Advances in Biochemical Engineering

48 Biotechnology

Managing Editor: A. Fiechter

Bioprocess Design
and Control

With contributions by
R. Aarts, M. Aynsley, J. E. Bailey, R M. Doran,
I. J. Dunn, K. Friehs, E. Heinzle, A. Hofland,
K. B. Konstantinov, C. Di Massimo,
G. A. Montague, A. J. Morris, K. E Reardon,
G. B. Ryhiner, T. Yoshida

With 52 Figures and 21 Tables

Springer-Verlag
Berlin Heidelberg NewYork
London Paris Tokyo
Hong Kong Barcelona Budapest
ISBN 3-540-56315-6 Springer-Verlag Berlin Heidelberg NewYork
ISBN 0-3 87-56315-6 Springer-Verlag NewYork Berlin Heidelberg

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, re-use 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 a copyright fee must always be paid.

Springer-Verlag Berlin Heidelberg 1993

Library of Congress Catalog Card Number 72-152360
Printed of Germany

The use of 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.

Typesetting: Macmillan India Ltd., Bangalore-25

Printing: Heenemann, Berlin; Bookbinding: Ltideritz & Bauer, Berlin
02/3020 - 5 4 3 2 1 0 Printed on acid-free paper
Managing Editor

Professor Dr. A. Fiechter

Institut fur Biotechnologie, EidgenOssische Technische Hochschule
ETH - HOnggerberg, CH-8093 Ztirich

Editorial Board

Prof. 1-1.W. Blanch University of California

Department of Chemical Engineering,
Berkeley, CA 94720-9989/USA

Prof. Dr. H.R. Bungay Rensselaer Polytechnic Institute,

Dept. of Chem. and Environment, Engineering,
Troy, NY 12180-3590/USA

Prof, Dr. Ch. L. Cooney Massachusetts Institute of Technology,

Department of Chemical Engineering,
Cambridge, Massachusetts 02139/USA
Prof. Dr. A. L. Demain Massachusetts Institute of Technology,
Department of Biology Room 56-123,
Cambridge, Massachusetts 02139/USA
Prof. S.-O. Enfors Department of Biochemistry and Biotechnology,
Royal Institute of Technology,
Teknikringen 34, S - 100 44 Stockholm
Prof. K.-E. L. Eriksson Department of Biochemistry,
The University of Georgia,
Athens, Georgia 30602/USA
Prof. Dr. K. Kieslich Gesellschaft ftir Biotechnolog. Forschung mbH,
Mascheroder Weg 1, W-3300 Braunschweig/FRG
Prof. Dr. A. M. Klibanov Massachusetts Institute of Technology,
Department of Chemistry,
Cambridge, Massachusetts 02139/USA
Prof. Dr. R. M. Lafferty Technische Hochschule Graz,
Institut ftir Biochemie und Technologie,
Schlrgelgasse 9, A-8010 Graz
Prof. B. Mattiasson Department of Biotechnology,
Chemical Center, Lund University,
P.O. Box 124, S - 221 00 Lund
Prof. Dr. S. B. Primrose General Manager, Molecular Biology Division
Amersham International plc., White Lion Road,
Amersham, Buckinghamshire HP7 9LL, England
VI Editorial Board

Prof. Dr. H. J. Rehm Westfalische Wilhelms Universit~it,

Institut fiir Mikrobiologie,
Corrensstr. 3, W-4400 Mtinster/FRG

Prof. Dr. P. L. Rogers School of Biological Technology,

The University of New South Wales, P. O. Box 1,
Kensington, New South Wales, Australia 203

Prof. Dr. H. Sahm Institut for Biotechnologie,

Kernforschungsanlage Jiilich, W-5170 Jiilich/FRG

Prof. Dr. K. Schiigerl Institut for Technische Chemie, Universit~it Hannover,

CallinstraBe3, W-3000 Hannover/FRG

Prof. Dr. U. von Stockar Institute de Genie Chirnique,

Ecole Polytechn. Federale de Lausanne,
EPFL - Ecublens, ETH Lausanne, CH - 1015 Lausanne

Prof. Dr. G. T. Tsao Director, Lab. of Renewable Resources Eng.,

A. A. Potter Eng. Center, Purdue University,
West Lafayette, 1N 47907/USA

Dr. K. Venkat Corporate Director Science and Technology,

H. J. Heinz Company U.S. Steel Building, P. O. Box 57
Pittsburgh, PA 15230/USA

Prof. Dr. C. Wandrey Institut fiir Biotechnologie, Inst. 2,

Forschungszentrum Jiilich GmbH,
Posffach 1913, W - 5170 Jiilich/FRG
Attention all “Enzyme Handbook” Users:

A file with the complete volume indexes Vols. 1 through 5 in delimited

ASCII format is availablefordownloading at no charge from the Springer
EARN mailbox. Delimited ASCII format can be imported into most
databanks.

The file has been compressed using the popular shareware program
“PKZIP” (Trademark of PKware INC., PKZIP is available from most
BBS and shareware distributors).

This file distributed without any expressed or implied warranty.

To receive this file send an e-mail message to:

SVSERV@DHDSPRI6BITNET.
The message must be: “GET/ENZHB/ENZ-HB .ZIP”.

SPSERV is an automatic data distribution system. It responds to your

message. The following commands are available:

HELP returns a detailed instruction set for the

use of SVSERV,
DIR (name) returns a list of files available in the
directory “name”,
INDEX (name) same as “DIR”
CD <name> changes to directory “name”,
SEND <j?lename> invokes a message with the file “filename”
GET -@lename> same as “SEND”.
Table of Contents

Artifical Intelligence and the Supervision of Bioprocess

M. Aynsley, A. Hofland, A. J. Morris, G. A. Montague,
C. Di Massimo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Host-Vector Interactions in Escherichia coli

J. E. Bailey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

Parameters Influencing the Productivity of Recombinant

E. coli Cultivations
I~'. Friehs, K. F. Reardon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Modeling and Control for Anaerobic Wastewater Treatment

E. Heinzle, I. J. Dunn, G. B. Ryhiner . . . . . . . . . . . . . . . . . . . . . . . . 79

Design of Reactors for Plant Cells and Organs

P. M. Doran .......................................... 115

Expert Systems in Bioprocess Control: Requisite Features

K. B. Konstantinov, T. Yoshida, R. Aarts . . . . . . . . . . . . . . . . . . . . . 169

A u t h o r I n d e x Volumes 1 - 48 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Artificial Intelligence and the Supervision
of Bioprocesses (Real-Time Knowledge-Based
Systems and Neural Networks)

M . A y n s l e y , A. H o f l a n d , A.J. M o r r i s , G . A . M o n t a g u e a n d C. D i M a s s i m o
Department of Chemical and Process Engineering, University of Newcastle,
Newcastle upon Tyne, NE1 7RU, England

Dedicated to Prof. Karl Schiigerl on the occasion of his 65th birthday

Abbreviations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 The Bioprocesses Studied . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 An Industrial Penicillin Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 An Industrial Mycelial Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3 Real-Time Knowledge-Based Supervision of Bioprocesses . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1 The Real-Time Knowledge-Based System Shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Bioprocess Supervision Control and Analysis (Bio-SCAN) . . . . . . . . . . . . . . . . . . . . . 7
3.3 Knowledge-Based Systems for Bioprocess Scheduling Operations . . . . . . . . . . . . . . . . 8
3.4 Bioprocess Database System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.5 System Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.6 System Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.7 Integration with Process Analysis and Control Software . . . . . . . . . . . . . . . . . . . . . . . 14
4 Application of Real-Time Knowledge-Based Bioprocess Supervision . . . . . . . . . . . . . . . . . 14
5 Bioprocess State Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
6 Artificial Neural Network Based Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
7 Applications of Neural Networks to Bioprocess Supervision . . . . . . . . . . . . . . . . . . . . . . . 20
8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
9 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

The ability to supervise and control a highly non-linear and time variant bioprocess is of
considerable importance to the biotechnological industries which are continually striving to obtain
higher yields and improved uniformity of production. Two AI methodologies aimed at contributing
to the overall intelligent monitoring and control of bioprocess operations are discussed.
The development and application of a real-time knowledge-based system to provide supervisory
control of fed-batch bioprocesses is reviewed. The system performs sensor validation, fault detection
and diagnosis and incorporates relevant expertise arid experience drawn from both bioprocess
engineering and control engineering domains. A complementary approach, that of artificial neural
networks is also addressed. The development of neural network modelling tools for use in bioprocess
state estimation and inferential control are reviewed. An attractive characteristic of neural networks
is that with the appropriate topology any non-linear functional relationship can be modelled, hence
significantly reducing model-process mismatch. Results from industrial applications are presented.

Advancesin BiochemicalEngineering
Biotechnology,Vol.48
ManagingEditor:A. Fiechter
Springer-VerlagBerlinHeidelberg1993
2 M. Aynsley et al.

Abbreviations

AI artificial intelligence
CER carbon dioxide evolution rate
ICP intelligent communication protocol
KBS knowledge-based system
OUR oxygen uptake rate
RQ respiratory quotient
RTKBS real-time knowledge-based system
SQL Structural query language
TCS Turnbull control system
Artificial Intelligenceand the Supervision of Bioprocesses

1 Introduction

The requirement to operate biotechnological processes in a cost effective

manner is becoming increasingly relevant as the bioprocessing industry devel-
ops and market competition increases. This situation is seen both in the more
established product areas, such as penicillin production, as well as in more
recently developed processes such as the production of protein. Historically, the
most important means of achieving increased productivity from bioprocess
plant has been through strain improvements and media developments. More
recently, however, significant advances have been made in the area of bioprocess
supervision and control. This alternative, complementary, route to process
improvement has many potential benefits, although many problem areas are
still to be overcome before the application of advanced supervision and control
techniques becomes common place.
Typically, little use is made of the large amounts of bioprocess on-line
and laboratory assay data after a batch is completed (Stephanopoulos and
Tsiveriotis [1]). Recently, attempts have been made using knowledge-based
systems (KBS) approaches to improve the quality of information presented to
operators (Karim and Halme [2]) and increase the level of automatic process
supervision (Cooney et al. [3]), although few industrial applications have yet
been reported. On a plant-wide perspective, the scheduling of process operations
is also traditionally manually undertaken by experienced personnel using a trial
and error approach with a planning board. This also appears to be a potential
area for the application of a KBS (Hofmeister et al. [4]). In the particular fed
batch bioprocess described, the "out of sequence" preferential harvesting to
maximize productivity, while ensuring that this would not result in lengthy
down-times for other process units, could yield a very significant improvement
in the economics of production.
Knowledge-based systems can be designed to cope with uncertainty and
allow the coupling of quantitative information with the qualitative or symbolic
expressions (in the form of heuristics) so as to reproduce the actions of an
experienced process operator. They might therefore be used as an intelligent,
on-line assistant to the process operator, or, with extra knowledge as a
supervisor in running and maintaining bioprocess operations within optimal
operating conditions. The main objective in this approach is to develop an
expert supervisory system which uses bioprocess and control knowledge in the
form of rules in conjunction with bioreactor state estimation (soft-sensing),
parametric identification and process control algorithms running in real-time.
In addition, provision needs to be made for process models and known
bioprocess behavioural characteristics to be incorporated in order to enhance
the overall supervisory control strategy. The complete system aims to perform
process monitoring, process control, sensor validation, fault detection and
diagnostic tasks and in addition, provide recovery advice so as to achieve
continuous on-line optimisation.
4 M. Aynsleyet al.

An integrated system for improved supervisory control and scheduling of

bioprocesses is proposed using a KBS approach. The system will comprise a
bioprocess supervisory knowledge base and a plant scheduling knowledge base,
both of which will be linked to an on-line relational database to provide a
maximum level of integrity and intelligence in a real-time environment.
Collateral with the above approach is the on-line determination of the
critical variables which govern process production (e.g. biomass concentration,
growth rate etc.). Halme [5], and Montague et al. [6] have reviewed the
measurement and estimation problems and techniques in bioprocess systems.
Halme [5] indicates three ways of reducing the difficulties encountered:

- better sensors
-better sampling and automatic analysis systems
-on-line estimation of difficult to measure, or unmeasurable, variables.

Although all of these are attracting research effort, the techniques of on-line
state and parameter estimation are receiving the greatest interest.
Whilst developments in biosensors are in some cases helping to contribute to
the on-line determination of bioprocess parameters, they are by no means at the
stage where general applicability has been achieved. Until this is the case the
development of on-line estimation techniques will provide the major means by
which on-line closed loop feedback control of critical bioprocess variables can
be achieved. (The need to obtain information from process variables which are
difficult to measure is not restricted to any one type of process. Problems can be
identified in such diverse industries as for example the processing of food,
chemicals, polymers, minerals, etc.)
Several different approaches can be adopted in the development of bio-
process estimation algorithms. The easiest, but potentially least accurate, way of
obtaining an estimate of a "difficult to measure" primary process variable is to
establish a correlation with a measurable secondary process variable, whilst
ignoring measurement errors, noise, etc. A more robust approach might be to
adopt a numerical estimation technique, either for estimating the parameters of
a pre-defined model structure, usually of a generalised linear time series form, or
directly to obtain an estimate of a difficult to measure primary process variable
(state estimation or soft-sensor). An alternative approach is to exploit the
nonlinear mechanistic structure of the bioprocess, if known, in the form of a
state observer.
A relatively new development in artificial intelligence is the area of neural
computing. Neural networks are dynamic systems composed of highly intercon-
nected layers of "simple" neurone-like processing elements. In the field of
process engineering it is their ability to capture process non-linearities which
offers the potential benefits. Other useful characteristics are their ability to
adjust dynamically to environmental and time-variant changes, to infer general
rules from specific examples, and to recognize invariances from complex high-
dimensional data. These properties provide neural networks with the potential
to out-perform other "learning" techniques. Given a series of examples the
Artificial Intelligence and the Supervision of Bioprocesses 5

network is able to establish the governing relationships in the training data. This
ability can be exploited to aid the nonlinear modelling, control and optimisation
of complex processes, as well as being used in existing predictive control
methods.
Two industrial systems are used to demonstrate the application of the AI
methodologies reviewed the fed-batch production of penicillin G and a large
scale mycelial process which operates in a continuous mode.

2 The Bioproeesses Studied

2.1 An Industrial Penicillin Process

The production of an established product such as penicillin is a highly com-

petitive business. Even small savings that could be achieved through the
application of improved supervision and control are important in the gaining of
a market edge. The industrial production of penicillin is achieved by a fed-batch
process in which two distinct operating regimes can be identified. In the early
stages the process is operated to produce large quantities of biomass, predomi-
nantly utilising the substrate in the initial batched media. Towards the end of
this phase the feed additions being made to the bioreactor are increased as the
initial substrate becomes exhausted. During the second phase the substrate
additions are maintained at a rate which keeps the substrate concentration at a
low level. As a consequence of the low substrate concentration and the resulting
low growth rate, penicillin is produced by the organism. In order to maximize
the yield of penicillin it has been observed that the growth rate should be
maintained above a pre-determined minimum value. The higher the growth rate
is above this minimum constraint the lower the yield of penicillin.
The present operating regime of off-line analysis to determine the bio-
reactor conditions results in a conservative feeding strategy since samples are
taken relatively infrequently. It is highly desirable therefore to gain some on-line
insight as to bioprocess behaviour at a higher frequency so as to be able to
operate closer to the constraint. To increase the frequency of information
routinely available would require a move from off-line analysis to on-line
measurement or estimation. Unfortunately, to date instrumentation is not
available which provides the necessary data. It is for this reason that an observer
has been developed to provide this information and utilise it within a control
strategy in order to improve overall bioprocess operation.

2.2 An Industrial Mycelial Process

This particular bioprocess (Fusariumgraminearum)is operated in a continuous

mode, on a large scale, to produce biomass which on processing is marketed as a
6 M. Aynsleyet al.

meat analogue. It is described in more detail in Edelman et al. [7]. Regulation of

the biomass concentration is required in order to achieve the desired product
quality. The present regulatory philosophy is based upon the use of off-line
biomass assays, carried out every four hours. The results are available to the
process operator one to three hours after sampling. Based upon the assay result,
the dilution rate is adjusted in an attempt to compensate for process distur-
bances. The shortcoming of this control scheme is the low sampling frequency of
biomass concentration in relation to the process dynamics. Furthermore,
because the laboratory analysis of dry weight is subject to significant error, a
biased averaging of the last three biomass assays is applied. Consequently, data
up to 15 h old is utilized for feedback control of biomass. It is not therefore
surprising that the control scheme results in excursions outside acceptable
operating conditions. Carbon dioxide evolution rate measurements are avail-
able to the process operators almost continuously and have been shown to be a
good indicator of the level of biomass present in the bioreactor. However, the
relationship between biornass concentration and carbon dioxide evolution rate
has not been quantitatively analysed.

3 Real-Time Knowledge-Based Supervision of Bioprocesses

Real-Time Knowledge-Based Systems (RTKBS) can provide novel solutions to

many problems encountered in the operation of time critical processes. Chantler
[8], Stephanopoulos and Stephanopoulos [9] and Moore et al. [10] have
discussed the basic requirements of RTKBS which are now becoming well
established in a wide range of application domains. It is only very recently,
however, that reports have appeared on the application of expert systems in the
field of bioprocess control [2, 3, 11].

3.1 The Real-Time Knowledge-Based System Shell

The G2 Real-Time Expert System [12] provides a flexible framework within
which to construct a knowledge base, but with much more powerful features
than those usually provided by a shell. The concept of real-time is represented
within the system in terms of update intervals which direct the supervisory
software to collect values for variables at regular time intervals; validity intervals
which specify the length of time the current value of a variable remains relevant,
and rule scanning intervals which inform the inference engine how often to
invoke any given rule. Currently valid data can be reasoned with directly or can
be analysed alongside past data using statistical features. It is also possible to
import functions written in "C" or Fortran which can be used to carry out
complex numerical calculations. Rules are written in a very expressive English-
like language and have an icon based graphical interface to create knowledge
Artificial Intelligence and the Supervisionof Bioprocesses 7

frames (i.e. an object-orientated approach). The system also supports arithmetic

and symbolic expressions and has a built-in simulation facility which can
be used in the "background" while the actual process is being controlled in
real-time.

3.2 Bioprocess Supervision Control and Analysis (Bio-SCAN)

Bio-SCAN is a real-time knowledge-based system, developed using G2, which
has been designed to provide a general purpose framework for the monitoring
and supervisory control of a range of bioprocesses. Internally, the system
incorporates knowledge of operating strategies (e.g. feeding regimes, process set-
points) and expected behaviours to enable rapid detection of any process faults.
When faults occur, the system attempts to diagnose the cause of the problem
and will offer recovery advice to the process operators. In addition to knowledge
encoded within Bio-SCAN, the utility of the system has been extended by
interfacing to conventional algorithmic methods of analysing and controlling
bioprocesses. A toolkit of on-line modules has been developed to incorporate a
number of features into the overall control strategy such as simulation and
prediction, statistical evaluation and bioreactor state estimation algorithms. An
underlying philosophy in the development of Bio-SCAN has been to make the
solutions as generic as possible (applicable across a broad spectrum of bio-
processes) and to provide the system with as much (high quality) information as
possible. This led to the development of a single data depository (database) for
on-line interrogation by Bio-SCAN. Further details about the development of
the system can be found in Aynsley et al. [13].
System development begins with the introduction of objects necessary for
the description of the process under study. The objects are assigned attributes
(usually symbolic or quantitative variables) and the source of their values. Rules
are written which inform Bio-SCAN about what to conclude or how to respond
to changing conditions in the domain defined by the objects. An intelligent end-
user interface is also provided in the form of read-out tables, graphs, meters,
dials and controls.
Bioprocesses employing different operating philosophies will require differ-
ent specific knowledge (rule-sets) for optimal supervisory control. These "super-
visory rule-sets" are used to govern the general operation of the process by
identifying growth or production phases and determining which rule-sets are
applicable to a given phase (i.e. supervisory rule-sets function as meta rules). The
supervisor is responsible for feed scheduling, or on-line modifications to the
schedule in response to changes in the process (i.e. faults) and for determining
the best time to terminate a reaction and harvest the product. Specific rule-sets
are also necessary in order to assess and respond to the metabolic state of a
process organism in any given bioprocess (i.e. detect "metabolic faults"). How-
ever, for the diagnosis of hardware faults generic rules can be used that apply to
all bioprocesses under Bio-SCAN control.
8 M. Aynsley et al.

The diagnostic knowledge base has been structured into a number of discrete
rule-sets with meta rules again determining which rule-set should be assessed
next. The decision to invoke a particular set of rules is based upon a comparison
of the current raw, or processed, data with that permitted or expected in the
current phase of the bioreaction (characteristic trajectories of any number of
variables of interest can be incorporated into the system in addition to con-
straints). This ensures that knowledge can be focused in response to the
occurrence of specific events.
Once a fault has been detected the operator is informed of the problem on a
message board and additional rule-sets are invoked in an attempt to establish
the nature of the fault. A list of possible causes is then presented to the operator
together with advice on how to correct the fault. If Bio-SCAN cannot solve the
problem automatically, the operator is requested to confirm a diagnosis (e.g.
contamination). Such advice is readily achieved when the process is sufficiently
overdetermined, since the knowledge-based system capitalises on redundancies
to test for data consistency and error identification. The source of these
redundancies is usually balances or load cells, although software sensor estima-
tion of variables reduces the need for hardware redundancy [-14, 15].
Since structured, generic rule-sets have been used in the development of the
system, coupled with an object-orientated approach to the description of the
bioprocess, it is hoped that much of the knowledge base will be generally
applicable across a wide range of bioprocesses.

3.3 Knowledge-BasedSystems for Bioprocess

Scheduling Operations
Commercial scheduling software is, in general, little used in the process in-
dustries since it is considered to lack the flexibility and level of detail required for
day-to-day plant operation, see DTI Report [-16]. This may, in part, be due to
the nature of the predominantly algorithmic methods used in these packages
(e.g. mixed integer linear programming) which may fail to converge in a
reasonable period of time or may require oversimplifications to be made in
defining the problem. Moreover, no allowance is made for constraint relaxation,
which may be the only method of generating a feasible solution under certain
circumstances. The idea of using a KBS approach is attractive in that these
systems are well suited to chaining through rules and constraints that govern the
heuristic approaches used on most plants. Such an approach may well achieve
the speed and flexibility required, albeit at the expense of the near optimal
solutions provided by the algorithmic methods [-17].
The industrial penicillin installation considered in this paper is a single
product plant comprising several production vessels with capacities in excess of
1000001, together with numerous storage facilities and a range of smaller seed
vessels. Once a seed vessel has been inoculated, the length of the batch can be
Artificial Intelligenceand the Supervision of Bioprocesses 9

manipulated, to a certain extent, during the first few hours but then becomes
committed to transfer its culture at a fixed time. An appropriate production
vessel must therefore be available at this time. Production batches then typically
run for around 200 h before harvesting and recovery of product. In order to
achieve maximum plant efficiency, production vessels must be utilized in such a
way as to ensure high potency media are always available at regular intervals so
that downstream processing facilities are continuously supplied. In theory this
simply requires an orderly sequence of operations involving seed inoculations,
growth of seed, transfer of culture to production vessels and subsequent growth
and product formation. In practice, however, the ideal schedule may not be
realized because of "upsets" caused by factors such as contaminations of seed or
production vessels, maintenance requirements, plant problems, etc. The prob-
lem is compounded in that there are different sizes of both seed and production
vessels and there may be constraints imposed on where a seed vessel can transfer
its culture. This can result in lengthy downtimes for production units if an
existing schedule cannot be changed by redirection of previously committed
seed cultures.
Further disturbances to the "orderly" harvesting of batches arises as a
consequence of the inherent batch to batch variability in production rate. At any
one time several batches on the plant may have achieved a sufficient titre to
make them candidates for harvesting. However, it is often found that total plant
productivity can be improved by harvesting a vessel out of its inoculated
sequence. For this reason an on-line forecaster has been developed to identify
poor performing batches based on off-line titre information. One method of
forecasting is based upon the classification of historical titre profiles and
identification of the current batch type to predict future behaviour. An alter-
native strategy is to employ a curve-fitting approach to a number of possible
model structures. Whatever strategy is used, the forecaster seeks to identify
which bioreactors should be dispatched to downstream processing as fast as
possible and which should be left to run-on. The scheduler must then decide if
harvesting a bioreactor out of sequence is workable and economically feasible
with regard to the current schedule in operation.
The (simplified) problem description outlined above is not of the conven-
tional flow or job-shop type described in the literature, see e.g. Ku et al. [18] and
is essentially one of on-line rescheduling. A solution to the problem using a KBS
approach is being developed which combines a model of the plant with a module
which will use rule-based heuristics to examine operating strategies. This does
not, however preclude the use of algorithmic methods if a mixed approach is
found to be needed. The model is being de.signed to incorporate a detailed
knowledge of plant structure and operation and is based on the activity-event
representations used in Critical Path Analysis [19]. In this approach however,
both activities and events exist as connected objects. Knowledge of how each
object exists in the context of plant operation is contained in an associated frame
representation. The model will be used to test what-if scenarios by process
operators in a manner similar to the system proposed by Bernstein et al. [20].
10 M. Aynsleyet al.

An "intelligent" user interface will allow operators to easily change constraints

or operating conditions and examine the effects. The knowledge base will be
particularly useful in analysing the feasibility of the harvesting sequence sugges-
ted by the on-line forecaster, and in training operator personnel. System output
will be in the form of text messages which can then be manually transferred to
the plant planning board.

3.4 Bioprocess Database System

The important problems to be tackled when developing bioprocess databases
are not merely how to collect large volumes of different types of data, but also
how to validate it before storage and then extract meaningful information. The
types of bioprocess data which are usually collected and stored have been
reviewed by Carleysmith [21]. In addition, information could also be held
concerning characteristic trajectories of key variables with their associated
standard deviations as an aid to process evaluation and fault detection by Bio-
SCAN. For the scheduling knowledge base the database would be required to
maintain a record of the current plant status with expected completion times of
process units (based on the current schedule in operation). Information con-
cerning the maintenance requirements of equipment and vessels could also be
incorporated [22], coupled with manpower availability and stock levels of
replacement equipment and raw materials. From a user-standpoint the database
must also be simple to interrogate, update and view data, and must be flexible in
design. Provision for report generation and the statistical analysis of data to
establish the relationships between variables would also be useful.
The system chosen to develop the database was INGRES 1, by Relational
Technology Inc., which operates a powerful relational database management
system [23]. INGRES proved simple to design and update via its form-based-
interfaces and supports a powerful query language (SQL) based on virtually any
combination of criteria. SQL can also be embedded in external host code, such
as "C', which proved important in establishing the link with the scheduling and
supervisory knowledge bases.

3.5 System Configuration

The experimental environment which has been developed is shown in Fig. 1. The
system is composed of standard industrial process control and signal processing
devices, an intermediate bioprocess control computer and a supervisory control
computer. The process control and signal processing devices are supplied by
Turnbull Control Systems (TCS), the intermediate bioprocess control computer

1INGRES is a trademark of Relational TechnologyInc.

Artificial Intelligenceand the Supervision of Bioprocesses l1

off-gas

Air flow
measurement
] C02
analyser
~ 1~. ample
_J off-gas
I
I
i
. . . . . . . . . ]
i
i

TCS Serial
link Ethernet link
signal -.4F -- --I~-

processor

Serial link
I
T
Air Feed I
pump
I- i Balance

Fig. 1. Experimentalenvironmentfor bioreactor supervisorycontrol

is an IBM PS2/70 and the supervisory computer is a SUN 3/60 Workstation.

The TCS devices provide the interface to the bioprocess under study. They are
linked via a multi-drop RS422/RS232 data highway to the bioprocess control
computer. This control computer communicates to the supervisory computer
via an ethernet network. The whole structure is flexible, reliable and reflects
those used in industry.
The bioprocess instrumentation is monitored at a fast sampling rate by the
TCS devices, with low level control actions being calculated and implemented
directly by the local process controllers and signal processors. The intermediate
bioprocess control computer interrogates the TCS devices at regular intervals.
The intermediate control computer also carries out signal filtering, data vali-
dation and inferential estimation using "soft-sensing" techniques. The condi-
tioned and validated data are then passed to the supervisory computer where it
is stored as a file.
On the supervisory computer the data from the intermediate control com-
puter is read by the Bio-SCAN system. This is achieved using the G2 Standard
Interface (GSI z) which provides facilities for the building of interfaces between
G2 and external processes and systems. The data read, and operated upon, by
Bio-SCAN is then written back into a file using the same GSI extension process.
This file is then read at regular intervals by the intermediate control computer
and the information in it is then passed on to the TCS devices.

2 GSI is a trademark of GENSYM.

12 M. Aynsleyet al.

For communication between Bio-SCAN and the I N G R E S database system

a different GSI extension process has been developed. Because of the very time
consuming nature of relational database management systems it was decided to
use a second SUN workstation as the database host machine. This posed an
additional communication problem which has been solved by splitting the
extension process in two separate processes, namely a Gateway and a Server.
The Gateway acts as a true GSI extension process on the Bio-SCAN host
machine 3, while the Server runs on the database machine. The processes
communicate with each other using the Remote Procedure Call mechanism
(RPC 4) which implements a machine architecture independent way of com-
munication between two computer systems connected by ethernet. Whenever
Bio-SCAN needs to interact with the database a request, together with associ-
ated information and timing data, is passed via GSI to the Gateway. The
Gateway passes this request on to the Server and monitors the progress of the
Server by means of the timing information supplied by Bio-SCAN. If for some
reason the Server fails to respond within the specified time, the Gateway returns
an error message back to Bio-SCAN. Under normal conditions the Server will
be able to respond to t h e request within the specified time. In this case the
Gateway returns a completion code and any additional information that may
have been sent by the Server. In order to improve robustness the Gateway and
Server have been designed to implement a stateless protocol. This means that
each request from Bio-SCAN to the database contains all the information that
is needed. Successful completion of a request does not depend on any previous
or subsequent requests. This allows Bio-SCAN to recover from machine and
network failures gracefully.
Both GSI interfaces described above have been implemented in "C" using
the hooks provided by GSI which are in the form of "C" function stubs. The
body of these function stubs needs to be written by the GSI user. Whenever
Bio-SCAN/G2 needs information from the GSI interfaces, it will instruct the
extension process to execute the appropriate "C" function. The file interface and
the Gateway are both examples of this approach. The Server has also been
written in "C". However, rather than being controlled by the GSI, it is
controlled by a R P C and the stateless protocol mentioned previously. For the
database operations, an embedded SQL has been used.

3.6 System Operation

The supervisory knowledge-based system (KBS), Bio-SCAN, has been designed
to oversee the operation of a range of (potentially different) bioprocesses

3 It is not strictly necessary to run an extension process on the same machine as the G2 real-time
expert system. G2 and GSI use the Intelligent Communications Protocol (ICP) which allows
extension processes to be run on different host computers and to communicate via ethernet.
4 RPC is a trademark of Sun Micro Systems.
Artificial Intelligence and the Supervision of Bioprocesses 13

simultaneously, by providing a permanent, consistent level of "expert" super-

vision and monitoring. Figure 2 indicates the structure of the software
environment.
When a seed or production reaction is to be initiated, requests will initially
be made to the database for retrieval of operating conditions and feeding
regimes for the strain of organism and mode of operation specified by the
scheduler. Once a reactor is running, data from the process is received by Bio-
SCAN at regular intervals and is analysed with respect to both expected and
recent values (there is provision for limited short-term storage of data within the
KBS), and for deviations from "optimal" behaviour. In checking for data
consistency, the supervisor also validates the data before it is passed to the
database for permanent storage using the protocol detailed above. During the
course of a run, requests may also be made for retrieval of, for example, a
standard profile of alkali addition or CO2 evolution to evaluate bioreactor
performance. When faults are detected and diagnosed, text messages will be sent
to the database specifying the nature of the fault and the action taken. Should a
contamination occur, a message is also sent to the scheduler which will then pass
back an instruction to stop feeding and terminate the run when an available slot
in downstream processing becomes available. Once a batch has been completed,
data collected during the course of the run is integrated into standard profiles
and abstractable data, such as the specific production rate, can be derived.

State Estimation] ~ , , . I
,( S o f .t S e n s i n. g ,~ t~ - - [ ~S i m u. l a t i o n n~ IOptimisationl
P r e d i c t i o n of' | ~
Future Control [ ~

#
Bio - SCAN
Monitoring and Supervisory Control
- Phase Identification
- Operating Strategy
- I n t e r n a l Models
- Expected Trajectories

-
-
- Fault Detection & Diagnosis
Sensor Validation
R e c o v e r y Advice ~ l n t e llige n t /
User
Interface
~,9 ..................
[

Local C o n t r o l C o m p u t e r ] Off-Line

Fig. 2. Supervisory monitoring and control software environment

14 M. Aynsleyet al.

3.7 Integration with Process Analysis and Control Software

Whilst the knowledge base described above is expected to be capable of

adequately supervising bioprocesses under most circumstances, the utility of the
system can be greatly extended by interfacing to conventional real-time algo-
rithmic methods of analysing and controlling bioreactions. For this reason a
toolkit of on-line modules is being developed that will incorporate the following
features into the overall supervisory control strategy:
Measured signal conditioning, statistical screening and interrogation to place
confidence bounds around data.
-Simulation using process models in imported subroutines.
- Identification of process model parameters using on-line recursive parameter
estimation algorithms.
Estimation of hidden states using 'soft sensors'.
-Prediction of future process outputs given pre-specified open-loop feeding
regimes. Use of adaptive-predictive control schemes to achieve closed-loop
profiling of bioprocess states.
On-line constrained optimisation using on-line identified process models that
capture the essential non-linear structure of the bioprocess.
Continuous statistical monitoring and process control to provide inter-
active assessment of overall bioreaction performance against pre-specified
specifications.
-Neural Network modelling of non-linear bioprocess models for hidden state
estimation, fault detection, and predictive control.

4 Application of Real-Time Knowledge-Based

Bioprocess Supervision

Results are presented using data monitored from an industrial fed-batch

penicillin process. In these bioreactions the respiratory quotient is usually high
and steady and so CO 2 evolution rate or 02 consumption rate may be used for
data analysis. Figure 3 compares the typical C02 profiles observed in two
"normal" industrial fed-batch penicillin reactions with those of from "abnor-
mal" runs. For proprietary reasons the abscissa scaling in this figure has been
omitted. An example of a very simple rule which will detect the possibility of an
abnormality in the bioreaction is of the general form:
IF the C02 production rate or rate of change of CO2 production is
greater or less than that expected
THEN there is evidence of a fault.
Clearly this rule needs to be expanded to incorporate the limits at which the rule
Artificial Intelligence and the Supervision of Bioprocesses 15

Contaminated Batch

,=,,
Feed Pump Failure
Z

8
(3
at

i i 410 i i i i f i i t i i i i I t l

20 60 80 1 O0 120 140 160 180 200

TIME (hi

Fig. 3. Carbon dioxide evolution rate test profiles (process Faults)

will fire, how these limits could change during the course of a reaction and
perhaps the probability of a process fault. The CO 2 must also be reconciled with
other process measurements such as the oxygen uptake rate and dissolved
oxygen concentration before diagnostic rule-sets are subsequently invoked. In
the contaminated bioreaction shown in Fig. 3, Bio-SCAN detected the in-
creased CO2 production rate early in the batch. Since this reading was found to
be reconcilable with other measurements and no other abnormalities in the
operation of the process could be detected, the system concluded that a
contamination had most likely occurred and informed the operator after 17 h. It
is interesting to note that this contamination was actually detected by the actual
process operators after 95 h when laboratory analysis identified the presence of
a yeast and terminated the reaction. Had the reaction been terminated earlier a
considerable saving in substrate feed costs and plant operation costs would have
been realized.
In the case of the second abnormal reaction shown in Fig. 3, Bio-SCAN
detected an atypical fall in the CO2 measurement after 150 h which was not
reconcilable with any mini-harvests. As all other operational parameters ap-
peared normal, the system invoked a rule-set which established from balance
measurements that substrate flow into the vessel was below normal. The
operator was advised to investigate the cause of this problem (e.g. a blocked line
or pump failure). Once the problem has been corrected Bio-SCAN is "informed"
via the intelligent user interface. Supervisory rules then modify the feeding
strategy in an attempt to bring microbial growth back onto a pre-defined
trajectory and recover the productivity of the reaction.
16 M. Aynsleyet al.

Analyser Recalibration

F-

z
8
o

[ I I I I I ~ I I I I I I I I I I

20 40 60 80 ~00 120 140 160 180 200

TIME (h)

Fig. 4. Carbon dioxide evolution rate test profile (analyser Recalibration

Figure 4 compares the original two normal bioreactions with a batch during
which the CO2 analyser was detected to have drifted off calibration. Once the
suspected calibration error had been confirmed it was corrected at time 81 h and
normal bioprocess operation resumed.

5 Bioprocess State Estimation

In contrast to academic research environments, most industrial bioprocess
control policies are based upon the use of off-line assay information for process
operator supervision. This usually involves the removal of samples from the
bioreactor, laboratory analysis and finally operator (or control system) action to
correct any undesirable process condition. This philosophy has several impor-
tant consequences for the control of the system. Off-line sample analysis, due to
the staff and laboratory support, is usually a costly operation. A common
outcome being that the sampling frequency is reduced so as to minimize costs.
This can result in poor process regulation with an inability to react quickly to
any process disturbances. The problem of slow sampling is further compounded
by the delay induced in the measurement due to analysis. Process operators
acting on old information can further degrade the quality of bioprocess super-
vision. A major industrial requirement would therefore appear to be an al-
gorithm which reduces the off-line analysis frequency whilst at least maintain-
ing, preferably improving, the quality of the information available to the
Artificial Intelligence and the Supervision of Bioprocesses 17

bioprocess operator. The philosophy being that more frequent and more up-to-
date information enables improved process operation.
Numerous techniques for the estimation of important "hidden" bioprocess
variables papers have appeared in the literature since the early 1980s. The
methodologies developed have, however, yet to be proved completely reliable
and find acceptance with the industrial community. Industrial systems are
inherently non-linear and it is expected that the use of a process model which
reflects this essential non-linear structure would prove to be most beneficial. As
a consequence the estimation algorithms developed should either be inherently
based upon the non-linear structure or alternatively approximate the non-
linearities using a linear model with adaptation. Standard non-linear estimation
techniques are available, such as the extended Kalman filter [24]. They can,
however, suffer from numerical problems and convergence difficulties. Despite
these complications many successful applications have been reported, see e.g.
Svrcek et al. [25]; Stephanopoulos and San [26, 27].
Several examples were presented at the 1st IFAC Workshop on Modelling
and Control of Biotechnical Processes [28] in 1982, e.g. Dekkers [29];
Swiniarski et al. [30], and the following IFAC Symposium held in the Nether-
lands in 1985 [31], e.g. Ghoul et al. [32]; Montague et al. [33]. A clear message
from these early studies was that although model based estimation methods
could achieve quite reasonable performance, varying growth regimes and ill-
known process disturbance and noise characteristics could invalidate the pro-
cess description. Much of this early work concentrated on an extended Kalman
filter approach coupled with a mechanistic model, the accuracy of which played
a major determining factor in the quality of estimation. Deficiencies in the
process model, such as unmodelled process dynamics and model parameter
variations, have to a certain extent been overcome by applying adaptive
estimation techniques, e.g. Leigh and Ng [34]; Shioya et al. [35]; Chattaway and
Stephanopoulos [36]. In some cases estimation of the model parameters and
process state variables simultaneously is often necessary in order to obtain
reliable results.
A prerequisite for any model based estimation technique is the development
of a balanced mathematical model of the system under study. The model must
be balanced in the sense that a compromise must be made between the
conflicting requirements of dynamic complexity and the ability to obtain the
parameters of the model either through estimation or experimental measure-
ment. Several techniques have been suggested by Dochain [37] for the develop-
ment of bioprocess observers, as well as controllers [38]. The observers de-
veloped are either of fixed parameter form, partially adaptive form or fully
adaptive. The experimental validation of an observer for the on-line state
estimation of bioreactor performance (biomass and product) has recently been
presented by Dochain et al. [39].
Whilst a large number of process models are available, two major problems
are faced. Firstly, the parameters of the models are usually difficult to determine
experimentally. Secondly, and perhaps more importantly, the simplifications
18 M. Aynsleyet al.

made in any process model usually result in a need to adapt the model
parameters to compensate for dynamic model inaccuracies (model-process
mismatch). One approach that has been applied to the penicillin process in order
to provide an estimate of biomass concentration [40], is to estimate the
parameters of a CER model off-line and use the resulting model in conjunction
with an on-line estimator. This results in a partially adaptive observer. That is, a
priori knowledge of some of the model parameter values is required for observer
application, although it is possible to estimate some of the parameters on-line. In
such an observer, the error in the prediction of CER acts as the driving force for
state and parameter updating. Although convergence and stability proofs have
been developed for observers of the form described, providing certain key
criteria are met, partially adaptive observers tend to be insensitive to initial
conditions [37]. Adaptive observers that do not require any prior knowledge of
the model parameters are known as being fully adaptive. Here the parameters
are assumed to be unknown and are identified along with the state estimation.
An alternative starting point for the development of adaptive estimation
algorithms is to adopt the philosophy that the bioprocess dynamics can be
represented by a general linear input-output model. This approach is similar to
that adopted for the development of well known adaptive control laws, except
that an additional term representing the secondary process measurement is now
included. In the bioprocess plant the secondary measured variable, v, is CO2 in
bioreactor off-gas, the infrequently measured controlled variable, y, is biomass
concentration assay, and the manipulated variable variables, u, are the bio-
reactor feeds. The algorithm thus provides a means by which frequent estimates

Disturbances

Point
+

I ~ . J P,ooess

I y
i
j . . . . I
I Parameter i~1 _ 1
I Estimator I~1. . . . . .

19
I I Parameter
,~/I z estimates

Estimator ~

Only active when primary measurement available

F i g . 5. G e n e r a l i s e d linear model based estimator

Artificial Intelligenceand the Supervisionof Bioprocesses 19

of biomass can be obtained (at the rate of off-gas measurement) without

laboratory assay delay hence improving the information available to the process
operator. A representation of the estimation scheme is shown in Fig. 5. The
estimator has been applied to the industrial continuous bioprocess so that given
infrequent, and irregular, laboratory dry weight data, and frequent CO 2 off-gas
analysis and dilution rate data, it is able to predict hourly biomass concentra-
tions. Further details of the derivation of the adaptive inferential state estimator,
together with some industrial applications, are set out in Guilandoust et al. [14]
and Tham et al. [15].

6 Artificial Neural Network Based Estimation

An alternative learning technique now beginning to be applied is that of neural

networks. These are dynamic systems composed of highly interconnected layers
of "simple" neurone like processing elements. The potential of modelling process
engineering systems with neural networks was initially discussed by McAvoy
et al. [41] and Bhat et al. [42]. Results from their studies indicated a tremendous
potential of neural networks for process modelling.
Whilst many artificial neural network architectures have been proposed one
structure has been predominant; that is the feedforward network. A feedforward
neural network is made up of interconnected nonlinear processing elements,
termed nodes. Scaled data is presented to the network at nodes of the input layer
from where it is propagated to the output through intermediate layers. Each
connection has associated with it a scalar weight which acts to modify the signal
strength. The nodes in the hidden and output layers perform two functions. In
the most simple implementation, the weighted inputs and a bias term are
summed and the resulting summation is passed through a sigmoidal activation
function.
Following specification of the network inputs, outputs and topology, the
network is trained by successive presentations of input-output data pairs.
During the training, the data is propagated forward through the network, which
adjusts its internal weights to minimize the error between the true output and
the output produced by the network. The minimisation of the error can be
achieved by the application of suitable optimisation algorithms. After training
the network weights are fixed and the model verified on test data prior to actual
application. Further details concerning the development of artificial neural
network based models can be found in Willis et al. [43].
It is the topology of the network, together with the activation function which
provides the potential for non-linear system approximation. Although the
network possesses a very rudimentary structure, studies have claimed that any
continuous non-linear function could be approximated by a network, with a
topology consisting of two hidden layers [44, 45]. These proofs are of major
20 M. Aynsleyet al.

significance, in that an artificial neural network with the appropriate topology

can theoretically be used to model any system. To date a few successful gen-
eral process and bioprocess applications of this modelling philosophy have
been reported, e.g. Bhat et al. [46], Lant et al. [47], Thibault et al. [48], Willis
et al. [43].

