HK, PFKl) use a large amount of ATP before the subsequent steps can
create ATP. In addition, the ATP level remains at this lower level due to further ATP-
consuming reactions stimulated by the increasing level of glycolytic intermediates.
This finding may be superimposed by both a reduced efficiency of oxidative phos-
phorylation and an increased ATP demand for maintenance. However, the total
amount of adenosine nucleotides has been found to decrease. A potential explanation
for this observation is an enhanced synthesis of RNA in consequence of stress trig-
gered by the glucose pulse. One can exclude the production of glycogen as a reason for
additional ATP consumption. Previous work showed that the glycogen level remains
constant although G6P increases. A possible explanation is the inhibition of glycogen
synthetase by increasing ortho-phosphate concentration. Figure 14.6d illustrates an
example for application of the stopped-flow sampling techniques (data points de-
picted with. indicate reaction times of 150, 300, 450, 600, and 800 ms respectively;
the open symbols 0 show the same measurements illustrated in Fig. 14.6b).
The data [40, 44] have been applied to identify the kinetic rate expressions and
parameters of the intracellular enzymes involved. The detailed results and the com-
plete dynamic model have been described in the paper of Rizzi et. a1. [45]. In the
following, more simple examples will serve to demonstrate some of the important
tools for identification of the structure of rate expressions as well as estimation for
the kinetic parameter for intracellular reactions at in vivo conditions.

14.3.3
Dynamics of the Pentose Phosphate Pathway - an Example for in vivo Diagnosis
of Intracellular Enzyme Kinetics

Transient experiments as described above were performed to study the in vivo dy-
namics of the PPP in S. cerevisiae [46]. Figure 14.7 summarizes the metabolites and
enzymes of interest.
Vaseghi et al. [46] have derived a complete model for the quantitative description
of the dynamics of the PPP. The first step towards quantitative analysis of the shunt
involves the assumption, that the first two reactions catalyzed by G6P-DH and 6PG-
DH are irreversible whereas the other reactions are assumed to be reversible and in
near-equilibrium.
The balance equation for 6PG is given by
dC6PG
-----;;[t = rG6P-DH - r6PG-DH - {l C6PG (14.12)

The last term in Eq. (14.12) represents the dilution caused by the growth of yeast.
The measurements required for the identification of the kinetics of the two enzymes
are summarized in Fig. 14.8.
In addition to the substrates G6P and 6PG we need to know the dynamic response
of the concentrations of the co-substrate NADP. Because NADPH is known as a pro-
duct inhibitor for both reactions, this co-metabolite must also be measured. A care-
ful inspection of the measured data illustrates that the level of G6P at steady state
already results in a substrate saturation of G6P-DH. Therefore, the increase of
NADPH and 6PG cannot be attributed to the influence of increased G6P concentra-
tions. A careful consideration of the literature and in vitro measurements with the
isolated G6P-DH [46] provided an interesting solution to the problem. ATP turned
out to be a strong inhibitor of both reactions. The increased flux through the PP-
454 14 Quantitative Analysis of Metabolic and Signaling Pathways

shunt can therefore be easily explained by the drop of ATP after the pulse of glucose
(Fig. 14.8). Summarizing the aforementioned characteristics of the two reactions, the
following rate expressions are suggested:

(14.13)
( 1 + CMgATP(t))
Kj,MgATP,1

Notice, that the concentrations at the right hand-side of this equation are time-
dependent. A comprehensive solution would require an extension of the balance
equations for NADP, NADPH, and ATP. To reduce the complexity of the problem,
measured data for the co-metabolites are approximated with the aid of approximate
analytical functions. The following functions have been used to fit the observed
changes of concentrations after the glucose pulse.
~ 44.1 t
C (t) = 0.9 +-----;;-Z (14.14)
G6P 48.0 + t + 0.45 t

~ 1.48 t
CNADP(t) = 0.17 - -9.-7-1-+-1-6-.1-t-+-0-.4-8-t-=-z ( 14.15)

~ 0.516 t
CNADPH(t) = 0.16 - 25.39 + 0.37 t + 0.5 tZ (14.16)

C (t) = 2.3 _ _ _ _2_9_.8_3_t_ _--:- (14.17)

ATP 29.77 + 13.42 t + 0.05 t Z

These functions are simply used to fit the data and do not have any mechanistic
background.
The next step towards estimation of the parameters is the calculation of the max-
imal rates ,.ma\ Estimates from measured enzyme activities under in vitro condi-
tions are questionable, because of the possible removal of effectors during cell dis-
ruption, shear sensitivity, effects of ion strength, protein-membrane, and protein-
protein interactions, etc. An alternative estimate is based on the rate equation and

Table 14.3. Results of the parameter identification (in vivo diagnosis) for the enzymes in-
volved in the pentose-phosphate pathway in S. cerevisiae

Parameter Value mmoll!

KNADP,l 0.116
K i,NADPH,l l.702
K i,MgATP,l 0.33
KNADP,2 l.848
K i,NADPH,2 0.055
K i,MgATP,2 0.109
14.3 Measurements of Intracellular Compounds 455

measured intracellular concentration under steady-state conditions. Let us split the

rate equation into two terms:

(14.18)

with concentration vector C containing all compounds that influence the activity and
P parameter vector of the reaction j. Let us next assume that rj has been estimated
from Metabolic Flux Analysis, C has been measured, and P is available as a first
estimate. The unknown maximal rates are given as

(14.19)

This approach assumes that during the transient enzyme concentrations remain at
their steady state value.
Combining Eqs. (14.13}-{14.19) leads to

(14.20)

and

(14.21)

which may be written as

rG6P-DH = rz:::-DH !G6P-DH(t, PG6P-DH) (14.22)

(14.23)

Membran

"c ~8
~
, : AMP
"
0v
.."
Cf6~Y3:> ... ~
: ' , .. ' mIJ )~
it.. ..-t...... . ,,.#fII>.. - - :>
'V<- - - - ~ - - - ,...
r;;-;;;-J~
~ A
,
'
. IT~P II ADP ! PF~\,,« _____ 0< - -: -- - -'
IF-l ,6.BPI',
'. "0 -c A :

I Q lAMP~ I Abp
.. i'
l:
I A~P ! '
'. -. -------- -. --
.. -----. - ------:>0
Fig. 14.9. The Ras-cAMP signal transduction pathway
456 14 Quantitative Analysis of Metabolic and Signaling Pathways

r-B o
~
Fig. 14.10. Mechanism of activation of PKA
+
o
Thus the balance equation for 6PG (Eq. 14.l3) takes the form

dC6PG
d t= max I' (
-r6PG-DH j6PG-DH t, P6PGDH
)
(14.24)
+ r'll6":-DH /G6P-DH(t,PG6PDH) - JL C6PG(t)
with/G6P-DH and!6PG_DH defined in Eqs. (14.20) and (14.21) respectively.
The software packages ACSL (Mitchel and Gauthier, Concord USA, integration
procedure Gear algorithm) and OptdesX (Design Synthesis Inc., Orem, UT, optimi-
zation strategy simulated annealing) have been applied for the estimation of the
parameter vectors PG6P-DH and P6PG-DH' The results of the parameter identification
procedure are shown in Table 14.3 The relative error square sum Crel for the time
course of the different metabolites and co-metabolites concentrations were calcu-
lated in order to assess the quality of the identification procedure. The functional
used for parameter estimation is

(14.25)

The comparison between measurements and model predictions for 6PG is shown
in Fig. 14.8.
This example not only illustrates the strategy for the in vivo identification of in-
tracellular kinetics. - a more detailed description of this approach is presented in the
papers of Rizzi et. al. [45] and Vaseghi et. al. [46] - it also indicates an interesting
result regarding the modular structure, which is important for complex network
analysis. According to the result of the identification procedure, the flux through
the PPP is independent from the concentration of the substrate G6P and only de-
pends on the ATP pool as well as the NADP/NADPH ratio. It therefore seems that
the flux is adjusted to the energy state and balanced to the demand for biosynthesis.
The PP-shunt thus acts as a functional unit, which is modulated by the energy state
of the cell and the demand in biosynthesis.

14.4
Quantitative Analysis of Glucose Induced Signal Transduction

Signal transduction is rapidly evolving into one of the major topics in biology [47].
Quantitative analysis based on in vivo measurements and mathematical modeling
will be crucial for fundamental understanding of regulation phenomena as well as
its integration into Metabolic Control Analysis (MCA). Much prior research rests on
in vitro analysis or in vivo measurements with mutants at physiological conditions,
which are not well defined. Actual focus of experimental investigations is the unra-
14.4 Quantitative Analysis of Glucose Induced Signal Transduction 457

I Pulse
\1/
0.005
a) cAMP
0.004 mM
0.003

IV--
0.002 _.
3.0
2.5
2.0

• b) ATP
1~:~
mM 0.5
r - - - - - - - - - - - - - - - - - - - - - - - \ 0.0
0.020

·0.015
c) rpFK2
0.010
mM/sec
0.005

0.000 ~r---------------------_l
0.016
d) rOeg,F26BP
0.012
mM/sec
0.008

0.004

i - - - - - - - - - - - - - - - - - - - - - - - , " 0.000
0.012 • e) F26BP
0.008
mM
0.004 •
•
0.000 - ' - - - - - , - - - - - . - - - - . - - - - - - - - - ; - - - - - - 1
o 500 1000 1500 2000

time sec
Fig. 14.11a-e. Time course of cAMP, ATP, PFK2 activity and the rate of F2,6bP degradation
after a glucose pulse in a glucose limited continuous culture at D=O.l h- 1 [43,56]

veling of the genes involved and much less emphasis is put on the kinetic analysis of
transduction cascades.
The glucose induced RAS-adenylate cyclase signaling pathway is one of the best
known and most intensively studied regulatory pathways in S. cerevisiae [47]. An
overview of the signaling pathway in response to the glucose-sensing system and
its interaction with the glycolysis is schematically illustrated in Fig. 14.9.
When glucose is added to derepressed cells of the yeast S. cerevisiae, a cAMP sig-
nal is induced which triggers a protein phosphorylation cascade. The mechanism by
which glucose initially activates the adenylate cyclase is still the subject of research.
458 14 Quantitative Analysis of Metabolic and Signaling Pathways

3H_cAMP ------+ 3H-cAMP-BP

cAMP + BP ~ cAMP-BP

Sample + 3H_cAMP + BP
!
Incubation
~
Adsorption of not bound
cAMP to charcoal

1
Centrifugation
1
Measurement of
3 H-cAMP
'"
in the supernatant

1
Calculation

Fig. 14.12. Assay for the measurement of cAMP; BP: binding protein

It is known, that both glucose addition and acidification trigger an increase of the
cAMP level.
There is increasing evidence that only acidification causes an increase in the GTP/
GDP ratio on the Ras proteins which in turn leads to the activation of adenylate cyclase
[48]. The stimulation by glucose seems to depend upon the presence of an additional
Ga-protein Gpa2. These different phenomena are important for the kind of experi-
ments described above, because the glucose pulse results in a drop of intracellular pH.
We therefore anticipate a superposition of the two effects, acidification and glucose.
The increased cAMP leads to an activation of the PKA. The mechanism of this
activation is illustrated in Fig. 14.10. The inactivated form of PKA is a tetramer con-
sisting of two subunits, the catalytic C-PKA and the regulatory R-PKA [50-55]. Two
cAMP molecules are reversibly bounded to R-PKA and are responsible for the dis-
sociation of the two catalytic subunits. These catalytic subunits are representing the
active form of PKA which in turn is catalyzing the phosphorylation of the target
proteins in the glycolysis as well as the cell cycle [47].
As far as the aforementioned signal cascade is concerned, the most important en-
zymes in the glycolysis are the PFK2, the F6P-synthetase, and the trehalose-6-phos-
phate-synthetase. The dynamics of the activation of PFK2 may serve as an example
for the quantitative analysis of such activation modules for enzymes. For in vivo
diagnosis of intracellular reactions this is a special challenge because the maximal
rate r max in the rate expression changes during the transient and, thus, the enzyme
does not stay any more in a "frozen" state.
PFK2 is activated by PKA causing an increase of fructose-2,6-bisphosphate. This
"metabolic messenger" stimulates PFK1 and strongly inhibits FBPasel. T6P-synthe-
tase, also shown in Fig. 14.9, is additionally modulated by PKA. The increase of T6P
caused by the stimulation of the enzyme results in an inhibition ofHK I and II [78].
Quantitative analysis of these regulation phenomena requires the in vivo measure-
ment of the dynamics of signal compounds such as cAMP and F2,6bP, as well as
14.4 Quantitative Analysis of Glucose Induced Signal Transduction 459

Protein-Phosphatases

Fig. 14.13. Phosphorylation of PFK2 through PKA

transient behavior of an enzyme activity (PFK2) as a rapid response to the glucose

stimulus. Both problems represent new challenges for the measurement of intracel-
lular compounds because of very low concentrations and application of modified
quenching and extraction methods to maintain enzyme activities, respectively.

14.4.1
Measurement of Intracellular cAMP

The results of the observed dynamic response illustrated in Fig. 14.11a [56] were
obtained by quenching samples in methanol at -70°C and afterwards extracting
with boiling ethanol [55]. The measurements of the cAMP concentrations were per-
formed with a commercialy available competitive protein binding assay (Amershan
International TRK 432). The method is schematically illustrated in Fig. 14.12. The
time course of the signal (Fig. 14.11a) indicate a rapid response to the glucose con-
centration which is in qualitative agreement with the observations of Beullens et al.
[57].
Unravelling of the action mechanism of this signal is complicated because most of
the experiments reported so far in the literature have been performed with glucose-
starved cells. Alternatively, glucose has been added to cells grown on a non fermen-
table energy and carbon source. As outlined by Thevelein [47], the effects observed
during these experiments might be masked by the superposition of the time course
of ATP level. This is particularly important if observations are extended to the dy-
namic behavior of protein kinases which are obviously affected by the ATP pool. As
such, the experiment reported here presents for the first time a quantitative analysis
of the dynamics of the cAMP signal at physiologically defined growth conditions.
This simultaneous observation leads to interesting clues regarding the influence of
intracellular pH and ATP upon the cAMP cascade. Due to the sharp drop in intra-
cellular ATP pool (Fig. 14.11b), pH will inevitably also drop. According to the obser-
vations of Colombo et al. [49] we must conclude that both activation effects upon

Detection at 340 nm

Fig. 14.14. Assay for the determination of F2,6bP [61]

460 14 Quantitative Analysis of Metabolic and Signaling Pathways

I Pulse
MgATP30 , -_ _-+-_ _ _ _ _ _ _ _ _ _ _ _-,
mM 2.5
I----<--i
2.0
1.5
1.0
•
0.5
• • •
0.0 - 1 - - - - - - + - - - - - - - - - - - - - - + 2.0 F6P
1.5mM

1.0

- ••--.----..o--J 0.5
F26BPo.020 0.0
mM 0.016 I
I
0.012 I
o.ooa
I
I
0.004 '~/ •
0.000
-30 0 30 60 90 120

time sec
Fig. 14.15. Time course of ATP, F6P and F2,6bP after a glucose pulse in a glucose limited con-
tinuous culture at D=O.1 h- 1 [43]

adenylate cyclase (acidification via Ras and glucose via Gpal protein) act simulta-
neously. However, we are also able to confirm quantitatively the conjectures of Co-
lombo et al. [49], that ATP drops to limiting levels for the adenylate cyclase. Careful
inspection of the time course of ATP (Fig. 14.11) clearly shows a prolonged period of
ATP levels which are far below the KM-value [49]. The increase of adenyl ate cyclase
activity may therefore be counterbalanced by the substrate limitation of ATP imme-
diately after the glucose pulse. The result is a delayed increase of the cAMP signal.

14.4.2
Measurement of the PFK2 Activity

A further step for elucidation of the interaction between the signaling pathway and
glycolysis is the observation of the dynamics of enzyme activation. In what follows
PFK2 will serve as a typical example. According to the scheme in Fig. 14.13, PFK2 is
phosphorylated by PKA which in turn is activated by cAMP. This example also
serves to introduce the experimental tools for rapid measurements of intracellular
enzyme activities. The quenching and extraction methods applied to measure en-
zyme activities are those suggested by Francois et al. [58]. Samples are taken after
the glucose pulse and quenched in methanol at -70°C. After centrifugation at -9°C
and resuspension in a buffer the cells are mechanically disrupted by intensive shak-
ing with glass beads. To prevent temperature increase, the disruption procedure is
interrupted four times to cool the sample down to O°C. After centrifugation, the
extract is incubated with the substrate F6P. Product concentration F26BP is deter-
mined with the assay described below. To measure the time course of product for-
14.4 Quantitative Analysis of Glucose Induced Signal Transduction 461

Fig. 14.16. The reactions around PFKl

R T

F6P F6P

Fig. 14.17. The mechanism of the M-model [65]

mation, the reaction is quenched at different times by adding NaOH and incubated
10 min at 80 o e.
The results of the measurements are shown in Fig. 14.11 c [43). The time course of
activity is similar to the cAMP signal in Fig. 14.11a. Attempts to correlate the two
signals with a simple Michaelis Menten type of kinetic (graded response of activated
protein to its stimulus) as well as taking into account a Hill type cooperativity
(switch-like response) [59) failed. However, recently Vaseghi could show [43) that
plotting the activity of PFK2 against cAMP resulted in a hysteresis, which points to
a "bi -stable true switch" [43, 59). Further measurements of PKA activities are re-
quired to elucidate the source of the positive feed-back, essential for this special type
of kinetics.

14.4.3
Measurements of F2,6bP

The measurement of F2,6bP concentrations rests on a special assay developed by van

Schaftingen et al. [60). This assay uses the activation of pyrophosphate dependent
PFK1 from potato tubers through F26bP. The formation of F26bP is measured with
the aid of the enzyme cascade illustrated in Fig. 14.14. Results of the measurements
after pulsing glucose to the steady state continuous culture are summarized in
Fig. 14.11c [43). It must be emphasized that only part of the time course is directly
correlated with the activity of the PFK2. The decline of the signal is also determined
by the degradation of F2,6bP, which has been studied with permeabilized cells using
a special in situ assay suggested by Vaseghi [43).
From the comparison of the time courses of ATP, F6P, and F2,6bP (Fig. 14.15) it
can be concluded that the rate of formation of F2,6bP is not influenced by the two
462 14 Quantitative Analysis of Metabolic and Signaling Pathways

ADP
T1
F-2,6-BP F-2 ,6-BP
r-··
~) -
_ r+'
\...T. /

AMP 11
F6P
AMP 11 F6P

ADP
[IJ ADP
R2
oMgATP T2 c:±::J
F-2,6-BP
e - F-2,6-BP
(0
O MgATP
e
CD -'>- +

AMP AMP

Fig. 14.18. The mechanism of the V-model [43]

substrates of the PFK2, ATP and F6P. We may therefore assert that the formation of
F2,6bP is determined by the activity of PFK2, which in turn is regulated via the
cAMP cascade [43].
Recently Vaseghi [43] has shown, that the quantitative information, which can be
extracted from these measurements could be significantly expanded by considering
the degradation of F2,6bP. This is an important issue for the dynamic transition to
the previous steady state of the culture because PFKl is strongly activated by this
signal.
The balance equation for F2,6bP reads:
dCP26bP
---;t-- = TpPK2 - TVeg,P2,6bP - p,CF2 ,6bP (14.26)

The rate of degradation TVeg,P2,6bP summarizes the action of the two specific phos-
phatases with different affinity and one unspecific phosphatase [61-64]. Using sui-
table approximation functions for C F2 ,6bP(t) and TPFK2(t), it is possible to estimate the
time course of the degradation rate. The results of these calculations are depicted in
Fig. 14.l1d [43]. The important observation is that the dilution of the F2,6bP pool
due to growth considerably contributes to the decrease of the signal [43]. This again
indicates the importance of the physiologically defined growth conditions for the
quantitative investigations of the dynamics of signal transduction processes.

14.5
Comparison Between in vitro and in vivo Kinetics -
Illustrated for the Enzyme PFK1 (Phosphofructokinase 1)

Unfortunately, most of the published work in the application of Metabolic Control

Analysis (MCA) - the important framework for rational design of cell factories - is
based on in vitro kinetics. Provided that intracellular pool concentrations are
known, we may ask the important question: to what extend does the situation inside
14.5 Comparison Between in vitro and in vivo Kinetics 463

the cell leads to kinetic behavior, which differs from the in vitro condition in which
the enzyme has been taken out of the intracellular milieu? Presumably the answer to
this question depends on the enzyme under study. Because of its complex structure
and function, the enzyme PFKl is one of the most challenging examples.
The kinetic behavior of this enzyme has attracted a lot of attention because of its
important role in the regulation of the glycolysis. The allosteric enzyme catalyses the
phosphorylation of F6P to Fl,6bP and is modulated by several effectors as schema-
tically illustrated in Fig. 14.16.
The most important attempts to model the allosteric behavior of this enzyme goes
back to the work of Monod et al. [65]. The basic structure of this model is illustrated
in Fig. 14.17.
Two conformations are considered: conformation T has a low- and confirmation R
a high affinity to the substrate F6P. In the case of PFK1, each enzyme molecule con-
sists of 8 subunits. The basic model of Monod et al.[65] only considers the allosteric
behavior with respect to the substrate F6P and a non-allosteric effect of ATP. Freyer
[66] additionally considered the modulation through AMP, ADP, and ATP and its
influence upon the allosteric effects of F6P. This model is also known as Reuter-
model [67]. In a similar way to that illustrated in the more sophisticated V-model
(Fig. 14.18), this model assumes that:
- Each protomer of the octameric protein appears in two basic conformations Rl
and R2 as well as two sub-conformations R2 and T2
- Substrate F6P is bonded to Rl and Tl with different affinities
- ATP binds to all different conformations with the same affinity
- ATP binds as inhibitor to Rl and Tl
- AMP and ADP bind as activators to Rl/Tl or R2/T2, respectively

A further approach to the kinetics of this enzyme has been suggested by Hofmann
and Kopperschlager [68] . These authors reduced the complexity of the model by
assuming that the low affinity of the enzyme in the Tl conformation can be ne-
glected.

V-Model

K 'I6lfJlJ =0
K .....,.. A =0
K"'IJI',I =0
K"·'16"~.,4 =0

F-Mode~

Fig. 14.19. Strategy for the reduction of the general V-model [43]
464 14 Quantitative Analysis of Metabolic and Signaling Pathways

[$l Sample Dilution in NaCI 0.9 %

-----+-6 ml Suspension bei 0.9-1.5 gDW/1

"""oo,~
Culture
I '" a:·:·:·":":::::· ;:;;;n~::~~::::rm, 4 min)

(0=0.1 h-1) Add 300 IJI

Toluol-Ethanol (1 :4, v/v)

r"-- Shake (5 min) mit Vortex

Centrifugation (5000 rpm, 4 min)

"'0= Store on
Ice
Resuspend
1. in 6 ml 0.1 M KPi-Buffer
2. in 1.2 ml O.SM MES

Fig. 14.20. Procedure for permeabilization of yeast cells [76]

None of these models considered the strong activation through F2,6bP [69]. Also
the competitive inhibition between F2,6bP and F1,6bP reported by several research
groups [70, 71] should be taken into account.
Vaseghi [43] has derived a complete model for PFK1 considering the concerted
action of all the aforementioned regulation phenomena. The model is schematically
illustrated in Fig. 14.18 and can be represented by the following set of equations:
aF2,6bP,I = C F2 ,6bP K F2 ,6bP,I
{ a F2 6bP A = C F2 ,6bP KF2 6bP A
(14.27)
a F ]'6bP'I = C F2 ,6bP K F ] ~bP ~
aF2:6bP:A = C F2 ,6bP KF~,6b;,A

aATP,S = CATP KATP,s

a ATP I = CATP K ATP I
aATP'A = CATP KAT; A
aAD;,I = C ADP K ADP :1
aADP,A = C ADP KADP,A
(14.28)
aADP,C = C ADP KADP,C
aAMP,I = CAMP KAMP,I
aAMP,A = CAMP KAMP,A
a F6P ,T = C F6P K F6P ,T
a F6P ,R = C F6P K F6P ,R

z= (l.0 + aATP,I) (l.0 + aADP,I) (l.0 + aAMP,I)

(14.29)
(l.0 + a F1 ,6bP,A + a F2 ,6bP,A)

N = (l.0 + aATP,A) (l.0 + aADP,A) (l.0 + aAMP,A)

(14.30)
( 1.0 + a Fl ,6bP,A + a F2 ,6bP,A)
14.5 Comparison Between in vitro and in vivo Kinetics 465

Table 14.4. Comparison of the parameters of the V-Model for in situ and in vitro conditions.
Except for Lo, all other parameters are expressed in mM- 1

Parameter in vitro in situ

Lo 40.57 3.4e-9
KATP,S 27.15 6.55
KATP,I 0.4 0.11
KATP,A 6.48 1.1
KF6P,RO 4.48 9.61
KF6P,TO 0.92 200.18
KADP,I 1.22 2.15
KADP,A 0.52 1.13
KADP,C 3.07 1.7e-8
KAMP,I Nl a Nla
KAMP,A Nl a Nl a
KF26BP,I Nl a Nl a
KF26BP,A Nl a Nl a
KFI6BP,I 28.43 0.0
KFI 6BP,A 273.2 1.74e+4
aNI: Not identifiable

ADP
CIJ (j MgATP
F-2,6-BP

0 R1

G
F-l ,6·0P
,I F6P
AMP

H
ADP
CD
F-2,6-BP .r
" .,:.
0 R2

G
F-l ,6-BP
F6P
AMP

Fig. 14.21. The reduced V-model for in situ conditions according to Vaseghi [43]

z
Q=- (14.31)
N

K F6P,R = K F6P,R,O Q (14.32)

K F6P,T = K F6P,T,O Q (14.33)

466 14 Quantitative Analysis of Metabolic and Signaling Pathways

00<- ----~
A
,

Fig. 14.22. Reactions involved in the balance of F6P

T 1 + (};P6P,T) 8
L=-=Lo ( (14.34)
R 1 + (};P6P,R

M _ _(};_P_6P_
P6P,R - (};P6P + 1
(14.35)

M _ C(};P6P
P6P,T - C(};P6P +1 (14.36)

Np6P = M p6P,R R + M p6P,T T (14.37)

(};
N _ ATP,S
ATP - (14,38)
1.0 + (};ATP,S + (};ADP,C

(14,39)

Figure 14.19 shows that this general model can easily be reduced to the existing
models of Monod et al. [65] (M-Model), Freyer [66] (F-Model), and Hofman and
Kopperschlager [68] (HK-Model).
For a more rational interpretation of the differences between in vitro and in vivo
conditions, it was decided to complement the experimental investigations with some
in situ experiments performed with permeabilized yeast cells. The procedure to pre-
pare permeabilized cells, which may offer an interesting tool for studying intracellu-
lar enzymatic rate expressions, is depicted in Fig. 14.20 according to Serrano et al.
[76].
The assay for determination of in situ activities for PFK1 has been described by
Gancedo and Gancedo [72]. It consists of an enzymatic coupling of the PFK1 reac-
tion with a glycerol-3-phosphate-dehydrogenase and a measurement of NADH.
Table 14.4 summarizes the results of the identification of the parameters of the
kinetic model using the in vitro data of Hofman and Kopperschlager [68] as well as
the in situ measurements of Vaseghi [43]. The differences in the values of the para-
meters are pronounced. According to Vaseghi [43] there are two striking phenom-
ena, which attract attention:
- Compared to the in vitro situation, the modulation strength of all the effectors are
less pronounced under in situ conditions
14.5 Comparison Between in vitro and in vivo Kinetics 467

F-2,6-BP

F6P

J1
F-2,6-BP
G)

F6P

Fig. 14.23. The reduced V-model for in vivo conditions according to Vaseghi [43]

- The parameter value of Lo tends to zero. This result leads to the interesting con-
clusion that the enzyme only acts in the R conformation (high affinity)
On the basis of these observations, the model can be reduced to a much more simple
structure illustrated in Fig. 14.21 and represented by the following expressions [43]:

Z = (1.0 + CXATP,I) (1.0 + CXADP,I) (1.0 + CXAMP,I)

(14.40)
(1.0 + CX p1 ,6bP,A + CXP2 ,6bP,A)

N = (1.0 + CXATP,A) (1.0 + CXADP,A) (1.0 + CXAMP,A)

(14.41 )
(1.0 + CXFl ,6bP,A + CXF2 ,6bP,A)
z
Q=- (14.42)
N

Kp6P,R = Kp6P,R,Q Q (14.43)

M _ CXp6P
P6P,R - -...::....:c+'----1 (14.44)
cxP6P

Np6P = MP6P ,R R + M p6P,T T (14.45)

CX
NATP -- ATP,S
(14.46)
1.0 + CXATP,S + CXADP,C

(14.47)
468 14 Quantitative Analysis of Metabolic and Signaling Pathways

Next, the identification of the kinetic parameters of the original V-model

(Fig. 14.18) is performed with the aid of the in vivo diagnosis [45], which has been
exemplary explained with the PPP in Sect. 14.3. For this purpose the measured time
courses of F6P, F1,6bP, F2,6bP, MgATP, ADP, and AMP (experimental results sum-
marized in Fig. 14.6b,c) are approximated by suitable analytical functions fi [43].
The balance equation for F6P (see Fig. 14.22) reads:

dCP6P max f. max

(
- d - = -rpPKI jPPKI t, C P6P , PPPKI
t
)
+ rpGI '
G6P jPGI t, C P6P , PPGI + rTAL t
f. ( ) ( )
(14.48)
+ rTKLZ(t) - p, C P6P (t)

The contribution from the reactions Transaldolase (TAL) and Transketolase

(TKL2) are computed from the model of the pentose phosphate shunt [46]
(Sect. 14.3). The other rate expressions are those described in the original paper of
Rizzi et al. [45]. The parameter were then estimated by minimizing the functional
A Z
_ ~ (CP6P(tk) - CF6P(tk ))
(14.49)
Crel,P6P - L C (t)
k P6P k

The results of this estimation procedure further support the conclusions from the
in situ experiments [43]. Again, the value of Lo tends to zero and indicates, that the
enzyme only acts in the R configuration. In case of the in vivo diagnosis, the sensi-
tivity of the individual parameters were further investigated by comparing the time-
dependent elasticity coefficient [43]. Accordingly, flux control coefficients are given
by [79]:

drpPKI
dPj
= f) rpPKI
f) Pj
+ [2: i
f) rpPKI f)
f) Ci f)
Ci]
Pj
(14.50)

By quantitative analysis of the hierarchy of these coefficients, Vaseghi [43] clearly

showed, that modulation of ADP, AMP, and F1,6bP is less important under in vivo
conditions. F6P, ATP, and F2,6bP are the metabolites and effectors with strongest
influence on the activity of the enzyme. Together with the result Lo= 0 and

_ T _
L---Lo
R
(1 ++
1 aF6P,R
8
a p6P,T) _
-0 (14.51 )

we conclude that cooperativity between the eight subunits of the enzyme may exist
but cannot be identified under the chosen conditions. The model therefore finally
reduces to the simple structure shown in Fig. 14.23 and rate expression:

(14.52)

(14.53)

Q=-
z (14.54)
N

KP6P,R = Kp6P,R,0 Q (14.55)

Appendix 469

M _ ap6P
P6P,R - ap6P +1 (1456)

N p6P = M p6P,R R + M p6P ,T T (14,57)

a ATP,S
NATP -_ (1458)
LO + aATP,S + aADP,C

(14,59)

To extrapolate the results obtained from the kinetic analysis of PFK1 to other en-
zymes, it is worthwhile to speculate about possible clues for these pronounced dif-
ferences between in vivo and in vitro conditions. A reasonable explanation needs to
consider the well known effects of homologous and heterologous protein interaction
inside the cell. The first effect is related to the concentration of the enzyme. Srere
[73] indicated that concentrations of glycolytic enzymes inside the cells are usually
1000 times higher than those normally applied to in vitro assays. According to Ara-
gon and Sanchez [74] the concentration of PFK1 in S. cerevisiae is of the order of
190-550 Itg/ml. The concentration of in vitro assays is 1-10 Itg/ml. The influence of
high intracellular concentrations of the enzyme on the regulation of PFK1 has been
shown by Aragon and Sanchez [74]. The second effect - heterologous intereactions -
stands for associations between enzymes and structure proteins of the cytoskeleton
[77]. Particularly for PFK1 in S. cerevisiae, KopperschHiger [75] has shown for the
first time an organized association between the enzyme molecule and the microtu-
buIes under in vivo conditions. The impacts of such associations on the kinetic be-
havior of the enzymes is not known yet. However, because of the strong relationship
between structure and function, these are strong candidates for the explanation of
the aforementioned differences between in vitro and in vivo conditions. Because of
the closer agreement between in situ and in vivo observations, the in situ experi-
ments could be viewed as interesting and alternative tools for the kinetic analysis
of intracellular reactions.

Appendix

Biochemical Reactions in the Cytosol

gk + atp = g6p + adp + h

g6p=f6p
f6p + atp = fdp + adp + h
fdp = gap + dhap
dhap = gap
gap + nad+ p = 3 pg+ nadh + h
3 pg + adp = g3p + atp
g3p = pep + h20
pep + adp + h = pyr+ atp
470 14 Quantitative Analysis of Metabolic and Signaling Pathways

glut + nM + atp = glum + adp + p + h

glut + atp + 2 nadph+2h=pro + adp + p + h20 + 2 nadp
nh4 + akg + nadph + h = glut + nadp + h20
2 atp + nh4 + co2 + h20 = carp + 2 adp + 3 h + P
om + carp + asp + atp + h20 = arg+ fum + amp + 3 P + 2h
g3p + glut + nad + h20 = ser + akg + p + h + nadh
ser + thf = gly + methf + h20
s04 + 2 atp +4nadph +2 h = sulfide+4nadp + amp + adp + 3 P + h20 +h
oac + glut = asp + akg
asp + glum + atp + 2 h20 = asn + 2h + amp + 2 P + glut
asp + atp + 2 nadph + 2h =hom + adp + p+ 2 nadp
hom+ atp + h20 = thr+ adp + p + h
2 pyr + glut + nadph + 2 h = val + akg + h20 + nadp + co2
thr + pyr + nadph + glut + 2 h = ileu + nM + nadp + h20 + co2 + akg
pyr + glut = ala + akg
2 pyr + nadph + 2 h = aki + co2 + nadp + h20
2 pep + e4p + nadph + atp = cho + adp +4 p + nadp
cho + glut + h = phen + akg + co2 + h20
cho + glut + nad = tyr + akg + co2 + nadh
2 glut + accoA + atp + nadph + h20 = om + akg + coA + adp + p + nadp + ac + h
2 glut + accoA + atp + 2 nadph + 2 nad + 2 h20 = Iys + co2 + akg + coA + amp +
2nadp+2 p+h+2nadh
cho + glum + prpp + ser = tryp + 2 P + co2 + gap + glut + h + pyr + h20
ribSp + atp = prpp + amp + h
prpp + glum + atp + 2nad+ Sh20 =his + aicar+ akg+ 2nadh + 7h+ Sp
hom + succoA + cys + mythf + 2 h20 = met + coA + suc + pyr + nh4 + h + thf + h20
ser + accoA + sulfide + h = cys + coA + ac
aki + glut + accoA + h20 + nad = leu + akg + coA + nadh + h + co2
pyr + atp + h20 + co2 = oac + adp + p + 2 h
pyr + h = ald + co2
ald + nadh + h = etoh + nad
ald + nad + h20 = ac + nadh + 2 h
dhap + nadh + h = glycerol3p + nad
glycerol3p + h20 = glycerol + p
prpp + 2 glum + gly + S atp + asp + fthf + 4 h20 + co2 = aicar + S adp + 7 P +
2 glut + thf + fum + 9 h
aicar + fthf = thf + imp + h20
imp + asp + atp = amp + adp + p + fum + 2 h
imp + nad+ 2 atp + glum + 3 h20 = gmp + 2 adp + 2 P + glut + nadh + 4h
carp + asp + nad+prpp =ump + nadh + co2 + 3 P + h
ump + 2 atp = utp +2 adp
utp + glum + atp + h20 = ctp + adp + p + 2 h + glut
ctp + 2 adp = cmp + 2 atp
amp + atp= 2 adp
Appendix 471

g6p + 2 nadp + h20 = ribu5p + 2 nadph + co2 + 2 h

ribu5p = rib5p
ribu5p = xy15p
rib5p + xy15p = sed7p + gap
sed7p + gap = f6p + e4p
xy15p + e4p = f6p + gap
ac + coA + atp + h20 = accoA + amp + 2 P + h
8 accoA + 7 atp + 14nadph + 6h + h20 = pal + 8 coA + 7 adp + 14nadp + 7p
pal + 2 nadph + h + accoA + atp = stear + adp + p + 2 nadp + coA
pal + nadh+ 02 + h= palen + nad+ 2 h20
stear + nadh + h + 02 = 01 + nad + 2 h20
01 + nadh + h + 02 = linol + nad + 2 h20
mal + nad= oac + nadh +h
oac + accoA + h20 = isocit + coA + h
thf + atp + nadh + co2 = fthf + adp + p + nad
thf+ co2 + 3 nadh + 3 h = mythf+ 3 nad+ 2 h20
methf + 2 nad + 2 h20 = thf + co2 + 2 nadh + 2 h

Synthesis of Macromolecules and Biomass

11 g6p + 10 atp + 10 h20 = polysacch + 20 P + 10 adp + 10 h

91872/317671 pal + 196680/317671 01 + 330660/31767l palen + 23661/31767l stear + 6001
31767llinol + 309540/31767l glycerol = lipids + 2 h20
193920123593 glut + 68175123593 glum + 106050123593 pro + 103020123593 arg + 1848301
23593lys + 119685/23593 ser + 187860/23593 gly + 4444/23593 cys + 1908901
23593 asp + 65650123593 asn + 124230/23593 thr + 31815123593 met + 1242301
23593 ileu + 293910123593 ala + 172710/23593 val + 199980123593 leu + 848401
23593 phen + 66660123593 tyr + 17574123593 tryp + 42420/23593 his + 400 atp + 300 h20 =-
protein + 400 adp + 400 h + 400 P
277027711189 amp + 8320832/3567 gmp + 364036411189 ump + 2800281
123 cmp + 10000 atp = rna + 10000 adp + 10000 P + 10000 h
57634192001228923467 lipids + 124756940001228923467 polysacch + 32021001
228923467 rna + 46500314001228923467 protein = 100 bio

Biochemical Reactions in the Mitochondria

atp + h20 = adp + p + h

pyr + nad + coA = accoA + nadh + co2
oac + accoA + h20 = isocit + coA + h
isocit + nad = akg + nadh + co2
akg + coA + nad = succoA + nadh + co2
succoA + adp + p = suc + atp + coA
suc + fad = fum + fadh2
fum + h20 = mal
mal + nad= oac + nadh + h
472 14 Quantitative Analysis of Metabolic and Signaling Pathways

Oxidative Phosphorylation

Komplex I: nadhmit + h mit + UQ + 3 h mit = UQHZ + nadmit + 3 h cyt

Komplex II: fadhZ mit + UQ = UQHZ + fad mit
nadh cyt + h cyt + UQ = UQHZ + nadcyt
Komplex III: UQHZ + CYTc + 3 h mit = CYTcHZ + UQ + 3 h cyt
Komplex IV: CYTcHZ+ 1/2 ozmit + 3 h mit = CYTc + hZo mit + 3 h cyt
ATP-Synthase: h cyt + 113 11 adpmit + 113 11 pmit + 113 11 h mit = h mit + 113 11 atpmit
+ 113 11 hZo m1t

Transport Mitochondria-Cytosol

adpmit = adpcyt
akg mit = akgcyt
atpmit = atpCyt
coZ mit = cozcyt
coAmit = coAcyt
fum mit = fum cyt
h mit = h cyt
hZo mit = hZo cyt
isocitmit = isocitcyt
mal mit = malcyt
ozmit = ozcyt
pyrmit+hmit + adpmit + pmit = pyrcyt + atpmit + hZomit
Z pmit + h mit + adpmit = pcyt + atpmit + h20 mit
SUC mit + Z h mit + Z adpmit + Z pmit = suc cyt + Z atpmit + Z hZo mit

Transport Cytosol-Extracellular

ac cyt = ac ext
bio cyt = bio ext
cozcyt = coZ ext
etoh cyt = etoh ext
glcCyt = glc ext
glycerol cyt = glycerolext
hcyt=hext
hZo cyt = hZo ext
ozcyt = ozext
nh4 cyt + h cyt + adpcyt + pcyt = nh4ext + atpcyt + hZo cyt
Z pcyt + h cyt + adpcyt = pext + atpcyt + hZo cyt
so4 cyt + hcyt + adpcyt + pcyt = so4 ext + atpCyt + hZocyt

Metabolites: Cytosol

accoA acetyl-CoA
Appendix 473

ac acetate
adp adenosine diphosphate
aicar S-amino-4-imidazolecarboxamide ribotide
akg alpha-ketoglutarate
aki alpha-ketoisovalerate
ala L-alanine
aid formaldehyde
amp adenosine monophosphate
arg arginine
asn asparagine
asp aspartate
atp adenosine triphosphate
bio biomass
carp carbamyl phosphate
cho chorismate
cmp cytidin-S-monophosphate
co2 carbon dioxide
coA coenzyme A
ctp cytidin-S-triphosphate
cys cystein E
dhap dihydroxyacetone phosphate
e4p erythrose 4-phosphate
etoh ethanol
f6p fructose-6-phosphate
fdp fructose-l ,6-phosphate
fthf formyltetrahydrofolate
fum fumarate
g6p glucose-6-phosphate
gap glyceraldehyde-3-phosphate
glc glucose
glum glutamine
glut L-glutarnate
gly glycine
glycerol glycerol
glycerol3p glycerol-3-phosphate
3pg 3-phospho-d-glyceroyl-phosphate
g3p glycerol-3-phosphate
gmp guanosin-S-monophosphate
h proton
h20 water
his histidine
hom homoserine
ileu isoleucine
imp inosin-S-monophosphate
isocit iso-citrate
leu leucine
linol cis-9,12-octadecenoate (linoleic acid)
lipids lipids
lys lysine
mal malate
met methionine
methf methylentetrahydrofolate
mythf methyltetrahydrofolate
nad diphosphopyridindinucleotide, oxidized
nadh diphosphopyridindinucleotide, reduced
nadp diphosphopyridindinucleotide-phosphate, oxidized
474 14 Quantitative Analysis of Metabolic and Signaling Pathways

nadph diphosphopyridindinucleotide-phosphate, reduced

nh4 ammonium ion
02 oxygen
oac oxalacetate
01 cis-octadecenoate (oleic acid)
orn ornithine
p inorganic phosphate
pal hexadecanoate (palmitic acid)
palen cis-hexadecenoate (palmitinoleic acid)
pep phosphoenolpyruvate
phen phenylalanine
polysacch polysaccharides
pro prolin
protein protein
prpp phosphoribosylpyrophosphate
pyr pyruvate
rib5p ribose 5-phosphate
ribu5p ribulose 5-phosphate
rna ribonucleic acid
sed7p sedheptulose 7-phosphate
ser serine
s04 sulfate ion
stear octadecenoate (stearic acid)
suc succinate
succoA succinyl CoA
sulfide sulfide
thf tetrahydofolate
thr threonine
tryp tryptophan
tyr tyrosine
ump uridin-5-monophophate
utp uri din -5-triphophate
val valine
xyl5p xylulose 5-phosphate

Metabolites: Mitochondria

accoA acetyl-CoA
adp adenosine diphosphate
akg alpha-ketoglutarate
atp adenosine triphosphate
co2 carbon dioxide
coA coenzyme A
CYTc cytochrome c, oxidized
CYTcH2 cytochrome c, reduced
fad flavin -adenine-dinucleotide, oxidized
fadh2 flavin -adenine-dinucleotide, reduced
fum fumarate
h proton
h20 water
isocit iso-citrate
mal malate
nad diphosphopyridindinucleotide, oxidized
nadh diphosphopyridindinucleotide, reduced
oac oxalacetate
p inorganic phosphate
References 475

pyr pyruvate
suc succinate
succoA succinyl CoA
UQ ubiquinone, oxidized
UQH2 ubiquinone, reduced

Cmol-Composition, Charge and Molecular Weight (MW) of Macromolecules and Biomass

Protein CI Hl.58 NO.271 charge: -0.00843 MW 22.5 g Cmol- I

0 0.315 SO.0032
RNA C I Hl.024 NO.383 charge: -0.2113 MW 33.6 g Cmol- I
0 0.747 PO.1056
Lipids C I Hl.798 charge: -0.5516 MW 16.0 g Cmol- I
0 0.134
Polysaccharides CI H1.682 charge: -0.0303 MW 28.5 g Cmol- I
0 0.894 PO.0152
Biomass
D=0.33 h- I CI Hl.61 NO.18B charge: -0.0187 MW 23,6 g Cmol- I
0 0.45 P0.0047
SO.00214
D=0.24 h- I C I Hl.59 NO.181 charge: -0.0302 MW 24.2 g Cmol- I
0 0.58 PO.OlO3
SO.00185
D=0.10 h- I C I Hl.61 NO.161 charge: -0.0267 MW 24.6 g Cmol- I
0 0.521 PO.0091
SO.00171

References
1. Oliver SG (1997) Curr Opin Genet Dev 7:405
2. Selkov E, Basmanova S, Gaasterland T, Goryanin I, Gretchkin Y, Maltsev N, Nenashev V,
Overbeek R, Panyushkina E, Pronevitch L, Selkov EJR, Yunus I (1996) Nucl Acids Res
24:26
3. Selkov E, Maltsev N, Olsen GJ, Overbeek R, Whitman WB (1997) Gene 197:GC11
4. Overbeek R, Panyushkina E, Pronevitch L, Selkov EJR, Yunus I (1996) Nucl Acids Res
24:26
5. Overbeek R, Larsen N, Smith W, Maltsev N, Selkov E (1997) Gene 191:GC1
6. Bork P, Koonin EV (1998) Nature Genet 13:313
7. Bailey JE (1991) Toward a science of metabolic engineering. Science 252:1668-1675
8. Stephanopoulos GN, Aristidou AA, Nielsen J (1998) Metabolic engineering. Academic
Press, New York
9. Marx A, de Graaf AA, Wiechert W, Eggeling L, Sahm H (1996) Biotechnol Bioeng 49:111
10. Zupke C, Stephanopoulos G (1995) Biotechnol Bioeng 45:229
11. Cortassa S, Aon JC, Aon M (1995) Biotechnol Bioeng 47:193
12. Nissen T, Schulze U, Nielsen J, Viladsen J (1997) Microbiology 143:203
13. Vanrolleghem PA, de Jong-Gubbels P, van Gulik WM, Pronk JT, van Dijken JP, Heijnen JJ
(1996) Biotechnol Prog 12:434
14. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
15. Nielsen JVJ (1994) Bioreaction Engineering Principles. Plenum-Press, New York
16. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KCHAM (1994) Biotech-
nol Bioeng 43:11
17. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KCHAM (1994) Biotech-
nol Bioeng 43:3
18. van der Heijden RRJM, Romein B, Heijnen JJ, Hellinga L, Lyben KCHAM (1994) Biotech-
nol Bioeng 44:781
476 14 Quantitative Analysis of Metabolic and Signaling Pathways

19. Theobald U, Mailinger W, Baltes M, Rizzi M, Reuss M (1997) Biotechnol Bioeng 55:305
20. Gancedo CSR (1989) Energy- yielding metabolism In: Rose AHH (ed) The yeasts, vol 3.
Academic Press, London, p 205
21. Kramer R, Palmierei F (1992) Metabolite carriers in mitochondria. In: Ernster L (ed)
Molecular mechanisms in bioenergetics. Elsevier Amsterdam, p359
22. Cockburn M, Earnshaw P, Eddy AA (1975) J Biochem 146:705
23. Biinger R, Mallet RT (1993) Biochim Biophys Acta 1151:223
24. de Vries S, Marres CAM (1987) Biochim Biophys Acta 895:205
25. de Robichon-Szulmajster H, Surdin-Kerjan H (1971) Nucleic acid and protein synthesis in
yeast:regulation of synthesis and activity. In: Rose AHH (ed) The yeasts, vol 2. Academic
Press, London, p 335
26. Maser F (1997) Experimentelle und modellgestiitzte Untersuchungen zum Lipid-und Pro-
teingehalt in S. cerevisiae unter definierten physiologischen Bedingungen. Institut fiir
Bioverfahrenstechnik, Universitat Stuttgart
27. Stryer L (1988) Biochemistry. Freeman, New-York
28. Fuchs H (1997) Makromolekulare Zusammensetzung von S. cerevisiae bei verschiedenen
Verdiinnungsraten im aeroben, glukoselimitierten Chemostaten. Institut fiir Bioverfahren-
stechnik, Universitat Stuttgart
29. Oura E (1972) Biotechnol Bioeng 28:415
30. Verduyn C, Postma E, Scheffers WA, van Dijken JP (1990) J Gen Microbiol136:395
31. Spieth A (1997) Experimentelle Untersuchungen zur Bestimmung des Gehalts an Polysac-
chariden, RNA und Asche in S. cerevisiae unter definierten physiologischen Bedingungen.
Institut fiir Bioverfahrenstechnik, Universitat Stuttgart
32. Mauch K, Arnold S, Posten C, Reuss M (1997) Computer algebra systems in model build-
ing and model analysis for bioprocesses. 15th IMACS World Congress
33. Heinrich R, Schuster S (1996) The regulation of cellular systems. Chapman and Hall,
New-York
34. Char BW, Geddes KO, Keith 0, Gonnet GH, Leong BL, Monagan MB, Watt SM (1991).
Maple V reference manual. Springer, Berlin Heidelberg New York
35. Gancedo JM, Lagunas R (1973) Plant Science Letters 1:193
36. Verduyn C, Stouthamer AH, Scheffers WA, van Dijken JP (1991) Antonie van Leeuwen-
hoek 59:49
37. Bailey JE (1998) Biotechnol Prog 14:8
38. Casti JL (1992) Reality rules: I. Picturing the world in mathematics, the fundamentals.
Wiley, New York
39. Casti JL (1992) Reality rules: II. Picturing the world in mathematics, the frontier. Wiley,
New York
40. Theobald U, Mailinger W, Reuss M, Rizzi M (1993) Anal Biochem 214:31
41. Schafer U, Weuster-Botz D, Wandrey C (1998) Dechema Jahrestagungen 98, 26-28. Mai
1998, 16. Jahrestagung der Biotechnologen, Band 1:353
42. Kopperschlager G, Augustin HW (1967) Fehlermoglichkeiten bei der Bestimmung von
Metabolitgehalten in Hefezellen. Experientia 23:623
43. Vaseghi S (1999) Modellgestiitzte Analyse des PFK1/PFK2 Systems in S. cerevisiae. PhD
Thesis, Universitat Stuttgart, Stuttgart, Germany
44. Theobald U (1995) Untersuchungen zur Dynamik des Crabtree-Effekts. PhD Thesis, Uni-
versitat Stuttgart, Stuttgart, Germany und VDI Verlag, Diisseldorf, Germany
45. Rizzi M, Baltes M, Theobald U, Reuss M (1997) Biotechnol Boeng 55:592
46. Vaseghi S, Baumeister A, Rizzi M, Reuss M (1999) Metabol Eng 1:128
47. Thevelein JM (1994) Yeast 10:1753
48. Thevelein JM, Beulens M, Honshoven F, Hoebeeck G, Detremerie K, Griewel B, den Hol-
lander JA, Jans AWH (1987) Gen Microbiol 133:2197
49. Colombo S, Ma P, Cauwenberg L, Winderickx J, Crauwels M, Teunissen A, Nauwelaers D,
de Winde JH, Growa M-W, Colavizza D, Thevelein JM (1998) EMBO J 17:3326
50. Hixson CS, Krebs EG (1980) J Bioi Chern 255:2137
51. Matsumoto K, Uno I, Oshima Y, Ishikawa T (1982) Proc Nat! Acad Sci USA 79:2355
52. Wingender-Drissen R (1983) FEBS Lett 163:28
References 477