7 Applications of Neural Networks to Bioprocess Supervision

As discussed earlier, a possible solution to the control problems encountered

with the mycelial process is to provide the process operators with a more
frequent measure of biomass concentration, say once an hour. On-line gas
analysis of carbon dioxide concentration provides the opportunity to obtain
these process variable estimates at an acceptable frequency. Using the Gen-
eralised linear estimator it has been shown that the gas analysis can be used in
combination with off-line biomass assays to estimate biomass concentration
[47]. The fundamental assumption of the estimator is that the relationship
between the process variables can be modelled by a linear time series. Although
in practice the process is non-linear, it is assumed that the dynamics can be
tracked by on-line adaptation of the linear process model. However, if an
improved process model was available then enhanced performance might be
expected.
With this aim in mind, a feedforward neural network was considered as a
possible vehicle for process model development. Carbon dioxide evolution rate
(CER) and reactor dilution rate have already been identified as important
variables in biomass estimation. Whilst other variables affect the biomass/
CER/dilution rate relationship, for example pH, temperature etc., tight envi-
ronmental regulation maintains a low variance in these variables. Network
inputs are therefore selected to be CER and dilution rate, with biomass
concentration being the network output. Not only is the current value of CER
fed to the network but also hourly and two hourly delayed values. This time
history serves to remove the necessity of dead-time specification, often a
problem in model construction. Six network inputs were specified with one
process output - the biomass concentration. The network topology consisted of
one hidden layer containing three neurones. The operation of the feedforward
network estimator can be related to that of the linear estimator, discussed in the
previous section, but with fixed parameters (i.e. non-adaptive). Referring to
Fig. 5, the neural network learning phase can be related to the "off-line"
identification of model (or estimator) parameters. Network prediction of bio-
mass concentration, given measurements of bioprocess feed rate and CER, is
similar to biomass estimation of a fixed parameter output estimator. With the
neural network, however, the "model" used for estimation is fixed parameter
and non-linear, rather than being based upon an adaptive linear model. The
Artificial Intelligence and the Supervision of Bioprocesses 21

quality of the model fit produced by the neural network approach can be seen
from the plots in Fig. 6, where the step-like curve is the laboratory (off-line)
biomass assay. The bias that can be observed between the biomass assay results
and the neural network estimate, from around 25 h onwards, is due to a drift in
the on-line CO2 analyser calibration. This was corrected at around 170 h. At
around 210 h a major plant steriliser failure occurs which takes about 5 h to
correct. It is encouraging to observe that the neural network estimator remains
reliable over the period that this large process disturbance is effective. The
variable frequency of the laboratory assays is also quite evident.
The continuous bioprocess considered is required to operate as closely as
possible to pre-defined "steady state" operating conditions. In this respect a
linear model approximation was acceptable in the vicinity of normal process
operating conditions. A fed-batch bioprocess, however, presents more of a
problem in terms of model development; the system passes through a wide range
of dynamic situations never achieving a steady state.
The measurement problems experienced in controlling the penicillin process
are similar to those experienced in the continuous bioprocess. It is desirable to
control the biomass growth rate at a low level to optimise the production of
penicillin. Whilst the growth rate can be adjusted by varying the rate of
substrate (feed) addition, it is unfortunately not possible to measure the rate of
growth on-line. The aim therefore is again to develop a model which relates an
on-line bioprocess measurement to the process variable of interest. In this case
CER was used as the major variable from which to infer biomass concentration.
It is worthy of note, however, that in other studies the oxygen uptake rate
(OUR) has been used equally successfully. (In the penicillin process after the first
20 h or so, the ratio of CER to OUR, known as the respiratory quotient (RQ),
remains constant at around 0.9.)

z
_o
I,,.--
Z
I..IJ
O
Z
O
O
69
<
_o Estimate
-- Measured

() I()0 2(?0 "3(?0 400

Time (hours)

Fig. 6. Neural network non-linear bioprocess model

22 M. Aynsley et al.

Process analysis reveals that a combination of three network inputs are

required - Carbon dioxide evolution rate (CER) measured in the air flow leaving
the reactor, substrate feed, and bioreaction age (i.e. time). Since there are in this
case two primary substrate feeds, utilized in a varying ratio, then it is useful to
separate the feed in its two components. The network topology was therefore
specified as being made up of four inputs, two hidden layers and one output. The
determination of the number of nodes in the hidden layers was based upon the
optimal method proposed by Wang et al. [45]. In this case four nodes per layer
was found to be appropriate. The network was then trained on two sets of
process data (not shown).
The quality of the neural network model estimator performance was as-
sessed by testing the network on four sets of "unseen" data. This is demonstrated
by Figs. 7, 8, 9 and 10. Here, the estimates of biomass have been obtained solely
with the network inputs. Neural network estimates are compared against
interpolated off-line biomass measurements. Although the biomass assay results
are corrupted with a high level of noise, it can be seen in Figs. 7 and 8 that the
neural network estimate is representative of the underlying process behaviour
under two quite different sets of operating conditions. This indicates that the
neural network model has been able to capture the non-linear process character-
istics spanned by the measured data. Figure 9 demonstrates the performance
when a contamination occurs towards the end of the batch, at around 120 h,
causing the process characteristics to change. The reaction was allowed to run
on in order to assess the performance of the neural network estimator over the
complete batch. Not surprisingly the network model is no longer applicable to
the system.
Figure 10 shows the estimator producing acceptable information until
around 60 h into the batch. At this point the estimates are observed to diverge

4-
4-
,4- 94- 4--I-
Z
o 4-§ -4- 4- 4-

4-
I,Ll
0
Z
0
0
g
4--4-
O
N
Estimated
Measured

2~ 4~ 6'0 I~o I~o i~o i~0 200

TIME (h)

Fig. 7. Neural network penicillin biomass estimation-Test Data Set 1

ArtificialIntelligenceand the Supervisionof Bioprocesses 23

z
o_
t~ .4-

uJ ,~. "4"
o
z
0
o
4-
,<
0
Estimated

+ + + + Measured

200

T~ME(~,l

Fig. 8. Neural network penicillin biomass estimation-Test Data Set 2

z
o

~
LI3
o
z
0
(,3

,<

+//~+ Estimated

+ + + + Measured

180

T~ME (h)

Fig. 9. Neural network penicillin biomass estimation-Test Data Set 3

from the laboratory assays. The operators, having established confidence in the
network estimator, were led to check the calibration of the on-line CO z
analyser. The analyser calibration had drifted resulting in a significant error in
the predicted biomass. This was corrected at around the 100 h point. Towards
the end of the reaction (140 h) the estimated biomass concentration again
diverges. In this case off-line laboratory assays confirmed a contamination. This
type of behaviour, considered alongside other recent work in process engineer-
ing, suggests that with appropriate reconciliation of data there is potential for
24 M. A y n s l e y et al.

q"~ ,4.

<

1.1/
z
0
o
g
<
0

Estimated

+ + + + Measured

2012
TIME (h)

Fig. 10. Neural network penicillinbiomassestimation-Test Data Set 4

use in fault detection. Bioprocess faults can of course be associated with reactor
and sensor hardware as well as the bioreaction itself.
Acceptable biomass concentration estimates have been achieved without the
need for corrective action from off-line assays. It is possible, however, to make
such updates. In a similar vein, it is also possible to make use of a first-principles
model of the penicillin process in concert with a neural network. In such a
scenario the network is used to take account of the process/first-principles-
model mismatch. Studies are on-going in these areas.

8 Discussion

The initial objective of the RTKBS study outlined in this paper has been the
development of a generally applicable real-time knowledge-based system for the
supervisory control of bioprocesses. The supervisory strategy also performs
sensor validation, fault detection and diagnostic tasks and provides advice to aid
productivity recovery following fault correction. In order to achieve this aim,
generic-type rules have been used as much as possible, particularly in the case of
hardware fault detection and diagnosis. Process specific knowledge can be
incorporated within the existing framework relatively easily owing to the
modularity and object orientated approach adopted in the development of this
system.
Further work is underway aimed at incorporating on-line modules to
perform process simulation, soft-sensing and control. Intelligent use of this
information should significantly improve the overall performance of the system
Artificial Intelligence and the Supervision of Bioproccsses 25

in providing for the continuous on-line optimisation of bioprocesses. Most

recently, studies into the use of knowledge-based methodologies in bioprocess
scheduling have been started.
Model-based bioprocess state estimators have tended to use relatively
simple process models. It is of course possible to develop quite complex first
principles bioprocess models. Such an approach is capable of delivering a
considerable improvement in the bioreaction data available for supervision and
control well beyond that provided by existing techniques. However, these
benefits must be weighed against the considerable effort which must be em-
ployed to model biochemical systems and the resulting high cost of doing so.
Care needs to be excercised, however, in that inappropriate assumptions
inevitably lead to problems in observer application.
A new and exciting approach, that of neural network modelling, provides a
generic non-linear modelling technique. Network models have been developed
for the on-line prediction of intersample values of low frequency biomass
measurements using higher frequency secondary measurements. The potential
of these networks to model dynamic non-linear bioprocesses, in order to provide
an on-line state estimator, has been demonstrated by application to two
industrial bioprocesses. Although there are a number of open issues such as
optimum network size, etc., the power of this new technique should be judged by
its ability to achieve acceptable state estimation and process control in a cost
effective manner.

Acknowledgements: The authors would like to acknowledge the support of

Gensym Corporation; Smith Kline Beecham, ICI Biological Products and
Marlow Foods, in terms of access to plant and process data and financial
support; and also the U.K. Science and Engineering Research Council for their
financial support.

9 References

1. Stephanopoulos G, Tsiveriotis C (1988) Towards a systematic method for the generalisation of

fermentation data. Preprints, 4th Int Cong on computer applications in fermentation techno-
logy, Cambridge, UK, September
2. Karim MN, Halme A (1988) Reconciliation of measurement data in fermentation using on-line
expert system. Preprints of 4th International Cong on computer applications in fermentation
technology, Cambridge, UK, September
3. Cooney CL, O'Conner GM, Sanchez-Rivera F (1988) An expert system for intelligent super-
visory control of fermentation processes. Preprints of 8th International Bioteeh Syrup, Paris,
July
4. Hofmeister M, Halasz L, Rippin DWT (1989) Comp Chem Eng 13:1255
5. Halme A (1987) Measurement and estimation in bioreactors. Proc Inteinational Symposium on
control of biotechnical processes, University of Newcastle, UK, October
6. Montague GA, Morris A J, Ward AC (1989) Fermentation monitoring and control:A perspect-
ive. Biotechnology and genetic engineering reviews, vol 7, December
7. Edelman J, Fewell A, Solomons GL (1983) Myco-protein - a new food. Nutrition abstracts and
reviews in clinical nutrition, Series A, vol 53, p 6471
26 M. Aynsley et al.

8. Chantler MJ (1988) Real-time aspects of expert systems in process control. In: Expert systems in
process control, IEE Colloqium, March 1988
9. Stephanopoulos G, Stephanopoulos G (1986) Artificial intelligence in the development and
design of biochemical processes. Trends in Biotechnology, p 241
10. Moore RL, Hawkinson LB, Levin M, Hofman AG, Matthews BL, David MH (1987) Expert
systems methodology for real-time process control. Proceedings of the 10th IFAC World
Congress, Munich, FRG, July
11. Qi C, Wang S-Q, Wang J-C (1988) Application of expert system to the operation and control of
industrial antibiotic fermentation process. Preprints of 4th International Congress on computer
applications in fermentation technology, Cambridge, UK, September
12. Moore R (1988) The G2 real-time expert system for process control. International symposium
on advanced process supervision and real-time knowledge-based control, University of New-
castle, UK, November
13. Aynsley M, Peel D, Morris AJ (1989) A real-time knowledge-based system for fermentation
control. Proc ACC, Pittsburgh, USA, June
14. Guilandoust MT, Morris AJ, Tham MT (1987) Adaptive inferential control. Proc IEE, vol 134,
PtD
15. Tham MT, Montague GA, Morris AJ, Lant PA (1991) Soft-sensors for process estimation and
inferential control. J Proc Cont, vol 13-14
16. D.T.I Report (1989) Batch process scheduling computer software. The R&D Clearing House
17. Bruno G, Elia A, Laface P (1986) IEEE Computer 32:32
18. Ku H-M, Rajagopalan D, Karimi I (1987) Scheduling in batch processes. Chem Eng Prog,
August, p 35
19. Barnetson P (1969) Critical path plann!ng. Butterworth, London
20. Bernstein G, Carlson EC, Felder RM (1989) Development of a simulation-based decision
support system for a multipurpose pharmaceutical plant. AIChE Annual Meeting, San
Francisco, CA, November
21. Carleysmith SW (1988) Data handling for fermentation development - An industrial approach.
Preprints, 4th Int Cong on computer applications in fermentation technology, Cambridge, UK,
September
22. Flynn DS (1982) Instrumentation for fermentation processes. Proc 1st IFAC Conf, modelling
and control of biotechnical processes, Helsinki, Finland
23. Codd EF (1970) Comm ACM 13:377
24. Anderson BDO, Moore JL (1979) Optimal filtering. Prentice Hall, New Jersey
25. Svrcek WY, Elliot RF, Zajic JE (1974) Biotech Bioeng 16:827
26. Stephanopoulos G, San KY (1981) State estimation for computer control of biochemical
reactors. Advances in biotechnology, vol 1, p 399
27. Stephanopoulos G, San KY (1984) Biotech Bioeng 26:1176
28. Halme A (ed) (1982) Proceedings 1st IFAC workshop on modelling and control of biotechnical
processes, Helsinki, August, Pergamon, Oxford
29. Dekkers RM (1982) State estimation of a fed-batch bakers yeast fermentation. Proc 1st IFAC
workshop, modelling and control of biotechnical processes. Helsinki, Finland, August
30. Swiniarski R, Lesniewski A, Dewshi MAM, Ng MH, Leigh JR (1982) Progress towards
estimation of biomass in a batch fermentation process. Proc 1st IFAC workshop, modelling and
control of biotechnical processes, Helsinki, Finland
31. Johnson A (ed) (1985) Proceedings 1st IFAC symposium on modelling and control of biotechni-
cal processes, Noordwijkerhout, December, Pergamon, Oxford
32. Ghoul M, Pons MN, Engasser JM, Border J (1985) Extended Kalman filtering technique for the
control of Candida utilis production. 1st IFAC symposium modelling and control of biotechni-
cal processes, Noordwijkerhout, p 165
33. Montague GA, Morris AJ, Wright AR, Aynsley M, Ward A (1985) Parameter adaptive control
of the fed-batch penicillin fermentation. 1st IFAC symposium modelling and control of
biotechnological processes, Noordwijkerhout, p 39
34. Leigh JR, Ng MH (1984) Estimation of biomass and secondary product in batch fermentation.
6th International Conference on analysis and optimisation of systems, Nice, France, p t9
35. Shioya S, Shimizu H, Ogata M, Takamatsu T (1985) Measurement of state variables and
controlling biochemical reaction processes. 1st IFAC symp modelling and control of biotechni-
cal processes, Noordwijkerhout, p 49
36. Chattaway T, Stephanopoulos G (1987) A new technique for estimation in bioreactors: adaptive
Artificial Intelligence and the Supervision of Bioprocesses 27

state observers. Proc international workshop on control of biotechnical processes. University of

Newcastle, UK, October
37. Dochain D (1986) PhD thesis. University of Louvain, Belgium
38. Bastin G, Dochain D (1988) Non-linear adaptive control algorithms for fermentation processes.
Proceedings of American Control Conference, Atlanta, USA, p 1124
39. Dochain D, de Buyl E, Bastin G (1988) Experimental validation of a methodology for on-line
estimation in bioreactors. 4th ICCAFT Conference- modelling and control of biotechnical
processes. Cambridge, UK, September
40. Di Massimo C, Saunders ACG, Morris AJ, Montague GA (1989) Non-linear estimation and
control of mycelial fermentations. Proceedings of the American Control Conference, Pittsburgh,
USA, p 1994
41. McAvoy T, Wang N, Bhat N (1989) Use of neural networks for interpreting biosensor data.
Proceedings of the American Control Conference, Pittsburgh, USA
42. Bhat N, Minderman P, McAvoy T, Wang N (1989) Modelling chemical process systems via
neural computation. International symposium - control for profit, University of Newcastle, UK,
November
43. Willis MJ, Di Massimo C, Montague GA, Tham MT, Morris AJ (1991) Artificial neural
networks in process engineering. Proc IEE, Pt D, 138, No 3, p 256-266
44. Cybenko G (1989) Approximations by superpositions of a signoidal function. Math. Cont.
Signal and System, 2, p. 303-314
45. Wang Z, Tham MT, Morris AJ (1992) Multilayer feedforward neural networks: Canonical
form approximation of nonlinearity, International Journal of Control
46. Bhat N, Minderman P, McAvoy TJ (1989) Use of neural nets for modelling of chemical process
systems, Preprints IFAC Symp Dycord + 89, Maastricht, The Netherlands, Aug. 21-23, p 147
47. Lant P, Willis MJ, Montague GA, Tham MT, Morris AJ (1990) A comparison of adaptive
estimation with neural network based techniques for bioprocess application. Proceedings of the
American Control Conference, San Diego, USA, p 2173
48. Thibanlt J, Van Breusegem V, Cheruy A (1991) Biotech Bioeng 36:1041
Host-Vector Interactions in Escherichia coli

James E. Bailey
Department of Chemical Engineering, California Institute of Technology,
Pasadena, CA 91125, USA

Dedicated to Prof. Karl Schiigerl on the occasion of his 65 birthday

Symbols and Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2 Interactions with DNA Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.1 Replication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 T r a n s c r i p t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Interactions with RNA Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1 Ribosome Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 RNA Degradation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Interactions with Protein Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.1 Regulatory Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Chaperones and Proteins of the Secretory Apparatus . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Protein Turnover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5 Interactions with Membrane Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1 Processing of Secreted Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.2 Influences on Membrane Energetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6 Interactions with Energy and Precursor Metabolism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
7 Effects on Specific Growth Rates . . . . . . . . . . . . . . . . . : ......................... 44
8 Inclusion Bodies and "Toxic" Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
9 Implications of Selection Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
10 Mathematical Models of Host-Plasmid Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
10.1 Small Structured Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
10.2 Large-scale Structured Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
10.3 Population Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
11 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
12 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

Introduction of a DNA vector into E. coli for the purposes of cloned gene expression can perturb
native cell functions at many levels. The presence of foreign D N A can alter regulation of
chromosomal D N A replication, regulation of transcription of chromosomal genes, ribosome
functions and RNA turnover, activities of regulatory proteins, chaperone and protease levels,
membrane energetics and protein post-translational processing, as well as energy and intermediary
metabolism of the cell. The combined effects of these interactions on the metabolic characteristics of
the host-vector system have major implications for yields of cloned biotechnological products and
interactions of genetically engineered organisms with their environment.

Advancesin BiochemicalEngineering
Biotechnology,Vol. 48
ManagingEditor: A. Fiechter
9 Springer-VerlagBerlinHeidelberg1993
30 J. E. Bailey

Symbols and Abbreviations

C time for chromosomal replication

D time between replication form termination and cell division
G~L number of ~-lactamase gene copies per cell
IPTG isopropylthiogalatoside
kdml3L specific degradation rate of [Mactamase message
[mRNA]~L intracellular concentration of 13-1actamase messenger RNA
~t specific growth rate
Tltx transcription efficiency of the [Mactamase gene
t time
Host-Vector Interactions in Escherichia coli 31

1 Introduction

The genome of an organism contains the genes and regulatory elements which
interact and function to establish the chemical, physical,.and functional charac-
teristics of that organism. "Wild-type" or "normal" attributes of the organism
are a consequence of these interactions and functions. In seeking to clone a
segment of DNA or to express a cloned gene, introduction of recombinant
vectors into the bacterium Escherichia coli is a common procedure. The
insertion of new DNA into the cell can perturb and in some instances severely
disrupt the normal balance of activities found in a wild-type cell. This review
considers experiments and mathematical modeling which have been conducted
to elucidate the types of perturbations which can arise.
Studies of the interlocking influences of vector and host cell upon each other
have major implications in biotechnology and basic science. For example, when
manufacturing a cloned therapeutic protein such as granulocyte colony stimu-
lating factor using genetically engineered E. coli, the reduction in specific growth
rate caused by high-level expression of the product must be taken into account
when designing the host-vector expression system and when operating the
process. From a basic science perspective, careful study of the metabolic and
regulatory characteristics of cells containing a recombinant DNA vector offers
insights into native cell function in the sense of a stimulus-response experiment.
Alterations in host cell metabolism which result from introduction of a cloning
vector are also of basic importance in the study of metabolic engineering. When
seeking to discern the influence on cell metabolism of expressing an active
cloned molecule in the cell, it is essential to recognize and to control for other
influences on cell functions which result from all other interactions with the
cloning vector.
This review will emphasize general points and will not endeavor to catalog
every experiment or study involving a different vector or host strain. Because all
data on interactions between hosts and vectors and their consequences are
obtained from in vivo studies, it is difficult if not impossible to dissect clearly the
most critical loci of the interactions. Many different types of perturbations and
changes in cell functions occur simultaneously. Nevertheless, in an effort to
emphasize the diverse ways by which introduction of new DNA can and does
alter cell function, the following comments will begin with the discussion of
perturbation of functions of host cell DNA, progress through perturbations in
RNA and protein functions and processing, and subsequently turn to other
special areas of metabolism before arriving at the most common subject of
vector effects on specific growth rates.
The key concept is the parasitic and perturbing influence of the added vector.
However, it is critically important that this parasitism be recognized as a
complex network of interconnections with many aspects of cell function and not
something as simple as "a demand for extra DNA synthesis" or "a demand for
more ATP". Considering alterations in the bacterial genome in a somewhat
32 J.E. Bailey

more general context, it is clear that there is a complete spectrum of responses

possible ranging from a silent mutation which changes the wild-type DNA
without influencing cell function whatsoever to the introduction of a lytic virus
which completely preempts native cellular activity resulting in virus synthesis
and cell death. Thus, a broad spectrum of responses must be considered, and the
cellular responses will vary based upon the particular characteristics of the
vector and the host involved. Except where otherwise noted, this review will
emphasize plasmid vectors, but many of the same comments and considerations
apply as well to phage vectors.

2 Interactions with D N A Functions

The most direct possible interaction of a cloning vector with the DNA of the
host and its functions occurs when introduction of the vector modifies the host
genome. This occurs for integrating vectors or if the autonomously replicating
vector contains a transposon or a region of homology with the host genome
such that some insertion is made into the host DNA. Such insertions can have
many effects, of course, ranging from silence to lethality by disrupting a function
essential for viability. Many other types of influences are possible by inactivating
important genes or regulatory elements. Such interactions are commonplace
with expression vectors for mammalian cells such that the resulting clones differ
widely in their growth characteristics, cell size, productivity, and other pheno-
typic characteristics, presumably due to different integration sites and their
subsequent consequences for host cell function. Such difficulties can be avoided
in E. coli, for example, by using vector and selection technology which localizes
the integration site such that a function unimportant for the growth conditions
of interest is targeted. Such a manipulation, for example, would involve integra-
ting a gene into the lacZ locus for a strain which would be grown on a glucose-
containing medium. Other types of interactions which can also exert profound
effects on the characteristics of the host-vector system are indirect in the sense
that the presence of the vector alters the native distribution of DNA-binding
proteins on the host DNA.

2.1 Replication
Bacterial plasmid vectors contain origins of replication consisting of regulatory
sequences which serve to control initiation of replication of the plasmid and also
the site of replication initiation. However, the vector typically does not encode
the necessary DNA polymerase and associated enzymes for DNA replication,
depending upon the host enzymes for these functions. Further, the plasmid may
depend upon other host proteins involved in regulation of chromosomal DNA
Host-Vector Interactions in Escherichia coil 33

replication with resultant perturbations on the replication of the host chromo-

some. No experiments have been designed to test the nature and extent of these
types of interactions exclusively. However, available data show significant effects
of plasmid presence on DNA and cell cycle regulation in E. coli, effects which are
likely due to the types of interactions just outlined.
In particular, Seo and Bailey determined the nuclear cell cycle parameters C
and D for a plasmid-free host strain and for a set of copy mutant vectors which
are extremely similar except for small domains which influence their copy
number (number of plasmids per cell) [1]. (Fig. 1 shows the structure of these
plasmids which have been employed in many investigations of host-vector
interactions which will be discussed in this review.) C is the time required after
initiation of chromosomal DNA replication for the corresponding replication
fork to reach the terminus of the chromosome. The time interval designated D is
the time which elapses from the time the replication fork reaches the terminus
until cell division occurs. By analyzing the distribution of single-cell DNA
contents in different recombinant strains of E. coli using flow cytometry and
analyzing those distributions with population model relationships, Seo and
Bailey showed that plasmid-carrying strains exhibited significantly smaller C
and D times than the corresponding plasmid-free host, E. coli HB101 (Table 1).
These results are interesting and somewhat unexpected in the sense that, because
of the general parasitic and disruptive effect of plasmid presence, one would tend

ORI 0RI
-_ pr

Barn HI digestion
and ligation

RNA imo' Apr

Fig. 1. Geneology of the plasmids used in many studies of host-plasmid interactions. Indicated in
parentheses is the copy number (plasmids/cell) of each plasmid in E. coli HB101 grown in Luria-
Bertani (LB) medium. Alternative nomenclature for these plasmids is based on these reference copy
numbers (e.g., P24 denotes pDM246, P60 denotes RSF 1050, etc.). Reprinted from [24]
34 J.E. Bailey

Table 1. Estimated C and D periods for recombinant E. coil HB101 strains. Estimates were obtained
by two independent methods. Flow cytometry analysis of the single-cellDNA content distribution
following chloramphenicol (CM) addition combined with theory enabled estimation of the number
of active replication forks in an exponentiallygrowing population and thereby,based on population
theory, the C and D times. Direct comparison results were obtained by least squares optimization of
C and D in the theoretical steady-state single-cell DNA distribution compared to corresponding
flow cytometry data. Growth conditions: LB medium at 37 ~ at pH 7.0. Reprinted from [1]
CM-treatment Direct comparison
C D C+D C D C+D
Strain (min) (rain) (min) (min) (rain) (min)
HB101 only 36.8 19.9 56.7 37.3 20.5 57.8
HB101 :pDM247 30.3 15.4 45.7 31.0 15.5 46,5
HB101:pDM246 30.9 16.7 47.6 31.8 15.0 46.8
HB101 :RSFI050 33.3 16.2 49.5 33.8 13.7 47.6
HB101:pDM248 32.1 15.0 47.1 33.5 14.6 48,1

to expect increases in these cell cycle times rather than decreases. Furthermore,
it is extremely interesting that the plasmid copy number exerts almost no effect
on the magnitude of the change of C and D; merely the presence of plasmids, in
this case all plasmids with a ColEl-type origin of replication, is sufficient to give
rise to these responses in the host-vector construct.

2.2 Transcription
E. coli RNA polymerase apoenzyme and holoenzyme bind nonspecifically to
DNA. Accordingly, introduction of additional D N A into the cell as a multi-
copy plasmid can be expected to cause significant parasitic partitioning of RNA
polymerase onto the plasmid DNA. Of course by simple mass action relation-
ships, this parasitic partitioning would not be significant for a small, low copy
plasmid but could be substantial for a large or high copy vector. A model of
cloned gene expression and its regulation which considers the influences of
nonspecific binding on vector D N A on availability of polymerase for gene
transcription clearly indicates substantial effects when plasmid D N A content of
the cell approaches the D N A content of the host cell genome (corresponding to
around 1000 copies of a vector of 4 kb [2]).
Moreover, plasmids carry several promoters. For example, in pBR322 there
are two promoters associated with the origin of replication (for the small
transcripts RNAI and R N A I I ) as well as the promoters associated with the
ampicillin resistance and tetracycline resistance genes. In a construct designed
for high level expression of a cloned protein, promoters of such great activity are
employed that they may, even when present at relatively low levels, compete
strongly for available RNA polymerase holoenzyme.
These disturbances in the distribution of RNA polymerase relative to that in
the wild-type cell are expected to influence the level of expression of various host
Host-Vector Interactions in Escherichia coli 35

Table 2. Determinations of relative levels of total RNA and ribosome components in E. coli HB101
carrying different plasmids from Fig. 1 (P0 denotes the plasmid-freestrain). Reprinted from [3]
RNA"/protein 70Sb 50S~ 30Sc
Strain (lag lag- 1) (AUFS per 103 dpm) (70S- 1) (70S- 1)
P0 0.26 - -
P24 0.25 3.6 0.23 0.16
P60 0.28 3.7 0.19 0.14
P120 0.23 2.3 0.14 0.09
a Calculated with Ab26 o of 1.0 to be equivalent to 40 p.g RNA per ml.
b Normalized with dpm of 3H-labelled cells and Ab42 o of P24 AUFS denotes absorption units full
scale.
c Height ratio from absorbance (254 nm) scan of ribosome fractions separated in a sucrose gradient

cell genes. Experiments to examine ribosomal RNA contents and relative levels
of ribosome components in cells carrying different members of the family of
plasmids shown in Fig. 1 have clearly shown the consequences of such perturba-
tions [3]. As-plasmid copy number increases, RNA and active 70S ribosome
levels first decline slowly, then drop rapidly (Table 2). These trends parallel
those for specific growth rate discussed below.
W o o d and Peretti also observed a decline in ribosomal RNA (rRNA) levels
for increasing copy number in experiments employing the plasmids in Fig. 1 in
E. coli HB101 [4]. Interestingly, their detailed study showed that this decline in
steady-state rRNA levels with increasing plasmid content was a consequence of
an increase in rRNA degradation which more than compensated for an increase
in rRNA synthesis rate with increasing copy number. A different trend was
observed in studies of the consequences of induction of the strong tac promoter
in a pUC-based expression plasmid [5]. In this case rRNA content increased as
the cloned promoter was induced because the increase in rRNA synthesis rate
with increasing induction dominated a smaller increase in rRNA degradation
rate.
If the pattern of individual proteins found in the cell is regarded as indicative,
at least in part, of the corresponding pattern of transcription, the alterations in
cellular content of m a n y individual host cell proteins which arise upon introduc-
tion of multicopy plasmids indicate significant perturbations in host cell gene
expression patterns which presumably result from shifts in distribution of RNA
polymerase. Based on computer image analysis of two-dimensional gel electro-
phoresis patterns of recombinant E. coli HB 101 extracts, levels of many host cell
proteins are altered by presence of multicopy plasmids (Table 3) [3]. Levels of
stress response proteins detected are generally higher, while many metabolic
enzymes exhibit reduced levels at the highest copy number studied. Lower
activities of glucose-6-phosphate dehydrogenase, fructose 1,6-diphosphate al-
dolase, and fructose 1,6-diphosphotase were observed in E. coli D H 5 ~ carrying
the higher copy members of the plasmid family in Fig. 1 [6].
The appearance of foreign or misfolded native proteins in the cell has been
suggested to be the primary signal for the stress or "heat shock" response of the
36 J.E. Bailey

Table 3. Relative amounts of individual proteins in plasmid-free (P0) E. coli HB101 and in HB101
carrying P60 and PI20 as determined by analytical two-dimensional protein gel electrophoresis.
MW is molecular weight (kDa) and pl is isoelectric pH. Results for P0 indicate the fraction of that
protein relative to all proteins detected. P0 numbers denote 3sS spot intensity on the basis of a total
intensity of 106 for all spots detected on the P0 gel. The values for P60 and P120 strains were
normalized by the corresponding values for P0 to indicate changes in protein levels relative to the
plasmid-free host. A value of 1.0 thus indicates no change of protein level in comparison to the
plasmid-free strain P0. Reprinted from [3]

Strain
Mr
Name (kDa) pI P0 a P60 b P120 b

Carbon metabolism
Oxoglutarate dehydrogenase
complex
Oxoglutarate dehydrogenase 110.0 6.36 1314 2.68 0.92
Dihydrolipoate succinyl-
transferase 48.0 5.74 2239 2.55 1.14
Dihydrolipoate dehydrogenase 49.3 6.25 2716 1.72 0.81
PEP Carboxylase 92.9 5.88 215l 0.99 1.64
Pyruvate dehydrogenase complex
Pyruvate dehydrogenase 99.6 5.70 4652 1.13 0.60
Dihydrolipoate transacetylase 74.7 5.12 10,593 0.90 0.75
Dihydrolipoate dehydrogenase 49.3 6.25 2716 1.72 0.81
Pyruvate kinase I 53.8 6.23 8631 0.97 0.65
Succinate dehydrogenase 71.3 6.17 199 3.73 2.20
Succinate thiokinase (~) 41.8 5.66 1270 2.26 1.09
Amino acid metabolism
Aspartate aminotransferase 41.7 5.93 945 0.75 1.35
Glutamine synthetase,
Adenylated 50.7 5.52 2721 0.93 1.51
Unadenylated 50.8 5.57 1167 1.23 0.34
Nucleotide metabolism
Aspartate transcarbamoylase 33.9 7.00 1556 2.03 0.57
Carbamoyl-phosphate
synthetase (~) 136.2 5.57 1754 1.20 1.77
Protein synthesis
translation
Aspartyl-tRNA synthetase 66.1 5.75 1278 0.96 0.25
Elongation factor
G 87.5 5.60 11,475 1.35 0.67
Tu 43.2 5.71 103,503 0.71 0.63
Ribosomal subunit proteins
S1 69.1 5.00 8608 - 0.58
$5 17.2 - 21,326 1.15 0.45
S6a 15.0 5.70 2195 2.02 0.08
S6b 15.1 5.63 10,567 1.07 0.31
$6c 15.2 5.53 7674 0.60
L1 34.4 - 6358 1.14 0.45
L3 28.7 - 10,163 1.12 0.26
L7/12 12.2 4.91 28,168 0.92 0.84
Lll 14.7 - 26,411 0.85 0.28
L25 11.0 - 12,972 0.81 0.69
Heat shock proteins
Dna K 71.3 4.96 7506 1.14 1.12
Htp,
G 69.3 5.24 563 1.52 2.05
Host-Vector Interactions in Escherichia coli 37

Table 3. (continued)

Strain
Mr
Name (kDa) pI P0 a P60 b P120b

H 35.9 5.65 355 0.59 1.33

Grp E 25.1 4.96 1889 1.02 2.27
Gro EL 58.0 5.01 13,770 1.16 1.63
Other
13-Lactamasec 29.6 5.77 - 2.19 8.52

" This is ppm as dpm (specific protein) per dpm total valid protein spots.
b This is ppm per ppm P0.
c This is normalized with respect to P24 which has 2889 ppm

Fig. 2A,B. Two dimensional gel electrophoresis of E. coli BGF1 before (A) and after (B) the
induction of the CSH 11 mutant 13-galactosidase.Cells were grown at 37 ~ and pulse-labeled before
(A) and 30 rain after (B) IPTG addition. The following stress proteins were identified according to
known databases (Phillips et al., 1987): (1) Protease La (H94.0, H94.1); (2) HtpM (F84.1); (3) DnaK
(B66.0); (4) HtpG (C62.5); (5) GroEL (B56.5);(6) GrpE (B25.3);and (7) GroES (C15.4). The position of
the abnormal [3-galactosidase is indicated with an arrow. The right side of the gels corresponds to
lower pH values. Samples loaded each had 100000 epm total. Reprinted from [9]

cell [7]. T h u s , the a p p e a r a n c e of a h e t e r o l o g o u s or m u t a n t h o m o l o g o u s p r o t e i n

r e s u l t i n g f r o m c l o n i n g a n d e x p r e s s i o n of its gene is expected to activate stress
r e s p o n s e gene expression. I n d e e d , this has b e e n o b s e r v e d in e x p e r i m e n t s with
several different r e c o m b i n a n t s t r a i n s of E. coli, i n c l u d i n g the e x p e r i m e n t s
s u m m a r i z e d in T a b l e 3. A n o t h e r s t u d y has revealed the sensitivity of the stress
r e s p o n s e to small q u a n t i t i e s of m u t a n t p r o t e i n . M. K o s i n s k i c o n s t r u c t e d E. coli
B G F 1 b y i n s e r t i n g i n t o the c h r o m o s o m e of a w i l d - t y p e s t r a i n ( M G 1 6 5 5 ) a single
38 J.E. Bailey

copy of a lacZ gene containing the CSH 11 mutation [8]. Addition of IPTG to a
culture of BGF1 resulted in elevated expression of several stress response
proteins as determined by analytical two-dimensional gel electrophoresis
(Fig. 2) [9]. As a warning on the difficulties of designing experiments which
clearly delineate only host-plasmid interactions, the recent discovery that
addition of IPTG to wild type E. coli alters metabolism and protein pattern
should be noted [10].
One possible strategy for reducing alterations in host cell gene transcription
resulting from parasitic partitioning of host cell RNA polymerase is introduc-
tion of a separate polymerase-promoter pair for cloned gene expression. Indeed,
this strategy is embodied to some extent in the expression strategy which
employs phage T7 polymerase and the corresponding phage T7 promoter for
cloned gene expression [11]. However, this promoter is so strong that cell
specific growth rate is reduced following induction, presumably as a direct and
indirect consequence of the high level of message produced.

3 Interactions with RNA Functions

The presence of plasmid vectors implies the presence of a corresponding set of

transcripts, some of which may interact with RNA binding functions of the host.
Also, presence of additional transcripts may alter the wild-type pattern or level
of activity of RNA degradation. Data on the specifics of such interactions are
limited but suggestive of the importance of parasitic plasmid interactions with
RNA functions.

3.1 Ribosome Partitionin9

In the detailed single-cell model of plasmid-carrying E. coli presented by Peretti
and Bailey [12], competition for active ribosomes between host cell messages
and plasmid-derived messages was considered as was the impact of such
competition on the total level of ribosome activity in the cell. Simulations, which
will be discussed in greater detail later, indicated potential significance of this
interaction, motivating an extensive subsequent study of a special recombinant
expression configuration employing specialized ribosomes.
Introduced by DeBoer and coworkers [13], introduction of a cloned gene for
a mutant 16S ribosomal RNA and corresponding modification of the Shine-
Dalgarno(SD) sequence on a messenger RNA enables construction of a sub-
population of specialized ribosomes which are competent only to translate the
correspondingly modified messages, Although the synthesis and assembly of
these specialized ribosomes is certainly a nontrivial perturbation of host cell
metabolism, the establishment of a separate class of ribosomes for translating
Host-Vector Interactions in Escherichia coli 39

plasmid message in principle could reduce the deleterious effects of high levels of
plasmid message and of strong ribosome binding sites on those messages on
translation of host-cell message. An extensive investigation of this strategy and
its consequences has recently been presented by W o o d and Peretti [14]. They
observed a 30% reduction in specific growth rate of the culture following
expression of the specialized 16S RNA; however, no further decline in growth
rate was observed following induction of transcription of cloned [3-galactosidase
message bearing the modified SD sequence.

3.2 R N A Degradation

M e a s u r e m e n t s of the rate of [Mactamase message synthesis and degradation

have been conducted for E. coli H B 101 carrying several of the plasmids of Fig. 1.
Estimation of the message-specific m R N A degradation rate was based upon the
following material balance for 13-1actamase m R N A concentration per unit
volume of cells averaged over the entire culture:
d[mRNA]~L
dt - T]txGl3L -- [kdml3L + lJ'] (1)

rhxG~L is the rate of message synthesis and [ m R N A ] denotes the intracellular

concentration of R N A in mols per unit volume of cells, kdml3L is the message-
specific m R N A degradation rate constant and p. is the specific growth rate of the
culture. F o r either a chemostat culture in steady state or a batch culture in
exponential balanced growth, the message concentration is expected to achieve
a steady state obtained by equating the left hand side of Eq. (1) to zero and
allowing an explicit calculation of the degradation rate constant in terms of
other parameters.
After measuring the relative steady state I3-1actamase message levels using
N o r t h e r n hybridization and the relative overall rates of 13-1actamase message

Table 4. Kinetic data for synthesis and degradation of !3-1actamasemessenger RNA in E. coil HB101
carrying the different plasmids shown in Fig. 1. The first column of data shows relative synthesis
rates. Division of these by the corresponding plasmid copy numbers gives the relative transcription
efficiencies rhx indicated. The relative [3-1actamasemessage specific degradation rate constants kdm!3L
were estimated from Eq. (1) making use of other measurements of relative steady-state ~3-1actamase
mRNA levels. Reprinted from [15]
Strain Counts per min a (10 -6) Yltx (10-8 cpm per gene kdm[3L (h 1)
per rain)
PI2 0.17 0.89 21
P60 0.72 0.75 2.8
P120 6.6 3.4 79
P408 690 110 330
aAmount of labeled [Mactamase mRNA on a per cell basis. Total RNA was pulse-labeled with [aH]
uracil, hybridized to nitrocellulose bound ~-lactamase DNA and digested with both RNase A and
RNase T 1 prior to scintillation counting
40 J.E. Bailey

synthesis using pulse-chase methods, estimates for the relative [3-1actamase

message degradation rate constants for E. coli HB101 carrying four of the
plasmids in Fig. 1 were obtained (Table 4) [15]. There is a major increase in
message degradation rate at the highest copy numbers. Similar trends of greater
degradation rates for rRNA with increasing copy number [14] and with
increasing cloned promoter induction [5] have been observed. Although the
basis for the significantly increased rate of RNA turnover in these recombinant
cultures is unknown, the effect is striking and must be considered as one of the
major consequences of plasmid presence.

4 Interactions with Protein Functions

Several influences of plasmid presence on the functions of proteins involved in

regulating and conducting DNA, RNA, and protein synthesis have already been
indicated. Further interactions of plasmids and their products with host-cell
protein functions are discussed next.

4.1 Regulatory Proteins

As noted earlier, plasmid replication often depends upon particular proteins
which are also implicated in controlling host-cell chromosome replication
initiation. In addition, expression plasmids are generally designed to afford
controlled and regulated expression of the product gene. Often, the proteins
required for regulation of these cloned promoter-operator systems are provided
by the host. Introduction of many copies of the cloned promoter-operator may
exhaust the available supply of regulatory protein, resulting in loss of control of
expression of the cloned product gene and also the corresponding gene (or
genes) in the host genome. Such loss of control is well known and well
documented for multicopy plasmids carrying the lac and tac promoter-
operators. However, similar effects can also be expected for other promoters
which depend upon host-cell proteins for their regulation.

4.2 Chaperones and Proteins of the Secretory Apparatus

Host cell chaperones are implicated in proper folding and assembly of numerous
host cell proteins, and also in preserving export-competent states of host-cell
polypeptides prior to membrane translocation [16, 17]. Expression of cloned
proteins may result in their interaction with host-cell chaperones, thereby
modifying the distribution of chaperones on host proteins and altering their
folding, processing, and export. As noted earlier, genetic introduction of either
Host-VectorInteractions in Escherichiacoli 41

multicopy plasmids or a small quantity of mutant protein into the cell activates
expression of stress response proteins including several chaperones. Besides
perturbing host cell protein metabolism, these interactions may also influence
the folding and targeting of the cloned protein.

4.3 Protein Turnover

Expression of a cloned protein may influence the level and distribution of

protease activities in the cell. Clear demonstration of this is provided in recent
experiments by Kosinski and coworkers which show major responses of cellular
proteolytic activity to expression of a very small quantity of a mutant E. coli
I~-galactosidase from a single chromosomally integrated gene [9]. In particular,
the activity of ATP-dependent proteases was found to increase upon induction
of mutant [3-galactosidase expression while the level of ATP-independent
protease activity was relatively unaffected. This finding is significant in view of
the implication of the ATP-dependent protease La in the initial, rate-limiting
step of intracellular proteolysis [18].

5 Interactions with Membrane Functions

5.1 Processin9 o f Secreted Proteins

Cloned proteins may be targeted for export through the cytoplasmic membrane
by gene fusion of a signal sequence with the product gene. Although this strategy
has not proven generally effective (only those proteins secreted in their native
hosts appear to be reasonable candidates for secretion in a heterologous host),
many such constructs have been employed effectively to obtain export of a
cloned protein into the periplasm of E. coll. In some cases, use of such fusion
constructs with strong promoters has resulted in lethality following induction.
This so called "overproduction lethality" has been attributed to interference of
the cloned preprotein with processing of host preproteins. That is, by mass
action, the large quantity of cloned preprotein occupies much of the limited
capacity of the host cell secretory apparatus, thereby impeding normal rates of
host cell preprotein processing and transport. Thus, there is again a parasitic
interaction and competition between the newly expressed cloned preprotein and
the normal processing of host cell secreted proteins. Another consequence of this
type of interaction may be reduction in integrity of the outer membrane ofE. coli
since its proteins are not processed sufficiently rapidly (relative to the cell
division rate) at the cytoplasmic membrane. This can result in leakage of
periplasmic proteins into the medium [see, for example, 19]. The latter phenom-
enon can be exploited to advantage in practice to produce cloned products
42 J.E. Bailey

extracellularly using an E. coli expression system. Whether overproduction

lethality and preprotein processing limitations might be circumvented by
overexpressing cloned proteins of the host secretory apparatus provides intri-
guing challenges for future research.

5.2 Influences on Membrane Energetics

Axe and Bailey have reported an in vivo nuclear magnetic resonance phospho-
rous-31 spectroscopy study of host and p!asmid-carrying E. coli [20]. Observa-
tions of intracellular and extracellular pH transients following addition of
glucose suggest a difference in membrane energetics between the transformed
and unmodified cells. Whether these changes result from differences in proton
pumping activities or differences in susceptibility to passive leakage of proteins
is not known. Also, it should be noted that the plasmid involved in these
experiments encoded pre-[3-1actamase, a protein new to the cell which must be
processed at the cytoplasmic membrane. Whether the observed effects on
transmembrane pH difference is connected with the expression of this cloned
preprotein is not established but worth considering.

6 Interactions with Energy and Precursor Metabolism

Alterations in the metabolic characteristics of host-cells transformed with

plasmids relative to the unmodified hosts are often ascribed somewhat casually
to additional demands for ATP or additional demands on precursors for DNA
synthesis. The alterations which plasmid presence causes in ATP and DNA
synthesis for use in biomass production can be well estimated by rigorous
methods. Stoichiometric analysis of bacterial growth pioneered by Stouthamer
and coworkers [21, 22] can be extended to analysis of host-plasmid systems.
Because of limited data available on the detailed composition of plasmid-
bearing strains, the stoichiometric analysis conducted by Da Silva and Bailey
assumed the host-cell biomass composition remained unaltered by plasmid
presence. Plasmids, plasmid-derived transcripts, and proteins encoded on the
plasmid were assumed additive to the host-cell composition. Then, following the
approach of Stouthamer, calculations were undertaken for a variety of scenarios
to estimate the effect of the presence of plasmids and the expression of plasmid
genes at different levels on the amount of ATP required for growth [23]. Some of
the results of these calculations are summarized in Table 5.
As might have been expected given the ATP demands for protein synthesis in
the wild-type host t21], it is the extent of plasmid-directed protein synthesis
which determines whether plasmid presence significantly increases the ATP
demand on the cell. From this stoichiometric viewpoint, even propagation of
Host-Vector Interactions in Escherichia coli 43

Table5. Theoretical estimates of the maximum cell yield from ATP (Y.... ATP) and of ATP
requirements for synthesis of different classes of cellular components. Six cases are considered for
recombinant cells. The parameters n and 6 denote the number of plasmids per cell and the fraction of
total protein synthesized which is encoded on the plasmid, respectively. The "product retained"
figures include plasmid-encoded protein as part of total cellular protein; the "product secreted"
figures do not. Thus the latter figures may also be interpreted as requirements and yields for active
biomass. Reprinted from [23]

Recombinant cell

Product retained Product secreted

ATPrequirement Normal 8 = 0.2 6 = 0.5 ~ = 0.5 ~ = 0.2 6 = 0.5 ~ = 0.5

for synthesis of cell n = 50 n = 50 n = 100 n = 50 n = 50 n = 100

Polysaccharide 20.52 18.12 13.45 13.44 20.49 20.49 20.46

Protein (f) 13.55 11.96 8.88 8.87 13.53 13.53 13.51
(p) 191.42 169.02 125.48 125.36 191.13 191.13 190.85
Lipid 1.40 1.24 0.92 0.92 1.40 1.40 1.40
RNA (f) 34.50 30.46 22.62 22.59 34.45 34.45 34.40
(p) 9.20 8.12 6.03 6,03 9.19 9.19 9.17
DNA (f) 8.64 7.63 5.66 5.66 8.63 8.63 8.61
(p) 1.92 1.70 1.26 1.26 1.92 1.92 1.91
Product (f) 0.00 2.99 8.88 8.87 3.38 13.53 13.51
(p) 0.00 42.26 125.48 125.36 47.78 191.13 190.85
Plasmid DNA (f) 0.00 0.35 0.26 0,53 0.40 0,40 0.80
(p) 0.00 0.08 0.06 0.12 0.09 0.09 0.18
mRNA Turnover 13.80 12.19 9.05 9.04 13.78 13.78 13.76
Sub-total 294.95 306.12 328.03 328.03 346.16 499.66 499.41
Transport of
NH,~ 42.42 44.60 48,84 48.85 50.43 74.40 74.37
K+ 1.92 1.69 1,26 1.26 1.92 1.92 1.91
P 7.75 6.88 5,11 5.13 7.78 7.78 7.82
Total ATP 347.04 359.30 383.24 383.27 406.29 583.76 583.51
requirement
Y~-~ (g cells per 28.8 27.8 26.1 26.1 24.6 17.1 17.1
tool ATP)

v e r y h i g h c o p y n u m b e r v e c t o r s d o e s n o t m a k e m u c h difference o n g l o b a l e n e r g y
m e t a b o l i s m . H o w e v e r , a c c o r d i n g to this s t o i c h i o m e t r i c analysis, p e r t u r b a t i o n s
in A T P r e q u i r e m e n t s o f the cell c a n b e c o m e i m p o r t a n t w h e n the e x p r e s s i o n level
o f the c l o n e d p r o t e i n s b e c o m e s significant r e l a t i v e to t h a t for all h o s t cell
p r o t e i n s . T h e fact t h a t s u b s t a n t i a l g r o w t h rate r e d u c t i o n s o c c u r w i t h h i g h Copy
p l a s m i d s a n d w i t h m o d e r a t e to l o w o v e r a l l e x p r e s s i o n o f c l o n e d p r o t e i n s [24] is
a clear i n d i c a t i o n t h a t the i n t e r a c t i o n s d i s c u s s e d in p r e v i o u s s e c t i o n s m a y
d o m i n a t e p l a s m i d effects o n c e l l u l a r A T P utilization,
T h i s t y p e o f a n a l y s i s of c o u r s e d o e s n o t i n c l u d e a n d c a n n o t e s t i m a t e the
effects of p l a s m i d p r e s e n c e o n m a i n t e n a n c e . As p o i n t e d o u t by S t o u t h a m e r a n d
others, m a i n t e n a n c e is a k i n e t i c r a t h e r t h a n a s t o i c h i o m e t r i c p r o c e s s and, in
E, coli, c o n s i s t s p r i m a r i l y of m a i n t a i n i n g the e n e r g e t i c c h a r g e of the c y t o p l a s m i c
m e m b r a n e [22]. U n f o r t u n a t e l y , there h a v e b e e n few definitive studies o n
44 J.E. Bailey

maintenance requirements for plasmid-bearing cells compared to the corres-

ponding plasmid-free hosts, probably because of the difficulty of long term
continuous culture of plasmid-bearing strains.
Some genetic engineering groups have advocated careful consideration of
"codon usage" when designing synthetic genes for high level expression in a
heterologous host. This means that, among the available codons, one should
choose those which are preferred in that particular host. Data on codon usage in
the host of interest are typically obtained by amino acid analysis of some of its
most highly expressed native proteins [25]. While the importance of codon
usage for achieving high level expression is a subject of ongoing debate,
suggestion of its importance implies another level of potential host-plasmid
interaction. In particular, if a cloned gene involves unusually high usage of a
particular codon which is relatively infrequent in the host, availability of the
charged tRNA corresponding to that codon may become limiting and may then
influence expression of some host-cell proteins. Clearly, one could identify many
other particular loci in the overall metabolic system of the cell at which similar
imbalances and perturbations could result as a consequence of plasmid-directed
activity.

7 Effects on Specific Growth Rate

Several if not all of the potential perturbations in regulation and activity of

cellular metabolic processes caused by plasmid presence occur simultaneously.
The most commonly studied and widely reported consequence of plasmid
presence is alteration in the specific growth rate of the cell relative to that of the
plasmid-free host.
Listed in Table 6 are the specific growth rates in two different media,
measured from exponential growth in shake flasks, of E. coli HB101 bearing the
different copy mutant plasmids illustrated in Fig. 1 [24]. There is a monotonic
decline in specific growth rate as the plasmid copy number increases. This
decline is gradual for low copy numbers and subsequently becomes more steep.
These studies and others have observed the expected consequence of introdu-
cing a vector which provides no useful products for the host. The resulting
extent of specific growth rate reduction follows the expected trends, increasing
with higher level expression of a cloned product [26, 27] and with higher copy
number vectors. However, it should be noted that an opposite influence has
been observed in the case of E. coli expressing the hemoglobin cloned from
ViweoscilIa [28]. When grown under oxygen-limited conditions, a strain engin-
eered to express Vitreoscilla hemoglobin can outgrow a similar strain which
does not express hemoglobin [29]. Of course this result can be viewed as a
particular instance of widely applied selection strategies designed to provide a
growth advantage for the plasmid-bearing strain in a particular growth environ-
Host-Vector Interactions in Escherichia coli 45

Table 6. Plasmid content effects on recombinant E. eoli HB101 maximum specific growth rates."
Reprinted from 1-24]

Reference Relative specific growth rates

copy
Plasmid number M9 LB M9C

pDM247 12 0.93 0.92 0.97

pDM246 24 0.88 0.91 0.94
RSF1050 60 0.88 0.87 0.88
pDM248 120 0.78 0.82 0.84
pFH118 408 0.68 0.77 0.82

" Values shown are the ratio of measured specific growth rate to specific growth rate of the plasmid-
fiee host in the same medium. E. coli HB101 specific growth rates: LB medium: 1.38 h - l ; M9C
medium: 0.79 h 1; M9 medium: 0.41 h - i

ment. However, in this case the "selective environment" is a condition of oxygen

limitation which is commonly encountered in bioprocessing.

8 Inclusion Bodies and "Toxic" Products

Many heterologous and some homologous proteins, when overexpressed,

aggregate into microscopic particles called inclusion bodies (IBs) or refractile
bodies. These are believed to be produced by interactions of partially folded
intermediates in the folding pathway which create an off-pathway diversion into
these aggregated structures [30]. It is common for aberrant elongated cell
morphologies to occur when extensive inclusion bodies synthesis is observed in
recombinant E. coli. For example, Fig. 3 shows a microphotograph of E. coli
carrying the plasmid pRED2 which expresses the Vitreoscilla hemoglobin
peptide driven by the native Vitreoscilla hemoglobin promoter [31] (also called
ORE and OxyPro TM in the literature). It is unknown how inclusion body
formation is coupled mechanistically to the formation of these elongated cells.
Presumably the presence of the inclusion body (or the processes of its synthesis)
interferes with transverse cell wall synthesis and other reactions associated with
the final stages of the cell division cycle. It is possible also that, by physical
interference with native intracellular events, inclusion bodies disturb and distort
normal cell cycle progression.
Experience in the biotechnology industry has shown that cell growth ceases
or is severely inhibited when certain heterologous proteins are expressed at very
low levels. Thus, for an effective process for producing such proteins, it is
essential that the protein be expressed from a regulated promoter and that
regulation in the inactivated state be extremely tight. Novel promoter regu-
latory configurations combining minimal preinduction activity and high trans-
cription rates post induction have recently been proposed based on molecular-
level mathematical models 1-32]. Why certain heterologous proteins in very
46 J.E. Bailey

Fig. 3A,B. Photomicrographs of E. coli JM101:pRED2 cells. A: phase contrast micrograph (bar is
1 pro). B: transmission electron micrograph (bar is 1 gm). In addition to refractile-type inclusion
bodies, cells also contain another inclusion body morphologydesignated floccule-type.Reprinted
from 1-31]

small amounts so greatly interfere with cell growth is also not known. Some
hypothetical causes include binding of the heterologous protein to crucial DNA
or RNA sequences, interaction of the heterologous DNA with chaperones or
proteases in the cell, or catalysis by the heterologous protein of reactions which
produce toxic components.
Host-Vector Interactions in Escherichia coli 47

9 Implications of Selection Pressure

In order to isolate clones containing vectors of interest or in order .to favor

growth in culture of the desired recombinant strains, it is common to design a
complementary genetic and medium formulation such that cells containing the
vector exhibit substantially more rapid specific growth rate than cells lacking
the vector. Common selection mechanisms used in this context are antibiotic
resistance encoded on a single gene on the vector coupled with addition of this
antibiotic to the medium. Another familiar motif is complementation in which
the host cell contains a mutation which inactivates a gene encoding an enzyme
essential for growth; provision of this gene on the plasmid then complements the
host cell defect and allows the vector-containing cells to grow.
It should be recognized, however, that the relief of the lethal situation
provided by the marker gene on the vector may not be complete. Thus, in a
culture growing in a high concentration of ampicillin in order to select for
recombinants containing the 13-1actamase gene marker, the presence of the
antibiotic may still interfere with cell envelope structure tO some degree.
Similarly, a cell which is complemented by a "good" copy of the mutated host
cell gene may still suffer of some level of "nutrient limitation" or even nutrient
excess because the synthesis of that enzyme may not be properly coordinated
with the entire cellular metabolic structure as is the wild-type chromosomal
copy.

lO Mathematical Models of Host-Plasmid Interactions

Mathematical representation of cellular activities which include some accoun-

ting for host-plasmid interactions have been formulated to address different
questions and to serve different purposes. As in any modeling of cellular
phenomenon, one can never include everything that is happening in the cell
which may be important. Instead, it is necessary to choose a particular subset of
cellular components and interactions to include in the model. These choices
must be based upon the intended use of the model, and, unfortunately, what is
included and what is neglected remains a subject of some degree of modeling art.
There is no rigorous, theoretical basis for choosing which variables and
interactions should be retained in the model in order to describe the essence
of the system and to avoid omission of some critically important component
or mechanism. The following discussion briefly considers the types of models
which have been formulated relating to host-plasmid interactions and appli-
cations of such models.
48 J.E. Bailey

10.1 Small Structured Models

In any description of cell kinetics, simple models which represent the cell by a
small number of pseudocomponents are convenient and often, by judicious
choice of a small number of parameters, can be adjusted to describe the growth
or product formation trajectories of interest. Although the components of such
small models are often assigned a loose physical interpretation such as total
nucleic acids or total protein or "total synthetic component", these labels are
more useful for conceptual guidance in proposing reasonable kinetic forms than
they are rigorously valid interpretations. A small structured model is basically
an empirical fit to the observed phenomena in which most if not all model
parameters have no clear physical relationship with the fundamental parameters
characterizing the reactions and equilibria actually occurring within the cell,
However, as a trade-off, the numbers of parameters involved and the com-
putational complexity of such models is minimized. Therefore, these types of
models are particularly useful for implementation in control schemes and for
representing experimental observations of the overall characteristics of the
cultures.
The four compartment model of Nielsen et al. includes variables for "active
biomass" interpreted to be primarily ribosomes, a lumped component consisting
of structural material and chromosomal DNA, plasmid DNA, and cloned
protein [33]. This model simulates many of the important qualitative features of
host-plasmid interactions including decreasing maximum specific growth rate
with increasing plasmid content and variations in RNA concentration and
plasmid content with culture specific growth rate. These trends are also well
simulated by an eight-compartment model formulated by Bentley and Kompala
[34].

10.2 Large-scale Structured Models

Large-scale computer models containing many variables (say of order 30) and
parameters (in excess of 100) serve a different purpose. These models represent
an attempt to calculate, in a systematic and coordinated fashion, the conse-
quences of many simultaneous interactions within the cell, some of which are
represented in great detail, perhaps even at the level of the known molecular
mechanism. These models have an important place for hypothesis testing in the
following sense: the model in itself is an embodiment of a certain set of variables
and corresponding interactions precisely and rigorously defined by the model
equations. The model thus defines a self-consistent "computer cell" which
possesses its own intrinsic characteristics based entirely on the model's structure
and parameter values. Therefore, the model calculations with the computer cell
show the extent to which such a structure is capable of manifesting the behavior
observed experimentally for biological cells. As such, agreement between model
and experiment is an indication that the structure implemented in the model is a
plausible one for interpretation of the experimentally observed phenomenon.
Host-Vector Interactions in Escherichia coli 49

However, it is extremely dangerous to extend this logic excessively and to

suggest that, because such a detailed model simulates even many aspects of
experimentally observed behavior in cells, the model itself is correct in all of its
details and that use of such models can be used to identify molecular mech-
anisms. This is not possible because of possible interactions in the model such
that errors in one segment of the model may compedsate for other errors in
other segments of the model.
A large-scale structured model of plasmid-bearing E. coli was constructed by
Peretti and Bailey [12, 35] based upon the plasmid-free single-cell E. coli model
developed by Shuler and several of his collaborators [36]. The Peretti model
was built to examine the implications of the types of parasitic interactions
shown schematically in Fig. 4. That is, the Peretti model explicitly calculates the
quantity of actively transcribing RNA polymerase as well as the quantity of
actively translating ribosomes in the cell and includes, in the case of plasmid-
bearing cells, the influence of plasmids on these quantities. Figure 5 depicts
schematically the components and processes considered in the Peretti-Bailey
model for RNA polymerase-promoter interactions and gives some indication of
the level of detail involved.
Table 7 lists simulated growth rates and cloned gene product levels of E. coli
carrying different copy number plasmids and comparable experimental data.
Clearly, the trends are captured correctly. This model also simulated reasonably
well the unexpected influence of plasmid presence on the C and D cell cycle
parameters observed experimentally [1, 12]. Because experiments showing the
phenomenon had not yet been accomplished, this model did not include a
strong effect of increasing plasmid copy number on decreased cloned gene
message stability which was subsequently discovered [15]. Possible major
influences of unknown interactions is always a potential pitfall of cell kinetics
modeling, but it must be clearly recognized when working with large models
which may, by virtue of their complexity, give a falsely assuring impression of
completeness and rigor.

Chromosomal REPLICATION ~_____ Plasmid

DNA ~ . ~ - DNA Polymerase DNA
= Deoxyribonucleotides

tRNA , ~
rRNA
mRNA c
TRANSCRIPTION
" RNA Polymerase
- Ribonucleotides
~ RNA p
mRNA p

Chromosomal ~ TRANSLATION
Plasmid
Protein " Ribosomes Protein
- AminoAcids

Fig. 4. A schematicrepresentation of the competition between chromosomal- and plasmid-directed

macromolecular systemsfor the cell's macromolecular synthetic machinery. Reprinted from [-121
50 J. E. Bailey

CORE POLYMERASE SIGMA RN& POLYMERASE HOLOENZYME

FACTOR

,o/ ,o/
PROMOTER PROMOTER

Fig. 5. RNA polymerase exists in many conformations. Binding of the sigma subunit to core
polymerase alters binding selectivityof enzyme.Resultant holoenzymehas much greater affinityfor
promoter sites and decreased affinity for nonpromoter regions relative to core polymerase. These
protein and DNA species and their relative affinitiesare explicitlycalculated in the Peretti-Bailey
single-cellmodel for E. coll. Reprinted from [35]

Table 7. Comparison of experimental and model simulation results for the specific growth rate and
the amount of cloned gene product per plasmid as a function of the relative number of plasmids per
cell. Reprinted from [12]
Cloned gene product, plasmid
Number of Relative specific growth rates specific synthesis
plasmids
per cell Experimental Calculated ExperimentaP Calculated
0 1 1
1 0.92 0.94 1 1
2 0.91 0.90 2.1 1.9
5 0.87 0.89 4.2 3.6
10 0182 0.87 5.4 4.1
20 0.66 - 5.2
34 0.77 - 7.0 -
Experimental values are taken from Ref. [24] for E. coli HB101 with the plasmid in Fig. 1 grown at
37 ~ in EB enriched with leucine, proline and thiamine and supplemented with 0.2% glucose

10.3 Population Models

P r o d u c t i v i t y of a cloned protein can be greatly c o m p r o m i s e d by r e a r r a n g e m e n t s
of the vector or by "shedding" of the vector from the culture as it grows. Several
investigators have formulated m a t h e m a t i c a l models which seek to describe the
changes in the plasmid c o n t e n t of a growing p o p u l a t i o n in batch a n d c o n t i n u o u s
bioreactors.
Simplest of these models b u t yet extremely useful has been an a p p r o a c h
which lumps all cells c o n t a i n i n g plasmids and all cells lacking plasmids into two
separate populations. These types models have proven quite useful in character-
Host-Vector Interactionsin Escherichia coli 51

izing the fundamental parameters of plasmid segregation and the overall

qualitative characteristics of the dynamics of plasmid content in bacterial
cultures [37, 38]. Other models have appended to this basic structure other
possible subpopulations which contain rearranged plasmids [39].
A few population balance models have been formulated and applied to
describe plasmid propagation and segregational instability (e.g., [40]). These
calculate not only the concentration of plasmid-free cells in the culture but the
concentrations in the culture of cells with differing plasmid contents, in fact often
over a contimlous distribution of plasmid numbers. Such models provide a clear
description of the relationship between the intrinsic instability of the plasmid
and the corresponding distribution of plasmid contents (including of course the
average plasmid content) in the recombinant culture.

11 Concluding Remarks

Although there has been an attempt in the previous presentation to decompose

the different elements of host-plasmid interactions, it must be recognized that
these several interactions arise simultaneously, greatly complicating any clear
interpretations of the mechanisms or dominant phenomena which determine the
physiological characteristics of a plasmid-bearing bacterium. The intrinsic
complexity of the situation is illustrated by a body of phenomenology which is
known to those who have worked with many different combinations of E. coli
strains and plasmids. For reasons which are not at all apparent, some plasmids
are more stable in some strains than others. Also, some proteins are expressed
much better in hosts with a particular genetic background than in other hosts.
Although of course in some cases hosts have been especially engineered to
enhance some important property of DNA stability or protein processing, in
many other cases the connection between the details of the host cell genotype
and the different physiological responses following introduction of a plasmid are
difficult to discern and in some cases almost impossible to conceive. Undoubt-
edly further research on the physiology of plasmid-bearing cells and on their
internal composition and regulatory characteristics will clarify some of these
perplexing observations. However, the complexity of the influences of a new
piece of DNA on all levels of cellular function is sufficiently great that some trial
and error optimization of vector and host will likely remain for some time an
integral part of the process to identify an effective expression system.

Acknowledgements: The author's research in the field of host-plasmid inter-

actions has been supported by the National Science Foundation and by the
Advancement Industrial Concepts Division of The United States Department of
Energy.
52 J. E. Bailey: Host-Vector Interactions in Escherichia coli

References

1. Seo J-H, Bailey JE (1987) Biotechnol Bioeng 30:297

2. Lee SB, Bailey JE (1984) Biotechnol Bioeng 26:1383
3. Birnbaum S, Bailey JE (1991) Biotechnol Bioeng 37:736
4. Wood TK, Peretti SW (1990) Bioteehnol Bioeng 36:865
5. Wood T K , Peretti SW (1991) Biotechnol Bioeng 38:397
6. Mason CA, Bailey JE (1989) Appl Microbiol Biotechnol 32:54
7. Ananthan J, Goldberg AL, Voellmy R (1986) Science 232:522
8. Kosinski MJ, Bailey JE (1991) J Biotechnol 18:55
9. Kosinski M J, Rinas U, Bailey JE (1992) Appl Microbiol Biotechnol (in press)
10. Kosinski M J, Rinas U, Bailey JE (1992) Appl Microbiol Biotechnol (in press)
11. Reznikoff WS, McClure WR (1986) In: Reznikoff W, Gold L (eds) Maximizing gene expression.
Butterworths, Boston, Mass, p 1
12. Peretti SW, Bailey JE (1987) Biotechnol Bioeng 29:316
13. DeBoer HA, Hui A, Comstock LJ, Wong E, Vasser M (1983) DNA 2:231
14. Wood TK, Peretti SW (1990) Biotechnol Bioeng 36:865
15. Peretti SW, Bailey JE, Lee JJ (1989) Biotechnol Bioeng 34:902
16. Rothman JE (1989) Cell 59:591
17. Collier DN, Bankaitis VA, Weiss JB, Bassford Jr PJ (1988) Cell 53:273
18. Goldberg AL, Swamy KHS, Chung CH, Larimore FS (1981) Meth Enzym 80:680
19. Georgiou G, Shuler ML, Wilson DB (1988) Biotechnol Bioeng 32:741
20. Axe DD, Bailey JE (1987) Bioteehnol Lett 9:83
21. Stouthamer AH (1973) Antonie van Leeuwenhoek 39:545
22. Stouthamer AH (1973) Biochem Biophys Acta 301:53
23. Da Silva NA, Bailey JE (1986) Biotechnol Bioeng 28:741
24. Seo J-H, Bailey JE (1985) Biotechnol Bioeng 27:1668
25. Bennetzen A, Hall B (1982) J Biol Chem 257:3026
26. Betenbaugh MJ, Dhurjati P (1990) Ann N Y Acad Sci 589:111
27. Siegel R, Ryu DDY (1985) Biotechnol Bioeng 27:28
28. Khosla C, Bailey JE (1988) Nature 331:633
29. Khosla C, Curtis J, DeModena J, Rinas U, Bailey JE (1990) Bio/Technology 8:849
30. Mitraki A, King J (1989) Bio/Technology 7:690
31. Hart RA, Rinas U, Bailey JE (1990) J Biol Chem 265:12728
32. Chen W, Bailey JE, Lee SB (1991) Biotechnol Bioeng 38:679
33. Nielsen J, Pedersen AG, Strudsholm K, Villadsen J (1991) Biotechnol Bioeng 37:802
34. Bentley WE, Kompala D (1989) Biotechnol Bioeng 33:49
35. Peretti SW, Bailey JE (1986) Biotechnol Bioeng 28:1672
36. Domach MM, Leung SK, Hahn RE, Cocks GG, Shuler ML (1984) Biotechnol Bioeng 26:203
37. Imanaka T, Aiba S (1981) Ann N Y Acad Sci 389:1
38. Ollis DF (1982) Phil Trans Roy Soc London B297:617
39. Lee SB, Bailey JE (1985) Biotechnol Bioeng 27:1699
40. Seo J-H, Bailey JE (1985) Biotechnol Bioeng 27:156
Parameters Influencing the Productivity
of Recombinant E. coli Cultivations

K. Friehs I and K. F. Reardon 2

1 Technische Fakult/it, AG Fermentationstechnik, Universit/it Bielefeld,
D-4800 Bielefeld, Germany
2 Department of Agricultural and Chemical Engineering, Colorado State
University, Fort Collins, Colorado 80523, USA

Dedicated to Professor Dr. Karl Schiigerl on the occasion of his 65th birthday

1 P a r a m e t e r s Relating to D N A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.1 P l a s m i d C o p y N u m b e r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.1.1 Increased C o p y N u m b e r by P l a s m i d Design . . . . . . . . . . . . . . . . . . 55
1.1.2 Influence of C u l t i v a t i o n C o n d i t i o n s . . . . . . . . . . . . . . . . . . . . . . 56
1.2 P l a s m i d Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.2.1 Segregational Genetic Stability . . . . . . . . . . . . . . . . . . . . . . . . 57
1.2.2 Structural Genetic Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2 P a r a m e t e r s R e l a t i n g to P r o t e i n Synthesis . . . . . . . . . . . . . . . . . . . . . . . . 60
2.1 P r o m o t e r s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.1.1 P r o m o t e r Strength . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.1.2 I n d u c t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 T e r m i n a t o r s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3 m R N A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4 R i b o s o m a l Binding Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.5 Stop C o d o n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.6 C o d o n U s a g e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3 P a r a m e t e r s Relating to P r o t e i n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1 Proteolysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1.1 Protease Deficient Strains . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1.2 I n h i b i t i o n of Proteases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.1.3 F u s i o n Proteins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.1.4 Protein E x p o r t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.1.5 R a p i d P r o d u c t F o r m a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.2 Inclusion Bodies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3,2.l A d v a n t a g e s and D i s a d v a n t a g e s of Inclusion Body F o r m a t i o n . . . . . . . . . . 65
3.2.2 Factors Influencing F o r m a t i o n . . . . . . . . . . . . . . . . . . . . . . . . 66
4 P a r a m e t e r s Relating to D o w n s t r e a m Processing . . . . . . . . . . . . . . . . . . . . . 67
4.1 Cell H a r v e s t and Cell D i s r u p t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.1 Cell Harvest . . : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1.2 Cell D i s r u p t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Protein T r a n s p o r t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2.1 Signal Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2.2 Secretion Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Protein F o l d i n g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Protein S e p a r a t i o n and Purification . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.5 I n a c t i v a t i o n of Biological W a s t e . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Advances in Biochemical Engineering

Biotechnology, Vol. 48
Managing Editor: A. Fiechter
9 Springer-Verlag Berlin Heidelberg 1993
 54 K. Friehs and K. F. Reardon

In the past 10 to 15 years, many of the promises of microbial genetic engineering have been realized:
the use of recombinant Escherichia coli has moved from the laboratory to the production facility,
and the manufacture of therapeutic recombinant proteins such as human growth hormone and
interleukins is a rapidly growing industry.
Along with this progress, however, have come new problems to solve: bioreactor operators have
discovered that large-scale cultivations of plasmid-containing bacteria do not behave in exactly the
same way as those of plasmid-free cells, plasmid stability has been recognized as a major hurdle, and
the protein product might not be present in a soluble form but rather as intracellular granules that
resist solubilization. These and other difficulties represent a new generation of challenges for genetic
engineering.
However, genetic engineering can do more than solve these problems. Molecular biological
techniques also have the ability to create new opportunities: to produce new compounds, to use
cheaper substrates, to facilitate downstream processing, and to optimize production in new ways.
The productivity of a cultivation can generally be expressed as the product of the cell density and
the specific biological activity. Both of these parameters are influenced by a variety of factors. For
recombinant cultivations, though, the level of biological activity, a reflection of the plasmid copy
number and expression efficiency, is the more interesting and important consideration and will
therefore be given more attention in our review. In this contribution, our general goal is to discuss
the factors that influence the productivity of recombinant E. coli cultivations, covering
--parameters relating to DNA;
- - parameters relating to protein synthesis;
- parameters
- relating to proteins; and
- parameters
- relating to downstream processing.
The object is not to tell the reader how to choose the perfect plasmid, host, and cultivation
conditions, but to make known the many variables involved in designing a recombinant process and
to point out recent and potential advances made possible by genetic engineering. The discussion
focuses on the production of a protein, but many of the same concepts apply to other cultivations of
recombinant E. coli, including cases in which the desired product is not a protein or the ceils have
been designed for a special metabolic capability such as pollutant biodegradation.
Parameters Influencingthe Productivityof Recombinant Cultivations 55

1 Parameters Relating to DNA

1.I Plasmid Copy Number

It is often true that an increase in the number of copies of a gene (the "gene
dosage") will result in an increase in the production of that gene's product
protein (the "gene dosage effect"). While this is not always the case - regulatory
mechanisms or a saturation effect may impose an upper limit [1] - increasing
the copy number of a plasmid is a common method to enhance the productivity
of a cultivation.

1.1.1 Increased Copy Number by Plasmid Design

Since the copy number is closely related to plasmid replication, it has been
possible to design higher copy number plasmids by modifying replication
functions. Examples of this include:
- - c o p and tom mutations (in ColE1 derivatives)
- - Multiple tandem gene repeats
--Runaway replication plasmids
- - Dual low/high copy number origins of replication
The first two of these provide a high plasmid copy number throughout a
cultivation. The cop and rom mutations in ColEl-type plasmids influence the
regulation of plasmid replication [2, 3]. These types of mutations have been
used to produce ColE1 derivatives maintained at a level of 500 copies per cell
[4]. Similar opportunities for exploiting mutations in replication regulation are
available for plasmids such as R1, R6K, and pSC101 [5, 6, 7]. The use of tandem
gene copies in a plasmid has also been successful; for example, E. coli that
contained a plasmid with four copies of the chloramphenicol acetyltrans-
ferase gene produced four times as much product as cells harboring a plasmid
with a single gene copy [8]. However, tandem gene copies are prone to
homologous recombination.
Since continuous maintenance of plasmids at high copy numbers is a burden
to the host cell, resulting in significantly reduced growth rates as well as
increased plasmid instability, plasmid designs allowing one to increase the copy
number at an appropriate point in a cultivation are often desirable. One such
design is a temperature-sensitive "runaway" replication mutant, with which the
copy number of a plasmid is low at 25~ but rapidly increases when the
temperature is shifted to 37 ~ [9]. A different method involves the use of both
high and low copy number origins of replication on the same plasmid. Here, the
idea is to insert a controllable promoter in front of the replication primer of the
high copy number plasmid. An example of this is pMG411, which is maintained
at 4 copies per cell at 30~ and at 140 per cell when the culture temperature is
increased to 42 ~ [10].
56 K. Friehs and K. F. Reardon

1.1.2 Influence of Cultivation Conditions

Cultivation parameters also play a role in determining the copy number; thus, it
is also possible to change the copy number of a plasmid by manipulating the
cultivation conditions. The effects vary with the plasmid being used, but some
trends are evident.
Seo and Bailey [11] utilized different media to change the growth rate (and
perhaps other influential parameters) in batch cultures, and observed that the
copy number increased with decreasing growth rate. The influences of nutrient
limitation in continuous culture have also been studied. When a minimal
medium was used in a chemostat with glucose as the limiting substrate, Jones
et al. [12] noted significant decreases in copy number after one week of
operation.
When using runaway replication plasmids, the culture conditions have a
high impact on the achievable copy number. For example, the influence of
substrate feeding during runaway induction of pOU140 is shown in Fig. 1 [13].
Clearly, the addition of the carbon source, lactose, led to a significantly greater
number of plasmid copies than in the case in which no substrate was added.

100
Induction
O 80

~' 60
E
:~ 40
O
o 20

0
0 5 10 15 20 25
Time (h)

Fig. 1. Runaway replication with addition of lactose medium at point of induction. ( 4 1 - with
addition; --t-- without addition)

1.2 Plasmid Stability

One of the most important issues affecting the productivity of recombinant

E. coli cultivations is the maintenance of plasmids within the host cells. Plasmid
instability is of two types: segregational (plasmid loss resulting from defective
partitioning during cell division) and structural (undesired plasmid modifica-
tions resulting from insertion, deletion, or rearrangement of DNA).
Factors influencing both types of instability and strategies for overcoming
these problems will be discussed in the following sections.
Parameters Influencing the Productivity of Recombinant Cultivations 57

1.2.1 Segregational Genetic Stability

Although some wild type, low copy number plasmids exhibit high stability, most
plasmids of industrial interest (high copy number and expressing heterologous
proteins) are lost at frequencies of 10 -2 to 10-s per cell per generation [14]. A
recombinant host-vector system with an appreciable level of segregational
plasmid instability will be less productive than desired for two reasons:
- - t h e lowered average copy number will generally result in lower specific
productivity; and
- - plasmid-free cells which eventually emerge have a higher specific growth rate
(since they are no longer metabolically burdened by the plasmids) and will
thus reach a higher concentration than the plasmid-containing cells.
In a model developed by Imanaka and Aiba [15], cells were grouped as either
plasmid-containing or plasmid-free (i.e. copy number was not considered). Two
parameters were used to determine the fraction F of plasmid-containing cells
remaining after N generations: the frequency of plasmid loss (formation of
plasmid-free cells) and the ratio of the growth rates of plasmid-free to plasmid-
containing cells. The value of F was much more sensitive to the growth rate
ratio than to the loss frequency.
Strategies for overcoming segregational plasmid instability are based on the
alteration of one of these key parameters. Several types of approaches can be
identified. These have been classified as either selective (based on eliminating
plasmid-free cells) vs non-selective [16], or cellular/molecular vs bioprocess
methods [17]. In the following paragraphs, methods for preventing or overcom-
ing plasmid instability problems will be discussed according to the level at which
they are implemented, i.e.
- plasmid construction
-

--plasmid copy number

cultivation conditions
- -

- bioreactor configuration
-

Influences of plasmid construction. The composition of a plasmid influences its

stability and thus provides the basis for the most common approaches for
enhancing plasmid maintenance. Perhaps the best known of these is the
inclusion of an antibiotic-resistance gene on the plasmid combined with the
addition of that antibiotic to the medium. Although this method is simple to
implement, it has a number of disadvantages on an industrial scale, including
the cost of the antibiotic and the need to separate the antibiotic from the desired
product.
In another selective approach, a gene essential to the host cell (usually a
mutant) is included in the plasmid. Examples of this are the serB gene (for serine
production) in a serB- host [18], the valS gene (encoding valyl tRNA-synthetase)
in a valS ts host [19], and the ssb gene (for the SSB protein) in a ssb- host [20].
This method has been very successful; for instance, plasmid stability was
58 K. Friehs and K. F. Reardon

maintained for more than 200 generations in the valS system [19]. However,
problems can arise when the mutated host cell is unable to grow rapidly or when
spontaneous reverse mutations occur.
Selection for plasmid-containing cells can also be accomplished by killing
plasmid-free cells. In one approach, the parB locus of plasmid R1 is incorporated
into the plasmid. This sequence includes two genes: hok, which produces a
potent bacteriocin, and sok, which forms an mRNA product that prevents the
translation of hok mRNA. Since the sok mRNA has a much shorter lifetime than
hok mRNA, ceils that lose the plasmid after division are killed by the bacteriocin
[21]. An alternative method of killing plasmid-free cells is to include the
bacteriophage ~ repressor on the plasmid and to infect the host cells with the
phage. Any plasmid-free cells that appear are then lysed [22].
The use of adjustable copy number (runaway replication) plasmids is a
different type of method that allows both good plasmid stability and high
productivity. Plasmid stability is maintained during an initial low copy number
growth phase due to a favorable ratio of growth rates between plasmid-
containing and plasmid-free cells. Following the growth phase, the copy number
is increased (e.g. by a temperature shift) and a period of high rates of product
formation occurs [9]. This approach can be applied to continuous cultivation
by utilizing a two-stage chemostat [23].
Most of the preceding methods are similar in that plasmid-containing cells
are given a growth advantage over plasmid-free cells without inherently increas-
ing the segregational plasmid stability. Another problem with selective tech-
niques is that they do nothing to maintain plasmid copy number but rather
require only that one plasmid per cell be present.
Non-selective, or genetic, approaches are more promising for ensuring that
high copy numbers are stably maintained. An example of these methods is the
inclusion of strong terminator sequences on the plasmid to prevent stability
problems posed by the strong promoters that are frequently used [24].
The par locus of pSC101 is responsible for stable maintenance of that
plasmid. When par has been cloned into other plasmids, their loss frequencies
have decreased [25, 26]. Meacock and Cohen [26] determined that the par locus
effectively increases the stability of low copy number plasmids but had little
impact on high copy number vectors. It should also be noted that insertion of
the par sequence can reduce the copy number [27].
Another example of a genetic method for enhanced plasmid stability is the
incorporation of the cer locus of the plasmid ColE1. When plasmids containing
cer are grown in a xer + (genes for site-specific recombination) host, multimeriz-
ation of plasmids is reduced, leading to increased plasmid stability [14].

Influences of plasmid copy number. The number of plasmids per cell has an
influence on the segregational plasmid stability of the culture. Although low
copy number plasmids with an active partition mechanism (e.g. par) are usually
very stable, high copy numbers generally result in greater stability when random
Parameters Influencing the Productivity of Recombinant Cultivations 59

partitioning occurs at cell division [2]. However, this stability advantage of a

high copy number system can be offset by the decreased growth rate associated
with such cultures, so that the generation of even a few plasmid-free cells can
lead to domination by the faster-growing segregant.
Influences of cultivation conditions. The growth environment of a recombinant
cell can have a significant effect on thesegregational stability of the plasmids it
carries. Variations of plasmid stability with differences in growth rate, medium
composition, dissolved oxygen concentration, and temperature have been ob-
served; however, it has been difficult to find real trends in most cases.
Plasmid instability has generally been observed to increase with decreasing
growth rate, primarily because the relative growth rate advantage of plasmid-
free cells over those containing plasmids is decreased under these conditions
[23].
The influences of medium composition are less clear, due in part to
differences between host strains. Nutrient effects are greater in chemostat
cultivations, where at least one nutrient is limiting, than in batch cultures [28].
Although exceptions have been reported, continuous cultures limited by phos-
phate and magnesium generally exhibit high plasmid instability [12]. Instability
under glucose limitation is often lower [29, 30], and nitrogen limitation does not
appear to affect plasmid maintenance [29]. Plasmid instability can often be
lessened with the addition of complex nutrients like casamino acids [31].
Limiting levels of dissolved oxygen have a detrimental effect on plasmid
stability; both short-term (oxygen shock) and long-term oxygen limitations have
been shown to increase the rate of plasmid loss [32, 33]. Finally, plasmid
stability is generally found to decrease with increased temperature [34].
Bioreactor configuration. Several novel bioreactor designs have been proposed
to increase the plasmid stability of a culture. Such alternatives to traditional
batch or chemostat cultivations include two-stage chemostats (e.g. for use with
runaway replication vectors [23]) and cycling of growth conditions between
different dilution rates [35], substrate concentrations [36], and temperatures
[37]. A special cell-recycle reactor has been used to maintain high levels of
plasmid-containing cells by taking advantage of a ftocculation sequence on the
plasmid [38]. Cell immobilization has also been shown to increase the plasmid
stability of a culture [39].

1.2.2 Structural Genetic Stability

Structural plasmid instability can be difficult to detect in a cultivation, since the

growth rate and marker phenotype are the same as those of the desired cells.
Several studies have shown that structural instability is affected by the cell's
environment; for example, Godwin and Slater found different types of structural
changes in glucose- and phosphate-limited chemostats [40].
60 K. Friehs and K. F. Reardon

2 Parameters Relating to Protein Synthesis

2.1 Promoters

The synthesis of a protein starts with the promoter. The initiation of trans-
cription is a rate-limiting process for mRNA synthesis. Comparisons of more
than 100 promoters of E. coli have shown that there are two regions of
conserved DNA sequences [41, 42]. These regions, located 10 and 35 base pairs
upstream from the transcription initiation site, strongly influence the strength of
a promoter, which in turn determines the rate of transcription initiation. During
the development of the molecular biology of E. coli, many different promoters
were investigated; some of them are currently used in biotechnology and others
may be useful after further study.