53. Cannon JF, Tatchell K (1987) Mol Cell Bioi 7:2653

54. Toda A, Cameron S, Sass P, Zoller M, Wigler M (1987) Cell 50:277
55. Toda A, Cameron S, Sass P, Zoller M, Scott JD, McBuilen B, Hurwitz M, Krebs EG, Wigler
M (1987) Mol Cell Bioi 7:l371
56. Macherhammer F, Vaseghi S, Rizzi M, Reuss M (1998) Modelling of glycolysis in yeast:
consideration of signal transduction dynamics. Poster presentation, Metabolic Engineer-
ing II, Elmau, Germany
57. Beullens M, Mbonyi K, Geerts L, Gladins D, Detremerie K, Jans AWH, Thevelein JM
(1988) Eur J Biochem 172:227
58. Fran<;ois J, van Schaftingen E, Hers H-G (1984) Eur J Biochem 145:187
59. Ferrel JE (1998) TIBS 23:461
60. van Schaftingen E, Lederer B, Bartons R, Hers H-G (1982) Eur J Biochem 129:191
61. Fran<;ois J, van Schaftingen E, Hers H-G (1988) Eur J Biochem 171:599
62. Kretschmer M, Schellenberger W, Hofmann E (1987) J Theoret Bioi 127:181
63. Kessler R, Gartner G, Schellenberger W, Hofmann E (1991) Biomed Biochim Acta 50:851
64. Purwin C, Laux M, Holzer H (1987) Eur J Biochem 164:27
65. Monod J, Wyman J, Changeux, J-p (1965) J Mol Bioi 12:88
66. Freyer R (1977) Phosphofructokinase aus Hefe: Kinetisches Verhalten verschiedener En-
zymformen und Diskussion eines mathematischen Modells. PhD Thesis, Karl-Marx-Uni-
versitat Leipzig, Leipzig
67. Reuter R, Eschrich K, Schellenberger W, Hoffmann E (1979) Acta Bioi et Med Germ
38:1067
68. Hofmann E, Kopperschlager G (1982) Meth Enzymol 90:49
69. Kessler R, Schellenberger W, Nissler K, Hofmann E (1988) Biomed Biochim Acta 47:221
70. Lederer B, Vissers S, van Schaftingen E, Hers H-G (1981) Biochem Biophys Res Commun
103:1281
71. Przybylski F, Otto A, Nissler K, Schellenberger W, Hofmann E (1985) Biochim Biophys
Acta 831:350
72. Gancedo JM, Gancedo C (1971) Arch Microbiol 109:132
73. Srere PA (1967) Science 158:936
74. Aragon JJ, Sanchez V (1985) Biochem Biophys Res Commun l31:849
75. Kopperschlager G (1999) personal communication
76. Serrano R, Gancedo JM, Gancedo C (1973) Eur J Biochem 34:479
77. Ovadi J (1995) Cell architecture and metabolism. Springer, Berlin, Heidelberg New York
78. Entian KD, Zimmerman FK (1997) Yeast sugar metabolism: biochemistry, genetics, bio-
technology and application. Technomic Publishing, Lancaster, Basel
79. Mauch K, Arnold S, Reuss M (1997) Dynamic sensitivity analysis for metabolic systems.
Chern Eng Sci 52:2589
80. Mailinger W, Baumeister A, Reuss M, Rizzi M (1998) J Biotechnol 63:155
81. Gonzales B, Frqancois J, Renaud M (1997) A rapid and reliable method for metabolite
extraction in yeast using boiling buffered ethanol. Yeast 13(14):1347-1355
15 Metabolic Analysis of Zymomonas mobilis

Albert A. de Graaf

15.1
Introduction

15.1.1
Zymomonas mobilis

Zymomonas mobilis is an anaerobic, Gram-negative bacterium producing ethanol

from glucose via the Entner-Doudoroff (2-keto-3-deoxy-6-phosphogluconate, KDPG)
pathway in conjunction with the enzymes pyruvate decarboxylase and alcohol dehy-
drogenase. The organism was originally isolated from fermenting sugar-rich plant
saps, e.g., in the mexican pulque drink made from agave sap, or in palm wines of
tropical Africa [1]. The carbohydrate metabolic pathways of Z. mobilis have been
reviewed recently [2] and are shown in Fig. 15.1.
Being an anaerobic bacterium, Z. mobilis still has retained functions of the elec-
tron transport chain and grows in micro aerobic environments [3], illustrating its
closeness to aerobic relatives Gluconobacter and Acetobacter [4]. Indeed, the beha-
vior of this anaerobic bacterium under aerobic conditions is currently receiving in-
creased attention in connection with its membrane-associated ATPase [5,6], respira-
tory chain-linked NADH oxidase system [7], and a d-Iactate oxidase component of
its respiratory chain [8].
The ethanol yield from glucose fermented by this remarkable organism can be as
high as 97-98% of the theoretical value [2, 9]. Only insignificant amounts of by-
products are produced by Z. mobilis growing on glucose [10, 11]. However, although
the pathway for glucose and fructose degradation is almost identical (it differs only
in the first two steps), biomass and product yields on fructose are lower than on
glucose and more by-products are generated. Especially at high fructose concentra-
tions, growth is slower, biomass yields are only half of those on glucose, and ethanol
yields are only 90% of the theoretical value [12-14]. This is due to an increased
formation of side products, especially glycerol and dihydroxyacetone [15-18]. While
these compounds were shown to derive from glyceraldehyde-3-phosphate via dihy-
droxyacetone phosphate and glycerol-3-phosphate [19], the metabolic regulations
leading to the increased by-product formation are not yet understood.
15.1 Introduction 479

o r r
Glucose

ATP

Glucose 6-P
3

..ill...
Fructose

0
Fructose 6-P
ATP
Xylose
\.~
~
Xylulose
ATP-\@
~ Xylulose 5-P
0 ..L-. NAD(P)H tt CO2 ~¥"
Rul5P
iI
!
6-P-Gluconolactone
(.;\5
~

x--
Rib5P I

@! :
\V 20
6-P-Gluconate k
GAP Sep7P
I I ~
KDPG I .. ---~
~
. ----')§:"
Ery4P
o ~
GAP.---" '--
@,t.NADH
1,3 di-p-glycerate

0,t.ATP
3-p-glycerate

@~
2-P-glycerate

ATP ~~
Pyruvate. \ PEP

~
I

@ NADH

9@
CO Acetaldehyde \ • Ethanol
2

Fig. 15.1. Scheme of the catabolic pathways in wild-type and xylose-degrading recombinant
Zymomonas mobilis. Abbreviations: KDPG=2-k~to-3-deoxy-6-phosphogluconate, GAP=glycer-
aldehyde-3-phosphate, PEP=phosphoenolpyruvate, Rul5P=Ribulose 5-phosphate, Rib5P=Ri-
bose 5-phosphate, Sed7P=sedoheptulose 7-phosphate, Ery4P=erythrose 4-phosphate. Enzyme
activities are as follows: 1, glucokinase; 2, fructokinase; 3, phosphoglucose isomerase; 4, glu-
cose 6-P dehydrogenase; 5, 6-P-gluconolactonase; 6, 6-P-gluconate dehydratase; 7, KDPG aldo-
lase; 8, glyceraldehyde-P dehydrogenase; 9, phosphoglycerate kinase; 10, phosphoglycerate
mutase; 11, enolase; 12, pyruvate kinase; 13, pyruvate decarboxylase; 14,15, alcohol dehydro-
genases (two isoenzymes); 16, xylose isomerase; 17, xylulokinase; 18, ribulose 5-P epimerase;
19, ribose 5-phosphate isomerase; 20, transketolase; 21, transaldolase; 22, 6-phosphogluconate
dehydrogenase. Adapted from [20, 159]

15.1.2
Substrate Spectrum Engineering

While Z. mobilis is superior to yeasts with respect to ethanol productivity (3-5 times
higher), sugar tolerance (up to 400 g r 1), and resistance to high ethanol concentra-
tions (12-13%), it has not become a serious competitor because of its narrow sub-
strate spectrum. The wild-type organism only grows on glucose, fructose, and su-
crose. Several attempts to engineer Z. mobilis into an efficient ethanol producer from
480 15 Metabolic Analysis of Zymomonas mobilis

abundant and renewable lignocellulosic biomass (wood, straw, milk whey, agricul-
tural residues) have been carried out. Studies until 1993, reviewed in [20] met with
limited success: only a partial and very slow conversion oflactose (which constitutes
up to 75% of the dry weight of whey [21]) and cellobiose (a main product of cellu-
lase-catalyzed breakdown of cellulose) could be achieved, mainly due to limitations
in the substrate uptake. However, in recent years several breakthroughs were
achieved. Zhang et al. [22] reported the successful construction of a Z. mobilis strain
that was able to convert xylose with an 86% of theoretical yield to ethanol. To this
end, Escherichia coli genes encoding xylose isomerase, xylulokinase, trans aldolase,
and transketolase had been introduced and expressed in Z. mobilis. Growth and pro-
ductivity on xylose, however, were about five times slower than on glucose. More
recently, the same research group succeeded in engineering an efficient arabinose-
fermenting Z. mobilis strain [23]. While xylose is the predominant pentose sugar
derived from the hemicellulose of most hardwood feedstocks, arabinose is an impor-
tant constituent of various agricultural residues and other herbaceous crops. The
arabinose-fermenting strain was constructed by introducing the E. coli genes for L-
arabinose isomerase, L-ribulokinase, L-ribulose-5-phosphate-4-epimerase, transal-
dolase, and transketolase in Z. mobilis. Growth and ethanol productivity on the
new sugar were only 2-3 times slower than on glucose and the ethanol yield based
on consumed substrate was as high as 98% of the theoretical yield. Weisser et al. [24]
constructed a strain that used mannose as a novel substrate after introduction of the
phosphomannose-isomerase gene from E. coli. Uptake of this substrate was shown to
be due to the glucose facilitator of Z. mobilis, while phosphorylation was due to a
side activity of the resident fructokinase. After plasmid overexpression of fructoki-
nase activity, growth rates of 0.25 h -I were obtained and the growth yield was better
than on fructose. The construction of an alternative xylose-degrading strain using
the genes encoding xylose isomerase and xylulokinase from Klebsiella pneumoniae
in addition to the E. coli genes for transaldolase and transketolase has also recently
been completed (G. Sprenger et al., manuscript in preparation).

15.1.3
Purpose

The recent progress in the engineering of a broader substrate spectrum for Z. mobilis
shows that the organism in principle is able to ferment different carbon sources to
ethanol with yields generally above 85-90% of theoretical yields. However, growth
rates, growth yields, and by-product formation vary widely between the reported
strains. This suggests that, although the glycolytic enzymes are always able to direct
the main carbon flux to ethanol, metabolic bottlenecks and/or a considerably dis-
turbed secondary metabolism occur in different strains or with different substrates.
Therefore, to a certain extent the intracellular metabolism of Z. mobilis may still be
regarded as black box machinery where we are able to exchange or modify some
parts, but of which the details of its functioning are poorly understood. It is the
purpose of this contribution first to give an overview of state-of-the-art methods
for metabolic analysis and, second, to give an overview of those studies that have
been devoted to the characterization by such techniques of the intracellular metabo-
lism of Z. mobilis, with the aim of providing a more detailed insight into the meta-
bolic processes and their regulation as they function in vivo in this intriguing mi-
croorganism. The availability of such knowledge is a prerequisite for a continuing
15.2 Methods for Metabolic Analysis 481

improvement of biotechnologically relevant strains by purposeful metabolic engi-

neering.

15.2
Methods for Metabolic Analysis

15.2.1
Introduction

The process of characterizing the turnover rates of intracellular metabolism can be

referred to as Metabolic Flux Analysis. For the purpose of this overview, only regula-
tion of metabolic rates on the level of enzyme activity is considered. Thus, we will be
dealing with metabolic events that are either in a steady state or represent a short-
term transient reaction to some change of environmental parameters. Studies focus-
ing on metabolic regulation at the molecular level and its long-term effects on gene
and enzyme expression are not covered. The relevant parameters to describe meta-
bolism then are, first, the metabolic reaction rates themselves, second, the enzyme
levels, and third, the concentrations of the total of all effectors influencing the activ-
ities of the enzymes involved. It is important to note here that enzyme activities
alone, especially as determined from cell-free extracts, can at best give a qualitative
impression of which pathways are active in vivo and do not allow one to draw con-
clusions about the relative activities of parallel pathways. Enzyme activity determi-
nation assays are more or less standard and will not be discussed here. It is never-
theless worth pointing out that determination of certain enzymes in Z. mobilis is far
from trivial and requires extensive treatment of the cell extract preparations to re-
move competing overwhelming activities of the glycolytic enzymes [25, 26). The
remaining parameters for a description of metabolism are metabolite concentrations
(i.e., pool sizes) and metabolic reaction rates (i.e., fluxes). Several approaches that
can be used to quantify these can be distinguished and will be reviewed now.

15.2.2
Metabolite Pool Determination

15.2.2.1
Invasive Approaches

The usual approach to measure the intracellular pool sizes is to take a sample of cells
from the reactor, stop the metabolism by some procedure, then extract the metabo-
lites of interest and finally determine their concentrations by standard analytical
methods like gas chromatography, capillary electrophoresis, HPLC, mass spectrome-
try, NMR, or via enzymatic methods. Superior sensitivity is obtained in recent ap-
proaches that employ enzymatic cycling [27-29). By far the least sensitive method is
NMR spectroscopy, which is however applied because of its high dynamic range and
its ability to distinguish many different compounds in a single sample, thus avoiding
the need for partial purification of samples to select only the wanted class of meta-
bolites. For instance, a large number of different sugar phosphates can be identified
from a single 31p NMR spectrum at suitable pH [30,31) and l3C NMR is very specific
in its identification of, e.g., sugars [32) and amino acids [33).
482 15 Metabolic Analysis of Zymomonas mobilis

The application of invasive sampling techniques to measure intracellular metabo-

lite pool sizes requires [34] that:
1. The sample can be taken without disturbing the metabolic state of the cells
2. The inactivation of the metabolism is complete and fast as compared to the intra-
cellular metabolic reaction rates
3. The extraction of the wanted metabolites and the denaturation and separation of
the intracellular enzymes is complete
4. The total procedure of sampling, inactivation and extraction does not affect the
stability of the metabolites
5. The dilution by inactivation and extraction is controlled, reproducible and mini-
mal

Z. mobilis is a very critical organism with respect to the second requirement (i.e.,
that the inactivation be fast as compared to the intracellular metabolic rates) because
it has an extremely high rate of glycolysis. Glucose uptake rates as high as
900l1molmin-l (g dry wtfl [35] have been observed with this organism. Assuming
an average intracellular volume of 2 p,l (mg dry wtr 1 [36], an intracellular sugar
phosphate pool of 0.1 mmoll- 1 would completely turn over once every 16 ms under
these conditions. Likewise, since two reactions of the ED pathway produce ATP, a
2 mmoll- 1 ATP pool would turnover once every 0.16 s. This illustrates the point: the
quenching of metabolism should be faster than these turnover times as a mere pre-
requisite to ensure a reliable preservation of the indicated pools. Therefore, the clas-
sical method for inactivation within a few ms has become the direct quenching of
the sample in liquid nitrogen or liquid CO 2 , where the heat transfer rate is maxi-
mized by the high temperature difference and by a large sample-liquid exchange area
obtained by spraying of the sample [37]. Recently, spraying the sample in 60%
methanol at -40°C has become popular [38, 39] (inactivation within 100 ms) as well
as the injection in precooled perchloric acid [29, 40] which results in inactivation
times of 200-500 ms [34].

15.2.2.2
In vivo Techniques

Since the relative errors associated with all consecutive steps involved in the meta-
bolite pool measurement procedures described in Sect.15.2.2.1 are cumulative,
methods that avoid some or most of these steps seem desirable. Thus, approaches
for in situ, in vivo pool measurements have attracted considerable interest. Two im-
portant representatives of this category are in vivo Nuclear Magnetic Resonance
spectroscopy [41] and in vivo fluorescence spectroscopy [42]. In vivo NMR has a
long and steadily increasing record of application to the characterization of bio-
chemical processes in living cells. A number of early in vivo NMR studies on micro-
organisms were conducted in the Shulman group and have become classical exam-
ples [43-45]. Since then, many more applications to microorganisms and also to cell
culture work, studies on animals and human subjects, have been reported [46]. Thus
31 P NMR allows the intracellular measurement of pH, sugar phosphates, inorganic
phosphate, polyphosphate, nucleoside di- and tri-phosphates, UDP-sugars, PEP, ni-
cotine-adenine dinucleotides, cell wall phospholipids, and phosphorylated sugar
polymers (Fig. 15.2). Currently, with the best systems a metabolite concentration (re-
lated to the total sample volume) of approximately 250 p,moll- 1 is needed to produce
15.2 Methods for Metabolic Analysis 483

I
Pi,ext
PP4
Pi,cyt

\ / P-mannan \
Pi,vac
/ ADP+ATP
NAD(P)(H)
+UDP-glucose
I
ADP+ATP
\ / PP3

\
sugar-P

UDP-glucose
PP2
\
\\ /
ATP \
\

o
I I I I I I
-5 -10 -15 -20 -25 ppm

Fig. 15.2. Typical in vivo 31 P NMR spectrum of Saccharomyces cerevisiae aerobically cultivated
at a density of 75 (g dry wt) ml- 1 in a hydro cyclone bioreactor system specifically designed for
NMR studies [47]. Assignments: sugar-P, sugar phosphates; Pi,cyt, cytoplasmic inorganic
phosphate; Pi,vac, vacuolar inorganic phosphate; Pi,ext, extracellular inorganic phosphate;
PEP, phosphoenolpyruvate; P-mannan, phosphomannan. Peaks labeled PP1-PP4 stem from
polyphosphate resonances [176]. The differences in the positions of the inorganic phosphate
resonances arise from differences in pH between the compartments [30]. Recording time was
2.5 min (Gonzalez B, de Graaf AA (1998) unpublished results)

a useful in vivo 31p signal (S/N=5) in 1 min [47]. Since the measurement time varies
inversely to the squared concentration, a 10-fold dilution of the cell suspension
would require a 100-fold longer measurement time for intracellular pools. Therefore,
in vivo NMR studies usually employ rather high cell densities (typically in excess of
100 mg dry weight per ml). This has triggered the development of in situ NMR/bior-
eactor systems that are capable of adequately maintaining highly concentrated cell
suspensions in a well-defined steady state for a prolonged period of time [47-49].
(For a recent review, see [34]). Signals in the in vivo 31 p NMR spectrum are generally
very broad, leading to strong overlap and a very limited capability to resolve indivi-
dual sugar phosphate signals (Fig. 15.2). lS N NMR spectroscopy in principle has con-
siderable potential for the in vivo monitoring of, e.g., amino acids and amidated
sugars (e.g., glucosamine) [50, 51] but its application is impeded by the very low
sensitivity and natural abundance of the lS N nucleus. The l3C nucleus has a large
chemical shift dispersion (approx. 200 ppm) and the spectra are generally well-re-
solved and easy to interpret (Fig. 15.3). Thus, in vivo l3CNMR has a great potential
for the measurement of many different intracellular metabolites (sugars, amino
acids, sugar phosphates, compatible solutes, etc.). However, since the intrinsic NMR
sensitivity of l3C is four times lower than that of 31 p and the natural abundance of
l3C is only 1.1 %, these advantages can only be exploited in tracer studies using l3C_
enriched precursors (see also Sect. 15.2.3.3).
In vivo fluorimetry for the purpose of monitoring intracellular NAD(P)H levels at
450 nm was introduced in 1957 [52]. The method has been applied for the on-line
monitoring of biomass concentration [53, 54], sugar [55], and phenol concentration
[56] during fermentation processes. It has also been used to characterize intracellu-
484 15 Metabolic Analysis of Zymomonas mobilis

Glu4

T1
I
\
T I
T3 T6

\\T2T4 I Glu3
c=c-o /1 Gln4/

I/p'
Gln2 Asp3 Ala3
~ \ Gln3/

/
I I I I I I I I I
100 90 80 70 60 50 40 30 20 ppm

Fig. 15.3. Typical in vivo 13CNMR spectrum of Saccharomyces cerevisiae aerobically cultivated
at a density of 75 (g dry wt) ml- 1 in a hydro cyclone bioreactor system specifically designed for
NMR studies [47], taken 1 h after the start of infusion of [1- 13 C] labelled glucose. Assignments:
T, trehalose; PM, phosphomannan; Glu, glutamate; Gin, glutamine; Asp, aspartate; Ala, alanine.
The numbers refer to the respective carbon atoms. Recording time was 5 min (Gonzalez B, de
Graaf AA (1998) unpublished results)

lar pH [57,58) in yeast. The tryptophan fluorescence was reported to correlate better
with cell mass than that of NAD(P)H [59). Fluorescence peaks are very broad in
general, making it difficult to identify more than a single compound when employ-
ing a limited measuring range. The use of a wide spectral range of excitation and
emission wavelengths, allowing one to detect a series of important compounds, e.g.,
riboflavin, FAD, FMN, NAD(P)H, pyridoxinic compounds, tryptophan, tyrosine, and
phenylalanine [60) was integrated in a new a technique for two-dimensional fluores-
cence spectroscopy [61). Thus, an increase ofNAD(P)H and a decrease ofFAD/FMN
could be monitored in Saccharomyces cerevisiae during the aerobic-anaerobic transi-
tion and the formation of biomass and ergot alkaloids in a culture of Claviceps pur-
purea could be quantified [61). The time resolution of the method is about 1 min.
While this technique certainly has significant potential, a wider and truly quantita-
tive application of in vivo fluorimetry is still awaited.

15.2.2.3
Rapid Sampling

The methods for metabolite pool determination as described above reveal very little
about the dynamic properties of the metabolism. Worse, any subsequent modeling of
the data has to rely very heavily on enzyme kinetic data as determined from in vitro
measurements. Dynamic methods with a time resolution in the subsecond range,
applied to a metabolic system that is perturbed from the steady state by, e.g., a sud-
den substrate pulse, may provide a wealth of information on the dynamic properties
of the metabolism because the system passes through a whole range of different
states during its reaction on the perturbation. Clearly, invasive techniques with very
15.2 Methods for Metabolic Analysis 485

short inactivation times (Sect. 15.2.2.1) must be applied in order to achieve a subse-
cond time resolution. As early as 1964, incubations of cells were performed with a
freeze-quench apparatus with a four-jet tangential mixer [37] and the changes of
glycolytic intermediates after a pulse of glucose to the carbon-starved culture were
measured one by one by varying the incubation time (0.015-5 s). However, this
method did not allow for a defined metabolic state of the microorganisms, whence
later developments concentrated on connecting rapid sampling devices to a con-
trolled bioreactor where the organisms are in a balanced steady state. The first such
device employed a miniature valve coupled with an HPLC capillary in a hypodermic
needle which was connected to the mixing zone of the fermentor [40]. Samples were
taken manually every 5 s with vacuum sealed, precooled glass tubes containing the
inactivation/extraction solution. The achieved sampling rate of 0.2 per second did
not however allow one to monitor fast dynamic changes. This system was adapted
later to generate several additional samplings at a rate of 5 per second during the
first phase of the dynamic response (M. Reuss, personal communication). In a dif-
ferent approach, continuous sampling of microorganisms from a controlled bioreac-
tor with rapid inactivation of metabolism and extraction of metabolites using pre-
cooled -40°C perchloric acid solution (35%) was achieved with a sampling tube
[29]. This method, which used a total sampling time of 200 s, resulted in the fixing
of fast dynamic reactions at a certain position in the tube. Immediately after stop-
ping the sampling, the tube was frozen in liquid nitrogen. The tube was then divided
into identical parts and the metabolites were analyzed enzymatically. After deconvo-
lution for the axial dispersion in the tube, a time resolution of approximately 70 ms
was achieved. Very recently, a completely automated rapid sampling device for mon-
itoring intracellular metabolite dynamics was described that employed sample flasks
fixed in transport magazines that were moved by a step engine such that subsequent
flasks were filled at a rate of 4 per second for a total time of about 40 s [62]. The
flasks contained 60% methanol of -50°C for quenching of the metabolism.
Analysis of the large number of samples resulting from these type of experiments
(typically 100) has to be performed by laboratory robots. Extensive macrokinetic
modeling must be applied to extract information on fast metabolic regulation from
the data (see Sect. 15.2.4).

15.2.3
Metabolic Flux Analysis

15.2.3.1
Basic Carbon Balancing

Besides pool sizes, a second important parameter to describe metabolism is consti-

tuted by the metabolic fluxes. The first step to quantitate these is the establishment
of elementary balances for, e.g., phosphorus, nitrogen, oxygen, and especially car-
bon. The measurement of the amounts of carbon source consumed, of products ex-
creted, as well as of biomass synthesized, offers a first important quantitative insight
into the metabolic transitions that must occur inside the cell. In the industrial envir-
onment, careful and exhaustive analysis of the composition of the fermentation med-
ium before and after a fermentative production run may give important clues as to
which medium components may be limiting productivity, or which components are
present in excess and can be saved on. Thus, although the cell at this stage is still
486 IS Metabolic Analysis of Zymomonas mobilis

largely considered as a black box, the fluxes in its metabolic network can be confined
within certain boundaries by this basic input-output analysis.

15.2.3.2
Metabolite Balancing

A more detailed flux analysis requires that the cell wall boundaries be crossed and a
quantitative analysis based on biochemical reaction network stoichiometries be per-
formed. This has culminated in the past decade in the appearance of numerous
methods and their computer implementations for so-called metabolic flux balancing.
The physical basis of this method is that it assumes a metabolic steady state, whence
metabolic fluxes into and out of every metabolite pool must exactly match in order
to comply with the conservation of mass. This principle is combined with detailed
quantitative knowledge about the cellular biomass composition and the available
biochemical knowledge of the pathways involved in biosynthesis. A detailed quanti-
tative description of Escherichia coli biomass composition and biosynthetic path-
ways can, e.g., be found in [63,64]. Thus, from the biomass growth rate, the biomass
protein content, and its amino acid composition, the fluxes through all amino acid
synthesis pathways can be calculated. Likewise, from the contents and composition
of DNAIRNA and cell (phospho )lipids, the fluxes through the biosynthetic pathways
of nucleotides and fatty acids, respectively, can be evaluated. All these biosynthetic
building blocks are synthesized from a rather small set of 11 elementary precursor
metabolites generated by the central metabolism, namely glucose-6-phosphate, fruc-
tose-6-phosphate, ribose-5-phosphate, erythrose-4-phosphate, glyceraldehyde-3-
phosphate, 3-phosphoglycerate, phosphoenolpyruvate (PEP), pyruvate, acetyl-coen-
zyme A, oxaloacetate, and a-ketoglutarate [64]. Therefore, these calculations of the
drain-offs from the central metabolism result in the end in a set oflinear constraints
for the distribution of the metabolic fluxes over glycolysis, pentose phosphate path-
way, citric acid cycle, glyoxylate pathway, and the anaplerotic C3-carboxylating reac-
tions in the cell. Together with the measured values of substrate uptake rate and
(by- )product excretion rates, the equation system for the complete metabolic flux
network usually is determined to within 1-4 degrees of freedom. This indeterminacy
can then be resolved either by assuming additional reaction stoichiometries (e.g.,
PIO ratio for oxidative phosphorylation or an exact balance of NADPH regenera-
tion), by assuming inactivity of a redundant pathway based on enzyme measure-
ments (e.g., the glyoxylate cycle during growth on glucose), or by acquiring addi-
tional data of a different class (e.g., isotopic tracer data, see Sect. 15.2.3.3).
The metabolite balancing approach, first formulated in the form outlined above by
Holms [63], was made a focus of the rapidly evolving field of metabolic engineering
by publications of Vallino and co-workers [65-67] that concentrated on the use of
intracellular fluxes for the elucidation of metabolic control. Since then, an ever in-
creasing number of applications of the method has been reported [68-70].

15.2.3.3
Stable Isotope Labeling

As mentioned above, the metabolite balancing method usually still leaves several
degrees of freedom for the flux network equation system. The application of addi-
tional assumptions on the metabolism to render the system (over)determined may
15.2 Methods for Metabolic Analysis 487

not be feasible. Moreover, even if the system is fully determined, the metabolite bal-
ancing approach does not allow one to draw any conclusions about the reversibility
of certain reaction steps in vivo, since such reversibility is simply not reflected in the
net mass balances. Tracer data represent an alternative source of information. For a
considerable period, the radioactive 14C isotope was used for metabolic flux quanti-
tation (see, e.g., [71] for a state-of-the-art review) ). While the 14C isotope offers a
very high sensitivity, experimental procedures using this isotope were extremely te-
dious since metabolites had to be extracted and chemically degraded to obtain the
specific enrichment values for the single carbon atom positions [71-73] that are ne-
cessary for a comprehensive analysis. A much more elegant, though less sensitive,
approach evolved that uses stable isotopes in combination with NMR spectroscopy.
Undoubtedly the most important of these is the l3C isotope, but it should also be
kept in mind that 2H [74, 75]and 15N [76, 77] have considerable potential for flux
determination. While the history of 13C NMR metabolic flux analysis dates back to
the early 1970s [78], with classical examples already from the late 1970s [43-45] until
about a decade ago [79-82], a significant improvement was achieved by the integra-
tion of metabolite balancing and l3C labeling techniques to analyze complete meta-
bolic networks, with pioneering work presented in [83-85]. The latter approach [85]
used a truly comprehensive isotopic labeling data set that was obtained using NMR
analysis of a variety of proteinogenic amino acids isolated from cells exclusively
grown on [1-l3C]glucose as the sole carbon source. Those integrated approaches
finally enabled to achieve an overdetermined system of equations that simulta-
neously describe the metabolic fluxes in vivo and the ensuing l3C labeling, thus sol-
ving the flux analysis problem for the stationary case [86]. Subsequent analyses
showed that l3C labeling data also allow one to quantify the degree of reversibility
of various intracellular bidirectional reaction steps [87-89]. While this adversely im-
plies that the number of unknowns in the flux determination procedures increases
considerably, this may be more than compensated for [90, 91] by the increase of
information content of the l3C labeling experiment that will result from the use of
multiply-labeled l3C substrates in combination with refined measurement techniques
for isotopomer analysis such as two-dimensional NMR [92, 93], heteronuclear spin-
echo NMR [94,95] ,and mass spectrometry [96, 97]. While the stationary flux ana-
lysis problem may thus be considered solved, it only represents a single point analy-
sis which using current approaches is still rather time-consuming. New methodolo-
gical developments therefore will concentrate on more efficient analytical techniques
that will enable one to perform dynamic flux analysis, such that, e.g., the evolution of
the metabolic flux distribution during a fed-batch process can be monitored.

15.2.3.4
NMR Magnetization Transfer

While the techniques described so far are more or less indirect in that the metabolic
fluxes are inferred either from measured changes of metabolite pool sizes or from
the isotopic labeling of metabolites, the NMR magnetization transfer method enables
the direct determination of unidirectional rate constants of metabolic reactions at
steady state. It does so by monitoring the transfer of magnetization from one mole-
cular species to another in real time. The accepting species may be another metabo-
lite or the same compound in a different compartment, where it has a different NMR
resonance frequency. Several examples of in vivo determinations of especially glyco-
488 IS Metabolic Analysis of Zymomonas mobilis

lytic fluxes [98, 99] and membrane transport rates (see [100] for a review) have been
published. The method, however, requires that the turnover rates of the involved
metabolic pools be from the same order of magnitude as the longitudinal relaxation
rates of their nuclear spins. This in practice limits the application of NMR spin
transfer experiments to reactions that involve rather high pool turnover rates of 0.1
to 10 per second. Moreover, the adequate supply of substrate and oxygen to the dense
bacterial suspensions needed for NMR often presents a major problem [101].

15.2.4
Metabolic Modeling

The experimental data resulting from the methods described above cannot be used
just as such to extract the relevant metabolic information. The structure, dynamics
and regulatory properties of metabolic networks are extremely complex. Mathema-
tical models are the only way that net consequences of simultaneous, coupled, and
often counteracting processes can be evaluated consistently and quantitatively [102].
Therefore, the importance of mathematical modeling of metabolic phenomena has
grown strongly in recent years. Within the scope of this review, only a few key refer-
ences to the various methods can be given.
The most frequently used modeling approach is mechanistic modeling, i.e., a
quantitative description of the system's dynamics based on enzyme kinetic proper-
ties. In this type of approach, each individual enzyme-catalyzed reaction or trans-
port step is represented by a kinetic equation for the reaction velocity as a function
of the concentrations of the involved substrates and effectors [103-105]. This results
in a set of balance equations for each intracellular metabolic pool that, depending on
the presentation, manifest themselves as stationary equations describing a stationary
state, or as differential equations describing transient phenomena (e.g., as used in
combination with rapid sampling procedures [106]). Several approaches based on
analytical methods from the field of nonlinear system dynamics have been devel-
oped for the theoretical analysis of such equation systems. They all require a com-
prehensive mechanistic model of the metabolic pathways considered as well as the
accurate and precise values of the reaction kinetic parameters and result in general
model structures of the type presented in [107, 108]. This general structure allows
one to calculate stationary states as well as to carry out stability analyses and sensitiv-
ityanalyses [104, 107, 109, 110]. The well-known Metabolic Control Theory [111-
115] can be considered as a special development of this technique. This theory is
based on a linearization of the system equations on a logarithmic scale. The ensuing
state sensitivity matrices are built up by control coefficients that allow a quantitative
characterization of the regulatory influences of the kinetic model parameters. Since,
however, detailed knowledge of the kinetic parameters is often at best fragmentary,
alternative approaches were developed. One of these is the important universal 5-
Systems modeling framework [116-119] where each intracellular material balance
is approximated by an exponential expression involving all metabolite concentra-
tions. Positive effectors of the reaction velocity enter with a positive exponent, nega-
tive effectors with a negative exponent. Non-participating metabolites enter with an
exponent equal to zero. This is equivalent to a linearization of the mechanistic model
on a logarithmic concentration scale and as such it is very similar to Metabolic Con-
trol Theory. The approach, which may be used independent of a mechanistic model,
is able to generate at least an approximate description of the system's behavior. It has
15.3 Metabolic Analysis of Zymomonas mobilis 489

served as a basis for the recent formulation of a production-oriented metabolic op-

timization framework [119, 120). Of a still more semi-quantitative nature are the
order-of-magnitude-calculation [121, 122) modeling and the approach to incorpo-
rate thermodynamic data as constraints for the intracellular reaction steps [123,
124). The latter approach can be integrated with Metabolic Control Theory [125).
A special case is formed by the stationary balancing approaches involved in meta-
bolic flux analysis (see Sect. 15.2.3.2). They employ rather straightforward matrix
formalisms to handle the linear equation systems [65,68,69) that describe the ma-
terial balances for all intracellular metabolite pools considered. The information ob-
tained from isotopic labeling experiments (see Sect. 15.2.3.3) may either be added as
a set of additional constraints to such equation systems (as was done, e.g., in [83, 84,
126), or it may be incorporated in a special isotopic labeling model. Classical exam-
ples of the latter can be found in [79, 81, 97). These and subsequent models consid-
ered only flux distributions over a limited number of pathways at a few specific
branchpoints in metabolism. Only recently, a truly comprehensive modeling frame-
work for stationary flux analysis based on isotope labeling was established [85, 87,
88, 127) which in fact represents nothing else than the formulation of the material
balances for all individual carbon atoms of the metabolic network in a single matrix.
The metabolite pool balances emerge as a special case from these. Basic elements of
this integrated isotopic labeling model, i.e., the so-called atom mapping matrices,
were described in [128). The modeling of l3C isotopomer data, a current focus of
metabolic flux analysis, also has some history. Classical examples can be found in
[80, 82, 129, 130). These models also suffer from a limited applicability. Recent de-
velopments therefore exploit matrix formulation to achieve a more general isotopo-
mer modeling framework [131-133). Very recently, Wiechert et al. [90) and Mollney
et al. [91) succeeded in establishing a completely general modeling framework for
isotopomer experiments based on a very fast, non-iterative solution strategy for the
isotopomer balance equations. The previously developed comprehensive model for
stationary flux analysis based On isotope labeling [85, 87, 88) emerges as a special
case within this new framework.
The coming decade will yield an enormous increase of information not only from
the familiar data sources like dynamic metabolite concentration measurements and
metabolic flux analysis but also from very different sources, i.e., genome sequencing,
gene expression studies using DNA microarrays and protein two-dimensional gel
electrophoresis. Therefore, methods of bioinformatics that combine these different
types of often imprecise and fragmentary data are expected to play an increasingly
important role in metabolic engineering. Several approaches for holistic modeling
now under development are already worth pointing out [134-136).

15.3
Metabolic Analysis of Zymomonas mobilis

15.3.1
Introduction

A specialty of Z. mobilis metabolism is that only 1 mol of ATP is formed per mol
glucose consumed. Accordingly, only 2% of the substrate carbon typically ends up
in the biomass. Nevertheless, this organism can grow at growth rates as high as
490 15 Metabolic Analysis of Zymomonas mobilis

0.2Sh- l . It succeeds in doing so by maintaining an extremely high glycolytic rate

(specific glucose uptake rates of 0.9-1.0 U (mg dry wtr l have been reported [137]).
A number of metabolic studies concentrated on the question of how this flux is
regulated and where eventually the rate-limiting step in the glycolytic pathway is
located. These included the substrate uptake (which proceeds via a glucose facilita-
tor) and product excretion (i.e., rapid membrane diffusion) steps.
Z. mobilis exhibits decreased growth and ethanol yields as well as increased by-
product formation during growth on fructose. Several studies therefore concentrated
on the differences between glucose and fructose metabolism. Adaptation of Z. mobi-
lis to highly osmotic media has been a topic of interest because the natural environ-
ment of the organism features high glucose and fructose concentrations. As the prin-
cipal product ethanol is known to become a limiting factor for Z. mobilis growth at
high concentrations (above 80gl- I), several studies concentrated on effects of this
compound on the glycolytic enzymes. Since the Entner-Doudoroff glycolytic path-
way of Z. mobilis is a linear pathway and the wild-type organism does not possess a
complete pentosephosphate pathway nor a closed citric acid cycle, the central meta-
bolism of this organism contains no important branching points. This certainly is a
reason why very few isotopic labeling studies for flux analysis have been carried out
with Z. mobilis. Only the recent construction of pentose-fermenting recombinant
strains has motivated research in this area. In the following review, establishment
of a more or less coherent view on Z. mobilis metabolism based on the information
provided by studies on the various topics indicated above will be attempted. The
discussion will be focused on ethanol production on media containing a single
monomeric sugar as carbon source. Thus, the biotechnologically relevant conversion
of glucose plus fructose to gluconolactone and sorbitol by the action of a glucose-
fructose-oxidoreductase in Z. mobilis [138] as well as the bioconversion of sucrose
([139] and references therein) are not covered in this contribution.

15.3.2
Enzymatic Studies

Specific activities of the enzymes involved in Z. mobilis glycolytic and other path-
ways have been reported in several studies [140-142], giving rather complete compi-
lations of data. However, data in [142] were obtained at 23 °C and therefore cannot
be used in a predictive manner for metabolic rates at 30 0c. Typical values are com-
piled in Table 15.1.
While a considerably reduced specific growth rate and cell yield were observed
during growth on fructose as compared to glucose, approximately equal specific con-
sumption rates for the two substrates of 10-11 g g-I h -I (equivalent to 0.9-1.0 U (mg
dry wtr l were found [137]. Scopes et al. [145] reported that glucose-6-phosphate
dehydrogenase was inhibited by its product, glucose-6-phosphate and glucose
6-phosphate dehydrogenase by ATP, but only with rather high inhibition constants
KI of 15 mmol r l for glucose 6-phosphate and 1.4 mmoll- I for ATP, respectively. No
effectors for the other five glycolytic enzymes glyceraldehyde 3-phosphate dehydro-
genase, phosphoglycerate kinase, phosphoglycerate mutase, enolase, and pyruvate
kinase, were identified [146].
From these studies and the data as shown in Table 15.1, it was originally inferred
that the phosphorylation of glucose by glucokinase (or of fructose by fructokinase,
respectively) and the oxidation of glucose-6-phosphate by glucose 6-phosphate dehy-
15.3 Metabolic Analysis of Zymomonas mobilis 491

Table 15.1. Average reported activities at 30°C of the enzymes involved in Z. mobilis sugar
catabolism determined from cell-free extracts. In converting values per g protein to values
per g dry mass, it was assumed that protein constitutes 60% of Z. mobilis dry mass [1]

Enzyme Typical activity Source

(U (mg dry wtr 1 )

Glucose facilitator 1.8" [143]

Glucokinase 1.0 [141]
Fructokinase 0.6b [141]
Phosphoglucose isomerase 0.8 b [141]
Glucose 6-phosphate dehydrogenase 1.4 [141]
6-phosphogluconolactonase 2.5 [141]
6-phosphogluconate dehydratase 1.8 [141]
KDPG aldolase 3.0 [141]
Glyceraldehyde 3-phosphate dehydrogenase 3.0 [141]
Phosphoglycerate kinase 6.0 [141]
Phosphoglycerate mutase 12.0 [141]
Enolase 1.5 [141]
Pyruvate kinase 4.5 [141]
Pyruvate decarboxylase 2.8 [141]
Alcohol dehydrogenase-l 3.0 [141]
Alcohol dehydrogenase-2 10.0 [141]
a Value for glucose. Vm= for fructose and xylose were reported to be 23% and 121 % higher, respectively
[24]
b Reported to be three times higher during growth on fructose [144]

drogenase are the major flux-controlling steps in the absence of high ethanol con-
centrations. This conclusion was recently confirmed with substantial evidence:
Snoep et al. [147] reported that glucose 6-phosphate dehydrogenase exerts a very
large flux control over the glycolytic flux in Z. mobilis, and two studies [148, 149]
demonstrated that the enzyme is indeed allosterically controlled by phosphoenol-
pyruvate (PEP), presumably in order to maintain a balance of glycolytic throughput
and ATP consumption in Z. mobilis.
A discussion of metabolic activity based on metabolite pool measurements (see
next section) must take into account the kinetic constants of the pathway enzymes.
A number of glycolytic enzymes from Z. mobilis have been kinetically characterized.
A compilation of kinetic constants reported in the literature is given in Table 15.2.
From this table, it can be seen that with the exception of pyruvate decarboxylase, all
glycolytic enzymes will reach substrate-saturating conditions at submillimolar con-
centrations of the substrates.

15.3.3
Metabolite Pool Measurements

15.3.3.1
Overview

A number of studies have been devoted to the study of intracellular glycolytic meta-
bolite pools in Z. mobilis. Lazdunski and Belaich [156] measured the static ATP con-
492 15 Metabolic Analysis of Zymomonas mobilis

Table 15.2. Literature data on kinetic constants for various glycolytic enzymes of Z. mobilis'.
Abbreviations: G6P, glucose 6-phosphate; PEP, phosphoenolpyruvate; 6PG, 6-phosphogluco-
nate; GP, glycerophosphate; Pi, inorganic phosphate; 3-PGA, 3-phosphoglycerate; 2-PGA, 2-
phosphoglycerate

Enzyme Km-value(s)(mmoll- 1 ) Source

Glucose facilitator 3.2 (glucose) [143]

39 (fructose) [150]
40 (xylose) [24]
Glucokinase 0.1 (glucose) [151]
0.2 (ATP)
0.05 (phosphate)
15 (G6P)b [145]
Fructokinase 0.7 (fructose) [145]
0.45 (ATP)
Glucose 6-phosphate dehydrogenase 0.2 (G6P) [149]
0.022 (NADP)
0.5-1.1 (NAD)d
0.05 (PEP)C
1.4 (ATP)b [145]
6-phosphogluconate dehydratase 0.04 (6PG) [152]
0.3 (D-a-Gp)b
2.5 (Pi)b
2.0 (3-PGA)b
KDPG aldolase 0.25 (KDPG) [153]
Phosphoglycerate kinase 1.5 (3-PGA) [146]
1.1 (ATP)
Phosphoglycerate mutase 1.1 (3-PGA) [146]
Enolase 0.08 (2-PGA) [146]
Pyruvate kinase 0.08 (PEP) [146]
0.17 (ADP)
Pyruvate decarboxylase 4.4 [154]
Alcohol dehydrogenase-1 0.086 (acetaldehyde) [155]
0.027 (NADH)
4.8 (ethanol)
0.073 (NAD)
Alcohol dehydrogenase-2 1.3 (acetaldehyde) [155]
0.012 (NADH)
27.0 (ethanol)
0.11 (NAD)
a No kinetic data for phosphoglucoseisomerase, 6-phosphogluconolactonase and glyceraldehyde 3-phos-
phate dehydrogenase were found
b Inhibitory constant, i.e., Ki
C PEP is an allosteric inhibitor; Hill coefficient varies from 1.35 at [PEP]=O /Lmol tl to 2.0 at

[PEP]> 100/Lmol tl
d Depends both on [PEP] and on [G6P]
15.3 Metabolic Analysis of Zymomonas mobilis 493

tent to describe the energetic cellular state at different uncoupling conditions of

growth. Barrow et al. [157] studied the levels of nucleoside triphosphates, sugar
phosphates, UDP-sugars, and Pi with a time resolution of 1 min in intact fermenting
cells as well as in perchlorate extracts using 31p NMR spectroscopy. Algar and Scopes
[140] studied metabolite concentrations in cell-free extracts of Z. mobilis to which
glucose was added continuously, in order to investigate the controls on the glycolytic
enzymes. Osman et al. [142] determined intracellular levels of glycolytic intermedi-
ates and nucleotides in several phases during a batch glucose fermentation using
31 p NMR. Strohhacker et al. [158] investigated ethanol inhibition of glucose catabo-
lism in Z. mobilis using 31 p NMR spectroscopy in vivo and of perchloric acid extracts
from cell suspensions incubated with various concentrations of ethanol. Weuster-
Botz [29] used a sampling tube device with rapid inactivation of metabolism for
continuous sampling of Z. mobilis cells from a controlled bioreactor after application
of a glucose pulse, enabling a dynamic investigation of Z. mobils glycolytic inter-
mediates with a time resolution of only 0.64s. De Graaf et al. [159] compared the
intracellular levels of nucleoside di- and triphosphates, sugar phosphates, UDP-su-
gars, and Pi in glucose- and fructose continuous cultures of Z. mobilis as measured
by 31 P NMR using a special membrane-cyclone NMR bioreactor system.
Intracellular sugar concentrations have also been studied. Struch et al. [160] deter-
mined intracellular glucose and xylose concentrations in an investigation of the phy-
siological basis of the exceptionally high sugar tolerance of Z. mobilis. In a study of
pentose metabolism in wild-type and recombinant Z. mobilis strains, Feldmann et al.
[25] determined intracellular concentrations of various pentoses and their phos-
phates. Schoberth and de Graaf [161] used in vivo l3e Nuclear Magnetic Resonance
spectroscopy to follow xylose uptake in Zymomonas mobilis. In a study using in vivo
NMR spin transfer of ethanol transmembrane diffusion in Z. mobilis, Schoberth et
al. [36] monitored intra- and extracellular glucose and ethanol concentrations simul-
taneously in intact fermenting cells in a single experiment.
In the next sections an attempt will be made to put together the fragmentary
information presented in the studies indicated above in order to get a more complete
view on Z. mobilis' metabolic properties.