2.1.1 Promoter Strength

Due to the importance of the promoter strength on the productivity of a

recombinant cultivation, genetic engineering is widely used to enhance the
strength. The sequence of well known promoters such as lacUV5 have been
changed and the effects on promoter strength examined [43]. New sequences are
often tested in order to find especially strong promoters like LP L and LP R from
the bacteriophage lambda.
While it is important to use strong promoters in the production of re-
combinant proteins, regulation of those promoters is essential since constitutive
overproduction of heterologous proteins leads to decreases in growth rate,
plasmid stability, and culture viability. Some promoters are regulated by the
interaction of a repressor protein with the operator (a region downstream from
the promoter). The most well known operators are those from the lac operon
and from bacteriophage ~. An overview of regulated promoters in E. coli is
presented in Table I.

2.1.2 Induction

A major difference between typical bacterial cultivations and those involving

recombinant E. coli is the technique of separating growth and production
phases. This method takes advantage of regulated promoters to achieve high
cell densities in the first phase (while the promoter is "off" and the metabolic
burden on the host cell is slight) and then high rates of heterologous protein
production in the second phase (following induction to turn the promoter "on").
For industrial bioprocesses, low-cost induction systems are desirable. This is
still a problem; for example, the widely used lac promoter system is induced with
IPTG, a relatively expensive compound. Another common induction technique
Parameters Influencing the Productivity of Recombinant Cultivations 61

Table L Regulated promoters in E. coli

Induction Promoter Operator Ref.

Temperature shift kve ~ 44

)~vR )~ 45
)~ tandem )~ 46
IPTG lac lac 47
tac lac 48
lac mutations lac 49
lpp trp tandem lac 50
mac lac 51
rac lac 52
rrnBP2 lac 53
Synthetic consensus lac 54
T7 gene 10 lac 55
IAA trp trp 56
lpp trp tandem trp 50
Arsenite ars ars 57
Dissolved oxygen trp trp 58
vgb vgb 59
CO2 limitation ? ? 60
Fe limitation ? ? 60
Mg limitation 9 9 60
NO32+ limitation 9 ? 61
PO4 a+ limitation ugp ugp 62
psi network 63
N limitation glnHP2 glnHP2 64
pH shift alx alx 65
)~PL ? 66
Osmotic pressure bet ? 67
otsA ? 67
otsB ? 67
proU proU 68
treA ? 67
Redox potential shift ? ? 69
Suecinate tna tna 70
Deoxyribose phosphate deo PIP2 ? 71
SOS response )~PL ? 72
Tryptophan limitation trp trp 73

used with p r o m o t e r s like ~,PL is a t e m p e r a t u r e shift, which requires large

a m o u n t s of energy a n d m a y also lead to the undesired f o r m a t i o n of stress
response proteins. I n d u c t i o n by a p H shift or by a d d i t i o n of a n inexpensive
i n d u c e r molecule would be of great biotechnological interest.

2.2 Terminators

The first step of protein biosynthesis ( m R N A formation) ends with the termin-
ation of D N A transcription, when the R N A polymerase reaches the t e r m i n a t o r
sequence. W i t h o u t correct t e r m i n a t i o n , the m R N A would form the w r o n g
proteins, leading to an overall decrease in productivity. The use of strong
62 K. Friehs and K. F. Reardon

promoters requires strong terminators to prevent transcription readthrough. A

variety of transcription terminators have been utilized, including the bacte-
riophage fd terminator [74], and some are commercially available (e.g. the trpA
terminator) [54].

2.3 mRNA

The first product of protein biosynthesis is mRNA, which is then translated to

form the protein of interest. In addition to structural factors of the mRNA that
influence its translation, the amount of translatable mRNA has a direct impact
on the overall productivity. The amount of mRNA in a cell is a function of the
rate of transcription (discussed above) and the rate of mRNA degradation. This
degradation rate depends on the presence of RNase recognition sequences,
especially those for 3'-exonucleases. Chan et al. have tried to increase the halflife
of mRNA by using non-essential regions of an intron [75]. This research led to
the development of the cloning vector pKTN-CAT, which increased the produc-
tion of chloramphenicol acetyltransferase three- to seven-fold by stabilizing the
mRNA molecules [76].
A second method to increase the lifetime of mRNA involves the product of
the ares (altered mRNA stability) gene. Strains that carry the temperature
sensitive ares-1 mutation have longer mRNA halflives at the non-permissive
temperature [77].
Another possibility for enhancing the halflife of mRNA is to use RNase-
deficient mutants [78]. However, such mutants are often difficult to cultivate.

2.4 Ribosomal Binding Sites

The translation of mRNA, the second step in protein biosynthesis, starts with
the binding of the ribosomes at specific binding sites (RBS) on the mRNA
molecule. Weak binding sites lead to a low level of expression. In many systems,
the expression of foreign genes utilizes the native RBS. An increase in the
expression level can be achieved by replacing this natural RBS with an altered,
more efficient sequence [793. For example, Olsen et al. were able to enhance the
expression of bovine growth hormone in E. coli by enriching the sequence
flanking the RBS with A and T nucleotides [80]. The distance between the RBS
and the AUG start codon also has an impact on the rate of initiation of
translation [81]. A more complete understanding of the influence of the
secondary structure of the RBS on the rate of translation should provide further
opportunities to optimize foreign gene expression {-82].
Parameters Influencingthe Productivityof RecombinantCultivations 63

2.5 Stop Codons

As in the case of transcription, translation also requires an efficient termination

sequence. These translation stop signals have been added to commonly used
vectors like pUC12 (forming pUC12-STOP) [83]. This vector was constructed
by inserting a DNA linker with TAA translational stop codons in all three
reading frames. Other terminators, such as the Universal Translation Termina-
tor [54], are commercially available.

2.6 Codon Usage

When using synthetic genes, the amino acid sequence of a protein is used to
develop a DNA sequence. Due to the nature of the genetic code, a choice of
codons is usually available. It has been shown that the use of particular codons
can play a role in gene expression. For example, highly expressed genes in
several species show a bias for certain synonymous codons [84]. Differences in
codon usage can also influence mRNA lifetimes.

3 Parameters Relating to Proteins

3.1 Proteolysis

Heterologous proteins produced by E. coli are usually subject to attack by a

variety of proteases. E. coli cells produce cytoplasmic, membrane-bound, and
periplasmic endoproteases. Relatively little is known about their induction,
substrate sequences, or kinetics.
Several methods have been devised to minimize proteolytic activity on the
cloned gene product. These include the use of low protease hosts, inhibition of
proteases, disguising the desired product by forming a fusion protein, excreting
or secreting the product to a "safer" location, and overwhelming the proteolytic
enzymes with a high rate of product formation. Modification of cultivation
conditions can also reduce protease levels; for example, some proteases are
synthesized in response to low dissolved oxygen or glucose conditions.

3.1.1 Protease Deficient Strains

The levels of protease activity vary among E. coli strains, and this may represent
one criterion for host strain selection. More directed efforts have focused on
mutants deficient in the production of one or more of the known proteases. The
best known examples are the lon- mutants, which cannot form the cytoplasmic
64 K. Friehs and K. F. Reardon

protease La [85]. Although cloned protein accumulation can be higher in Ion-

hosts [86], removal of one protease does not eliminate proteolytic activity.
Drawbacks to the use of protease mutant hosts include their altered physiology;
for example, lon- strains are UV sensitive and some exhibit a mucoid phenotype
that makes cultivation problematic.

3.1.2 Inhibition of Proteases

Although it is more difficult to inhibit intracellular proteases than those present

in the cultivation medium, at least one such effort has been successful. In this
case, the pin sequence (encoding a protease inhibitor produced during bacte-
riophage T4 infection) was cloned onto a plasmid, effectively reducing the
degradation of the desired product [87].

3.1.3 Fusion Proteins

It is often possible to protect a heterologous protein from proteolysis by

producing it as a hybrid product with a native protein such as [3-1actamase [88]
or special foreign proteins like ubiquitin [89]. The use of fused sequences can
also enhance translation initiation, and the hybrid products can be designed to
facilitate purification. However, the cleavage of the fused molecule to obtain the
desired product requires additional processing and may be difficult. Also,
expression of a fused protein may be lower than expected.

3.1.4 Protein Export

Another method to reduce proteolysis of the product protein is to engineer its

excretion from the cytoplasm into the periplasmic space or its secretion into the
medium. Transport into the periplasmic space can be accomplished by fusing
leader peptides for periplasmic or outer membrane proteins to the product
protein. Although the process of directed transport is not well understood and
selection of the best leader sequence is done by trial and error, this approach has
proven successful in many cases (e.g. [90]). Additional aspects of protein
secretion and excretion are presented in Sect. 4.2.

3.1.5 Rapid Product Formation

Finally, producing the cloned gene product at high rates so as to saturate the
proteolytic enzymes is an effective method of minimizing degradation. This can
be accomplished with runaway replication vectors or with plasmids containing
inducible promoters. Since this strategy has other benefits (e.g. increased
plasmid stability), it is frequently implemented.
Parameters Influencing the Productivityof RecombinantCultivations 65

Rapid production of heterologous proteins often results in the formation of

insoluble inclusion bodies. Although these protein aggregates offer further
protection from proteases, recovering active product molecules can be
problematic.

3.2 Inclusion Bodies

In E. coli (and other microorganisms), the product of a foreign gene frequently

appears in the form of proteinaceous aggregates called "inclusion bodies" or
"refractile bodies". The desired protein can make up from 40 to 95% of the total
protein content of these particles; contaminants include other proteins (espe-
cially outer membrane proteins), lipopolysaccharides, and membrane fragments
[91, 92]. Inclusion bodies can be found in the cytoplasm or in the periplasmic
space [93].
The implications of inclusion body formation are mixed for cultivations of
recombinant E. coli in which a protein is the desired product. (Of course, when
the foreign protein is not the desired product, formation of inclusion bodies is
detrimental.) Although total protein productivity may be enhanced, the process
economics may be adversely affected. These tradeoffs, as well as the factors
influencing inclusion body formation, are discussed in the following paragraphs.

3.2.1 Advantages and Disadvantages of Inclusion Body Formation

The production of inclusion bodies can enhance protein production in re-

combinant E. coli cultivations in two areas. First, protein accumulation often
increases, primarily because proteins in an inclusion body are resistant to
protease attack. The formation of inclusion bodies can also allow the produc-
tion of high levels of proteins that are toxic to the host cell when soluble.
Due to their physical characteristics, inclusion bodies also offer the oppor-
tunity to improve recovery of the protein product from a cultivation. A high
degree of purification can be achieved by cell lysis and simple low-speed
centrifugation. Since proteins in inclusion bodies are already denatured, addi-
tional purification steps can utilize conditions (e.g. detergents) and techniques
(e.g. gel filtration chromatography) that are less effective or inappropriate for
active proteins.
However, protein production in the form of inclusion bodies can also have
significant disadvantages for the bioprocess. One problem involves the release of
endotoxins when cells are disrupted to free the aggregates. This is a concern for
the purification of any intracellular protein and can only be avoided by secreting
the product into the medium.
The larger difficulty posed by inclusion bodies is the renaturation of the
protein to obtain an active product; in addition to the technical problems
inherent in this task, the process economics can be negatively affected by the
special renaturation steps. Following solubilization of the protein particles with
66 K. Friehs and K. F. Reardon

concentrated solutions of urea, organic solvents, and/or chaotropic salts (e.g.

guanidine-HC1), refolding of the totally denatured protein is required. At
present, determining the best process is a trial-and-error procedure. Typical
steps are careful removal of the solubilizing agent and the addition of com-
pounds (e.g. thiols) to control the formation of disulfide bonds. Substances like
polyethylene glycol have also been reported to improve folding [94, 95]. The in
vitro usage of chaperone proteins has been suggested and is an area of active
investigation [96, 97]. In general, however, renaturation is still problematic.
Although high yields of renaturated proteins have been reported [98], recovery
of activity has been low in some cases, especially for larger proteins. In addition,
the kinetics of the refolding process are often slow.
An interesting approach to improve the yield and kinetics of renaturation is
suggested by the work of Orsini et al. [95], in which removal of a portion of the
prourokinase sequence fortuitously resulted in the doubling of the refolding rate
and active product yield. If problematic regions (e.g. cysteine-rich) can be altered
without affecting the desired activity, similar process improvements might be
possible for other proteins.

3.2.2 Factors Influencing Formation

Although the mechanism of inclusion body formation is not well understood, a

number of factors influencing their production have been identified. These
include aspects of the protein molecule, the rate of protein synthesis, the
cultivation conditions, and the host strain used.
Researchers have sought to find a correlation between the formation of
inclusion bodies and various protein characteristics. There does not appear to
be a connection between the molecular mass, number of disulfide bonds, or
hydrophobicity of the protein product [99]. However, an interesting trend has
been observed for proteins consisting of subunits; when all necessary subunits
are produced concurrently in a cell, they are typically present in the soluble
form. On the other hand, inclusion bodies result when individual subunits are
produced separately [99].
The rate of protein expression appears to have a major role in determining
whether or not soluble protein is produced. High rates of expression generally
yield inclusion bodies [100]. Thus, the use of inducible promoters or runaway
replication vectors should be expected to lead to the production of inclusion
bodies.
Cultivation conditions also affect the form in which foreign proteins appear.
Temperature has a significant influence; higher temperatures result in increased
inclusion body formation [101]. This could be not only due to an increased rate
of synthesis, but also to the relationships between temperature, the protein
folding kineticsl and the growth rate of the cells. Similarly, media compositions
and pH values that reduce the growth rate have been observed to lead to lower
levels of inclusion body formation [93]. In some cases, special medium compo-
Parameters Influencingthe Productivityof Recombinant Cultivations 67

nents such as metal cofactors or nonmetabolizable sugars may exert direct

influences on protein folding [102, 103].
Finally, the choice of host E. coli strain has been shown to have an impact on
the formation of soluble vs particulate protein [-104]. As with other aspects of
recombinant cultivations that vary among strains, the basis for this influence is
not known.

4 Parameters Relating to Downstream Processing

4.1 Cell Harvest and Cell Disruption

If the product is not secreted into the medium or at least exported into the
periplasmic space, the cells must be harvested and disrupted to release the
product. This is the case for both inclusion bodies and soluble intracellular
proteins.

4.1.1 Cell Harvest

The harvesting of cells following a cultivation is basically a liquid-solid separa-

tion. Although this is a very common step in the biotechnology industry, it is by
no means a simple operation on a large scale. The most frequently used methods
are centrifugation and filtration [105]. For large-scale use, only continuous
centrifuges such as the tubular bowl or disc models are practical. Common
large-scale filtration methods include rotary vacuum drum filters, plate filter
presses, and tangential and microfiltration.
It may be possible to employ genetic engineering approaches to improve this
separation. One example is a mutation in the pil operon that results in the
overproduction of pili and subsequent cell flocculation [38]. This leads to an
increase in the sedimentation rate and should translate into more rapid centrifu-
gation as well. A similar method that has been developed for yeast involves the
cloning of a cell surface protein responsible for flocculation [106].

4.1.2 Cell Disruption

Both mechanical and non-mechanical methods are used to disrupt cells in large-
scale processing [105, 107]. Mechanical methods such as high pressure homo-
genization and bead mills often result in product losses through inefficient
disruption or thermal denaturation of proteins. Non-mechanical methods in-
clude those based on physical processes (e.g. sonication) and chemical effects
(e.g. organic solvents and enzymatic lysis). Each of these have disadvantages; for
68 K. Friehs and K. F. Reardon

example, large-scale use of organic solvents has an associated explosion risk

[108] and lytic enzymes are expensive [109].
Due to the importance of cell disruption for recovery of recombinant
proteins, several genetic approaches have been proposed recently to improve the
process with respect to efficiency, active protein yield, waste chemicals, and cost.
One such possibility is the creation of mutant host strains that have a more
permeable outer membrane. For example, the amount of porin in the outer
membrane of a strain of E. coli K12 was altered to increase the permeability
[110]. This led to an increase in the export of substances out of the cell.
However, mutations such as these frequently result in unhealthy cells that are
unable to grow well.
A more elegant technique is to increase the permeability or disrupt the cells at
an appropriate time in the cultivation (analogous to the concept of induced
transcription). An example of this involves the kil gene of ColE1, the product of
which leads to total cell lysis [111]. This gene has been placed under the control
of the lac promoter and incorporated into a plasmid. Thus, lysis can be induced
with IPTG after overexpression of the recombinant protein [112]. A similar
system has been developed using the lysis gene E from bacteriophage qbX174,
under control of the kPL promoter [113]. When this system is used together
with the temperature sensitive repressor ci857, cell lysis can be induced via a
temperature shift to 42 ~

4.2 Protein Transport

In many cases, it would be desirable to secrete the recombinant product into the
medium. Purification would be simpler than for an intracellular protein since
the product would not be contaminated with cytoplasmic components. (It
would be necessary to handle larger volumes, but this problem has been lessened
by newer chromatographic methods.) In addition, the formation of inclusion
bodies could be avoided and the toxic effects of some protein products on the
host cell could also be reduced. Protein secretion would also reduce proteolysis,
unless exoproteases are produced by the transport system.
Excretion into the periplasmic space of E. coli also provides many of these
benefits. Proteins can be freed from the periplasmic space with gentle treatments
that remove the cell wall and outer membrane.
A significant disadvantage facing protein secretion is that protein folding
may be incorrect or may not occur. As discussed in the following section, this is a
problem in the production of all recombinant proteins. The investigation of
special reactors for promoting folding is underway in several laboratories.
Two important factors influence the transport of proteins: the type of cellular
transport system and the nature of the signal (or leader) sequences that allow a
protein to use that transport system.
Parameters Influencingthe Productivityof RecombinantCultivations 69

4.2.1 Signal Sequences

Signal sequences are short peptides that allow the protein to be transported. In
bacteria, these signal sequences are typically 15 to 30 amino acids long [114,
1t5] and are located at the N-terminus of proteins, although some, like that of
hemolysin, are positioned at the C-terminus. Signal sequences form positively
charged heads that are able to pass through the membrane. Between the signal
peptide and the preprotein, a cleavage site for specific signal peptidases is
located to allow formation of the mature protein after transport.
In addition to the leader sequences, transported proteins often have other
sequences that are important for secretion. One type, located inside the protein
sequence, aids in the formation of a transport competent structure. The function
of other sequences, found in membrane proteins, is to stop transport. Thus, the
fusion of a gene to a signal sequence does not guarantee that the product will be
transported [116].
Some signal sequences that have been used to transport recombinant
proteins are shown in Table 2. Well known signal sequences like bla (~-
lactamase), malE (maltose binding protein), ompA (outer membrane protein),
and phoA (alkaline phosphatase) come from proteins that are exported into the
periplasm. Fusion proteins formed with one of these are therefore mainly
transported into the periplasm, but some secretion has also been observed. Of
special interest are the signal sequences spA (Staphylococcus protein A) and
malE, since both can be used as purification tags in affinity chromatography (see
Sect. 4.4).

4.2.2 Secretion Systems

Different organisms can be used as a secretion system. Bacterial hosts include

Bacillus subtilis, Staphylococcus aureus, Streptomyces lividans, and, increasingly,
E. coli. Important eukaryotic secretion systems are yeasts and cell cultures.
The transport of recombinant proteins by one of these hosts can be
accompanied by several difficulties [114, 115]. For example, the secretion
system may not be able to transport proteins that are too large or have an
unfavorable distribution of charges, the wrong hydrophobicity, or similar
structural burdens. Other potential problems include the saturation of export
sites, competition for the signal peptidases, or lack of the proteins that support
secretion.
Genetic engineering can be used to overcome some of the disadvantages of
the host. The secretion machinery can be optimized or the permeability of
membranes and cell walls can be changed. For E. coli, a series of different
proteins involved in export to the periplasm are known. Most of these belong to
the sec product family, and can be coexpressed with the recombinant protein to
produce an increase in secretion. Alternatively, the membrane can be changed to
70 K. Friehs and K. F. Reardon

Table 2. Signalsequences and target proteins for transport in E. coil; the final protein location is
shown as periplasmic space (P) or medium (M)

Signal sequence Target protein Location Ref.

amy amylase (B. stearothermophilus) M 116

bla proinsulin (human) ? 117
IgG (mouse) ? 118
[Mactamase M 119
epidermal growth factor (rat) P/M 120
triosephosphatase (chicken) M 121
malE gene 5 protein (phage M13) P 122
Klenow polymerase P 123
nuclease A (S. aureus) P 123
ompA colony stimulation factor (human) ? 124
superoxide dismutase (human) P 125
e~2interferon ? 126
antiviral proteins (Mirabilis) M 127
~-sarcin P 128
prokallikrein (human) P 129
nuclease A (S. aureus) P 125
ompF [~-endorphin M 130
phoA trypsin inhibitor (bovine) ? 131
epidermal growth factor (human) ? 132
fusion: [3-galactosidase-alk.phosphatase M/P 133
~-neo-endorphin P 134
fusion: maltose binding protein-13- P 135
galactosidase
ribonuclease T1 P 136
phoS growth hormone release factor (human) P 137
spA parathyroid hormon (human) M 138
insulin-like growth factor (human) M 139
Ovalbumin ovalbumin ? 140
Pullulanase [3-1actamase M 141
Preproinsulin ( r a t ) proinsulin (rat) ? 142
Enterotoxin LTA epidermal growth factor (human) ? 143
Synthetic ~2 interferon P 144
Metalloprotease and metalloprotease(with helper protein) M 145
helper protein
BRP as helper insulin-like growth factor (human) M 131
cloacin M 146
hly hemolysin M 147

Other abbreviations: amy = amylase, bla = [Mactamase, BRP = bacteriocine release protein, hly
= hemolysin,M = medium, malE = maltosebinding protein, ompA = outer membrane protein A,
ompF = outer membrane protein F, P = periplasmic space, phoA = alk. phosphatase, phoS
= phosphate binding protein, spA = Staphylococcus aureus protein A.

yield leaky m u t a n t s . U s i n g this approach, alkaline phosphatase [148], [3-

lactamase [149], a n d rat p r o i n s u l i n [150] were secreted into the m e d i u m from
E. coli. The m a j o r disadvantage to this m e t h o d is that leaky m u t a n t hosts are
very sensitive to their e n v i r o n m e n t a n d are difficult to grow. An interesting
o p t i o n is the use of bacteriocin release protein. W h e n expression of this protein
was induced, h u m a n growth h o r m o n e , which had a c c u m u l a t e d in the periplasm
of E. coli due to its fusion to a signal sequence, was secreted into the m e d i u m
[1311.
Parameters Influencing the Productivityof RecombinantCultivations 71

The export of proteins into the periplasmic space is of increasing interest.

The advantages of the periplasmic space, which accounts for 20 to 40% of the
total cell volume, include protection against cytoplasmic proteases (Sect. 3.1.4),
reduced frequency of inclusion body formation (Sect. 3.2), and the proper
environment for correct protein folding (Sect. 4.3). The use of the periplasmic
space in industrial scale processes depends on the development of methods to
open the cell wall and outer membrane without releasing cytoplasmic sub-
stances. At present, the outer membrane is often made more permeable by the
addition of chemicals or osmotic shock. Reports of secretion systems in which
the outer membrane leakiness is increased by plasmid-encoded products have
also been published 1-151]. Enhanced membrane permeability may also occur
fortuitously as a result of high-level expression of a foreign protein.

4.3 Protein Folding

The correct folding of a protein is essential for its biological activity. Although
issues of recombinant protein structure and overproduction receive a great
amount of attention, relatively little work has been done on protein folding,
despite its immense commercial impact and status as an important fundamental
question [ 152, 153]. The factors influencing protein folding are diverse and must
be considered when the production of active proteins is desired 1-154, 155, 156].
The production of recombinant proteins leads to at least two additional
problems with regard to protein folding: incorrect folding and the need to
renature proteins produced as inclusion bodies.
Some proteins, especially those from eukaryotic sources, are not folded
correctly in the oxidative milieu of the bacterial cytoplasm. One way to address
this problem is to export the protein into the periplasmic space of E. coli (via
fusion to a signal sequence). The periplasmic space is a reducing environment
and supports correct folding of eukaryotic recombinant proteins.
The need to refold proteins from inclusion bodies can be obviated by
preventing their formation. As discussed earlier (Sect. 3.2.2), this can be achieved
by using low-expression systems, by changing the cultivation conditions, or by
fusion to a signal peptide, among other methods.

4.4 Protein Separation and Purification

A major portion of the production costs in recombinant protein cultivations is

due to separation and purification needs. Thus, improvement of these steps via
genetic engineering approaches is becoming more common. Purification can be
simplified by the secretion of proteins into the medium or excretion into the
periplasmic space (Sect..4.2), but even if this can be achieved, it is still necessary
to isolate and purify proteins from a relatively complex mixture.
 72 K. Friehs and K. F. Reardon

A good technique for improving protein purification is the use of tags fused
to the protein product. The tag (e.g. Staphylococcusprotein A) is chosen to allow
the use of efficient separation techniques such as affinity chromatography. In
order to facilitate removal of the tag following purification, a specific cleavage
site can be placed between the tag and the desired protein. In Table 3, a number
of different tags and their chromatographic ligands are presented.
An example of the power of this method is presented in Fig. 2. In this case,
Staphylococcus protein A (SPA) was used as a fusion tag for the production of

Table 3. Examples of purification methods using tagged fusion proteins

Interaction Tag Ligand Target protein Ref.

Immunoaffinity 13-galacto- anti-13-gal,ab proline carrier 157

sidase protein
Pseudoimmuno- spA IgG EcoRI 158
affinity
Substrate affinity [3-galacto- APTG DNA binding 159
sidase protein
Common binding malE starch gene 5 protein 160
affinity phage M13
Metal chelate affinity his6 Ni(II)-NTA DHFR 161
Charge arg5 cation exchanger 13-urogastrone 162
Hydrophobic phe 11 phenyl groups 13-galacto- 163
interaction sidase
Cysteine thiol cys4 thiol groups galactokinase 163
interaction

Abbreviations: ab = antibody, APTG = p-aminophenyl-[3-D-thiolgalactoside, DHFR = dihydro-

folic acid reductase, IgG = immunoglobulin G, male = maltose binding protein, NTA = nitrilo-
triacetic acid, spA = Staphylococcus aureus protein A.

101 15
8 4"--
E
t-

O o
00 6 3 ~
Cq o
"r"
c-
O v

4 2
L
0 . M

r
..Q
< 2

0
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Fraction

Fig. 2. Affinity chromatography of EcoRI-SpA fusion on IgG 9 Total protein (Abs.) [] EcoRI
activity
Parameters Influencing the Productivity of Recombinant Cultivations 73

0.1
[////

r////

er i////

0.01
V/A
////

i/ill
I/// i/ill
J / / / / ~
0.001
Zephirol Formal- Formic acid Phenol Tego
dehyde

Fig. 3. Inactivation of plasmids by use of disinfectants. Rate of transformation relative to untreated

samples. Concentrations: [] 1% [] 0.1%

the restriction enzyme EcoR1 [1583. The fusion protein could be enriched and
purified in one step using affinity chromatography with IgG as the ligand.

4.5 Inactivation of Biological Waste

For safety reasons and for public acceptance, it is necessary to inactivate the
biological waste from the production of recombinant proteins. While this is also
done for many other cultivations, inactivation of recombinant microorganisms
involves not only killing whole cells but also significantly reducing the level of
recombinant DNA (such as transformable plasmid DNA).
A series of experiments was performed to determine the efficiency of plasmid
inactivation by different methods. The intact biological activity of the plasmids
was measured by comparing the transformation rate of isolated plasmid DNA
before and after treating the cells. The results of these experiments (shown in
Fig. 3) demonstrate that several commonly used compounds, including formal-
dehyde, formic acid, phenol, Tego, and Zephirol cannot inactivate plasmids
when used in normal concentrations [1643.
The only truly efficient method to inactivate plasmid DNA with respect to
transformation is thermal incubation. This can be accomplished continuously,
just as sterilization is.

5 Conclusion

In this review, we have presented a wide variety of factors and techniques that
influence the productivity of recombinant E. coli cultivations. These parameters
74 K. Friehs and K. F. Reardon

exert their effects o n m a n y levels, from increased plasmid stability to facilitated

p r o d u c t purification.
This i n f o r m a t i o n could be used to improve productivity by guiding the
design of a vector, the choice of a host strain, or the selection of cultivation
conditions a n d bioreactor type. I n addition, the examples provided here suggest
additional strategies by which the p r o d u c t i o n of r e c o m b i n a n t proteins could be
enhanced. The rapid pace of research in this direction should result in m a n y
exciting developments in the next few years.

6 References

1. Dennis K, Srienc F, Bailey JE (1985) Biotechnol Bioeng 27:1490

2. Twigg AJ, Sherratt D (1980) Nature (London) 283:216
3. Lacatena RM, Cesareni G (1981) Nature (London) 294:623
4. Tacon WCA, Bonass WA, Jenkins B, Emtage JS (1983) Gene 23:255
5. Terawaki Y, Itoh Y (1985) J Bacteriol 162:72
6. Sherratt D (1986) Control of plasmid maintenance. In: Booth IR, Higgins CF (eds) Regulation
of gene expression - 25 years on. Cambridge University Press, Cambridge, p 239 (39th Symp
Soc Gen Microbiol)
7. Scott JR (1984) Microbiol Rev 48:1
8. Shibui T, Kamizono M, Teranishi Y (1988) Agric Biol Chem 52:2235
9. Uhlin BE, Molin S, Gustafsson P, Nordstrom K (1979) Gene 6:91
10. Yarranton GT, Wright E, Robinson MK, Humphreys GO (1984) Gene 28:293
11. Seo J-H, Bailey JE (1985) Biotechnol Bioeng 27:1668
12. Jones IM, Primrose SB, Robinson A, Ellwood DC (1980) Mol Gen Genet 180:579
13. Edler C, Friehs K, Sch/igerl K (1989) DECHEMA Biotechnol Conf 3:549
14. Summers DK, Sherratt DJ (1984) Cell 36:1097
15. Imanaka T, Aiba S (1981) Ann NY Acad Sci 369:1
16. EnsleyBD (1986) CRC Crit Rev Biotechnol 4:263
17. Kumar PKR, Maschke H-E, Friehs K, Schfigerl K (1991) Trends Biotechnol 9:279
18. Anderson DM, Herrrmann KM, SomervilleRL (1983) US Patent No. 4371614
19. Skogman SG, Nilsson J (1984) Gene 31:117
20. Porter RD, Black S, Pannuri S, Carlson A (1990) Bio/Technology 8:47
21. Gerdes K, Rasmussen PB, Molin S (1986) Proc Natl Acad Sci USA 83:3116
22. Rosteck Jr PR, Hershberger CL (1983) Gene 25:29
23. Siegel R, Ryu DYD (1985) Biotechnol Bioeng 27:28
24. Chert J, Morrison DA (1987) Gene 55:179
25. Nishimura N, Taniguchi T, Komatsubara S (1989) J Fermentation Bioeng 67:107
26. Meacock PA, Cohen SN (1980) Cell 20:529
27. Chambers SP (1989) PhD Thesis, University of Warwick
28. Ryan W, Parulekar SJ (1991) Biotechnol Bioeng 37:415
29. Naock D, Roth M, Geutner R, Muller G, Undisz K, Hoffmieir C, Gaspar S (1981) Mol Gen
Genet 184:121
30. Jones SA, Melling J (1984) FEMS Microbiol Lett 22:239
31. Brownlie L, Stephenson JR, Cole JA (1990) J Gen Microbiol 136:2471
32. Hopkins DJ, Betenbaugh MJ, Dhurjati P (1987) Biotechnol Bioeng 29:85
33. Caunt P, Impoolsup A, Greenfield PF (1989) Biotechnol Lett. 11:5
34. Wouters JTM, Driehuis FL, Polaczek PJ, Van Oppenraag MLHA, Van Andel JG (1980)
Antonie van Leeuwenhoek 37:311
35. Impoolsup A, Caunt P, Greenfield PF (1989) Bioteehnol Lett 11:605
36. Stephens ML, Lyberatos G (1985) Biotechnol Bioeng 31:464,
37. Di Pasquantonio VM, Betenbaugh MJ, Dhurjati P (1987) Biotechnol Bioeng 29:513
Parameters Influencing the Productivity of Recombinant Cultivations 75

38. Henry KL, Davis RH, Taylor AL (1990) Biotechnol Prog 6:7
39. Berry F, Sayadi S, Nasri M, Thomas D, Barbotin JN (1990) J Biotechnol 16:199
40. Godwin D, Slater H (1979) J Gen Microbiol 111:201
41. Hawley DK, McClure R (1983) Nuc Acd Res 1l:2237
42. Horwitz MSZ, Loeb LA (1990) Prog. Nuc. Acd, Res. Mol. Biol. 38:137
43. Kobayashi M, Nagata K, lshihama A (1990) Nuc Acd Res 18:7367
44. Reinikainen P, L/ihde M, Karp M, Suominen I, Markkanen P, M/ints/il/i P (1988) Biotechnol
Lett 10:149
45. Botterman J, H6fte H, Zabeau M (1987) J Biotechnol 6:71
46. Elvin CM, Thompson PR, Argall ME, Hendry P, Stamford NPJ, Lilley PE, Dixon NE (1990)
Gene 87:123
47. Shapira SK, Chou J, Richaud FV, Casadaban MJ (1983) Gene 25:71
48. Amann E, Brosius J, Ptashne M, (1983) Gene 25:167
49. Makoff AJ, Oxer MD (1991) Nuc Acd Res 19:2417
50. Perez L, Vega J, Chuay C, Menendez A, Ubieta R, Montero M, Padron G, Silva A, Santizo C,
Besada V, Herrera L (1990) Appl Microbiol Biotechnol 33:429
51. Vidal-Ingigliardi D, Raibaud O (1985) Nuc Acd Res 13:1163
52. Pistillo, JM, Vishwanatha JK (1990) Biochem Biophys Res Comm 169:1129
53. Lukacsovich T, Orosz A, Baliko G, Venetianer P (1990) J Biotechnol 16:49
54. Pharmacia (1991) Pharmacia LKB Biotechnology, Uppsala, Sweden
55. Deng T, Noel JP, Tsai MD (1990) Gene 93:229
56. Yasukawa K, Saito T (1990) Biotechnol Lett 12:419
57. Owolabi JB, Rosen BP (1990) J Bacteriol 172:2367
58. Flores N, de Anda R, Guerec L, Cruz N, Antonio S, Balbas P, Bolivar F, Valle F (1986) Appl
Microbiol Biotechnol 25:267
59. Dikshit KL, Dikshit RP, Webster DA (1990) Nuc Acd Res 18:4149
60. Scanlan DJ, Bloye SA, Mann NH, Hodgson DA, Carr NG (1990) Gene 90:43
61. Berg BL, Stewart V (1990) Genetics 125:691
62. Su TZ, Schweizer H, Oxender DL (1990) Gene 90:129
63. Metcalf WW, Steed PM, Wanner BL (1990) J Bacteriol 172:3191
64. Claverie-Martin F, Magasanik B (1991) Proc Natl Acad Sci USA 88:1631
65. Bingham RJ, Hall KS, Slonczewski JL (1990) J Bacteriol 172:2184
66. Poindexter K0 Gayle III RB (1991) Gene 97:125
67. Repoila F, Gutierrez C (1991) Mol Micorbiol 5:747
68. Lucht' JM, Bremer E (1991) J Bacteriol 173:801
69. Unden G, Trageser M, Duchene A (1990) Mol Microbiol 4:315
70. Terasawa N, Masayuki I, Uchida Y, Kobayashi M, Kurusu Y, Yukawa H (1991) Appl
Microbiol Biotechnol 34:623
71. Fischer M, Fytlovich S, Amit B, Wortzel A, Beck Y (1990) Appl Microbiol Biotechnol 33:424
72. Mott JE, Grant RA, Ho Y, Platt T (1985) Proc Natl Acad Sci USA 82:88
73. Iijima S, Kawai S, Mizutani S, Taniguchi M, Kobayashi T (1987) Appl Microbiol Biotechnol
26:542
74. Stanley KK0 Luzio JP (1984) EMBO J 3:1429
75. Chan WKY, Belfort G, Belfort M (1988) Gene 73:295
76. Chan WKY, Belfort G, Belfort M (1991) J Biotechnol 18:225
77. Claverie-Martin F, Diaz-Torres MR, Yancey SD, Kushner SR (1991) J Biol Chem 266:2843
78. Donovan WP, Kushner SR (1986) Proc Natl Acad Sci USA 83:120
79. Schmucker R, Gfilland U, Will M, Hitlen W (1989) Appl Microbiol Bioteehnol 30:509
80. Olsen MK, Rockenbach SK, Curry KA, Tomich C-SC (1989) J Biotechnol 9:179
81. Surek B, Wilhelm M, Hillen W (1991) Appl Microbiol Biotechnol 34:488
82. Spanjaard RA, van Dijk MCM, Turion AJ, van Duin J (1989) Gene 80:345
83. Br6ker M, Amann E (1986) Appl Microbiol Biotechnol 23:294
84. Ernst JF (1988) Trends Biotechnol 6:196
85. Maurizi MR, Trisler P, Gottesman S (1985) J Bacteriol 164:1124
86. Boss MA, Kenton JH, Wood CR, Emtage JS (1984) Nucl Acids Res 12:3791
87. Simon LD, Randolph B, Irwin N, Binowski G (1983) Proc Natl Acad Sci USA 80:2059
88. Kadokura H, Yoda K, Mitsunobu I, Yamasaki M (1990) Appl Environ Microbiol 56:2742
89. Wittliff JL, Wenz LL, Dong J0 Nawaz Z, Butt TR (1990) J Biol Chem 265:35
90. Quaas R, McKeown Y, Stanssens P, Frank R, B16cker H, Hahn U (1988) Eur J Biochem
173:617
76 K. Friehs and K. F. Reardon

91. Hartley DL, Kane JF (1988) Biochem Soc Trans 16:101

92. Langley KE, Berg TF, Strickland TW, Fenton DM, Boone TC, Wypych J (1987) Eur J
Biochem 163:313
93. Georgiou G, Bowden GA (1991) Inclusion body formation and the recovery of aggregated
recombinant proteins. In: Prokop A, Bajpai RK, Ho CS (eds) Recombinant DNA technology
and applications. McGraw-Hill, New York, p 333
94. Spalding BJ (1991) Bio/Technology 9:229
95. Orsini G, Brandazza A, Sarmientos P, Molinari A, Lansen J, Cauet G (1991) Eur J Biochem
195:691
96. Gatenby AA, Viitanen PV, Lorimer GH (1990) Trends Biotechnol 8:354
97. Horwich AL, Neupert W, Hartl F-U (1990) Trends Biotechnol 8:126
98. Burton SJ, Quirk AV, Wood PC (1989) Eur J Biochem 179:379
99. Kane JF, Hartley DL (1988) Trends Biotechnol 6:95
100. Kopetzki E, Schumacher G, Buckel P (1989) Mol Gen Genet 216:149
101. Takagi H, Morinaga Y, Tsuchiya M, Ikemura H, Inouye M (1988) Bio/Technology 6:948
102. Claassen LA, Ahn B, Koo H-S, Grossman L (1991) J Biol Chem 266:11380
103. Bowden GA, Georgiou G (1988) Biotechnol Prog 4:97
104. Kenealy WR, Gray JE, Ivanoff LA, Tribe DE, Reed DL, Korant BD, Petteway SR (1987) Dev
Ind Microbiol 28:45
105. Lee S-M (1989) J Biotechnol 11:103
106. Mota M J, Teixeira JA (1990) Current Microbiol 20:209
107. Harrison STL (1991) Biotechnol Adv 9:217
108. Kula MR, Sch/itte H (1987) Biotechnol Prog 3:31
109. Scawen MD, Hammond PM, Sherwood RF, Atkinson T (1990) Biochem Soc Trans 18:231
110. Sampson BA, Benson A (1987) J Ind Microbiol 1:335
111. Altieri M, Suit JL, Fan M-LJ, Luria SE (1986) J Bacteriol 168:648
112. Luria SE, Suit JL, Jackson JA (1991) US Patent 4948735
113. Dabora RL, Eberiel DT, Cooney CL (1989) Biotechnol Lett 11:845
114. Rasched I, Oberer-Bley E (1990) Chem Lab Biotech 41:36
115. Hsiung MH, Becket GW (1988) Biotechnol Genet Eng Rev 6:43
116. Lee C, Li P, Inouye H, Brickman ER, Beckwith J (1989) J Bacteriol 17:4609
117. Alexander P, Oriel PJ, Glassner.DA, Grulke EA (1989) Biotechnol Lett 11:609
118. Emerick AW, Bertolani BL, Ben-Bassat A, White TJ, Konrad MW (1984) Bio/Technology
2:165
119. Zemel-Dreasen O, Zamir A (1984) Gene 27:315
120. Georgiou G, Shuler ML, Wilson DB (1988) Biotech Bioeng 32:741
121. Ohagi H, Kumakura T, Komoto S, Matsuo Y, Oshiden K, Koide T, Yanaihara C, Yanaihara
N (1989) J Biotechnol 10:151
122. Summers RG, Knowles JR (1989) J Biol Chem 264:20074
123. Blondel A, Bedouelle H (1990) Eur J Bioehem 193:325
124. Bedouelle H, Duplay P (1988) Eur J Biochem 171:541
125. Libby RT, Braedt G, Kronheim SR, March CJ, Urdal DL, Chiaverotti TA, Tushinski RJ,
Mochizuki DJ, Hopp TP, Cosman D (1987) DNA 6:221
126. Takahara M, Sagai H, Inouye S, Inouye M (1988) Bio/Technology 6:195
127. Barbero JOL, Buesa JM, Penalva MA, Perez-Aranda A, Garcia JL (1986) J Biotechnol 4:255
128. Habuka N, Akiyama KJ, Hideaki T, Miyano M, Matsumoto T, Noma M (1990) J Biol Chem
265:10988
129. I-Ienze P-PC, Hahn U, Erdmann VA, Ulbrich N (1990) Eur J Biochem 192:127
130. Shibui T, Matsui R, Uchida-Kamizono M, Okazaki H, Kondo J, Nagahari K, Nakanishi S,
Teranishi Y (1989) Appl Microbiol Biotechnol 31:253
131. Sherwood R (1991) Trends Biotechnol 9:1
132. Marks CB, Vasser M, Ng P, Henzel W, Anderson S (1986) J Biol Chem 261:7115
133. Oka T, Sumi S, Fuwa T, Yoda K, Yamasaki M, Tamura G, Miyake T (1987) Agricult Biolog
Chem 5:1099
134. Ohsuye K, Nomura M, Tanaka S, Kubota I, Nakazato H, Shinagawa H, Nakata A, Noguchi T
(1983) Nuc Acd Res 11:1283
135. Guan C, Li P, Riggs PD, Inouye H (1988) Gene 67:21
136. Fujimura T, Tanaka T, Kanako O, Morioka H, Uesugi S, Ikehara M, Nishikawa S (1990)
FEBS 269:71
Parameters Influencing the Productivity of Recombinant Cultivations 77

137. Anba J, Baty D, Lloubes R, Pages J-M, Joseph-Liauzum E, Shire D, Roskam W, Lazdunski C
(1987) Gene 53:219
138. Hogset A, Blinsmo OR, Saether O, Gautvik VT, Holmgren E, Hartmanis M, Josephson S,
Gabrielsen OS, Gordelaze JO, Alestrom P, Gautvik KM (1990) J Biol Chem 265:7338
139. Wadenstein H, Ekebacke A, Hammarberg B, Holmgren E, Kalderen C, Tally M, Moks T,
Uhlen M, Josephson S, Hartmanis M (1991) Bioteehnol Appl Bioehem 13:412
140. Fraser TH, Bruce TJ (1987) Proc Natl Aead Sci USA 75:5936
141. Kornacker MG, Pugsley AP (1990) Molec Microbiol 4:1101
142. Talmadge K, Brosius J, Gilbert W (1981) Nature 294:176
143. Morioka-Fujimoto K, Marumoto R, Fukuda T (1991) J Biol Chem 266:1728
144. Legoux R, Leptlatois P, Joseph-Liauzun, E (1991) US Patent 4945047
145. Letoffe S, Delepelaire P, Wandersman C (1991) J Bacteriol 173:2160
146. van Putten AJ, de Graaf FK, Oudega B (1987) Proc 4th Europ Con Biotechnol 4:593
147. Mackman N, Nicaud JM, Gray L, Holland IB (1986) Cur Top Microbiol Immunol 125:159
148. Lazzaroni JC, Portalier RC (1982) Eur J Appl Microb Biotechnol 16:146
149. Arian D, Portalier RC (1984) Eur J Appl Microb Biotechnol 19:5
150. Mosbach K, Birnbaum S, Hardy K, Davies J, Bulow L (1983) Nature (London) 302:543
151. Murakami Y, Furusato Y, Kato C, Habuka N, Kudo T, Horikoshi K (1989) Appl Microbiol
Biotechnol 30:619
152. Bialy H (1987) Bio/Technology 5:883
153. Thatcher DR (1990) Biochem Soc Trans 18:234
154. Fischer G, Schmid FX (1990) Biochem 29:2205
155. Dill KA (1990) Biochem 29:7133
156. Jaenicke R (1987) Prog Biophys Molec Biol 49:117
157. Hanada K, Yamamoto I, Anraku Y (1987) J Bio Chem 262:14100
158. Maschke H-E, Hebenbrock K, Friehs K (1990) DECHEMA Biotechnol Conf 4:379
159. Ullman A (1984) Gene 29:27
160. Maina CV, Riggs PD, Grandea AG, Slatko BE, Moran LS, Tagliamonte JA, McReynolds LA,
di Guan C (1988) Gene 74:365
161. Hochuli E, Bannwarth W, D6beli H, Gentz R, Stfiber D (1988) Bio/Technology 11:1321
162. Smith JC, Derbyshire RB, Cook E, Dunthorne L, Viney J, Brewer SJ, Sassenfeld HM, Bell LD
(1984) Gene 32:321
163. Persson M, Bergstrand MG, Billow L, Mosbach K (1988) Anal Biochem 172:330/337
164. Stiegelmeier C, Friehs K, Schilgerl K (1992) Acta Biotechnol (in press)
Modeling and Control for Anaerobic
Wastewater Treatment

Elmar H e i n z l e 1, I r v i n g J. D u n n and Gerhard B. R y h i n e r 2

Biological Reaction Engineering Group, Chemical Engineering Department,
Swiss Federal Institute of Technology, Zfirich, Switzerland

Dedicated to Professor Dr. Karl Schfigerl on the occasion o f his 65th birthday

List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2 Recent Reviews of the Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3 Reactions in Anaerobic Digestion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4 Dynamic Mass Balance Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.1 Biomass Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2 Substrates and Product Balances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5 Stoichiometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6 Kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7 Ion Charge Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
8 Modeling Biofilm Kinetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
9 Gas Phase Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10 Comparison of Simulations with Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10.1 Steady State Comparison of Experimental and Simulated Carbon Balances . . . . . . 96
10.2 Comparison of Dynamic Experimental and Simulation Results . . . . . . . . . . . . . . . . 96
11 On-line Measurement and Observation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
12 Control of Anaerobic Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
12.1 Simple Controller Design Using Simulations of a Complex Process Model . . . . . . . 100
12.2 Proportional-Integral-Differential Control of pH . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
12.3 PID Control of Organic Acid Concentrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
12.4 Control of Dissolved Hydrogen Concentration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
12.5 Control Based on the Methane Content of the Gas . . . . . . . . . . . . . . . . . . . . . . . . . 105
13 Adaptive Control and Optimization of Anaerobic Processes . . . . . . . . . . . . . . . . . . . . . . 106
14 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 ll
15 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1t 2

The recent literature on the modeling and control of anaerobic wastewater treatment processes is
reviewed. An example from the author's personal work is used to describe how a dynamic simulation
model can be developed from the basis of multi-organism growth kinetics and mass balancing
techniques. This included consideration of the organic acid dissociation equilibria important for pH
calculation and the thermodynamic influence of hydrogen on the reactions involving propionic and
butyric acids. Liquid phase balances were linked to gas phase balances by gas-liquid transfer
considerations. It is shown in detail how the model was applied to one and two stage experimental
reactors for the design and tuning of controllers. Both conventional PID controllers and adaptive
optimizing controllers, employing simple input-output models with an objective function, were
tested.

l To w h o m all correspondence should be addressed

2 Present address: Sulzer Chemtech, 8401 Winterthur, Switzerland

Advances in Biochemical Engineering

Biotechnology, Vol. 48
Managing Editor: A. Fiechter
~S) Springer-Verlag Berlin Heidelberg 1993
80 E. Heinzle, I. J. Dunn and G. B. Ryhiner

List of Symbols

Symbol Unit Name

A [m 2] Reactor cross sectional area
ai, bi, c [-] Model parameter
C [mol m - 3] Concentration of inorganic compound
C [C-mol m 3] Concentration of organic compound
EF [-] Equilibrium factor
F [m a h - 1] Liquid flow rate
G -mol h - 1] Gas flow rate
AGO' -kJ tool- 1] Free Gibb's enthalpy
AG o -kJ mol- *] Standard free enthalpy of formation
Hi -mol m - 3 bar- 1] Henry-coefficient
Ks -C-mol m - 3] Saturation constant
Kl -C-mol m - 3] Inhibition constant
KA -mol m - 3] Acid dissociation constant
KB -mol m - 3] Base dissociation constant
Kw -mol 2 m 6] Water dissociation constant
Kr [*] Controller gain
ka [h-*] Death rate of organisms
kLa [h -1] Gas liquid phase mass transfer coefficient
N [mol m - a h - 1] Mass transfer rate
ni [moll Amount of gas i in reactor gas phase
P [bar] Total pressure
PI [] Performance index
bar m 31
R Universal gas constant
r [mol m -a h -1] Reaction rate
rp [C-mol m -3 h -1] Product formation rate
rs [C-mol m - 3 h - 1] Substrate consumption rate
rx [C-tool m - 3 h- 1] Growth rate
T [K] Temperature
UB [m h -1] Bubble rising velocity
U [*] Manipulated process input
V [m33 Volume
X [C-mol m - 3] Biomass concentration
xi [-] Molar fraction of gas i in bubbles
Yx/s [C-mol C-tool- 1] Biomass yield from substrate S
YP/S [C-mol C-mol- 1] Product yield from substrate S
Y [*] Process output

Greek symbols:
[*] Gain adaptive optimizer
[3 [,] Working point optimizer
Modeling and Control for Anaerobic Wastewater Treatment 81

[mol m 3] Different quantity in Eq. (14)

Fractional gas hold up
~"min [*l Minimum forgetting factor optimizer
[h -1] Specific growth rate
~max Fh - l l Maximal specific growth rate
v~ [-] Stoichiometric coefficient
[*] Sensitivity optimizer
"[sl [h] Solid retention time

[*]: Units varying depending on application

Indices:
AH acid
A- dissociated AH
Ac acetic acid
An- anion
B base
Bu butyric acid
D differential
g gas phase
HAc undissociated acetic acid
HBu undissociated butyric acid
HPr undissociated propionic acid
i refers to component
I integral
K+ cation
L liquid phase
P product
Pr propionic acid
S substrate
sl solid
titr titrator base
tot total
W water
X biomass
Z surplus cations
0 reactor feed stream

Abbreviations:
COD Chemical oxygen demand
GC Gas chromatograph(y)
P Proportional
PI Proportional-integral
PID Proportional-integral-differential
VFA Volatile fatty acids
82 E. Heinzle,I. J. Dunn and G. B. Ryhiner

1 Introduction

Anaerobic processes have obvious advantages for wastewater treatment, espe-

cially high strength wastewaters. These are a sludge production decrease (one-
tenth that of aerobic treatment), no aeration energy requirement, reduction of
odors in a closed system, and methane energy production. Distinct disadvan-
tages remain: methanogenic organisms grow slowly, the stability of anaerobic
processes can be upset either by toxic substrates or by overloading, and the
process is not completely understood.
Much progress has been made in the understanding of anaerobic processes,
from both the microbiological and from the theoretical standpoint. This contri-
bution will attempt to discuss and evaluate recent efforts in the mathematical
modeling and in the design of control strategies for anaerobic wastewater
treatment. In some respects the two areas are related, since modeling leads to
quantitative understanding and control system design benefits from this. Simple
control methods are based on simple qualitative ideas, but it will be shown that
a model can be useful in simply reducing the time necessary to test the
controller. The slow rate of anaerobic processes makes it useful to minimize
experimentation and to replace it as far as possible with simulation studies. This
combined approach using simulation and experimentation to develop control
strategies and to determine control parameters will be emphasized here.
In discussing the model developments it must be kept in mind that in
anaerobic wastewater treatment one deals with a multiorganism system, of yet
unknown complexity, with respect to the number of reactions and their inter-
mediates. In addition, the spatial configuration of the individual organisms in
flocs, granules and biofilms is largely unknown and would itself have a
pronounced influence on the overall process, for example due to hydrogen
scavenging. Comparing the multiorganism-system to that of a single organism
and remembering that no single kinetic model will describe even all the
complexities of pure culture dynamics, then it must be a combination of naivety
and fearlessness that allows engineers and scientists to attempt such model
development.
Modelling must be viewed in the light of its purpose. If the understanding of
the process complexities is the goal, then a detailed model may be justified,
especially if the many variables are actually measured and compared with the
model. Complex models may be very useful for design of state estimators and
controllers. For the purpose of model based state estimation, it is necessary to
simplify a complex mechanistic model to the point that it can be treated with the
available theory. A curve-fitting non-mechanistic model may be another ap-
proach possible if on-line measurements are available.
Modelingand Controlfor AnaerobicWastewaterTreatment 83

2 Recent Reviews of the Field

Andrews [1] developed a simulation model based on Monod kinetics with

substrate inhibition for growth, using a constant yield for consumption and
production. Undissociated fatty acid was considered as the substrate for the
formation of methane, and the dissociation equilibrium was considered. In
addition the bicarbonate equilibrium and a charge balance was used to obtain
dissolved CO2 and pH. The transport rate of CO 2 to the gas phase was
considered to obtain the total gas rate. The simulations gave dynamic results in
terms of un-ionized acid substrate, bicarbonate, total gas flow, gas composition.
Thus it was possible to consider the response of the continuous digester to step
changes in substrate feed. This model permitted the simulation of various
control strategies, such as recycling of thickened sludge, and the scrubbing of
carbon dioxide from the liquid phase using recycle methane. This early work
represented the state of the art in anaerobic process modeling, part of which was
done on a hybrid analog-digital computer.
One of the leaders in the applied anaerobic treatment field has reviewed the
research and development in anaerobic treatment during the last century [34].
From this review the microbial complexities are evident, and it is also clear that
much has still to be done regarding basic research and process innovation.
Rozzi et al. [50] has more recently used the Andrews model with some
modifications: gas-liquid equilibrium was assumed, and the fatty acid substrate
was considered to be acetic acid which reacted with the bicarbonate. Control of
pH with the addition of NaOH, Na2CO 3 and Na2HCO3, and of bicarbonate
alkalinity with NaOH addition was studied with simulation. Also considered
was an early warning on the basis of CO 2 in the off-gas for conventional pH
control with the base addition. A later paper [51] considered bicarbonate
alkalinity control in more detail and discussed its automatic measurement. This
work concluded that bicarbonate alkalinity control is promising. Wiesmann
[69] gave an overview with special emphasis on acetic acid consumption and
control of pH. Schfirbfischer and Wandrey [55] reviewed many aspects of
anaerobic digestion including reaction schemes, contribution of various organ-
ism groups, modeling, on-line measurement and estimation as well as control.
Barnes and Fitzgerald [5] gave an overview on anaerobic reactions and
kinetics with special emphasis on the influence of toxic compounds (metals and
organics).
An extended overview on the biology of various anaerobic processes is given
in the books edited by Zehnder [70] and by B61aich et al. [54]. In an earlier
book [22] many aspects including modeling are extensively treated. The most
important chapters related to the topics of the present paper are on the
modeling and its application for start-up and control [58] and on energetics and
biomass yields of anaerobic processes [59].
A recent review [41] stated that with the exception of the hydrolysis of solids
all other subprocesses of anaerobic digestion have been modeled successfully
84 E. Heinzle, I. J. D u n n and G. B. Ryhiner

Table 1. Rate parameters as reported from the literature by Pavlostathis and Giraldo-Gomez [41].
VSS - volatile suspended solids

Substrate
~maxCs~
k Ks gmax Y
KsY /
g COD g COD g VSS
d-1
g VSS d m3 g COD

Butyric acid 6.2-17.1 12-500 0.13-1.2 0.01-0.27

Propionic acid
Acetic acid 2.6-26 11-930 0.08-0.7 0.01-0.054
H 2 and CO2 1.92-90 4.8 x 10-5-0.6 0.05-4.07 0.017-0.13

with Monod-type kinetics. This includes the conversion of sugars to organic

acids, the conversion of acetic acid to methane and the reduction of carbon
dioxide by hydrogen to methane. The reasons given for the wide variation of
kinetic parameters shown in Table 1, where the widely varying conditions, and
possibly inaccurate measuring procedures. Certainly the complexity of anaer-
obic processes is also reflected in these widely varying kinetic constants.

3 Reactions in Anaerobic Digestion

Anaerobic degradation is a very complex multi-substrate, multi-organism pro-

cess. Although there are still some mysteries, most mechanistic models consider
basically the four steps shown in Fig. 1.
Here the polymer materials (carbohydrates, proteins or lipids) are hydro-
lyzed to yield the monomer compounds (e.g. amino acids, sugars and fatty
acids). In some cases particles are contained in the wastewater. Structured
modeling of the digestion of particulate material has been addressed by Bryers
[8] and Tschui [62]. Anaerobic hydrolysis of fats and proteins is a complex
process, involving various groups of organisms but is quite well understood
[37]. Many compounds are not easily degraded to short chain carboxylic acids.
These include especially aromatic compounds like lignin-derived materials [9].
In a second acidogenic step these compounds are transformed to organic
acids (mainly acetic, propionic and butyric acid) carbon dioxide and hydrogen.
In the third acetogenic step organic acids with more than three atoms of carbon
per molecule are converted to acetic acid and hydrogen [20]. The last meth-
anogenic steps convert acetic acid to methane and reduce carbon dioxide with
hydrogen to methane, a process well investigated but incompletely understood
on the biochemical level [63].
In Table 2 the most important reactions are summarized with whey as the
starting substrate. Only the elements C, H and O are considered in these
reactions.
Modeling and Control for Anaerobic Wastewater Treatment 85

Polymers (Proteins,
fats, carbohydrates ] '~

Monomers (sugars, [
amino acids,...) [

IHBu - =

Fig. 1. Reaction scheme of anaerobic degradation. Dashed arrows indicate gaseous compounds
transferred out of the liquid phase. HPr - propionic acid; Pr- - propionate; HBu - butyric acid;
Bu - butyrate; HAc - acetic acid; Ac- - acetate

Table 2. Most important reactions in the anaerobic degradation of whey. Reaction R1 is given in a
C-mol basis (one carbon atom per molecule); vi stoichiometric coefficients.
(AGo' f o r T = 2 5 ~ p H = 7 , p = l b a r , xn2o = 1, c = l m o l k g - 1 )

Hydrolysis and acidification of whey

CH1.8500.853 + vH2oH20 ~ vBuCH2Oo. 5 + VprCH20o.667
-F VAcCH20 + Vco~CO 2 + vlt2 H 2 (R1)

Acetification
CH3(CH2)2COO- 4- 2HzO ,~-2CH3COO- + 2 H z + H + AG ~ + 71.7kJ (R2)
CH3CH2COO + 2 H20 ~- CH3COO- + CO 2 -~ 3 H 2 AG O' = + 48.3 kJ (R3)

Methanogenesis
CH3COO- + H + CH 4 4- CO 2 AG O' = - 35.8 kJ (R4)
CO 2 + 4 H 2 ~- CH 4 + 2 H20 AG O' = - 130.7 kJ (R5)

R e f e r r i n g t o T a b l e 2, t w o t y p i c a l p r o c e s s c a s e s m a y b e d i s t i n g u i s h e d .

If solid matter has to be degraded, then the hydrolysis step may be rate
l i m i t i n g , a n d o v e r l o a d i n g will n o t b e i m p o r t a n t .
- O v e r l o a d i n g o f t h e r e a c t o r m a y o c c u r if e a s i l y d e g r a d a b l e m a t e r i a l is t o b e
t r e a t e d . I t is a g r e e d t h a t h y d r o g e n p l a y s a n i m p o r t a n t r o l e i n t h e p r o c e s s
s t a b i l i t y . D u e t o t h e p o s i t i v e free e n e r g y o f r e a c t i o n , r e a c t i o n s R 2 a n d R 3 in
86 E. Heinzle,I. J. Dunn and G. B. Ryhiner

Table 2 can only proceed to the right, if the products, especially H2, are
continuously removed. Thus if the final methanogenesis steps are inhibited due
to toxicity or pH then the hydrogen and acetic acid accumulation will stop the
reactions, R2 and R3. Because the concentration of H 2 influences the equilib-
rium relation to the second (R2) and even third power (R3), it is evident that
the hydrogen partial pressure must be kept sufficiently low. This is accom-
plished by the methanogenic scavengers, whose rates cannot keep up in the
event of substrate overloads. H2 consuming organisms can remove H 2 locally
to concentration levels which make the reaction proceed exergonically [11].

According to Thauer /-61] and Archer [2] the oxidation of propionic to

acetic acid should thermodynamically only be possible at pH 2 < 10 -4 bar.
Experimental measurements of dissolved hydrogen in this range are difficult to
obtain. Experiments by Kaspar [28], Kaspar and Wuhrmann [29] and Denac
and Dunn /-15] applying pure H 2 did not show any significant inhibition.
Kaspar [28] and recently Whitmore et al. [68], who actually measured dis-
solved H 2 down to the 1 gmolar range, pointed out that this may be caused by
diffusion-reaction phenomena. Local high consumption rates of H 2 in a biofilm
or floc may reduce local H 2 concentration to allow the endergonic reactions R2
and R3 to proceed. The importance of hydrogen as controlling intermediate was
extensively reviewed by Harper and Pohland/-26] and in recent book [54].
From a series of observations it seems to be evident that disturbances in the
methanogenic step, which generally is considered to be the most sensitive one,
leads to accumulation of I-I2 [2, 26, 35, 67]. Additionally the state of an
anaerobic reactor may be characterized by its volatile fatty acid levels, from the
C H 4 / C O 2 ratio and from the total gas production rate.
The compounds shown in Fig. 1 and Table 2 are not the only ones occurring
in anaerobic digestion. Investigating the influence of hydrogen on the thermo-
dynamics of reaction in a reactor continuously fed with ethanol and propionate,
it was found that a shock load of feed did not just stop the propionate
conversion, as expected from hydrogen considerations, but that it induced the
reduction of the propionate to propanol [57].
A new modeling concept was proposed by McCarty and Mosey [36] based
on complex population dynamic interactions of "butyric bacteria", which
produce butyric acid at low pH and acetic acid normally from glucose, and
"propionic bacteria", which produce propionic and acetic acid at neutral pH.
The "propionic bacteria" are seen to be mainly Enterobacteria that grow well at
high concentrations and neutral pH, while the "butyric bacteria" are Clostridia
that grow well at low concentrations, producing acetic acid but produce butyric
acid at low pH. Also involved are the parallel reactions of Syntrophomonas,
which converts higher organic acids to propionic and acetic acids, and Syn-
trophobacter, which utilizes only propionic acid. Both rely on the hydrogen
consumption of Methanobacterium for their thermodynamics. Thus a reactor
overload would first cause the acetate and propionate to increase, lowering the
pH; this would then cause the formation of butyrate. Hydrogen would quickly
accumulate and be consumed, as would acetate. Propionate would fall slowly
Modeling and Control for Anaerobic Wastewater Treatment 87

because only the Syntrophobacter can use it. One thing is clear from the new
modeling concepts presented in this paper: many mysteries still remain with
regard to the microbiology and its kinetic modeling.
Smith and McCarty [57] used a four bacteria-type model for mixed
substrates ethanol and propionate. Two types were used for oxidation of the two
substrates to acetate and the other two types for methanogenesis. Thus four
component balances and four biomass balances were used.
Costello et al. [12] used the Mosey model [-39] as a basis but also included
lactic acid as an important intermediate. Six groups of bacteria were considered.
One group formed acetic, butyric or lactic acid and a second degraded lactic
acid into propionic or acetic acid. Two groups were acetogenic and formed
acetic acid from butyric and from propionic acid. Methane was formed by two
groups, one by hydrogen-consuming bacteria reducing carbon dioxide and
another from acetate.
A four organism Mosey-type model was used by Dochain et al. [19] to
develop an adaptive control algorithm for the control of hydrogen concentra-
tion. A quasi-steady state assumption regarding the glucose concentration
allowed relating the accumulation rate for hydrogen to the glucose inflow rate
and the hydrogen outflow gas rate. A further assumption allowed neglecting the
propionate reaction path. For the simulated results both hydrogen and acetate
were assumed to have inhibiting effects on the methanogenic steps and the
originally neglected propionate reaction was included. Without control the
hydrogen-consuming methanogens would be washed out of the reactor.
Ryhiner [52] used the reaction scheme shown in Fig. 1 and Table 2 including
a six organism group model to describe anaerobic digestion of whey.

4 Dynamic Mass Balance Equations

For the purpose of modeling, many reactors used in anaerobic degradation can
be considered as well mixed. The basic balance equation for a well mixed liquid
phase is
mass transfer from -
F accumulation ~ = input - output + reaction + the gas to the
[_of component i j
liquid phase
d C L.
V L d ~ = FL (CLio -- CLi) + ri VL + Ni VL (1)

4.1 Biomass Balance

The simplest approach considers biomass to be homogeneously suspended in

the reactor. In many reactor systems the biomass residence time is made longer
88 E. Heinzle, I. J. Dunn and G. B. Ryhiner

than the hydraulic residence time, either by biomass recycling or by immobili-

zation. The biomass retention can be modeled using a sludge retention time (zsl)
which is always larger than the hydraulic retention time. The resulting biomass
balance is:
~loss in~
accumulation = production - death - [_efftuentl

dX i Xi
dt - rx, - Xi'kd, -- - - (2)
Zsl
where
rxl = ~iXi (3)
where i indicates populations feeding on the various substrates (whey, butyric,
propionic and acetic acid, hydrogen). The term Xikd~ also accounts for mainte-
nance and endogenous reactions.

4.2 Substrates and Product Balances

For non-volatile substrates, intermediates and products Eq. (1) simplifies to

accumulation ] = flow _ consumption + production

of component i
dCLi
VL dt - FL (CLio -- CLi) + ~rsiVe + ~re~Ve (4)

Each substrate may be consumed and produced by several reactions which are
taken into account by the summation. The corresponding balances for volatile
components (H2, CO2 and CH4)include mass transfer term as shown in Eq. (1).
The relation of substrate consumption is

-- ~iXi (5)
rs~ -- Yxi/si

Formation of a product j from substrate i is usually assumed proportional to

biomass formation. This is expressed using yield coefficients

t'tiXiYPjsi (6)
rvj -- Yx~/si

5 Stoichiometry

Stoichiometric coefficients for reactions R2 to R5 in Table 2 are well defined.

Only the stoichiometric coefficients (vi) of reaction R1 and corresponding other
Modeling and Control for Anaerobic Wastewater Treatment 89

hydrolysis reactions are not known a priori. Experiments under defined condi-
tions are necessary to determine these coefficients. The six unknown stoichio-
metric coefficients of the acidification of whey were determined after measure-
ment of acid production and COD reduction in the acidification reactor of the
two stage system [52]. In this reactor no methane was produced, and reactions
R2 and R3 were assumed not to proceed under these conditions. The experi-
mentally determined values were: vBu = 0.5; Vpr = 0.12; VAc = 0.15; VCO2 = 0.23;
Vn2 = 0.24; Vn20 = 0.095.
More difficult is the estimation of biomass yield coefficients of individual
reactions. One successful approach uses ATP balancing. The yield of biomass
from ATP (Yx/ATP) is quite constant for many organisms [-4, 481. The average
value given there is YXmTP ~ 10 g tool 1. If the production of ATP from
metabolic reactions is known, the biomass yield can be estimated in a straight-
forward manner. Smith and McCarty [-57] developed these stoichiometric
coefficients from the free energy of the catabolic reactions. Growth was related
to energy availability by calculating the difference between the free energy of the
catabolic reaction and the energy involved in the conversion of the substrates to
pyruvate. It was noted that at least 6 reactions occurred during the experiments,
which involved reduced intermediates that were not included in the model.
Erickson [21] presented an electron balance and discussed it as a means of
estimating biomass yield coefficients in anaerobic processes.

6 Kinetics

The hydrolysis of especially particulate material is a very complex process. Often

first order kinetics with respect to particles are applied [62]. In the work of
Ryhiner [-52] the kinetics of biomass growth are described by simple Monod
relationships for whey, butyric acid, propionic acid and hydrogen.

~maxi Csi
lai - Cs i + Ks ~ (7)
or with an additional substrate inhibition term (acetic acid)

I-tmax~Csi
gi = Cs i + Ks ~ + Csi2/Kii (8)
The acidic form was taken to be the kinetically determining form, as was done in
previous work [14, 32, 69]. The concentrations of the actual substrates (un-
dissociated carboxylic acids) are pH dependent. At neutral pH-values, only
C O 2 - / H C O 3 - and in sulfur-containing wastewaters HzS/S- have a reasonable
buffer capacity. All other acids mainly exist in their salt form.
In the acetogenic step, acetic acid, hydrogen and carbon dioxide are
produced from propionic and butyric acid. The thermodynamic limits for these
90 E. Heinzle, I. J. D u n n and G. B. Ryhiner

reactions are incorporated by estimating the chemical equilibrium limits for

butyric acid and propionic acid using an empirical factor which was varied as a
function of the ratio of the time varying equilibrium constant, based on the
instantaneous concentrations, to the true value based on AG o' . The value of the
factor varied as an S-shaped function from 0 to 1 [52]. Another but similar
approach was taken by Mather [33], who also used Eq. (7) to describe specific
growth rate, but set ~max K AG. Here AG was the actual free reaction
=

enthalpy at actual concentration values and K was an empirical constant, which

caused ].lmax to increase monotonously with increasing substrate and decreasing
product concentrations. No real saturation value for growth rate at high
substrate concentration is obtained.
In a recent review [41] it was stated that with the exception of the hydrolysis
of solids all other subprocesses of anaerobic digestion can be modeled success-
fully with Monod-type kinetics.

Table 3. A n a e r o b i c stoichiometric a n d kinetic p a r a m e t e r s from the literature. (C-molecular weight

of biomass: 29.5 g C - t o o l - 1)

Substrate ~l,max Ks K~ Yx,s kd Reference

[h -1] [ C - m o l m 3] [ C - m o l m -3] [C-mol [h -1]
C - m o l - 1]

Whey 0.36 0.7 - 0.101 _ [30]

0.25 - 0.03 - [33]
0.40 0.25 - 0.03 0.0004 [52]
Butyric 0.0113 0.0000851 0.17 0.013 - [14]
acid
0.0096 12.5 - 0.051 0.0004 [8]
0.0154 0.33 - 0.083 0.00011 [25]
0.011 0.0322. 3 _ _ 0.05 0.0004 [52]
Propionic 0.0108 0.00082 0.22 0.023 - [14]
acid
0.0033 21.4 - 0.023 0.0004 [8]
0.013 1.6 - 0.06 0.0004 [25]
[33]
0.003 - 0.06 [52]
0.0053 0.00742, 3 _ 0.04 0.0004
Acetic 0.0145 0.003 a 0.33 0.043 0.018 [14]
acid
0.0142 15.6 - 0.041 0.00013 [8]
0.0142 5.16 - 0.0567 0.0006 [25]
[69]
0.01'* 0.12,4 1.42,4 - - [52]
0.0083 0.12 1.42 0.05 0.0004
Hydrogen s 0.0583 0.0375 - 0.021 0.0004 [8]
0.0583 0.038 - 0.0283 0.0004 [25]
0.001 - [39]
0.0002 - 0.03 [33]
0.058 0.001 - 0.026 0.0004 [52]

i A s s u m p t i o n 50% content of C in biomass; 2 free acid; 3 experiment; '~ m e a n value; 5 d i m e n s i o n for

K s [ m o l m - 3 ] , Yx,s [ C - m o l tool - 1 ]
Modelingand Controlfor AnaerobicWastewaterTreatment 91

Costello et al. [12] used Monod-type kinetics to determine substrate uptake

rates, and the growth rates were related to either the product rates or the
substrate rates. Hydrogen regulation and inhibition was considered for all acid-
forming groups in the manner of Mosey [39] by multiplying the rates of
substrate and product formation by hydrogen pressure functions in such a way
that the stoichiometry was satisfied. Inhibition by pH was achieved by multiply-
ing the substrate rates by a suitable factor. Product inhibition was modeled with
two modified Monod forms on the substrate uptake rates, using the respective
acid product concentration for the four groups.
A companion paper [13] tested the model against three different sets of
anaerobic reactor data from the literature. The authors found the data from the
experiments in the literature to be insufficient for testing the model completely.
Particularly lacking was influent alkalinity as well as effluent organic and
inorganic carbon, which would have made a carbon balance possible. The
following conclusions could be made: hydrogen inhibition at high partial
pressures was found to require adjustment of various parameters, the impor-
tance of lactic acid as an intermediate was confirmed and well-modeled, it was
suggested that other bacterial groups may be required for a complete model,
adaptation to low pH was not modeled and may be required, and a 10 day solids
residence time described biofilm systems well.

7 Ion Charge Balance

Calculation of the available undissociated organic acid substrates involves

considering the pH, which greatly influences anaerobic kinetics.
A recent review of anaerobic modeling concepts [36] has stressed the
importance of three buffer systems to determine the pH: carboxylic acid-
bicarbonate, acetic acid-acetate if bicarbonate ions are absent, and ammonia-
ammonium if carbon dioxide is absent. The Mosey model [39] utilized these to
model operation over a wide range of pH with a computer solution that chose
the dominating buffer system according to the reactor conditions. The pH
sensitivity of the organism growth was modeled by a simple variable factor and
the organism death rate pH sensitivity.
Costello et al. [12] used the Mosey model [39] as a basis but also included
lactic acid as an important intermediate. Dynamic mass balances for organic
acids and inorganic carbon were used. The equilibrium relations described the
dissociation of carbon dioxide and all organic acids. A dynamic sodium ion
balance was used in conjunction with a charge balance to calculate pH.
In any liquid element an ion charge balance yields

2(Ccationi • chargei) = ~(C,nionj X charges) (9)

92 E. Heinzle, I. J. Dunn and G. B. Ryhiner

In the pH range of interest (around pH = 7) all strong acids and strong bases are
completely dissociated. Weak acids and bases are only partly dissociated.
AH~A- +H + (10)
The degree of dissociation is determined by the dissociation constant as
CA_H +
K A -- - - (11)
CA
where: C A concentration of the undissociated acid; C A - concentration of the
corresponding base (dissociated acid).
Moderately strong acids occurring in anaerobic systems are: acetic:
KAc ---- 1.73 x 10 - 2 [ m o l m - 3 ] , propionic: Kpr = 1.31 x 10 -2 [mol m - 3 ] ,
butyric: KAc = 1.44x 10 -2 [mol m-3], carbon dioxide: KAc = 3.02x 10 -2
[mol m-3], hydrogen sulfide: KH2s = 1.26 x 10 -4 [mol m -3] and ammonium:
KNH4+ = 5.28 x 10 -7 [mol m-3]. The only moderately strong base is ammonia:
KNH3 = 1.85 X 10 .2 [m01 m-3]. At pH values below 7 ammonia is practically
completely dissociated (cNn3-< CNH4+*0.005). Correspondingly, the buffering
effect of ammonia is negligible. The existing cations in anaerobic degradation
process are therefore cations (K +) from strong bases (e.g. Na +, K +, NH4+,
Ca2 +). In the expected pH range one always has ZCK+ >>CH+,where ZCK+ is the
total cation concentration. Negative ions are mainly from strong acids (e.g. C1-,
SO #-) and from weak acids (Ac-, Pr-, Bu-, HCOj). The concentration of
CO 2- is always much smaller than that of HCO 3 (e.g. at pH = 7;
Cco~-/CHco~ ~ 4 x 1 0 - 4 ) . From this one obtains

KA.
2 K,~Kw -C, ..... + 2 CK+ = 2 KAi~'CH + CA.... ~+ ~CAn- (12)
KB~ + CH+

KA~ Acid dissociation constant (e.g. Knc); KBI - Base dissociation constant (e.g.
-

KNn3); K w - Dissociation constant of water; C B..... total concentration of base

i, i indicates ammonia; CAto,,i - total concentration of acid i; i indicates CO2,
acetic, propionic, butyric acid and hydrogen sulfide. From Eq. (12) the pH can
be estimated for any situation solving a non-linear equation provided the total
concentrations of weak acids (C A..... ), weak bases (CB..... ), cations of strong
bases (CK+) and of anions of strong acids (CAn-) are known.
From the above terms, CO 2 is always important. If at 1 bar total pressure
30% of the gas is CO 2, the equilibrium liquid phase concentration will be
around 4 [mol m-3]. At pH below 7.25, less than 1% of ammonia nitrogen
exists in the free base form. If the total concentration of ammonia is rather low
( < 1.5 [mol m-3]) the corresponding terms can be neglected. The H2S/HS-
system has its highest buffering capacity at pH 6.9. If the pH is below 6, this
buffer system is not important. The experiments described by Ryhiner [52] were
typically around pH ~ 6, which reduces the importance of this system for
buffering. The concentration of sulfur in whey waste water is quite low. The
M o d e l i n g a n d C o n t r o l for A n a e r o b i c W a s t e w a t e r T r e a t m e n t 93

measured gas concentrations were usually below 2%. The corresponding H2S
concentration in the liquid phase was therefore always < 0.6 [mol m-3].
Neutralization reaction is a very fast reaction reaching equilibrium almost
instantaneously. During pH control strong base is added. The addition of strong
alkali causes an increase in ZCK+ which, by solving Eq. (12), results in a decrease
of CH+.
It is possible to use only the difference between cations and anions
Cz = ~CK+ -- 2CAn (13)
For the numerical solution it is useful to rearrange Eq. (12) and use Eq. (13) in
the form,
Ki KBi
6 = ~ Ki + Cn+ Ctot,~- Z Kw C, ..... - C z (14)
KBi +
CH+

CH+ is then varied iteratively until 6 approaches sufficiently closely a value of 0.

This numerical solution is not always trivial using conventional methods for
non-linear algebraic equations (e.g. Newton Raphson, Regula falsi as incorpo-
rated in the simulation language ACSL). Bellgardt [7] used a very robust
algorithm which has been found to converge in any case. The iteration shown in
Fig. 2 was started at CH+ = 10 -14 k m o l m -3. CH+ was then increased by a
factor fn (usually 10) until 6 < 0. This corresponds to moving along one single
curve in Fig. 2 until the line intersects with 6 = 0. Then C , + was decreased by fn
and fn+ 1 = ~fn. NOW, the previous procedure was repeated again until 6 < 0.

10

-10 .... , .... , ....

5 10 15

pI-I

Fig. 2. E s t i m a t i o n of p H for different cation c o n c e n t r a t i o n s C z using Eq. (22) CAr = 4 tool m - 3;

Cco 2 = 4 tool m - 3 ; CH2s = 0.6 mol m - 3 ; CNH3 = 4 mol m - 3 ; C z = 0, 2, 4, 6, 8 mol m -3. The a r r o w
indicates increasing c a t i o n c o n c e n t r a t i o n
94 E. Heinzle, I. J. D u n n and G. B. Ryhiner

Titration

Cation :1
Dynamic Cation __.~
Cation uent " - ' - - ~ M a s s Balance
Ion Charge
Balance
(Algebraic
equation)
Acid Feed Dynamic Acid M a s s Balances
(Total m a s s of week acids and
Acid Effluent b a s e s - acetic, CO2, NH3,...)

Production,
Consumption

CA
r=f(CA) [~ Concentration
of free acids

Fig. 3. Schematic of the ion charge balance iteration within a dynamic simulation. C A - free acid
concentration

The iteration was stopped when reaching an accuracy criterion, usually

< 0.01. Convergence is always guaranteed. If initially pH but not Cz is

known, C z can be estimated from Eq. (14) with 5 = 0.

In addition to the mass balance equations for total weak acids (acetic,
propionic, butyric, carbonic . . . . ) and weak bases (ammonium), a balance for
cations of strong bases (K +, Na + . . . . ) and anions of strong acids (CI-,
S O 2 - , . . . ) is necessary as follows:

vdCz
L dt = FL (CZ~ - C Z ) + Ftitr C Z .... (15)

Figure 3 gives an information flow diagram of how the solution of the dynamic
mass balances depends on the reaction kinetics. The kinetics in turn depend on
the undissociated acids, which are given by the total acids and the pH, according
to the equation. The charge balance represents an algebraic loop in the
otherwise sequential dynamic integration.

8 Modeling Biofilm Kinetics

In those reaction systems where biofilms or flocs are formed (UASB, fluidized
bed, fixed bed) diffusion and direct inter-species transfer of intermediates (e.g.
hydrogen) may be important. The detailed biofilm structure and formation is
Modeling and Control for Anaerobic Wastewater Treatment 95

very complex which can easily be seen in electron microscopic pictures. It can be
assumed that the biofilm structure is also dynamic but with longer time
constants than the hydraulic retention time. Quantitative data on biofilm
structure and inter-species hydrogen transfer are practically not existing [36].
It is also difficult to determine a representative diffusion coefficient. Examples
of biofilm modeling in the literature are: Denac et all [17]; Bryers [8]; Wang
et al. [65].

9 Gas Phase Modeling

The transport of gaseous components (methane, hydrogen and carbon dioxide)

from the liquid phase to the gas phase can be modeled as a well-mixed gas phase
with mass transfer driven by the difference between solubility and liquid phase
concentration [52]. The flow rate leaving the reactor was calculated from the
bubble velocity and gas volume fraction considerations. Thus,
Accumulation = Transport - Gas flow

XiUBA~gp
dni - VLkLa(XiHip - cLi) (16)
dt RT
In addition the time delay of the gaseous components due to the head space was
modeled with a total gas balance. The gas liquid mass transfer coefficient, kLa,
was modeled as a function of fractional gas hold-up, which was a function of the
gas production rate. Mather [33] used a similar concept to model the gas phase
of an anaerobic reactor.

10 Comparison of Simulations with Experiment

It is very difficult to set up a model which qualitatively and quantitatively agrees

with experimental data. Therefore, comparisons of experimental data with
simulation results are rather scarce.
The model of Smith and McCarty [57] predicted the cyclic methane
production that occurred after shock loading, due to the inhibition of hydrogen
on the propionate conversion to acetate and due to the biomass population
dynamics in the continuous tank reactor with suspended culture. This work has
shown that the present knowledge of the detailed kinetics is still not adequate to
establish a complete model. Tschui [62] set up a complex model of anaerobic
sludge stabilization and compared simulated results with substrate pulse ex-
periments. The qualitative agreement of all parameters was reasonable.
96 E. Heinzle, I. J. Dunn and G. B. Ryhiner

The model of Ryhiner [52] was used to compare simulated results with
experimental data from whey treatment. Steady state and dynamic comparison
was made as given below.

10.1 Steady State Comparison of Experimental

and Simulated Carbon Balances
From steady state simulations and experiments, carbon fluxes were determined,
and carbon balances were estimated. The amount of carbon in the feed was
taken as reference (= 100%). A comparison of these values for the single stage
system gave good agreement between experiment and simulation. The devi-
ations for the total carbon balance were below 5%. Also other values (methane
and CO 2 in the gas phase, dissolved CO2, carboxylic acids) agreed very well.
The total carbon balances for the two stage system also agreed well but showed
larger deviations for individual compounds.

10.2 Comparison of Dynamic Experimental

and Simulation Results
Dynamic responses to step changes in feed concentration (+ 42%) were investi-
gated experimentally and by simulation. Examples from the single stage reactor
are shown for pH response in Figs. 4 and 5 and the response in the gas
composition in Figs. 6 and 7. The response in dissolved hydrogen concentration
(Fig. 8) was only available from simulation.

6.2 .8

pH
Gas Production
6.1
._~
-6 E~
6.0

"1- -4 n"
O- 5.9
20
5.8
Y
"O
o

2 IX.

5.7-

5.6 9 . 9 ~ 9 9 9 ~ 9 9 9 ~ 9 . . ~ .

1 2 3 4

Time [hi

Fig. 4. Response of pH and gas production rate on concentration step change (1.4 kg m -3
--* 2.0 kg m 3 whey powder) in single stage reactor (experiment)
Modeling and Control for Anaerobic Wastewater Treatment 97

6.2
pH
6.1, Gas Production
i

6 .c_
E
6.0

"-r
5.9 4 rr
g
5.8
StepC h ! n ~ o
2 13.
5.7 69
(.g

5.6 ' 9 ' i i i i

0 1 2 3 4

Time [h]

Fig. 5. Response of pH and gas production rate on feed concentration step change (1.4 kg m -3
2.0 kg m- 3 whey powder) in single stage reactor (simulation)

68 4O

~176
64
36 o'-s
.. 62-
34
O
60 84

32
58
StepChange ~ CH4
CO~
56 9 9 , , , , . 9 9 , 9 9 . 30
O 1 2 3 4 5

Time [h]

Fig. 6. Response of gas composition on feed concentration step change (1.4kg m 3

-* 2.0 kg m- 3 whey powder) in single stage reactor (experiment)

11 On-line Measurement and Observation Methods

The off-line a n d on-line m e a s u r a b l e variables for a n a e r o b i c m e t h a n e - p r o d u c i n g

processes have been extensively reviewed [60]. The variables considered for on-
line a p p l i c a t i o n were the following: pH, organic acids, alkalinity, redox, poten-
98 E. Heinzle, I. J. Dunn and G. B, Ryhiner

68 9 40

66
38

64

-2.'
36
o "6
62
-r-
0 -34 0
60 C)

. 32
58 .l CH4
Step Change ~ C(~

5 6 9 . 9 , . 9 . , 9 9 . , 9 9 , 9 . . 3 0

0 1 2 3 4

Time [h]

Fig. 7. Response of gas composition to a feed concentration step change increase (1.4 kg m -3
--, 2.0 kg m - 3 whey powder) in single stage reactor (simulated)

9e-5

8e-5

E
"6 7e-5
E
-'r-

68-5

t
Step Change
5 e - 5 '

1 2 3 4

Time [h]

Fig.& Response of dissolve hydrogen to a feed concentration step change increase (1.4kg m -3
2.0 kg m- 3 whey powder) in single stage reactor (simulated)

tial, sulfide ion activity, gas production rate, gas composition, hydrogen gas, and
carbon monoxide gas. Bicarbonate alkalinity is determined by acid titration to
defined p H between (4.0 _< pH _< 5.75), as discussed in detail by Powell and
Archer [44] and Di Pinto et al. [43]. Switzenbaum et al. [60] felt that carbon
monoxide and hydrogen gas responded sufficiently fast to overloads to make
them useful for control. A later report by the same authors [27] described their
experimental work on carbon dioxide and hydrogen responses and concluded
that CO gave information on the acetate consumption, while Hz was involved
mainly in the CO2 reduction step. Pauss et al. [40] measured continuously the
Modelingand Controlfor AnaerobicWastewaterTreatment 99

dissolved hydrogen in the liquid phase using a commercial hydrogen/air fuel

cell-based membrane probe. This probe was highly sensitive to hydrogen but
exhibited interference with HzS in high sulfur containing waste streams. It was
further shown that dissolved hydrogen concentration deviated from gas phase
equilibrium by a factor of about 50. The gas phase hydrogen sensor used by
Tschui [62] required absorption of HaS to avoid poisoning but was useful to
follow substrate pulses in mesophilic anaerobic sludge stabilization.
Denac et al. [17] utilized the NaOH consumption rate of the pH control
system to devise a control scheme which manipulated the influencing feed rate to
the reactor. Since the algorithm required turning off the feed during overloads,
equalization tank capacity would be required. In addition this principle was
used to maintain a constant range of effluent organic acids concentration.
Sch/irbfischer and Wandrey [55] listed on-line measurement methods for
anaerobic digestion including experimental problems and benefits of each
method. They also showed the application of a Kalman filter to estimate non-
measurable variables.