15.3.3.2
Glycolytic Intermediates

The concentration time profiles of Z. mobilis glycolytic intermediates after a glucose

substrate pulse as described in [140, 157, 158] reveal several common features that
are sketched in Fig. 15.4. The NTP content rises rapidly to a plateau value of 3-
5 mmoll- I reached already after 1-2 min. Glucose 6-phosphate and glyceraldehyde
3-phosphate both rise to a maximum value within 30 s. Hereafter, they decline again,
such that glucose 6-phosphate reaches a steady-state value at t=1 min while glycer-
aldehyde 3-phosphate continues to decrease to a level of 0.3-0.2 mmoll-I, reached at
t=4 min. 6-Phosphogluconate also rises within 1 min to a plateau value of about
0.5 mmoll- I • 3-Phosphoglycerate rises to a steady-state value of about 2 mmol r l in
2-3 min, i.e., slower than glucose 6-phosphate, 6-phosphogluconate, and glyceralde-
hyde 3-phosphate. In accordance with the NMR studies, Weuster-Botz [29] also re-
ported a rapid increase and subsequent decline of glucose 6-phosphate, glyceralde-
hyde 3-phosphate, and 3-phosphoglycerate after application of a glucose pulse. How-
ever, due to the fact that this study applied extremely rapid sampling in a well-con-
494 15 Metabolic Analysis of Zymomonas mobilis

5.0

4.0

::::§:
.s
___ G6P
3.0
-o-6PG
Q)
~GAP
N
'iii 2.0 --tr-3-PGA
(5 --+- NTP
0
fl.
1.0

0.0
0.0 2.0 4.0 6.0 8.0
Time (min)
Fig. 15.4. Time courses for selected glycolytic intermediates and NTP (ATP plus UTP) in Z.
mobilis after application of a glucose pulse of several hundred mmoll- 1 to very dense cell
suspensions (100-200 mg dry wt. ml- 1 ). The traces represent smoothed averages of concentra-
tions reported in [157, 158] determined by 31pNMR. More or less similar patterns, although
on a roughly tenfold longer time scale due to crude extract dilution, are reported in [140]

trolled bioreactor with a mixing time of about 500 ms it was possible to demonstrate
that glucose 6-phosphate and 3-phosphoglycerate reach their peak values essentially
within 1 s, where in the NMR studies this occurred only after 30 s. This probably
reflects a much longer mixing time of the glucose pulse in the dense cell suspensions
used in these studies. In the rapid sampling experiment, glyceraldehyde 3-phosphate
reacted slower and reached a peak value after 25 s [29]. Unfortunately, no data on
other glycolytic intermediates and ATP are given. In metabolic studies on cell-free
extracts of Z. mobilis, Algar and Scopes [140] demonstrated that the levels of the
glycolytic intermediates depended strongly on the ATPase level added to the extracts,
with limiting ATPase activity causing accumulation of metabolites from the Entner-
Doudoroff pathway, especially glucose 6-phosphate. ATP levels were very low both at
limiting ATPase levels and at excess ATPase levels, and high at intermediate levels.
This partly paralleled earlier results [156] showing that ATP levels in rapidly grow-
ing Z. mobilis (corresponding to high ATPase activity) were significantly lower than
in cells that were growth-limited by panthotenate.
A first macrokinetic model of Z. mobilis catabolism, based on the enzyme charac-
teristics given in Tables 15.1 and 15.2 (except for the allosteric inhibition of glucose
6-phosphate dehydrogenase by PEP), was reported by Wulf et al. [103] and used for
dynamic simulations of glucose and fructose metabolism. This model qualitatively
reproduced the accumulation of glucose 6-phosphate found initially after application
of a substrate pulse.
The increased by-product formation by Z. mobilis during growth on fructose [13]
has long remained a puzzling fact since both sugars share essentially the same gly-
colytic pathway (Fig. 15.1). Only very recently was a clue to a possible explanation
reported [159]. In vivo 31 PNMR experiments on Z. mobilis growing in a controlled
bioreactor showed that inorganic phosphate, NDP, NTP, and UDP-sugar levels were
two-fold lower during fructose metabolism, and that the total sugar phosphate pool
was almost five times higher than during glucose metabolism. Subsequent 31 p NMR
15.3 Metabolic Analysis of Zymomonas mobilis 495

measurements on cell-free chloroform extracts demonstrated elevated levels of fruc-

tose 6-phosphate, ribose 5-phosphate, sedoheptulose 7-phosphate, 3-phosphoglyce-
rate, and dihydroxyacetonephosphate. It was hypothesized [159] that this global al-
teration of the levels of intracellular phosphorylated metabolites is primarily caused
by an elevated concentration of intracellular fructose-6-phosphate during growth on
fructose. This hypothesis is supported by the finding that overexpression of fructo-
kinase results in a severe growth retardation of Z. mobilis on fructose media [24].

15.3.3.3
Sugars

From the enzyme activities as given in Table 15.1, it is to be expected that the uptake
of glucose, fructose, and xylose is faster than the maximum glycolytic rate. Thus, it
is to be anticipated that the substrate accumulates intracellularly, even during high
fermentative activity. Several authors indeed reported this observation. Belaich et al.
[143] already concluded from micro calorimetric measurements that an excess of glu-
cose entry, compensated by an outflow, most probably occurs in Z. mobilis exponen-
tially growing cells. struch et al. [160] were the first to demonstrate elevated intra-
cellular levels of glucose in metabolizing Z. mobilis. These authors found that the
intracellular glucose concentration in cultures with a specific glucose uptake rate of
4.3 g g-l h-l (i.e., about half-maximal) was always about 200 mmol tl lower than the
extracellular concentration. This gradient constituted the driving force for the net
glucose uptake. The equilibration of intra- and extracellular glucose was shown to
provide an almost complete osmotic balance between internal and external space
[160], which explains the exceptionally high sugar tolerance of Z. mobilis (up to
40% glucose [1]. Using an in vivo NMR method developed by Schoberth and de
Graaf [161], Schoberth et al. [36] were able to monitor intracellular levels of both
the a and the ~ anomers of glucose in 3-ml cultures of intact Z. mobilis converting
780 Ji,illol of glucose to ethanol within 20 min. The total intracellular glucose concen-
tration was as high as 160mmolt 1 when the extracellular level was 400mmolt 1•

15.3.3.4
Ethanol

The studies of Schoberth et al. [36] not only measured intracellular sugar concentra-
tions but also generated unique data on intracellular ethanol concentrations and uni-
directional rate constants of ethanol transmembrane diffusion. From the data given
[36] it can be calculated that ethanol efflux would match ethanol production in cells
fermenting glucose at the highest reported specific uptake rates (0.9-1.0 U (mg dry
Wt)-l [137]) already when the intracellular ethanol concentration is only 2 mmoll- 1
higher than the extracellular concentration. Thus, this study ruled out any possibi-
lity that a limitation in ethanol efflux leading to strong intracellular accumulation
would block glycolysis in Z. mobilis.
Nevertheless, ethanol at high concentrations obviously affects metabolism ad-
versely. Accumulation of ethanol during fermentation was reported to cause a de-
crease of the growth rate and the specific ethanol production rate [9]. In the presence
of 100 g tl ethanol, glycolysis is severely slowed down and 3-phosphoglycerate accu-
mulates [158]. Data from [140] obtained with cell-free extracts suggest that glycer-
aldehyde 3-phosphate, KDPG, acetic aldehyde, and pyruvate levels also accumulate
496 15 Metabolic Analysis of Zymomonas mobilis

and that the NAD concentration increases with increasing ethanol concentration.
Under the conditions tested (i.e., cells grown on glucose concentrations of 10-12%,
showing enzyme activities as given in Table 15.1), inhibition of enolase by ethanol
might be the primary responsible cause of these effects [158]. However, in fermenta-
tions involving higher ethanol concentrations (i.e., 70-110 g 1-1) the effects of ethanol
on Z. mobilis metabolism very likely are much more diverse. Relevant studies indi-
cate that ethanol primarily exerts inhibitory effects on cellular growth [162, 163].
Responsible mechanisms include an increase of maintenance energy requirement
[164], irreversible deactivation of anabolic enzymes [26], and a loss of cofactors
and metabolites [165]. Hermans [166] showed that the energy charge and the ATP
concentration remained high in Z. mobilis continuously cultivated at 100 g 1-1 etha-
nol, but that the ability to maintain pH homeostasis decreased due to inhibition of
the membrane-bound H+-ATPase. Moreover, it was observed that cell division as
well as synthesis of DNA and fatty acids in Z. mobilis were completely inhibited at
ethanol concentrations above 80 g 1-1 while biomass still increased due to protein
synthesis. At ethanol concentrations above 110 g tt, severe leakage of metabolites
into the medium indicates that the cells are in fact extracted [166].

15.3.4
Flux Analyses

15.3.4.1
Overview

As pointed out in Sect. 15.2.3.1, the first step in the quantitation of metabolic reac-
tion rates is the establishment of elementary balances, especially for carbon. Several
of the physiological studies of Z. mobilis referred to in the preceding sections include
a basic level of this type of flux analysis, in the form of fermentation balances. Since
these typically indicate a conversion of 96-98% of the substrate into equimolar
amounts of ethanol and carbon dioxide [1, 18], there has been little interest in a
characterization involving detailed metabolic balancing procedures (see
Sect. 15.2.3.2) of the remaining 2-4% of the carbon. Likewise, the largely unbranched
structure of the Entner Doudoroff pathway did not make Z. mobilis a promising
object for investigation by stable isotope labeling methods and NMR, since the po-
tential of these methods is only fully exploited with metabolic networks involving
multiple branching points and reversible reaction steps (Sect. 15.2.3.3). Hence, only
a rather limited number of flux analysis studies dealing with Z. mobilis have been
reported.

15.3.4.2
Metabolite Balancing

Variable data on the macromolecular biomasss composition [1, 167] and the protein
composition [167, 168] of Z. mobilis have been reported. After combination of these
with those on another Gram-negative organism, E. coli [64], the following overall
composition of Z. mobilis dry mass can be proposed: protein 60.5%; RNA 19.5%;
DNA 2.7%; lipid 8.5%; peptidoglycan 2.5%; glycogen 2.5%; polyamines 0.3%; meta-
bolites and ions, 3.5% [159]. From these data, following the approach of Neidhardt et
al. [64], a monomeric composition showing the detailed precursor requirement for
Table 15.3. Precursor requirements for biomass synthesis of Z. mobilis calculated from data in [1,64, 167, 168]. Adapted from [159]. Abbreviations: G6P,
glucose 6-P; F6P, fructose 6-P; RI5P, ribose 5-P; E4P, erythrose 4-P; GAP, glyceraldehyde-3-P; PGA, phosphoglycerate; PEP, phosphoenolpyruvate; PYR,
......
pyruvate; AcCoA, acetyl-CoA; OAA, oxaloacetate; AKG, 2-oxoglutarate VI
;....,
Precursor stoichiometry (mol mol-I) s::
<b
Amino acid Amount G6P F6P RI5P E4P GAP PGA PEP PYR AcCoA OAA AKG CO 2 III
'"
0"
(ILmol g-I) g.
;:;.
Ala 1088 >
::s
Arg 181 ~
Asx 478 ~.
en
Cys 20 0
,..,.,
Glx 343 ~
Gly 920 ~
c
His 82 ~
c
Ile 369 1 -1 ;::!
I::>
Leu 369 2 -2 '"~
Lys 249 -1 c
Met 81 g;
;;:.
Phe 11 2 -1
Pro 210
Ser 202
Thr 224
Trp 54 -1
Tyr 70 2 -1
Val 569 2 -1
polymer G6P F6P RI5P E4P GAP PGA PEP PYR AcCoA OAA AKG CO2
Protein 0 0 136 135 0 1142 216 3582 369 1401 734
RNA 600 350 250 600
DNA 87 44 44 87
Lipids 120 120 1976 -1976
Peptidoglycan 55 28 83 55 28 28 -55
Glycogen 154
CI-Units 49
Polyamines 59 '-l
'"""
total 154 55 823 135 120 1705 244 3665 2400 1723 821
498 15 Metabolic Analysis of Zymomonas mobilis

biomass synthesis of Z. mobilis as given in Table 15.3 can be calculated [159]. These
data provide a basis for detailed metabolite balancing studies of Z. mobilis.

15.3.4.3
NMR and Stable Isotope Labeling

l3C NMR has been used in Z. mobilis research for the identification and quantitation
of the major lipid components [169], for the structure elucidation of a novel poly-
fructoside (-a-fructofuranosyl-(2-1)-~-fructofuranosyl-(2-6)-) [170], and for the
identification of sorbitol in sucrose and fructose plus glucose fermentations [171].
However, these studies used the natural abundance l3C present in these metabolites.
In fact, true stable isotope labeling has only been applied in very few cases to eluci-
date specific metabolic characteristics of Z. mobilis. Barrow et al. [172] studied the
deuterium-labeling patterns and stereochemistry of the ethanols produced from the
metabolism of [1-2H]glucose and unlabelled glucose in 2H 20 by a combination of
2H, l3C, and lHNMR. The labeling patterns were explained in terms of enzyme me-
chanisms and stereospecificity of the dehydrogenases involved in the conversion of
sugars to ethanol in Z. mobilis, and metabolite enolization. Rohmer et al. [173] stu-
died the biosynthesis of the side-chain of bacteriohopanetetrol and of a carbocyclic
pseudopentose from l3C-labeled glucose as well as the early steps in isoprenoid bio-
synthesis [174] in Zymomonas mobilis using l3C NMR.
As pointed out already in Sect. 15.1.1, the intracellular differences between glucose
and fructose metabolism leading to increased by-product formation in Z. mobilis
during growth on fructose have largely remained unexplained. While a hypothetical
biosynthetic pathway for the main by-products glycerol and dihydroxyacetone was
already proposed in 1986 [13], the formation of these compounds has only recently
been elucidated [19] using a combination of enzymatic methods and 31 p _ as well as
l3C NMR spectroscopy. The latter was used to study the incorporation of label from
[2-13C]fructose into glycerol, dihydroxyacetone, lactate, and acetoin. These studies
identified a novel pathway for the formation of glycerol 3-phosphate which branches
off the Entner-Doudoroff pathway at the intermediate glyceraldehyde 3-phosphate
and proceeds via dihydroxyacetone phosphate, dihydroxyacetone, glycerol to glycer-
ol 3-phosphate [19].
To date, only a single comprehensive metabolic flux analysis study combining me-
tabolite balancing and 13CNMR of chemostat-grown Z. mobilis has been carried out
[159]. This study involved stable isotope-aided flux analysis of Z. mobilis wild type
growing on glucose and fructose, and of a xylose-degrading recombinant strain that
was constructed in a similar approach to [22] but using the genes for xylose isomer-
ase and xylulokinase from Klebsiella pneumoniae in addition to the genes for trans-
ketolase and transaldolase from E. coli [175]. The recombinant strain was able to
grow on xylose and to produce ethanol at 86% of the theoretical yield but specific
growth and production rates were 4-5 times lower than those typically observed on
glucose. The flux analysis of [2-1 3C]glucose- and [2-l3C]fructose-grown Z. mobilis
shows clearly that 95% of the sugar is rapidly metabolized over the ED pathway with
very little drain-offs to by-products and cell mass synthesis. While this is evident
from the carbon balances alone, the analyses also give detailed insight in the meta-
bolic activities of the non-oxidative pentose phosphate pathway enzymes in vivo.
Only negligible net fluxes are carried by the transketolase, 6-phosphogluconate, ri-
bulose-5-phosphate epimerase, and phosphoribose isomerase reactions, but the flux
15.3 Metabolic Analysis of Zymomonas mobilis 499

Xylose (" '" .. Xyl~ol

,~
@~
Xylulose

@
Glucose 6-P ~ Fructose 6-P CO,
~
~ulose 5-P

~ r--·RuI5P

@! ~
6-P-Gluconolactone
Rlb5P

@)!
6-P-Gluconate
GAP Sed7P

KDPG
Ery4P

~ GAP+---'

~
1,3 di-p-glycerate

~@
3-P-glycerate

@~
2-P-glycerate

~@
Pyruvat~e
.~--@])""9-9.""8PEP

161.8 @
CO. Acetaldehyde .....;;-~.~ Ethanol

Fig. 15.S. Intracellular flux distribution during [1-1 3 C]xylose-fed chemostat fermentation
(D=O.OI8 h- 1) of a recombinant strain of Zymomonas mobilis in which heterologous activities
of xylose isomerase, xylulokinase, transketolase and trans aldolase were expressed. All fluxes
are given as JLmol per gram dry weight and minute. Data taken from [159]. Fluxes were deter-
mined by non-linear least squares fitting of a flux model to joint NMR and fermentation data.
Fluxes towards minor by-products as well as precursor fluxes for biomass synthesis, included
in the original analysis [159], are not shown

analysis predicts a high degree of reversibility of the transketolase, ribulose-5-phos-

phate epimerase, and phospho ribose isomerase reactions as well as of the phospho-
glucose isomerase reaction. Because the principal l3C label in ribose 5-phosphate
was found in C-2, it was concluded that this precursor is synthesized primarily via
the non-oxidative pentose phosphate pathway in Z. mobilis [159], in accordance with
an earlier study [173]. The flux analysis of the [1- 13 C]xylose-grown recombinant
strain gives a detailed view of the heterologous catabolic activities of the pentose
phosphate pathway enzymes (Fig. 15.5). Moreover, this study identified heterologous
xylulokinase as the rate-limiting enzyme of the strain and hypothesizes that stronger
overexpression of this enzyme would probably completely prevent accumulation of
the by-products xylitol and xylulose [159]. This would improve the ethanol yield to
96%, equal to that on glucose.
500 15 Metabolic Analysis of Zymomonas mobilis

15.3.5
Summary

Clearly, no consistent single set of pool measurement data covering all glycolytic
intermediates as well as pyruvate, acetaldehyde, ethanol, and the nucleoside mono-,
di- and triphosphates measured under defined conditions in Z. mobilis is available.
However, the material discussed in Sects. 15.3.2. and 15.3.3 allows one to establish a
coherent picture of Z. mobilis glycolytic metabolism.
The capacity of the uptake of the monomeric sugars glucose and fructose is (far)
greater than the capacity of the glycolytic system for sugar concentrations above the
Km of the facilitated diffusion uptake system (3.2 mmoll- 1 and 39 mmoll- 1 for glu-
cose and fructose, respectively (Table 15.2.)). This strong overcapacity enables Z.
mobilis to maintain effortlessly an osmotic balance in its high-sugar natural environ-
ments.
The maximum for the glycolytic rate is ultimately set by the glucokinase Vmax
(Table 15.1). However, the consistent presence of glucose 6-phosphate in concentra-
tions well above Km for glucose 6-phosphate dehydrogenase (0.2 mmoll-I, Ta-
ble 15.2) [29, 157, 158] together with the finding that glucose 6-phosphate dehydro-
genase is allosterically inhibited by PEP [148, 149] identifies this enzyme as the
principal rate-controlling step of glycolysis in non-optimal conditions, as was also
concluded from metabolic control analyses [147]. Apparently, the overall regulatory
control structure of Z. mobilis catabolism established by the reported characteristics
of its glucose 6-phosphate dehydrogenase is that the pace of glycolysis is set by the
total ATPase activity resulting from, e.g., cellular growth processes and maintenance
of transmembrane gradients of pH and ion concentrations. The ATP breakdown by
this ATPase activity must stoichiometrically match the overall ATP production of
1 mol ATP per mol glucose consumed in glycolysis [140, 156]. When ATPase activity
is less than the glucokinase potential, e.g., due to limiting concentrations of certain
growth factors in the medium, the level of regenerated ADP will fall, thereby limiting
pyruvate kinase activity which will lead to an accumulation of PEP. The increased
PEP level, together with the increased ATP concentration, downregulates glucose
6-phosphate dehydrogenase activity [141, 145] (Table 15.2). The resulting accumula-
tion of glucose 6-phosphate will then slow down glucokinase activity (Table 15.2) to
the appropriate level, even in the presence of saturating ATP concentrations for this
enzyme.
At elevated ethanol concentrations, no simple view of the regulation of metabolism
of Z. mobilis can be established since a variety of effects occur simultaneously (see
Sect. 15.3.3.4). However, the inhibition of the membrane-bound H+-ATPase as well as
of DNA- and fatty acid synthesis, together with the observation that ATP concentra-
tions seem little affected by ethanol [166], suggest that an inhibition of the effective
total ATPase activity of the cell may be an important causal factor in the progressive
limitation of the glycolytic rate in Z. mobilis with increasing ethanol concentrations.
The flux analyses performed thus far with Z. mobilis show that the non-oxidative
pentose phosphate pathway plays an important role in the biosynthesis of ribose
5-phosphate. Furthermore they indicate that properly engineered recombinant
strains of Z. mobilis will always be able to direct about 95% of the carbon to the
product ethanol, irrespective of the type of monomeric sugar chosen as substrate.
References 501

15.4
Concluding Remarks

The many physiological studies conducted with Zymomonas mobilis in the past
15 years have yielded a basic understanding of the short-term regulation of the cata-
bolic fluxes in this organism. Interestingly, several very important metabolic charac-
teristics (i.e., the high flux control by and allosteric regulation of glucose 6-phos-
phate dehydrogenase as well as the overall alteration of phosphorylated pools during
growth on fructose) of Z. mobilis have only very recently been discovered.
Z. mobilis typically converts sugars at extremely high rates with over 95% yield
into ethanol. While worldwide attempts to transform this organism into an efficient
ethanol producer from abundant and renewable carbon sources have long failed, the
recent successful engineering of xylose-, arabinose-, and maltose-fermenting strains
now brings this perspective into closer view. First flux analyses reveal the remark-
able fact that Zymomonas mobilis demonstrates the same high yield and capacity of
ethanol formation on these sugars as it does on glucose and fructose. These charac-
teristics guarantee an ongoing interest in and use of this organism and its genes for
metabolic engineering purposes.

Acknowledgements. The author wishes to thank S. Bringer and G. Sprenger for helpful com-
ments, and Prof. H. Sahm for continuous support.

References
1. Swings J, de Ley J (1977) Bacteriol Rev 41:1
2. Sprenger GA (1996) FEMS Microbiol Lett 145:301
3. Kalnenieks U, de Graaf AA, Bringer-Meyer S, Sahm H (1993) Arch Microbiol 160:74
4. Sprenger G (1992) The genus Zymomonas. In: Balows A, Triiper HG, Dworkin M, Harder
W, Schleifer KH (eds) The prokaryotes, vol III, 2nd edn. Springer, Berlin Heidelberg
New York, p 2287
5. Reyes L, Scopes RK (1991) Biochim Biophys Acta 1068:174
6. Toh H, Doelle H (1997) Arch Microbio1168:46
7. Kim YJ, Song KB, Rhee SK (1995) J Bacteriol 177:5176
8. Kalnenieks U, Galinina N, Bringer-Meyer S, Poole RK (1998) FEMS Microbiol Lett
168:91
9. Rogers PL, Lee KJ, Scotnicki ML, Tribe DE (1982) Adv Biochem Eng 23:37
10. Viikari L (1988) CRC Critical Rev Biotechnol 7:237
11. Amin G, Van den Eynde E, Verachtert H (1983) Eur J Appl Microbiol Biotechnol18:1
12. Rogers PL, Lee KJ, Scotnicki ML, Tribe DE (1982) Adv Biochem Eng 23:37
13. Viikari L, Korhola M (1986) Appl Microbiol Biotechnol 24:471
14. Toran-Diaz I, Jain VK, Baratti JC (1984) Biotechnol Lett 6:389
15. Dawes EA, Ribbons DW, Large PJ (1966) Biochem J 98:795
16. Johns MR, Greenfield PF, Doelle HW (1991) Adv Biochem Eng 44:97
17. Viikari L, Korhola M (1986) Appl Microbiol Biotechnol 24:47l
18. Viikari L (1988) CRC Critical Rev Biotechnol 7:237
19. Horbach S, Strohhacker J, Welle R, de Graaf AA, Sahm H (1994) FEMS Microbiol Lett
120:37
20. Sprenger GA (1993) J Biotechnol 27:225
21. Reiser J, Kappeli 0, Fiechter A (1988) Bio/Technol 6:1335
22. Zhang M, Eddy C, Deanda K, Finkelstein M, Picataggio S (1995) Science 267:240
23. Deanda K, Zhang M, Eddy C, Picataggio S (1996) Appl Environ Microbiol 62:4465
24. Weisser P, Kramer R, Sprenger GA (1996) Appl Environ Microbiol 62:4155
25. Feldmann SD, Sahm H, Sprenger GA (1992) Appl Microbiol Biotechnol 38:354
502 15 Metabolic Analysis of Zymomonas mobilis

26. Bringer-Meyer S, Sahm H (1989) Appl Microbiol Biotechnol 31:529

27. De Koning W, van Dam K (1992) Anal Biochem 204:118
28. Bergmeyer HU (1985) Methods of enzymatic analysis, vols VI and VII. VCH, Deerfield
Beach, Florida
29. Weuster-Botz D (1997) Anal Biochem 246:225
30. Moon RB, Richards JH (1973) J BioI Chern 248:7276
31. Salhany JM, Yamane T, Shulman RG, Ogawa S (1975) Proc Nat! Acad Sci USA 72:4966
32. Breitmaier E, Voelter W (1989) Carbon-13 NMR spectroscopy, 3rd edn. VCH, Weinheim
New York
33. Barrett GC, Davies JS (1985) Nuclear Magnetic Resonance spectra of amino acids and
their derivatives. In: Barrett GC (ed) Chemistry and biochemistry of the amino acids.
Chapman and Hall, London New York, p 525
34. Weuster-Botz D, de Graaf AA (1996) Reaction engineering methods to study intracellu-
lar metabolite concentrations. In: Scheper T (ed) Advances in biochemical engineering/
biotechnology, vol 54: Metabolic engineering. Springer, Berlin Heidelberg New York,
p76
35. Lee KJ, Skotnicki ML, Tribe DE, Rogers PL (1980) Biotechnol Lett 2:339
36. Schoberth SM, Chapman BE, Kuchel PW, Wittig RM, Grotendorst J, Jansen P, de Graaf
AA (1996) J Bacterio1178:1756
37. Chance B, Eisenhardt RH, Gibson QH, Louberg-Holm KK (eds) (1964) Rapid mixing
and sampling techniques in biochemistry. Academic Press, New York
38. De Koning W, van Dam K (1992) Anal Biochem 204:118
39. Seiler M, Sauer U, Bailey JE (1994) Intracellular metabolite determination in Escherichia
coli. Institute of Biotechnology, ETH Zurich
40. Theobald U, Mailinger W, Reuss M, Rizzi M (1993) Anal Biochem 214:31
41. Gadian DG (1995) NMR and its applications to living systems, 2nd edn. Oxford Univer-
sity Press, Oxford
42. Srivastava AK, Volesky B (1991) Appl Microbiol Biotechnol 34:450
43. Shulman RG, Brown TG, Ugurbil K, Ogawa S, Cohen SM, Den Hollander JA (1979)
Science 205:160
44. Den Hollander JA, Brown TG, Ugurbil K, Shulman RG (1979) Proc Nat! Acad Sci USA
76:6096
45. Ugurbil K, Rottenberg H, Glynn P, Shulman RG (1978) Proc Nat! Acad Sci USA 75:2244
46. Gillies R J (1994) NMR in physiology and biomedicine. Academic Press, London New
York
47. Hartbrich A, Schmitz G, Weuster-Botz D, de Graaf AA, Wandrey C (1996) Biotechnol
Bioeng 51:624
48. de Graaf AA, Wittig RM, Probst U, Strohhacker J, Schoberth SM, Sahm H (1992) J Magn
Reson 98:654
49. Gillies RJ, Galons J-p, McGovern KA, Scherer PG, Lien Y-H, Job C, Ratcliff R, Chapa F,
Cerdan S, Dale BE (1993) NMR Biomed 6:95
50. Haran N, Kahana ZE, Lapidot A (1983) J BioI Chern 258:12,929
51. Altenburger R, Abarzua S, Callies R, Grimme LH, Mayer A, Leibfritz D (1991) Arch
Microbiol 156:471
52. Duysens LNM, Amesz J (1957) Biochim Biophys Acta 24:19
53. Zabriskie DW, Humphrey AE (1978) Appl Eur Microbiol 35:337
54. Luong JHT, Carrier DJ (1986) Appl Microbiol Biotechnol 24:65
55. Einsele A, Ristroph DL, Humphrey AE (1978) Biotechnol Bioeng 20:1487
56. Boyer PM, Humphrey AE (1988) Biotechnol Lett 2:193
57. Slavik J, Kotyk A (1984) Biochim Biophys Acta 766:697
58. Pena A, Ramirez J, Rosas G, Calahorra M (1995) J Bacterioll77:1017
59. Horvath JJ, Glazier SA, Spangler CJ (1993) Biotechnol Prog 9:666
60. Schulmann SG (1985) Molecular luminescence spectroscopy, methods and applications.
In: Chemical analysis, vol 77. Wiley, New York, p 170
61. Marose S, Lindemann C, Scheper T (1998) Biotechnol Prog 14:63
62. Schafer U, Boos W, Takors R, Weuster-Botz D (1999) Anal Biochem 270:88
References 503

63. Holms WH (1986) Curr Top Cell Reg 28:69

64. Neidhardt FC, Ingraham JL, Schaechter M (1990) Physiology of the bacterial cell: a mo-
lecular approach. Sinauer Associates, Sunderland, Massachusetts
65. Vallino JJ (1991) PhD Thesis, Massachusetts Institute of Technology, Cambridge, MA
66. Vallino JJ, Stephanopoulos G (1994) Biotechnol Progr 10:320
67. Vallino JJ' Stephanopoulos G (1994) Biotechnol Progr 10:327
68. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KChAM (1994) Bio-
technol Bioeng 43:3
69. van der Heijden RTJM, Heijnen JJ, Hellinga C, Romein B, Luyben KChAM (1994) Bio-
technol Bioeng 43:11
70. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
71. Blum JJ, Stein RB (1982) On the analysis of metabolic networks. In: Goldberger R (ed)
Biological regulation and development, vol 3A. Plenum Press, New York, p 99
72. Di Donato L, Des Rosiers C, Montgomery JA, David F, Garneau M, Brunengraber H
(1993) J Bioi Chern 268:4170
73. Cohen SM, Rognstad R, Shulman RG, Katz J (1981) J Bioi Chern 256:3428
74. London RE (1992) In: Berliner LT, Reuben J (ed) Biological magnetic resonance, vol 11.
Plenum Press, New York, p 277
75. Ross BD, Kingsley PB, Ben-Yoseph 0 (1994) Biochem J 302:31
76. Kanamori K, Weiss RL, Roberts D (1988) J Bioi Chern 263:2871
77. Tesch M, de Graaf AA, Sahm H (1999) Appl Environ Microbiol 65:1099
78. Eakin RT, Morgan LO, Gregg CT, Matwiyoff NA (1972) FEBS Lett 28:259
79. Walker TE, Han CH, Kollman VH, London RE, Matwiyoff NA (1982) J Bioi Chern
257:1189
80. Chance EM, Seeholzer SH, Kobayashi K, Williamson JR (1983) J Bioi Chern 258:13,785
81. Walsh K, Koshland DE (1984) J Bioi Chern 259:9646
82. Malloy CR, Sherry AD, Jeffrey FMH (1988) J Bioi Chern 263:6964
83. Sharfstein ST, Tucker SN, Mancuso A, Blanch HW, Clark DS (1994) Biotechnol Bioeng
43:1059
84. Zupke C, Stephanopoulos G (1995) Biotechnol Bioeng 45:292
85. Marx A, de Graaf AA, Wiechert W, Eggeling L, Sahm H (1996) Biotechnol Bioeng 49:111
86. Wiechert W, de Graaf AA (1996) In vivo stationary flux analysis by I3C labeling experi-
ments. In: Scheper T (ed) Advances in biochemical engineering/biotechnology, vol 54:
Metabolic engineering. Springer, Berlin Heidelberg New York, pili
87. Wiechert W, de Graaf AA (1997) Biotechnol Bioeng 55:101
88. Wiechert W, Siefke C, de Graaf AA, Marx A (1997) Biotechnol Bioeng 55:118
89. Follstad BD, Stephanopoulos G (1998) Eur J Biochem 252:360
90. Wiechert W, Mollney M, Isermann N, Wurzel M, de Graaf AA (1999) Biotechnol Bioeng
66:69
91. Mollney M, Wiechert W, Kownatzki D, de Graaf AA (1999) Biotechnol Bioeng 66:86
92. Szyperski T (1995) Eur J Biochem 232:433
93. Szyperski T, Bailey JE, Wuthrich K (1996) Trends Biotechnol 14:453
94. Wendisch VF, de Graaf AA, Sahm H (1997) Anal Biochem 245:196
95. de Graaf AA, Mahle M, Mollney M, Wiechert W, Stahmann P, Sahm H (2000) J Biotech-
nol77:25
96. Park SM, Shaw-Reid C, Sinskey AJ, Stephanopoulos G (1997) Appl Microbiol Biotechnol
47:430
97. Inbar L, Lapidot A (1987) Eur J Biochem 162:621
98. Brown TR, Ugurbil K, Shulman RG (1977) Proc Natl Acad Sci USA 74:5551
99. Alger JR, Shulman RG (1984) Q Rev Biophys 17:83
100. Kirk K (1990) NMR Biomed 3:1
101. de Graaf AA, Wittig RM, Probst U, Strohhacker T, Schoberth SM, Sahm H (1992) J Magn
Reson 98:654
102. Hatzimanikatis V, Emmerling M, Sauer U, Bailey JE (1998) Biotechnol Bioeng 58:154
103. Wulf G, Wiechert W, de Graaf AA, Kramer R, Posten C, Sprenger G, Weuster-Botz D,
Munack A (1997) Modelling, simulation and measurement of metabolic pathways in
504 15 Metabolic Analysis of Zymomonas mobilis

Zymomonas mobilis. In: Gertler JJ, Cruz JB, Peshkin M (eds) Proceedings of the 13th
World Congress, International Federation of Automatic Control, on CD-ROM. Elsevier
Science, Amsterdam
104. Torres NV (1994) Biotechnol Bioeng 44:104
105. Wright BE, Butler MH, Albe KR (1992) J BioI Chern 267:3101
106. Rizzi M, Baltes M, Theobald U, Reuss M (1997) Biotechnol Bioeng 55:592
107. Reder C (1988) J Theor BioI 135:175
108. Liao JC, Lightfoot EN (1988) Biotechnol Bioeng 31:847
109. Shiraishi F, Savageau MA (1992) J BioI Chern 267:22,926
1l0. Torres NV (1994) Biotechnol Bioeng 44:112
Ill. Heinrich R, Rapoport TA (1974) Eur J Biochem 42:89
112. Heinrich R, Rapoport TA (1974) Eur J Biochem 42:97
113. Kell DB, Westerhoff HV (1986) FEMS Microbiol Rev 39:305
114. Kaeser H (1988) Regulation and control of metabolic pathways. In: Bazin MJ, Prosser JI
(eds) Physiological models in microbiology. CRC series in mathematical models in mi-
crobiology, vol 2. CRC Press, Boca Raton, FL, USA
115. Fell DA (1998) Biotechnol Bioeng 58:121
116. Voit EO (ed) (1991) Canonical nonlinear modeling. S-system approach to understanding
complexity. Van Nostrand Reinhold, New York
117. Shiraishi F, Savageau MA (1992) J BioI Chern 267:22,912
118. Torres NV, Voit EO, Gonzalez-Alcon C (1996) Biotechnol Bioeng 49:247
119. Hatzimanikatis V, Floudas CA, Bailey JE (1996) Biotechnol Bioeng 52:485
120. Hatzimanikatis V, Floudas CA, Bailey JE (1996) AIChE Journal 42:1277
121. Mavrovouniotis ML, Stephanopoulos G (1988) Comp Chern Eng 12:867
122. Mavrovouniotis ML, Stephanopoulos G (1989) Biotechnol Bioeng 34:196
123. Mavrovouniotis ML (1993) Identification oflocalized and distributed bottlenecks in me-
tabolic pathways. In: Hunter L, Searls D, Shavlik J (eds) Proceedings of the First Inter-
national Conference on Intelligent Systems for Molecular Biology. AAAI Press, Menlo
Park, CA, USA
124. Mavrovouniotis ML (1993) Identification of qualitatively feasible metabolic pathways. In
Hunter L (ed) Artificial intelligence and molecular biology. AAAI Press, Menlo Park,
CA, USA
125. Nielsen J (1998) Biotechnol Bioeng 58:125
126. Sauer U, Hatzimanikatis V, Bailey JE, Hochuli M, Szyperski T, Wuthrich K (1997) Nat
Biotechnol 15:448
127. Marx A, Striegel K, de Graaf AA, Sahm H, Eggeling L (1997) Biotechnol Bioeng 56:168
128. Zupke C, Stephanopoulos G (1994) Biotechnol Progr 10:489
129. Katz J, Wals P, Lee W-NP (1993) J BioI Chern 268:25,509
130. Lee W-NP (1993) J Bioi Chern 268:25,522
131. Schmidt K, Carlsen M, Nielsen J, Villadsen J (1997) Biotechnol Bioeng 55:831
132. Zupke C, Tompkins R, Yarmush D, Yarmush M (1997) Anal Biochem 247:287
133. Schmidt K, Nielsen J, Villadsen J (1999) J Biotechnol 71:175
134. Yen J, Lee B, Liao JC (1996) In: Proceedings of the 5th IEEE International Conference on
Fuzyy Systems, vol 1. IEEE Press, Piscataway, NJ, USA, p220
135. Edwards JS, Palsson BO (1998) Biotechnol Bioeng 58:162
136. Breuel G, Gilles ED, Kremling A (1995) 6th Conference on Computer applications in
Biotechnology, DECHEMA, p 199
137. Lee KJ, Scotnicki ML, Tribe DE, Rogers PL (1981) Biotechnol Lett 3:207
138. Hardman MJ, Scopes R (1988) Eur J Biochem 173:203
139. Doelle MB, Greeenfieid PF, Doelle HW (1990) Appl Microbiol Biotechnol 34:160
140. Algar EM, Scopes RK (1985) J Biotechnol 2:275
141. Scopes RK (1987) Austral J Biotechnol 1:58
142. Osman YA, Conway T, Bonetti SJ, Ingram LO (1987) J Bacteriol 169:3726
143. Belaich JP, Senez JC, Murgier M (1968) J Bacteriol 95:1750
144. Hesman TL, Barnell WO, Conway T (1991) J Bacterio1173:3215
145. Scopes RK, Testolin V, Stoter A, Griffiths-Smith K, Algar EM (1985) Biochem J 228:627
References 505

146. Pawluk A, Scopes RK, Griffiths-Smith K (1986) Biochem J 238:275

147. Snoep JL, Arfman N, Yomano LP, WesterhoffHV, Conway T, Ingram LO (1996) Biotech-
nol Bioeng 51: 190
148. Anderson AJ, Dawes EA (1985) FEMS Microbiol Lett 27:23
149. Scopes RK (1997) Biochem J 326:731
150. Weisser P, Kramer R, Sahm H, Sprenger GA (1995) J BacterioI177:3351
151. Scopes RK, Bannon DR (1995) Biochim Biophys Acta 1249:173
152. Scopes RK, Griffiths-Smith K (1984) Anal Biochem 136:530
153. Scopes RK (1984) Anal Biochem 136:525
154. Hoppner TC, Doelle H (1983) Eur J Appl Microbiol Biotechnol17:152
155. Kinoshita S, Kakizono T, Kadota K, Das K, Taguchi H (1985) Appl Microbiol Biotechnol
22:249
156. Lazdunski A, Belaich JP (1972) J Gen Microbiol 70:187
157. Barrow KD, Collins JG, Norton RS, Rogers PL, Smith GM (1984) J Bioi Chern 259:5711
158. Strohhacker J, de Graaf AA, Schoberth SM, Wittig RM, Sahm H (1993) Arch Microbiol
159:484
159. de Graaf AA, Striegel K, Wittig RM, Laufer B, Schmitz G, Wiechert W, Sprenger GA,
Sahm H (1999) Arch Microbio1171:371
160. Struch T, Neuss B, Bringer-Meyer S, Sahm H (1991) Appl Microbiol Biotechnol 34:518
161. Schoberth SM, de Graaf AA (1993) Anal Biochem 210:123
162. Lee KJ, Rogers PL (1983) Chern Eng J 27:B31
163. Joebses IML, Roels JA (1986) Biotechnol Bioeng 28:554
164. Joebses IML, Hiemstra HCH, Roels JA (1987) Biotechnol Bioeng 29:502
165. Ingram, LO (1986) Trends Biotechnol 4:40
166. Hermans M (1992) PhD Thesis, University of Diisseldorf
167. Low KS, Rogers PL (1984) Appl Microbiol Biotechnol19:75
168. Sutter B (1991) PhD Thesis, Universite de Haute-Alsace, Mulhouse, France
169. Barrow KD, Collins JG, Rogers PL, Smith GM (1983) Biochim Biophys Acta 753:324
170. Barrow KD, Collins JG, Rogers PL, Smith GM (1983) Eur J Biochem 145:173
171. Barrow KD, Collins JG, Leigh DA, Rogers PL, Warr RG (1984) Appl Microbiol Biotechnol
20:225
172. Barrow KD, Rogers PL, Smith GM (1986) Eur J Biochem 157:195
173. Rohmer M, Sutter B, Sahm H (1989) J Chern Soc Lond, Chern Commun 19:147l
174. Rohmer M, Knani M, Simonin P, Sutter B, Sahm H (1993) Biochem J 295:517
175. Laufer B (1998) PhD Thesis, University of Diisseldorf, Germany
176. Navon G, Shulman RG, Yamane T, Eccleshall R, Lam KB, Baronsfski J, Marimur J (1979)
Biochem 18:4487
16 Metabolic Flux Analysis of Corynebacterium
glutamicum
Albert A. de Graaf

16.1
Introduction

Corynebacterium glutamicum is an aerobic gram-positive bacterium intensively used

in the industrial production of a variety of amino acids. L-Glutamate and L-lysine are
produced in annual quantities of several hundreds of thousands of tons [1] by this
organism. In addition, processes for producing L-tryptophan, L-tyrosine, L-phenyla-
lanine [2], L-valine [3], and histidine [4] have been patented. The organism also has
a high potential for the production of L-threonine [5, 6] and L-isoleucine [7], and
probably for other amino acids as well. Extensive process development and genetic
engineering has been carried out over the years with C. glutamicum and relevant
studies are well-documented in several recent review articles, e.g., specifically on
glutamate production [8], L-isoleucine biosynthesis [7], L-lysine biosynthesis [9],
and aspartate-family amino acid synthesis [10, 11], or more generally on recombi-
nant DNA technology [12] and metabolic engineering [13, 14].
The physiology and mechanism of secretion of amino acids has been intensively
studied [15-19] and several transport genes have been identified [20], including that
of the recently found lysine export carrier [21] which structurally represents a new
type of translocator.
The central carbon metabolic pathways in Corynebacterium glutamicum are
shown in Fig. 16.1; they include the Embden-Meyerhoff-Parnas (EMP) pathway, the
pentose phosphate pathway (PPP), the tricarboxylic acid (TCA) cycle, as well as sev-
eral enzymes for acetate catabolism.
Of the glycolytic pathway, the gene cluster encoding glyceraldehyde-3-phosphate
dehydrogenase, 3-phosphoglycerate kinase, triosephosphate isomerase, and phos-
phoenolpyruvate carboxylase has been transcriptionally analyzed [22] and the genes
for fructose-l,6-bisphosphate aldolase [23], the three enzymes glyceraldehyde-3-
phosphate dehydrogenase, 3-phosphoglycerate kinase, and triosephosphate isomer-
ase [24], and pyruvate kinase [25] have been cloned.
Very recently, the C. glutamicum trans aldolase gene was cloned [26] and the en-
zymes glucose-6-phosphate dehydrogenase and 6-phosphogluconate dehydrogenase
from the oxidative branch of the PPP were kinetically characterized [27].
The TCA cycle genes for citrate synthase [28], isocitrate dehydrogenase [29], and
2-oxoglutarate dehydrogenase [30] have been cloned. During growth on acetate, ac-
tivities of phosphotransacetylase and acetate kinase as well as of the glyoxylate cycle
enzymes isocitrate lyase and malate synthase are strongly elevated. The respective
genes have all recently been cloned [31-33].
16.1 Introduction 507

co,
IGlucose I 7CD'\ ~ G6P~6PG
@4@
Rul5P
PEP Pyr @t.. ®I \®
Rib5P Xul5P
~
@.
F6P

F16BP
Sed7P
<ill
GAP

~
DHAP'trGAP
7 t@
@t 3PGA ++ 13BPGA
@J
@

!@
2PGA+'+ PEP

co, Pyr
l~
~CO,
Ij:
AcCoA .....f-_@""'30"--_Acetyl_P

Fig. 16.1. Scheme of the pathways of central metabolism in Corynebacterium glutamicum. En-
zymes involved: (1) phosphotransferase system; (2) glucose-6-phosphate dehydrogenase; (3) 6-
phosphogluconate dehydrogenase; (4) phosphoglucose isomerase; (S) phosphofructokinase;
(6) fructose-I,6-bisphosphate aldolase; (7) triose phosphate isomerase; (8) phosphoribose iso-
merase; (9) ribulose-S-phosphate epimerase; (10) transketolase; (ll) transaldolase; (12) trans-
ketolase; (13) glyceraldehyde-3-phosphate dehydrogenase; (14) phosphoglycerate kinase; (1S)
phosphoglycerate mutase; (16) enolase; (17) pyruvate kinase; (18) pyruvate dehydrogenase
complex; (19) citrate synthase; (20) aconitase; (21) isocitrate dehydrogenase; (22) 2-oxogluta-
rate dehydrogenase; (23) succinate thiokinase; (24) succinate dehydrogenase; (2S) fumarase;
(26) malate dehydrogenase and/or malate:quinone oxidoreductase [1l4J; (27) isocitrate lyase;
(28) malate synthase; (29) acetate kinase; (30) phosphotransacetylase; (31)-(35) anaplerotic
enzymes (see Fig. 16.2). Abbreviations: G6P, glucose 6-phosphate; 6PG, 6-phosphogluconate;
RuISP, ribulose-S-phosphate; RibSP, ribose-S-phosphate; XuISP, xylulose-5-phosphate; Sed7P,
sedoheptulose-7-phosphate; GAP, glyceraldehyde-3-phosphate; Ery4P, erythrose-4-phosphate;
F6P, fructose 6-phosphate; FI6BP, fructose-I,6-bisphosphate; DHAP, dihydroxyacetone phos-
phate; 13BPGA, I,3-bisphosphoglycerate; 3PGA, 3-phosphoglycerate; 2PGA, 2-phosphoglyce-
rate; PEP, phosphoenolpyruvate; Pyr, pyruvate; AcCoA, acetyl-CoenzymeA; Cit, citrate; leit,
isocitrate; AKG, 2-oxoglutarate; SucCoA, succinyl-CoenzymeA; Suc, succinate; Fum, fumarate;
Mal, malate; OAA, oxaloacetate; acetyl-P, acetyl phosphate. Double-pointed arrows indicate re-
actions that are likely to operate in a reversible manner in vivo
508 16 Metabolic Flux Analysis of Corynebacterium glutamicum

co,

Fig. 16.2. Scheme of the pathways involved in the anaplerosis of Corynebacterium glutamicum.
Enzymes involved: (31) PEPcarboxylase; (32) pyruvate carboxylase; (33) PEPcarboxykinase;
(34) oxaloacetate decarboxylase; (35) malic enzyme

Regarding the anaplerotic reactions, Corynebacterium glutamicum possesses a

rather complete spectrum of enzymes that interconvert C3- and C4-metabolites
(Fig. 16.2). The anaplerotic enzymes phosphoenolpyruvate(PEP) carboxylase [34],
pyruvate carboxylase [35], and malic enzyme [36] were all shown to be present.
Moreover, the organism possesses PEP carboxykinase [37], oxaloacetate decarboxylase
[38], and probably also PEP synthetase [39]. Of these, the genes for PEP carboxylase
[40,41] and pyruvate carboxylase [42,43] have been cloned.
Efficient nitrogen assimilation is of key importance for amino acid production.
When Corynebacterium glutamicum is grown with a sufficient nitrogen supply, the
important nitrogen sources ammonium and urea cross the cytoplasmic membrane
by passive diffusion. Under conditions of nitrogen starvation, energy-dependent up-
take systems for urea and ammonium are synthesized [44,45]; the (methyl)ammo-
nium carrier has been cloned [44]. C. glutamicum was shown to possess glutamate
dehydrogenase [46,47] and glutamine synthetase [48] as the principal ammonium-
assimilating enzymes; genes for these enzymes were cloned [47,48]. Under condi-
tions of ammonium limitation or in glutamate dehydrogenase-deficient mutants, the
glutamine:2-oxoglutarate aminotransferase system is expressed as alternative gluta-
mate-producing enzyme [49, 50]. Figure 16.3 shows the principal ammonium-assim-
ilating enzyme system of C. glutamicum.
The biosynthesis of the amino acids from the aspartate family in C. glutamicum
(Fig. 16.4) is one of the best studied pathway complexes in micro-organisms. Never-
theless, despite the detailed knowledge of the genetics, physiology, and regulation of
the pathways for lysine [9] and isoleucine [7], reported selectivities for these pro-
ducts are still significantly lower than theoretical maxima. In the case of isoleucine,
the transport step may be limiting [51]. In the case of lysine, several authors have
claimed that selectivities are restricted by a non-optimal coordination of central me-
tabolism and amino acid biosynthesis, resulting in a limitation due to, e.g., inade-
quate precursor supply [52] or energy excess [53]. These suppositions have greatly
stimulated the detailed investigation of the interplay between central metabolism
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 509

NH, + H,O+
NADPH+H' NADP'

2-0X09lutarateU9lutamate ymass
[J C@
<F«
.OO

-«-
-y-
coo-
DH
«00" /
tI,N-«-
'9-
-9.-
coo'

Fig. 16.3. Scheme of the principal ammonium-assimilating pathways in Corynebacterium glu-

tamicum. Abbreviations: GDH, glutamate dehydrogenase; GS, glutamine synthetase; GOGAT,
glutamine 2-oxoglutarate aminotransferase

and amino acid biosynthesis in C. glutamicum. This contribution presents a state-of-

the art overview and analysis of studies that employed intracellular metabolic flux
analysis by means of metabolite balancing and/or isotopic labeling combined with
NMR spectroscopy in order to correlate activities of the central metabolic pathways
in vivo with amino acid overproduction.

16.2
Fundamentals of Intracellular Metabolic Flux Analysis in Corynebacterium
glutamicum

16.2.1
Metabolite Balancing

This part presents the elementary data and flux model necessary to perform meta-
bolite balancing with C. glutamicum.

16.2.1.1
Biomass Composition

During non-limited aerobic growth of C. glutamicum, up to 60% of the carbon may

be converted to biomass. Therefore, the first step in metabolic flux analysis is the
detailed calculation of the precursor requirements for anabolic purposes. For C. glu-
tamicum, this has been done, e.g., in [54-56]. These analyses were all based on the
510 16 Metabolic Flux Analysis of Corynebacterium glutamicum

i
L-Aspartate 1

ATP----,J
ADP---.CD
L-Aspartylphosphate
NADPH-------.J '2'
NADP_+'6J @
L-Aspartate- ~""';=:""'-"", Homoserine • • • 1 i.-Methionine 1
P t
yruvae~@
semialdehyde
@! . .

L-2,3-Dihydrodipicolinate Homoserine-
NADPH:==:::t @ phosphate
NADP L-Piperideine- @ ~
suc~~a:., ~rbOXYlate I@rrninel

N-Succinyl-2-amino- 2-Ketobutyrate

-----J® ®
6-ketopimelate
NADP:S G',lama', '16" 1 _ Pyruvate
NAOPH oxOglutarate~ NH~'" ~t
NH - N-Succinyl-2,6-L,L- NADPH Acetohydroxy-

t
• diaminopimelate NADP butyrate

S,,,inata ~ (j) ®
L'L-Diamin~8imelate Dihydroxymethyl-
valerate

D,L-Diaminopimelate
@!
Ketomethylvalerate

t@
L-Lysine
@)!
L-Isoleucine
~® @)~
I L-Lysinel I L-Isoleucine I

Fig_ 16_4_ Scheme of the biosynthetic pathways of amino acids from the aspartate family in
Corynebacterium glutamicum. Enzymes and genes involved: (1) aspartate kinase (lysC); (2)
aspartate semialdehyde dehydrogenase (asd); (3) dihydrodipicolinate synthase (dapA); (4) di-
hydrodipicolinate reductase (dapB); (5) tetrahydrodipicolinate succinylase (dapD); (6) N-suc-
cinyl aminoketopimelate transaminase; (7) N-succinyl diaminopimelate desuccinylase (dapE);
(8) diaminopimelate epimerase; (9) diaminopimelate dehydrogenase (ddh); (10) diaminopime-
late decarboxylase (lysA); (ll) permease (lysE); (12) homo serine dehydrogenase (hom); (13)
homoserine kinase (thrB); (14) threonine synthase (thrC); (15) threonine dehydratase (ilvA);
(16) acetohydroxyacid synthase (ilvB, ilvN); (17) isomeroreductase (ilvC); (18) dihydroxyacid
dehydratase (ilvD); (19) transaminase; (20) permease, Cofactors are shown only for the lysine
biosynthetic pathway

approach and data of Neidhardt et aL [57] for E. coli, a Gram-negative organism.

Since C. glutamicum is a Gram-positive organism, the data have to be corrected for
the different cell wall composition, i.e_, the higher peptidoglycan content. This can
be done on the basis of the diaminopimelate content of C. glutamicum, as given in
[56]. The calculated precursor requirements are given in Table 16_l.
This table can also be used to calculate that C. glutamicum requires 9816 ttmollg
dry wt of nitrogen for its elementary composition, of which 1773 ttmollg dry wt are
accounted for by glutamine amidotransferase reactions in the synthesis of arginine,
histidine, tryptophan, N-acetylglucosamine, and the pyrimidine and purine nucIeo-
tides of RNA and DNA. The remaining 8043 ttmollg dry wt are accounted for· by
glutamate dehydrogenase (7137 ttmollg dry wt) as well as a number of directly ami-
nating enzymes (906 ttmollg dry wt).
Table 16.1. Precursor as well as carbon dioxide requirements in {lmol/g of dry weight for biomass synthesis of Corynebacterium glutamicum. The amino .....
acid composition was determined from C. glutamicum strain MH20-22B leuA grown in continuous culture [56]. The peptidoglycan content was inferred 0\
N
from the diaminopimelate (Dap) content. The high intracellular glutamate and glutamine pools of C. glutamicum [50] were explicitly taken into account. "I'l
The distinction between glutamate and glutamine in the cell protein was made based on C. glutamicum codon usage data (Eikmanns BJ, Tesch M, de ~
::s
Graaf AA, unpublished results). Relative amounts of the (deoxy)nucleotides were estimated taking into account that the relative GC content of C.
glutamicum is 56%. The data on other cellular constituents were taken from the literature [57]. NADPH stoichiometries were inferred from data pre- (1)
~
sented in [113] Abbreviations: PEP, phosphoenolpyruvate; AcCoA, acetyl-coenzymeA a
e.
<J>
Precursor stoichiometry (mol/mol amino acid) o...,
Amino Amount(p,mol!g G6P F6P Ri5P E4P GAP PGA PEP Pyr AcCoA OAA AKG CO2 NADPH ...,
acid dry weight) a...
$>l

Ala 606 1
$>l
Arg 189 4
~
...
Asx 399 s:::
Cys 87 5 ~
Glu 360
go
g.
GIn 147 ;::;'
Glu_pool 250
GIn_pool 49 ~
Gly 361 ~
His 71
Ile 202 -1 5
-[
<J>

Leu 440 2 -2 2
Lys 202 -1 4
Met 146 8
Phe l33 2 -1 2
Pro 170 3
Ser 225
Thr 275 3
Trp 54 -1 2
Tyr 81 2 -1 2
Val 284 2 -1 2 U1
Dap 146 4
Table 16.1. Continued VI
......
N

Precursor amount (ttmol/g dry weight)

Polymer G6P F6P Ri5P E4P GAP PGA PEP Pyr AcCoA OAA AKG CO2 NADPH

Protein total 0 0 125 268 0 673 482 2604 440 1370 1165 -1647 10548
RNA ATP 152 152 152 456
GTP 218 218 218 436
UTP 125 125 125 125
CTP 135 135 135 135
DNA dATP 22 22 22 88
......
dGTP 28 28 28 84 0\

dCTP 28 28 28 56 ~
dTTP 22 22 22 44 ~
~
Lipids 129 129 2116 -2116 3612 2.-
LPS 51 16 24 24 24 329 -329 470 n'
Peptido- 292 146 0 292 0 0 -292 146
glycan ~
Glycogen 154 0 :>
::l
~
C1-Units 49 49
Poly- 59 180
'[
en
amines 0
....,
total 205 308 879 268 129 1295 652 2604 3177 1680 1224 -3654 16429 (]
c
~
;:::
<J-
'"
:;:,
"....<Ii
;,:'
::1
<>.3..
.:
;:;-
::1
"',
.:
::1
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 513

16.2.1.2
Condensed Bioreaction Network

The condensed bioreaction network of C. glutamicum can be obtained from

Figs. 16.1, 16.2, and 16.4 by lumping all reactions that occur in a linear sequence
without branching points as well as all sets of reactions that cannot be mutually
discriminated because they have identical overall reaction equations. This results in
the following set of equations describing central metabolism (water is omitted, upt
denotes glucose uptake via the PTS system, ace denotes acetate uptake and activa-
tion, ppp denotes pentose phosphate pathway, emp denotes Embden-Meyerhof-Par-
nas pathway, ana denotes anaplerosis, tcc denotes citric acid cycle, gs denotes glyox-
ylate shunt):
upt: Glucose+ PEP--+G6P+ Pyr
ace: Ace+ATP+CoA--+AcCoA+ADP
pppl: G6P+2 NADP--+RuISP+CO z+2 NADPH
ppp2: RuISP--+XuISP
ppp3: RuISP--+Rib5P
ppp4: XuISP+ Ery4P--+ F6P+GAP
pppS: XuISP+ RibSP--+Sed7P+GAP
ppp6: Sed7P+GAP--+Ery4P+F6P
empl: G6P--+F6P
emp2: F6P+ATP--+FI6BP+ADP
emp3: FI6BP--+2 GAP
emp4: GAP+ADP+NAD--+PGA+ATP+NADH
empS: PGA--+PEP
emp6: PEP+ADP--+Pyr+ATP
anal: PEP+COz--+OAA
ana2: Pyr+C0 2 +NADPH--+Mal+NADP
ana3: Pyr+C0 2 +ATP--+OAA+ADP
teel: Pyr+CoA + NAD--+AcCoA +CO z+NADH
tcc2: AcCoA +OAA --+ Icit+CoA
tcc3: Icit+ NADP--+AKG+CO z+NADPH
tcc4: AKG+NAD+ADP--+Suc+COz+NADH+ATP
teeS: Suc+FAD--+Mal+FADH
tce6: Mal+NAD--+OAA+NADH
gsl: Icit----+Suc+GlyOx
gs2: GlyOx+ AcCoA --+ Mal+CoA

The sequence emp6-ana3 is completely identical to the single reaction anal. There-
fore, anal and ana3 must also be lumped, i.e., ana3 is omitted from the set of equa-
tions. The sequences emp6-ana2 and anal-tcc6 (reversed) are distinguishable in
principle because of their different cofactor requirements.
The two different pathways for lysine synthesis in C. glutamicum [S8J are lumped
into:
lys: OAA+Pyr+ATP+4NADPH--+Lys+COz+ADP+4NADP

The NADPH is used directly as shown in Fig. 16.4, or indirectly in the regeneration
of glutamate from 2-oxoglutarate. Glutamate is nitrogen donor in the transamination
of oxaloacetate to aspartate, a precursor of lysine.
514 16 Metabolic Flux Analysis of Corynebacterium glutamicum

NAD, FAD, and ATP are regenerated by oxidative phosphorylation with different
P/O stoichiometries Sl and S2 (water is again omitted):
oxp1: NADH+1/2 02+S1 ADP-+NAD+S1 ATP
oxp2: FADH+ 1/2 02+S2 ADP-+FAD+S2 ATP

The reactions that describe the withdrawal of precursor metabolites (G6P, F6P,
Rib5P, Ery4P, GAP, PGA, PEP, Pyr, AcCoA, OAA, AKG) and the concomitant use of
NADPH as well as the fixation and release of CO 2 can be written using the stoichio-
metries given in Table 16.1.

16.2.1.3
Approaches to Resolve Network Underdeterminacy

Rather than by using matrices, the extent to which the condensed bioreaction net-
work can be determined from extracellular (i.e., involving net import into or output
from the cells) measurements can be judged rather conveniently from Fig. 16.1, as
described in the following. Once the biomass yield on the substrate is measured, all
reactions removing precursors for anabolic purposes are known from the yield value
and the stoichiometries given in Table 16.1. Now consider as a typical network node
glucose 6-phosphate (G6P) (Fig. 16.5). At metabolic steady state, the G6P pool does
not change, whence the total flux into the pool must exactly match the total efflux
from the pool. Or,

upt =pppi + emp1 + bs_G6P (16.1)

where the latter denotes the anabolic precursor requirement of G6P. Since the speci-
fic glucose uptake rate is also measured, upt and bs_G6P are both known. Therefore,
only the sum pppl+empi is determined by the measurements. Two conclusions can
be drawn at this stage. First, the anabolic precursor requirement fluxes do not have
to be considered in the judgment of the network determinacy, since all are known
(i.e., omitting bs_G6P from Eq. (16.1) still leaves only the sum ppp1+empl deter-
mined by the measurements). Second, a value for pppi must be chosen in order to
determine empl. Having done this and applying the first conclusion, it can easily be
verified from Fig. 16.1 that ppp2, ppp3, ppp4, ppp5, ppp6, emp2, emp3, emp4, and
emp5 are all determined by the measured specific glucose uptake rate, the measured
biomass yield, and our choice of pppi in combination with the reaction stoichiome-
tries. For the next branching point, phosphoenolpyruvate (PEP), we then have:

emp5 = anal + upt + emp6 (+bs_PEP), (16.2)

bs G6P r
.4~-~---1 G6P I--~~~
ppp1

tmp1
Fig. 16.5. Metabolite fluxes at the glucose-6-phosphate node. The flux names are as defined in
the text
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 515

forcing us to choose a value for anal in order that emp6 be determined. At the next
node, pyruvate (Pyr), a value for ana2 must be chosen in order that tcc1 be deter-
mined. This then suffices to determine all fluxes since the total of oxaloacetate and
malate synthesized via anal, ana2, and gs2 must equal the precursor amounts of
oxaloacetate plus 2-oxoglutarate, i.e., gs2 is determined by our choice of anal and
ana2. With tcc1, gs2, anal, and ana2 known, tcc2, gsl, tcc3, tcc4, tcc5, and tcc6 all
follow consecutively.
Summarising, the metabolic network at this stage still has three degrees of free-
dom: pppl, anal, and ana2 (using different choices at the branch points G6P, PEP,
and Pyr, one could consider as completely equivalent empl, emp6 and tcc1 as the
free fluxes in the model). Unless further action is taken, the flux analysis problem
cannot be solved. If no further measurements can be performed, basically three op-
tions remain:
1. Further lumping of reactions
2. Exclusion of reactions based on biochemical considerations
3. Inclusion of other balances than carbon, e.g., redox or energy balances
In the case of C. glutamicum, these three approaches can all, respectively, be used:
anal and ana2 may be lumped into a single reaction originating from a combined
PEP/pyruvate pool; gs2 can be assumed zero during growth on glucose based on
enzyme measurements; and pppi can be determined by assuming that no net synth-
esis or consumption of NADPH occurs in the network defined by the reactions pre-
sented in the previous paragraph. The latter in fact creates one additional balance
equation:
2 x pppi + tcc3 =ana2 + 4 x lys + bs_NADPH (16.3)
where bs_NADPH denotes the total anabolic precursor requirement of NADPH. The
combined application of all three measures results in a completely determined equa-
tion system. As an alternative to the lumping of anal and ana2, one of these two
reactions may be assumed absent, enabling one to retain separate pools of PEP and
pyruvate. One might also include oxpl and oxp2 in order to create an additional
balance equation based on NADH conservation. Application of this procedure re-
quires the measurement of the oxygen consumption rate in addition to knowledge
of the stoichiometries Sl and S2 (see preceding paragraph).

16.2.1.4
Theoretical Lysine Selectivity

An important benefit of the metabolite balancing procedure is that it allows the the-
oretical consideration of optimal product yields as well as of the flux configurations
necessary to produce these. Different maximum product yields will be arrived at,
depending on the constraints applied. Lysine synthesis is a good example. For the
purpose of this illustration, let it be assumed that we are dealing with a population
of non-growing C. glutamicum.
If one considers only the carbon balance and disrespects all biochemistry, an over-
all reaction stoichiometry of
1 glucose---+ 1 lysine
would be allowed, or a molar yield of 100%. Considering that 4 moles of NADPH are
consumed per mole of lysine formed, and assuming that NADPH can only be regen-
516 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Glucose ~ G6P --~-_ __

IATPI~
F6P+---'<!;:':!.3....
~ATPI
F16BP

A
GAP +------"~

~IATPI
~INADHI
PEP

.---PYR
~IATPI

~
TP ~~
CO c~o,
AcCoA
, -F/3-2

L ~AA ~Icit
ATP ~HJ NADPH CO,
M~DH'I AKG
~'"
Lysine Suc~CO,
Fig. 16.6. Reaction network with cofactor stoichiometries for non-growing, I-lysine-producing
Corynebacterium glutamicum. Symbols in ovals denote molar fluxes

erated in the oxidative pentose phosphate pathway with concomitant loss of COl> one
arrives at the following reaction stoichiometry:
4 glucose-+ 3lysine+6 COl>

or a molar yield of 75%.

Now consider a somewhat more realistic calculation that takes into account ATP,
NAD, and NADP. Figure 16.6 shows the reaction network in the absence of growth. It
is assumed that pyruvate carboxylase is the only active anaplerotic reaction. The
glucose uptake rate is normalized to 1. It is easily verified that the fluxes as given
in Fig. 16.6 yield a closed carbon balance. The need for a closed NADPH balance
yields the following requirement for the activity F of the oxidative pentose phos-
phate pathway as a function of the lysine production rate L:

F=(18 x L-6)/5 (16.4)

Inspection of the flux distribution (cf. Fig. 16.6) shows that even at the imaginary
yield of 100% (i.e., L=1), no fructose-l,6-bisphosphatase activity is required. This
result differs from that of [53].
Setting up the balance for NADH and substituting Eq. (16.4) yields a net synthesis
ofNADH:

RNADH =8 - 4 x F/3 - 6 x L = (48 - 54 x L)/5 (16.5)

For FADH we have analogously

16.2 Fundamentals of Intracellular Metabolic Flux Analysis 517

RFADH = 2 - F/3 - 2 x L = (12 - 16 x L)/5 (16.6)

The net requirement for ATP is given by

RATP =2 x L (16.7)

This ATP has to be generated by oxidative phosphorylation (reactions oxpl and

oxp2). Assuming a P/O ratio of2 [59], i.e., P/O stoichiometry S1=2 in oxpl and S2=1
in oxp2, the oxidation of NADH and FADH yields an amount OATP of ATP:

OATP = 2 X RNADH + RFADH = (108 - 124 x L)/5 ( 16.8)

Requiring that OATP-RATP>O and substituting Eqs. (16.7) and (16.8) yields an
upper value of 108/134=0.806 for L based on the ATP balance. Inspection of
Fig. 16.6 reveals, however, that this would imply a reductive operation of the TeA
cycle and a negative pyruvate dehydrogenase flux. Therefore, a more or less realistic
uppermost lysine yield results when all pyruvate is channeled to lysine. From
Fig. 16.6, it is easily seen that this implies

2 x L=2-F/3 (16.9)

which after substitution of Eq. (16.4) yields a value of 0.75 for L and a net synthesis
of 1.5 mol ATP per mol glucose. The translocation of protons for lysine transport
[60] requires only half that amount of ATP if an effective stoichiometry of 1 mol
ATP per mol of lysine transported is adopted (it is likely to be less). The required
pentose phosphate pathway activity at the yield L=0.75 according to Eq. (16.4) is
equal to 1.5. No transhydrogenase activity would be necessary.

16.2.1.5
Limitations

The fundamental limitation of the metabolite balancing approach is of course that

the assumptions necessary to arrive at a fully determined equation system may not
be valid. Membrane-associated additional NADPH oxidase activity not accounted for
in the analysis would for instance affect the NADPH balancing and result in an un-
derestimation of the pentose phosphate pathway activity. Likewise, a residual activity
of the glyoxylate pathway not accounted for would result in an overestimation of the
anaplerotic carboxylating reaction rates.
A second, also very important limitation of the technique, is that it is not able to
resolve a lumped pathway as well as the degree of reversibility of reversible enzyme
reactions. Thus, metabolite balancing can for instance give no clues as to which ana-
plerotic enzyme is the predominant supplier of oxaloacetate for lysine biosynthesis
in vivo, a very important issue in the chemical engineering of C. glutamicum [61].
Also, metabolic cycles cannot be allowed for by metabolite flux balancing since the
technique considers only net fluxes. However, metabolic cycles may play an impor-
tant regulatory role, as well as having an impact on the energy balance of the cell. To
answer questions related to such topics, stable isotope labeling approaches are re-
quired.
518 16 Metabolic Flux Analysis of Corynebacterium glutamicum

16.2.2
Isotopic Labeling Combined with NMR Spectroscopy

The use of stable isotopes allows the metabolic fate of single atoms to be traced
through the metabolic network. The analysis of timecourses from dynamic labeling
experiments allows one to calculate directly absolute fluxes for the pathways in-
volved. The measurement of label at isotopic steady state allows the determination
of the relative fluxes, since the fractional enrichment of a metabolite is determined
by the ratio of the fluxes coming from different source metabolites and by the frac-
tional enrichment of these source molecules. Thereby, the limitations of the metabo-
lite balancing technique mentioned above can be overcome. NMR spectroscopy ap-
pears to be particularly suitable for the analysis of isotopic labeling experiments
since it does not require chemical derivatization of the labeled compounds.
This part introduces the application of 13 C and ISN labeling to the study of meta-
bolic rates in C. glutamicum. It will be shown that 13 C labeling data from pyruvate,
oxaloacetate, and lysine (or from their successors, e.g., alanine, aspartate) allow one
to resolve the flux distribution at the glucose-6-phosphate and PEP/pyruvate branch-
points, in addition to the flux distribution over the parallel diaminopimelate path-
ways in lysine biosynthesis. ISN labeling allows one to determine the flux distribu-
tion over the primary ammonium-assimilating enzymes of C. glutamicum.

16.2.2.1
Isotopic Atom Balancing

In order to calculate metabolic fluxes from measured isotopic enrichments, mathe-

matical expressions relating both must be formulated. The basic procedure is to es-
tablish specific isotope mass balances around single atoms. This is exemplified here
for the case of!3e. Thus, for an arbitrary atom Q of an arbitrary metabolite pool that
receives inputs Fk from several other metabolite carbon atoms with fractional l3C
enrichments Pk while Q itself is supplying several other pools with rates Ej
(Fig. 16.7), we can write for the rate of change of the isotopic content:

d( l3 C_Q)/dt = FI x PI + F2 X P2 + ... + FN X PN
- (E I + E2 + ... + EM) x ( l3 CQ)/[Q] (16.10)

where l3C_Q denotes the absolute 13 C content of pool Q, and [Q] denotes the pool
size (concentration) of Q. Equations of this type can be used for dynamic simula-
tions. At isotopic steady state, d(l3C_Q)/dt=O and Eq. (16.10) transforms to

(16.11)

[~~r' FN !..........!
Fig. 16.7. Principle of carbon atom balancing. See text
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 519

The formulation of these simple-type isotopic balance equations allows one to gain
important basic insights in flux analysis by isotope labeling as will be shown in the
next sections.

16.2.2.2
Resolving Glycolysis and Pentose Phosphate Pathway

As a first example, the flux partition over glycolysis and the pentose phosphate path-
way at the branchpoint glucose 6-phosphate will be analyzed under the simplifying
condition that all fluxes are unidirectional. It will be shown how the 13C label in C-3
of glyceraldehyde 3-phosphate (precursor of pyruvate) under isotopic steady state
conditions reflects the pentose phosphate pathway activity. In the following equa-
tions, the l3C fractional enrichment of the i-th carbon atom of metabolite X is de-
noted by X_i.
The relevant part of the metabolic network, along with the definition of the fluxes,
is shown in Fig. 16.8. The substrate glucose is labeled only in C-1, with enrichment
La. Since the fructose-1,6-bisphosphate pool is supplied only from fructose 6-phos-
phate, its l3C-labeling at steady state is identical to that of the latter. The pentose
phosphates are assumed to be in rapid equilibrium. GAP _3 results from three differ-
ent sources for which equations must be formulated: F16BP _1 (equal to F6P _1),
F16BP _6 (equal to F6P _6), and Xul5P_5 (in the two transketolase reactions, see
Fig. 16.8).
For glucose 6-phosphate, Eq. (16.11) translates to
G6P _1 = fax Lo/(fl + f2) = La

G6P_6 = fax O/(fl + f2) = 0

Since the pentose phosphates are only synthesized from glucose 6-phosphate, they
will be unlabelled in all positions. Sedoheptulose 7-phosphate is only synthesized
from the pentose phosphates and will therefore also be unlabelled. Erythrose 4-phos-
phate, which is only synthesized from sedoheptulose 7-phosphate, will also be un-
labeled. We then have for fructose 6-phosphate (Eq. 16.11, Fig. 16.8):
F6P_1 = (fl x G6P _1 + f3 x Sed7P_1 + f3 x Xul5P _1 )/(fl + 2f3) = fl x Lo/(fl + 2f3),

F6P _6 = (fl x G6P_6 + f3 x GAP_3 + f3 x Ery4P_ 4)/(fl + 2f3) = f3 x GAP _3/(fl + 2f3),
and for glyceraldehyde-3-phosphate:
GAP_3 = {2f3 x Xul5P_5 + (f1 + 2f3) x (F16BP_1 + F16BP_6)}/(2f1 + 6f3)
= {(f1 + 2f3) x (F6P_1 + F6P _6) }/(2f1 + 6f3)
= {f1 x La + f3 x GAP_3}/(2f1 + 6f3)'
Rearranging terms and substituting f3=f2/3 yields
GAP_3 = La x f1/(2f1 + 5f3) = La x (fa - f2)/(2fo - f2/3) (16.12)
Figure 16.9 shows GAP _3 as a function of f2. According to this graph, NMR mea-
surement of the l3C enrichment in C-3 of glyceraldehyde-3-phosphate (or one of its
successors such as pyruvate) enables the sensitive and unequivocal determination of
the pentose phosphate pathway activity.
520 16 Metabolic Flux Analysis of Corynebacterium glutamicum

1
G6P ' -1'
. ~-®
5
• 6

F6P

f~ GAP
F16BP ~ 1 <~-.-- .. -----.+>,
. ~ -:::::::::=~
f5
.
':'
~
4 GAP
6 5
6

Fig. 16.8. Carbon atom transitions in the pentose phosphate pathway. The rectangular boxes
represent individual carbon atoms, with numbers according to their position in the original
glucose molecule. Small numbers next to the boxes indicate the carbon numbering in the re-
spective metabolites. Double-pointed arrows symbolize complete equilibration of different la-
beled species within one metabolite pool. The fluxes fo-f3 and fIx are used for modeling pur-
poses (see text). The bidirectional flux of the phosphoglucose isomerase reaction fIx is only
used in Eq. (16.17) and following. Abbreviations: G6P, glucose 6-phosphate; F6P, fructose 6-
phosphate; F16BP, fructose-1,6-bisphosphate; GAP, glyceraldehyde 3-phosphate; RuI5P, ribu-
lose 5-phosphate; Rib5P, ribose 5-phosphate; XuI5P, xylulose 5-phosphate; Sed7P, sedoheptu-
lose 7-phosphate; Ery4P, erythrose 4-phosphate

16.2.2.3
Resolving the Parallel Lysine Biosynthetic Pathways

The parallel lysine biosynthetic pathways (see Fig. 16.4), not accessible to metabolite
balancing, are readily resolved using l3C labeling and NMR. Figure 16.10 shows the
relevant labeling scheme (adapted from (62)). Operation of the succinylase variant of
the diaminopimelate pathway results in label scrambling since a symmetric molecule
is involved. For the labeling oflysine we have (Eq. 16.ll, Fig. 16.10):
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 521

50

f 40
~
~
30
M
I 20 ... ......... ,-
a..
04:
(!)
10

o ~----~~~--~~--~--~--~
o 10 20 30 40 50 60 70 80 90 100
f, (%)
Fig. 16.9. Fractional 13C enrichment GAP of C-3 of glyceraldehyde 3-phosphate as a function
of the activity f2 of the oxidative pentose phosphate pathway (expressed as % of the glucose
uptake rate), calculated using Eq. (16.12)

L-Aspa~

L-2-Amino-6-
ketopimelate
Dtamrnopimelate
dehydrogl!f1ase
pothwa, _ - - - ---_~--....

N-Succinyl-L,L-2,6- ,
diaminopimelate

co~
L- Lysine
L~L"i" ~
Fig. 16.10. Carbon atom transfer scheme for the biosynthesis oflysine from its precursors as-
partate and pyruvate. Black and shaded boxes indicate the routes of aspartate C-2 and pyruvate
C-2, respectively. The intermediate N-succinyl-I,I-2,6-diaminopimelate is symmetric, which
leads to label scrambling in the succinyl diaminopimelate pathway

Lys_6 = (fD x Pyc2 + 1/2 x fs x Asp_2 + 112 x fs x Pyc2)/(fD + fs)

Addition of the two equations yields

or
522 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Subtracting the two equations yields

Lys_2 - Lys_6 = fD x (Asp_2 - Pyr_2)/{fD + fs)
Substitution of Asp_2 and rearrangement of terms yields for the flux partition of
the dehydrogenase variant of the diaminopimelate pathway:
(16.12A)
Thus, NMR measurement of the l3C-Iabel in the relevant carbons of pyruvate-de-
rived alanine as well as of lysine resolves the fluxes.

16.2.2.4
Reso/ving Anap/erosis, Citric Acid Cycle, and the G/yoxy/ate Shunt

The analysis of the fluxes over anaplerosis (fA)' citric acid cycle (fT), and glyoxylate
pathway (fG) using simple isotopic atom balances is somewhat more demanding. The
relevant labeling scheme along with the fluxes are shown in Fig. 16.11. It is assumed
that pyruvate carboxylase is the only active anaplerotic carboxylating enzyme. The
total of the anaplerotic and glyoxylate pathway fluxes fA +fG must equal the sum of
the precursor fluxes of 2-oxoglutarate plus oxaloacetate, bs_AKG+bs_OAA. It is im-
portant to note that the reaction from succinate to oxaloacetate involves complete
scrambling of the l3C label between C-1 and C-4, and between C-2 and C-3.
The strategy is to express the labeling of the oxaloacetate carbons as a function of
the labeling of pyruvate. For our purpose, it suffices to consider only C-2 and C-3 of
oxaloacetate, denoted as 0_2 and 0_3, and C-2 and C-3 of pyruvate, denoted as P_2
and P_3, respectively.
The basic equations are (Eq. 16.11, Fig. 16.11):
0_2= {fAX P_2 + 112 x fT x {0_2+ P_3) +
112 x fG x (0_2 + P_3) + fG x 0_3}/{fA+ fT + 2fG) (16.13)

0_3 = {fA X P_3+ 112 X fT x {0_2+ P_3)+

112 x fG x (0_2 + P_3) + fG x P_3}/{fA+ fT + 2fG). (16.14)
For ease of description it is assumed that pyruvate is labeled only in C-3, i.e.,
P_2=0. Subtracting Eq. (16.14) from Eq. (16.13) gives
0_2 - 0_3 = {fG x 0_3 - (fA + fG) x P_3}/{fA+ fT + 2fG)
or
0_3 = {(fA + fT + 2fG) x 0_2 + (fA + fG) x P_3}/{fA + fT + 3fG) (16.15)
Substitution of Eq. (16.15) in Eq. (16.13) yields the desired function
0_2 = (R/S) x P_3 (16.16)
with
R = fT + fG + 2fG X (fA + fG)/{fA + fT + 3fG)
and
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 523

co,
1
'I'
, ,
,,
Pyr

I OM

H
r ~
3
U
2
3 II

'6 CO
'u
'2
'
• J 3 AKG

'~ '4 ~_AKG 4 CO,

Fig. 16.11. Model of carbon atom transfer in the anaplerotic carboxylation reactions (ana; flux:
fA)' the glyoxylate shunt (gs; flux: fG) and the tricarboxylic acid cycle (tee; flux: fT). The rec-
tangular boxes represent individual carbon atoms, with numbers according to their original
position in pyruvate (dark-shaded boxes) or oxaloacetate (shaded boxes). Small numbers next
to boxes indicate the carbon numbering in the molecule. Abbreviations: Pyr, pyruvate; AcCoA,
acetyl-coenzymeA; !cit, isocitrate; AKG, 2-oxoglutarate; Sue, succinate; OAA, oxaloacetate. The
precursor requirement fluxes of OAA and AKG for anabolic reactions are represented by
bs_OAA and bs_AKG, respectively

Figure 16.12 shows a set of curves obtained using Eq. (16.16) for realistic values of
P_3 (i.e., 30%) and the total anaplerotic flux (fA +fG) (i.e., 20% of the glucose uptake
rate). Considering that the typical label measurement inaccuracy is 0.5% 13C,
Fig. 16.12 illustrates that the isotopic labeling approach enables a good resolution of
the fluxes over anaplerotic carboxylation and the glyoxylate pathway, which is im-
possible by the metabolite balancing approach.

16.2.2.5
Resolving the Principal Ammonium-Assimilatory Pathways

By monitoring the preferred site of incorporation of 15N label upon incubation with
15N-labeled ammonium, researchers have been able to identify the principal ammo-
nium-assimilating enzymes in several microorganisms [63-65]. For our purpose, the
524 16 Metabolic Flux Analysis of Corynebacterium glutamicum

35.0

-
~
~
30.0

25.0

e... 20.0

15.0 o f y =10%
N
I A 1,=20%
0 10.0 .1,=40%
• 1,=80%
5.0
... 1,=160%
0.0
0.0 5.0 10.0 15.0 20.0
fG (%)
Fig. 16.12. Fractional 13C enrichment 0_2 of C-2 of oxaloacetate as a function of the activity fG
of the glyoxylate shunt (expressed as % of the glucose uptake rate) at various citric acid cycle
fluxes fTas indicated, calculated using Eq. (16.16) under the assumption that C-3 of pyruvate is
30% 13C-enriched, and that the total anaplerotic flux is 20% of the glucose uptake rate (molar
units)

50.0

40.0

IS
~ 30.0

(")1 20.0 o f,x=O%

11. o 1,,=15%
«C) I:l. f,x =40 %
10.0 • 1,,=100%
.... 1,,=400%
0.0 +----'---;---'--.-;.---'--;-.--'---;.---''---'0
0.0 20.0 40.0 60.0 80.0 100.0
f2 (%)
Fig. 16.13. Fractional 13C enrichment GAP_3 of C-3 of glyceraldehyde-3-phosphate as a func-
tion of the activity f2 of the oxidative pentose phosphate pathway at various bidirectional
fluxes fIX of the phosphoglucoseisomerase reaction (see Fig. 16.8) as indicated, calculated
using Eq. (16.23). The broken line represents the curve that results when glucose 6-phosphate
and fructose 6-phosphate are in extremely fast equilibrium. All fluxes are expressed as % of
the glucose uptake rate

dynamic behavior of the 15N labeling of the glutamate and glutamine nitrogen atoms
is modeled with a differential equation approach analogous to Eq. (16.10). Three
principal ammonium-assimilating pathways exist in C. glutamicum ([46-50] and re-
ferences therein): the glutamate dehydrogenase (GDH) pathway, the glutamine
synthetase (GS) pathway, and the glutamine 2-oxoglutarate-aminotransferase (GO-
GAT) (see Fig. 16.3). Denoting the corresponding fluxes with FGDH , FGs , and FGOGAT,
respectively, and the precursor fluxes of glutamate and glutamine for biomass bio-
synthesis as BGlu and BGln> respectively, the following differential equations describe
the simplified system for the case that only net fluxes are considered.
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 525

For the Na pool of glutamate:

[Glu] x dlsNaGlu/dt = FGDH x ISNH 4 + FGOGAT x CSNaGln + ISNoGln}-
{FGS + BGlu } x ISNaGlu
For the Na and No pools of glutamine:
[Gin] x d ISNaGln/dt = FGS x ISNaGlu - {FGOGAT + BG1n } X ISNaGln

[GIn] x dlsNoGln/dt = FGS x ISNH 4 - {FGOGAT + BG1n } X ISNoGln

As initial condition, the ISN content of all intracellular nitrogen pools is assumed
to be zero, and the labeling of ammonium, i.e., ISNH 4 , is set to the constant nitrogen-
15 substrate enrichment. Substituting this in the above set of equations, we obtain
for the first point on the time courses (i.e., the first measuring point):
[Glu] x ISNaGlu ~ FGDH x ISNH 4 · ~t = Ml

[GIn] x lsNoGln~FGs x ISNH 4 x M=M 2,

where Ml and M2 denote the measured ISN contents (in mmoll- l ) after a short time
~t of incubation with the labeled ammonium.
Thus, provided the time resolution is fast enough and the ISN quantitation is suf-
ficiently accurate, the ratio MdM2 of the relative rates of ISN accumulation in the
amino nitrogen of glutamate and the amido nitrogen of glutamine, respectively, di-
rectly reflects the ratio of the GDH and GS fluxes.
For the second iteration step, the values for ISNaGlu, ISNaGln, and ISNoGln derived
from Ml and M2 at the first time point are substituted in the differential equations
for the Na and No pools of glutamine. This, after some rearrangement, yields for the
ratio M3/M4 of the increments of the glutamine No and Na signals in the second time
interval:
M3/M4 = [GIU]ISNH 4 /M I - ([Glu]/[Gln]) x (M2/Md x (FGOGAT + BGln)/FGS ,
showing that now FGOGAT is also reflected in the time courses of ISN incorporation.
The fluxes FGDH , FGs , and FGOGAT are then unequivocally determined from the mea-
sured ISN incorporation time courses together with the measured specific ammo-
nium uptake rate and values for BGlu and BGln inferred from Table 16.1.

16.2.2.6
Influence of Reaction Reversibility

The presence of reversible reaction steps may have a pronounced effect on isotopic
labeling depending on conditions. This on the one hand complicates the analysis
since the simplified pictures drawn up above are no longer valid. On the other hand,
it offers the possibility of identification of such reversible reactions, which may have
significant informative value for metabolic engineering purposes. As an illustrative
example the analysis of the pentose phosphate pathway will be redone, now includ-
ing a variable reversibility of the phosphoglucose isomerase reaction. Figure 16.8
again applies, however now with the additional bi-directional flux fix interconvert-
526 16 Metabolic Flux Analysis of Corynebacterium glutamicum

ing glucose 6-phosphate and fructose 6-phosphate included. For glucose 6-phos-
phate we then have
(16.17)

(16.18)
i.e., the label on C-6 of glucose 6-phosphate is no longer o.
Since the pentose phosphates are only synthesized from glucose 6-phosphate, they
will only be labeled in C-5. Sedoheptulose 7-phosphate is only synthesized from the
pentose phosphates and will only be labeled in C-7. Erythrose 4-phosphate, which is
only synthesized from Sedoheptulose 7-phosphate, will be labeled only in C-4. We
then have for fructose 6-phosphate (Fig. 16.8):
F6P_1 = (fl + fIX) x G6P_1I(fi + fIX + 2f3) (16.19)

F6P_6 = {(fl + fIx) x G6P _6 + f3 x GAP_3 + f3 x Ery4P_4)}/(fl + fIX + 2f3) (16.20)

Substitution of Eq. (16.17) in Eq. (16.19) yields
F6P_1 = fo x Lo X (fl + fIX)/{(fi + fIX + f2) X (fl + fIX + 2f3) - (fl + fIX) x fIX} (16.21)
Since Ery4P_4=Sed7P_7=Rib5P _5=G6P _6 we can rearrange Eq. (16.20) after sub-
stitution of Eq. (16.18) into
(16.22)
where
Q = f3 X (fl + fIx + f2)/{(fl + fIX + 2f3) X (fl + fIX + f2) - (fl + fIX) X fIX - £3 x fIX}
For glyceraldehyde-3-phosphate we again have
GAP_3 = {2f3 x Xul5P _5 + (fl + 2f3) x (F16BP _1 + F16BP_6)}1(2fI + 6f3)
= {2f3 x G6P _6 + (fl + 2f3) x (F6P_1 + F6P _6) }/(2fI + 6f3)
Substitution of Eqs. (16.18), (16.21), and (16.22) after rearranging of terms yields
(16.23)
with

S = {2f3 X fIX + (fl + 2f3) X (fl + fIX + f2) }/(fl + fIX + f2)
and Q as defined in Eq. (16.22). As the final step, f3=f2/3 is to be substituted in these
expressions. It can be verified that Eq. (16.23) transforms to Eq. (16.12) in the case
when fIX=O.
Figure 16.13 shows a set of curves of GAP_3 as a function of f2 for various values
of fIX' obtained using Eq. (16.23). Obviously, the degree of reversibility of the phos-
phoglucose isomerase reaction strongly influences the estimation of the pentose
phosphate pathway activity from the J3C label in glyceraldehyde 3-phosphate, or
pyruvate.
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 527

16.2.2.7
/sotopomers

The fact that the degree of reversibility of enzyme reactions may be quantitated from
its effects on the 13C label distribution adversely implies that the number of un-
knowns in the flux determination procedure increases considerably. Therefore, it is
of crucial importance to increase the information content of the labeling experiment
as much as possible. Positional isotopic enrichments clearly are not the optimal
choice, since a molecule containing N carbon atoms in this case can provide at most
N independent pieces of information. In contrast, a molecule with N carbon atoms
can exist as 2N different isotope isomers, or isotopomers, since each carbon may be
present as 12C or as 13e. Thus, a small molecule like pyruvate can provide up to 8
independent pieces of information from its 13C isotopomer distribution. This num-
ber may be extended even further by combining 13C labeling with heteronuclear
stable isotope labeling, e.g., IS N, 2H , or 17 0. Modern, refined measurement techni-
ques such as two-dimensional NMR [66,67], heteronuclear spin-echo NMR [68,69],
or mass spectrometry [70, 71], have helped isotopomer analysis to evolve into an
extremely powerful tool for metabolic flux analysis. As a qualitative example, con-
sider again the labeling scheme in Fig. 16.11. It may be helpful to think of isotopomer
analysis as a method to measure the relative abundance of intact carbon backbone
fragments (Fig. 16.14). It is now assumed that pyruvate exists in two isotope isomeric
forms: [12 C3]pyruvate (90%) and [13 C3]pyruvate (10%). [13 C3]Pyruvate entering the
citric acid cycle via pyruvate dehydrogenase is decarboxylated, leaving an intact
13C2-fragment that is subsequently transferred to oxaloacetate via one turn of the
citric acid cycle. It has a probability of only 10% to form a C-C bond with another
13C-Iabeled carbon. This probability can easily be corrected for [66] and is not con-
sidered here. In contrast, [13 C3]pyruvate entering the citric acid cycle via pyruvate
carboxylase will appear as an intact 13C3 fragment in oxaloacetate. Thus, to a first
order approximation, the ratio of the pyruvate carboxylase flux to the total of ana-
plerotic and citric acid cycle fluxes into the oxaloacetate pool is simply proportional
to the relative isotopomeric abundance of intact 1,2,3-13C3 fragments within the ox-

;--------------.------ ... ---- 2_ 13C [IB]J

:............:--

---------- 2,3- 13C2

---------- 1,2-13C2
-
[IBII
11B0

52_0 51.5 51.0 50.5 ppm

Fig. 16.14. Part of the one-dimensional 100 MHz I3C NMR spectrum of alanine isolated from
hydrolyzed protein of Corynebacterium glutamicum grown on I3C-Iabeled glucose. The reso-
nance of C-2 is shown, with different multiplet contributions from the various isotopomer
species as indicated
528 16 Metabolic Flux Analysis of Corynebacterium glutamicum

aloacetate pool (or within an oxaloacetate-derived amino acid such as aspartate).

The proof of the possible presence of back reactions of oxaloacetate to the sym-
metric intermediate fumarate also seems almost trivial, since this would give rise to
unique intact 2,3,4_13C3 fragments within the oxaloacetate pool. The presence of the
respective l3C3 fragments can unequivocally be inferred from multiplet structures in
the l3C NMR spectrum (Fig. 16.14).
These rather qualitative arguments already clearly illustrate the potential of isoto-
pomer measurements for metabolic flux analysis. The mathematical analysis, how-
ever, is rather involved.

16.2.2.8
Sources of Isotopic Measurement Data

Collecting the biogenic material from which the labeling data are to be measured
presents an important experimental problem in isotopic labeling studies. The sim-
plest approach is to use the culture supernatant and to analyze the labeling patterns
in metabolites excreted by the cells. However, under optimal conditions for exponen-
tial growth, aerobic C. glutamicum may produce only carbon dioxide, from which (at
least at isotopic steady state) little or no information can be derived. Or, under opti-
mal production conditions, only a single or at best a few metabolites are excreted
that can be used to obtain labeling data from. Moreover, the labeling of excreted
metabolites may be diluted by the presence of unlabeled material accumulated before
addition of the labeled substrate, or resulting from the metabolization of unlabelled
storage compounds. Thus, analysis solely of excreted compounds generally results in
a very small isotopic labeling data set with a poor and possibly biased information
content. A better, but more elaborate, procedure is to perform a cytoplasmic extrac-
tion of the cells. This gives access to the labeling of a larger number of compounds
with high information content, such as, e.g., alanine, aspartate, and glutamate. These
amino acids are derived from their precursors in a single enzymatic step. At isotopic
steady state, the labeling state of the carbon backbones of these amino acids there-
fore is identical to that of their respective precursors pyruvate, oxaloacetate, and
2-oxoglutarate. When using cytoplasmic extracts, one has to take care that the meta-
bolism is quenched quickly enough in order to prevent significant alterations of the
metabolic pools and/or their labeling state during the extraction procedure. Also, it
must be assured that a true isotopic steady state of all cytoplasmic metabolites has
been established. This may take several hours in C. glutamicum due to the presence
especially of a very high cytoplasmic glutamate pool (250 mmoll- 1 in the wildtype
ATCC13032 [50]). Since, however, a very large number of different metabolites with
very different concentrations are present in the cytoplasmic extract, the determina-
tion of l3C fractional enrichments by NMR often requires chemical isolation and
purification of compounds, since otherwise the NMR spectra are largely unsuited
for quantitative analysis. In practice, the use of cytoplasmic extracts for labeling
determinations is also rather limited because many important metabolites are pre-
sent in too Iowa concentration to be accessible by NMR. The analysis typically re-
quires 1 /.lmol of the labeled compound. This implicates for a compound with a cy-
toplasmic concentration of x mmoll- 1 that a cytoplasmic volume of 1/x ml must be
extracted with 100% efficiency, i.e., at least approximately 0.5/x g dry weight of cells.
Because of the limited availability of labeled biogenic material even in cytoplasmic
extracts, current most efficient methods [66,56] use the macromolecular cell consti-
16.2 Fundamentals of Intracellular Metabolic Flux Analysis 529

tuents to retrieve the labeling information, since these function as storage devices
for precursor metabolites. Hydrolysis of the cellular protein for example gives access
to the complete range of amino acids in which the carbon backbones of the respec-
tive precursor metabolites (Table 16.1) are preserved in a defined and well-known
way. Thus, erythrose 4-phosphate, ribose 5-phosphate, 3-phosphoglycerate, PEP,
pyruvate, acetyl-CoA, oxaloacetate, and 2-oxoglutarate can all be accessed using this
procedure. Glyceraldehyde-3-phosphate, 3-phosphoglycerate, and acetyl-CoA can be
accessed via the cell lipids, and the cellular RNA constitutes an additional source of
especially ribose 5-phosphate labeling data. An added advantage of using the macro-
molecular components is that very large amounts of precursor metabolites are stored
in them; i.e., hydrolysis of only 10 mg of dry biomass yields sufficient material of a
precursor with a very modest content of 100 p,mollg dry weight for NMR analysis.
Sole requirement is that the macromolecular constituents have been synthesized en-
tirely from precursors that were in an isotopic steady state, and that this state did not
vary significantly during the incubation period. These conditions are best satisfied
when using cells incubated with the labeled substrate for at least three dilution times
in chemostat culture, cells exponentially grown in very thinly-inoculated batch cul-
tures, or excreted proteins.

16.2.2.9
A Comprehensive Modeling Framework

The metabolite balancing approach essentially amounts to the solution of a set of

linear equations (typically 20-30). It is therefore readily appreciated that this can
be done most conveniently by standard matrix calculation procedures [54, 72]. The
isotopic labeling experiment for which only positional enrichments at steady state
are considered can be regarded as a metabolite balancing procedure for single car-
bon atoms (cf. the examples given above). As such, it also gives rise to a set of linear
equations (typically 70-120). Again in this case, one would expect that the most
efficient approach to handle these equation systems would be the application of ma-
trix calculation methods. This, however, was not realized until fairly recently, when
the first procedures involving so-called atom mapping matrices appeared in the lit-
erature [73]. While these formulations did not immediately reveal the analogy with
the metabolite balancing solution approaches, this was clearly the case in the simul-
taneously developed formalism of Wiechert and co-workers [56, 74, 75]. This proce-
dure enables a fully integrated analysis of mixed data resulting from measurements
of fluxes as well as positional l3C enrichments, and includes efficient numerical pro-
cedures for the estimation of statistical variances of the calculated fluxes. Other ap-
proaches also developed. A recent overview that describes analytical methods for
l3C-Iabeled metabolites, 13C labeling strategies, metabolite balancing approaches, as
well as their synergetic integration can be found in [76].
The modeling and analysis of isotopomer data is much more complicated since
bimolecular reactions give rise to bilinear terms in the isotopomer balance equations
[77]. Only in special cases will these equations transform to linear equations, namely
when no bimolecular reactions are present in the network, or when for each bimo-
lecular reaction the labeling state of one of the inputs is given. This for instance
applies to a model that involves the citric acid cycle, the glyoxylate cycle and an
anaplerotic carboxylation reaction in the case that the labeling state of the common
substrate of all bimolecular reactions involved, i.e., pyruvate, is given. The reason is,
530 16 Metabolic Flux Analysis of Corynebacterium glutamicum

that this isotopomer distribution can be considered as a constant which results in

linear equations relating the remaining unknown isotopomers and fluxes. This sub-
sequently allows one to formulate the appropriate convenient analytical expressions
[78, 79). It has long been thought that only iterative numerical strategies allow one to
calculate the isotopomer distributions in the general case. Several generalized mod-
eling approaches involving so-called atom mapping and isotopomer mapping ma-
trices [80, 81) were devised in order to facilitate the modeling. However, Wiechert
and co-workers [82, 83) very recently demonstrated that the isotopomer balance
equations can indeed also be solved for the general case using standard matrix ap-
proaches and following a recursive procedure after suitable coordinate transforma-
tion. This enabled to construct a completely general isotopomer modeling frame-
work where only the basic carbon atom transitions in the various reaction steps have
to be specified by the user. The strongly enhanced computational efficiency allows
one to carry out optimal design studies of labeling experiments by searching for
minimal simulated standard deviations of calculated fluxes resulting from the use
of different labeled substrates in silico.

16.3
Metabolite Balancing Studies

16.3.1
Overview

After their key article [52) introducing the concepts of flux balancing and network
rigidity in relation with metabolic engineering, Vallino and Stephanopoulos wrote
an important paper [59) about metabolic flux distributions in C. glutamicum during
growth and lysine overproduction. Together with their first paper on the subject
[54), these publications have been highly instrumental in the establishment of meta-
bolic flux analysis as the invaluable tool for metabolic engineering purposes it is has
become today. In order to apply a network rigidity analysis as proposed in [52),
these authors applied several metabolic perturbations that enabled one to character-
ize the flexibility of the pyruvate and glucose-6-phosphate branch points in lysine-
producing C. glutamicum ATCC 21253 [61,84).
A series of flux analyses at different growth rates in lactate-limited as well as in
glucose-limited chemostat cultures of C. glutamicum ATCC 17965 were reported by
Cocaign-Bousquet and co-workers [55, 85). These authors used reported and mea-
sured enzyme activities to decide which pathways are active in vivo, so as to yield a
determined system of equations for metabolic balancing. This group also reported
metabolite balancing studies performed in order to characterize the metabolic
changes associated with a transient period of oxygen-limited growth of batch cul-
tures of C. glutamicum ATCC 17965 [36). However, the complexity of the observed
changes together with the enormous variation of the calculated carbon flux distribu-
tion within the central metabolic pathways does not permit a clear analysis of the
results.
In the following sections, an illustrative selection of the most important results of
the reported metabolite balancing studies of C. glutamicum are given and discussed.
16.3 Metabolite Balancing Studies 531

16.3.2
Comparison of Fluxes During Growth and Lysine Production

Vallino and Stephanopoulos [59] studied a control lysine fermentation of C. glutami-

cum ATCC 21253 by analyzing the culture broth at several time points during the
initial growth phase and during several following different lysine production phases,
while applying the metabolite balancing approach. This strain produces lysine under
threonine limitation. Two illustrative flux distributions are shown in Fig. 16.15. Dur-
ing the growth phase (Fig. 16.15a), the main products from glucose are biomass,
carbon dioxide, and trehalose, while only trace amounts of lactate, acetate, and ly-
sine are observed. For the analysis it was assumed that the glyoxylate pathway is
inactive, that no other NADPH-regenerating enzymes than isocitrate dehydrogenase,
glucose-6-phosphate dehydrogenase, and 6-phosphogluconate dehydrogenase are ac-
tive, and that PEP carboxylase is the only active anaplerotic carboxylation enzyme.
The two different lysine biosynthetic pathways (Fig. 16.4) were lumped. This strategy
results in a fully determined equation system as explained above in Sect. 16.2.l.3.
Since measurements of the oxygen and ammonium uptake rates as well as the carbon
dioxide evolution rate were also included, the network was in fact overdetermined.
This analysis indicates a significant activity of the oxidative pentose phosphate path-

Trehalose Trehalose

t®
~
G.= ' \ :$S\--
@C:J' ~t®®y,
Glucose G6\P~se_P ...