12 Control of Anaerobic Processes

Control of anaerobic processes may be necessary because of possible detri-

mental disturbances of the process by overloading or toxification. Applied
control algorithms were mostly of simple conventional PID type and there are
only few experimental results in the literature.
Several variables were manipulated by the controllers applied. Most control-
lers acted on the feed flow. This, however, requires a buffer tank to allow
intermediate storage of the wastewater. Another control strategy uses addition
of a base (e.g. CaCO3) to keep pH at a desired value. This creates significant
operational costs. Another suggested method of pH control uses separation of
CO 2 from the biogas and recycling of methane to strip CO2 from the reactor.
No experimental results, however, have been reported. CO 2 separation also may
be quite costly. Simple recycling of the product biogas without CO2 removal
gives only a very limited possibility of increased rate of stripping CO 2 by
increasing kLa. A forth strategy uses controlled sludge recycling. This may be
used in contact processes, where surplus sludge is stored in a special tank for
later reuse for control.
Various variables were used as measured process output variables. In
principle any variable giving information on intermediates like acids, hydrogen,
CO2, etc. can be used as measured variables, pH measurement is quite simple,
and it can be kept at a desired set point value by direct manipulation of one of
the above mentioned variables. Bicarbonate alkalinity is determined by titration
to a pH value between pK of organic acids and carbonic acid. It was shown to
give a fast and sensitive response which can be used for control of anaerobic
processes [50].
100 E. Heinzle,I. J. Dunn and G. B. Ryhiner

Partial pressure of C O 2 (Pco2) in the off-gas was reported to be a good

stability indicator but a bad control variable [50]. The same was found for PcH4
[52]. Hydrogen partial pressure in the off-gas requires very sensitive sensors but
provides fast information on the process condition [71]. It was claimed that
there is no long-term response [38]. Dissolved hydrogen concentration was
measured by mass spectrometer and made control of anaerobic processes
possible [66, 68, 31]. Volatile fatty acid accumulation is a clear indication of
overloading or toxification and allows control of an anaerobic bioreactor. The
measurements are, however, expensive and relatively complex involving on-line
GC measurement of the liquid phase [46, 52, 72]. Other controllers use model
based estimation of controlled variables. This includes the application of a
Kalman-filter as described by Schiirbiischer and Wandrey [55]. Table 4 lists
recent simple-type control studies. Of these 15 studies, 9 were experimental (4
pilot or industrial scale, 5 laboratory scale) and 6 involved simulations. One
study involved both simulations and lab-scale experiments.

12.1 Simple Controller Design using Simulations

of a Complex Process Model
As experiments with anaerobic systems are very time consuming, it is desirable
to design controllers using a complex process model. An example of such a study
[52, 53] will be described in more detail below. The general procedure is
sketched in Fig. 9. The control was implemented as PID-type feed back control
using a discrete digital control algorithm in the position form. The sampling
time was 15 min which allowed for complete analysis of gas (CO2 and CH4) and
liquid phases (acetic, propionic and butyric acid). The simulation model
permitted the determination of the suitable sampling time and controller
settings by the Ziegler-Nichols method.
Since setting the controller parameters required additionally a certain
amount of trial and error, the simulation model was valuable in determining an
approximate setting. The integral and differential settings were identical in
experiment and simulation, whereas the proportional constants needed re-
adjustment.

12.2 Proportional-Integral-Differential Control of pH

The control of pH by manipulating the feed rate using PID digital control was
investigated to counteract a step change in the feed concentration in single and
two stage fluidized bed reactors. For the single stage system the feed rate was
increased by a factor of two and for the two stage system by a factor of three. The
differential part was found to be ineffectual. Comparisons of the experimental
and simulated PI control responses for step changes in the flow rate are given in
Figs. 10 and 11 for the single stage. Here the controller settings were essentially
Modeling and Control for Anaerobic Wastewater Treatment 101

Table 4. Examples of simple control studies in the literature

Control Control Controller Advantages ( + ) System Ref.

variable action type disadvantages
(-)
pH Feed flow P ( - ) Holding Anaerobic [10]
tank contact
Recycle of P ( - ) Restriction process [10]
sludge from a of control Simulations
2nd stage action
COz [i]
scrubbing and
gas recycle

Base addition On-/Off [32]

[33]
Base addition On-/Off ( - ) robust Primary [14]
Feed flow PI ( - ) efficient sediment-
ation (solids) [58]
and digester
Simulations

Feed flow On-/Off Anaerobic [45]

filter
Industrial
scale

Feed flow PI ( + ) simple, fast One- and [52]

( - ) depending two-stage
on buffer anaerobic
capacity fluidized bed
reactors
Lab scale and
simulations

pH or Bi- Base On-/Off ( + ) BA faster Anaerobic [50]

carbonate addition than pH contact [51]
alkalinity process
(BA)
pH and Base addition On-/Off ( + ) fast Simulations
Pco2 response of
Pco2

Pco2 Base addition On-/Off ( - ) poor Anaerobic [51]

control variable contact
process
Simulations

PCH4 Feed flow PI ( - ) poor Anaerobic [52]

control variable fluidized
beds
Lab scale
Rate of Recycle of On-/Off ( + ) at presence Anaerobic [i]
methane sludge from a of toxics contact
production 2nd stage process
Simulations
102 E. Heinzle, I. J. Dunn and G. B. Ryhiner

Table 4. (continued)

Control Control Controller Advantages ( + ) System Ref.

variable action type disadvantages
(-)
H 2 (gas) (step ( + ) fast Anaerobic [38]
and pH response) response of filter
hydrogen Lab scale
( + ) long-term
response of pH

H 2 (gas) (step ( + ) simplicity Anaerobic [3]

response) of measurement contact
( - ) transfer digester
liquid-gas Pilot scale
H 2 (liquid) Feed flow On-/Off ( + ) fast sensitive Anaerobic [66]
( - ) mass digester
spectrometer Lab scale
required
Feed flow PI Simulations [52]
H2 (gas) (step ( - ) not Anaerobic [71]
response) sensitive filter
Industrial
scale
Volatile (step ( + ) simple Anaerobic [47]
acid to response) measurement digester
alkalinity (titration) Lab scale
ratio ( + ) no calibration
Bicarbonate Alkali ( + ) simple Anaerobic [43]
addition measurement digester
Alkalinity (titration) Lab scale
Volatile Feed flow Adaptive ( - ) knowledge Anaerobic [18]
acid controller of influent digester
concentration concentration Pilot scale [46]
and
stoichiometry
Feed flow Generic ( - ) knowledge Anaerobic [72]
model of influent digester
control concentration Simulations
and
stoichiometry

Feed flow PID ( + ) fast One- and [52]

( - ) on-line two-stage
measurement anaerobic
of VFA fluidized bed
reactors
Lab scale and
simulations

identical. For the two stage system control the controller gain in the experiment
was set at 40% of the value found suitable by simulation. In spite of this, some
instability was observed, possibly due to the higher buffering in the experiment
(twice the amount of dissolved CO2).
Modeling and Control for Anaerobic Wastewater Treatment 103

,omass . nr, s
Balances Law I

Gro~. ~ /~o.. I Trans,er

I Kinetics
~rx, ~ ,
Cc~ ' Rate
~ji I
Gas
I .,e,.s I I ,o!s
~rsi .a,anc~ I I "a'~nc~l
XCH4
~ C~lP~ Csi XC02

Con,ro,,er~ ~c.,
Fig. 9. Calculation flow diagram for the simulation of on-line control of the anaerobic process

6.0 10

Setpoint 8
5.9
-'T
.E
6 E
-r
Q_ 5.8 0~
4
o
5.7 E
2
pH
FlowRate
5,6 i
0 1
Time [hi

Fig. 10. PI-Control of pH after feed concentration step change (1.4 kg m 3 __,2.8 kg m -3 whey
powder), single stage reactor (experiment). ( K r = 0.7 m 6 mol- 1 h- 1, .q = 0.2 h)

12.3 PID Control of Organic Acid Concentrations

T h e o r g a n i c acid r e s p o n s e was r a p i d (here with 15 m i n s a m p l i n g time) b u t

requires r a t h e r expensive a n a l y t i c a l m e t h o d s . S h o w n here (Figs. 12 a n d 13) are
s i m u l a t e d a n d e x p e r i m e n t a l results from P I D c o n t r o l b a s e d o n acetic acid
m e a s u r e m e n t ; the c o n t r o l l e r c o n s t a n t s were s i m i l a r b u t n o t identical. T h e g a i n
104 E. Heinzle, I. J. D u n n and G. B. Ryhiner

6.0 10

Setpoint

J
'8

5.9,

.=_
.6 E
5.8.
Step C~hang!
I
I
'4 rr

F-
o

5.7. It

pH
Flow Rate
5.6 9 r i 0

1 2 3

Time [h]

Fig. 11. PI-Control of pH after feed concentration step increase (1.4 kg m=3---~ 2.8 kg m -3 whey
powder), single stage reactor, simulation. (Kr = 0.6 m 6 tool- 1 h - 1, Xl = 0.2 h, % = 0.05 h)

50 25

t r, --.-- Acet,0Ac,f0
= = -- -- _ _ -~

<O
15

(g
o 20. rr
o

10 ~-
10,

0 9 9 9 i 9 9 9 i 9 , 9 i , 9 9 i 9 9 9 i 9 9 9 i 9 9 9 5

0 1 2 3 4 5 6 7

Time [h]

Fig. 12. Acetic acid set-point, PID-control of the two stage system, simulated. (K r = 3 • 10 -4,
~ = 0.25, zD = 0.2). Step change concentration increase (1.4 kg m 3 ~ 4.2 kg m -3 whey powder)

was set somewhat high in the experiment and caused oscillations, Similar results
were obtained with propionic acid control.

12.4 Control of Dissolved Hydrogen Concentration

Hydrogen was not measured, but a simulation of hydrogen control with a PI
controller demonstrated that this was a sensitive control variable (Fig. 14).
Modeling and Control for Anaerobic Wastewater Treatment 105

40 - 35

/
9 Acetic Acid

30 ~ . - Flow Rate 9 30

-- c

20 20

"~ 0~
< 15
0 10'

-.m.~ S t e p C~Change u ~ 10 n~_

"

0'
5

-10 , 9 9 , 9 9 9 ~ 9 9 , ~ 9 9 9 , 9 9 9 ~ 9 9 9 t 9 9 9 0
0 1 2 3 4 5 6 7

Time [hi

Fig. 13. Acetic acid set-point, PID-control of the two stage system, experimental. (K r = 3 x 10 -4,
zl = 0.3, ~D = 0.05). Step change concentration increase (1.4 kg m - 3 ~ 4.2 kg m - 3 whey powder)

le-4 14

H 2

8e-5, ~ . Flow Rate 12

E
6e-5
JL_ _-==== .
Set Point
. . .
10
"-
i
c

r 4e-5 n" ~"

a: o ~o

2e-5
4

~ ] Feed Concentration I
0e+0 9 . 9 ~ 9 9 9 ~ 9 9 9 , 9 9 9 ~ 9 9 9 , 9 9 9 , 9 , 9 2
1 2 3 4 5 6

Time [h]

Fig. 14. Simulated response of the PI hydrogen controller to a square wave in feed concentration for
the single stage system. (K r = 2,5 m 6 mol- ~ h- 1 ~ = 0.2 h). Step change concentration increase
(1.4 kg m -3 ~ 2.8 kg m 3 whey powder)

12.5 Control Based on the Methane Content of the Gas

F o r h i g h r a t e a n a e r o b i c t r e a t m e n t , the l i q u i d p h a s e flow rate g r e a t l y influences

the a m o u n t of c a r b o n d i o x i d e in the gas phase. T h u s t h e difference in solubilities
of m e t h a n e a n d c a r b o n d i o x i d e m a k e s this t y p e of f e e d b a c k c o n t r o l difficult
 106 E. Heinzle, I. J. Dunn and G. B. Ryhiner

to implement. In principle a computer control system could correct for the

dissolved carbon dioxide in the effluent. Basing the control on the mass flow of
methane in the gas phase could be more successful.

13 Adaptive Control and Optimization of Anaerobic Processes

Optimal operation of anaerobic process should allow minimization of fixed,

variable and operating costs according to a defined performance index. Basically
two types of optimization method may be used. Steady-state methods estimate
the steady state gain and use a gradient method with steepest descent to
approach the variable optimum [49, 52, 53]. A general disadvantage of this
method is that it is relatively slow, but uses simple linear models as described
later for the case of whey treatment. Dynamic methods [42] or model-based
methods [24, 56] allow a faster response at the expense of a more complex
model setup. Only steady state methods have been applied for anaerobic
digestion processes as far as has been reported.
On-line measurements of the variable to be controlled permit taking control
action with simple conventional feed-back methods. However, the character-
istics of anaerobic treatment of wastewater continually change, and such simple
methods are therefore not always successful. Set-point control is also usually not
sufficient when it is of interest to maintain the process at an optimum which is a
function of two or more process variables. In such cases a model for the process
is necessary if the controller is to adapt to changing conditions. In adaptive
optimal control, a model is used for a changing process in order to seek a
maximum in a performance index function.
A recent book [6] reviews the principles of adaptive control methods and
their application to biological reactors. The literature contains relatively few
applications of optimization methods for anaerobic wastewater treatment sys-
tems, most systems are simulated, and experimental work is lacking (Table 5).
These systems are of particular interest because of the potential for widely
varying feed composition and biomass adaptation.
A four organism model Mosey-type model was used by Dochain et al. [19]
to develop an adaptive control algorithm for the control of hydrogen concentra-
tion, which was tested by simulation. In this control scheme the hydrogen level
was to be controlled by measuring the inflow glucose, the outflow hydrogen gas
and the liquid phase hydrogen concentration on-line and by manipulating the
inflow rate. A quasi-steady state assumption regarding the glucose concentra-
tion allowed relating the accumulation rate for hydrogen to the glucose inflow
rate and the hydrogen outflow gas rate. A further assumption neglected the
propionate reaction path. It was possible to prove mathematically the stability
of the proposed control system without knowing the kinetics or yields. For the
simulation both hydrogen and acetate were assumed to have inhibiting effects
Modeling and Control for Anaerobic Wastewater Treatment 107

Table 5. On-line optimization of anaerobic processes

Control Performance Optimization System Ref.
action index
Recycle ratio Cost function: fixed Simplex Two-stage [23]
costs (HRT) and search anaerobic
variable costs Off-line digester with
(organics and settler
substrate in the Simulations
effluent)
Feed flow Deviation of pH and Fibonacci Anaerobic [1o]
feed flow rate search digester
On-line Simulations
Base addition same plus base flow
(feed-
forward/
feed-
backward
control)
Feed flow Biogas production Closed-loop Chemostat [64]
(residence (space-time yield) On-line Pilot scale
time)
Feed flow Productivity Adaptive Bakers' yeast [49]
on-line fermentation
optimization, Lab scale
Steepest
descent
Feed flow Biogas production Closed-loop Chemostat 1-64]
(residence (space-time yield) On-Line Pilot scale
time)
Feed flow Gas production rate Empirical Stirred tank [33]
questing laboratory
routine reactors
Feed flow Methane production Adaptive One and two- [52]
minus VFA in the on-line stage
effluent optimization, anaerobic
Steepest fluidized bed
descent reactors
Lab scale and
simulations

on the methanogenic steps and the neglected propionate reaction was included.
Without control the hydrogen consuming methanogens are washed out of the
reactor. Numerous simulations tested the controller under a range of conditions,
including a comparison with simple P! control. This approach appears to be a
promising way of including some of the model complexities into an adaptive
controller design.
Detailed mechanistic models for adaptive control are generally too complex,
and would require simplification before implementation. However they have the
advantage of containing most information and may be necessary if not all
108 E. Heinzle, I. J. Dunn and G. B. Ryhiner

variables can be measured. When all the important process variables can be
measured, then empirical models of a simple, linear form can be used to
represent the relations between manipulated inputs and measured outputs [493.

yl(t) + al, i y~(t - 1) + a~,z yl(t - 2) + . . . . . + al, n yl(t - n)

=b,, i u ( t - l - d l ) + c 1 (17)
y2(t) + a2,1 yz(t - 1) + a2,2 Y2(t - 2) + . . . . . + a2,n Y2(t - n)
= b2,i u(t - 1 - de) q- C2 (18)
where d 1 and d 2 are integer indices which refer to the number of sampling times
that the output is delayed, due for example to the residence time of the process
or to the measurement delay. The nomenclature ( t - n) refers to the sample
taken n sample times earlier. The sample time is determined by the process
dynamics and the analytical instruments. The input u may be, for example, feed
flow rate; the outputs Yl and Y2 could be measured concentrations or uptake
rates.
Such empirical models have the advantage that very little process under-
standing is necessary, other than knowing what to measure and what to
manipulate. The theory behind the method allows the process steady state to be
estimated from the dynamic data, making it unnecessary to wait for the steady
state. The optimal inputs are found from this information, which is an important
advantage for slow biological processes.
The elements of an adaptive algorithm are summarized in Fig. 15.
A complex mechanistic simulation model as described earlier was used to
find the maximum of the performance index (PI), which was a compromise
between high methane production and low effluent acids concentration 1-52].
The feed flow rate was adjusted to maximize a desired performance index,
taken as a compromise between high methane production and low effluent

Feed
Changes
F ] Reactor ] Measured

~_~Linear I Data f
Process Model ~ Compare
[Updated
IPar~m.
I I Parameter l_
Estimation ]-

F ]Optimizer l- Objective
] ~ Function

Fig. 15. Continual parameter updating allows the use of a simple model. The optimizer calculates
the flow rate required to reach the maximum
Modeling and Control for Anaerobic Wastewater Treatment 109

concentration. In this application the aim was to maximize the rate of anaerobic
degradation without allowing the organic acid concentrations to become ex-
cessively high. Thus the performance index, PX, was formulated as follows:
Methane
Total 1
[Objective~ = production organic acids
- Constant x
L function J rate concentration

PI = Yl - 13 x Y2 (19)
where Yl = rCH4and Y2 = '~CAi' The shape of the performance index was found
by the use of the differential equation simulation model of the process. From this
the influence of feed rate on the performance index at steady state could be
determined. This function, as shown in Fig. 16 is dependent on feed concentra-
tion and feed rate.
Simulation results of the reactor with the estimation (Eqs. 17 and 18)
indicated that the simplest model was sufficient for obtaining good estimates of
Yl and Y2,
yl(t) = b1,1 u(t - 1 - d~) + cl (17)
Y2(t) = b2,1 u(t - I - d2) + c2 (18)
This was therefore used in the experiments. This simple model considers only the
most recent output data directly, and although much of the dynamic informa-
tion was not considered by this simplification, this was found to be of no
disadvantage. Influence of older data was included by variable forgetting factor
for parameter values.
Simple input-output models were used for the adaptive optimization of one
and two stage anaerobic fluidized bed reactors.

15
.1[ " yl : Methane rate [mmol h -1]

10

~= 0

-5 ~. PI = y l - 0.5 y2
9 PI = y l - l . 0 y2
-10 -------13----PI = yl-l.5 y2
o i ~
Feed rate [1h -1]

Fig. 16. The performance index depends on the penalty function constant for the organic acid
concentration, here for a single stage system, simulated (Feed concentration - 42 C-mol m 3)
1 10 E. Heinzle, I. J. D u n n and G. B. Ryhiner

The strategy for control was as follows:

1. Sampled data on organic acid and methane concentrations (GC) and gas
flows were used to manipulate the feed flow rate.
2. Control by adaptive on-line optimization was implemented using a linear,
single input (feed flow rate) - double output (methane rate and total organic
acids concentration) model. Thus as given in Eqs. (17) and (18), the feed flow
rate (u) was manipulated and caused changes in both the methane rate (Y0
and total organic acids concentration (Y2). The parameters of the model bi
and ci were continually updated to give a fit to the current time period. The
model was then used to calculate which feed flow changes would move the
process toward the desired optimum. If the process changed, say, due to a
change in substrate feed composition, then the outputs would move away
from the optimum. This would be sensed, incorporated into the model
parameters, and the feed flow rate would be changed to move towards the
new optimum.
The values for the controller parameters were found by simulation using the
mechanistic model; they compared favorably with the values found from
experiment, with the differences probably due to influence of measurement
noise. Simulated results for the single stage reactor are given in Figs. 17 and 18.
In the simulation the optimizer quickly responded to the feed concentration
increase disturbance keeping the performance index (PI) nearly constant. Feed
concentration decrease did not allow keeping the performance index constant.
Instead of that the flow rate was kept constant and only increased after

0.007
AC

0.006
NC
7.00e-4 _
0.005
6.00e-4
r
90.004
5.00e-4 _

4.00e4. NC 0.003

3.00e-4,

2.00e-4 : i i

0 20 40 60 80
T i m e [hi

Fig. 17. Simulated optimizer response of the single stage process to a step change (42 to 63
C-mol m - 3) in the feed concentration. AC - adaptive control, N C no control. (~ = 10- 6, 13 = 1.0,
Zo.1 = 10-v, •o,2 = 10 7, ~min = 0.1)
Modelingand Control for Anaerobic WastewaterTreatment 111

0.8

0,7

e-,
1 6
~D
0,6
5 .~=
0.5
E

0.4 l ~D
2
0.3
Feed concentration [ I
0.2
'2 4 ; 8 1'0 1'2
Time [h]

Fig. 18. Controlled response of the single stage process to a step change (42 to 63 C-tool m 3) in the
feed concentration (experiment). (~ = 2 x 10 -6, 13 = 1.0 Z o 1 = 10-6, Zo,2 = 10 -6, ~min : 0.1)

considerable delay. The results obtained using the adaptive controller are
compared to uncontrolled system response. A feed concentration increase by
75% (42 --, 73.5 C-mol m-3) was handled much better by the adaptive control-
ler. Comparison of response to concentration decrease (time = 40 h) is difficult
from Fig. 17 since the starting points were different. In the experiment (Fig. 18)
the controller reacted more quickly in both cases. The flow rate decrease by the
controller corresponded nearly to keeping constant the substrate loading rate of
the reactor. The experimental results for the two stage reactor were similar to the
simulated results but had a significant time delay and correspondingly larger
overshooting.
It is quite remarkable that the detailed mechanistic simulation model was so
useful for designing and testing, the adaptive controller. The controlled process
could be simulated for all controller-types, allowing the controller parameters to
he found. The performance index as defined was satisfactory for control and the
algorithm from the literature [49] was usable.

14 Concluding Remarks

Recent developments in modeling, simulation and control of anaerobic pro-

cesses have been reviewed. Despite the fact that anaerobic processes are very
complex multi-organism, multi-substrate processes, it seems to be possible to
describe the essential features of anaerobic degradation process dynamics. A
relatively complex model has been successfully used to describe dynamic
responses to changing feed conditions and to design simple feedback and
112 E. Heinzle, I. J. Dunn and G. B. Ryhiner

advanced optimizing controllers. Controller design using models is important in

anaerobic processes partly because of lengthy experimentation time.
Although much has been achieved, there is still much to do to understand
anaerobic processes in more detail. There is little information on the kinetics of
degradation of less common substrates such as aromatic compounds frequently
occurring in industrial waste streams. Microorganism interaction is only rudi-
mentarily understood and therefore crudely modeled. It is still not yet clear
which method is best for controlling the stable operation of anaerobic reactors.
It may be that the method may depend on the individual case. Simple, reliable
and robust measurements are required, and they will need more complex
control methods to be successful.

15 References

1. Andrews JF (1978) The development of a dynamic model and control strategies for the
anaerobic digestion process. In: Mathematical models in water pollution control. J Wiley,
Chichester, p 281
2. Archer DB (1983) Enzyme Microb Technol 5:570
3. Archer DB, Hilton MG, Adams P, Wiecko H (1986) Biotechnol Lett 8:570
4. Atkinson B, Mavituna F (1983) Biochemical engineering and biotechnology handbook. Nature
Press, New York
5. Barnes D, Fitzgerald PA (1987) Anaerobic wastewater treatment processes. In: Forster CF,
Wase DAJ (eds) Environmental biotechnology. Ellis Horwood, Chichester, p 57
6. Bastin G, Dochain D (1990) On-line estimation and adaptive control of bioreactors. Elsevier,
Amsterdam
7. Bellgardt KH (1991) personal communication
8. Bryers JD (1985) Biotechnol Bioeng 27:638
9. Colberg PJ (1988) Anaerobic microbial degradation of cellulose, lignin, oligoligninols, and
monoaromatic lignin derivatives. In: Zehnder AJB (ed) Biology of anaerobic microorganisms.
Wiley, New York, p 333
10. Collins AS, Gilliland BE (1974) J Environ Eng Div 100:487
11. Conrad R, Schink B, Phelps TJ (1986) FEMS Microbiology Ecology 38:353
12. Costello DJ, Greenfield PF, Lee PL (1991a) Water Res 25:847
13. Costello DJ, Greenfield PF, Lee PL (1991b) Water Res25:859
14. Dalla Torre A, Stephanopoulos G (1986) Biotechnol Bioeng 28:1106
15. Denac M, Dunn IJ (1988a) Biotechnol Bioeng 32:159
16. Denac M, Miguel A, Dunn IJ (1988b) Biotechnol Bioeng 31:1841
17. Denac M, Lee PL, Newell RB, Greenfield PF (1990) Water Res 24:583
18. Dochain D, Bastin G (1985) Environ Technol Lett 6:584
19. Dochain D, Perrier M, Pauss A (1991) Ind Eng Chem Res 30:129
20. Dolfing J (1988) Acetogenesis. In: Zehnder AJB (ed) Biology of anaerobic microorganisms.
Wiley, New York, p 417
21. Erickson LE (1988) Bioenergetics and yields for anaerobic digestion. In: Erickson LE and Fung
DY-C (eds) Handbook on anaerobic fermentations. Dekker, New York, p 325
22. Erickson LE, Fung DY-C (1988) Handbook on anaerobic fermentations, Dekker, New York
23. Fan T, Erickson LE, Baltes JC, Shah PS (1973) J Water Pollut Control Fed 45:591
24. Golden MP, Ydstie BE (1989) AIChE J 35:1157
25. Gujer W, Zehnder AJB (1983) Water Sci Technol 15:127
26. Harper SR, Pohland FG (1986) Biotechnol Bioeng 28:585
27. Hickey RF, Switzenbaum MS (1991) J Water Pollut Control Fed 63:129
Modeling and Control for Anaerobic Wastewater Treatment l 13

28. Kaspar HF (1977) Untersuchung zur Kopplung von Wasserstoff- und Methanbildung im
Faulschlamm. Diss ETH, Zfirich
29. Kaspar HF, Wuhrmann K (1978) Appl Environ Microbiol 36:1
30. Kissalita WS, Lo KV, Pinder KL (1989) Biotechnol Bioeng 33:623
31. Lloyd D, Whitmore TN (1988) Lett AppI Microbiol 6:179
32. MS~rklH, Mather M, Witty W (1983) Mess- und Regeltechnik bei der anaeroben Abwasserrein-
igung sowie bei Biogasprozessen. In: Rehm H-J, Reed G (eds) Biotechnology Vol. 2,
R. Oldenburg, M/inchen, p 369
33. Mather M (1986) Mathematische Modellierung der Methang~irung. In: Fortschrittsberichte-
VDI, Reihe 14: Landtechnik/Lebensmitteltechnik Vol. 28, VDI Verlag, Dfisseldorf
34. McCarty PL (1981) One hundred years of anaerobic treatment. In: Hughes ef al. (eds) Anaerobic
digestion. Elsevier, Amsterdam, p 1
35. McCarty PL, Smith DP (1986) Environ Sci Technol 20:1200
36. McCarty PL, Mosey FE (1991) Water Sci Technol 24:17
37. McInerney MJ (1988) Anaerobic hydrolysis and fermentation of fats and proteins. In: Zehnder
AJB (ed) Biology of anaerobic microorganisms. Wiley, New York, p 373
38. Moletta R (1989) Environ Technol Lett 10:173
39. Mosey FE (1983) Water Sci Technol 15:209
40. Pauss A,'Beauchemin C, Samson R, Guiot S (1990) Biotechnol Bioeng 35:492
41. Pavlostathis SG, Giraldo-Gomez E (1991) Water Sci Technot 24:1519
42. Petersen JN, Whyatt GA (1990) Biotechnol Bioeng 35:114
43. Di Pinto AC, Limoni N, Passino R, Rozzi A, Tomei MC (1990) Anaerobic process control by
automated bicarbonate monitoring. In: Briggs R (ed) Instrumentation, control and automation
of water and wastewater treatment and transport systems. Pergamon, Oxford, p 5l
44. Powell GE, Archer DB (1989) Biotechnol Bioeng 33:570
45. Da Pra E (1987) Rechnerunterstfitzte Optimierung des Abbaus yon Rfibenschwemmwasser
mittels eines Anaerobfilters Diss. Univ, Zfirich
46. Renard P, Dochain D, Bastin G, Naveau H, Nyns EJ (1988) Biotechnol Bioeng 31:287
47. Ripley LE, Boyle WC, Converse JC (1986) J Water Pollut Control Fed 58:406
48. Roels JA (1983) Energetics and kinetics in biotechnology. Elsevier Biomedical Press, Amsterdam
49. Rolf M J, Lim HC (1984) Chem Eng Commun 29:1236
50. Rozzi A, Eng M (1984) Trans Inst M C 6:153
51. Rozzi A, Di Pinto AC, Brunetti A (1985) Environ Technol Lett 6:594
52. Ryhiner G (1990) Regelung und adaptive On-Line Optimierung von ein- und zweistufigen
anaeroben Fliessbett-Reaktoren. Diss ETH, Zfirich
53. Ryhiner G, Dunn IJ, Heinzle E, Rohani S (1992) J Biotechnol 22:89
54. B61aich JP, Bruschi M, Garcia JL (1990) Microbiology and biochemistry of strict anaerobics
involved in interspecies hydrogen transfer. FEMS Symp. No. 54. Plenum New York
55. Schfirb~scher D, Wandrey Ch (1991) Anaerobic waste water process models. In: Rehm HJ, Reed
G (eds) Biotechnology Vol. 4, VCH, Weinheim, p 441
56. Shi Z, Shimizu K, Watanabe N, Kobayashi T (1989) Biotechnol Bioeng 33:999
57. Smith DP, McCarty PL (1990) J Water Pollut Control Fed 62:175
58. Stephanopoulos G (1988) Modelling mixed culture interactions in anaerobic digestion. In:
Erickson LE and Fung DY-C (eds) Handbook on anaerobic fermentations. Dekker, New York,
p 597
59. Stouthamer AH (1988) Bioenergetics and yields with electron acceptors other than oxygen. In:
Erickson LE and Fung DY-C (eds) Handbook on anaerobic fermentations. Dekker, New York,
p 345
60. Switzenbaum MS, Giraldo-Gomez E, Hickey RF (1990) Enzyme Microb Tecbnol 12:722
61. Thauer, Jungermann, Decker (1977) Bacteriol Rev 41:147
62. Tschui M (1989) Dynamisches Verhalten der mesophilen anaeroben Schlammstabilisierung.
Dissertation. ETH, Zfirich
63. Vogels GD, Keltjens JT, Van der Drift C (1988) Biochemistry of methane production. In:
Zehnder AJB (ed) Biology of anaerobic microorganisms. Wiley, New York, p 707
64. Wandrey C, Aivasidis A, Bastin KH (1986) Ann NY Acad Sci 469:421
65. Wang YT, Suidam MT, Rittman BE (1986) J Environm Eng 112:975
66. Whitmore TN, Lloyd D (1986) Biotechnol Lett 8:203
67. Whitmore TN, Lazzari M, Lloyd D (1985) Biotechnol Lett 7:203
114 E. Heinzle et al.: Modeling and Control for Anaerobic Wastewater Treatment

68. Whitmore TN, Jones G, Lazzari M, Lloyd D (1987) Methanogenesis in mesophilic and
thermophilic anaerobic digesters: Monitoring and control based on dissolved hydrogen. In:
Heinzle E, Reuss M (eds) Mass spectrometry in biotechnological process analysis and control.
Plenum, New York, p 143
69. Wiesmann U (1988) Chem Ing Tech 60:464
70. Zehnder AJB (1988) Biology of anaerobic microorganisms. Wiley, New York
71. Ricker NL, Slater WR, Merigh M (1988) Process monitoring and control strategies for
anaerobic wastewater treatment. DECHEMA Biotechnol Conf Vol. 2, p 413
72. Costello DJ, Lee PL, Greenfield PF (1989) Bioprocess Eng 4:119
Design of Reactors for Plant Cells and Organs

Pauline M. Doran
Department of Biotechnology, University of New South Wales,
P . O . B o x 1, K e n s i n g t o n N S W 2033, A u s t r a l i a

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2 Suspended Plant Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
2.1 Properties of Plant Cultures Relevant to Reactor Design . . . . . . . . . . . . . . . . . . . 118
2.1.1 Rheology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
2.1.2 Aggregate F o r m a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
2.1.3 Shear Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2.2 Stirred Versus Airlift Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2.3 Engineering Analysis of Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
2.3.1 Mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
2.3.2 Mass Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
2.3.3 Shear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
2.4 Calculation of Time C o n s t a n t s in Plant-Cell Reactors . . . . . . . . . . . . . . . . . . . . . . 137
2.4.1 Oxygen C o n s u m p t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
2.4.2 Mass Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
2.4.3 Mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
2.4.4 Shear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
2.5 C o m p a r i s o n of Reactor Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
2.6 I m p r o v e m e n t of Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
2.6.1 Airlift Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
2.6.2 Stirred T a n k s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
2.7 Alternative Reactor Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3 Immobilised Plant Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
3.1 M a s s Transfer Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
3.1.1 External Mass-Transfer Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
3.1.2 Internal Mass-Transfer Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3.2 Self-Immobilised Plant Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.3 Immobilised-Cell Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.4 Effect of I m m o b i l i s a t i o n on Plant-Cell Metabolism . . . . . . . . . . . . . . . . . . . . . . . . 153
4 Differentiated Plant Tissue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.1 H a i r y Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.2 E m b r y o s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.3 Shoots, Buds and Plantlets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5 P h o t o a u t o t r o p h i c Plant Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6 Conclusions and F u t u r e O u t l o o k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Advances in Biochemical Engineering

Biotechnology, VoI.48
Managing Editor: A. Fiechter
9 Springer-Verlag Berlin Heidelberg 1993
116 P.M. Doran

To be economically feasible, production of most secondary metabolites by plant-cell

culture requires high cell densities. In standard reactor configurations such as stirred-tank
and airlift vessels, large-scale culture of plant cells encounters problems of shear damage,
poor mixing, and inadequate mass transfer of oxygen. Immobilised cells and organised
tissue such as embryos, roots and shoots often give enhanced levels of production;
however, design of appropriate reactors is at an early stage. This paper identifies the
rate-limiting processes in plant-cell culture, and discusses ways for improving reactor
performance.
Design of Reactors for Plant Cells and Organs 117

1 Introduction
Cultured plant cells are capable of producing a range of valuable secondary
metabolites. Research into large-scale plant-tissue culture is aimed mainly at
commercial manufacture of these compounds, although reactor culture of orga-
nised tissue can also be used for plant propagation. The shikonin process developed
by Mitsui Petrochemical Industry Ltd in Japan demonstrated in 1983 that
industrial-scale culture of dispersed plant-cells is technologically feasible. A major
effort is still being made to define the requirements of suspended plant-cell cultures
and their responses to different reactor conditions. This has proven to be a difficult
and challenging area of engineering research. Familiar problems of reactor
hydrodynamics, mixing, and oxygen transfer are coupled with plant-cell shear-
sensitivity and aggregation, formation of non-Newtonian suspensions at relatively
low cell densities, and poor definition of conditions required for secondary
synthesis. Although reactors up to 20000 litres have been used to culture plant
cells, the task of designing an appropriate reactor system for secondary-metabolite
production has not been completed. Yields of most secondary compounds in vitro
are still very low and decline even further after scale-up. In most cases, economics
demands that high cell densities be used to raise volumetric productivity; a reactor
capable of supporting 100 kg m - a dry weight suspended plant-cells without mixing,
mass transfer or shear-damage problems is yet to be found. While immobilised
cells and differentiated plant tissue are rapidly replacing suspended plant-cells as
the focus of research, recent progress with enzyme regulation and genetic
manipulation of plant cells may eventually improve the commercial prospects of
large-scale suspended-cell culture.
Interest in immobilising plant cells followed from the theory that secondary
metabolites are more likely to be produced by slow-growing cells with high cell-cell
contact than by dispersed suspended cells. Early work showed that product yield
and excretion could be greatly improved by immobilisation, with the added benefit
of reduced shear damage to cells. This promoted several studies of immobilisation
methods and bioreactors suitable for large-scale operations. The natural tendency
of plant cells to clump together is being exploited; self-immobilised or aggregated
plant-cells can be grown in packed- and fluidised bed reactors. Surface-attachment
of plant cells in membrane and biofilm reactors is also being investigated. Provision
of adequate oxygen transfer and maintenance of aseptic conditions during reactor
operation are two important factors in immobilised plant-cell reactor design.
Once it was demonstrated that hairy-root cultures produce high levels of
secondary metabolites there was a need for appropriate reactors. Work in this
new area of bioreactor design is progressing rapidly. A major problem with
large-scale culture of roots is poor distribution of biomass. Novel systems are now
being developed to fix the roots in place so the vessel can be completely filled
with tissue. In addition, spray reactors are being used instead of submerged culture;
this mode of culture promotes oxygen transfer and provides low-shear conditions.
Another innovation is use of bioreactors to grow somatic embryos; this technology
could be applied for mass propagation of plants and production of artificial
seed as well as for large-scale phytochemical synthesis. Differentiated shoots and
l 18 P.M. Doran

plantlets have recently been cultivated in reactors. New ideas for providing light
to photoautotrophic plant cells in large-scale vessels are also now appearing.
The aim of this review is to identify problems associated with large-scale culture
of various forms of plant tissue and to summarise the most recent developments
in reactor technology. Emphasis is given to the engineering aspects of plant culture.
More general treatment of applied plant-tissue culture can be found in previous
articles by Fowler [1], Misawa [2], Kurz and Constabel [3], and Rhodes et al. [4].

2 Suspended Plant Cells

2.1 Properties of Plant Cultures Relevant to Reactor Design

Most of the technological challenges associated with large-scale culture of
suspended plant cells can be traced to certain biological characteristics. High-
density plant-cell slurries are difficult to mix and keep in suspension, resulting in
poor distribution of oxygen and nutrients in reactors. Non-Newtonian rheology,
tendency to aggregate and shear sensitivity compound these problems. Details of
these parameters and their influence on reactor design are described in the following
sections.

2.1.1 Rheology
Plant culture rheology remains poorly characterised. The morphology of plant
cells in vitro is highly varied; near-spherical and elongated ceils co-exist in
suspension cultures, and there is usually a wide size-distribution. These properties
mean that high-density plant-cell suspensions become viscous, to the order of
several hundred centipoise. High viscosity leads to longer circulation times in
reactors, and formation of dead zones where the liquid remains stagnant. Reduced
turbulence in viscous suspensions has a pronounced effect on oxygen transfer as
air bubbles are not dispersed as easily as in non-viscous media.
Kato et al. [5] determined that Nicotiana tabacum cultures acted as non-
Newtonian fluids even at low cell concentrations of 0.9 kg m - 3. The culture filtrate
after removal of cells maintained a Newtonian character through the entire culture
period. As the cell concentration grew from 0.9 to 13 kg m-3 dry weight, filtrate
viscosity increased from 0.9 mPa s to only 2.2 mPa s even though apparent viscosity
of the whole cell suspension increased by a factor of 27.5. This small increase in
filtrate viscosity indicates that excretion of extracellular material by plant cells
does not contribute significantly to the overall viscosity; the viscous and non-
Newtonian character of the suspension is due mainly to the solids content.
Several models are available to describe non-Newtonian fluids; those applicable
to microbial cultures are discussed by Roels et al. [6]. Bingham and power-law
models have been used to describe plant-cell suspensions. The equations are:

Bingharnfluid: 'c = "co + K~ n (1)

Power-law fluid: "c = K'~n (2)
Design of Reactors for Plant Cells and Organs 119

where ~ is shear stress, % is the yield stress, K is the consistency index, j~ is shear
rate, and n is the flow behaviour index.
By analogy with Newton's law of viscosity, an apparent viscosity gapp can be
defined from Eq. (2) for power-law fluids:

g, pp = K? ~- ~ (3)

so that:

z = Gpp3' 9 (4)

Unlike Newtonian fluids, the apparent viscosity of plant-cell cultures depends on

shear intensity.
Measurement of empirical constants such as %, K and n is difficult for fluids
containing solids. Practical limitations associated with viscosity measurements are
discussed by Charles [7] and Metz et al. [8]. For suspensions, these include
destruction of pellets and flocs during analysis, settling of solids, and formation
of less dense layers at the walls of the measuring device. Some of these problems
can be alleviated by use of a turbine impeller for viscometry studies rather than
conventional instruments such as the rotating cylinder [6].
Wagner and Vogelmann [9], Kato et al. [5], Tanaka [10] and Scragg et al. [11]
have demonstrated that plant-cell cultures are pseudoplastic or shear-thinning at
shear rates up to 1000 s -1, so that n in Eqs. (1)-(3) is <1. Time-dependent
thixotropic behaviour has been reported [9, 11]; this is probably due to progressive
break-up of cell aggregates when shear levels are maintained. Non-Newtonian
behaviour involving a yield stress (t0) has been shown to occur in Morinda citrifolia
[9] and Catharanthus roseus [12] cultures; however, Scragg et al. [11] report that
evidence of a yield stress for Catharanthus roseus is dependent on the measurement
technique employed.
Shear thinning in plant-cell cultures means that apparent viscosity is lower in
areas of high shear, e.g. in the impeller region of stirred reactors. Air bubbles will
rise rapidly through these regions; away from the impeller dead zones will occur
more readily as apparent viscosity increases. If yield stress is a feature of plant-cell
theology, this may have implications for aeration. Small bubbles rising in the
culture liquid may not exert a sufficiently high stress on the surrounding fluid and
will remain fixed in the same fluid for long periods of time. Bubbles circulating
with the fluid rather than following the rising path of larger bubbles will become
depleted of oxygen, so that values for total gas hold-up in the reactor will give a
distorted picture of mass-transfer conditions.
In some cases viscosity models such as Eqs. (1) and (2) are valid only under
limited conditions; K and n values may vary with sufficiently large changes in
shear intensity. Because the rheological behaviour of plant-cell cultures is of critical
importance in reactor design, investigation of these parameters is required. The
limited information currently available is summarised in Table 1.
In studies conducted so far, the power-law index n for plant cells has remained
largely independent of aggregate size and cell concentration. The consistency index
120 P. M. Doran

Table 1. Rheological parameters for plant-cell suspensions

Species n % Reference
(Pa)

Nicotiana tabacum 0.69-0.74 - [5]

0.70 -- [10]
0.73 - [25]
Vinca rosea 0.53 - [10]
Cudrania tricuspidata 0.53 - [1t3]
Licorice 0.7 - [25]
Morinda citrifolia - 2 4 [9]
Catharanthus roseus - 300 [12]

K on the other hand is strongly dependent on cell concentration. Absolute values

of K have not been reported for plant cells; however, for biomass levels between
10 and 15 kg m -3 dry weight, Tanaka [10] showed that K values for several plant
species varied in proportion to cell concentration raised to a power between 6
and 7. This relationship is similar to that found for pelleted cells of the mould
P a e c i l o m y c e s varioti [13], although there is a general lack of consensus about the
relationship between biomass concentration and viscosity for mycelial cultures [8].
If K is so strongly dependent on plant-cell concentration, apparent viscosity of
plant-cell suspensions can increase tremendously over the course of batch culture.
Few researchers have reported operating viscosities in plant:cell reactors. Scragg
et al. [11] estimate the apparent viscosity of C a t h a r a n t h u s roseus cells at a concen-
tration of 250 k g m -3 fresh weight (which would correspond to 10-15 k g m -3
dry weight) and a shear rate of 609 s- 1 as approx. 3 m P a s. Significantly greater
apparent viscosities can be expected at higher cell densities and at the lower
shear-rates found in airlift and stirred reactors. Even so, maximum values would
be an order of magnitude less than those achieved in many mycelial cultures.