@\~.
GAP Sed7P

~ @ ~4P'
F6P @ F6P @

~P/PGA'
.~
. GAP/PGA· .
Alanine
~
Alanine
C@+
Valine PEP Valine PEP' ...............•
Lactate ~@
~:=:==:t=::::=:;:Pyr .....
Lactate
=-=~:j::::::~Pyr
~@
CO, ®l ~® Acetate co, @ I@ Acetate
AcCoA4iJ..
'" 4JJ
OAA~ OM~'
@
.
L-Iysme
~~.~)t~~uta.
-(
Mal
@
0 mate
j
®
L-Iysine
~~.~)'t~;luta_
-(
Mal
@ mate
j
\. . AKG....-
~ :KG....-
a "'" Sue 59 b Sue +

Fig. 16.15a,b. Illustrative flux distributions in Corynebacterium glutamicum ATCC 21253 dur-
ing: a the growth; b early lysine production phase, as determined by metabolite balancing.
Data were taken from [59]. Fluxes are normalized by the glucose uptake rate. Dotted arrows
indicate precursor fluxes for biomass synthesis. While these fluxes were not explicitly stated in
[59],60% of total carbon was used for biomass synthesis in A and 15% in B
532 16 Metabolic Flux Analysis of Corynebacterium glutamicum

way (PPP) already under conditions of balanced growth: 25% of glucose 6-phosphate
are metabolized over the PPP.
As soon as threonine in the medium was depleted, lysine production started at a
high rate while growth continued but gradually slowed down. The metabolic flux
distribution determined towards the end of this phase is given in Fig. 16.15b. The
main products are now biomass, lysine, and carbon dioxide. Trehalose synthesis is
increased and there is still a minor excretion of acetate and lactate. Lysine is pro-
duced at 29% molar yield. A number of very characteristic changes in the flux dis-
tribution are apparent when comparing Fig. 16.15a and 16.15b: an almost twofold
increase of the activity of the PPP and the anaplarotic carboxylation and a signifi-
cant increase of the glutamate dehydrogenase activity. The citric acid cycle rate is
little affected [59].

16.3.3
The Search for Yield-Limiting Flux Control Architectures

16.3.3.1
The Pyruvate Branch Point

Two experiments to assess the control structure around the pyruvate branch point
were carried out [61]. In the first experiment, a permanent potential increase of the
pyruvate availability for lysine synthesis was created by selecting a mutant with 98%
attenuated pyruvate dehydrogenase activity. The metabolic flux distribution map ob-
tained from this experiment is shown in Fig. 16.16. Since it very closely resembles
the flux distribution obtained with the control fermentation (Fig. 16.15b), it was con-
cluded that the availability of pyruvate is not a limiting factor for lysine synthesis.
However, the attenuation of the pyruvate dehydrogenase activity resulted in an over-
all flux attenuation in the network of 65-75% while metabolite excretion patterns
were unchanged. This was considered evidence for the presence of a rigid branch
point somewhere else in the metabolic network [61].
In the second experiment, an instantaneous increase of the pyruvate availability
for lysine synthesis was created by the addition of fluoropyruvate at the onset of
lysine production. Fluoropyruvate is a strong inhibitor of the pyruvate dehydrogen-
ase complex. The flux distribution resulting from the analysis reveals a pattern that
is largely identical to the one obtained with the control fermentation at the same
production phase, with the exception that approximately 40% of the pyruvate pre-
viously metabolized in the citric acid cycle is diverted from that pathway and ex-
creted [61]. The fact that the lysine production is virtually unaffected again demon-
strates that the lysine yield is not limited by pyruvate availability.

16.3.3.2
The G/ucose-6-Phosphate Branch Point

Two experiments to assess the control structure around the glucose-6-phosphate

branch point were carried out [84]. In the first experiment, a permanent potential
increase of the glucose-6-phosphate availability for metabolism over the oxidative
PPP was created by selecting a mutant with 90% attenuated glucose-6-phosphate
isomerase activity. The results of the analysis indicated an attenuation of fluxes
throughout the metabolic network, although without a significant alteration of flux
16.3 Metabolite Balancing Studies 533

Trehalose

@ t@ @~l'
~ __ G6~P~e-P
GAP Sed7P

, ® ~P.
F6P @
.~
GAP/PGA ......... .
Alanine
~ p~@ ....... .
Valine

Lactate ...:::::====t==~J.. ... ..

CO, ® f§ff !® Acetate

L-Iysine Mal
® mate
J
\. - AKG~

"'" Sue 78' ~

Fig. 16.16. Flux distribution in Corynebacterium glutamicum FPS009 during the early lysine
production phase, as determined by metabolite balancing. In this strain, pyruvate dehydrogen-
ase activity was only 2% of the activity present in the parent strain ATCC 21253. Data were
taken from [61]. Fluxes are normalized by the glucose uptake rate. Dotted arrows indicate
precursor fluxes for biomass synthesis. While these fluxes were not explicitly stated in [61],
20.5% of total carbon was used for biomass synthesis. Compare with Fig. 16.15b

partitioning. This again indicated the presence of a rigid branchpoint elsewhere in

the metabolic network [84]. In the second experiment, the glucose-6-phosphate
branch point was essentially bypassed by using gluconate as the primary carbon
source. Gluconate enters the PPP after phosphorylation to 6-phosphogluconate. The
backreaction of 6-phosphogluconate to glucose 6-phosphate must be considered ki-
netically unfeasible for thermodynamic reasons. However, it appeared that without
the inclusion of an additional NADPH-oxidizing reaction in the network this con-
straint could not be fulfilled due to an overproduction of NADPH resulting from the
forced metabolization of 100% of the substrate over the oxidative PPP. The flux dis-
tribution resulting after this network modification is given in Fig. 16.17. The gluco-
nate consumption was almost as rapid as the glucose consumption in the control
fermentation and the flux distribution map other than the PPP was again very simi-
lar to the control fermentation (Fig. 16.1Sb) except that the catabolism of gluconate
produces a substantial excess of NADPH. This excess, however, does not result in an
increase in lysine yield, thus ruling out a limitation by NADPH availability.
534 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Trehalose
@ t@(jJ/J'
~ __ G6\P~se_p
GAP Sed7P
@)~
f.:!j4P
F6P ~

Alanine
~/PGA
c@>t
Valine PEP
Lactate ~@
!® ~
~~=J:::~Pyr
co, @ Acetate
AcCoA~

OAA~
® ,~?):~,..
L-Iysine Mal (~;te 0
\. - AKG..-

"--sue 91

Fig. 16.17. Flux distribution in Corynebacterium glutamicum NFG068 during the early lysine
production phase, as determined by metabolite balancing. In this strain, phosphoglucose iso-
merase activity was only 7% of the activity present in the parent strain ATCC 21253. Data were
taken from [84]. Fluxes are normalized by the glucose uptake rate. No biomass synthesis took
place; in fact, 1.1% of the carbon was derived from biomass hydrolysis [84]. Compare with
Fig. 16.16

16.3.4
Growth Rate-Dependent Modulation of the Central Metabolic Fluxes

16.3.4.1
Growth on Lactate

Figure 16.18 shows two representative flux distributions for growth on lactate at low
and high growth rates [55]. At the low growth yield, NADPH for anabolic purposes
is supplied by the pentose phosphate pathway (PPP) and the isocitrate dehydrogen-
ase reaction, while the anaplerotic flux was found to be distributed over PEPcarbox-
ylase and the glyoxylate pathway. At the high growth rate, the anaplerotic flux is
supplied by a pyruvate-carboxylating enzyme while a relatively large flux catalysed
by malic enzyme generates additional NADPH. Moreover, a significant overflow of
pyruvate was observed (Fig. 16.18). The rationale behind the seemingly arbitrary
choice of pathways acting in vivo is as follows. Enzymatic determinations in cell-free
extracts of Brevibacterium flavum, a close relative of C. glutamicum, had indicated
16.3 Metabolite Balancing Studies 535

@'
_\jjZ.e <9J
co,

GAP Sed7P
.. @.

"' 6.0 ~4P' .......@ .•

. F6P @
@
. ~
GAP/PGA .. @.
Pyruvate
QM)t .....@ ..•
Pi®
@.

·@r
!@
Pyr··· .
co, @
AcCoA . @..•

Mal ____ -"-"~

@.
a
~ Sue
Fig. 16.18a,b. Illustrative flux distributions in Corynebacterium glutamicum ATCC 17965 at di-
lution rates of: a 0.17 h- 1; b 0.28 h- 1 in chemostat cultures during growth on lactate, as deter-
mined by metabolite balancing. Data were taken from [55]. Fluxes are normalized by the lac-
tate uptake rate. Dotted arrows indicate precursor fluxes for biomass synthesis. Note the pro-
posed changes in the operation of the anaplerotic enzymes. At the high growth rate, malic
enzyme covers the increased anabolic demand for NADPH

that the key enzyme of the oxidative PPP, glucose-6-phosphate dehydrogenase, was
expressed approximately 7-20-fold lower during growth on organic acids (lactate,
acetate, glutamate) than during growth on glucose [86]. Therefore, the authors as-
sumed that at all growth rates, the oxidative PPP supplies precursors for ribose
5-phosphate and erythrose 4-phosphate only to the exact amounts that are needed
for biomass synthesis. This activity results in a basic supply of NADPH (i.e., 2 mol
per mol of glucose 6-phosphate metabolized over the oxidative PPP).
At the low growth rate, the additional NADPH required for biomass synthesis was
assumed to be provided only by the isocitrate dehydrogenase reaction, in the exact
amount needed. By this assumption, the flux through the citric acid cycle is fixed
and the carbon balance must be closed by adjustment of the fluxes over the follow-
ing two alternative routes with different overall stoichiometries: an anaplerotic car-
boxylation reaction (overall stoichiometry Pyr+COz--+OAA) and the glyoxylate cycle
(with overall stoichiometry 2 Pyr---+OAA+2 CO 2 ), This adjustment is completely de-
termined since, first, the total molar amount of oxaloacetate formed in these two
reactions must exactly equal the sum of the molar amounts of the precursor meta-
bolites oxaloacetate and 2-oxoglutarate required for biomass synthesis, and secondly,
the total amount of available pyruvate is fixed. The choice of PEP carboxylase as the
anaplerotic reaction is in fact arbitrary.
536 16 Metabolic Flux Analysis of Corynebacterium glutamicum

At the high growth rate (Fig. 16.18b), the demands for additional NADPH and
anaplerotic precursors are much larger than can be supplied by the isocitrate dehy-
drogenase and PEP carboxylase reactions, respectively. From their enzyme measure-
ments, which indicated a decrease of PEP carboxylase activity along with a strong
increase of malic enzyme and oxaloacetate decarboxylase activity when going to
higher growth rates [55], the authors assumed that PEPcarboxylase does not contri-
bute to the anaplerotic flux at the high growth rate. Instead, it was concluded that
malic enzyme provides the additional NADPH needed for biomass synthesis, and
that the measured oxaloacetate decarboxylase activity in reality represents a pyru-
vate carboxylase activity. The latter provides for the recycling of pyruvate formed by
malic enzyme and fulfills the anaplerotic precursor requirement. It was also assumed
that the glyoxylate cycle is inactive, since an increased pyruvate concentration had
been shown to repress the synthesis of the glyoxylate bypass enzymes.
The resulting pathway configuration then yielded the fully determined flux estima-
tion shown in Fig. 16.18b. The rather constant ratio of pyruvate dehydrogenase ac-
tivity vs the calculated flux through this enzyme, as well as the pyruvate overflow
observed at high growth rates, were taken as evidence that this enzyme was sub-
strate-saturated under all conditions [55].

16.3.4.2
Growth on Glucose

Figure 16.19 shows a representative flux distribution for growth on glucose at a very
high growth rate of 0.59h- 1 [85]. NADPH for anabolic purposes is supplied to the
exact amount required, but only by the pentose phosphate pathway (PPP) and the
isocitrate dehydrogenase reaction. Pyruvate carboxylase was assumed to be the only
active anaplerotic enzyme at this growth rate. With these choices, the flux estimation
is completely determined as explained above under in Sect. 16.2.1.3. A flux of almost
60% of available glucose 6-phosphate into the PPP was calculated. This study [85]
did not supply other explicit complete flux distributions, but these may easily be
calculated from the data given, except for the precise flux distributions over the al-
ternative anaplerotic enzymes PEP carboxylase, malic enzyme, and oxaloacetate dec-
arboxylase (thought to mask pyruvate carboxylase activity). These enzymes are
shown to be present in significant amounts, but with relative activities that vary
considerably with the growth rate. None of them may be considered inactive at any
growth rate and one can only speculate about the actual activities of the enzymes in
vivo. Here, clearly a limitation of flux analysis by the metabolic balancing method is
encountered.

16.3.S
Summary

The metabolite balancing studies referred to above have yielded valuable insights in
the interdependency of the metabolic fluxes in the central metabolism of c. glutami-
cum. While there may remain some doubt as to whether PEP carboxylase is really the
principal anaplerotic enzyme in C. glutamicum as assumed in [54, 59, 61], the ex-
perimental evidence gathered in the perturbation studies of the pyruvate branch
point [61] leaves little or no alternative than to support the conclusion that pyruvate
availability does not limit lysine yield. Likewise, irrespective of whether an addi-
16.3 Metabolite Balancing Studies 537

@ CO,

IGlucosel~G6~p
@j~""""",, . @ .
~ @ G~7P
Lt P .....@~

F6P @
@
..1.5 . ~GAP/PGA .
"@'N
~PEP . . . @.•

CO,--.Ifi: 7@ ...... .®.

~<@

i~
AcCoA· ..... ® ..
~® ..... OM~

~
.. ~G®'
~~
Fig. 16.19. Flux distribution in Corynebacterium glutamicum ATCC 17965 at a dilution rate of
0.59 h- 1 in chemos tat culture during growth on glucose, as determined by metabolite balan-
cing. Data were taken from [84]. Fluxes are normalized by the glucose uptake rate. Dotted
arrows indicate precursor fluxes for biomass synthesis. The 1.44 mol NADPH per mol glucose
required for biomass synthesis is supplied by the pentose phosphate pathway and the isocitrate
dehydrogenase reaction. Malic enzyme was concluded to be inactive [84]

tional NADPH oxidase activity is present in C. glutamicum or not, the experimental

evidence presented in [84] unequivocally proves that lysine yield is not limited by
NADPH supply. In contrast, the results of the analyses of the growth rate-dependent
modulation of the central metabolic fluxes in C. glutamicum leave considerably more
questions that ask for a definitive answer. Many choices regarding the operationality
of pathways in vivo made in these studies were based on enzyme measurements.
However, activities of the key enzymes of the oxidative pentose phosphate pathway
and the glyoxylate pathway were not determined, whence important information ne-
cessary to support the assumptions is missing. But even if this information would be
available, any conclusion about the relative activity of, e.g., anaplerotic pathways in
vivo based on enzyme measurements alone must be treated with great care. Thus,
additional data that give access to the true metabolic activities in vivo are required
before any definitive answers can be given to such exciting questions as to whether
an NADPH-generating malic enzyme-pyruvate carboxylase metabolic cycle exists in
C. glutamicum [55]. In the next section it will be demonstrated that such data can in
fact be provided by isotopic labeling experiments in combination with NMR or mass
spectrometry (MS).
538 16 Metabolic Flux Analysis of Corynebacterium glutamicum

16.4
Studies Based on Isotopic Labeling and NMR

16.4.1
Overview

Glutamate-excreting bacteria, which are today all known by the species name C.
glutamicum, were isolated starting in 1957 [87]. From relatively early on, isotopic
labeling studies of C. glutamicum and its close relatives Brevibacterium flavum and
B. ammoniagenes have been performed. While at first radiolabel methods were em-
ployed [88-90], studies using stable isotope labeling in combination with NMR and
MS started to emerge in the 1980s. They were initiated by a 15NNMR study of Bre-
vibacterium lactofermentum [91] and a l3C NMR study of the biosynthesis by Micro-
bacterium ammoniaphilum of glutamate selectively enriched with carbon-13 [92].
This study already used mathematical expressions relating l3 C NMR multiplet ratios
to activities of PEP carboxylase, the citric acid cycle, and the glyoxylate cycle, and as
such presents at once the first isotopomer study of a C. glutamicum-related bacter-
ium. Ishino et al. subsequently used l3C labeling and NMR to establish the presence
of the dual pathway for lysine synthesis (Fig. 16.4) in vivo [93] and to characterize a
histidine-fermenting C. glutamicum mutant [94]. First attempts to model and quan-
tify the principal fluxes of the central metabolism during lysine production by l3C
labeling were reported in [95] for C. glutamicum using l3C NMR and in [96] for B.
flavum using a combination of 1H NMR, 13C NMR, and MS. Ishino et al. [97] per-
formed a comparative study of the flux distribution over the Embden-Meyerhof
pathway and the pentose phosphate pathway during lysine and glutamate fermenta-
tion by c. glutamicum using l3C-Iabeled glucose and l3CNMR. Sonntag et al. [62]
presented a detailed l3C labeling study of the flux distribution over the diaminopi-
melate pathways in lysine biosynthesis of wildtype and lysine-producing C. glutami-
cum in dependence on culture conditions. Using a model adapted from [92, 96], the
principal fluxes in the central metabolism of C. glutamicum ATCC17965 during ex-
ponential growth on l3C-Iabeled glucose were estimated [98]. Marx et al. [56] pre-
sented a first fully integrated metabolite balancing/ l3 C labeling flux analysis of ly-
sine-producing C. glutamicum MH20-22B in chemostat culture with [1-l3C]glucose
as the sole carbon source. In order to obtain an isotopic enrichment data set as large
as possible, this method used a comprehensive set of amino acids isolated from the
l3C-Iabeled cell protein. This approach was subsequently refined and applied to ex-
ponentially growing, glutamate-producing, and lysine-producing C. glutamicum in
batch cultures [99], to exponentially growing, and glutamate-producing isogenic
strains derived from C. glutamicum MH20-22B in continuous cultures [100], and to
lysine-producing strains with different glutamate dehydrogenase cofactor depen-
dency derived from C. glutamicum MH20-22B in continuous cultures [101]. Using
different l3C-Iabeled substrates and modeling procedures, two independent studies
of the anaplerotic reactions in PEPcarboxylase-deficient C. glutamicum were re-
ported [70, 102]. Using a membrane cyclone bioreactor devised for in vivo NMR
spectroscopy with high microbial cell densities, intracellular fluxes in lysine-produ-
cing C. glutamicum were studied with the help of l3C-Iabeled glucose and in vivo
l3C NMR [103]. Finally, this same bioreactor system was employed to quantify the
principal ammonium-assimilating fluxes in C. glutamicum ATCC 13032 and a gluta-
16.4 Studies Based on Isotopic Labeling and NMR 539

mate dehydrogenase mutant with the help of lsN-labeled ammonium and in vivo
lsN NMR spectroscopy [50].
In the next sections, the most important results of the reported studies using
stable isotope labeling and NMR are presented and discussed.

16.4.2
The Dual Pathways of Lysine Biosynthesis

The dual pathways of lysine biosynthesis in C. glutamicum (see (Fig. 16.4) have re-
ceived considerable interest. At the tetrahydrodipicolinate branchpoint of the lysine
biosynthetic pathway of C. glutamicum, carbon can be shunted either through the
one-step meso-diaminopimelate (Dap) dehydrogenase pathway, or through the four-
step succinnylase pathway to form the cell wall component and immediate precursor
to lysine, meso-Dap [58, 93]. Studies using mutants either of the one-step dehydro-
genase pathway [58] or of the four-step succinylase pathway [104, 105] indicated that
each pathway is in principle dispensable for growth of C. glutamicum, as long as the
other remains unaffected. The dehydrogenase pathway was shown to be essential for
high lysine productivity [58], whereas the succinylase pathway was very recently
shown to be dispensable for lysine production [105]. However, an initial analysis of
lysine-producing C. glutamicum ATCC 21543 using l3C stable isotope labeling had
shown that the flux over the dehydrogenase pathway was approximately only half as
large as that over the succinylase pathway [95]. Consequently, the correlation of the
flux distribution with lysine productivity and culture parameters has been investi-
gated in more detail [62].

16.4.2.1
Correlation with Lysine Production

In the specialized study [62], four C. glutamicum strains were compared: ATCC
13032 (wildtype), DG 52-5 (a moderate lysine producer), Asn (a mutant of strain
DG 52-5 missing the one-step pathway), and MH20-22B (a good lysine producer).
The cells were incubated in batch cultures with [6-l3C]glucose, and harvested at the
end of the lysine production phase. Lysine and alanine were isolated from the culture
supernatant, or from the cytoplasm in case of the non-excreting wildtype strain. It
was checked that the l3C labeling of pyruvate-derived alanine could be used to re-
present the labeling of pyruvate. The l3C fractional enrichments in C-2 of alanine
and C-2 as well as C-6 of lysine were determined by proton NMR spectroscopy of
the isolated amino acids. The model using simple carbon atom balances presented
above (Eq.16.12A) was then used to analyze the l3C enrichments. Table 16.2 shows
the results. The experiment with strain Asn served to check the model validity, i.e.,
that the lysine fractional enrichments of C-2 and C-6 as well as of C-3 and C-5 were
mutually identical (within experimental error) upon the sole operation of the succi-
nylase pathway.
The results clearly showed that, contrary to expectations, the lysine productivity
did not correlate with the participation of the dehydrogenase pathway.
540 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Table 16.2. Fractional 13C enrichments P2 of alanine C-2, and L2 and L6 oflysine C-2 and C-6,
respectively, as well as calculated contribution f of the one-step pathway to the total lysine
synthesized with different strains of C. glutamicum. Data were taken from [62]. nd=not deter-
mined

Strain Accumulated L-lysine P2 (%) L2 (%) L6 (%) f (%)

(mmol rl)

ATCC 13032 0.0 4.2 lO.3 8.0 24±2

ASn 19.7 nd 7.6 7.9 0
DG 52-5 29.2 1.7 5.4 3.6 32±7
MH 20-22B 2lOa 2.7 7.2 4.9 33±3
aMH20-22B was incubated with 525 mmol rl glucose, all other strains with 200 mmol rl glucose

16.4.2.2
Correlation with Culture Parameters

Sonntag et al. [62] subsequently studied the variation of the relative contributions of
the dehydrogenase and succinylase pathways to lysine synthesis during a batch fer-
mentation of C. glutamicum MH20-22B and found that the instantaneous flux dis-
tribution over the dehydrogenase pathway changed from an initial 72% to a final 0%
towards the end of the fermentation, resulting in an overall contribution to the total
synthesized lysine of 33% as found before (Table 16.2). Interestingly, in strain DG
52-5 cultivated with glutamate as the sole nitrogen source, the dehydrogenase path-
way did not contribute at all to lysine synthesis. Analysis oflysine produced in short-
term fermentations with strain MH20-22B upon incubation with different ammo-
nium concentrations (Table 16.3) then unequivocally revealed that the flux over the
dehydrogenase pathway was controlled by the ammonium concentration. Since no
regulatory phenomena for the involved enzymes had been found [106, 107], it was
concluded that this is a purely kinetic effect of the ammonium availability, and that
the rather low percental participation of the one-step pathway reflects the low affi-
nity of meso-diaminopimelate dehydrogenase for NH4+ (typical KM 36mmoll- 1 )
[108].

Table 16.3. Relative contribution f of the dehydrogenase pathway to lysine synthesis by C. glu-
tamicum MH20-22B at different ammonium concentrations in the culture medium. Data were
taken from [62]

Ammonium concentration Lysine produced f (%)

(mmoll- 1 ) (mmoll- 1 )

23 1.9 0
38 3.7 6±3
150 5.6 24±4
600 4.3 51 ±3
16.4 Studies Based on Isotopic Labeling and NMR 541

16.4.3
Distinct Metabolic Modes: Growth, Glutamate Production, and Lysine Production

Several isotope labeling studies with C. glutamicum or related strains had indicated
that the contribution of the pentose phosphate pathway is high (44-69%) during
lysine synthesis [95, 97], rather low (11-40%) during glutamate synthesis [88-90,
92,97], and intermediate (45%) during growth [98]. Since the reported values vary
considerably, and since the employed modeling approaches were in a need of im-
provement as seen in the light of the appearing new metabolite- and carbon balan-
cing methods, it was decided to perform a series of studies employing the newest
techniques of integrated metabolite balancing/ l3 C isotope labeling [56, 74, 75, 77] to
characterize the metabolic flux distributions in C. glutamicum during growth, gluta-
mate production, and lysine production [56, 99, 100].

16.4.3.1
Comparing Isogenic Strains in Continuous Cultures

Metabolic flux analysis in continuous culture offers the advantage that the organism
is cultivated in a reproducible stationary state, and that the establishment of closed
elementary balances such as is necessary for metabolic balancing is a relatively
straightforward standard procedure. Moreover, once the cells have been incubated
for at least three dilution times with a labeled substrate, the macromolecular cell
components can be isolated and hydrolyzed to yield an abundance of amino acids
or nucleotides for l3C isotope labeling analysis by NMR. Thus, metabolic flux analy-
sis was performed with two isogenic strains: lysine-producing C. glutamicum
MH20-22B, leuCD lysCSer381Phe [107] featuring a feed-back resistant aspartate ki-
nase [56], and C. glutamicum LE4, leuCD lysC+ featuring a feedback-sensitive aspar-
tate kinase [100]. The latter strain was grown both in glucose-limited chemostat and
in combined glucose- and biotin-limited chemostat cultures. Biotin limitation is one
of several possibilities to trigger glutamate excretion in any C. glutamicum strain
[109].
The carbon balances of the cultures are shown in Table 16.4. During growth, 45%
of the carbon is converted into biomass, and 52% into carbon dioxide. Glutamate
production results in a significantly increased CO 2 yield, and an almost twofold re-
duction in biomass yield. Growth-associated lysine production results in a reduction
of both CO 2 and biomass yields, but not as severe (Table 16.4).
After harvesting the cells, separate hydrolyses and chemical preparations of the
cellular protein and RNA fractions, respectively, were carried out. Amino acids and
nucleotides were finally separated by cation exchange chromatography in order to
obtain fractions of the single compounds [56, 62, 100] that were sufficiently pure to
allow for highly accurate quantitative NMR analysis. A representative set of l3C la-
beling data obtained with non-excreting strain LE4 is reproduced in Table 16.5.
These enrichments were determined by lHNMR, exploiting the carbon-13 satellite
signals in lH NMR spectra [56, 62, 68, 100].
Flux estimates for the three cultures, obtained by non-linear least squares fitting to
the flux and labeling measurement data [56, 74, 75, 77], are summarized in
Fig. 16.20. In the flux model, the PEP and pyruvate pools were merged since no dif-
ference in labeling between these pools could be detected. The results first of all
542 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Table 16.4. Carbon balances of C. glutamicum MH20-22B and C. glutamicum LE4 in glucose-
limited chemostat cultures (Data taken from [100])

Strain/conditions
LE4 LE4 Biotin limitation MH20-22B
Carbona Non-excretion L-Glutamate-production L-Lysine-production

Glucose 100 100 100

Biomass 45 26 39
Product 3 IS 19
CO 2 52 59 43
Sum 100 103 101
Y Xl5 (gxxgs- I ) 0.46 0.27 0.3S
Y P/S (molpxmol 5- 1 ) <0.01 0.22 0.19

Yx/s (Y p / s) represent biomass (X) and product (P) yield on glucose (S), respectively
aExpressed as percentage of the consumed glucose carbon

Table 16.5. Representative data set of measured l3 C enrichments in precursor metabolites from
non-excreting chemostat-grown C. glutamicum LE4. The data were obtained from proteino-
genic amino acids and RNA-derived nucleotides as follows: Rib5p (ribose 5-phosphate) from
guanosine, Ery4p (erythrose 4-phosphate) from phenylalanine, GAP (glyceraldehyde-3-phos-
phate and 3-phosphoglycerate) from glycine and serine, Pyr (pyruvate) from alanine and va-
line, AKG (2-oxoglutarate) from glutamate, and OAA (oxaloacetate) from aspartate and threo-
nine. The values in italics are the corresponding enrichments as predicted by the solution of
the mathematical model for flux analysis (Data taken from [100]). nd=not determined

l3 C enrichment (%) at carbon

Metabolite C-1 C-2 C-3 C-4 C-5

Rib5p 19.5 ±O.S 2.0 ± 1.0 2.0 ± 1.0 2.0 ± 1.0 19.7 ±0.6
Rib5p 19.3 1.9 3.1 2.3 19.5
Ery4p 1.5 ± 1.0 1.5 ± 1.0 1.5 ± 1.0 lS.7±2.0
Ery4p 2.6 3.0 2.2 lS.l
GAP nd 3.6±0.5 2S.4± 0.5
GAP 4.0 2.S 27.7
Pyr 3.9 ± 1.0 4.4±0.5 27.0±0.5
Pyr 5.0 4.6 26.S
AKG 14.1 ± 1.0 22.6±0.2 12.7 ±0.2 26.5 ± 0.2 3.0±2.0
AKG 14.5 22.5 l2.5 26.S 4.6
OAA 9.6±0.5 12.1 ± 0.3 22.5 ± O.S 14.1 ± 0.7
OAA 9.4 12.5 22.5 14.5

illustrate that a very detailed flux quantification is possible using the integrated me-
tabolite balancing/ l3 C labeling approach. In addition to the net fluxes over glycoly-
sis, the pentose phosphate pathway, the anaplerotic carboxylation reactions, the
glyoxylate pathway, the citric acid cycle, and the reactions withdrawing glucose
6-phosphate, fructose 6-phosphate, ribose 5-phosphate, erythrose 4-phosphate, gly-
ceraldehyde-3-phosphate, 3-phosphoglycerate, PEP, pyruvate, acetyl-CoA, oxaloace-
tate, and 2-oxoglutarate shown in Fig. 16.20, also bidirectional fluxes over phospho-
glucose isomerase, transketolase, trans aldolase, phospho ribose isomerase, ribulose-
16.4 Studies Based on Isotopic Labeling and NMR 543

tr @..!~.e-p
G6\P
l co,
... @,
;r~
G6~P
@.J~ose-P
co,
@.,
;r
G6~P
l

@)~tose-P@'
co,

@ GAP Sed7P @ GAP Sed7P @ GAP Sed7P

@ 1!!rp . . @ .., @ ~4P ...... @ .., @ ~4P@'

.F6P CD F6P CD F6P ®
® @ .. , @ ~IPGA ®. ~ .. @,
@!
);P/PGA" .. );P/PGA

@t ~

lF
co,17 46
PEP/Pyr
'-~oA35..·· . .
""" ...®..
®
., co
1fPEr/Pyr.

'
96 ~@
AcCoA ....
(iV,

@"
co, 31 68
PEP/Py
l~· .,
®

AcCoA ..... @.,

- ( -® ® -(~~" -(~®.
Mal Mal .~) 'iO' L-lysine Mal 1

'( ~~G " V . '"

a Sue b Sue C Sue

Fig. 16.20a-c. Flux distributions in Corynebacterium glutamicum during: a glutamate produc-

tion with strain LE4 under biotin limitation (fL=0.05 h- I ); b growth with strain LE4 (fL=O.1 h-
I); c I-lysine production with strain MH20-22B (fL=O.1 h- I ) in glucose-fed chemostat cultures,
as determined by integrated 13e NMR/metabolite balancing analysis. Data were taken from [56,
100]. Note the presence of simultaneous forward and reverse fluxes in the PEP/pyruvate and
oxaloacetate-interconverting pathways, with significantly different values in the three cases. No
NADPH constraints were applied in the analyses

5-phosphate epimerase, as well as bidirectional fluxes interconverting glyceralde-

hyde-3-phosphate and PEP/pyruvate, oxaloacetate/malate, and PEP/pyruvate, and
oxaloacetate/malate and fumarate could be quantified from the data. Typical values
for the degree of reversibility (defined here as the ratio of the backward and forward
fluxes) of various reactions as observed in the experiments [56, 100, 110] are given
in Table 16.6. There is obviously considerable variability in these estimates. This re-
flects, of course, the differences in metabolic states considered, but also the some-
times limited precision of the estimates. Estimates of the reversibility of the ana-
plerotic reactions are generally precise, indicating that the variability reflects differ-
ences during growth, lysine production, and glutamate production. However, espe-
cially the bidirectional fluxes in glycolysis show rather large 90% confidence inter-
vals [110], indicating that their estimate is imprecise.
The flux distributions (Fig. 16.20) reveal an impressive flexibility of C. glutamicum
central metabolic pathways even in isogenic strains. This is illustrated in more detail
for the glucose 6-phosphate and pyruvate branchpoints in Fig. 16.21. Most striking,
the contribution of the PPP to glucose 6-phosphate conversion ranges from 25%
during glutamate production to 66%, i.e., almost three times higher, during lysine
production (Fig. 16.21). While the total flux entering PEP/pyruvate varies only by
10% (Fig. 16.20), the relative flux over PEP/pyruvate anaplerotic carboxylating reac-
tions (PEPcarboxylase, pyruvate carboxylase, malic enzyme) ranges more than dou-
ble from 26% (glutamate production) to 62% (growth), that via pyruvate dehydro-
544 16 Metabolic Flux Analysis of Corynebacterium glutamicum

Table 16.6. Typical degree of reversibility (i.e., 100 x backward flux/forward flux) observed for
bidirectional reactions in C. glutamicum MH20-22B and LE4 [56, 100, 110]

Reaction(s) Enzyme/metabolites involved Degree of

reversibility (%)

empl phosphoglucose isomerase 10-90

ppp2 ribulose-5-phosphate epimerase >95
ppp3 phosphoribose isomerase >95
ppp4 transketolase 80-90
ppp5 transketolase 30-50
ppp6 trans aldolase 20-70
emp4-emp5 glyceraldehyde-3-phosphatePEP/pyruvate 10-50
anal, ana2, ana3 PEP Ipyruvateoxaloacetate/malate 40-75
tcc5-tcc6 succinate/fumarateoxaloacetate/malate 1-30

Glutamate Lysine
a b Growth c
production production

.~
* *
.. @ .G6P---@+ .@ •. @ ·G6P--@--.
7
63 33

Fig. 16.21a-c. Detailed flux partitionings at the glucose-6-phosphate and PEP/pyruvate branch
points for the three flux distributions of Fig. 16.20. Fluxes are normalized to the glucose uptake
rate and the net glycolytic flux, respectively, for these two branch points. Dotted arrows at the
pyruvate and oxaloacetate nodes include lysine biosynthesis, and the dotted arrow at the acet-
yl-coenzymeA node includes the glyoxylate cycle activity. These flux distributions demonstrate
the high flexibility of the central metabolism of Corynebacterium glutamicum

genase loS-fold from 53% (lysine production) to 77% (glutamate production). The
back flux in the anaplerotic reactions is increased more than fourfold during growth
as compared to glutamate production. The glyoxylate pathway was not or only very
weakly active in all cases, in accordance with the fact that the activities of the re-
sponsible enzymes are reduced 50-fold during growth on glucose as compared to
acetate as the sole carbon source [31,32]' and in accordance with [98].
16.4 Studies Based on Isotopic Labeling and NMR 545

16.4.3.2
Comparing Different Strains in Batch Cultures

Metabolic flux analysis in C. glutamicum during growth and glutamate production

(strain ATCC 13032) as well as lysine production (strain MH20-22B) in batch cul-
tures was performed using a slightly modified procedure [99]. Only cytoplasmic glu-
tamate and alanine were available for NMR analysis of 13~ fractional enrichments.
Therefore, in order to obtain a reliable estimate of the pentose phosphate pathway
flux, each experiment was performed in duplicate: first, with [l-13C]glucose, and
second, with [6-13C]glucose as a substrate. The fraction of glucose 6-phosphate me-
tabolized over the ppp relative to that over glycolysis was calculated as (I-AlB) [92],
where A and B represent the 13C labeling in C-3 of pyruvate obtained from the ex-
periments with [l-13C]glucose and [6-13C]glucose, respectively. Since pyruvate could
not be isolated from the cell extracts, the 13C labeling of C-3 of alanine and/or C-4 of
glutamate (which are both directly derived from C-3 of pyruvate) were used instead.
The flux ratio of the ppp vs glycolysis was then supplied as an additional measure-
ment value to the flux analysis software, in addition to the measured glucose uptake
rates, product excretion rates, and precursor fluxes for biomass synthesis [99]. The
results obtained are summarized in Fig. 16.22 as flux distributions at the glucose 6-
phosphate and PEP/pyruvate branching points. While absolute values differ, several
tendencies observed in the experiments with chemostat cultures are also apparent
from the batch cultures. PPP activity during lysine production is again increased

a Glutamate b Growth c Lysine

production production

* * *
.@ G6P~ ,·0·· G6P----@)-+ 'CD·G6P~
83 58 52

®. @.

Fig. 16.22a-c. Detailed flux partitionings at the glucose-6-phosphate and PEP/pyruvate branch
points for flux distributions determined from batch cultures of: a glutamate-producing Cory-
nebacterium glutamicum ATCC 13032; b exponentially growing Corynebacterium glutamicum
ATCC 13032; c lysine-producing strain MH20-22B. The analyses involved metabolite balancing
as well as NMR analrsis of cytoplasmic alanine and glutamate upon incubation with
[l-13C)glucose and [6_ 1 C]glucose in separate experiments. Data were taken from [99]. Fluxes
are normalized to the glucose uptake rate and the net glycolytic flux, respectively, for the two
branch points. Dotted arrows at the pyruvate and oxaloacetate nodes include lysine biosynth-
esis
546 16 Metabolic Flux Analysis of Corynebacterium glutamicum

almost threefold as compared to glutamate production, and the anaplerotic carbox-

ylation during glutamate production is almost twice as low as during lysine produc-
tion. The back reaction from oxaloacetate/malate to PEP/pyruvate is again lowest
during glutamate production. However, forward- as well as reverse fluxes connecting
PEP/pyruvate and oxaloacetate/malate are now highest during lysine production,
whereas this was not the case with the continuous cultures. Also, the flux from suc-
cinate to oxaloacetate during glutamate production is almost twice as low as during
growth and lysine production, whereas in the chemostat cultures it was higher.
While these differences may partly arise due to significant imprecision of the flux
estimates for the anaplerotic reactions (not assessed in [99]), they also certainly re-
flect metabolic differences between non-carbon-limited batch and glucose-limited
chemos tat cultures. That such differences must exist can be deduced from the sig-
nificantly altered growth yields YXIS during lysine and glutamate production, which
are strongly reduced to 0.15 and O.Ogxxgs-I, respectively. As with the continuous
cultures, the glyoxylate pathway was not active in all cases. Comparing all experi-
ments, the situation with glutamate production is obviously rather a special one,
since activity of the PPP as well as of anaplerotic carboxylation reactions is consis-
tently low. The low PPP activity can easily be explained by the low demand for
NADPH during glutamate production (see also next section). A possible explanation
for the low anaplerotic carboxylation activity would be that the activity of biotin-
dependent pyruvate carboxylase is strongly reduced during biotin-limitation-in-
duced glutamate production, which would lower the overall anaplerotic rate and re-
duce the need to recycle excess oxaloacetate to PEP/pyruvate.

16.4.4
Perturbations of the Redox Metabolism

The flux analyses described in the previous two sections, that are based on indepen-
dent l3C labeling data, allow one to perform an independent analysis of the
NADP(H) redox balance for each experiment. To this end, the NADPH stoichiome-
tries as presented in Table 16.1 can be used. These differ slightly from those used in
[56, 100, lOll due to newly incorporated (deoxy)nucleotide and peptidoglycan data.
Thus, an estimated 16,385/Lmol of NADPH are required in the biosynthesis of 1 g c.
glutamicum biomass from glucose.
Furthermore, lysine biosynthesis requires effectively 4 mol NADPH per mol lysine,
and glutamate 1 mol NADPH per mol glutamate. In C. glutamicum, NADPH is regen-
erated from NADP only in the oxidative PPP (2 mol NADPH per mol glucose 6-phos-
phate converted) and in the isocitrate dehydrogenase reaction (disregarding malic
enzyme). Taking this into account, the relative fluxes of oxidation and regeneration
of NADPH as given in Table 16.7 are arrived at. From Table 16.7, there seems to be a
positive correlation between the total amount of NADPH oxidized and the amount
regenerated in the PPP, and a negative correlation between the total amount of
NADPH oxidized and the amount regenerated in the isocitrate dehydrogenase reac-
tion. This prompted study of the effect of a defined modification of the redox meta-
bolism of C. glutamicum [lOll. In this study, glutamate dehydrogenase was chosen
as the target because this single enzyme accounts for approximately 50% of the total
NADPH oxidized in the synthesis of biomass. A change of the cofactor dependency
of this enzyme therefore is expected to result in a significantly altered NADPH-re-
lated redox metabolism. Thus, a glutamate dehydrogenase-deficient mutant of the
 16.4 Studies Based on Isotopic Labeling and NMR 547

Table 16.7. Relative fluxes of oxidation and regeneration of NADPH (mol per mol of glucose con-
sumed) calculated from the flux distributions presented in [56,99, 100, 110] and the NADPH stoichio-
metries given in Table 16.1

Condition NADPH (mol/mol)

Oxidized Regenerated
Biomass Gluta- Lysine Total PPP Isocitrate Total Net
synthesis mate synthesis dehydro- regene-
synthesis genase rated
Glutamate production (batch) 0 0.66 0 0.66 0.34 1.28 1.62 0.96
Growth (batch) 1.09 0 0 1.09 0.80 1.08 1.88 0.79
Lysine production (batch) 0.44 0 0.96 lAO 0.94 1.06 2.00 0.60
Glutamate production 0.80 0.21 0 1.01 0.50 1.11 1.61 0.60
(chemostat)
Growth (chemostat) 1.36 0 0 1.36 0.72 0.93 1.65 0.29
Lysine production 1.09 0 0.75 1.84 1.32 0.61 1.93 0.09
(chemostat)

lysine-producing strain MH20-22B was complemented in one case with the homo-
logous NADP-dependent plasmid-encoded glutamate dehydrogenase, and in another
case with an heterologous, NAD-dependent glutamate dehydrogenase from Peptos-
treptococcus asaccharolyticus. The two strains are henceforth termed homologous
mutant and heterologous mutant, respectively.
Continuous fermentations with [1-l3C]-labeled glucose as the sole carbon source
were carried out, and integrated metabolite balancing/ 13 C stable isotope labeling flux
analyses were performed [101]. The resulting flux distributions at the glucose
6-phosphate and PEP/pyruvate branching points are given in Fig. 16.23. Obviously,
the defined alteration of the redox metabolism leads to remarkable changes in the
flux distribution. The effect is most striking for the PPP, of which the activity is
reduced threefold in the heterologous mutant. The flux distribution at the PEP/pyr-
uvate and oxaloacetate/malate branching points is less affected, showing a 36% in-
crease in citric acid cycle flux and also a significant increase in oxaloacetate/malate
de carboxylating flux in the heterologous mutant [101]. With the fluxes thus deter-
mined, balances of NADPH-oxidizing and -regenerating fluxes can again be estab-
lished, on the assumption that the heterologous glutamate dehydrogenase is strictly
NAD-dependent (in fact, enzyme measurements indicated a 10% NADP-dependent
side activity). In setting up the balances, special care must be taken to correct the
NADP stoichiometries (Table 16.1) of the various biosynthetic reactions for the al-
tered glutamate dehydrogenase cofactor dependency. For example, the I-step and
4-step pathways in lysine biosynthesis (Fig. 16.4) will now have different NADPH
stoichiometries: while synthesis of lysine from pyruvate and oxaloacetate in the
homologous mutant requires 4 NADPH for either pathway, synthesis via the dehy-
drogenase (I-step) and succinylase (4-step) pathway requires 3 NADPH and
2 NADPH, respectively, in the heterologous mutant. Since the flux distribution over
these pathways was also determined in the integrated analysis, a rigorous book-keep-
ing of the NADPH-oxidizing and -regenerating fluxes can be performed. The results
are shown in Table 16.8. A very strong and flexible response of C. glutamicum central
metabolism is the answer to the twofold decrease in requirement for NADPH regen-
 548 16 Metabolic Flux Analysis of Corynebacterium glutamicum

a b

Fig. 16.23a,b. Detailed flux partitionings at the glucose-6-phosphate and PEP/pyruvate branch
points for flux distributions determined from chemostat cultures of glutamate dehydrogenase
mutants oflysine-producing Corynebacterium glutamicum MH20-22B equipped with plasmid-
encoded: a homologous (i.e., NADP-dependent); b heterologous (i.e., NAD-dependent) gluta-
mate dehydrogenases. The analyses involved integrated I3 e NMR/metabolite procedures. Data
were taken from [101]. Fluxes are normalized to the glucose uptake rate and the net glycolytic
flux, respectively, for the two branch points. Dotted arrows at the pyruvate and oxaloacetate
nodes include lysine biosynthesis. The flux partitioning at the PEP/pyruvate node is much less
affected than that at the glucose-6-phosphate node

Table 16.8. Relative fluxes of oxidation and regeneration of NADPH (mol per mol of glucose con-
sumed) during lysine production with glutamate-dehydrogenase mutants of C. glutamicum MH20-
22B, calculated from the flux distributions presented in [101] and the NADPH stoichiometries given
in Table 16.1 after adaptation for the NAD-dependent glutamate dehydrogenase in the heterologous
mutant

Strain NADPH (mol/mol)

Oxidized Regenerated
Biomass Gluta- Lysine Total PPP Isocitrate Total Net re-
synthesis mate synthesis dehydro- generated
synthesis genase

Heterologous mutant 0.55 0 0.44 0.99 0.53 0.86 1.39 0.40

Homologous mutant 0.83 0 1.20 2.03 1.53 0.58 2.11 0.07

eration. Very interestingly, the PPP adapts quantitatively to the change. However, the
excess NADPH regenerated in the isocitrate dehydrogenase reaction results in a net
over-regeneration of NADPH in the heterologous mutant, whereas (within experi-
mental error) no excess regeneration of NADPH is present in the homologous mu-
tant. The last column of Tables 16.7 and 16.8 show data regarding the NADPH over-
regeneration for all cases analyzed. A clear tendency in the rate of over-regeneration
of NADPH can be seen, namely that it shows a negative correlation with the rate of
16.4 Studies Based on Isotopic Labeling and NMR 549

2.5
Q)
8
.=! 2.0
C>
o
E

.s15 1.5
~
~ 1.0
c:
Q)
C>
~
::c 0.5
Il.
Cl
~
0.0 4£--~iL----;'---~--';----';
0.0 0.5 1.0 1.5 2.0 2.5
NADPH oxidised (mol/mol glucose)

Fig. 16.24. Calculated fluxes (mol per mol of glucose consumed) of regenerated NADPH by the
oxidative pentose phosphate pathway plus isocitrate dehydrogenase (circles) and by the oxida-
tive pentose phosphate pathway alone (squares), as a function of the calculated total flux of
NADPH oxidation (mol per mol of glucose consumed) for flux distributions presented in Ta-
bles 16.7 and 16.8. The continuous line represents points where oxidizing and regenerating
fluxes match exactly. The broken line suggests a trend in the data points with square symbols.
An excess of NADPH regeneration is apparent that seems to increase with diminishing
NADPH oxidation

NADPH oxidation, i.e., the higher the rate of oxidation, the less NADPH is over-
regenerated. This is more evident from the graphical representation of selected data
from Tables 16.7 and 16.8 in Fig. 16.24. Apparently, the PPP consistently regenerates
NADPH to a rate 0.5 mol!molless than the rate of oxidation. The additional NADPH
regeneration by the isocitrate dehydrogenase reaction matches the required 0.5 mol!
mol at high NADPH oxidation rates (i.e., 1.8-2.0 mol!mol), resulting in a good match
of total oxidation and regeneration rates, but it is increasingly in excess of this
amount the less NADPH is oxidized. Several speculative explanations are possible.
One is that at lower oxidation rates, additional oxidative processes are active that are
not accounted for in the calculations. These processes may serve as a kind of redu-
cing power sink, which the organism possibly needs in order to maintain a basic
level of NADPH-turnover. While the cycle pyruvate-malate-oxaloacetate-pyruvate in-
volving malic enzyme might theoretically act as a transhydrogenase system oxidizing
NADPH, a functioning of malic enzyme in the anaplerotic sense in C. glutamicum to
date could not be demonstrated. Growth-related processes are also unlikely candi-
dates, since no correlation with biomass synthesis is apparent from the combined
Tables 16.7 and 16.8. The only positively identified candidate to date is a recently
discovered superoxide-generating NADPH oxidase system in the respiratory chain
of c. glutamicum [111]. Another theoretical possibility would be that isocitrate de-
hydrogenase, contrary to reports, is not strictly NADP-dependent, or that an NAD-
dependent isoenzyme exists in C. glutamicum. However, no experimental data sup-
porting either one of these hypotheses is currently available.
550 16 Metabolic Flux Analysis of Corynebacterium glutamicum

16.4.5
The Ammonium-Assimilating Fluxes

A glutamate dehydrogenase mutant of C. glutamicum ATCC 13032 did not show any
phenotype [46]. This seemed to be incompatible with knowledge of the regulation of
the principal enzymes for ammonium assimilation, i.e., glutamate dehydrogenase
(GDH), glutamine synthetase (GS), and glutamine 2-oxoglutarate-aminotransferase
(GOGAT), which indicated that the alternative GS/GOGAT pathway was not ex-
pressed during ammonium abundance. Since it was unclear whether alternative
routes are operative in the GDH-mutant or the regulation of GS/GOGAT is altered,
the principal ammonium-assimilatory fluxes in C. glutamicum were quantitated
using in vivo 15N NMR in a dynamic labeling study [50]. C. glutamicum ATCC
13032 (wildtype) and its GDH-mutant were cultivated and analyzed with on-line
NMR using a specially developed NMR membrane-cyclone bioreactor system [103,
112] which is suited for the continuous aerobic cultivation of C. glutamicum at cell
densities above 50 g dry wt/l. At metabolic steady state, the culture ammonium con-
centration was increased within 2 min from 40 mmoll- 1 to 150 mmoll- 1 by the addi-
tion of 15N-Iabeled ammonium sulfate. The time courses of the ensuing incorpora-
tion of the label in the Na nitrogen of glutamate and the Na and No nitrogens of
glutamine were monitored with a time resolution of 8-15 min. After proper calibra-
tion of the NMR data, the parameters of a nitrogen flux model were then fitted to the
data. The differential equation model included the following reactions: GDH (rever-
sible), GS, GOGAT (reversible), glutamate aminotransferases, glutamine amidotrans-
ferases, and incorporation of glutamate and glutamine in protein and cytoplasmic
pools. Tests on the GDH mutant with a GOGAT inhibitor showed that no other pri-
mary ammonium-assimilating pathways are operative in C. glutamicum. The pools
of glutamate and glutamine were found to be strongly altered in the GDH mutant:

a b
Glutamate
Wildtype dehydrogenase
mutant

Fig. 16.2Sa,b. Nitrogen flux distributions in: a Corynebacterium glutamicum ATCC 13032; bits
~lutamate dehydrogenase mutant, determined in a dynamic labelling study [50] using in vivo
sN NMR. TA, transaminases; GDH, glutamate dehydrogenase; GOGAT, glutamine:2-oxogluta-
rate aminotransferase; GS, glutamine synthetase; AT, glutamine amidotransferases
16.4 Studies Based on Isotopic Labeling and NMR 551

82 p,mollg dry wt and 93 p,mollg dry wt, respectively, as compared to 250 p,mollg dry
wt and 49 p,mollg dry wt in the wildtype. The flux distributions resulting from the
analyses [50] are represented in Fig. 16.25. They indeed showed a marked difference
between the two strains; whereas in the wildtype, 72% of the ammonium was assimi-
lated via GDH and 28% via GS, 100% were assimilated via GS in the GDH mutant. In
the wildtype, glutamate resulted from operation of GDH (74%) and glutamine trans-
aminases (26%); GOGAT was inactive. In the GDH mutant, the task of GDH was
completely taken over by GO GAT, via which 72% of the glutamate was produced.
The validity of the analysis was confirmed by the fact that it predicted zero activity
of GDH for the mutant. The absence of GOGAT flux predicted for the wildtype was
completely in accordance with enzyme determinations. Thus, nitrogen fluxes can
also be quantified independently in great detail by 15N labeling and NMR.