2.1.2 A g g r e g a t e F o r m a t i o n
Aggregation is generally a poorly-controlled parameter in plant-cell suspensions;
aggregates and flocs range in size from two cells to clumps of 0.5 cm diameter or
more. Wagner and Vogelmann [9] observed that plant-cell cultures forming flocs
or pellets caused fewer macromixing problems and less shear damage than
suspensions containing mostly single cells. Tendency towards clumping and wall
growth is somewhat species dependent; reducing the Ca 2+ level in the medium
[14, 15] and addition of sorbitol and cell-wall-degrading enzymes such as pectinase
and cellulase [16] have been reported to reduce aggregate formation. In bioreactors,
the size of aggregates depends also on the shear levels present [17]. Control of
clumping may be achieved to some extent by changing reactor geometry
and operating conditions.
The range of aggregate size and specific gravity in several plant cultures is
shown in Table 2. These data can be compared with the density of the suspending
liquid. Murashige and Skoog medium containing 3% sucrose has a specific gravity
of about 1.013; this reduces to about 1.000 at the end of batch culture [10].
Design of Reactors for Plant Cells and Organs 121

Table 2. Size and specific gravity of plant-cell aggregates in suspension culture. (Adapted
from Tanaka [10])

Species Aggregate size Aggregate specific

(gin) gravity

Nicotiana tabacum 150-800 1.005 1.015

Vinca rosea 44 500 1.002-1.005
Agrostemma githago 44 2000 1.015-1.026
Cudrania tricuspidata 44- 2000 1.018-1.023

2.1.3 Shear Sensitivity

Data on plant-cell shear sensitivity show a wide range of responses. Wagner and
Vogelmann [9] reported that Catharanthus roseus cells were completely destroyed
within 5 days at a turbine-impeller speed of 28 rpm; based on the correlation of
Metzner and Otto [18] this corresponds to an average shear rate of only 5 s -1.
In more recent studies, Hooker et al. [19] found extreme shear damage in Nicotiana
tabacum cultures in a 3-1itre reactor with a flat-blade impeller operated at 200 rpm.
Scragg et al. [20] tested the response of Catharanthus roseus to shear in a 3-1itre
stirred tank operated at 1000rpm. The average shear rate was 167s -1 corre-
sponding to a maximum shear rate of about 1500 s - ~ at the impeller tip. At the
end of 5 h, 30% of the cells had been destroyed; the remainder were viable and
survived subsequent culture [11].
As well as mechanical disruption and loss of viability, there are other more
subtle effects of shear stress on plant cells. A study of Catharanthus roseus showed
that hemicellulose, cellulose and pectin levels in the cell wall depended on the
level of hydrodynamic stress imposed during culture [17]. The mass ratio of cell
wall to whole cell also increased with intensity of shear. Plant cells are reported
to be more susceptible to shear during late exponential- and early stationary-
phases when the cells are of relatively large size and contain large vacuoles [9, 21].
Variation in resistance to shear between plant species and the ability of cultured
cells to develop shear tolerance raise questions about whether plant cells are always
shear sensitive. Scragg et al. [11] and Allan et al. [22] report that cells initially shear
sensitive developed shear tolerance after 2 - 5 years cultivation in vitro. Most
studies of shear sensitivity in plant-cell cultures have been carried out over short
periods of time relative to the duration of batch culture. However, long-term shear
responses for four plant species: Catharanthus roseus, Nicotiana tabacum, Cinchona
robusta and Tabernaemontana divaricata, have been tested recently over 10-14 d
in a 3-1itre reactor with a 6-blade Rushton impeller [23]. The stirrer was operated
at 250 and 1000 rpm. C. roseus and N. tabacum were shear resistant under these
conditions; both C. robusta and T. divaricata suffered growth impairment and
cell disruption above 250 rpm. Shear tolerance was correlated in this study with
stability of growth characteristics and time since initiation of culture.
Sensitivity to shear has been determined in most cases by varying the shear
rate, ~. In work by Midler and Finn [24] on disruption of protozoa, cell suspensions
122 P.M. Doran

of high apparent viscosity suffered more damage at constant shear rate than
suspensions with low apparent viscosity. Accordingly, shear stress rather than
shear rate appears to be the important variable in assessing shear sensitivity of
cells. When high local velocity-gradients are present, such as at the tip of the
impeller in stirred systems, the extent of cell damage may also depend on the
maximum rather than average shear stress. Maximum shear stresses withstood
by plant cells in laminar flow in a high-shear Couette rheometer have been reported
by Chen et al. [25]. For tobacco cells the critical shear level was 2.5 Pa and for
licorice cells 8 Pa. These figures indicate a significantly lower tolerance to shear
stress than that observed by Meijer [23] with plant cells subjected to stirrer speeds
of 250-1000 rpm.
The actual mechanism of shear damage to plant cells has not been studied.
However, hydrodynamic effects on animal cells have been analysed by several
groups [26-34]; the results may also be applicable to suspended plant-cells. A
complete picture of shear effects in stirred and aerated reactors has not yet been
determined unequivocally. In stirred microcarrier-cultures of animal cells where
there is no entrainment of gas bubbles, interactions between cells and turbulent
eddies are considered most likely to cause shear damage [26, 35]. If the size of the
eddies is approximately the size of the microcarriers, the eddies dissipate their
energy against the surface of the microcarriers and cause damage to the attached
cells. The dimension of the smallest eddies in turbulent flow is given by the
Kolmogorov scale:

n = (~) 1/4
(5)

where 1"1is eddy size, v is fluid kinematic viscosity, and a is local energy dissipation
rate per unit mass of liquid. In large-scale equipment with low-viscosity liquids,
rl is of the order 30-100 gm; smaller eddies are often produced in laboratory-scale
vessels. A relationship between cell damage in microcarrier cultures and T1 has
been demonstrated; decreasing the Kolmogorov scale below 1/2-2/3 the microcar-
rier diameter increases cell damage [26, 28]. Croughan et al. [28] also showed that
increased viscosity, which results in increased Kolmogorov scale, reduces damage.
Since plant cells have dimensions in the vicinity of 30-100 Ixm and can form
clumps of even greater size, it is likely that interaction with turbulent eddies also
causes shear damage in plant suspensions. To date, however, there has been no
experimental investigation of these effects.
In aerated systems, shear damage can occur at much lower impeller speeds than
those producing eddies of the dimensions described above. There is evidence that
in systems with bubble entrainment, significant shear damage occurs at the liquid
surface as a result of bubble disengagement [32]. Data of Handa-Corrigan et al.
[29] relating cell viability to bubble-column height and formation of a stable
protective foam support this theory. Investigations by Kunas and Papoutsakis
[30] have distinguished shear-damage mechanisms in reactors with and without
bubble entrainment. At low agitation speeds between 150 rpm and 600rpm,
damage to hybridoma cells was insignificant if there were no gas-liquid interface
Design of Reactors for Plant Cells and Organs 123

in the vessel. Damage at these stirrer speeds occurred only when bubbles were
allowed to burst at the liquid surface. At 800 rpm, which corresponded to a
Kolmogorov-eddy size comparable with the size of the cells, damage to the cells
increased but was still less significant in the absence of a gas interface.
Studies of shear damage in plant-cell suspensions have concentrated exclusively
on the effects of mechanical agitation. While different species can be expected to
have different shear sensitivities, the extremely wide variation in shear response
from complete cell destruction at 28 rpm [9] to tolerance of 1000 rpm [23] may
be due to intrusion of other factors not considered in these studies. The shear
stress produced by bursting bubbles may be one such factor.

2.2 Stirred Versus Airlift Reactors

Much of the literature on plant-cell reactors has focused on experimental

comparison of airlift and stirred-tank designs [9, 10, 36-38]. In early work, Wagner
and Vogelmann [9] measured cell growth and alkaloid production by Catharanthus
roseus and Morinda citrifolia in a selection of bioreactors including shake flasks,
airlift vessels and stirred tanks with different impeller designs. They concluded
that airlift reactors provided the best conditions; mechanically agitated devices
caused excessive shear damage. After considering the low oxygen demand of plant
cells and predicting that mass transfer requirements could be adequately supplied
by pneumatic agitation, Fowler, Scragg and co-workers conducted extensive
studies of airlift reactors up to 80-1itres capacity for culture of Catharanthus roseus
[39-42]. Maximum biomass concentration in these trials was about 23 kg m-3
dry weight; the reactors were usually operated below 12 kg m - 3.
Economic analysis of commercial plant-cell culture shows that high cell densities
(up to 100 kg m - 3 ) a r e required to compensate for the low production rates and
yields normally found with dedifferentiated tissue [43]. There are several reports
[36, 37, 44] that performance of airlift reactors becomes severely limited by poor
gas disengagement, impeded liquid circulation and development of unmixed zones
at biomass densities above 20-30 kg m - 3 dry weight. Increasing aeration rates
to improve mixing has not been successful; over-ventilation of cultures reduces
levels of dissolved carbon dioxide and other gaseous components such as ethylene,
so that growth and production declines [39, 40]. These apparent shortcomings
of airlift reactors have re-awakened interest in stirred reactors for plant tissue
culture. Tobacco cells have been grown successfully in a 20000-1itre stirred tank
[45], while Mitsui Petrochemical Industries Ltd. uses two stirred reactors of
200-1itres and 750-1itres capacity for commercial production of shikonin [46].
Stirred vessels tested by Mitsui for industrial production of berberine are able to
support Coptisjaponica cell densities of 7 0 k g m -3 dry weight [47]. In 1986 a
German company DIVERSA set up a cascade of five stirred tanks in series to
study the economic aspects of large-scale plant-cell culture for secondary metabolite
production [48]. Reassessment of stirred reactors has also been encouraged by
recent findings that cultivated plant cells are able to develop resistance to shear
[22, 49].
124 P.M. Doran

Because of their widespread application, airlift and stirred-tank reactors have

been the subject of extensive chemical-engineering analysis. Equations developed
for reactor design should be applicable to plant-cell cultures and provide
information about the effects of changing culture conditions on reactor perfor-
mance. Predictive equations for stirred tanks containing non-Newtonian fluids
were developed 15-20 years ago; hydrodynamics, mixing and mass transfer in
airlift reactors have been studied much more recently. The following section
summarises available design information for stirred and airlift vessels; several of
the equations presented are applied in a theoretical analysis of these devices for
large-scale culture of suspended plant cells.

2.3 Engineering Analysis of Reactors

Figure 1 shows standard configurations of airlift and stirred-tank reactors.

In stirred-tank reactors, mixing and dispersion of air is achieved by mechanical
agitation; this requires a relatively high energy input per unit volume which is
ultimately dissipated as heat. High mass-transfer rates can be achieved, although
for a given mass-transfer coefficient stirred-tank reactors impart higher levels of
shear than non-mechanically agitated vessels [36]. For long-term aseptic operation
good-quality sterile seals are required where the agitator shaft enters the vessel.
Temperature, pH, dissolved oxygen concentration and nutrient levels are easy to
control in stirred tanks. Stirred reactors are usually equipped with baffles to reduce

Off-Gas Air Off-Gas Off-Gas

t r
t- - ' l _ _
I I
cl riser
O0 0 O]

o [o -downcomer vdownoomer
oO
~176176
o
i o o o ~

-<
Air Air

Internal loop
External loop
with draft tube

STIRRED REACTOR AIR IFr REACTORS

Fig. 1. Stirred-tank and airlift reactor configurations

Design of Reactors for Plant Cells and Organs 125

vortexing; a wide variety of impeller sizes and shapes can be used to produce
different flow patterns inside the reactor. In tall vessels, installation of two or more
sets of impellers improves mixing.
In airlift reactors, air sparging is used for both mixing and aeration. The reactor
is divided into two distinct zones. Gas is sparged into the riser section; the increased
gas hold-up and decreased fluid density in this section cause the fluid to rise. Gas
disengages at the top of the reactor leaving heavier liquid which recirculates
through the downcomer. Low and relatively uniform shear fields, low power input
and extended aseptic operation are important advantages of airlift reactors for
plant-cell operations.
As shown in Fig. 1 there are two basic types of airlift reactor:
a) the internal-loop airlift in which riser and downcomer are separated by an
internal baffle or draft tube; and
b) the external- or outer-loop airlift where riser and downcomer are separate
tubes connected by short horizontal sections at top and bottom.
Both types of airlift usually have circular cross-sections. In external-loop reactors
the head-space region at the top of the vessel can be modified to change the
efficiency of gas-liquid separation; this determines whether bubbles are recirculated
or not and can affect gas hold-up, liquid velocity and turbulence. In internal-loop
vessels either the draft tube or annulus may be sparged. Changing the distance
between the lower edge of the baffle and the bottom of the reactor can affect the
pressure drop in this region with consequences for liquid recirculation velocity,
gas hold-up and mixing. Gas-liquid separation depends on the depth of submersion
of the upper edge of the draft tube in the liquid. In internal-loop vessels, since the
downcomer and riser connect directly at the top of the draft tube, gas-liquid
separation is generally not as effective as in external-loop devices and gas bubbles
are often drawn into the downcomer. The capabilities of internal- and external-loop
configurations are compared in Table 3 in terms of various performance indicators
[S0l.

Table 3. Relative performance of internal- and external-loop airlift reactors. (Adapted from
Chisti [50])

Parameter Reactor

Internal-loop ExternaMoop

Mass transfer coefficient higher lower

Overall gas hold-up higher lower
Gas hold-up in riser higher lower
Gas hold-up in downcomer higher lower
Superficial liquid velocity in riser lower higher
Circulation time higher lower
Liquid Reynolds number (shear) lower higher
Heat transfer probably lower probably higher
126 P.M. Doran

For plant-cell suspensions it is crucial that the reactor provides good mixing
and suspension of solids, adequate oxygen transfer, and non-damaging levels of
hydrodynamic shear.

2.3.1 Mixing
Good mixing achieves homogeneous distribution of oxygen and other nutrients
in the bulk fluid; dead zones where anoxia or nutrient starvation can occur are
prevented. Mixing also reduces localised build-up of substances such as sugar
which can affect cell metabolism if allowed to accumulate in unmixed regions.
Both heat and mass transfer are greatly influenced by mixing and the intensity of
turbulence in the reactor.
The efficiency of mixing is usually quantified in terms of mixing time. Mixing
time (tmx) is defined as the time required to reach some arbitrary level of uniformity
in the liquid being mixed. In a recirculating-flow mixed reactor, the number of
circulations required for complete mixing depends on the degree of turbulent
diffusion. For a stirred tank with several baffles and small impeller, a satisfactory
relationship between circulation time and mixing time is [51]:

t ~ = 4tc (6)

where tc is the time taken for liquid to travel one complete circulation loop in the
reactor. For internal-loop airlift reactors where gas is sparged into the annulus:

(Ad) ~
tmx = 3.5to \At,/ (7)

and for external-loop airlifts:

(A)OS
tmx = 5.2to \ A r J (8)

where Ad is the downcomer cross-sectional area, and A r is the riser cross-sectional

area [52].
For the same total power input per unit volume (P/VL) specific mixing times
(tm~/VL) for water in either internal- or external-loop airlifts are three to five times
longer than in stirred tanks [52]. The mechanisms of mixing in stirred tanks and
airlift reactors are different and are reviewed separately.

2.3.1.1 Stirred Tanks

A wide range of mixing equipment is available for mechanical agitation; classifica-
tion is usually based on liquid viscosity. Figure 2 shows the recommended viscosity
ranges for a number of impeller designs [53]. Some of these designs are illustrated
schematically in Fig. 3; the most widely studied impeller is the turbine disc with
6 flat blades pitched at 90 ~
Design of Reactors for Plant Cells and Organs 127

10 7

10 6

10 5

10 4

or
r~ 10 3

10 2

IMPELLER TYPES
Fig. 2. Viscosity ranges for different impeller designs. (Adapted from Holland and Chapman
[53])

Metzner and Otto [18] determined that average shear rate (gay) in a stirred tank
is directly proportional to agitation speed:

JZav = kNi (9)

where the value of k depends on impeller design and type of fluid, and N i is
rotational speed. This relationship has been tested for both Newtonian and
non-Newtonian fluids.
Flow visualisation studies with non-Newtonian fluids have shown that flow
patterns are different from those set up by the same equipment in Newtonian
liquids [54]. For pseudoplastic fluids, apparent viscosity in the impeller region is
rather low because of the higher shear rates; viscosity increases with distance from
the impeller. The recommended geometry for mixing operations is also different
with non-Newtonian fluids. For an ungassed low-viscosity fluid the optimum ratio
of tank diameter to impeller diameter (Dr:D0 is 2.5-3.0: 1. Reduction of this
ratio to 1.5 2.0:1 [55, 56] or use of a close-clearance anchor or helical impeller
gives better mixing with viscous non-Newtonian fluids. At low viscosities the
volume of fluid set in motion by the impeller is of the order of several cubic
impeller-diameters. In viscous fluids this volume decreases markedly. The thickness
of the fluid layer moved by the impeller may become only a fraction of the
128 P. M. Doran

ANCHOR PROPELLER 6-FLAT-BLADE TURBINE

PADDLE GATE ANCHOR HELICAL SCREW

Fig. 3. Common impeller designs for stirred tanks

impeller diameter, so that impellers of small diameter are not likely to provide
adequate mixing.
Mixing time in non-aerated Newtonian fluids in baffled cylindrical tanks has
been correlated with system geometry and operating parameters by Norwood and
Metzner [57]. Flat-blade turbine impellers were used. The dimensionless group,
fm:
tmx(NiD2)2/3 gl/6D~/2
fm = (10)
H1/21~3/2
L "L'T

was found to be approximately constant and equal to 6 above a Reynolds number

(Rei) of 1.8 x 105, where:

Di2NiQL
Rei - (11)

In Eqs. (10) and (11) tmx is mixing time; Ni is impeller rotational speed; Di is
impeller diameter, g is gravitational acceleration; HL is liquid height; Da- is tank
diameter; QL is fluid density; and gL is fluid viscosity.
Norwood and Metzner [57] proposed that the relationship between fm and Rei
for Newtonian liquids could be used to calculate approximate mixing times for
pseudoplastic fluids at high Reynolds numbers. For non-Newtonian fluids and
viscous cultures of Streptomyces niveus, Wang and Fewkes [58] found that mixing
times measured in stirred tanks were in good qualitative agreement with the
Norwood-Metzner correlation, although the value of fm at Reynolds numbers
Design of Reactors for Plant Cells and Organs 129

above 104 was approx. 20. The discrepancy may be due to aeration effects; aeration
of suspensions increases mixing times for a given Reynolds number [59]. In addition,
Wang and Fewkes calculated the non-Newtonian impeller Reynolds number using
the equation proposed by Calderbank and Moo-Young [60]:

Rel D2 n0,( , )n
- OAK ~ 2 (12)

where K is the consistency index and n is the flow behaviour index for a power-law
fluid. This change should not, however, affect the correlation at high Re i. Ap-
plication of the Norwood-Metzner correlation to non-Newtonian cell suspensions
is discussed further by Charles [7].

2.3.1.2 Airlift Reactors

Mixing time in airlift reactors depends on the liquid circulation velocity. Higher
circulation rates are obtained in external-loop devices, because gas entrainment
in the downcomer is greater in internal-loop vessels. Liquid velocities are also
dependent on geometric parameters. For internal loops, minimal mixing times
have been reported for 0.6 < DffDo < 1.0, where Dr is riser (draft-tube) diameter
and D e is total column diameter. Equal riser and downcomer cross-sectional area,
i.e. Dr/D c --- 1/~/2 = 0.71, is therefore a reasonable design goal [61]. An internal-
loop airlift reactor built to these specifications has been used to culture plant cells
up to 25 kg m - 3 dry weight without development of stagnant zones [37]. Mixing
times also vary considerably as distance between the top of the draft tube and
the liquid level is altered; Weiland [62] measured a three-fold increase in mixing
time as this distance was decreased from 0.5 m. Sparger positioning for optimal
gas-bubble distribution and liquid circulation has been investigated by Chisti and
Moo-Young [63]; Fig. 4 shows proper sparger positioning for both internal- and
external-loop devices. The cross-hatched areas in Fig. 4 indicate zones of the
reactor which, when filled in, improve liquid flow and prevent biomass settling.
Frictional loss at the bottom of internal-loop reactors depends on the area for
flow between the riser and downcomer; the effect on liquid velocity of varying
geometry in this region of the vessel has been analysed by Chisti et al. [64]. Mixing
in airlift reactors can sometimes be improved considerably by attention to
geometric details.
Liquid circulation in airlift contactors declines with increasing viscosity [65].
For plant cells, increasing viscosity during batch culture can produce substantial
changes in mixing conditions. Although several studies of circulation rates in airlift
reactors have been made at low viscosities there are a limited number of correlations
for non-Newtonian fluids. Popovic and Robinson [66] developed a relationship
between liquid velocity, gas flow rate, reactor geometry and fluid properties
for heterogeneous flow in external-loop airlifts at superficial gas velocities
>O.04m s-~:
( A d ~ 0"97 . -0.39
uLr = 0.23u~ 2 \ ~ j }a,app (13)
130 P.M. Doran

Riser 0

o
,o
00

'o
0 ] I ]~
Downcomer e 0
- Riser
o0.0

g.?,'
0r o~
,oOOoOo -- 9
Downeomer
' ~ ~ oo~
"~2 0000200,0

t
GAS b
t
GAS

0 0 0 0 ~~

J~
0 o o
00 0 ~ o O~
'o 0 i
oo o
0 0 0 O0 9
i 0 00 o .
e 0 O0 O0 *
0 00 0 0 4
o~176176

II

II
GAS
t
e GAS d

Fig. 4a-d. Influence of sparger location on gas distribution in airlift reactors. Poor
distribution of gas in a) internal- and b) external-loop reactors. Proper positioning of the
sparger is shown in e) and d). Cross-hatched areas indicate zones which should be filled in
to improve liquid flow and prevent biomass settling. (Adapted from Chisti and Moo-Young
[63])

where u ~ is superficial liquid velocity in the riser, m s-1; uGr is superficial gas
velocity in the riser, m s - 1; Ad is downcomer cross-sectional area, m2; A r is riser
cross-sectional area, m2; and gapp is apparent viscosity, Pa s. Equation (13) was
developed using viscous pseudoplastic suspensions of 0.5-1.5 wt% carboxymethyl
cellulose (CMC) in 0.5 M Na2SOr in external-loop devices with Aa/Ar ratios of
0.111, 0.250 and 0.444. The liquid height in these reactors was 1.88 m. The apparent
viscosity was estimated using the correlation for shear rate in bubble columns
reported by Nishikawa et al. [67]:

~, = 5000u~ (14)

where u~ is superficial gas velocity, m s- 1; and ~ is shear rate, s - 1. This relationship

was originally developed for bubble columns with C M C suspensions and
Design of Reactors for Plant Cells and Organs 131

u~ _> 0.04 m s -1. From Eq. (14) the value of •app in Eq. (13) for pseudoplastic
fluids can be taken as:

l,,lapp = K(5000u~)"- 1 (15)

where ~app is apparent viscosity, Pa s; uar is superficial gas velocity in the riser,
m s-1; n is the flow behaviour index; and K is the fluid consistency index, Pa sn.
When applied to Aspergillus niger suspensions containing up to 20 kg dry weight
per m 3 cells, Eq. (13) over-predicted ULr especially at higher Uar values [68].
Equation (13) was developed for a specified reactor height; however, since liquid
circulation relies on the difference in hydrostatic pressures between riser and
downcomer, liquid velocity also depends on the height of the reactor. Most reported
data on the effects of reactor height are derived from small equipment; these are
difficult to extrapolate because of the predominance of end effects. Rousseau and
Bu'Lock [61] determined for 0.68 < HL(m) _< 1.64 that:

t ~ oc HL~'7 (16)

where H E is liquid height in the reactor, m. This relationship was developed using
air-water in an internal-loop device with Dr/D c between 0.2 and 0.6.
Other equations for liquid velocity have been developed from theoretical analysis
of the hydrodynamics in external- and internal-loop airlift reactors [50]. An energy
balance for non-Newtonian liquids leads to the equation:

ULrAr(APvr + APFa) - ~LgHuULrAr(~;r-- ~;d)

+ 2
E KT
( l ~ g , ) z + KB \ A J
]
(1 -- aa)2 = 0
(17)

where uLr is superficial liquid velocity in the riser, m s- 1; APFr is frictional pressure
drop in the riser, Pa; APFa is frictional pressure drop in the downcomer, Pa; EL
is density of the liquid, kg m 3; g is gravitational acceleration, m s-2; HD is
gas-liquid dispersion height, m; ~r is gas hold-up in the riser; ea is gas hold-up in
the downcomer; K T is frictional loss coefficient for the top of the reactor; KB is
frictional loss coefficient for the bottom of the reactor; Ar is riser cross-sectional
area, m/; and Ad is downcomer cross-sectional area, m 2. To use Eq. (17), gas
hold-up in the riser er must be known or estimated from independent correlations,
APFr and APFd are determined from shear stress calculations, and KB and KT are
estimated from correlations involving the area for flow between riser and
downcomer. Solution for uLr requires trial-and-error procedures.
Once uLr is determined, the superficial liquid velocity in the downcomer ULd
can be estimated from continuity:

ULrAr = ULdAa. (18)

132 P.M. Doran

For internal-loop airlift reactors and for external-loop reactors where the lengths
of horizontal top and bottom sections are small, circulation time can be determined
from ULr, ULe and geometric parameters:

Hr Hd
tc = - - + -- (19)
ULr ULd

where H r is height of the riser and H a is height of the downcomer. Mixing time
can then be calculated from Eq. (7) or (8).

2.3.2 Mass Transfer

Of all the nutrients supplied to plant-cell cultures, oxygen is often the most difficult
to provide in non-limiting quantities. The critical dissolved-oxygen concentration
for plant cells has been reported as 1.3-1.6 x 10- 3 kg m - 3 [69, 70]; this is roughly
20% saturation under average culture conditions. Above this level growth is not
oxygen-limited and follows first-order kinetics. Below the critical concentration
growth is linear with respect to time [71, 72]. Oxygen supply also affects secondary
metabolite production [73, 74] and organogenesis [69]; these effects are more
difficult to generalise or predict.
The extent to which reactors provide favourable conditions for mass transfer
is reflected in the value of the overall liquid-phase mass transfer coefficient, kLaL:

O T R = kLaL(C* -- CL) (20)

where O T R is oxygen transfer rate, kg m - 3 s-1; kL is liquid-film mass-transfer

coefficient, m s - 1; aL is interracial area per unit volume of unaerated liquid, m - 1;
C* is equilibrium concentration of oxygen in the liquid, kg m - 3; and CL is actual
oxygen concentration in the liquid, kg m -3. Equation (20) is valid for local C*
and C L values; normally C* and CL are assumed constant throughout the reactor
so gas and liquid phases need to be well mixed. This is acceptable for most
low-viscosity liquids but m a y not be true in high-viscosity systems. In particular,
the assumption of well-mixed gas phase is difficult to justify in tall vessels with
height:diameter ratios ~> 1 [75].
Measured values for mass-transfer coefficients can be expressed as either kLaL
or kLaD, where aD is the gas-liquid interracial area per unit dispersed volume, m - 1.
These two mass-transfer coefficients are related by Eq. (21):

kLaD
kLaL -- (21)
1 -- eT

where aT is total gas hold-up.

For the same total power input per unit volume (P/VL), much higher kLa L
values are produced in pneumatically agitated reactors than in stirred tank reactors;
this applies also to non-Newtonian fluids [76]. As cell concentration and apparent
Design of Reactors for Plant Cells and Organs 133

viscosity increase, the rate of mass transfer can drop to as low as 15% of its initial
value [77]. Prediction and improvement of oxygen mass transfer in reactors relies
on the prediction and enhancement of kLaL.

2.3.2.1 StirredTanks
In aerated stirred reactors one of the functions of the impeller is to break up and
disperse the air bubbles. With pseudoplastic fluids, shear thinning near the impeller
causes gas to channel in toward the impeller rather than into the remainder of
the tank. This non-uniform distribution is difficult to prevent and has adverse
effects on both mass transfer and mixing. Less bubble break-up is observed with
viscous non-Newtonian fluids; kLae values tend to be considerably lower than in
Newtonian systems under the same operating conditions [78].
Correlations for kLaL values in stirred tanks are of two types. Correlations
based on power input are reviewed by Schiigerl [76] who also discusses effects of
gas flow rate, fluid properties and impeller speed on power requirements for
non-Newtonian fluids. Other kLae correlations are presented in terms of dimen-
sionless groups representing system geometry, fluid properties and operating
conditions.
Rates of oxygen desorption in pseudoplastic non-Newtonian fluids were
measured by Yagi and Yoshida [78] in an agitated tank with four baffles and
6-blade turbine impeller. The stirrer speeds tested were 5 s ~ to 10 s z; superficial
gas velocities ranged between 0.002 and 0.008 m s- 1. The correlation obtained for
kLaL in non-Newtonian fluids is:

kLaLD2--O'o6(D2NiQL)l"5(DN2)~176176176
\~app ,/ k,~i--J \ ~ L J
/ N D \o.32
x( i~] [1 + 2.0(ZN0~176 (22)
\uo/

where kLaL is the mass transfer coefficient, s-1; Di is impeller diameter, cm; ~L
is diffusivity in the liquid, cm 2 s - 1; Ni is rotational speed of the impeller, s- 1; ~L
is liquid density, g cm- 3; Papp is apparent viscosity, g cm- 1 s- 1; g is gravitational
acceleration, cm s-2; UG is superficial gas velocity, cm s-~; c~ is surface tension,
g s-Z; and )~ is characteristic material time, s. For non-Newtonian fluids % is a
constant defined as the reciprocal of the shear rate at which the ratio of apparent
viscosity to zero-shear viscosity is 0.67. The correlation of Eq. (22) is limited by
determination of )~; viscosity at zero-shear is difficult to measure accurately.
Apparent viscosity in Eq. (22) is calculated using the Metzner and Otto [18]
relationship for average shear rate as outlined in Eqs. (3) and (9).
An alternative to Eq. (22) is the correlation of Perez and Sandall [79]:

kLaLD~ - 212 11 ( apo (Diuo O 4 7( oo94

@L
(23)
134 P.M. Doran

where gg is gas viscosity, g cm- 1 s- 1; the other symbols and units of Eq. (24) are
the same as for Eq. (22). Note that Eq. (23) is a dimensional equation. Apparent
viscosity in Eq. (23) is determined by the method of Calderbank and Moo-Young
[60]:
K (3n + 1~ ~
~app -- (11Ni) 1 - n \ ~ 4 n - n // 9 (24)

Equation (23) was developed for carbon dioxide absorption in a 0.25% Carbopol
solution using a tank equipped with four baffles and a 6-flat-blade turbine impeller
operated at 2810 _< Re i _< 26700; Re i is defined by Eq. (11) with the viscosity
given by Eq. (24). kLaL values calculated using Eq. (23) are for carbon dioxide
and must be corrected for oxygen transfer. The correction based on the two-film
theory of mass transfer is [50]:

(kLae)o2 = (krae)co2 No~ (25)

~CO2

where N is diffusivity. The ratio ~o2/Nco2 is equal to approx. 1.2. The applicability
of Eq. (23) is discussed further by Yagi and Yoshida [78].

2.3.2.2 Airlift Reactors

Overall mass-transfer coefficients in external- and internal-loop airlift reactors
have been studied by a number of investigators. The majority of correlations for
kLaL and gas hold-up involve a simple power-law function of superficial gas
velocity:
kLaL oc ug. (26)

The fundamental hydrodynamic basis for this type of relationship is discussed by

Chisti and Moo-Young [80].
Bello et al. [81] found that in internal-loop airlifts gas hold-up in the downcomer
was 80-95% that in the riser, while in external-loop reactors downcomer gas
hold-up was much lower at 0-50% of riser hold-up. Consequently, the contribution
of the downcomer in external-loop devices to oxygen transfer is minimal. Thus,
while low downcomer gas hold-up in external-loop airlift reactors promotes liquid
circulation, this effect is accompanied by a reduction in oxygen-transfer capability.
Complete or near complete gas disengagement in the head-space of external-loop
reactors may be harmful to organisms which are sensitive to oxygen concentrations
below the critical level. Chisti [50] suggests that in this case the downcomer could
be sparged separately with either fresh or recirculated gas.
Gas hold-ups and kLaL values in airlift contactors decline with increasing
apparent viscosity [68, 82]. Presence of suspended solids also has a significant
effect on kLaL due to dampened turbulence, enhanced bubble coalescence and
interracial blanketing; addition of solids above l % (w/v) has been found to cause
an 80% reduction in gas hold-up and oxygen transfer [83].
Design of Reactors for Plant Cells and Organs 135

Popovic and Robinson [84] have proposed a correlation for kta ~ in external-loop
airlift reactors:

kLaD ~--- 0.5 X 10 -2"uGr0'52,'r~0.5_

1.032/2L
~L _(1 q-~rr//Ad~- 0'85 "p.app-0'89--0"75tO (27)

where kLao is the mass transfer coefficient based on dispersed volume, s-1; uG~
is superficial gas velocity in the riser, m s- 1; @t is the diffusivity of oxygen in the
liquid, m 2 s- 1; ~i. is liquid density, kg m - 3; Ad is downcomer cross-sectional area,
m2; A~ is riser cross-sectional area, m2; btappis apparent viscosity, Pa s; and cr is
interfacial tension with respect to air, N m - 1. Reactor geometry used to develop
Eq. (27) is the same as that described for Eq. (13); apparent viscosity is calculated
using Eq. (15). Popovic and Robinson [84] tested Eq. (27) with a range of viscous
and non-Newtonian fluids; prediction of kLao was within 13 % of measured values
in each case.
Values of kLaD from Eq. (27) can be converted to kLat using Eq. (21). Gas
hold-up in airlift reactors is not uniform; eT can be evaluated using Eq. (28):

8T = (AdSd + ArEr)/(A d + A~) (28)

where ~d is gas hold-up in the downcomer and ar is gas hold-up in the riser.
For external-loop vessels ad can be taken as approximately 0.3at; ar is given by:

Ad) - 1.o6. -o.:oa (29)

~r = 0-465u~ 1 + A-~] IX,pp 9

Equation (29) was developed by Popovic and Robinson [84] for viscous CMC
solutions, but has been shown by Allen and Robinson [68] to apply to AspergilIus
niger suspensions at 0.02 _< uar (m s- 1) < 0.2 and 0.01 < ~app (Pa s) < 0.5.
An alternative equation for kLaL in external-loop airlift reactors has been
proposed by Chisti et al. [83]. This correlation involves solids concentration
explicitly:

kLae ---- 1 + (0.349 -- 0.102Cs) u~ 37•176176 (30)

where kLaL is the mass transfer coefficient, s - : ; Ad is downcomer cross-sectional

area, m2; Ar is riser cross-sectional area, mZ; UOr is superficial gas velocity in the
riser, m s - : ; and C~ is solids concentration, % (w/v) or g dry weight per 100 ml.
Equation (30) was determined using Aa/Ar ratios of 0.25 and 0.44, and 0.026 _< uor
(m s- 1) < 0.21. The solids tested were Solka Floc fibres at concentrations between
1% and 3% in 0.15 M NaC1.

2.3.3 Shear
Mechanical damage in plant-cell suspensions depends on the intensity of shear
developed in the reactor. A limited number of correlations for hydrodynamic shear
in stirred and airlift reactors is available.
136 P.M. Doran

2.3.3.1 Stirred Tanks

Equation (9) is widely accepted as a means for calculating average shear rates in
stirred vessels. Variation of shear has been investigated by Metzner and Taylor
[54] and Oldshue [85]; maximum shear levels occurring at the impeller tip are
significantly higher than those given by Eq. (9).
Wichterle et al. [86] measured maximum shear rates on the front surface of a
Rushton 6-blade turbine in non-Newtonian pseudoplastic fluids. For Re i > 10
the following relationship was determined:

9 / N ? - nD.2,, ",,1/(1+n)
7ma, (1 + 5.3n)1/"~ ' ~ '~L) (31)
Ni

where n is the flow behaviour index for power-law fluids and K is the consistency
index.

2.3.3.2 Airlift Reactors

At the present time, reliable estimation of average shear levels in airlift reactors
is difficult. Correlations reported for pneumatically agitated columns vary conside-
rably, as summarised in Table 4. Of these relationships, the equation by Nishikawa
et al. [67] is used almost exclusively; it was developed by analysing heat-transfer
coefficients in Newtonian and non-Newtonian liquids in a bubble column at
0.04 _< u~ (m s -1) _< 0.1. The correlation of Henzler [87] was determined by
fitting literature data on mass transfer coefficients obtained with non-Newtonian
liquids. The broad lack of agreement between the shear-rate correlations of Table 4
causes difficulties in estimation of mass transfer coefficients and mixing times;
correlations such as Eqs. (13) and (27) depend on estimation of laavp which, for
non-Newtonian fluids, relies on knowledge of j,.
Constant but relatively low shear-stress is produced in air-driven reactors by
flow of fluid at the walls of the vessel. Tramper et al. [88] estimated the shear stress

Table 4. Effectiveshear rate correlations for pneumatically agi-

tated reactors. (Adapted from Schumpe and Deckwer [211])

Correlation Reactor type Reference

~, = 5000u6 bubble column [67]

4/= 1500u~ bubble column [87]
~/= 2800u~ bubble column [2ll]
~, oc u~/D c bubble column [82]
"i' = uB/Db bubble column [2121
j/= ( P ~ ) 1/~n+l) bubble column [213]
Design of Reactors for Plant Cells and Organs 137

associated with frictional wall losses in an airlift reactor using the friction-factor
equation:

"Cw =
1 2
~ ~LULkw
(32)

where Zw is shear stress at the wall, Pa; eL is liquid density, kg m - 3 ; uL is fluid

velocity, m s-1; and kw is a resistance coefficient. Maximum wall shear-stress was
assumed to occur where the direction of fluid flow changes abruptly. In external-
loop airlift vessels this happens at the top and bottom of the vessel in the
connections between riser and downcomer. In internal-loop reactors energy losses
at the top will usually be minimal relative to the bottom because the head-space
of internal-loop devices is like an open channel whereas the flow path at the
bottom is much more constricted. For maximum shear stress at the vessel wall,
the value for kw in Eq. (32) can be taken to be 1.3 as for a sharp bend.
Much higher levels of shear are associated with gas sparging in air-driven
bioreactors than with fluid flow. Shear stress produced in airlift reactors by injection
of air bubbles, rising of the bubbles through the liquid, and bubble bursting have
been analysed by several groups [32-34, 89]. From calculations [89] and supporting
experiments [32], it appears that bursting bubbles create the highest shear stresses
in bubble columns. Shear levels associated with bubble break-up have been
estimated at 104N m -a [89] and 2 0 0 - 3 0 0 N m -z [34]. Duration of these high
shear forces is of the order of milliseconds [34].