16.4.6
Summary

Flux analysis studies based on isotopic labeling and NMR detection have provided a
much refined insight in the metabolic processes as they occur in vivo in a number of
different strains of C. glutamicum and under a variety of conditions. These studies
have given undoubted independent proof that the pentose phosphate pathway is very
important for an adequate supply of reducing equivalents during amino acid produc-
tion, and that this pathway reacts extremely flexibly to the actual demand for
NADPH in any given situation [56,99-101]. The experiments also gave independent
proof that the glyoxylate pathway in C. glutamicum is not or only very weakly active
when glucose is the sole carbon source [56,99, 100]. Furthermore, an important in-
sight in the regulation of the flux distribution at the tetrahydrodipicolinate branch-
point of the lysine biosynthetic pathway of C. glutamicum was provided [62], show-
ing that ammonium is a key factor in the activation of the one-step dehydrogenase
pathway which is essential for high lysine productivity. Several reversible reaction
steps were identified from the l3e labeling data. Typical degrees of reversibility for
the trans aldolase, transketolase, phosphoribose isomerase, ribulose-S-phosphate epi-
merase, phosphoglucose isomerase, as well as for the reactions interconverting gly-
ceraldehyde-3-phosphate and PEP/pyruvate, and oxaloacetate/malate and fumarate,
could be calculated from the data.
The integrated metabolite balancing/isotopic labeling analysis revealed a first clue
to the elucidation of the mechanisms that contribute to the rigidity of the flux dis-
tribution observed around the PEP/pyruvate/oxaloacetate branching points in C.
glutamicum metabolism: the presence of a very significant reverse flux of oxaloace-
tate/malate to PEP/pyruvate that varies considerably depending on conditions [56,
100, 101]. The available data suggest an important role of these backreactions in the
regulation of the relative levels of the pools of PEP, pyruvate, and oxaloacetate, which
are chief amino acid precursors. The l3e labeling approach has also been highly
instrumental in the identification of pyruvate carboxylase as important anaplerotic
enzyme in C. glutamicum [70, 102].
The in vivo 15N NMR dynamic labeling flux analysis study led to the discovery that
glutamine, together with glutamate, plays a key role in the ammonium assimilatory
metabolism of C. glutamicum.
As a whole, the results from the studies based on isotopic labeling and NMR de-
monstrate that the central metabolism of C. glutamicum is extremely flexible in
552 16 Metabolic Flux Analysis of Corynebacterium glutamicumReferences

adapting to situations with strongly different precursor demands, and that both glu-
tamine and glutamate are key metabolites in C. glutamicum nitrogen metabolism.
Furthermore, the complex set of PEP/pyruvate carboxylating and oxaloacetate-dec-
arboxylating reactions is identified as a promising object for further analysis.

16.5
Concluding Remarks

While extensive genetic knowledge about the central metabolic and amino acid bio-
synthetic pathways in C. glutamicum has accumulated already over many years, me-
tabolic flux analysis studies have now begun to yield deeper insights into the inter-
play of amino acid biosynthesis and the central metabolism which generates the pre-
cursors, the energy, and the reducing power necessary to produce these compounds.
Several targets for genetic engineering have been identified; some of these have al-
ready been probed while others will be studied in the near future. Metabolic flux
analysis has thus developed into a very powerful measurement tool. The strategies
employing integrated metabolite balancing and stable isotope-labeling combined
with NMR-detection described in this contribution have significantly contributed
to this advancement. Present developments which aim to maximize the information
content of the stable isotope labeling experiment by using 13 C-isotopomer analysis,
and which seek to integrate the flux information with cytoplasmic pool measure-
ments, are likely to strengthen further the role of metabolic flux analysis within the
rapidly evolving field of metabolic engineering in the near future.
Acknowledgements. The author wishes to thank W. Wiechert, A. Marx, L. Eggeling,
M. Tesch, K. Striegel, K. Sonntag, V. Wendisch, and A. Hartbrich for their excellent
contributions to the NMR and flux analysis studies covered in this contribution, and
Prof. H. Sahm for continuous support and stimulation.

References
1. Eggeling L, Sahm H, de Graaf AA (1996) Quantifying and directing metabolite flux:
application to amino acid overproduction. In: Scheper T (ed) Advances in biochemical
engineering/biotechnology, vol 54: Metabolic engineering. Springer, Berlin Heidelberg
New York, p 2
2. Katsumata R, Ikeda M (1997) Process for producing I-tryptophan, I-tyrosine or I-phe-
nylalanine. US Pat 5,605,818
3. Katsumata R, Hashimoto S (1996) Process for producing I-valine. US Pat 5,521,074
4. Mizukami T, Katsumata R, Oka T (1990) Process for producing histidine. US Pat
4,927,758
5. Malumbres M, Martin JF (1996) FEMS Microbiol Lett 143:103
6. Reinscheid DJ, Kronemeyer W, Eggeling L, Eikmanns BJ, Sahm H (1994) Appl Environ
Microbiol 60:126
7. Eggeling L, Morbach S, Sahm H (1997) J Biotechnol 56:167
8. Kimura E, Kawahara Y, Nakamatsu T (1997) Tanpakushitsu Kakusan Koso (Japan)
42:2633
9. Eggeling L (1994) Amino Acids 6:261
10. Sahm H, Eggeling L, Eikmanns BJ, Kramer R (1996) Ann N YAcad Sci 782:25
11. Eikmanns BJ, Eggeling L, Sahm H (1993) Antonie van Leeuwenhoek 64:145
12. Jetten MSM, Sinskey AJ (1995) Crit Rev Biotechnol 15:733
l3. Sahm H, Eggeling L, Eikmanns BJ, Kramer R (1995) FEMS Microbiol Rev 16:243
14. Jetten MSM, Follettie MT, Sinskey AJ (1994) Ann N YAcad Sci 721:12
References 553

15. Kramer R (1994) FEMS Microbiol Rev l3:75

16. Kramer R (1996) Analysis and modeling of substrate uptake and product release by
prokaryotic and eukaryotic cells. In: Scheper T (ed) Advances in biochemical engineer-
ing/biotechnology, vol 54: Metabolic engineering. Springer, Berlin Heidelberg New York,
p 31
17. Palmieri L, Berns D, Kramer R, Eikmanns BJ (1996) Arch Microbiol165:48
18. Hermann T, Kramer R (1996) Appl Environ Microbiol 62:3238
19. Zittrich S, Kramer R (1994) J BacterioI176:6892
20. Eggeling L, Kramer R, Vrljic M, Kronemeyer W, Sahtn H (1996) Ann N Y Acad Sci
782:191
21. Vrljic M, Sahm H, Eggeling L (1996) Mol Microbiol 22:815
22. Schwinde JW, Thurn-Schmitz N, Eikmanns BJ, Sahm H (1993) J Bacterio1175:3905
23. von der Osten CH, Barbas CF, Wong CH, Sinskey AJ (1989) Mol Microbiol 3:1625
24. Eikmanns BJ (1992) J Bacterio1174:6076
25. Gubler M, Jetten MSM, Lee SH, Sinskey AJ (1994) Appl Environ Microbiol 60:2494
26. Ikeda M, Okamoto K, Katsumata R (1999) Appl Microbiol Biotechnol 51:201
27. Moritz B, Striegel K, de Graaf AA, Sahm H (2000) Eur J Biochem (submitted)
28. Eikmanns BJ, Thurn-Schmitz N, Eggeling L, Ludtke KU, Sahm H (1994) Microbiol
140:1817
29. Eikmanns BJ, Rittmann D, Sahm H (1995) J Bacterioll77:774
30. Usuda Y, Tujimoto N, Abe C, Asakura Y, Kimura E (1996) Microbiol 142:3347
31. Reinscheid DJ, Eikmanns BJ, Sahm H (1994) J Bacterio1176:3474
32. Reinscheid DJ, Eikmanns BJ, Sahm H (1994) Microbio1140:3099
33. Reinscheid DJ, Schnicke S, Rittmann D, Zahnow U, Sahm H, Eikmanns BJ (1999) Micro-
bioi 145:503
34. Kinoshita S (1985) Glutamic acid bacteria. In: Demain A, Solomon N (eds) Biology of
industrial microorganisms. Benjamin/Cummings Publishing, London, p 115
35 Peters-Wendisch PG, Wendisch VF, Paul S, Eikmanns BJ, Sahm H (1997) Microbiol
143:1095
36. Dominguez H, Nezondet C, Lindley ND, Cocaign M (1993) Biotechnol Lett 15:449
37. Jetten MSM, Sinskey AJ (1993) FEMS Microbiol Lett 111:183
38. Jetten MSM, Sinskey AJ (1995 Antonie Van Leeuwenhoek 67:221
39. Jetten MSM, Pitoc GA, Follettie MT, Sinskey AJ (1994) Appl Microbiol Biotechnol 41:47
40. O'Regan M, Thierbach G, Bachmann B, Villeval D, Lepage P, Viret JF, Lemoine Y (1989)
Gene 77:237
41. Eikmanns BJ, Follettie MT, Griot MU, Sinskey AJ (1989) Mol Gen Genet 218:330
42. Peters-Wendisch PG, Kreutzer C, Kalinowski J, Patek M, Sahm H, Eikmanns BJ (1998)
Microbiol144:915
43. Koffas MA, Ramamoorthi R, Pine WA, Sinskey AJ, Stephanopoulos G (1998) Appl Mi-
crobiol Biotechnol 50:346
44 Siewe RM, Weil B, Burkovski A, Eikmanns BJ, Eikmanns M, Kramer R (1996) J Bioi
Chern 271:5398
45. Siewe RM, Weil B, Burkovski A, Eggeling L, Kramer R, Jahns T (1998) Arch Microbiol
169:411
46. Bormann-EI Kholy ER, Eikmanns BJ, Gutmann M, Sahm H (1993) Appl Environ Micro-
bioi 59:2329
47. Bormann ER, Eikmanns BJ, Sahm H (1992) Mol Microbiol6:317
48. Jakoby M, Tesch M, Sahm H, Kramer R, Burkovski A (1997) FEMS Microbiol Lett 154:81
49. Tesch M, Eikmanns BJ, deGraaf AA, Sahm H (1998) Biotechnol Lett 20:953
50. Tesch M, de Graaf AA, Sahm H (1999) Appl Environ Microbiol 65:1099
51. Eggeling L, Morbach S, Sahm H (1996) Appl Environ Microbiol 62:4345
52. Stephanopoulos G; Vallino JJ (1991) Science 252:1675
53. De Hollander JA (1994) Appl Microbiol Biotechnol 42:508
54. Vallino JJ, Stephanopoulos G (1990) In: Sikdar SK, Bier M, Todd P (eds) Frontiers in
bioprocessing. CRC Press, Boca Raton, p 205
55. Cocaign-Bousquet M, Lindley ND (1995) Enz Micr Techno117:260
554 16 Metabolic Flux Analysis of Corynebacterium glutamicum

56. Marx A, de Graaf AA, Wiechert W, Eggeling L, Sahm H (1996) Biotechnol Bioeng 49: III
57. Neidhardt FC, Ingraham JL, Schaechter M (1990) Physiology of the bacterial cell: a mo-
lecular approach. Sinauer Associates, Sunderland, Massachusetts
58. SchrumpfB, Schwarzer A, Kalinowski J, Puhler A, Eggeling L, Sahm H (1991) J Bacteriol
173:4510
59. Vallino JJ, Stephanopoulos G (1993) Biotechnol Bioeng 41:633
60. Broer S, Kramer R (1991) Eur J Biochem 202:136
61. Vallino JJ, Stephanopoulos G (1994) Biotechnol Progr 10:320
62. Sonntag K, Eggeling L, de Graaf AA, Sahm H (1993) Eur J Biochem 213:1325
63. Kanamori K, Weiss RH, Roberts JD (1987) J Bioi Chern 262:11038
64. Kanamori K, Weiss RL, Roberts JD (1987) J Bacteriol 169:4692
65. Kanamori K, Weiss RL, Roberts JD (1988) J Bioi Chern 263:2817
66. Szyperski T (1995) Eur J Biochem232:433
67. Szyperski T, Bailey JE, Wuthrich K (1996) Trends Biotechnol14:453
68. Wendisch VF, de Graaf AA, Sahm H (1997) Anal Biochem 245:196
69. de Graaf AA, Mahle M, Miillney M, Wiechert W, Stahmann P, Sahm H (2000) J Biotech-
no177:25
70. Park SM, Shaw-Reid, Sinskey AJ, Stephanopoulos G (1997) Appl Microbiol Biotechnol
47:430
71. Inbar L, Lapidot A (1987) Eur J Biochem 162:621
72. van Gulik WM, Heijnen JJ (1995) Biotechnol Bioeng 48:681
73. Zupke C, Stephanopoulos G (1994) Biotechnol Progr 10:489
74. Wiechert W, de Graaf AA (1997) Biotechnol Bioeng 55:101
75. Wiechert W, Sietke C, de Graaf AA, Marx A (1997) Biotechnol Bioeng 55:118
76. Szyperski T (1998) Q Rev Biophys 31:41
77. Wiechert W, de Graaf AA (1996) In vivo stationary flux analysis by !3 C labelling experi-
ments. In: Scheper T (ed) Advances in biochemical engineering/biotechnology, vol 54:
Metabolic engineering. Springer, Berlin Heidelberg New York, pill
78. Klapa MI, Park SM, Sinskey AJ, Stephanopoulos G (1999) Biotechnol Bioeng 62:375
79. Park SM, Klapa MI, Sinskey AI, Stephanopoulos G (1999) Biotechnol Bioeng 62:392
80. Schmidt K, Carlsen M, Nielsen J (1997) Biotechnol Bioeng 55:831
81. Zupke C, Tompkins R, Yarmush D, Yarmush M (1997) Anal Biochem 247:287
82. Wiechert W, Mollney M, Isermann N, Wurzel M, de Graaf AA (1999) Biotechnol Bioeng
66:69
83. Mollney M, Wiechert W, Kownatzki D, de Graaf AA (1999) Biotechnol Bioeng 66:86
84. Vallino JJ, Stephanopoulos G (1994) Biotechnol Prog 10:327
85. Lindley ND, Cocaign-Bousquet M, Guyonvarch A (1996) Appl Environ Microbiol62:429
86. Sugimoto SI, Shiio I (1987) Agric Bioi Chern 51:101
87. Kinoshita S, Udaka S, Shimono M (1957) J Gen Appl Microbiol (Tokyo) 3:193
88. Shiio I, Otsuka S-I, Tsunoda T (1960) J Biochem 47:414
89. Oishi K, Aida H (1965) Agric Bioi Chern 29:83
90. Otsuka S-I, Miyajima R, Shiio I (1965) J Gen Appl Microbiolll:285
91. Haran N, Kahana ZE, Lapidot A (1983) J Bioi Chern 258:12,929
92. Walker TE, Han CE, Kollman VH, London RE, Matwiyoff NA (1982) J Bioi Chern
257:1189
93. Ishino S, Yamaguchi K, Shirahata K, Araki K (1984) Agric Bioi Chern 48:2557
94. Ishino S, Kuga T, Yamaguchi K, Shirahata K, Araki K (1986) Agric Bioi Chern 50:307
95. Yamaguchi K, Ishino S, Araki K, Shirahata K (1986) Agric Bioi Chern 50:2453
96. Inbar L, Lapidot A (1987) Eur J Biochem 162:621
97. Ishino S, Shimomura J, Yamaguchi K, Shirahata K, Araki K (1991) J Gen Appl Microbiol
37:157
98. Rollin C, Morgant V, Guyonvarch A, Guerquin-Kern J-L (1995) Eur J Biochem 227:488
99. Sonntag K, Schwinde J, de Graaf AA, Marx A, Eikmanns BJ, Wiechert W, Sahm H (1995)
Appl Microbiol Biotechnol 44:489
100. Marx A, Striegel K, de Graaf AA, Sahm H, Eggeling L (1997) Biotechnol Bioeng 56:168
101. Marx A, Eikmanns BJ, Sahm H, de Graaf AA, Eggeling L (1999) Metab Eng 1:35
References 555

102. Peters-Wendisch PG, Wendisch VF, de Graaf AA, Eikmanns BJ, Sahm H (1996) Arch
Microbiol 165:387
103. Hartbrich A, Schmitz G, Weuster-Botz D, de Craaf AA, Wandrey C (1996) Biotechnol
Bioeng 51:624
104. Wehrmann A, Phillipp B, Sahm H, Eggeling L (1998) J Bacterio1180:3159
105. Shaw-Reid CA, McCormick MM, Sinskey AJ, Stephanopoulos G (1999) Appl Microbiol
Biotechnol 51 :325
106. Cremer J, Treptow C, Eggeling L, Sahm H (1988) J Gen Microbiol 134:3221
107. Schrumpf B, Eggeling L, Sahm H (1992) Appl Microbiol Biotechnol 37:566
108. Misono M, Soda K (1980) J Bioi Chern 255:10,599
109. Shiio I, Otsuka S-I, Takahashi M (1962) J Biochem (Tokyo) 52:108
110. Marx A (1997) PhD Thesis, University of Bonn, Germany
111. Matsushita K, Yamamoto T, Toyama H, Adachi 0 (1998) Biosc Biotechnol Biochem
62:1968
112. Weuster-Botz D, de Graaf AA (1996) Reaction engineering methods to study intracellu-
lar metabolite concentrations. In: Scheper T (ed) Advances in biochemical engineering/
biotechnology, vol 54: Metabolic engineering. Springer, Berlin Heidelberg New York,
p.75
113. Gottschalk G (1986) Bacterial metabolism, 2nd edn. Springer, Berlin Heidelberg New
York, chap 3
114. Molenaar D, van der Rest ME, Petrovic S (1998) Eur J Biochem 254:395
CHAPTER 17

17 Analysis of Metabolic Fluxes in Mammalian Cells

Neil S. Forbes, Douglas S. Clark, Harvey W. Blanch

17.1
Applications of Metabolic Flux Analysis in Mammalian Cells
The analysis of metabolic fluxes provides a unique perspective on the functioning of
mammalian cells. Metabolic flux analysis is the use of a mathematical description of
central metabolism to quantify the flow of carbon through the cell. Compared to
other biochemical techniques, which reduce cellular function to specific enzymes
and pathways, metabolic flux analysis views the cell as a whole. Concurrently mea-
suring the fluxes through multiple biochemical pathways can identify interactions in
the regulatory machinery of the cell. When pathways are observed independently,
their interactions must be inferred. However, associations between pathways can
only be seen when their respective fluxes are compared. Understanding the interac-
tions between pathways enables the prediction of cellular responses to changes in
nutrient composition, the response to the genetic events involved in transformation,
and the response to external stimuli and signals.
This chapter reviews the modeling methods used to quantify metabolic fluxes in
mammalian cells. We describe the different systems and cell types to which metabolic
flux analysis has been applied. The examples considered all employ a biochemical path-
way model of the whole cell to determine metabolic fluxes. We describe the common
assumptions used to create viable models and the corresponding mathematical de-
scriptions to deconvolute typical experimental data into fluxes. Model building and
mathematical manipulation is used because many fluxes can be quantified without di-
rect measurements. Compared to measuring each flux independently, the use of a path-
way model requires much less experimental effort. Studies of carbon fluxes through
single pathways, independent of central metabolism, will not be considered here.
Metabolic flux analysis is not new. Because of numerous technological advances,
many previously unanswered questions and new issues can now be addressed. Two
experimental tools, !3C_ and l4C-tracer studies, have traditionally been employed to
study metabolic pathways. Most of the biochemical pathways of central metabolism
were first elucidated using 14C-Iabeling in the 1920s and 1930s [1). With the advent
of more powerful superconducting magnets in the 1970s, nuclear magnetic reso-
nance (NMR) spectroscopy using !3C-enriched substrates provided a new route to
study biological systems [2). For example, in 1983 Chance et al. [3) incorporated
isotopomer 1 data derived from acid extraction of perfused rat hearts into a relatively

An isotopomer, as defined later in the review, is a metabolite with a specific pattern of

labeled carbons. There are 2n isotopomers in a molecule with n carbon atoms.
17.1 Applications of Metabolic Flux Analysis in Mammalian Cells 557

simple model to quantify the flux of the TCA cycle. At that time, even the modest
complexity of their model required the computing power of a Cray mainframe com-
puter for its solution. In contrast, Chatham et al. [4] reused Chance's model in 1994,
with added pathways, and solved the system of equations on a laptop 486 personal
computer. The striking development of more powerful computers in the last decade
has allowed a re-examination of whole-cell flux analysis. The experimental techni-
ques have not changed very much, but now more information can be derived from
isotope distributions (and other measurements) using more complex models of me-
tabolic pathways.
To date, there have been three major applications for flux analysis of whole cells:
optimization of therapeutic protein production, understanding metabolic regulation
in cancerous cells, and understanding metabolic regulation in normal cells. These
applications will now be discussed in turn.

17.1.1
Optimization of Protein Production

Large-scale mammalian cell culture is used almost exclusively to produce therapeu-

tic proteins. Typically, a desired protein comprises only about 10-20% of the total
cellular protein [5]. Clearer understanding of the interactions between substrate con-
sumption, product formation, and energy management could enhance productivity
and reduce the costs of culturing mammalian cells. Based on this information,
growth media could be rationally designed, and bioreactor design and control could
be improved [6]. A more complete picture of metabolism would include limiting
pathways, inter-relationships between pathways, and environmental effects on nutri-
ent requirements [5]. As an example of the power of flux analysis to generate pro-
duction-increasing strategies, Fitzpatrick et al. [7] measured metabolic fluxes in hy-
bridomas and examined how the interaction between the consumption of glucose
and glutamine affected energy production. The results allowed these authors to hy-
pothesize that limited ATP formation might limit antibody production. With incom-
plete glucose oxidation much of the energy contained in the glucose is lost. These
authors proposed that to increase ATP production and produce more protein would
involve inhibiting lactate formation and inducing pyruvate dehydrogenase (PDH).
These applications mirror ongoing efforts to optimize bacterial fermentations. An
even greater challenge than identifying the optimal environment for cell growth is
the genetic manipulation of a cell line to shift its metabolism towards higher produc-
tion rates. Efforts with bacteria have shown this to be a difficult task, because of
complex inter-regulation of the central metabolic enzymes [8]. Increasing the flows
through a given metabolic pathway often does not result in an overall increase in
production of a desired end product, due to unexpected limitations and controls
elsewhere in the cell. Similar efforts are compounded in mammalian cells by the
difficulty of introducing stable genetic alterations.

17.1.2
Metabolic Regulation in Transformed Cells

Since primary metabolism provides the energy and cellular components necessary
for proliferation, quantifying metabolic fluxes may shed light on the altered mechan-
isms of growth in cancerous cells. It has been known for some time that transformed
558 17 Analysis of Metabolic Fluxes in Mammalian Cells

cells have high rates of aerobic glycolysis [9], a characteristic typical of rapidly pro-
liferating cells. Warburg postulated in 1956 that a cure for cancer would be found if
the causes of the high aerobic glycolysis rates were discovered [10]. Since then, many
of the genetic mutations leading to the onset of cancer have been uncovered, invali-
dating much of Warburg's conjecture. However, the causes of aerobic glycolysis in
proliferating cells have yet to be determined even though it remains a defining char-
acteristic of the transformed state. High aerobic glycolytic rates have been observed
in all of the cancerous cell studies included in this review [5-7,11-15]. One possible
explanation for these high glycolytic rates is an altered composition of glycolytic
isozymes. The alteration of isozyme composition is so closely associated with trans-
formation that the isozyme composition is used as a clinical marker of progression
to the transformed state [16]. Changes in isozyme expression imply that enhanced
glycolytic rates involve genetic modifications. Another explanation for the high gly-
colytic flux is that it is induced by changes in other metabolic pathways. Portais et al.
[14, 17] and Merle et al. [13] studied how activity of the TCA cycle relates to glyco-
lytic rates by using NMR to quantify the metabolic fate of 1_l3C glucose in C6 glioma
cells. These authors used the C6 cells because they have "representative tumoral
characteristics:' implying that fluxes measured in these cells are useful as a model
of metabolic regulation in all transformed cells. The usefulness of flux quantification
to the study of cancer (here specifically breast cancer) was described by Singer et al.
[15]:
Although considerable effort has been devoted to determining the genetic events
in breast cell carcinogenesis, the relationship of the changes to the biochemical
phenotype of the cancer cell is essentially unknown. A more complete understand-
ing of the biochemical phenotype of normal cells versus cancer cells may allow for
a more focused study of important genetic changes leading to these biochemical
changes.

Unfortunately, achieving such a goal relies on an accurate description of normally

functioning cells, and this is difficult to obtain.

17.1.3
Metabolic Regulation in Non-Transformed Cells

Unlike cancerous cell lines, which are immortal, normal mammalian cells only sur-
vive a few passages in culture, making them difficult to study. Conditions in experi-
ments on immortalized or transformed cell lines can be carefully controlled and
reproduced with cells derived from a single source. On the other hand, normal cells
must be studied as short-lived primary cultures, which are more problematic to con-
trol and reproduce. Because it is difficult to duplicate physiological conditions accu-
rately ex vivo, normal cells are often studied as intact tissues or in their natural
environment within animal subjects.
Various rationales were proposed by the authors reviewed here for quantifying
metabolic fluxes in non-transformed mammalian cells. Hyder et al. [18] quantified
glycolytic and TCA cycle flux rates in rat brains to test the hypothesis that "addi-
tional energy required for brain activation is provided through non-oxidative glyco-
lysis:' Malloy et al. [19] determined the fluxes of the TCA cycle and the activity of the
anaplerotic enzyme, pyruvate carboxylase. Anaplerotic activity is relevant because it
17.2 Experimental Techniques 559

increases the concentrations of TCA cycle metabolites, and changes in these concen-
trations are associated with many "clinically relevant states;' e.g., diabetes and ische-
mia. Martin et al. [20) characterized the metabolic consequences of consumed glu-
cose and glutamine in rat cerebellum astrocytes and granule cells. These authors
sought to understand how the proposed two compartmentalized TCA cycles in the
brain are dependent on intercellular regulations. Chance et al. [3), Chatham et al. [4),
and Schrader et al. [21) only quantified the metabolic flux though glycolysis and the
TCA cycle in rat hearts [3,4) and human erythrocytes [21) without defining a pur-
pose beyond metabolic characterization.

17.2
Experimental Techniques

This section describes the different experimental measurements from which fluxes
are calculated. They include the direct measurement of external metabolite concen-
trations, the determination of l3C and l4C distribution patterns by NMR spectro-
scopy and scintillation counting, respectively, and measurement of enzyme activity
following extraction. Most researchers have employed more than one technique,
since single techniques alone are not generally sufficient to quantify cellular fluxes
uniquely [22). The data generated by different techniques are translated into meta-
bolic fluxes by the mathematical schemes and models described in Sects 17.3 and
17.4, respectively.
The different types of cultures and environments in which metabolic fluxes in
mammalian cells have been measured are outlined in Table 17.1.
The mammalian cell lines used commercially to produce proteins include baby
hamster kidney (BHK) cells, Chinese hamster ovary (CHO) cells, and murine hybri-
domas. All of these cell lines have been adapted to grow in suspension culture and
industrial practice is to employ batch systems, although continuous and perfused
systems are also used.
Tumor cell lines used for cancer research are anchorage dependent. These cells
will only grow when attached to a solid substrate, typically flasks, bottles, perfusion
reactors, and polymer beads. Anchorage-dependent cells must be adapted to growth
in suspension. This adaptation generates cells with de-differentiated genetics and
metabolism, which can undermine the purpose of the research. Tissues, excised

Table 17 .1. Different cell culture types and their uses

Culture Type Use Refs.

Suspension protein production 5-7, 12, 16

(both batch and continuous)
Culture flask and roller bottle protein production and metabolic 13,14,17,20
(batch) research
Perfusion reactor protein production and metabolic 23, 24
research
Perfused tissue metabolic research 3,4, 19,25
Animal subject metabolic research 18
560 17 Analysis of Metabolic Fluxes in Mammalian Cells

from animal subjects, are usually investigated by perfusing them with specified buf-
fers, e.g., Krebs-Henseleit buffer for the perfused rat heart [3, 4, 19]. The measure-
ment of external fluxes in perfused tissues is done in a similar manner to perfused
cell line cultures. An animal subject can also be considered as a perfused culture,
although the measurement of nutrient usage is more complex [18].

17.2.1
Direct Measurement of Extracellular Production and Consumption Rates

Direct measurement of extracellular metabolites is the simplest route for flux quanti-
fication, because many extracellular metabolites exist in large amounts relative to in-
tracellular metabolites. All that is required to determine a production or consumption
flux (in addition to cell concentration measurement) is the measurement of the bulk
concentration of an extracellular metabolite as a function of time. Historically, obser-
ving changes in these fluxes provided the first insights into metabolic regulation by
distinguishing between altered physiologies induced by external stimuli. Further
characterization of metabolic regulation, by determining the flux through internal
pathways, cannot be achieved by measurement of external metabolites alone. The
branched nature of intracellular pathways requires that additional measurements be
made, e.g., 14C_ or l3C-tracer measurements, or assays of extracted enzyme activities.
In the context of metabolic flux analysis, "external flux" refers to transportation of
a metabolite across the membrane of individual cells. However, metabolic flux ana-
lysis is not limited to characterizing homogeneous populations of individual cells.
Metabolic flux analysis has been applied to tissues perfused with medium [3, 4, 19,
25] and specific cells functioning within animal subjects [18]. In these cases, "exter-
nal flux" describes the rate of exchange of a metabolite in the cytoplasm of a group
of specific cells with the surroundings.
The external metabolites typically measured in cell culture are glucose, lactate,
amino acids, ammonia, carbon dioxide, oxygen, fatty acids, sterols, protein products,
fructose, pyruvate, and acetate. The nutrient composition of culture medium is typi-
cally designed to resemble mammalian blood plasma. Therefore, the consumption
rates of cells in animal subjects are found by measuring similar external metabolites
whenever possible. Because measured external fluxes only describe transmembrane
uptake or excretion and not necessarily metabolic utilization [5, 7, 12], the effects of
nutrient storage, e.g., glycogen production, must be considered for some cell lines.
External fluxes are calculated in a different manner for each of three different
types of cultures: continuous suspension culture, perfusion culture, and batch cul-
ture. In both continuous and perfusion cultures at steady state, nutrients are steadily
supplied and wastes are removed; thus concentrations in the bulk solution do not
vary with time. The rates of production and consumption are dependent on the flow
rates for delivery and removal. In batch cultures there is no replenishment of nutri-
ents, so they are depleted over time, and wastes accumulate. Cells within animal
subjects are treated like perfused cultures; it is assumed that tissues are supplied
with a constant supply of nutrients by the blood.
17.2 Experimental Techniques 561

17.2.1.1
Continuous Suspension Culture

Metabolic studies of suspension cultures are best performed using a chemostat [26].
At steady state, the specific flux of any extracellular metabolite (e.g., p,mol gluco-
se hr- l (10 9 cellrl) can be determined from the following mass balance:

[; _ F(C-1 - C-1,0 ) I
1,ex - I V. X (17.1)

where F=volumetric flow rate through the chemostat, V=liquid volume of the che-
mostat, X=cell concentration (cells/ml), Ci=steady state concentration of metabolite
i, (e.g., mmol rl), and Ci,o=feed concentration of metabolite i. In the context of me-
tabolic flux analysis, a flux is defined as the rate of appearance of a reaction product
and is therefore always positive. (Although a nutrient is consumed, and its flux is
traditionally defined as negative, the corresponding transmembrane flux is always
positive.) Hence the inclusion of the absolute value in Eq. (17.1).

17 .2.1.2
Perfused Culture

Fluxes for perfused cultures are calculated in a similar manner as for suspension
continuous culture [24]:

£, _ F(C-1 - C-1,0 ) I
1,ex - I N (17.2)

where N is the cell number within the perfusion device/reactor. For some devices
this value can be quite difficult to determine, and much effort has been made to
calculate it accurately. For example, Mancuso et al. [27] described an NMR technique
to determine the cell number within a hollow fiber bioreactor (HFBR) by measuring
the number of sodium atoms excluded by the cells in a sodium rich medium.
An important consideration for perfused cultures is the effect of cell growth on cell
number, N, which increases as a function of time when the system is not in true
steady state. This issue can be avoided by allowing the cells to reach confluence, at
which point cell growth essentially stops. Perfused tissues and animal subjects usual-
ly also have negligible cell growth. However, for non-confluent cell lines, it is gener-
ally assumed that during the long doubling time much of the nutrients are consumed
for maintenance. The doubling time of many mammalian cell lines is often of the
order of 1-2 days, much greater than the 20-60 min typical of bacterial systems.
From kinetic tracer studies, most metabolic processes, with the exception of bio-
synthesis of large macromolecules, reach steady state after approximately 30 min.
This two order-of-magnitude difference in time scales between growth rate and me-
tabolic rate implies that, even in the presence of growth, the cell number can be
considered constant with regard to metabolic reactions.
562 17 Analysis of Metabolic Fluxes in Mammalian Cells

17.2.1.3
Botch Culture

Mammalian cells are most typically grown in batch. Nutrient concentrations in typi-
cal batch medium are high enough to maintain saturation in the rate-limiting steps
of cellular transport and utilization (assuming that transport and the initial reac-
tions of substrate utilization are regulated by enzymes that exhibit saturation ki-
netics, e.g., a glucose transport protein or hexokinase). Saturation can be confirmed
by observing a linear concentration vs time profile of excreted and consumed meta-
bolites in batch culture. Linear profiles lasting over 12 h have been observed for
some cultures [28]. Specific fluxes can be determined from these linear profiles.
Least-squares fitting is one method for determining the flux at a given point in time:

f _I~
I,ex -
dCil
N dt (17.3)

17.2.2
Detection of Isotope Distribution by 13(-NMR

There are two levels of l3C isotope identification in intracellular metabolites: the
fractional enrichment of specific carbons and the isotopomer distribution of the en-
tire metabolite. Fractional enrichment is the ratio of metabolite labeled at a specified
carbon atom to the total metabolite concentration. A molecule with n carbon atoms
has n fractional enrichments. An isotopomer is a metabolite with a specific labeling
pattern. There are 2ll possible isotopomers in the molecule with n carbons 2 • Isotopo-
mer analysis considers the position of label in multiply-labeled molecules, informa-
tion that would be lost by fractional enrichment analysis. For example, a molecule
labeled at the first and third carbons is a different isotopomer than one labeled at the
first carbon alone. An experimentally useful property of fractional enrichments and
isotopomer distributions is that for metabolites homogeneously compartmentalized
within the cell, any portion of the metabolite will have the same enrichment or iso-
topomer pattern.

17.2.2.1
Measuring Fractional Enrichments

Portais et al. [14] described three different techniques to measure fractional enrich-
ments, briefly described here in turn.

Calculation from Absolute Labeling and Metabolite Concentrations

One key advantage of NMR for determining label distributions is its ability to detect
separate chemical species. In addition to the ability to distinguish different molecu-
lar species, NMR can distinguish the individual carbons on a given molecular spe-
cies. Many carbons of metabolic molecules have distinct and identifiable chemical

2 As defined here, an isotopomer is an actual chemical species. A model based on isotopomer

information will in fact use an isotopomer fraction, which is the ratio of isotopomer con-
centration to total metabolite concentration.
17.2 Experimental Techniques 563

shifts in typical NMR spectra. Measuring the amount of label at a given carbon on a
molecule distinguishes that molecule from among the ensemble of differently labeled
molecules. Care must be taken when determining concentrations from NMR spectra,
however, and various phenomena must be accounted for, including irradiation band-
width [29], incomplete relaxation [14], and irregular NOE [14] or DEPT [29] en-
hancement.
The absolute label concentration is calculated by referencing to a standard. Total
metabolite concentrations are determined by off-line analysis (enzymatic assay,
HPLC, etc.). When the total metabolite concentration is not known (e.g., in in vivo
experiments) only relative fractional enrichments can be measured. A relative frac-
tional enrichment is the ratio between the concentrations of a molecule labeled at
two different positions. However, relative fraction enrichments are not useful for
metabolites with only one labeled and observable carbon. For multiply labeled spe-
cies, one less relative fractional enrichment is available than the number of labeled
and observable carbons. Much potentially useful information is obviously lost when
metabolite concentrations are not known.

Determination by 1 H-NMR Spectroscopy

Purification of the metabolites by chromatography is usually necessary to discern

the peaks of interest in the IH-NMR spectrum. Proton peaks will be split by nearest
neighbor l3C atoms. The ratio of the satellite peaks to the total proton peak is the
fractional enrichment.

Mass Spectrometry

Portais et al. [14] also describe determination of the fractional enrichment of an

aspartate carbon not detectable by NMR from the enrichment of specific ionization
fragments.

17.2.2.2
Measuring Isotopomer Fractions

Determination of isotopomer fractions relies on another property of NMR spectro-

scopy. In a similar manner to proton splitting due to nearest l3C neighbors, l3C
peaks will split in proportion to the amount of neighboring l3C. Martin et al. [20]
describe the deconvolution of three glutamate peaks into isotopomer distributions.
An atom with one neighbor will appear as a triplet - the sum of a singlet and a
doublet. The singlet represents label at the home carbon alone and the doublet cor-
responds to label at both the home and neighboring carbon. An atom with two
neighbors will appear as a quintuplet, provided that the coupling constants for the
two pairs are different. Most of the isotopomers can be determined by quantifying
these peaks and satellites. However, carbons separated by more than one bond will
not affect each other. Therefore, there are isotopomers not discernible by 13C_NMR.
This ambiguity must be accounted for when using isotopomer data in models of
metabolism.
564 17 Analysis of Metabolic Fluxes in Mammalian Cells

17.2.2.3
In Vivo NMR

In vivo and extraction NMR are two methods by which intracellular isotope distri-
butions are determined. We define in vivo NMR as a technique whereby an isotope
distributed throughout the metabolites of central metabolism is detected in living
cells. In order for the cells to function normally, an experimental apparatus is
needed to provide the cells with a controlled environment. Two common approaches
are growth on micro spheres [30,31] and growth in hollow fiber bioreactors (HFBRs)
[23,24]. The microsphere method involves growing cells on agarose beads or some
other suitable matrix within standard NMR tubes supplied with oxygenated growth
media. The HFBR method similarly perfuses cells with oxygenated growth medium
through the lumina of numerous (greater than 1000) porous fibers of a tubular re-
actor. The cells are grown in the extracapillary space of the reactor; solutes diffuse
across the fiber. Reactors are scaled to fit within specially modified NMR probes.
An important advantage of in vivo NMR is that it can detect metabolites within
living cells. Because cells are metabolically active throughout the analysis, label con-
centrations are derived directly from the actual concentrations within the cell. In
addition, changes in the intracellular environment can be observed as they occur.
In vivo NMR is one of only a few techniques able to monitor molecular events in
living cells in real time. However, in addition to the complex apparatus required, it
is necessary to increase cellular density and overcome the limitations of low intra-
cellular metabolite concentrations. Another drawback is that only relative label frac-
tions can be detected, because pool sizes are difficult to measure by NMR. The in-
ability to detect absolute fractional enrichments limits the amount of data that can
be provided with this approach.

17.2.2.4
Extraction NMR

On the other hand, extraction NMR, defined as the separation of specific cellular
components from non-viable cells, is a much simpler procedure. It is possible to
extract metabolites from cells grown on numerous types of growth supports, as well
as from tissues excised from organisms [3]. The extraction method typically involves
lysing cells with an acid solution (often perchloric acid) that precipitates nucleic
acids, proteins, and triglycerides. The lysate is neutralized, lyophilized to remove
cellular water, and reconstituted in D2 0 for NMR analysis. Greater sensitivity can
be achieved in extraction samples because it is easier to use more cells, and because
metabolites contained within extracts can be concentrated by separation. A major
drawback of extraction, however, is the possibility that the extraction process will
affect the label distribution. This will, for example, occur if metabolites are con-
tained in different cellular compartments. Likewise, isotope distributions between
cellular compartments cannot be detected by in vivo NMR unless imaging techni-
ques are applied. An additional disadvantage of extraction NMR is that measure-
ments require the sacrificing of cells and thus only represent a single test condition.
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 565

17.2.3
Radio-Isotope Tracer Studies and Enzyme Activity Assays

Radio isotope tracer studies and enzyme activity assays are both classic methods of
investigating metabolic pathways. Both of these methods have been recently com-
bined in attempts to quantify whole cell metabolism [7, 12].
Radio isotope enriched substrates can be used to determine many cellular rates
including the rate of glucose membrane transport, the rate of glycolysis, the rate of
glucose carbon entering the TCA cycle, the rate of the pentose phosphate pathway,
the rate of pyruvate carbon entering the TCA cycle, and the rate of glutamine oxida-
tion [7, 12]. These assays are all based on time-course measurements of either 14C02
or 3H20 production. For each flux of interest, a substrate labeled at a specific carbon
with l4C is used. For example, the flux of glucose carbon entering the TCA cycle is
determined by using 6_ 14 C-glucose.
Neerman and Wagner [12] and Fitzpatrick et al. [7] described a method to deter-
mine the maximum activities of metabolic enzymes. This method first involves lys-
ing cells with specific extraction buffers to maintain select enzyme activities. The
activity can then be determined spectrophotometrically, at physiological conditions,
and in the presence of excess substrates. For this method, an extensive list of en-
zymes must be investigated, since it is through comparison of the activities and
comparison to the carbon flow that an understanding of cellular regulation can be
obtained.
By combing the results from both of these methods, enzymes with activities clo-
sest to the measured pathway flux are designated as rate controlling. It is proposed
that the flux through that pathway should be similar to the activity of the rate-limit-
ing step. Enzymes with activities higher than the measured flux are termed "open"
and have minimal control over the flow. However, making these designations as-
sumes that only one enzyme is rate-controlling. Experiments with recombinant bac-
teria, investigated by metabolic control analysis [8], have shown that the concept of a
single rate controlling enzyme in primary metabolism is not tenable.
For these types of analyses, it is assumed that CO 2 is mostly produced and con-
sumed in a few reactions; the evolution rate of 14C02 provides fluxes directly without
mathematical manipulation. If 14C results were incorporated into a flux model, this
assumption would not be needed, as more sites of CO 2 production could be consid-
ered. The coupling of 14C-tracer studies and enzyme extraction to other methods of
flux determination would no doubt give a more complete picture of metabolism and
regulation.

17.3
Mathematical Descriptions to Quantify Fluxes in Metabolic Models

The central metabolism of most cell types contains branched pathways that recom-
bine internally (e.g., the anaplerotic reactions, the pentose phosphate pathway, etc.),
which makes measuring the uptake and production rates of external metabolites in-
adequate to characterize internal metabolic fluxes. Two techniques described herein
to calculate metabolic fluxes, cometabolite measurement and 13C_NMR, overcome
this limitation. A metabolic model is then used to calculate internal fluxes from these
additional measurements. By using metabolic models, neither of these methods re-
566 17 Analysis of Metabolic Fluxes in Mammalian Cells

quires the measurement of each internal flux explicitly, which is necessary with the
14COz-evolution technique.
Measuring the production of cometabolites, i.e., metabolites produced or con-
sumed concurrently with other metabolites, provides additional constraints on the
set of internal fluxes. The number of cometabolite measurements needed to solve for
the fluxes in a given model can be found by assessing the degrees of freedom (DOF)
within the system:

degrees of freedom (DOF) = nunknown fluxes - nmetabolite balance equations

- nmeasured fluxes (17.4)

where nunknown fluxes is the number of fluxes (variables) in the model,

nmetabolite balance equations is the number of metabolites (equations) in the model, and
nmeasured fluxes is the number of measured external fluxes (constraints), including
those of cometabolites. To obtain a unique set of fluxes, the degrees of freedom must
be greater than or equal to zero. Since nunknown fluxes and nmetabolite balance equations are
both defined by the system, the minimum number of measurements that have to be
made, nmeasured fluxes> is found when DOF=O in Eq. (17.4). When there are no degrees
of freedom (i.e., DOF=O) the system of equations is said to be "exactly determined:'
If no branches existed, then the number of unknown fluxes would equal the number
of balances, and no cometabolite measurements would be necessary. Note that any
measured flux could be considered another equation and combined with metabolite
balances to form a group of "defining equations:' This distinction is made to draw
attention to the fact that one set of equations is inherent in the structure of the pre-
sumed pathway model and the other requires experimental measurement.
The 13C_NMR technique provides additional information in the form of fractional
enrichments. However, using fractional enrichments requires the inclusion of iso-
tope or isotopomer balances, which adds unknown fractional enrichment variables
to the calculation of the system's degrees of freedom:

DOF = llunknown fluxes + llunknown fractional enrichments - nmetabolite balances

- nisotope balances-llmeasured fluxes - llmeasured fractional enrichments (17.5)

Models based on tracer experiments inherently introduce non-linear equations in

the form of isotope label balances, whereas cometabolite measurements do not.
Thus, there is a clear association between linear models and cometabolite measure-
ment, and non-linear models and tracer studies.
As an example, consider the sample pathway model in Fig. 17.l. It contains 25
unknown fluxes, 11 metabolite balances, and 7 measurable external fluxes, leaving
7 degrees of freedom. This implies that a minimum of seven fractional enrichments
is necessary to close the system and render it determinable.

17.3.1
Determining Fluxes Using Cometabolite Measurements

Cometabolites are metabolites produced or consumed concurrently with other meta-

bolites within cyclic reactions. How they can be used to determine intracellular
fluxes is illustrated by the following example of Bonarius et al. [5] (Fig. 17.2). If only
the external consumption and production rates of S, A, and Bare measured (exam-
ple I), internal fluxes f2 - 4 cannot be uniquely determined. However, if the rate of
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 567

01-02-03-04-05-06

~
C~

~
f1
G l-G2-03-04-G5-G6-P Rul-Ru2-Ru3-Ru4-Ru5-P
f3
n +
FI ~~r-F4-F5-F6P

Cytoplasm P-Fl·F2-F3-F4-F5-F6-P

f51\
OAI-GA2-GA3-P

[Ser,Gly]
"!
PGJ-PG2-P-P03
Fatty Acids
+ fl3

f6 + A2·AI·CoA

ALI-AL2-AL3~ \P:~-IPE2-P-PE3C92
\
tt
01-02T-04 ~f13
~ t '- no C5-C4-C3-C2·Cl
L1-L2-L3 ... PI-P2-P3 ..
________________ ___________
~~ 1________________ Ml-M2-M3-M4
_
fl9 ,
_ - - - PI·P2-P3
[fatty ~C
+amino acids] , - ~
~A2-AI-COA
n4~

f
I
01-02-03-04 f14 C5-C4-C3-C2-CI

~
C~ Priteins

f23
f16 nI Tt m f25
f9 Kl-K2-K3-K4-K5~Tl-T2-T3-T4-T5-!>

Ml-M2M3-M4 TCA cycle tfl5

I
t n6
[Arg,His,Prol

NI-N2-N3-N4-N5

Mitochondria f17

fl8

Fig. 17.1. Example pathway model. Abbreviations: A, acetyl-CoA; AL, alanine; C, citrate; F,
fructose; G, glucose; GA, glyceraldehyde; K, a-ketoglutarate; L, lactate; M, malate; N, gluta-
mine; 0, oxaloacetate; P, pyruvate; PE, phosphoenolpyruvate; PG, phosphoglycerate; Ru, ribu-
lose; T, glutamate

appearance of cometabolite C is also measured (example II), a unique solution for

the internal fluxes can be found.
The cometabolites measured by Bonarius et al. [5] and Zupke and Stephanopoulos
[6] were NADH, NADPH, O2 , CO 2 , and NH 3 • Due to the unmeasurable activity of
568 17 Analysis of Metabolic Fluxes in Mammalian Cells

I
s s

Fig. 17.2. Cometabolite measurement can determine intracellular fluxes

transhydrogenase, Bonarius et al. [5] lumped NADH and NADPH into one balance.
Alternately, Zupke and Stephanopoulos [6] considered them as separate pools.
Neither NADPH nor NADH is easily measured directly, but since all of the major
pathways utilizing them are known, their balances are closed by the oxygen uptake
rate (OUR).
Cometabolites are metabolites that participate in reaction pathways with multiple
products or substrates. For our purposes, the cometabolites most useful to measure
are those produced or consumed within cyclic pathways, that have accurately mea-
surable fluxes, and that have closeable mass balances. For a cometabolite to have a
closeable balance, all of the major fluxes that produce or consume the cometabolite
must be included in the pathway model. Because Bonarius et al. [5] believed that
there are too many biochemical reactions utilizing ATP to close its balance, these
authors concluded that an ATP balance could not be used. In contrast, Zupke and
Stephanopoulos [6] lumped all generic ATP utilization reactions into the biomass
reaction, thus closing the balance.

17.3.1.1
Solution of the Stoichiometric Matrix

The linear model used to evaluate cometabolite measurements is made up of the

three components in Eq. (17.4): unknown fluxes, metabolite balances, and measured
external fluxes. The metabolite balances relate the unknown fluxes to the external
fluxes [5]:

(17.6)

where rA is the rate of change of metabolite A. The right hand side is the sum of the
contributions of each pathway flux, fj, to the flow through metabolite A. The reaction
coefficients, CXj,A' for pathway j and metabolite A have negative values if A is a sub-
strate and positive values if A is a product. Typically, they have values of 1 or -1
although some reactions introduce fractional and integer values. Rates of change
for excreted and absorbed metabolites are synonymous with fi,ex in Eqs. (17.1)-
(17.3). For internal metabolites, the rates of change are assumed to be zero because
of the pseudo-steady-state approximation (PSSA) [6].
Zupke and Stephanopoulos [6] showed that the PPSA is valid if the rate of change
of a given metabolite's concentration is small relative to the flux through the meta-
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 569

bolite. To illustrate this concept, consider internal metabolite M, which has flows in
and out, fin and fout:

The differential form of the metabolite balance (Eq. 17.6) around M is given by
(17.7)
where V is the intracellular volume (assumed constant) and eM is the intracellular
concentration of M. An arbitrary tolerance, 0, can be defined and used along with
the following criterion to determine whether the PSSA is valid:
(17.8)
The right hand side of Eq. (17.8) is the average flow through M multiplied by the
tolerance, 0. Zupke and Stephanopoulos [6] estimated the cell volume (V) and aver-
age metabolite flux (fin and fout ) for CRL-1606 hybridomas to be 9x 10- 13 l/cell and
_10- 10 mmol/cell/h, respectively. These values indicate that dCM/dt must be less than
10 mmol/l/h to satisfy the PSSA, if a tolerance, 0, of 10% is assumed. Most cellular
concentrations are in the range 0-10 mmol/l with an average being approximately
1 mmol/l, which implies that the PSSA is generally true. If a metabolite has a con-
centration of 10 mmol/l and a rate of change of 10 mmol/l/h then that pool is chan-
ging from completely "empty" to "full" in the course of an hour. It is difficult to
envisage a metabolite pool that a cell would completely cycle in an hour in the ab-
sence of an environmental stimulus. Given this, V(dCM/dt):::;oO, and Eq. (17.7) is re-
duced to fin=fout.
Equation (17.6) can be written in matrix form to express the entire system of equa-
tions. Solution involves the inversion of the sparse coefficient matrix ~ [5, 6]:

~f=! (17.9)

(17.10)
where ~ is the stoichiometric matrix containing the reaction coefficients, Uj,i' Each
row in A represents one metabolite balance; there is a column for each flux, and f
and! are both column vectors containing all of the variable fluxes and the measured
extracellular production/consumption rates, respectively. An exactly determinable
system will have a square stoichiometric matrix, A.
If A can be inverted, then the vector f can be obtained exactly. However, the inver-
sion of large sparse matrices has been the focus of much study and is not necessarily
straightforward. Therefore, the method ofleast-squares fitting is often employed to
estimate f. Various computational methods [32], including steepest descent, Gauss-
Newton, Marquardt'S method, etc., can be used to find the values of f that minimize
the weighted sum of squares, F:

(17.11)

The squares of the deviations are weighted by the variance, o,?, of the measure-
ments, rio
570 17 Analysis of Metabolic Fluxes in Mammalian Cells

17.3.1.2
The Objective Function

If fewer measurements are made than the minimum required to solve the system of
equations for the assumed biochemistry, the system is underdetermined. To solve
such an underdetermined system, Bonarius et al. [5, 33] applied an objective func-
tion [34, 35] to their stoichiometric matrix. Examples of objective functions include
maximization of ATP production [36], maximization of NAD(P)H production [36],
maximization of biosynthesis [36], and maximization of total cellular flux [33].
The appropriateness of such objective functions depends on the subsequent use of
the resultant fluxes. If they are used for hypothesis testing of cells in differing states,
care must be taken not to incorporate inadvertently the expected results into the
objective function. For example, if the effect of different glutamine feed concentra-
tions on the transport of NADH is being tested, it does not make sense to minimize
or maximize the production of NADH. For cancerous cell lines, it may not be rea-
sonable to claim that the cells are performing optimally for any criterion. If cell
functions were optimal, then the cells would either not be cancerous or they would
behave like bacteria. Clearly, mammalian cells reproduce more slowly and have
slower metabolic rates than bacteria because of certain advantages this might pro-
vide the organism. The inclusion of any teleological argument potentially biases the
outcome of any analysis or investigation. Is it possible to know how the purpose of a
cancerous cell's existence differs from the purpose of a normal cell? Ideally, a mea-
surement should report the state of a system without such assumptions.

17.3.2
General Principles of Isotope Balancing

17.3.2.1
Steady State Flux Analysis

In many ways 13C-NMR is the most powerful of the techniques reviewed here. Steady
state flux analysis is often unjustly criticized by claims that fluxes cannot be com-
puted from non-transient fractional enrichments, since the relationship between
steady state isotope distribution and pathway flux is not directly obvious. The simple
illustration taken from Wiechert and de Graaf [22] in Fig. 17.3 demonstrates this
relationship.

Fig. 17.3. Use of fractional enrichment data for flux determination. Sand P are a consumed
nutrient and an excreted product respectively; A and B are intracellular metabolites. The circles
represent carbons; filled ones are isotopically labeled. The carbons of product P are partially
labeled
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 571

External measurement alone can only determine the consumption flux fl' which is
equal to the production flux f4 at steady state. Fluxes f2 and f3 remain unknown.
However, reactions f2 and f3 scramble the carbons of S when producing the internal
metabolites A and B. If the label distribution of P were measured, then fluxes f2 and
f3 could be found.
The ratio of label in P directly relates to the ratio of the internal fluxes:
(17.12)

where PI and P2 are the fractional enrichments of P on the first and second carbon,
respectively. Notice that only relative flux information can be directly determined
from fractional enrichment data. To find the absolute fluxes f2 and f3 it is necessary
to measure also fl (re-emphasizing the statement that the combination of techniques
is necessary for complete analysis). In this example the fractional enrichments, PI
and P2, sum to one. Combining this relationship with the defining material balance
around S, fl=f2+f3' and Eq. (17.12) produces:
(17.13)

(17.14)

This illustrates how the numerous branched pathways of central metabolism ne-
cessitate the use of techniques beyond direct excretion measurements. Because they
recombine internally, the relative fluxes through many branches (such as f2 and f3 in
Fig. 17.3) are not discernible by measured external fluxes. Evaluation of the label
distribution following labeled nutrient infusion, as demonstrated, has the ability to
determine these relative fluxes. While writing Eq. (17.13) from Fig. 17.3 is intuitive, a
more formalized approach is needed for any realistic and thus more complex model.

17.3.2.2
The Isotope Balance

The construction of a useful mathematical description requires that the purpose of

the model and its assumptions be well defined. Clearly, the goal is quantification of
fluxes in a whole cell metabolic model. Here the initial information is a set of exter-
nal fluxes and enrichment fractions. The external fluxes relate directly to the internal
fluxes as seen in the discussion of cometabolite utilization and Fig. 17.3. The isotope
balance is used to interpret the enrichment data.
Figure 17.4 shows two pathways that both produce metabolite C from metabolites
A and B through reactions 1 and 2. Metabolite C is consumed by reaction 3.
In this example, the first carbons of A and C are both partially labeled and the first
carbon of B is not. Intuitively, because of the dilution contributed by reaction 2, the
first carbon of C is less enriched than the first carbon of A. The fractional enrich-
ment of A on the first carbon, MAl' is defined as
conc. A labeled at position 1
MAl == - - - - - - - - , - - - - = - - - - (17.15)
total conc. A
which can be rearranged as

MAl CA=LAI (17.16)

572 17 Analysis of Metabolic Fluxes in Mammalian Cells

Fig. 17.4. An example to demonstrate fractional enrichment, label flux, and the label balance.
Each circle represents carbon atoms in one of three molecules: A, Band C. The darkened circles
are partially enriched by l3 C. Three-carbon molecules, of which there are only a few in central
metabolism, were used to simplify the notation. The three carbon atoms of each molecule are
numbered 1-3 starting at the top. This particular set of reactions represents the recombination
of two pathways. Many of the metabolites in Fig.17.I are sites where pathways unite. Two
examples of such sets of reactions are the combining of pyruvate and malate carbons in oxa-
loacetate through pyruvate carboxylase and malate dehydrogenase, and the combining of glu-
tamate and citrate carbons in a-ketoglutarate through isocitrate dehydrogenase and glutamate
dehydrogenase

where CA and LAI are the total intracellular concentration of A and the intracellular
concentration of A labeled at position 1, respectively. The absolute labeling, LAl> is
measured by l3C-NMR. The fractional enrichment, MAl> is detected as split peaks in
lHNMR following l3C infusion (see Sect.I7.2.2.1). Taking the derivative of both
sides with respect to time and dividing by the cell number shows how fractional
enrichments relate metabolite fluxes, fA~c, to label fluxes, AAl~Cl' Note that fluxes
are by definition directional; hence, the designation of starting, AI, and ending, Cl,
positions:

(17.17)

The initial and final metabolite notation can be dropped and substituted with the
numeric flux notation of Fig. 17.4 with the following consideration: The flow of iso-
tope through a unidirectional pathway is relative to the fractional enrichment of the
initial and not the final metabolite, because of the effect of diluting flows into the
product. For this reason it is also necessary to define the flux, fA~C' relative to the
rate of change of the initial metabolite concentration. The numbered fluxes are nor-
mally defined as the specific rate of metabolite production, thus requiring the iso-
tope balance to be scaled by the stoichiometric coefficient, alA:

(17.18)

Here alA is defined similarly to Eq. (17.6); it is the ratio of the moles of products
formed to the moles of reactant consumed. There is only one flux, f5' in Fig. 17.1 for
which alA does not equal 1: two glyceraldehyde-3Ps are produced from each fruc-
tose-I,6biP.
Equations(17.I6) and (17.17) demonstrate, albeit simply, how fractional enrich-
ments relate both metabolite fluxes to label fluxes and metabolite concentrations to
label distributions. This relationship permits the development of additional equa-
tions to reduce the degrees offreedom of the system (Eqs.I7.4 and 17.5) when frac-
tional enrichment data are available.
Just as the mass of metabolites and carbon must be conserved, so must the mass
of label. In Fig. 17.4 the total flow oflabel into C must equal the flow out:
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 573

(17.19)
(It is not necessary to define an end-point for the out flow of label. The in flows
must balance all of the out flows.) Using Eq. (17.18), Eq. (17.19) can be written in the
standard form:
(17.20)
Here we write MAl as Al to simplify the subsequent equations. In the example in
Fig. 17.4, two more balances can be written, one for each of the three carbon atoms
in C.
There are 44 label balances in the model in Fig. 17.1 if only atoms through which
label can pass after 1_!3C glucose infusion are considered. Note that these equations
must contain the manner by which metabolic enzymes transfer carbons from meta-
bolite to metabolite. Historical isotope (mostly 14C) experiments determined the car-
bon transition patterns for all of the reactions in Fig. 17.1.

17.3.3
Least Squares Fitting of the Algebraic Form

Portais et al. [14] fitted their data to the complete set of label balances. The pathway
model used by these authors was simpler than the model in Fig. 17.1; they reduced
the number of label balances from 44 to 24. Many of the metabolites in the complete
set of label balances are present at low levels within the cell, rendering their frac-
tional enrichments difficult to measure. Therefore, these unknown fractional enrich-
ments do not reduce the degrees of freedom in the system. In the approach of these
authors, unmeasured enrichment fractions were left as variables, and the fitting al-
gorithm was used to determine their values.
The procedure used by Portais [14] to solve for pathway fluxes can be summarized
by the following steps:
1. Generate "differential equations describing time dependent variation" [14] of me-
tabolites and !3C labeled species. These equations are the same as the metabolite
balances in the linear model description. Initially the accumulation term was re-
tained so that transient data could potentially be incorporated
2. Apply the PSS assumption (Sect. 17.3.1.1), which sets all time derivative (accumu-
lation) terms to zero. All of the differential equations generated in step 1 are now
reduced to algebraic equations and are identical to metabolite and label balances
3. Use least squares fitting to "fit" the fluxes to the label fraction data
4. Calculate errors for the fluxes by the method of support planes as described in
Sect. 17.3.7.2.

17.3.4
Atom MappingITransition Matrices

Zupke and Stephanopoulos [37] provide an alternative approach. The first step in
this method is to arrange the enrichments of all atoms of each metabolite into vector
form. The vectors have the same number of elements as each metabolite has carbons.
In vector form a label balance is written for each metabolite rather than each atom.
Any changes to the model can be made by rearranging the enrichment vectors in the
matrix label balance equations. The use of vector notation reduces the number of
574 17 Analysis of Metabolic Fluxes in Mammalian Cells

A2-AI-CoA
\ C6 /f13
If Co..
-L
'-!14 I
01-02-03-04 ----""'~--II.~ C5-C4-C3-C2-Cl
f l6 J
~

Fig. 17.S. Example of atom mapping matrices: balance around citrate extracted from Fig. 17.1

equations needed to describe the label flow in the model in Fig. 17.2 from 44 to 16.
The information that is lost in reducing the number of equations is how the atoms
move from one metabolite to another via enzyme activity. Individual label balances
necessarily contain transition information. The form proposed by Zupke and Stepha-
nopoulos compensates for this by the inclusion of atom mapping matrices (AMMs)
[37].
AMMs describe the transfer of carbons from reactants to products and are gener-
ated using knowledge of specific biochemical pathways and enzyme reactions.
AMMs are constructed such that the multiplication of the AMM by the reactant en-
richment vector produces the product enrichment vector. For any reaction there is
an AMM for each reactant-product pair. An example is provided by the balance
around citrate in Fig. 17.1 as shown in Fig. 17.5.
For this example, there are three enrichment vectors:
Cl
C2
C3
Cit = (17.21)
C4
CS
C6
Since there are two reactants and one product, there are two AMMs for this reac-
tion:
0 0 0 0 0
0 0 0 0 0 1
0 0 0 0 0
[OAA > Cith4= [AcCoA > Cith4= (17.22)
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 0
Here the subscript 14 refers to flux f14, the reaction of interest. Steady state isotope
balances equations are formulated as the "sum of the products of the mapping ma-
trices and the reactant (enrichment) vectors, weighted by the corresponding reaction
flux" [37]. In the example in Fig. 17.5 the label balance around citrate is
[fl3 + f16 ] Cit =f14 [OAA > Cit] 14 OAA + fl4 [AcCoA > Cith4 AcCoA (17.23)
An equation in the form of Eq. (17.23) can be written for each metabolite in the
pathway.
Zupke and Stephanopoulos [6] applied this method to data from a hybridoma
culture. They formulated a considerably simplified model pathway which contained
only two branches. Two flux fractions were defined to describe the flow through each
of the branches from the branch points. The hypothetical measurements were two
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 575

relative fractional enrichments of lactate (Ll/L2 and L3/Ll). The isotope distribu-
tions for all possible flux fractions (from 0 to 1) were solved iteratively using the
Gauss-Seidel method.
This is a reverse solution strategy; fluxes are provided as input data and the iso-
tope distribution calculated. A graphical method was described to determine the
fluxes from isotope data. The solution space (both relative enrichment fractions)
was represented as overlaid contours on a plot with both flux fractions as axes. The
intersection of the label ratio contour lines (determined experimentally) fixed both
flux fractions. For the previous example there were two parameters and two solu-
tions (the relative isotope distributions), so it was possible to analyze graphically.
However, visualizing of more than three-dimensional space is difficult; this is the
major drawback of using a reverse solution and graphical analysis.
It is not possible to use the AMM notation and determine fluxes from a label
distribution without incorporating additional equations. Within each enrichment
vector there can be both measured fractional enrichments and unknowns. In the
approach of Portais et aI., unmeasured fractional enrichments remained vari-
ables, and measured enrichments were used to compute the sum of errors
squared. Any enrichment vector that is entirely measured or unmeasured can
be treated in this fashion. However, any vector composed of both would have
to be separated into its scalar elements. This effectively eliminates the elegance
of the AMM approach.

17.3.5
Isotopomer Mapping Matrices

Given the ability to experimentally distinguish metabolite isotopomers, the inclusion

of isotopomer distribution information into the mathematical description can more
accurately characterize the fluxes than fractional enrichments alone. However, nu-
merical manipulation of isotopomer information is cumbersome because of the ex-
ponential increase in the number of system variables; there are zn isotopomer states
per metabolite containing n carbon atoms as opposed to n fractional enrichment
states.
In contrast to fractional enrichment analysis, isotopomer distribution analysis
must be viewed within an entire-metabolite framework. Balances cannot be written
around each atom, since the state of neighboring atoms must be considered. And as
with label balances, any set of balance equations would contain many more isotopo-
mers than can be detected. A conservative estimate yields about fifteen detectable
isotopomers in a mammalian cell system (it would be much higher for any bacterial
system excreting detectable product) whereas there are over 600 total isotopomers in
Fig. 17.1. In order to handle this computationally daunting problem a method simi-
lar to the atom mapping matrices has been employed [38]. The problem would be
greatly simplified if measured isotopomers were deconvolute from unknown isoto-
pomers to remove excess isotopomers from the problem.
Schmidt et al. [38] introduce a notation to describe the relation between isoto-
pomers. The distribution of isotopomers in a metabolite is contained in isotopo-
mer distribution vectors (IDVs) that, "in contrast to (enrichment) vectors, do not
contain fractional enrichments at individual carbon atom positions but rather
mole fractions of metabolite molecules that are labeled in a specific pattern
[38]:' Each isotopomer of a metabolite is represented by a binary number of
576 17 Analysis of Metabolic Fluxes in Mammalian Cells

length equal to its number of carbons. The labeled and unlabeled atoms are re-
presented by ones and zeros respectively. When converted to a decimal number,
each of the 2ll possible isotopomer is represented. Equation (17.24) is Schmidt's
notation applied to acetyl-CoA (aca) in Fig. 17.5, which has two carbons and four
possible isotopomers, is

(17.24)

The "bin" notation signifies that the preceding number is a binary number. Each
element of an IDV is an isotopomer mole fraction, all of which sum to one:

L Iaca(i) = 1
3
(17.25)
i=O

As with enrichment vectors, a formulation is necessary that describes the car-

bon transitions between the IDVs, in order to incorporate them into the metabo-
lite model. For this purpose, Schmidt generated isotopomer mapping matrices
(IMMs). Given a specific labeling pattern of reactants and knowledge of how car-
bons transition from the reactants to the products for the given reaction, the dis-
tribution of label in the products is determinable. Contrary to fractional enrich-
ment transitions, it is possible for multiple substrate isotopomers to result in one
single product. Schmidt pointed out that isotopomer mole fractions could be
viewed as probabilities. In a reaction, the probability of one isotopomer form of
a reactant combining with one isotopomer form of another reactant is equal to
the product of their mole fractions. In Fig. 17.5 unlabeled acetyl-CoA and OAA
would produce unlabeled citrate:

(17.26)

Acetyl-CoA labeled at the second position and OAA label at the second and fourth
would produce citrate labeled at the second, third, and fifth carbon:

(17.27)

The large number of possible isotope transition equations can be reduced by ex-
pressing them in matrix form. As with AMMs, columns and rows represent substrate
and product labeling patterns, respectively. For reactions with multiple reactants and
products, an IMM is needed for each substrate-product pair. An IMM for a reaction
can be calculated if the AMM is known by systematically generating each possible
reactant and determining the products. The IMMs are sparse matrices usually con-
taining ones and zeros.
Reactions with single substrates are expressed mathematically in a manner similar
to label balances written using AMMs. Figure 17.6 is pyruvate dehydrogenase [fd
from Fig. 17.l.
The mv for acetyl-CoA derived from pyruvate is given by

Iaca = IMMpyr> aca Ipyr (17.28)

17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 577

c~

PI-P2-P3
f12 ~ f14
~ A2-AI-CoA--'-'--"~

fatty aCids~~
Fig. 17.6. IMM example: pyruvate dehydrogenase example from Fig. 17.1

where
o o 0 1 o 0
o 0 o 1 0
(17.29)
o o o o 1
o o o o 0
The complete isotopomer balance around acetyl-CoA is therefore
(17.30)

where 0 is a null vector with four elements; it assumes no label is present in fatty
acids.
For multiple substrate reactions, the operator @ must be introduced:
(17.31)

where Schmidt et al. define the @ operation as the "elementwise multiplication of

two equally long column vectors [38)." The IMM-IDV product for a single sub-
strate-product pair does not contain information about any other substrate. The ele-
mentwise multiplication of the IMM-IDV products scales the probability for each
pattern of the product appropriately. The isotopomer balance around citrate in
Fig. 17.5 is
(17.32)
An isotopomer balance, like Eqs. (17.30) and (17.32), can be written around each
metabolite and can be used in a manner similar to the balances written using AMMs.
Schmidt's implementation used fluxes as input data and solved for isotopomer distri-
butions in a manner similar to Zupke and Stephanopoulos's [6] implementation of
AMMs. When NMR is used to measure the isotopomer distribution, the problem is
reversed: metabolite fluxes must be calculated from the isotopomer balances. The
IDVs of metabolites with greater than three carbons cannot be completely determined
by NMR, since only nearest neighbors can be detected, e.g., a three carbon metabolite
labeled on the first and third carbon can not be distinguished from those labeled on
the first and third alone. Since no IDV can be completely determined, additional equa-
tions relating the known scalar components are necessary. This is a similar limitation
to the use of AMM notation. However, in the isotopomer case where a large number of
individual balances would be needed to describe the system, the use of matrix nota-
tion is a great simplification. While the addition of equations relating individual vec-
tor elements diminishes the elegance of a system entirely represented by matrices,
their addition outweighs the cumbersome nature and potential inaccuracy of using
individual isotope balances (e.g., the equations of Malloy et al. [19]).
578 17 Analysis of Metabolic Fluxes in Mammalian Cells

17.3.6
Transient NMR Measurement

A classic approach to flux measurement using isotopic labeling is the detection of

the transient appearance of the isotope throughout metabolic pathways. To obtain
meaningful results from the transient response necessitates information be provided
about the kinetics of the enzymes and/or the pool sizes of the metabolites involved
in the reactions. Both of these quantities can be elusive. The kinetics of enzymes
may change once removed from the cell; physiological levels of product inhibition
and allosteric effectors can be difficult to simulate. Pool sizes are difficult to measure
because metabolites are present in small concentrations and are distributed among
various cellular compartments. Many methods have been employed to overcome
these limitations, including direct measurement of pool sizes [3, 18,21], and compi-
lation ofliterature values for both pool sizes [4, 18] and kinetic constants [18].

17.3.7
Errors in the Determination of Fluxes

The process of flux determination discussed in the preceding section needs to be

complemented by analysis of experimental errors. Knowledge of the error distribu-
tion for the resultant fluxes has four uses:
1. The accuracy of the flux results can be evaluated.
2. The value of alternative pathway models can be discerned based on predetermined
physiological boundaries. Many models produce physiologically unreasonable re-
sults. Given that a physiological range of a flux can be defined (for most fluxes it is
simply the requirement that they be positive), the errors can be directly calculated.
From the set of probabilities for all the fluxes in a model, the probability that the
model produces physiological meaningful results can be assessed.
3. The sensitivities of each flux to each measurement can be established. Sensitivities
are useful in experiment design. The measurements with the greatest influence on
the results must be performed the most accurately, or repeated to increase confi-
dence.
4. A statistical significance can be assigned to observed changes in flux results be-
tween different test conditions. One means of calculating significance is the stu-
dents-t test.

17.3.7.1
Errors in Linear Models

Random errors in experimental technique or gross errors introduced by incorrectly

formulated pathways can only be detected in linear models if redundant measure-
ments are made [6]. A redundant system is over-determined, having more measure-
ments than the minimum required for the assumed biochemical pathways.
Within any organism there are more than 500 possible fluxes to consider. While it
is imperative that this number be reduced to the number of measurements to find a
solution, additional fluxes must be eliminated to allow for redundant measurements.
In doing so, the advantages of error determination are offset by the disadvantage
that fluxes contributing to the carbon flow are eliminated, making calculations less
17.3 Mathematical Descriptions to Quantify Fluxes in Metabolic Models 579

accurate. For the purposes of error analysis, Zupke and Stephanopoulos [6, 37]
eliminated both the PPP and the anaplerotic reactions. Both of these pathways have
been shown to be active in most cell lines and are critical to metabolic regulation.
However, the flow of carbon through both is typically small.
With more system-defining equations than unknown fluxes, the stoichiometric
matrix A is not square; it has more rows than columns. Gaussian elimination of A
produces the permutation matrix £ which is defined such that -

(17.33)

where! is the identity and.@. is the null matrix. The flux vector f can then be deter-
mined in the presence of redundant measurements in a manner similar to the deter-
mination of f by Eqs. (17.9) and (17.10). Both sides of Eq. (17.9) are multiplied by ~
to generate Eq. (17.34): -

(17.34)

Equation (17.34) can be partitioned into two matrix equations; the first is used to
determine f with the square upper portion (~) of £:
(17.35)

Matrix ~ can be considered equivalent to A-I in Eq. (17.10). The second portion of
Eq. (17.34) defines the redundancy matrix f
~!=O (17.36)

The number of rows in ~ is equal to the number of redundant measurements (also

the difference between the number of rows and columns in A). The product ~! re-
presents a series of equations that relate how various measurements, ri, scaled by the
coefficients of~, should sum to zero. Experimental errors and incorrectly formulated
models shift the results of each redundancy equation in the product ~! from this
optimal result. -

17.3.7.2
Errors in Non-linear Models

Either Monte Carlo simulations or the method of support planes can be used to
determine the error in non-linear models. Both of these techniques are described
by Chandler et al. [32] and Duggleby [39]. The Monte Carlo method involves the
following steps:
1. The set of best-fit fluxes (f) are calculated from the input data by minimization of
the sum of squares.
2. Random data is generated from the measured values and their experimental er-
rors. If the measured data set is assumed to be representative of all possible data
sets, a random data set would be contained within a Gaussian or other distribution
about the measured set.
3. The flux parameters are fit to the random data, using the best-fit fluxes (f) as
initial guesses.
580 17 Analysis of Metabolic Fluxes in Mammalian Cells

4. Steps 2 and 3 are repeated M times. The more times the simulation is repeated the
more accurate the calculated errors.
5. The standard error (Ei) of a particular flux parameter (0 is the root mean square
of its deviations from best-fit:

(17.37)

When the parameter space is flat around the minimum sum-of-squares in the di-
rection of one parameter, the error in that parameter is large. Support planes, as
described by Duggleby [39], are the positive and negative values of a given para-
meter (flux) which cause the minimum sum-of-squares value to increase by a factor
(1 + lIDOF), where DOF is the number of degrees of freedom in the metabolite and
label balance equations (an exactly determined model has zero degrees of freedom,
so support planes cannot be used). For linear systems, the (1 + lIDOF) factor exactly
determines the standard error. For non-linear systems (especially those that are lin-
ear in a parameter near the minimum) it is a reasonable approximation.
The method used to find the support planes is an iterative process. The chosen
parameter is constrained to a guessed value. The remaining parameters are opti-
mized to find the new minimum sum-of-squares. New guesses are made, and the
process is repeated until the sum-of-squares equals (1 + 1IDOF) times the original
minimum sum-of-squares.

17.4
Biochemical Pathway Model Formulation and Reduction

A comprehensive model of primary and secondary metabolism of mammalian cells

is depicted in Fig. 17.7. All of the pathways considered in this review are included.
Also, many of the pathways connected to common metabolites were included based
on known biochemistry [1,40]. It is necessary to eliminate many of these pathways
to produce models useful for flux analysis.
Two assumptions were made to develop the model in Fig. 17.7:
1. All of the major pathways of central metabolism are known. This implies that the
pathways have been identified in the literature or detected by direct measurement.
Exclusion of major pathways significantly producing or consuming a given meta-
bolite would invalidate the metabolite balance.
2. The pseudo steady-state approximation applies for each metabolite. This implies
that the flow of total carbon (or any element) in is equal to the flow of carbon (or
other element) out. Zupke and Stephanopoulos [6] make a very reasonable argu-
ment substantiating the PSSA, discussed in Sect. 17.3.1.1.

17.4.1
Reduction of Comprehensive Models

The "comprehensive" model of Fig. 17.7 cannot be used to convert l3C-tracer and co-
metabolite data into intracellular fluxes; it contains 98 fluxes and 45 metabolites, only 39
of which have closed mass balances. As such the system is under-determined and would
require 59 additional measurements to solve for all intracellular unknowns. The deter-
17.4 Biochemical Pathway Model Formulation and Reduction 581

---:--I> Gl-GZ-G3-G4-G5-G6
fl f3 f2 ~ CC2
Total Carbohydrate_ GI-G2-G3-G4-G5-G6-P L .. Rul-Ru2-Ru3-Ru4-Ru5-P

f4li f5 fI1 flY"{13

FI-Fl-F3-F4-P5-F6P ~;Z;;~ XI-X2-X3-X4-X5-P RI-R2-R3-R4-R5-P

Cytoplasm
f6! ~14
P-FI-Fl-F3-F4-F5-F6-P [RNA.DNA]

P-Dl-D~G:1~A3-P~~~~~~~::
. flO f21 It f22 fI9

It
Protems ..,1
1.3 DPG f25 f20
f31 f32 f2Jtf24 -.......:.:::;; 2.3 DPG Proteins ~ Asparagine
f30
~ Serine
f33
~PGI-PG2-PG3-P
tl -;26 f77 Jif7~ FatllY:;idS

r;======~===~~=====Z Aspartate =--------t>

f28 f68 It f76 A2-AI-CoA

f38 Ji
Proteins
f37
f75 "'I f69
01-02-03-04
~

<r---lH-ALl-AL2-AL3 ~ \ f29! C~ f661t f67 [6 f72

f39 ~Pl P2 P3
E f40 f36 0( "-.174 MI-M2)3-M4 C5-C4-C3-C2-Cl
f4: __ ~~-~:-~~_ ~::~~ ____ --!--~ft42 ____________ _
" PI-P2-P3
:' C r;.;- ~44
COz ~ +ammo aCIds] - - CC2~
f62 f63 f46
"--I>
f47~
A2-A I-CoA
f48
f60
Aspartate ~+==:;01-02-03-04 f48

"~f "':\~ f83

MI-M2 M3-M4 KI-K:~~-K4~~6 f84 : f88) Kl-=:tT~:4-:7

, o~ CM~,,~'"fL:'TITIT<"< ! ":'7i"j-1T3~:lf~
t. Mitochondria
,
FUl-FU2-FU3-FU4
(/ ~~-S3-S4
;:I f5'
"
J Pr~!£f92
f93 f94
[Arg.His.Pro]

\, / f5 [I1e. Val. Thr] , f98

'. _____ [:~~~h:~ _~~5_____ ~5_4____________________________ / NI-N2-N3-N4-N5 <r-

Fig. 17.7. Comprehensive model of all pathways reviewed. Abbreviations in addition to

Fig. 17.l: D, dihydroxyacetone phosphate; DPG, diphosphglycerate; E, erythose; FU, fumarate;
R, ribose; S, succinate; Se, sedoheptulose; X, xylulose

mination of cometabolite production rates, fractional enrichments, and isotopomers

can only provide 18 additional measurements (the number is dependent on the exten-
siveness of the experiment). The intractability of this system makes it necessary to re-
duce the number of fluxes in the model. This is done by deciding which pathways are
active in the particular cell line, and eliminating those with only small contributions.
582 17 Analysis of Metabolic Fluxes in Mammalian Cells

The validity of various assumptions can be tested. Particular sets of data can be
analyzed using a number of reduced models. For a given cell line and culture condi-
tions, "physiologically reasonable" ranges for intracellular fluxes can be defined
(e.g., flux through an irreversible reaction cannot have a negative value; the rate of
the TCA cycle cannot be orders of magnitude larger than glycolysis, etc.). Simula-
tions can be used to determine the probability that a given set of data and a given
model produce a realistic physiological result.
The creation of a pathway model is not usually discussed from the standpoint
of reducing a comprehensive model. A model containing all of the pathways
known to exist in a mammalian cell would be quite daunting. Models are con-
structed in degrees of increasing complexity. The usual reasons for including
pathways include:
1. Pathways are included that are obviously affected by a given measurement, e.g., the
fractional enrichment of 4- l3 C-glutamate following 1-13C-glucose infusion is
strongly related to pyruvate flow into the TCA cycle.
2. Pathways are included that are important for a hypothesis, e.g., pyruvate carbox-
ylase (PC) is active. If the model containing the PC flux more accurately fits the
data and has a plausible flux, then it is most likely active.
3. In 13C-NMR experiments, pathways are often included that dilute the flow of iso-
tope through the system, even if they do not affect the overall carbon flows, e.g.,
the exchange of amino acids into total cell protein. These pathways must be in-
cluded because they affect the accuracy of other fitted fluxes.

17.4.2
Pathway Inclusion and Reduction Assumptions

Table 17.2 contains each of the pathways in Fig. 17.7 as well as past studies including
and eliminating those pathways. Following the table are reasons for inclusion and
arguments for elimination of pathways in various cell types and conditions. These
arguments were used to reduce the comprehensive model in Fig. 17.7 to the simpli-
fied model in Fig.I?1.
Several pathways, for the most part, are conserved in all of the models. They are
glycolysis (or the Embden, Meyerhof, Parnas pathway), which is defined as the con-
version of glucose to pyruvate starting with hexokinase (HK) and ending with pyr-
uvate kinase (PK); pyruvate dehydrogenase (PDH); lactate dehydrogenase (LDH);
and the TCA cycle, which includes the reactions from citrate synthase (CS) to malate
dehydrogenase (MDH).
Some models do not include these pathways, usually because of special experimen-
tal conditions. For example, Malloy et al. [19] and Chance et al. [3] examined rat
hearts perfused with l3 C-Iabeled acetate and pyruvate. In these cases no glycolysis
or LDH activity would be detected and their inclusion in the model would not be
relevant. Likewise, Schrader et al. [21] studied human erythrocytes, which do not
have functioning TCA cycles when mature; PDH would not be active either.
Pyruvate carboxylase (PC) is the major anaplerotic enzyme. Its major function is
the replenishment of TCA cycle intermediates. For steady or pseudo-steady state sys-
tems, PC would only be active if there was a depletion of TCA cycle metabolites.
There are two sources of such depletions [1]: biosynthesis of fatty acids [f73 ] and
non-essential amino acids [f78 ,81,92], and oxidation following conversion to pyruvate
[f74,75]'
 17.4 Biochemical Pathway Model Formulation and Reduction 583

Table 17.2. Pathways included and eliminated in the literature

Pathway Corresponding fluxes Studies including Studies

in Fig. 17.7. pathway actively eliminating
pathway

Pyruvate Carboxylase (PC) f46 5,12,14,19,23 4,6, 18

Malic Enzyme (ME) f45,74 4,5,12,14,23,37 3, 18, 19
Aminotransferase f89,90; f35,36; f68,69 3-5,12,14,18,
and f85,86; f60,61 19,23,37
Malate/aspartate shuttle f58-69 4
Glutaminase f93,94 5,12,18,23
Pentose phosphate pathway fll - 20 5,12,14,21 ,23 37
Macromolecular synthesis f3,14,73 5
Non-essential amino acids f30,39,76,97'. f 33,34,79,80'. 5,14,23,37
f31 ,32,37,38,77,78,81,82,91,92
Fatty acid and cholesterol f n ,73 5,14,23
synthesis
DNA and RNA synthesis f14 5
Macromolecular carbo- f1,3 5,12,14,18,21
hydrate
Additional catabolic f47; f52,55,96 3-5,14,18,23
pathways
Enzymatic channeling f53,54,56,57 14 3,4,23,37
of TCA cycle intermediates
Phosphoenolpyruvate f75 12
carboxykinase
2,3 diphosphoglycerate f21-26 21
bypass

Hyder and co-workers [18] justified the elimination of PC because the rat brains
they studied were incapable of growth and therefore had low requirements for bio-
synthesis. Chatham et al. [4] argued that the PC reaction, whose purpose is to increase
TCA cycle intermediates, should not be active in their steady-state experiments. They
claimed that an isotope distribution would not be affected by a small PC activity be-
cause of the rapid forward and backward rates of fumarase. The flow through PC is
typically small, so it does not significantly affect carbon flows. This is not necessarily
true in rapidly growing cells, however, where PC is more active [16] in order to supply
biosynthetic pathways with more TCA cycle intermediates. The only other appreciable
source of TCA cycle intermediates is the consumption of glutamine.
Malic enzyme (ME) has the opposite function of PC; it is the critical step in the
oxidation of TCA cycle intermediates, converting malate into pyruvate. Anaplerotic
reactions, including PC and catabolism of some of the amino acids (i.e.,
[f52 ,55,61,94,96]), must be balanced by active ME or biosynthetic fluxes. Because malic
enzyme is the only enzyme able to produce pyruvate from TCA cycle intermediates,
it is included in almost every model.
An important function of ME has been described as the production of the NADPH
that participates in de novo lipogenesis [41]. The association between malic enzyme
and de novo lipogenesis has also been made because de novo lipogenesis relies on
citrate transport out of the mitochondria [f70 ,71], resulting in cytosolic OAA which
may only be oxidized after conversion to pyruvate by ME [f74 ].
584 17 Analysis of Metabolic Fluxes in Mammalian Cells

In the works of Malloy et al. [19), Hyder et al. [18), and Chance et al. [3), no ME
was included. Because Malloy et al. [19) included a PC flux, it is most likely that the
overall mass balance did not close. In normal, non-cancerous cells, e.g., rat brains
[18) and hearts [3), the consumption of glutamine is low and PC is inactive, so ME
need not be considered.
There is some controversy about the location of ME, i.e., whether or not it is pre-
sent within the mitochondria. Most models do not consider metabolite partitioning
across the mitochondrial membrane [5, 12, 14,37). When these include ME, its loca-
tion is unimportant. Portais et al. [14) include a generic flux described as "efflux
from the citric acid cycle;' which is most likely flux through malic enzyme.
The presence of mitochondrial compartmentalization is one of the two major dif-
ferences between the metabolism of prokaryotic and mammalian cells (the other is
the absence of the glycoxylate shunt). Chatham et al. [4) modeled ME as occurring
inside the mitochondria [f45 ) for rat hearts and Sharfstein et al. [23) modeled it out-
side [f74 ) for hybridomas. It has been shown that the location of ME is tissue depen-
dent [41,42). Newsholme and Leech [42) describe kidney metabolism where "malate
[is) transported out of the mitochondria and is converted... to pyruvate via 'malic
enzyme';' and malate is converted to pyruvate in the intestine by a "mitochondrial
reaction catalyzed by a decarboxylating malate dehydrogenase (ME):' Biochemistry
texts [1, 40) generally assume that malic enzyme is active only in the cytosol. An
extramitochondrial ME, in contrast to an intramitochondrial ME, would have differ-
ent effects on the flows through the malate/aspartate shuttle; the out flow of aspartate
would be greater than the in flow of malate.
Aminotransferase (AT) reactions are all linked by the nitrogen mass balance [5).
Glutamate (glu) can be converted to a-ketoglutarate (a-kg) by three reactions: alanine
aminotransferase (alaAT), aspartate aminotransferase (aspAT), and glutamate dehy-
drogenase (GDH). Glutamate dehydrogenase releases the glutamate amine as NH 3;
alaAT and aspAT transfer it to pyruvate and OAA, respectively. When ammonia pro-
duction is measured, the rate of GDH can be deconvoluted and a net flux through the
transferases determined [5, 6). In the absence of ammonia measurement, it is often
assumed that the forward and back reactions are fast, allowing the amino and a-keto
acids to be treated as a single pool [4,14,18,19). Because of the equilibrium between
malate and OAA, a single pool of malate, OAA, and aspartate has been suggested [5).
The aminotransferase reactions are important for 13C_NMR experiments. They
provide data on the labeling of a-keto acids, all of which are centrally located meta-
bolites and whose labeling data is required to solve most models. However, the con-
centrations of the a-keto acids are often too small to be detectable. If the amino-
transferase reactions are assumed to be in rapid equilibrium, then the labeling pat-
tern of the more abundant amino acids will be the same as that of their correspond-
ing a-keto acids (19).
It is not necessary for fluxes f85 and f86 to be rapid for the relative labeling of
glutamate and a-ketoglutarate to be identical. If the source of label is "upstream" of
isocitrate dehydrogenase [f49 ) then a-ketoglutarate is the only source of label for
glutamate. If the system is at metabolic steady state, then any label transferred to
a-ketoglutarate would also be transferred to glutamate. A slow transferase reaction
between glutamate and a-kg is one possible way to explain the "dilution" oflabel in
the TCA cycle. If the difference between the forward and back reactions [f85 and f86)
is large compared to the mean of the two values, then a reduction or "dilution" in the
absolute label in glutamate would be observed.
17.4 Biochemical Pathway Model Formulation and Reduction 585

The model of Chatham et al. [4] included the two sets of aspartate aminotrans-
ferases present in mammalian cells, an extramitochondrial [f68,69; f89,90] and an in-
tramitochondrial [f60,61; f85,86]. Both of these aspATs are needed for a model to con-
tain a complete malate/aspartate shuttle.
The malate/aspartate shuttle transports reducing equivalents (NADH) produced
in the cytosol into the mitochondria. It was included in the model of Chatham et
al. [4] because it is "the only mechanism for labeling the cytosolic glutamate pool
from a very small mitochondrial a-ketoglutarate pool:' A major effect of an active
malate/ aspartate shuttle is that it provides a means by which TCA intermediates
can reach the cytosol. Alternately, citrate can be transported by the tricarboxylate
transport system [f70 ,7l] or malate can be transported by the malate/aspartate shut-
tle operating in reverse [f64 ]. This has been called the malate shunt [23]. It has been
proposed that cancerous cells have limited ability to transport reducing equivalents
[43]. The malate/aspartate shuttle accounts for the oxidation of NADH produced
under aerobic glycolysis, a function also performed by the glycerol phosphate shut-
tle [4].
Glutaminase is more active in transformed cells, which consume large amounts of
glutamine. The reverse reaction [f93 ] is glutamine synthetase. Hyder et al. [18] ac-
counted for loss of glutamate label in the rat brain by an equilibrium between gluta-
minase and glutamine synthetase. Deamination of unlabeled glutamine would dilute
label in glutamate.
The pentose phosphate pathway (PPP) is active in most cell types. However, there
is some discrepancy in the literature about how much it contributes to the flux from
glucose to pyruvate [44]. The magnitude of the flux has been shown to be dependent
on the method of measurement used. Schmidt et al. [44] demonstrated that cometa-
bolite measurement has a tendency to overestimate this flux, to the extent of not
being physiologically feasible.
Many of the models treat the PPP as a single irreversible flux [5, 14, 23]. This
treatment is plausible if it is assumed that the intermediates in the PPP are not in-
volved in any other reactions of primary metabolism. Then the system of fluxes can
be lumped together as a single flux as in the model of Fig. 17.1. When the DNA/RNA
biosynthetic rate [fI4 ] is measured, this drain on ribose can be added [5]. Schrader et
al. [21] extensively studied this system of pathways in human erythrocytes using
kinetic NMR and various labeled glucose species. After treatment with 1_l3C glucose,
these authors detected 1_l3C pentose 5-phosphate (P5P), which could only have been
produced from 1_l3C fructose 6-phosphate (F6P). This implies that transketolase
[fI6 &17] and trans aldolase [fI9 ] are both reversible [fI5 ,18,20]. As with most reversible
fluxes, treating them as net fluxes does not affect the general metabolite balance, but
does affect the label distribution. Additionally, the rate through these reverse reac-
tions is most likely very small compared to the glycolytic flux (about 10% of the total
PPP flux [21]).
Zupke and Stephanopoulos [37] did not include the PPP. From literature values
these authors expected it to be less than 5% of the glycolytic flux. Even in the pre-
sence of high PPP flux, the effect on carbon flow to pyruvate is small; equal fluxes
through glycolysis and the PPP only produce 1/6 less pyruvate. The most significant
effect is on the labeling distribution following 1_13C glucose infusion, in which all of
the isotope is lost as 13C02 in the phosphogluconate dehydrogenase reaction of the
PPP. The PPP activity Schrader et al. [21] determined for human erythrocytes was
similar to the assumption of Zupke and Stephanopoulos: 6.5-17%.
586 17 Analysis of Metabolic Fluxes in Mammalian Cells

Macromolecular synthesis was shown by Bonarius et al. [5] to be necessary for mod-
els of fast growing cells. These authors measured cell composition changes and in-
corporated biosynthetic fluxes into their model. These measurements have been seen
previously as "corrections" to the energy metabolism fluxes, since their values are
often assumed to be small. However, this was not the case with the hybridoma line
employed by Bonarius et al. in continuous culture [5]. Labeling experiments [14, 18,
23], which do not independently measure macromolecular synthesis, often include
"generic" synthesis fluxes to explain losses in label.
Non-essential amino acids have two fates once synthesized: excretion and incor-
poration into total cellular protein. It is generally believed that some amino acids are
excreted (usually alanine) by transformed cells to reduce the internal concentration
of NH 3. Incorporation of the amino acids that are in exchange with the a-keto acids
of central metabolism (alanine, aspartate, and glutamate) into proteins
[f37,38,77,78,91,92] is an explanation for the loss of label in the TCA cycle often observed
in 13C-NMR experiments [14,23]. It is generally assumed that the rate of incorpora-
tion (translation) is equal and opposite to the rate of production (protein degrada-
tion). This is a reasonable assumption in non-growing cells [14, 23] These fluxes
cause label dilution because labeled amino acids are incorporated into proteins dur-
ing the course of the experiment, and are replaced by unlabeled amino acids incor-
porated into proteins before isotope infusion.
From many 13C_NMR experiments [14, 23], label appeared to be diluted in the
TCA cycle in both an "overall" and a "per turn" manner. Overall dilution results
from the loss of the total amount of label in the TCA cycle compared to the amount
that should have entered via pyruvate dehydrogenase (PDH). Determination of the
flux through PDH and pyruvate label enrichment show a dilution in the label enrich-
ment of glutamate. This apparent "overall" dilution of label in the TCA cycle may
result from the loss of glucose's first carbon in the PPP, slow aminotransferase reac-
tions, uptake of unlabeled lactate, and catalysis of fatty acids.
The early model of Malloy et al. [19] did not contain protein exchanges and fatty
acid catabolism. This model predicts that glutamate would remain equally labeled in
the second and third positions. In practice, this is not observed. For label enrich-
ments to be different, dilution fluxes from unlabeled metabolites and fluxes to un-
detectable macromolecules must enter and leave the TCA cycle as the label pro-
gresses from the third to the second position of glutamate on each turn. Hence the
concept "per turn." There are two mechanisms that could diminish the "per turn"
label in the TCA cycle: protein exchange of aspartate [f77,78] or of glutamate [f91 ,92]'
These two are experimentally indistinguishable without additional measurements.
The values of some TCA cycle fluxes [f48 ,49] are slightly affected by the choice of
location for this dilution [14].
Fatty acid and cholesterol synthesis are dependent on the tricarboxylate trans-
port system [f7o,7d which is the means by which acetyl-CoA is transported out of
the mitochondria. The precursor for both fatty acid and cholesterol synthesis is
cytosolic acetyl-CoA. It is somewhat misleading to think of citrate transport as
directly exchanging mitochondrial for cytosolic acetyl-CoA because of the deplet-
ing effect it has on TCA cycle intermediates if the carbons return through ME
[f74 ] and PDH [f441. Citrate efflux must be coupled with an anaplerotic reaction
(either PC or glutamine catabolism) to maintain steady state, or the carbons must
return via malate transport [f651. This synthetic flux is negligibly active in slow
growing cells.
17.4 Biochemical Pathway Model Formulation and Reduction 587

DNA and RNA synthesis both require ribose produced in the pentose phosphate
pathway. This flux is often assumed to be small, even in rapidly doubling cultures.
Macromolecular carbohydrate synthesis and degradation are responsible for most
observed differences between glucose uptake and glycolytic flux. The most promi-
nent macromolecular carbohydrate is glycogen, whose production can be measured
[5] or used as a fitting flux in label experiments [14]. Membrane transport of glucose
can be estimated from the literature [18], or measured using radio-labeled substrates
[7, 12]. Transformed cells have such high rates of glycolysis that they are unlikely to
store glycogen.
Additional catabolic pathways make small contributions to the metabolite bal-
ances, but they introduce unlabeled carbon that can dilute label enrichments. This
second effect is more important because calculated fluxes are sensitive to relatively
small changes in label enrichments. For this reason, many authors have included the
breakdown offatty acids and cholesterol into acetyl-CoA [fd [3, 4, 14, 18].
Measuring the amino acid composition in culture medium is a straightforward
process. Catabolism of amino acids can be determined by direct measurement, in a
manner similar to extracellular metabolites (see Sect. 17.2.1). Protein synthesis can
be considered as an accumulation of amino acids. Caution must thus be exercised
when attributing cellular uptake of essential amino acids to requirements for energy
generation. In the presence of preferred energy sources (e.g., glucose or glutamine)
alternate catalytic pathways are often suppressed (e.g., LDH). Therefore, it is likely
that essential amino acid uptake contributes primarily to protein synthesis.
Enzymatic channeling of TCA cycle intermediates was invoked by Portais et a1.
[14] to explain observed label distributions. Srere [45] postulated that the enzymes
of the TCA cycle are contained within a metabolon, which is a supramolecular
complex bound to the inner surface of the mitochondrial membrane. Metabolites
would be transferred directly from enzyme to enzyme, without being allowed to
diffuse and rotate freely in the mitochondrial fluid. This is important for reactions
involving succinate and fumarate, which are both symmetric molecules. With free
rotation, label on the second and third positions of succinate would be equivalent
as substrates for succinate dehydrogenase [f53 ,54]' Sumegi et a1. [46, 47] cite consid-
erable evidence to support the metabolon concept and performed experiments de-
monstrating this "orientation-conserved transfer" [46]. These authors fed mamma-
lian cells labeled substrates and observed the labeling pattern in specific metabo-
lites. Two such experiments were 3_ 13 C propionate to alanine [47], and 5_ 13 C gluta-
mate to aspartate [46]. In both cases, a preferred orientation of the product was
detected (2_13C succinyl-CoA resulted in only 2_13C malate. If free rotation were
possible 3_ 13 C malate would have been detected). Despite the evidence supporting
this theory, many of the models still contain freely rotating succinate and fumarate
[3, 4, 23, 37]. In Zupke's model "both succinate and fumarate are symmetrical, so
that when fumarate is converted to malate the 1 and 4 carbons will have the same
labeling [37]:'
Phosphoenolpyruvate carboxykinase (PEPCK) is the transcriptionally controlled
enzyme that regulates gluconeogenesis. Pyruvate kinase [f29 ] is considered irreversi-
ble, which implies that flux through PEPCK is the only route for carbons to enter
glycolysis and produce glucose. It is assumed to be inactive in most cells, except
those derived from the liver, gluconeogenesis being one of the liver's major func-
tions. In most transformed cell lines, high rates of glycolysis indicate that gluconeo-
genesis is not operational.
588 17 Analysis of Metabolic Fluxes in Mammalian Cells

The 2,3 diphosphglycerate (2,3DPG) bypass was included in the model of Schrader
et al. [21] because the rate of label appearance in 2,3DPG determined the PPP flux.
Being a cycle with only a single input and output, flux through the 2,3DPG bypass
does not affect total carbon flow; nor does it affect label distribution.
The determination of bi-directional reaction rates is one obvious advantage of iso-
tope labeling experiments coupled with non-linear analysis [22]. However, inclusion
of both fluxes (forward and back) invariably requires that additional measurements
be made, since no further metabolite balances are added. Reversible reactions can be
represented by net fluxes without loss of accuracy. However, considerable informa-
tion is contained in many of the exchange fluxes. As demonstrated in the discussion
on aminotransferases, a positive net flux from glu to a-ketoglutarate [f8s] without
any consideration of the reverse flux [f86] precludes any label from reaching glu from
a-ketoglutarate. If the exchange is assumed to be very rapid then the fractional en-
richment of the two can be made equal. Alternately, the reverse flux could be in-
cluded with no assumptions needed.

17.5
Observed Metabolic Flux Patterns in Mammalian Cells

It has long been known that cancerous cells have high rates of aerobic glycolysis, a
characteristic typical of rapidly proliferating cells [9]. Aerobic glycolysis is ineffi-
cient, wasting much of glucose's energy potential. The magnitude of aerobic glyco-
lysis is characterized by the molar yield of lactate from glucose (Ylac/gluc)' Neerman
and Wagner [12] found that Ylac/gluc ranged from 1.31 to 1.56 in BHK and CHO cells.
They also found that the amount of either pyruvate or glucose carbons entering the
TCA cycle via PDH in BHK cells was below the detection limit of the 14COz-evolution
assay. This observation was confirmed by the low activity detected for extracted
PDH. In T47D breast cancer cells, Ylac/gluc was approximately 1.4 [28].
There are numerous hypotheses explaining why cancer cells survive and flourish
with such apparently inefficient metabolic characteristics. In 1956 Warburg [10] pos-
tulated that the high cancerous rates of aerobic glycolysis were due to damaged re-
spiration. Both Singer et al. [15] and Neerman and Wagner [12] found an elevated
NADH/NAD+ ratio in primary breast cancer cells and CHO/BHK cells compared to
normal breast epithelial cells and insect cells, respectively. It was postulated that a
high NADH/NAD+ ratio drives the high production rate of lactate. Mazurek et al.
[11] and Eigenbrodt et al. [43] postulated that the transport of reducing equivalents
into the mitochondrion to drive oxidative phosphorylation is hampered. Mazurek
also postulated that an accelerated glycolytic flux would increase the concentration
of phosphometabolites upstream of pyruvate kinase; many of these metabolites are
precursors for biosynthesis, and their increase would induce proliferation. All of
these hypotheses can be further examined by flux analysis experiments.

17.5.1
Linkage of Glycolysis to the Tricarboxylic Acid Cycle

Regardless of which hypothesis most effectively explains aerobic glycolysis in can-

cerous cells, all of them involve a relationship between cytosolic gylcolysis and the
17.5 Observed Metabolic Flux Patterns in Mammalian Cells 589

mitochondrial TCA cycle. This relationship includes the transport of both reducing
equivalents and carbons across the mitochondrial membrane.
Lia and Behar [48] postulate that in brain cells (neurons, astrocytes, and oligoden-
drocytes) the transfer of reducing equivalents into the mitochondria is the mechan-
ism that mediates the coupling of glycolysis to the TCA cycle. These authors showed
that in the presence of an inhibitor of aspAT (~-methylene-DL-aspartate), cortical
slices have reduced oxygen uptake and increased lactate production [49]. This result
implies that impairing the malate/aspartate shuttle induces aerobic glycolysis. By
measuring the NADH/NAD+ ratio, this hypothesis could be further confirmed.
In BHK and CHOs, Neerman and Wagner [12] did not detect any activity in the
mitochondrial "linking" enzymes PEPCK, PC, and PDH. Without PDH activity, glu-
cose cannot be oxidized, and without PEPCK and PC activity, carbon cannot be ex-
changed between glycolysis and the TCA cycle. In contrast, Portais et al. [14] found
by !3 C-NMR that the oxidation of glucose produced 30% of the ATP in C6 glioma
cells and claimed that oxidative phosphorylation constituted the major energy
source for these cells.

17.5.2
Reducing Equivalents

Chatham et al. [4] included two reducing equivalent (NADH) transporters in their
model of rat heart metabolism: the malate/asp shuttle and the glycerophosphate
shuttle. As opposed to the malate/aspartate shuttle, which can be accounted for in a
carbon flow model, the glycerophosphate shuttle only involves the cycling of dihy-
droxyacetone phosphate (DHAP) and 3-phosphoglycerol [1]. Therefore its activity
cannot be detected by a carbon flow, and it can be considered a "generic" transpor-
ter, whose only observable effect is the direct transfer of cytosolic NADH to the
mitochondrial inner membrane electron transport chain.
Chatham et al. [4] measured both isotopomer distributions and oxygen consump-
tion; together these allowed the formulation of an NADH balance. When the glycer-
ophosphate shuttle was not considered, the rate of the malate/asp shuttle was too
high to explain the rate of label appearance in glutamate and it was too slow to
account for the NADH produced by glycolysis. These authors concluded that the
glycerophosphate shuttle had to be included in order for the NADH balance to close.
Bonarius et al. [5] detected active transhydrogenase in hybridomas by comparison
of both the NADH and NADPH balances. They found that it operated in the direc-
tion producing NADPH and NAD+ from NADP+ and NADH. Transhydrogenase ac-
tivity reduces the NADH/NAD+ ratio, inducing glycolytic flux, and increases the
concentration of NADPH, which is necessary for many biosynthetic reactions. In
the literature there is some controversy about the activity of transhydrogenase; most
authors consider it inactive. In mammalian cells further investigation of transhydro-
genase is needed because of its key role in pathway regulation.

17.5.3
Glutaminolysis

Glutaminolysis is defined as the production of pyruvate from consumed glutamine.

It is sometimes implied that the pyruvate is excreted as lactate, but there are many
fates possible for pyruvate depending on its location within the cell. Cell lines that
590 17 Analysis of Metabolic Fluxes in Mammalian Cells

do not appreciably oxidize glucose are dependent on the oxidation of glutamine for
energy. When glutamine carbons are consumed, they are first converted to the TCA
cycle intermediate a-ketoglutarate by glutaminase and GDH. They are then oxidized
to OAA by a truncated TCA cycle, producing reducing equivalents (2 NADH, and
FADH z ) within the mitochondrion. Glutamine consumption thus has an anaplerotic
nature, contributing carbons to the pool of TCA cycle intermediates.
There are three general fates for glutamine-derived OAA carbons:
1. They can be passed from the mitochondrion as aspartate where cytosolic malic
enzyme can convert them to pyruvate for excretion as lactate or alanine. Pyruvate
can also be transported to the mitochondria and oxidized, a route that requires
PDH activity.
2. Alternatively, a mitochondrial ME could directly convert the OAA (after MDH)
into pyruvate, which is then oxidized. As stated earlier, there is disagreement in
the literature about the existence of a mitochondrial ME. Chatham et al. [4] did
not observe any l3C label in either alanine or lactate. They speculated that, because
label was not present outside the mitochondrion, mitochondrial ME was active in
rat hearts. It has also been shown that the level of mitochondrial ME activity cor-
relates with the degree of de-differentiation in cancer cells [16]. Portais et al. [14]
did not see appreciable levels of l3C_2 lactate in C6 glioma cells and concluded that
ME was not active. These authors speculated that this low ME activity was due to
"a lack of enzyme or a low level of extracellular malate [14]."
3. Another consequence of the glutamine carbons is the fueling of the biosynthesis of
cellular components derived from TCA cycle intermediates (non-essential amino
acids or fatty acids).

Studies with U_ 14 C glutamine showed that BHKs [12] and hybridomas [7] oxidize
18% and 36% of consumed glutamine, respectively. This range of glutamine con-
sumption is typical for most cultured cell lines. Both Neerman and Wagner [12]
and Fitzpatrick et al. [7] reported little PDH activity, implying that a mitochondrial
ME is active to oxidize the glutamine carbons. Extracted citrate synthase was ob-
served to be 2- to 3-fold less active than aspAT, supporting the concept of a trun-
cated TCA cycle producing aspartate from glutamine carbons. Most of the energy
generated in the TCA cycle would be formed by the conversion of glutamine to as-
partate. The truncated cycle would produce nine ATPs, whereas a full turn of the
cycle produces 12. Fitzpatrick et al. [7] reported a consumption of glutamine in hy-
bridomas half that of the glucose uptake rate. That translates to 55% of the cellular
energy being derived from glutamine and 45% from glucose. Literature reports en-
ergy production from glutamine in cultured cell lines ranging from 30% to 98% [7]
of that obtained from glucose.

17.5.4
Pyruvate Carboxylase

Baggetto [16] describes how glycolytic cancer cells preferentially excrete citrate from
the mitochondrion. This citrate produces cytosolic acetyl-CoA through the action of
ATP-citrate lyase, which fuels lipid and cholesterol synthesis. For cells to excrete
carbons from the TCA cycle, an anaplerotic counter-balance is needed at steady
state. This can arise through the action of either PC or glutaminolysis. Both Martin
et al. [20] and Merle et al. [13] found that PC was twice as active in astrocytes as in
17.5 Observed Metabolic Flux Patterns in Mammalian Cells 591

granule cells. They used this result to explain the greater dependence on glutamine
in granule cells.

17.5.5
Pentose Phosphate Pathway

Martin et al. [20] determined the flux through the PPP in astrocytes and granule
cells to be 11% and 29% of the glucose flux, respectively. Bonarius et al. [5] observed
much higher levels in hybridomas, ranging from 76% to 92% of the total glucose
uptake. In the cells with 92% of the incoming glucose flowing through the PPP, a
negative flux from F6P to G6P was calculated. It was speculated that these high va-
lues of the PPP flux were required to supply NADPH to the rapidly growing contin-
uous hybridoma culture. NADPH is consumed primarily in biosynthetic reactions,
and is therefore essential for rapidly proliferating cells. Prior to the work of Bonarius
et al. [5], the range of the PPP flux relative to glucose uptake was given as 0.7% to
11% in hybridomas [5,23].
Schmidt et al. [44] questioned the validity of this high PPP flux based on metabo-
lite balancing alone. These authors used bacteria cultured on two different nitrogen
sources as a model to demonstrate their skepticism. In a single culture, both the
cometabolite measurements necessary for metabolite balancing and isotope tracers
were employed. They reported the shift in PPP flux when the nitrogen source was
changed from NH3 to N0 3 to be 34% to 120% of the glucose uptake by metabolite
balancing, and 53% to 60% by l3C-NMR measurements. They attribute the discre-
pancy between these results to a dependence of metabolite balancing on an NADHI
NADPH balance. It was assumed by Bonarius [5] that all of the reactions either pro-
ducing or consuming NADH and NADPH were included in the model in order to
close both balances. Transhydrogenase was also assumed to be activity. Neither of
these assumptions is necessary in tracer experiments. Bonarius et al. [33] defend
metabolite balancing, claiming that it is far simpler and less expensive for industrial
on-line testing. Zupke and Stephanopoulos [6] found little difference between the
fluxes determined by metabolite balancing and l3C-NMR.
Schrader et al. [21] found a reverse flux of both TA and TK in the presence of a
positive PPP flux in human erythrocytes. Using 1_l3C glucose, they detected 1_l3C
lactate, which could only have been formed by rearrangement in the PPP. Of the total
labeled lactate, 6% was labeled at the first position; this ratio approximates the rela-
tion of the reverse TK and TA flux to the total glycolytic flux.

17.5.6
Tumors as Nitrogen Sinks

Baggetto [16] claimed that GDH is not active in tumor cells, glutamate being prefer-
entially converted to a-ketoglutarate by aminotransferases. GDH is strongly inhib-
ited by GTP, as well as its enzymatic product NH 3, both of which are abnormally
high in cancerous cells. The dependence of cancer cells on glutamine oxidation re-
sults in a high concentration of nitrogen inside the tumor, where it remains trapped,
starving the host organism. Bonarius et al. [5] saw similar events in hybridoma cul-
ture. The GDH flux measured by these authors ranged from 0% to 1% of the glucose
uptake rate. In their system, glutamine carbons were supplied to TCA cycle a-kg
through aspAT and alaAT. The high concentration of intracellular NH3 forced GDH
592 17 Analysis of Metabolic Fluxes in Mammalian Cells

to produce glutamate. Glutamate production has also been observed in T47D breast
cancer cells [28]. Bonarius et al. [5] also saw a large proline synthesis flux. They
attribute this to accumulation of glutamate within the cell and propose it as another
means of reducing the intracellular NH3 concentration. It is postulated that proline
synthesis in hybridomas is a vestige of the intercellular proline cycle. In the presence
of glutamate, NADPH, and ATP, proline is produced by blood cells, from which hy-
bridomas are originally derived.

17.5.7
Oxidative Glycolysis in the Rat Brain

Hyder et al. [18] concluded that oxidative as opposed to non-oxidative glycolysis is

the major energy source in the brain following cortical stimulation. This was drama-
tically demonstrated by the more rapid increase of 4_13C glutamate in stimulated as
opposed to resting rats. It was found that upon stimulation the rate of oxidative
glucose utilization and rate of oxygen uptake increased 195% and 201 %, respectively.
These identical shifts in rates allowed the authors to conclude that the non-oxidative
rate did not contribute to the increase in glucose usage.

17.6
Specific Uses of Flux Pattern Information

Within most discussions of metabolic flux analysis, a common question often arises:
how is this knowledge about carbon fluxes through internal pathways of mammalian
cells useful? A common accusation is that metabolic flux analysis is a solution with-
out a problem. In other words, what do these lists of numbers representing flows of
carbon through mammalian cells actually mean? Regardless of whether this criti-
cism was valid in the past, it is now outmoded. Metabolic flux analysis exemplifies
the recent change in how biological systems are analyzed: from a reductionist to a
holistic approach. Many of the processes that control the functions of living cells can
only be fully understood when viewed in the context of other cellular processes.
While metabolic flux analysis by no means considers all cellular processes, it is one
of the first techniques to model whole cells. As experimental and modeling techni-
ques advance, models will soon account for more cellular processes, thereby increas-
ing the accuracy and significance of the results.
Recent advances in metabolic flux analysis have mostly been generated by so-
called metabolic engineers [50, 51] seeking explanations for the unexpected results
of planned genetic manipulations. For example, it has often been observed that sin-
gular increases in expression of what were believed to be rate-limiting enzymes did
not increase production of a desired metabolite. From metabolic control analysis [8,
52, 53] it was realized that many enzymes, not just one alone, control the flux of a
given pathway. This discovery of the inaccuracy of the rate-limiting step concept was
the driving force behind the development of the more holistic metabolic flux analy-
sis. Because understanding and modifying an isolated pathway proved to be insuffi-
cient to increase product synthesis, it became necessary to understand how pathways
interact. Metabolic flux analysis provides the means by which such interactions will
be illuminated.
References 593

Metabolic flux analysis will continue to play an important role in discovering reg-
ulation mechanisms and new functions within mammalian cells. Coupled with the
forthcoming automation and miniaturization of biological assays, flux analysis will
enable re-evaluation of old biological questions in a new and holistic manner. The
simultaneous comparison of multiple pathways will clarify intracellular regulation
mechanisms. Initially, metabolic flux analysis will be critical to defining cellular phe-
notypes and interpreting newly discovered genetic information. The ability to quan-
tify internal pathway fluxes enables flux analysis to define phenotypes in a way no
other method can. Palsson [54] termed this new field phenomics.
In addition, there is new interest in understanding the modifications of trans-
formed cells connected to their uncontrolled proliferation. While the concept that
an altered metabolism is the cause of proliferation [9] is outdated, there is obviously
a strong link between metabolism and growth rate. It is becoming more apparent
from genomics research that metabolism comprises a large portion of cellular activ-
ity. As Edwards and Palsson point out, "a majority of the assigned open reading
frames relate to metabolic functions" [55]. Causality notwithstanding, significant
changes in the expression and function of the enzymes of central metabolism must
surely be necessary for cancer cells to sustain increased and stable proliferation.
Determining changes in metabolic fluxes will point to corresponding changes in
gene expression.
Clearly, there are numerous potential benefits to obtaining a more detailed picture
of metabolic regulation, for both industrial and medical applications. Industrially,
knowing the factors that induce product formation would invariably improve pro-
cess efficiency. And medically, altered isozyme expression in tumor cells, detected as
changes in metabolic fluxes, may be novel drug targets. Baggetto [16] remarked
about observed flux differences between normal and cancerous cells: "such meta-
bolic particularities of cancer cells should stimulate researchers in finding specific
drugs to reach and kill cancer cells:' The exciting future for metabolic flux analysis is
that it will uncover regulatory mechanisms within mammalian cells, which in turn
will be used to develop novel drug treatments and enhance protein production.

References
1. Voet D, Voet JG (1990) Biochemistry. Wiley, New York
2. Matwiyoff NA, London RE, Hutson JY (1982) The study of the metabolism of I3C labeled
substrates by the I3C_ NMR spectroscopy of intact cells, tissues, and organs. In: Levy GC
(ed) NMR spectroscopy: new methods and applications. American Chemical Society, Wa-
shington DC, chap 8
3. Chance EM, Seeholzer SH, Kobayashi K, Williamson JR (1983) J BioI Chern 258:13,785
4. Chatham JC, Forder JR, Glickson JD, Chance EM (1995) J BioI Chern 270:7999
5. Bonarius HPJ, Hatzimanikatis V, Meesters KPH, de Gooijer CD, Schmid G, Tramper J
(1996) Biotechnol Bioeng 50:299
6. Zupke C, Stephanopoulos G (1995) Biotechnol Bioeng 45:292
7. Fitzpatrick L, Jenkins HA, Butler M (1993) Appl Biochem Biotech 43:93
8. Fell DA (1998) Biotechnol Bioeng 58:12l
9. Warburg 0 (1929) Biochem Z 204:482
10. Warburg 0 (1956) Science 123:309
11. Mazurek S, Michel A, Eigenbrodt E (1997) J BioI Chern 272:1491
12. Neerman J, Wagner R (1996) J Cell Physiol 166:152
13. Merle M, Martin M, Portais JC, Schuster R, Canioni P (1995) J Chim Phys 92:1783
14. Portais JC, Schuster R, Merle M, Canioni P (1993) Eur J Biochem 217:457
594 17 Analysis of Metabolic Fluxes in Mammalian Cells

15. Singer S, Souza A, Thilly WG (1995) Cancer Res 55:5140

16. Baggetto LG (1992) Biochimie 74:959
17. Portais JC, Merle M, Labouesse J, Canioni P (1992) J Chim Phys 89:209
18. Hyder F, Chase JR, Behar KL, Mason GF, Siddeek M, Rothman DL, Shulman RG (1996)
Proc Natl Acad Sci 93:7612
19. Malloy CR, Sherry AD, Jeffrey FMH (1988) J Bioi Chern 263:6964
20. Martin M, Portais JC, Labouesse J, Canioni P, Merle M (1993) Eur J Biochem 217:617
21. Schrader MC, Eskey C1, Simplaceanu V, Ho Chien (1993) Biochim BioRhys Acta 1182:162
22. Wiechert W, de Graaf AA (1996) In vivo stationary flux analysis by C labeling experi-
ments. In: Scheper T (ed) Advances in biochemical engineering. Springer, Berlin Heidel-
berg New York, p lO9
23. Sharfstein ST, Tucker SN, Mancuso A, Blanch HW, Clark DS (1994) Biotechnol Bioeng
43:1059
24. Mancuso A, Sharfstein ST, Tucker SN, Clark DS, Blanch HW (1994) Biotechnol Bioeng
44:563
25. Fernandez CA, Rosiers CD (1995) J Bioi Chern 270:lO,037
26. Blanch HW, Clark DS (1996) Biochemical engineering. Marcel Dekker, New York
27. Mancuso A, Fernandez EJ, Blanch HW, Clark DS (1990) Bio/Technology 8:1282
28. Forbes NS, Clark DS, Blanch HW (1998) University of California, Berkeley (unpublished
data)
29. Mancuso A (1993) PhD Thesis, University of California, Berkeley
30. Furman E, Rushkin E, Margalit R, Bendel P, Degani H (1992) J Steroid Biochem Molec
Bioi 43:189
31. Neeman M, Degani H (1989) Cancer Res 49:589
32. Chandler JP, Hill DE, Spivey HO (1972) Comput Biomed Res 5:515
33. Bonarius HPJ, Timmerarends B, de Gooijer CD, Tramper J (1998) Biotechnol Bioeng
58:258
34. Savinell JM, Pals son BO (1992) J Theor Bioi 154:421
35. Savinell JM, Palsson BO (1992) J Theor Bioi 155:421
36. Pramanik 1, Keasling JD (1997) Biotechnol Bioeng 56:398
37. Zupke C, Stephanopoulos G (1994) Biotechnol Prog lO:489
38. Schmidt K, Carlsen M, Nielsen J, Villadsen J (1997) Biotechnol Bioeng 55:831
39. Duggleby RG (1980) Eur J Biochem lO9:93
40. Stryer L (1988) Biochemistry, 3rd edn. WH Freeman, New York
41. Brdiczka D, Pette D (1971) Eur J Biochem 19:546
42. Newsholme EA, Leech AR (1983) Biochemistry for the medical sciences. Wiley, Chiche-
ster
43. Eigenbrodt E, Fister P, Reinacher M (1985) New perspectives on carbohydrate metabolism
in tumor cells. In: Bietner R (ed) Regulation of carbohydrate metabolism. CRC Press,
Boca Raton, chap 6
44. Schmidt K, Marx A, de Graaf AA, Wiechert W, Sahm H, Nielsen J, Villadsen J (1998)
Biotechnol Bioeng 58:254
45. Srere PA (1990) Trends Biochem Sci 15:411
46. Sumegi B, Sherry AD, Malloy CR, Srere PA (1993) Biochemistry 32:12,725
47. Sumegi B, Sherry AD, Malloy CR (1990) Biochemistry 29:9lO6
48. Lia JCK, Behar KL (1993) Dev Neurosci 15:181
49. Fitzpatrick SM, Cooper AJL, Duffy TE (1983) J Neurochem 41:1370
50. Bailey JE (1991) Science 252:1668
51. Stephanopoulos G, Sinskey AJ (1993) Tibtech 7:392
52. Heinrich R, Rapoprt TA (1974) Eur J Biochem 42:89
53. Kaeser H, Burns JA (1973) Symp Soc Exp Bioi 27:65
54. Palsson BO (1998) Metabolic engineering II. Elmau, Germany
55. Edwards JS, Palsson BO (1998) Biotechnol Bioeng 58:162
Subject Index

A anaerobic processes 17
abiotic phase 74, 86 anaplerosis 508, 522
accuracy 125 anchorage-dependent cells 559
acetate balance 186 animal cell culture 158, 163, 164
acetate formation 387 arginine 363
acetate limitation constant 191 Arrhenius-relation 72
acetate production 389 Arrhenius-type function 395
acetate uptake rate 191 arthrospores 413,417,420
acetic acid, formation 382 artificial neural network 326, 335
acetoin 241 aspartate familiy 510
acid titration rate 185 aspartate kinase, feed-back resistant 541
acid tolerance 370 asymptotic observer 154, 161
activation 46 ATL (airlift tower loop) 34
active part 366 ATL reactor 34, 36
activity, fermentative 279, 283 pilot plant 37-39
adaptive linearizing control 155, 156, 164 atom mapping matrices 114,489,529,574
aeration 173 ATP consumption 361
aerobic 61 ATP-citrate lyase 590
aerobic batch 291 automation 167,321
aerobic conditions 278 axial dispersion 261
aerobic process 16,51 coefficient 36
age distribution model 101,299,308
p/agitation controll 197 B
air sparging 247 Bacillus subtilis 241
airlift 293, 311 back flushing 371
reactor 10,247 bacterial spoilage 368
airlift tower loop (ATL) 34 Baker's yeast 37,65
reactor 21 balance boundary 375
algebraic slip model 212,227 balance equation, isotopic 519
allosteric activator 358 balance for the buffer 187
amino acid 359, 506 Balling formula 323
non-essential 586 base titration rate 185
aminotransferase 584 batch 25, 35, 80, 92, 283, 418, 422, 426, 560
ammonia 359 batch cultivation 13, 63
balance 187 batch culture 545, 562
ammonium carrier 508 batch kinetics 364
ammonium concentration 540 batch process 47
ammonium limitation 508 Be (bubble column) 21,30,34,247
ammonium-assimilating enzymes 523 beer fermentation 321,323
ammonium-assimilating pathway 509 bimolecular reaction 529
AMMs (atom mapping matrices) 574 biochemical engineering model 172
anabolism 45 biodegradable ester 368
anaerobic 58, 71 biofilm 54, 55, 90
anaerobic conditions 278 bioinformatics 489
anaerobic digestion 157, 160 biological behaviour 173
596 Subject Index

biological reaction process 173 cell

biomass composition 486,496 aggregate 28, 54
biomass estimation 382 anchorage-dependent 559
bioprocess 167 density cultivation, high 170
bioreaction network 513 lysis 362
bioreactor 207 mass balance 186
BIOSIM 194 recirculation 370
BIOSTAT ED5 194 cell-specific glucose uptake rate 190
biosynthetic machinery 367 cell-specific glycerol uptake rate 191
biosynthetic pathway 510 cell-specific growth rate 189
biotechnical production line 5 cell-specific maintenance rate 189, 192
biotic phase 12,74,86,283 cell-specific product formation rate 195
biotin limitation 541 cell-specific reaction rate 186, 192
bipolar electrodialysis 371 cellular compartment 564
black-box model 70 central metabolism 507
bon enrichment analysis 113 Cephalosporin C 92
bottle-neck 46, 190, 303 channelling 587
boundary condition 38, 40, 55, 102 cheese manufacture 353
branchpoint chemostat 14,25,61,77,113,291,351,561
glucose-6-phosphate 519,532 cultivation 281
pyruvate 532 Chen-Kim model 215,228
rigid 532 Chen-Kim modified k-E model 215
Brevibacterium flavum 538 Chinese Hamster Ovary 119
bubble column (BC) 21,30,34,247 chloroform extract 495
bioreactors 247 churn-turbulent flow 36
bubble diameter 229 circulation cells 35
bubble size distribution 28, 229 circulation flow pattern 22
bubble swarm 226 circulation model
bubbly flow regime 36 liquid 41
budding cells, fraction 298, 299 multiple cell 36
budding cycle 279 circulation time 22
bulk chemical 350, 368 citric acid 371
butanediol 241 cycle 522
butyric bacteria 110 market 369
by-product formation 494 CNMR 483
C-nuclear magnetic resonance (NMR)
C spectroscopy 112
I-BC]glucose 545 C02 development rate 332
BC fractional enrichment 528 C02 mass fraction 180
BC isotopomer data 489 C02 removal rate 181
BC-labeled cell protein 538 C02 transfer rate 182
6-BC]glucose 545 cometabolite measurement 565
9 14C 487 cometabolites 566, 568
calibration model 129 commodity chemical 368
cAMP 457,458,459,460,461 compartment model 32, 33, 73, 77, 78, 80, 283
CAN-bus 194 compartmentalization 367
carbon atom transfer 523 computational deficiency 530
carbon atom transmission 520 computational fluid dynamics 41,210
carbon balancing 485 concentration gradients 239
carbon dioxide mass balance 181 confidence limits 128, 130
carbon dioxide transfer rate 183 constitutive equation 74,76
carbon enrichment analysis 113 continuous 25, 283
carbon-13 satellite signals 541 continuous cultivation 14
cascade control system 178 continuous stirred tank reactors (CSTR)
casein peptone 364 ideal 23
catabolic pathway 367 ideally mixed 23
catabolism 44, 49 control 145,146,147,386,395
catabolite repression 192 linear 152
Subject Index 597

nonlinear 153, 154 DNA synthesis 587

sophisticated 170 DNA technology, recombinant 106
control analysis, metabolic 565 dormant cells 367
control loops 7 double jacket system 177
control strategies, applied 167 doubling time 561
cooling heat exchanger 177 down comer 10
cooling water entry heat resistance 178 down-stream operation 369
corn steep liquor 368 downstream processing 7,311,426
Corynebacterium glutamicum 112, 118, 506 drag coefficient 226
Crabtree-effect 287,292,308,310,313,316 drag coefficient cD 226
Crabtree-effect (gro) 279 drag force 212, 226
critical growth rate 381 drift flux model 36
cross validation 327 Dupres-Rankine 179
CSTR (continuous stirred tank reactors) 23 dynamic model 146,147,157,150
cultivation 5 dynamic simulations 518
batch 13
continuous 14, E
fed-batch 14 economic viability 371
culture supernant 528 Eddy viscosity 213
cybernetic model 81,83,283,287 effective viscosity 184
cybernetic variables 82 electrodialysis 371
cytoplasmic extract 528 electro neutrality condition 188
cytosol 438 elementwise multiplication 577
emulsifying agent 368
D endogenous metabolism 66, 88
dairy product 349 energetic parameter 361
declining phase 48 energy balance 376,517
dedicated fructose-PTS 355 enrichment fraction 573
degree of freedom 440 enrichment vectors 574
dehydrogenase pathway 539 Entner-Doudoroff 478
dehydrogenase reaction, isocitrate 546 environmental growth control function 191
DEPT 563 enzymatic cycling 481
design of production plants 169 equilibrium 148
design strategy, model-based 167 fast 524
development of complex processing equilibrium point 149
strategy 202 errors
a-d-glucose 365 experimental 578
p-d-glucose 365 random 126
diacetyl 327 standard 128
diaminopimelate pathway 520 systematic 126
diauxic behavior 195 erythrocytes 591
diauxic growth 58, 66, 80 Escherichia coli 110, 119, 199,374,375
diffusion coefficient 56 growth process 180
effective 56 Jl estimation 198
diffusion equation 54, 406 ethanol 478
Digital Control Unit (DCU) 194 inhibition 493
dilution rate 61,87,89,91,291,303 transmembrane diffusion 493,495
2,3 diphosphglycerate 588 Eulerian macro length scale 216
dispersion model 393 Eulerian two-fluid model 225
dissolved carbon dioxide 333 Euler-Lagrange approach 258
balance 186 experimental evaluation 195
dissolved oxygen extended Kalman Filter 339
balance 186 extraction 564
concentration 382
distorsions 137 F
distribution model 53, 303 fatty acid and cholesterol synthesis 586
age 299 FCC (flux control coefficient) 120, 121
disturbances 147,148 FDP 358
598 Subject Index

aldolase 355 gas-liquid mass transfer 259

fed-batch 25, 93, 283, 292, 299, 302, 309, 313, gas-mixing station 180
317,393,395,397,403,411,415,420,422,426 germination 409,411
cultivation 14, 65 (glc)-PTS 351
operation 35,37,50 gluconeogenesis 278
strategy 374, 386 gluconolactone 490
feed and titration vessels system 187 glucose 479
feed concentration 380 balance 186
feed-back 151 effect 279, 288
control 421,424 feeding profile 383
feedforward 151, 153 limitation constant 190
feeding rate 379 metabolism 494
fermentation unit 369 pulse 494
fermentative 295 glucose-6-phosphate branch point 532
fermentative activity 279, 283, 309 glutamate 550
fermentative metabolism 278 dehydrogenase 524, 546
FIA 130 production 546
filtration 370 glutaminase 585
final control element 173 glutamine 550
fitting 573 glutamine 2-oxoglutarate-
flocculating microorganism 28 aminotransferase 524
flow models 35 glutamine synthetase 524
fluorescence spectroscopy 482 glutaminolysis 589
flux glycerol balance 186
analysis 107 glycerol limitation constant 191
bi-directional 543 glycerophosphate 589
control 107,491 glycolysis 278, 284, 582
external 560 aerobic 558, 588
metabolite 572 glycolytic enzymes 491
quantification 107 glycolytic flux 491
flux control coefficient (FCC) 120, 121 glycolytic rate 490
food industry 368 glyoxylate pathway 522
formalkinetic model 51 glyoxylate shunt 524
formic acid 357 graphical analysis 575
fraction of budding cells 302,308,309,311 group control coefficient 120
fractional enrichment 562, 566, 572 growth-diauxy 292
freedom, degrees 566 growth energetics 363
Fru-6P/Glc-6P pool 365 growth kinetics 363
fructose 352, 365, 479 growth phase, exponential 47
metabolism 494 growth rate control 378,381
full scope model 172, 173 growth, synochrized 282
functional state model 65 gylcolysis 519
fuzzy controller 342
fuzzy reasoning 325 H
fuzzy rule 325 2H 487
Hansenula polymorpha 38
G harvest rate 185
galactose 353, 365 heat generated by the stirrer 178
gas film resistance 28 heat produced by microorganism 178
gas hold up 22, 229, 248, 250 heat transmission coefficient 176
gas phase 12,378 heat transmission resistance 176
C02 mass balance 182 heuristic knowledge 325, 326
model 25 HFBRs 564
02 mass balance 182 high capacity fructose-PTS 352
gas removal 173 high cell cultivation 199
gas residence time distribution 37 high cell density cultivation 195
gas velocity, superficial 251 high purity lactide polymer 369
gas-liquid interfacial area 28 high-cell density cultivation 374,375
Subject Index 599

hybrid model 331 kinetic model 364

hybrid modelling 337 linear 120
hydro cyclone 370 (log-linear) 120
hydrolysed whey protein 370 kLa correlation 184
hydroxonium hydrogen ion 188 knowledge-based system 133
hyphae 391,396,402,403,406,409,410,412, k-Emodel 255
417,418,420
hyphal growth unit 402,403,408 L
L.lactis MG 1363 350
label balance 572-574
in vitro enzyme activity 109 label enrichment 112
in vivo diagnosis 446 label fluxes 572
in vivo NMR 482 label scrambling 520
inactivation time 485 labeling study, dynamic 550
induction 46, 82, 85, 288 lactate dehydrogenase 582
industrial scale 350 lactic acid bacteria 349
information content 527 lactic copolymer 368
inhibition 46,58,81,82,85,97,99,100,288,292, Lactobacillus 350
393,395,398,415 Lactococcus 350
inhibition constant for acetate uptake 191 Lactoscan process 370,371
inhibition constant for glycerol uptake 191 lactose consumption 353
inhomogenity 235 lag-phase 47,64,77,288,418,421
initial conditions 40 Large-Eddy-Simulation 254
inorganic N-source 359 LDH 358
input variables 4 leakage of metabolites 496
interface area, special 249 least squares 573
intermediary metabolism 44 least squares fitting, non-linear 541
intracellular ethanol 495 Leloir pathway 354
intracellular glucose 495 Leuconostoc 350, 355
intracellular resistance 28 limitation, multiple 59
intrinsic balance 93,417,418,421 linear combination 376
intrinsic concentrations 74,75,76,78,82,85 linear control 152
intrinsic rate 57 linear programming (LP) 110
intrinsic reaction 54,67,68,84 linearly dependent 360
rate 76 liquid circulation model 41
investment cost 369 liquid film resistance 28
ion-exchanger 371 liquid phase 12
ionic product of water 188 reaction system 178
isocitrate dehydrogenase reaction 546 liquid-phase model 24
isotope balancing 570,571,277 local controller 173
isotoper distribution analysis 115 LP (linear programming) 110
isotopic balance equation 519 Luedeking and Piret 366
isotopic steady state 113,518,528 Luedeking-Piret equation 67,68
isotopic-tracer techniques 112 lumping 513
isotopomer 113, 527, 556, 563 lysate 564
analysis 527 lysine biosynthetic pathway 520
balance 566, 577 lysine fermentation 531
equation 489 lysine selectivity 515
distribution 562
vectors 575 M
mapping matrices 530, 575, 576 M'Kendrick-von Foerster equation 102,297
isotype 112 macro kinetics 51
isozyme 558 macro length scale 216
macro mixer 32, 33
K macrokinetic model 494
Kalman filter 338, 424 macrokinetics 409
kinetic constant 492 macromolecular carbohydrate 587
600 Subject Index

macromolecular synthesis 586 microorganism

maintenance metabolism 66,68, 76 flocculating 28
maintenance requirement 284 pellet-forming 28
malate shunt 585 micro spheres 564
malate/aspartate shuttle 585, 589 micro tubules 469
malic enzyme 534, 583 mitochondria 438
mannose 480 mitochondrial compartmentalization 584
mannose-PTS 355 mitochondrial malic enzyme 590
man-PTS 351 mixing 173, 248
mass balance 376 mixing characteristics 235
mass balance aerobic growth process 180 mixing model, two-region 33
mass transfer 249 mixing time 23, 239
coefficient 30,243 mode number 305, 307
gas-liquid 28 model reduction 159
mass transport behavior 173 model structure 380
massflow controller 180 model
mass-spectrometry 112 calibration 129
material balance 234,377 compartment 73, 77, 78, 80, 283
equations 235 cybernetic 81,83,283,287
mathematical model 3 distribution 53
matrix formalism 489 dynamic 146, 147, 150, 157
matrix, stoichiometric 569 formalkinetic 51
maximum 02 saturation concentration 183 non-interacting 61
maximum 02 transfer rate 183 segregated 52,55,86,87,94,96
maximum oxygen supply rate 182 single-cell 73
maximum specific growth rate 147,148 stoichiometric 286
MCA (Metabolic Control Analysis) 120,436, structured 52, 73, 77
446,456,462,565,592 unstructured 52, 54, 56, 86
measurement, redundant 578 model-based design methods 168
mechanistic modeling 488 modelling cycle 446
membrane reactors 11 modes of operation 13, 16, 27
membrane transport 488 contino us 16
metabolic balancing 486 batch 16
metabolic bottleneck 480 batch 16,25
metabolic burden 388 chemostat 25
Metabolic Control Analysis 120,436,446,456, continuous 25
462,565,592 fed-batch 16,25
Metabolic Control Theory 488 micro-aerobic 17
metabolic coordinator 84 molar C02 total concentration 184
metabolic cycles 517 molar H+ concentration 184
metabolic engineering 106,435,506 molar respiration quotient 182
metabolic flexibility 111 molecular diffusion 30
Metabolic Flux Analysis (MFA) 107,375,481 Monod model 38, 70
metabolic genotype 11 0 Monod-Blackman model 38
metabolic network 106 Monod-kinetics 57,58,88,98,99,287,395,407,
metabolic regulation 83,85,393,415,481,500 421
metabolic steady-state 108 Monte Carlo simulations 579
metabolism morphological differentiation 50, 69, 86, 92
endogenous 88 morphology 404
fermentative 278 morphology (of Acremonium
respiro-fermentative 278 chrysogenum) 412
metabolite balances 108, 568 morphology (pellet) 410
studies 530 morphology of cell 391,392
metabolite fluxes 572 multi-phase conservation equation 225
metabolome 436 multiple impeller 223
metabolon 587 multiple limitation 59
micro mixer 32, 33 multiple reference frame 211
micro kinetics 51 multiple-cell circulation model 36
Subject Index 601

multiply-labeled Be substrate 487 ornithine 363

multisubstrate reaction scheme 190 oscillation 282,296,299,302-307
mutation 96, 99 outlier test 138
mycelia 406 output variables 4
overdetermined 531
N overflow metabolism 387
15N 487 overshoot of p 357
15N labelling 551 over-supply of substrate 197
NADP(H) redox balanced 546 oxaloacetate decarboxylase 536
NADPH 533 oxidative 295, 302
regeneration 547, 548 oxidative growth 278
NAH/NAD+ ratio 358 oxidative phosphorylation 443,453
Navier-Stokes equation 213 oxidoreductive growth 302
system 252, 253 oxygen
network rigidity analysis 530 anaerobic 16
network theory 375 micro aerobic 16
network underdeterminancy 514 oxygen balance 378
neural networks 131 oxygen concentration 269
nitrogen assimilation 508 dissolved 267
nitrogen fluxes 551 oxygen consumption rate 22
nitrogen uptake 359 oxygen feeding strategy 385
N-limited batch fermentation 366 oxygen limitation 61,241,313,393,394, 4ll,
NMR (Nuclear Magnetic Resonance) ll2,562 422
bioreactor 493 oxygen mass balance 181
fine structure of ll2 oxygen supply 381
membrane-cyclone bioreactor rate 181
system 550 oxygen transfer capacity OTC 183
spectroscopy ll3 483 oxygen transfer profile 383
spectroscopy 481 oxygen transfer rate 183, 295, 382
spin transfer 488 oxygen uptake rate 317
two-dimensional ll8 oxygen-limited 287
node rigidity ll8
NOE 563 P
non-dissociated C02 concentration 183 PIO ratio 444,517
non-interacting model 61 parameter 4
nonlinear control 153, 154 parameter identification 5, 68, 456
non-Newtonian media 32 parameter model, lumped 31
Nuclear Magnetic Resonance (NMR) ll2,562 Pasteur-effect 279, 288
pathway 478
o pathway flux 568
02 bottle-neck 192 pathway inclusion 582
02 concentration in liquid bulk 183 pathway manipulation 121
02 equilibrium concentration 183 PDH catalysed pathway 356
02 mass fraction 180 pellet 54,56, 39l, 396,399, 406-4ll
02 supply rate 182 pellet-forming microorganism 28
02 transfer mass flux 198 Penicillin 93
02 transfer rate 182 Penicillium chrysogenum 34
02-maintenance rate 201 pentanedione 327
objective function 570 pentose phosphate pathway 355,519,585,591
observation 386 PEP carboxylase 536
of growth rate 198 peptidoglycan 510
p-observer 198 perfused culture 561
on-line measurement system 173 perfusion 560
on-line simulation 168, 194 perturbation, singular 159
operation and visualization 173 pH 70,92
optimality criterion 81 auxostat 14
optimality principle 47 buffering agent 368
optimization 395,397,421,423-426 equation 188
602 Subject Index

phenomics 593 R
phenotypes 593 radial liquid velocity profiles 37
phosphoenolpyruvate carboxykinase 587 radial-tangential jet 211
phosphoglucose isomerase reaction 525 random errors 126
phosphoric acid balance 187 rapid sampling 484
phosphorylation of the sugar 349 rate determining step 189,192
phosphotransferase system (PTS) 351 rate limiting 413, 417, 418
physicochemical behavior 173 steps 46,84,283,288-291,402,565,592
physiological model 375 rate of C02 evolution 333
PI (Proportional Integral) 152, 153 reaction
pilot plant ATL reactor 37,38,39 coefficient 568, 569
PNMR 482 model 189,377
PolV (specific power input) 23 specific 286
p02/agitation-control 199 reaction rates, bi-directional 588
p02/feed con troll 197 reaction steps, reversible 525
polylactide 369 reactor flow model 210
polymer film 368 real-world application 168
pool size 578 recirculation model 32
population balance 404 recombinant DNA 199
model 101, 403, 409 recombinant organism 95
power input 28 recombinant protein 374,387-389
precursor 393,400,403,413 redox metabolism 546
preparation 5 reducing equivalent 589
primary metabolites 49,67 redundancy 107
principal nodes 118 matrix 579
process control 339 redundant measurement 578
system 7 regression 81
process model 4 regulation 46,77,81,84, 150,279,283,287,416
process optimisation 322 regulator, metabolic 287, 288
process supervision 322,331 reliability 125
product formation 67,68,69,70 removal oflactic acid 370
kinetics 363 repression 46, 82, 85, 288, 292, 293, 417
production phase 49, 50 repression (catabolic) 421
production rate 360 residence time distributions (RTDS) 23,37
professional training 170 respiration 45
profit function 426 respiratory chain 284,288
program package SLDGL 40 respiratory quotient 316
Programmable Logic Control 7 retentate 371
propionic acid bacteria 11 0 reuse of acetate 195
Proportional Integral (PI) 152, 153 reversibility, degree 543
protein production 389 Reynolds equation 213
proteome 436 Reynolds-stress 21 0
proton NMR spectroscopy 539 rigid branchpoint 533
pseudo-steady-state approximation 568, 580 Riser 10
pulse experiments 235 RNA level 358
pumping capacity 228 RNA synthesis 587
purification 37l robustness 384
pyruvate branch point 532 RTD probability functions 34
pyruvate carboxylase 527,582 RTDS (residence time distributions) 23,37
pyruvate carboxylase 590 Runge-Kutta formula 172
pyruvate dehydrogenase 582 Rushton turbine 207
pyruvate, recycling of 536
S
Q Saccharomyces cerevisiae 38, 119,435,483,484
QIV (volumetric gas flow) 23 sampling rate 185
quenching of metabolism 482 sampling tube device 493
saturation 562
Sauter diameter 250
Subject Index 603

scale-up 207 stirred tank reactor 9, lO, 21, 22

secondary acidification 353 stirrer power input 184
secondary metabolites 50, 69 stoichiometric coefficients 70
sedimentation 370 stoichiometric model 286
segregated model 52,55,86,87,94,96,396,417, Streptococcus 350
419 structural part 366
segregational instability 96, 98 structured kinetic model 366
self-adaptive difference method 40 structured model 52,73,77,418
sensitivity function 380, 381 submerse bioreactor 21
shunt, malate 585 submodel of the bioreactor process 176
shuttle system 439, 451 double jacket 176
side product 478 liquid phase reaction system 176
signal transduction 122, 456 thermostat cooling 176
simple ha correlation 185 thermostat heating 176
simulation 379 substrate
ofHCDC 384,388 alternative 60, 64, 79
simulation systems, components 172 distributions 241
single fluid particle, trajectory 266 essential 60,61,64
single-cell model 73 feed rate 185
sliding mesh technique 211 growth-enhancing 63
slug flow 36 interacting 61
snapshot approach 211 limitation 197
solid-phase 12 maintenance rate 201
sophisticated control 170 multiple 63
sorbitol 490 non-interacting 61
special interface area 249 spectrum engineering 479
specific growth rate 53,57,61,63,66,67,69,72, toxic 58
73,76,82,84,88,93,97,147,152,155,158, substrate feeding 378
287,288,290,292,298,299,316,397,402,404, strategy 384
424 succinylase pathway 539
specific interfacial area 30 sucrose 479
specific power input 30 sugar flux 356
specific power input (PoN) 23 sugar metabolism 350
specific rates 53, 57 sugar tolerance 493
specific substrate consumption rate 324 sum-of-squares 580
specific surface area 243 superficial gas velocity 184
spinecho NMR 487 support panes 579
splitting 563 suspension culture 559-561
spore germination 404, 405 synchronization 302, 304
S-systems modelling framework 488 systematic errors 126
stability 148, 149
stable isotope labeling 486 T
standard addition method 130 Tagatose-6 P pathway 354
standard deviation 127 tank reactors, stirred 283
residual 129 TCA (tricarboxylic acid) 110
standard error 128 cycle 582
standard k-£ model 213 temperature 70,71,72,78,90,92,395,397,423
]I-stat problem 195 control 173
]I-stat procedure 202 tetrahydrodipicolinate branchpoint 539
state estimation 395 thermal transport time constant 176
state prediction 327 thermostat circulation heat resistance 178
state variables 4 thermostat cooling system 177
stationary phase 48 thermostat heating system 177
statistical test 127 tight regulatory control 359
statistical variance 529 tracking 150
steady state 148,364 training with virtual reaction process 194
isotopic 518, 528 training-simulation 171
stimulus response technique 33, 36 transaldolase 498
604 Subject Index

transcription 439,445 variance 128

efficiency 78 velocity pattern 264
transcriptome 436 vicinal diketone 327
transfer flux 26 virtual bioreactor 168
transhydrogenase 517,589 virtual mass force 212,227
system 549 virtual process 173, 194
transient response 578 virtual reality 170
transketolase 498 viscosity, effective 259
translation 439,445 volumetric gas flow (Q/V) 23
efficiency 78 volumetric gas transfer coefficient 28
transparent polymer 368 volumetric -mass transfer coefficient 22,30,243
transport genes 506 volumetric 02 transfer coefficient 183
transport system 438 volumetric reaction rate 186
trehalose synthesis 532 volumetric thermal capacity 176
tricarboxylate transport system 586
tricarboxylic acid (TCA) 110 W
cycle 278, 284 water evaporation rate 185
Trichosporon cutaneum 32 whey protein 364
triosephosphates 356
turbidostat 14 x
turbulence fluctuations 21,22 xanthan production 32
turbulence model 210 xylose 480
turnover 482 xylose isomerase 498
two-dimensional NMR 487 xylulokinase 498
two-fluid model 257
two-region mixing model 33 y
yeast extract 364
U yield coefficient 53,66,68, 75
UBICON 194 yields Ysp 349
unstructured model 52, 54, 56, 86, 395, 423
Z
V zymomonas mobilis 478
validating the automation hardware 202
validation'169