2.4 Calculation of Time Constants in Plant-Cell Reactors

Comparison of time constants for different reactor functions can be used to

determine which processes are rate-limiting. Time constant or characteristic time
is defined as the ratio of capacity to flow; a small time constant represents a fast
process.
In the following analysis, time constants are used to assess the suitability of
stirred-tank and airlift reactors for culture of plant-cell suspensions. The procedure
follows closely the technique of Kossen [90]. Typical geometries for the reactors
and other system parameters are assumed, as specified in Table 5. Calculations
are performed for an external-loop airlift reactor and for a baffled stirred-tank
operated with a flat-blade turbine-impeller and a constant air superficial velocity
of 0.05 m s- 1. The volume of both reactors is 10 m3.'
The time constants used in this comparison are those for oxygen conversion
(trx,), mass transfer (tmt), and mixing (tmx). Processes of heat transfer and heat
production are not considered since they are not expected to cause problems
during plant-cell culture. For stirred reactors the time constants are presented as
functions of the main operating variable, the stirrer speed N i. For airlift devices
the operating variable is the gas superficial velocity in the riser, U~r. Since the
effectiveness of mixing and mass transfer in plant-cell reactors changes considerably
as biomass levels increase, the analysis is performed using conditions for two
different times during batch or fed-batch culture. Near the beginning of the culture
138 P. M. Doran

Table 5. Parameter values for evaluation of time constants

Parameter Stirred tank External-loop

airlift reactor

C* solubility of oxygen in culture liquid 8 x 10- 3 8 x 10 - 3

(kg m- 3)
C~ solids concentration (%)
near beginning of growth phase 0.5 0.5
near end of growth phase 3 3
Da downcomer diameter (m) - 0.78
Di stirrer diameter (m) 0.78 -
HL diffusivity of oxygen in culture liquid 2.3 x 10 -9 2.3 x 10 .9
(m ~ s - ~)
Dr riser diameter (m) - 1.2
DT diameter of stirred tank (m) 2.3 -
g gravitational acceleration (m s-2) 9.8 9.8
Hc height of liquid in stirred tank (m) 2.4
Ha downcomer height (m) - 6
Hr riser height (m) - 6
ua superficial gas velocity (m s- l) 0.05 varies
VI~ volume of liquid in the vessel (m 3) 10 10
X cell concentration (kg dry weight m- 3)
near beginning of growth phase 5 5
near end of growth phase 30 30
~L density of culture liquid (kg m -3) 1015 1015
cy surface tension (kg s-2) 0.04 0.04

the .cell concentration is relatively low at 5 kg m -3 dry weight; the situation is

then re-assessed for conditions at the end of growth when 30 kg m -3 cells are
present.
The equations which will be used to calculate the time constants have been
presented in the preceding sections. As already noted, reactor-design equations
for non-Newtonian liquids are not as well defined as for low-viscosity Newtonian
fluids, and correlations for mass-transfer coefficients are generally limited due to
the sensitivity of this parameter to medium composition and method of measure-
ment. Application of equations other than those used below may give different
results; however, the overall picture does not change dramatically. None of the
available correlations for mixing and mass transfer were developed with plant-cell
cultures; virtually all bioreactor studies of the effect of solids and high viscosity
have used cellulose-fibre suspensions or homogeneous solutions such as carboxy-
methyl cellulose (CMC) to simulate cell slurries. The deficiencies of homogeneous
fluids as a substitute for cell Suspensions have been discussed by Allen and Robinson
[68] and, although Solka Floc fibre suspensions show a strong resemblance to
mycelial cultures such as Aspergillus and Penicillium [91], they are not a close
replica of plant-cell suspensions. This rheological dissimilarity represents a major
limitation of the following theoretical study. However, until further work is
completed with plant cells, correlations derived from simulation studies are the
only ones available for modelling plant-cell reactors.
Design of Reactors for Plant Cells and Organs 139

2.4.1 Oxygen Consumption

During growth, the time constant for oxygen consumption t~n can be expressed as:

C 9
trx, - (33)
ro 2

where C* is the equilibrium concentration of oxygen in the culture liquid, kg

O2 m-3, and to2 is the volumetric oxygen consumption rate, kg O2 m - 3 s -1.
The value for ro2 will depend on the biomass concentration X, kg m-3, and the
observed specific oxygen consumption rate qo2, kg 02 kg-1 s-1. Note that qo2 is
an observed parameter and so incorporates a maintenance contribution.
In the literature, oxygen consumption by plant cells in batch culture does not
show a consistent pattern. Data reported by Kobayashi et al. [74] for Thalictrum
minus suspensions and by Hong et al. [38] for strawberry suspension cultures show
that specific oxygen-uptake rates increase during growth to reach a maximum
near the onset of stationary phase. In contrast, Townsley et al. [92] and Drapeau
et al. [93] demonstrated with Triptergium wildordii, Catharanthus roseus and
Dioscorea deltoidea cultures that specific rate of oxygen consumption declines
progressively throughout batch culture. Quantitative results from these previous
studies are summarised in Table 6.
In the present analysis qo~ will be considered constant over the growth period
and equal to an average value of 10 -6 kg 02 kg dry weight -1 s -1. Combining
these figures with the biomass concentrations, at the beginning of growth ro2 is
5 • 1 0 - 6 kg O 2 m - 3 s - 1 and trx. is 1600 S; at the end of growth ro; is 3 x 10 -5 kg
0 2 m - 3 s - 1 and t~xn is 270 s. The calculated O2 uptake rate at the end of growth
is greater than the maximum rate of 1.6 x 10- 5 kg O 2 m - 3 s - 1 reported by Scragg
and Fowler [94] and Drapeau et al. [93]; this is probably due to the higher cell

Table 6. Literature values for observed specificoxygen uptake rate qo2 in suspended plant-cell
cultures

Plant %2 Reference
kg 0 2 (kg dry weight)-1 s-x x 10 6

Strawberry 1.1-1.5 [38]

(increasing during growth)
Thalictrum minus 0.18-4.3 [74]
(increasing during growth)

qo2
kg 0 2 (kg fresh weight)- 1 s- 1 x 106

Catharanthus roseus 0.21-0.03 [93]

(decreasing during growth)
Dioscorea deltoidea 0.51-0.08 [93]
(decreasing during growth)
140 P.M. Doran

10000

trxn

1000
tmt
tmx \ Jinx
trxn _\
100
(s)

....................... ~_ tmt

10 I q -I-'~ ...... .J

0.1 0.2 0.3 0.4 0,5

a UGr (ms "1)
10000

~, trxn

1000
tmt
t rnx
~tmt
trxn
I0 0 "":k
(s)
';:i~......... tmx
".%... ...............................

0 5 10 15 20
b N i (s1 ) )'~
Fig. 5a, b. Mixing, mass transfer and oxygen consumption in bioreactors at a suspended
plant-cell concentration C s = 5 k g m -3 dry weight a) external-loop airlift reactor; b)
stirred-tank reactor

density. The calculated time constants for oxygen uptake are plotted in Figs. 5
and 6. For comparison, time constants for microbial oxygen consumption are of
the order 9-45 s [90].

2.4.2 Mass Transfer

The expression for the mass-transfer time constant is:
1
trot -- (34)
kLaL
where kLaL is the mass transfer coefficient based on gas-liquid interracial area per
unit volume of liquid.
Design of Reactors for Plant Cells and Organs 141

10000 , , , ,

1000
tmt "",~"............................ ~,trnx
" , , , t rxnJk ................................
tmx
trxn 100
(s)

10 I I P P

0 0.1 0.2 0.3 0.4 0.5

UGr(mS "1)
10000 i . . . . i . . . . I . . . . I . . . .

l
\

1 1000

100 "........ "~ .................... ~mt

~_ ............. tmx .........

10 ~ .............
0 5 10 15 20
b N i (s"1)
Fig. 6a, b. Mixing, mass transfer and oxygen consumption in bioreactors at a suspended
plant-cell concentration Cs = 30kgm -3 dry weight, a) external-loop airlift reactor; b)
stirred-tank reactor

2.4.2.1 Airlift Reactors

For airlift reactors the value of kLaL can be calculated from Eqs. (27)-(29) or
from Eq. (30). In plant-cell cultures, high viscosity is due mainly to the solids
content, so both Eqs. (27) and (30) apply.
kLaL values have been calculated from Eq. (30) with C s = 0.5% and C S = 3%.
Corresponding values for the mass-transfer time constant (trot) for gas velocities
between 0.01 m s - 1 and 0.5 m s - i are plotted in Figs. 5a and 6a. Final kLaL values
at the end of growth are about 14% those near the beginning.
IfkLa L results calculated from Eqs. (27)-(29) and (30) are equivalent, g~pv values
can be determined to check the flow conditions. Depending on the gas flow rate
and therefore the shear rate in the reactor, apparent viscosities at the beginning of
142 P.M. Doran

culture range from 36 mPa s at u~r = 0.01 m s- 1 to 13 m P a s at uGr = 0.5 m s- a.

At the end of the growth period, the corresponding ~app values are 320 Pa s and
110 m P a s. The fluid consistency index K and flow behaviour index n can also
be calculated from Eqs. (3) and (15): K is 0.10 Pa sn at the beginning of the
culture and 0.99 Pa s" at the end; n = 0.71 irrespective of cell concentration.
These calculated values for apparent viscosity, K and n are realistic for plant-cell
suspensions.

2.4.2.2 Stirred Tanks

Since data on the viscoelastic properties of plant-cell suspensions are not available
Eq. (23) cannot be easily applied for calculation of kLaL. Instead, with the above
values for K and n, Eqs. (23)-(25) are used to evaluate kLaL for oxygen transfer
in stirred vessels. Results for the mass-transfer time constants are plotted in Figs.
5b and 6b. Depending on the stirrer speed between 0.2 s-1 and 20 s-1, g,vp values
are 88 m P a s to 23 mPa s at the beginning of culture and 850 mPa s to 220 m P a s
at the end.

2.4.3 Mixing

2.4.3.1 Airlift Reactors

Mixing times for external-loop airlift reactors can be calculated from Eqs. (13),
(16), (18), (19) and (8). The results are plotted in Figs. 5a and 6a.

2.4.3.2 Stirred Tanks

Mixing times for stirred-tank reactors are obtained from the Norwood-Metzner
correlation of Eq. (10). The value of fm is taken to be 20 as determined by Wang
and Fewkes [58] with Streptomyces niveus cultures. Under turbulent conditions
the mixing time in stirred reactors is largely independent of viscosity; consequently
t~,x values are the same at the end of the plant-cell culture as at the beginning.
The results are plotted in Figs. 5b and 6b.

2.4.4 Shear

2.4.4.1 Airlift Reactors

Shear stress in air-driven reactors is currently the subject of extensive investigation.

As discussed in Sect. 2.3.3, recent studies have suggested that highest shear is
associated with bubble disengagement. However, the duration of these forces is
extremely short, of the order of milliseconds. Although these conditions might be
sufficient to cause damage to fragile animal cells without cell walls, it is unclear
whether the same mechanism is responsible for plant-cell damage. In this analysis,
it is assumed that the less intense but more prolonged shear-forces associated with
fluid flow determine whether shear damage is suffered by suspended plant cells in
airlift reactors.
Design of Reactors for Plant Cells and Organs 143

10000 , , ,

1000 Cs= 0.5%

~" ......................
09

09

j/j~176176
.1" ~
go 100
~~_-f~ cs=3%

10 I [
0.1 0.2 0.3 0.4 .5
& UGr (ms"1)
10000
c~ = 0.5% .......... :

1000
f - ~ .....3s= " ~

lOO

1 0 i i I r p i r i I I , r i i I r i I i

0 5 10 15
b N i (s"1)

Fig. 7a, b. Liquid shear stress produced in bioreactors containing suspended plant-cell
concentrations C~ = 5 kg m- 3 dry weight (0.5%) and Cs = 30 kg m- 3 (3%). a) airlift reactor;
b) stirred-tank reactor

Fluid shear-stress is maximum in airlift reactors in the top and bottom sections
between riser and downcomer where flow directions change abruptly. The
magnitude of this stress is calculated from Eq. (32). Results for the beginning and
end of plant-cell culture are plotted in Fig. 7a.

2.4.4.2 Stirred Tanks

In stirred reactors, maximum shear levels associated with the action of the impeller
are used to determine whether plant cells are likely to be disrupted by hy-
drodynamic forces. Shear stress at the impeller tip is calculated from Eq. (31) for
a 6-flat-blade Rushton-type agitator. Results for the beginning and end of the
culture are shown in Fig. 7 b.
144 P.M. Doran

2.5 Comparison of Reactor Performance

The general trend of calculated results shown in Figs. 5 and 6 gives a realistic
picture of reactor performance.
The mixing and mass-transfer situation in stirred and airlift reactors at a plant-
cell concentration of 5 kg m - 3 dry weight is represented in Fig. 5. As reported
in the literature, performance of the airlift reactor at this biomass density is
adequate; both mass-transfer and mixing have shorter time constants than oxygen
consumption. In the stirred reactor conditions are similar; the results show that
only a low stirrer speed is required for adequate oxygen mass transfer. Because
the correlations for stirred-tank kLaL have limited applicability at low Rei below
Ni -- 0.3 s- 1, the actual stirrer speed required for equal values oftrxn and tmt should
not be read directly from Fig. 5b.
By the end of the culture when cell concentration is 30 kg m - 3, the situation
is somewhat different. Figure 6a shows how mixing becomes limiting in the airlift
reactor at high cell densities; at gas velocities below 0.5 m s -1 mixing times are
larger than the oxygen-consumption time so that oxygen gradients will develop.
This is consistent with experimental findings that mixing problems develop in
airlift vessels at biomass densities above about 25 kg m - 3 dry weight [36, 37, 44].
Mass transfer requirements are satisfied at gas flow rates above about 0.08 m s- 1.
From Fig. 6b, mixing in the stirred vessel is very effective even at low stirrer
speeds; however the mass-transfer time constant is greater than the reaction time
constant until N i --- 6 s-1. At N~ > 6 s-1 adequate oxygen transfer and mixing
are provided.
These results for mixing and mass transfer must be considered in conjunction
with predicted shear levels in the reactors (Fig. 7). The maximum shear stress
tolerated by the cells will set upper limits on N~ and Uar. Unfortunately, as discussed
above, tolerance of plants cells to shear is unclear at present. However, as an
example, assume the critical shear stress for growth of plant cells is 100 Pa, 25-100
times higher than that reported for insect cells [27]. From Fig. 7a, operation of
airlift reactors at the beginning of the culture must therefore be limited to gas
velocities less than 0.05 m s-1; from Fig. 5 a this does not adversely affect mixing
or mass transfer. As cell density increases, higher gas velocities up to about
0.3 m s - 1 can be used since the shear stress developed in the liquid decreases with
increasing viscosity. In the stirred reactor (Fig. 7b), N i must be less than about
0.5 s - 1 at the beginning and less than 1.5 s- 1 towards the end of growth to avoid
shear damage to the culture. Although the reactor will function reasonably well
at low stirring speeds at the beginning of the culture (Fig. 5b), the requirement
for low Ni at the end of growth conflicts directly with the necessity for N~ > 6 s- 1
for proper mixing and mass transfer (Fig. 6b). At the assumed level of shear
sensitivity, stirred reactors cannot provide non-destructive conditions for plant-cell
culture.
The above analysis is subject to limitations in addition to those already discussed.
If plant-cell aggregates are present, transport of oxygen depends not only on kLaL
but also on diffusion through the cell pellet. Accordingly, adequate gas-liquid mass
transfer is provided when trot < trxn; however oxygen deficiencies may still occur
Design of Reactors for Plant Cells and Organs 145

in the interior of cell aggregates. In addition, the analysis focuses solely on

requirements for growth; tr~n values were calculated using specific oxygen-uptake
rates during active growth. In interpretation of Figs. 5 and 6 it was assumed that
if time constants for mixing and mass transfer are smaller than for oxygen
consumption then reactor conditions are favourable. In the case of secondary
production however, high rates of mass transfer may not be desirable as oxygen
deficiency is known to stimulate secondary-metabolite synthesis in some cases. In
theory, a similar set of calculations could be carried out to test whether the reactors
are able to provide mixing, mass-transfer and shear conditions for optimal
productivity. This is difficult at the present time because the conditions required
for secondary synthesis are generally not known.

2.6 Improvement o f Reactors

In the above analysis, the various functions of bioreactors: mixing, mass transfer
and provision of shear, were considered separately to determine which limits
growth under different culture conditions. Once the limiting process is identified,
measures can be taken to improve it.

2.6.1 Airlift Reactors

At high cell.density, performance of airlift reactors is limited by mixing. Although
adequate mixing is provided at high gas velocities, increasing the gas flow rate
may not be a practical solution. Foaming and foam separation of cells is a problem
in plant-cell suspensions; to avoid formation of cell meringue at the top of the
vessel high gas flow should be avoided. Antifoams non-toxic to plant cells can be
used to relieve this problem [95, 96]; however significant reduction in k L can occur
after addition of antifoam [97]. Scragg, Fowler and colleagues have identified
another problem with high gas throughput in plant-cell cultures: volatile media
components are stripped from the liquid at high ventilation rates to the detriment
of cell growth [39, 40].
Mixing in airlift reactors must be increased by means other than raising the
gas flow rate. In some cases, attention to geometry: Ad/A r ratio, aspect ratio, draft-
tube clearances and flow-path smoothing (Fig. 4), will give some improvements
to liquid circulation, as will choice of external-loop rather than internal-loop
designs. Multiple gas-injection points in the reactor may also be beneficial. Chisti
[50] suggests that separate sparging of gas into the downcomer could be useful;
study of this mode of operation is needed to identify instability conditions and
monitor the effects on mass transfer and liquid flow. Adding an impeller to the
airlift design should also improve liquid circulation as long as shear levels are
minimised. An axial-flow propeller installed inside the draft tube of an airlift
reactor has been tested with plant-cell suspensions; for Beta vulgaris a stirrer speed
of 28 rpm caused sufficient shear damage to reduce biomass and product yields [9].
Because increased cell content is responsible for the decline of mixing efficiency
in airlift reactors, another strategy would be to reduce cell concentration and
culture viscosity. Dilution of plant-cell suspensions by small volumes of water or
146 P.M. Doran

fresh medium can significantly reduce viscosity. According to the relationship

between viscosity and plant-cell concentration reported by Tanaka [10], a 10%
dilution will lower apparent viscosity by about 50% as reported previously for
suspensions of filamentous mycelia [55]. Semi-continuous processes, where a
volume of culture is withdrawn periodically from the reactor and nutrient medium
added, would be advantageous for plant cells as this allows a far greater reduction
in viscosity than continuous feeding. Draw/fill methods for airlift reactors have
been tested by Scragg et al. [42] with Catharanthus roseus.
Although improvements are possible, if plant-cell densities substantially greater
than 30 kg m - 3 dry weight are required for economic feasibility, it is likely that
even the optimal airlift reactor will not be able to provide adequate mixing and
mass transfer conditions.

2.6.2 Stirred Tanks

In stirred vessels, sufficient mass transfer and mixing occur at stirrer speeds above
about 6 s-1. Although this is a relatively low speed for most microbial processes,
for shear-sensitive cells the shear stress imposed by a Rushton-type impeller under
these conditions would be destructive. Improvement of stirred reactors for
plant-cell culture can therefore be distilled down to reduction of shear levels.
Some studies of low-shear impellers have been carried out with plant-cell
suspensions. Tanaka [36] tested a large modified paddle-type impeller in a
bioreactor with four small baffles; this stirrer was operated at 60 rpm and, after
sucrose feeding, a cell density of 30 kg m - 3 was achieved.
Helical screw impellers (Fig. 8) have been studied extensively [51, 98] for mixing
viscous materials; mixing is accomplished without high liquid velocity streams.
The helical screw usually functions by carrying liquid from the vessel bottom and
discharging it at the liquid surface; alternatively, the screw may be operated in
reverse to pull liquid to the bottom of the vessel. Baffles are generally required in
screw-agitated reactors to promote turbulence and avoid dead zones near the
vessel wall. Helical agitators have been tested with plant-cell suspensions [37, 99].
A helical stirrer operated at 100 rpm produced high levels of rosmarinic acid from
Coleus blumei cells without shear damage and with kLaL values between 1.5 x 10- 3

~ r Fig. 8a. Typical configuration of helical

screw impeller; b) Typical flow pattern
produced by a helical impeller. (Adapt-
It, b ed from Bourne and Butler [98])
Design of Reactors for Plant Cells and Organs 147

and 6.0 x 10- 3 s 1. In the same study an anchor impeller operated at a maximum
speed of 40 rpm produced significantly lower product levels.
Hong et al. [38] installed 25 mm-wide teflon ribbons between the blades of two
turbine impellers in a 7-1itre stirred bioreactor in an attempt to reduce shear and
improve mixing. The modified agitator was tested for culture of strawberry cells;
growth did not occur and the viability of the cells dropped from 80% to less than
25% in 4 days. Better results were obtained by Hooker et al. [19] with a large
flat-blade impeller with and without sail-cloth blades. Higher maximum growth
rates were obtained compared with a smaller flat-blade turbine. Power consump-
tion by large sail-cloth impellers is considerably greater than with Rushton turbines
[511.

2.7 Alternative Reactor Designs

As well as modifying airlift and stirred reactors to improve plant-cell culture,

alternative reactor configurations have been sought. The most promising of these
is the rotating-drum reactor investigated by Tanaka et al. [100]. This design was
found to be superior to mechanically agitated vessels; oyxygen transfer was
sufficient to achieve a biomass density of 20 kg m - 3 dry weight under conditions
of high viscosity and low hydrodynamic stress. Tanaka and co-workers determined
a correlation for kLaL in the drum as a function of rotational speed, number of
baffles and baffle height, tank diameter and fraction of the tank volume filled with
liquid. The decrease in kLaL due to apparent viscosity was smaller in the rotating
drum than in mechanically agitated reactors.

3 Immobilised Plant Cells

Early reports by Brodelius et al. [101] showed that phytochemical synthesis is

enhanced if plant cells are first entrapped in alginate gel. The feasibility of
immobilised plant cells has since been demonstrated for single- and multiple-step
biotransformations [102-104] and for de novo synthesis of secondary metabolites
[105, 106].
Aspects of immobilised plant-cells have been extensively reviewed in recent
years. Discussion of potential technical and physiological advantages can be found
in articles by Shuler et al. [107], Lindsey and Yeoman [108] and Robins et al. [109].
Established techniques for immobilising plant cells by entrapment or surface-
attachment are also described [106, 110-114]. Of the natural gels such as alginate,
agarose, agar, gelatin and carrageenan commonly used for cell entrapment, alginate
appears to be most beneficial for plant cells. Studies comparing the effects of these
materials have shown that better secondary product formation and excretion occur
when alginate is used [115, 116]. This effect has not yet been explained; Hulst et al.
[117] determined that the oxygen-transfer characteristics of alginate were not
significantly different from other gels. Ayabe et al. [118] report that production of
retrochalcone by freely suspended Glycyrrhiza echinata cells is increased if calcium
148 P.M. Doran

alginate beads are added to the medium; Prenosil et al. [119] have postulated that
alginate, being similar to the polycarboxylic acids in cell walls, mediates cell-cell
interactions in the presence of calcium ions to stimulate secondary synthesis.
Enhanced secondary product synthesis has also been reported for plant cells
immobilised in reticulated polyurethane foam [106, 120, 121].
For most efficient operation, the product of interest should be excreted from
immobilised cells into the medium so that the biomass does not have to be
periodically destroyed. There is evidence that immobilisation enhances release of
secondary products from plant cells [101,122]. Alternatively, release of compounds
can be induced by permeabilising agents such as DMSO [123], although limited
success has been achieved using this approach. Once in the medium, products can
be extracted in situ using adsorptive resins such as Amberlite XAD-7 [124] or
immiscible solvents such as vegetable oil [125] or n-hexadecane [126]. Continuous
product extraction encourages further synthesis and minimises degradative meta-
bolism.
Most techniques for immobilising plant cells produce an environment in which
there are substrate, oxygen, hormone and other concentration gradients. This
applies even to methods involving surface attachment; once a monolayer of plant
cells is laid down, other cells readily adhere to it to form a thick biofilm. Designing
large-scale immobilised-cell processes requires knowledge of growth and produc-
tion kinetics. To determine whether mass transfer limits plant-cell activity, analysis
of diffusion processes must be carried out.

3.1 Mass Transfer Considerations

Equations for diffusion and reaction are applied to immobilised cells for prediction
of effective reaction rates. Intraparticle diffusion is the main resistance to mass
transfer in entrapped-cell systems, although external boundary-layers surrounding
the particle can also contribute. Despite the relatively slow metabolic rate of plant
cells, effects of diffusion on activity of immobilised plant ceils cannot generally be
neglected.
As well as reducing the overall reaction rate, diffusional limitations mask intrinsic
kinetics. Care must be exercised when comparing the performance of different
immobilised-cell systems; apparently higher rates of growth or product synthesis
in one system may reflect an absence of diffusional effects rather than any intrinsic
metabolic response.

3.1.1 External Mass-Transfer Resistance

External mass-transfer resistance is largely under the control of reactor hy-
drodynamics; increasing turbulence and mixing reduces the thickness of the liquid
boundary-layer and improves supply of nutrients to the particle surface. External
mass transfer is most likely to contribute to overall diffusion resistance when
reactors are not well mixed, for example in packed beds.
The significance of external diffusion in reactor systems can be determined. In
most cases, if external mass transfer is affecting overall reaction rate, improved
Design of Reactors for Plant Cells and Organs 149

mixing increases the observed reaction velocity [127]. On the other hand, if the
rate of external diffusion is much more rapid than the rate of reaction, the observed
rate should be largely independent of mixing intensity.

3.1.2 Internal Mass-Transfer Resistance

Intraparticle diffusion in immobilised plant-cell systems has been analysed by
Hulst et al. [117, 128], Mavituna et al. [129], Furusaki et al. [130] and Pras et al.
[131]. Whether or not internal mass-transfer of substrate affects reaction rate
depends on several factors:

a) specific cellular demand for substrate, mol cell- 1 s- 1;

b) reaction kinetics, i.e., the relationship between reaction rate and substrate
concentration;
c) biomass loading in the matrix (or biofilm), cells m-3;
d) particle diameter (or film thickness), m;
e) effective diffusivity of substrate molecules through the immobilisation matrix
(or biofilm), m 2 s - l ; and
f) concentration of substrate in the medium at the exterior surface of the particle
(or biofilm), tool cm-3.

Since these parameters vary considerably from system to system, results from one
analysis cannot be readily applied to other immobilised-cell situations.
The molecular species most likely to cause diffusional limitations in immobilised
plant-cells must be determined before the analysis can proceed. Mavituna et al.
[129] analysed diffusion of glucose into reticulated foam containing Capsicum
frutescens cells and determined that mass transfer of this substrate was not an
important problem. Furusaki et al. [130] measured rates of sucrose, oxygen and
codeinone consumption by immobilised Papaver somniferum cells; while effective-
ness factors for these substrates were close to unity, calculations showed that the
concentration of oxygen decreased considerably within the particles. Pras et al.
[131] concluded that limitations in oxygen transport were most likely after
measuring the effective diffusion coefficients and cellular demands for oxygen and
L-DOPA precursors by Mucuna pruriens entrapped in alginate gel.
In the case of oxygen as limiting substrate, for a spherical particle containing
plant cells, if the following assumptions can be made:

a) the cells are identical and their distribution in the particle is uniform;
b) oxygen transport in the particle occurs only by diffusion;
c) oxygen gradients occur only in the radial direction;
d) the effective diffusion coefficient for oxygen is independent of oxygen concentra-
tion;
e) respiration can be described using a Michaelis-Menten kinetic model; and
0 the number of plant cells is constant;
150 P.M. Doran

mass transfer and consumption of oxygen at steady state is described by the

mass-balance equation:

(1 d (r 2 dC))_ VmaxC
(35)

with the usual boundary conditions:

dC/dr = 0 at r = 0 (centre of the particle)

C = Ce at r = R (surface of the particle).

In Eq. (35), r is radial distance measured from the centre of the particle, m; C is
oxygen concentration within the particle, mol m-3; ~e is the effective diffusion
coefficient for oxygen in the particle containing cells, m 2 s- 1; Vmaxis the maximum
reaction velocity, mo! m - 3 s- 1; Km is the Michaelis-Menten constant, mol m - 3;
Ce is the oxygen concentration at the surface of the particle, mol m - 3 and R is the
radius of the particle, m. Equation (35) can be solved only by numerical integration
to give the oxygen profile in the particle, C as a function of r. A similar though
simpler equation is obtained for a biofilm which can be assumed infinitely long
and in which diffusion of oxygen occurs only in one direction.
Hulst et al. [128] calculated using Eq. (35) the particle size at which the oxygen
concentration in the centre is zero. The analysis was performed for cell aggregates
containing 100% Tagetespatula cells with a dry-weight cell-density of 10 kg m -3,
and intrinsic Michaelis-Menten constants vmax = 4 x 10 .3 m o l m -3 s -1 and
K m = 3 . 5 x l 0 - 2 m o l m -3. ~ e = 1 . 9 x 1 0 - 9 m - Z s -1 and C e = 0 . 2 5 m o l m -3.
Under these conditions, the critical particle diameter for depletion of oxygen at the
centre was 3 mm.
Some restrictions apply to use of the diffusion-reaction analysis represented
by Eq. (35). Equation (35) is a valid description of mass transfer and reaction
under conditions of constant biomass concentration in the immobilised-cell
particle. However, for systems in which growth occurs, the value of Vmaxvaries
according to the quantity of cells present. When cell concentration increases in
alginate beads, reticulated foam and membrane units, oxygen concentration
gradients and diffusional limitations will become more severe with time. In
addition, use of Eq. (35) assumes knowledge of the intrinsic kinetic parameters
for immobilised cells, Vmaxand K m. Experimental determination of these values is
not straightforward [128, 131] and use of suspended-cell parameters may not be
valid.
Although observed rates of oxygen consumption are often lower than in
suspension due to diffusion limitations, immobilised plant-cells still perform
secondary synthesis. The exact response to oxygen starvation depends on the
particular cell system and product of interest. Mavituna et al. [125] report that
capsaicin synthesis by foam-entrapped Capsicumfrutescens is enhanced when the
bulk-liquid dissolved-oxygen concentration is almost zero; if aeration is restored
capsaicin in the medium disappears. Pras et al. [131] showed that production of
Design of Reactors for Plant Cells and Organs 151

L-DOPA by immobilised Mucuna pruriens continues in the presence of significant

oxygen limitation. Kargi et al. [132] measured oxygen profiles in biofilms of
Catharanthus roseus and found that, although the oxygen concentration was close
to zero approx. 2 mm from the top surface, indole-alkaloid formation was
stimulated and highest alkaloid levels were found in biofilms of about 4 mm
thickness. It appears that although oxygen supply in immobilised cells is often
restricted due to diffusion limitations, oxygen concentration does not necessarily
control productivity of the cells. Alternatively, as suggested by Lindsey and Yeoman
[108], secondary metabolism and cell differentiation may be enhanced in response
to strong oxygen concentration gradients.

3.2 Self-Immobilised Plant Cells

The ability of plant cells to spontaneously aggregate into macroscopic clumps is

well known and often causes operating problems during large-scale suspension
culture. However tendency to clump can in some cases eliminate the need for
artificial immobilisation supports and could be exploited for large-scale phyto-
chemical synthesis. The success of this approach relies on formation of compact
propagative aggregates of regular shape without large quantities of dispersed cells
also being present. Self-immobilisation of plant cells has not received as much
attention as gel- or foam-entrapment; the range of species which can be
self-immobilised and the long-term feasibility of this type of plant biocatalyst is
unclear. However, there is an increasing number of reports describing formation
and application of aggregated cells, including Populus alba [133], Jasrninum species
[134], Coffea arabica [119], Tagetes patula [128], Cinchona ledgeriana [135], and
Solanum aviculare [136].
General procedures for self-immobilisation of plant cells are described by Fuller
and Bartlett [137]. Enhanced alkaloid synthesis compared with suspended cultures
has been found [135, 136]; these authors also report organisation of cells within
the aggregates and evidence of tracheid differentiation. Nutrient starvation in
larger aggregates causes cell lysis so that the centres are hollow. Hulst et al. [128]
correlated thiophene production with aggregate diameter; mean thiophene levels
increased with size up to 12 mm.

3.3 Immobilised-Cell Reactors

Bioreactor design for scale-up of immobilised plant cells is at an early stage;

suitability of different configurations for immobilised-cell culture is still being
tested. Application of established reactor designs such as stirred tanks, packed
and fluidised beds and airlift vessels has been described in previous review papers
[113, 114, 138]. Self-immobilised plant cells have requirements similar to artificially-
immobilised cells in terms of reactor design, and have been cultured successfully
in fluidised-bed [119] and conical bubble-column [136] reactors.
152 P.M. Doran

Important criteria for bioreactor design for immobilised plant-cells include:

a) good bulk-fluid mixing to minimise external mass-transfer resistances;

b) adequate exposure to air or aerated medium; and
c) minimal shear damage.

Mode of aeration in immobilised plant-cell reactors appears to have an important

influence on cell activity. In early work, Veliky and Jones [103] found that
production of 5[3-hydroxygitoxigenin in a packed-bed reactor was improved if air
were sparged directly into the column; if the medium were aerated away from the
reactor and recirculated through the bed, bioconversion was lower and cells quickly
lost viability. Kobayashi et al. [139, 140] devised a modified packed-bed reactor
for alginate-entrapped Thalictrum minus cells to provide direct contact with air.
Operation involved periodic exposure of the beads to liquid medium and then to
air; optimal cycle conditions for berberine production were 2 min exposure to
medium then 30 s air supply. Direct contact with air appeared crucial for synthesis
of berberine; when beads were continuously soaked in medium saturated with
oxygen, production was almost completely suppressed.
As well as the standard types of bioreactor, novel designs have also been tested.
Special reactors have been devised for surface-attached cells and biofilms;
surface-immobilisation of plant cells has the aim of reducing diffusion limitations
while maintaining high levels of cell-cell contact. Archambault et al. [141, 142]
attached Catharanthus roseus cells to non-woven, short-fibre polyester; the fabric
was wound into a square-spiral configuration on stainless-steel supports and placed
into airlift and mechanically-agitated vessels. A similar arrangement was also
tested with Papaver sornniferum cells for sanguinarine production [143]. Plant cells
were immobilised in situ; for P. somniferum, suspended cells became attached to
the matrix after 10-15 min exposure to the fabric and the entire surface was
covered within 6d. Kargi 'et al. [132] describe a horizontal-biofilm reactor contain-
ing plant cells growing on nutrient agar over which liquid medium is continuously
recirculated. The optimum biofilm thickness for maximum alkaloid production
was determined using this device.
The problem of intracellular storage of products in immobilised plant-cell
systems has been addressed by Tramper et al. [144]. A new type of density-difference
bioreactor called a liquid-impelled loop reactor has been designed; organic solvent
immiscible in water is introduced into the vessel to cause circulation and mixing
in much the same way as air does in airlift reactors. The solvent can be either
lighter or heavier than water; in either case the product of interest should partition
into the organic phase in preference to the aqueous phase. This has several
advantages: in situ extraction of product can be achieved for easier downstream
processing, product excretion is enhanced as concentrations in the aqueous phase
are reduced, and the chances of product inhibition or further metabolism are lower
if direct contact between cells and metabolite is removed. Organic solvents can,
however, have a negative effect on cell activity. Final selection of organic solvent
must also consider the density difference between the solvent and aqueous phase;
a high density difference facilitates mixing in the reactor and phase separa-
Design of Reactors for Plant Cells and Organs 153

tion. The liquid-impelled loop reactor has been tested with free and immobilised
cells of Rubia tinctorum and Morinda citrifolia, with hexadecane as solvent
[145].

3.4 Effect o f Immobilisation on Plant-Cell M e t a b o l i s m

Interest in immobilising plant cel!s is motivated to a large degree by reports of

enhanced secondary production. There is an expectation that increased cell-cell
contact and chemical and electrochemical gradients will mimic conditions in
the whole plant to the benefit of secondary synthesis [108]. While enhanced
secondary production has been observed in some cases, the actual reasons for this
response have not yet been elucidated. Any metabolic changes caused by
immobilisation seem to be reversible; upon release of plant cells from calciumal-
ginate gel it has been found that production rapidly returns to the same level as
in suspended cells [146].
Little work has been carried out to determine the properties of immobilised
plant cells beyond analysis of growth, sugar and oxygen consumption, and alkaloid
synthesis. Brodelius [147] compared phosphate uptake and metabolism by free
and immobilised Catharanthus roseus cells using 31P-NMR; however, no significant
differences were found. Mass-transfer limitations in most immobilised plant-cell
systems make it difficult to draw conclusions about the effects of immobilisation
on cell function. Kargi et al. [132] measured RNA and protein levels in a biofilm
of Catharanthus roseus attached to agar; cells in the lower layers of the biofilm
contained less RNA than those at the top which had better access to nutrients
and oxygen. The interior ceils were also slower-growing and contained less protein
than top-layer cells. Linear growth kinetics and lower biomass yields measured
by Archambault et al. [141] for immobilised Catharanthus roseus are also most
likely due to diffusional limitations; similar responses to oxygen-limited conditions
were described previously by Kato et al. [71] and Pareilleux and Vinas [72] for
suspended cells. There is evidence, however, that plant-cell immobilisation changes
the proportions of secondary products synthesised: Vanek and Macek [148] report
that relative yields of three biotransformation products changed depending on
whether Solanurn aviculare cells were freely suspended or immobilised in any of
five different support matrices. Increased tendency to secrete non-polar products
and synthesis of compounds not detected in suspended-cell culture have been
observed for immobilised Tagetes minuta [116].
The effects of immobiiisation on cell-cycle function and kinetics have not been
determined for plant Cells. Significant changes in growth patterns and DNA content
following surface attachment of microorganisms have been reported [149] but it is
not known whether immobilisation of plant cells interferes with cell-cycle operation
in a similar way. Ketel et al. [116] have suggested that the rigidity of calcium
alginate imposes physical stress on plant cells which stimulates production and
excretion of secondary metabolites. It is also possible that spatial restrictions due
to the surrounding solid matrix affect cell expansion and division, and limit growth.
154 P.M. Doran

The ability of plant cells to differentiate in response to concentration gradients

increases the difficulty of separating nutrient-limitation effects from intrinsic
metabolic changes caused by contact with the support matrix.

4 Differentiated Plant Tissue

A strong correlation between secondary-metabolite production and morphological

differentiation has been shown repeatedly in plant tissue-culture studies; secondary
synthesis is often poor in callus and suspensions but recovers immediately after
organogenesis [150, 151]. Organogenesis also alleviates problems of instability in
de-differentiated plant cells [152, 153]. Large-scale culture of roots is an attractive
proposition for root-derived products; roots are able to extend and grow
indefinitely as a result of cell reproduction at the tip meristem. Several studies
have demonstrated that product levels in roots cultured in vitro are at least as
high as those found in the whole plant [154-156]. Embryos are another potential
route for secondary-metabolite manufacture; large-scale culture of somatic em-
bryos has additional applications for plant propagation and production of artificial
seed. The ability of a large number of plants to form multiple shoots from axillary
and shoot-tip meristems could also be used for phytochemical production and
for large-scale propagation of elite genotypes. Exploitation of these forms of
organised plant tissue on a commercial scale requires design of appropriate
reactors. This area is developing rapidly.

4.1 Hairy Roots

Although in vitro culture of excised plant roots was first demonstrated as early
as 1934 [157], slow growth of most roots in liquid medium and the requirement
for exogenous growth factors are major drawbacks for commercial application.
Growth rates can be increased significantly by genetic transformation with
Agrobacteriumrhizogenesbacteria to form 'hairy roots'; these can then be cultured
axenically without external hormones. Hairy-root cultures have been tested for
production of a broad range of secondary products; members of the Solanaceae
family known to synthesise alkaloids in their roots have been studied extensively.
Review articles are available discussing the plant host-range for Agrobacterium
rhizogenes, procedures for infection, methods of hairy-root culture, and the range
of compounds produced in vitro [158-161].
In an effort to identify appropriate reactor configurations, several studies have
been carried out with hairy roots in batch [162, 163] and continuous-flow
air-sparged reactors [164, 165]. Final biomass densities of 10-11 kg m - 3 dry weight
have been obtained [162, 164]. When continuous liquid-flow operation is used,
this figure can be increased to about 40 kg m -3 dry weight [165]. An important
problem with reactor-culture of hairy roots.is poor distribution of biomass in the
vessels. Growth of hairy roots, whether in shake-flasks or in reactors, usually
results in formation of a dense root ball [163, 166]. Nutrient and oxygen limitations
Design of Reactors for Plant Cells and Organs 155

in the centre of the ball reduce cell activity and can lead to necrosis. Poor
distribution of roots also means that the total volume of the reactor is not used
efficiently. To overcome these problems, roots must be more evenly spread and
mass transfer improved. Use of conventional stirred reactors is generally not an
option for freely-suspended hairy roots; root tissue is easily damaged by mechanical
forces and excessive shear in reactors causes callus generation and loss of
productivity [165]. Several modified and novel designs have been proposed recently.
A common feature of these reactors is that the roots are immobilised on supports
within the vessel.
Jung and Tepfer [155] cultured Calystegia sepium hairy roots in a 30-1itre stirred
vessel containing a stainless-steel basket for attachment of the roots. Taya et al.
[162] report culture of horseradish hairy-roots immobilised on polyurethane foam
in a column reactor. The foam stood vertically in the reactor and air was pumped
into the vessel at the bottom. As well as submerged culture, trickling-column
operation and periodic filling and withdrawal of medium were tested. Growth of
carrot hairy-roots and the relationship between growth rate, oxygen mass-transfer
coefficient and reactor design have been investigated using three different bio-
reactors: air-driven, rotating-drum, and modified stirred-tank [167]. Growth in
the air-driven vessel was slow; mass-transfer coefficients were significantly reduced
in the presence of 10 kg m - 3 roots. Roots in the rotating drum were damaged
until a polyurethane-foam sheet was attached to the inner wall of the rotating
drum to provide protection and support. When 10 kg m -3 roots were immobilised
on the foam, mass-transfer coefficients increased significantly compared with those
measured in the absence of cells. In the stirred-tank reactor, a stainless-steel mesh
was installed to separate roots from the turbine impeller; mass transfer in this
vessel was high and unaffected by root growth. In most reactors, oxygen supply
to immobilised hairy-roots depends not only on bulk gas-liquid mass-transfer but
also on diffusion into the tangled root mass.
A stirred reactor in which roots were isolated from the impeller by wire mesh
was also tested by Hilton and Rhodes [168]. In this case a cylindrical mesh cage
was used as support matrix for Datura stramonium hairy roots; this allowed a
more even distribution of biomass up the length of the vessel. With continuous
feeding of medium, packing densities reached 70% (drained weight per volume);
however, mixing under these conditions was not ideal so that localised regions of
low oxygen and nutrient levels are likely to have been formed.
One solution to the problem of oxygen transfer to hairy roots is use of mist
reactors. In these vessels roots are not submerged in nutrient medium but are
sprayed with fine droplets; the liquid drains through the biomass and is collected
for recycle. Hairy roots respond well to direct exposure to air; Wilson et al. [169]
have compared performance of Datura strarnoniurn hairy roots in droplet- and
submerged-culture reactors. A 7-d lag phase observed in submerged culture was
eliminated in the mist reactor and final specific hyoscyamine levels were slightly
higher.
In addition to biomass distribution and mass transfer there are other considera-
tions affecting feasibility of commercial-scale hairy-root culture. These have been
identified by Wilson et al. [169] as:
156 P.M. Doran

a) aseptic transfer of roots between inoculum and growth vessels; and

b) harvesting of the biomass.
Because of the tangled nature of root cultures, inoculation of large-scale production
vessels is difficult to achieve aseptically. Roots grown on wire trellises or
polyurethane foam are also difficult to remove from the support either for
inoculation of a larger reactor or for product extraction. One solution to the
problem of biomass harvesting is a collapsible root-immobilisation matrix
described in a recent patent application by Wilson et al. [170]. A series of vertical
stainless-steel rods make up the matrix; even after roots grow over the wires, the
rods can be easily withdrawn by sliding them out of the root ball. This system
has been tested with hairy roots in a 500-1itre pilot-scale mist reactor.

4.2 Embryos
Mass culture of somatic embryos can be used for either plant propagation or
production of phytochemicals.
Somatic embryogenesis is a proven means for producing large numbers of plants
in vitro. Suspension cultures in which embryos float freely in the medium are
especially amenable to large-scale mechanical handling. Somatic embryos of some
plants do not respond well in liquid-culture; nevertheless, propagation using
embryo suspensions has been demonstrated for vegetables such as celery and
carrot and for several other species which readily form embryos in a liquid medium
[171-173], The ability to produce and grow embryos on a large scale is necessary
for commercial application of this technology. Theoretically, large bioreactors
would not be necessary since each embryo develops into an entire plant and more
than 100000 embryos can be cultivated in a few litres of medium [174, 175]. In
practice, however, embryo cultures contain a heterogeneous mixture of cells, cell
clusters, embryos in various states of development and deformed embryos. Luckner
and Diettrich [176] report that many Digitalis lanata embryos at the bipolar,
heart-shaped or torpedo stage developed into plantlets with roots and leaves, but
fewer than 1% developed into normal plants after further cultivation. Procedures
for identifying viable embryos and separating them from the remainder of the
culture are required.
Another use of embryos is direct sowing in the field; each embryo is a potential
seedling and can be used as a seed substitute. Encapsulation of somatic embryos
in gel at a saitabie stage of development prevents premature desiccation and
provides growth regulators which control seedling development. The benefits of
artificial seeds include eiimination of seed-borne diseases and development of
improved strains in a relatively short time. Plant Genetics, Inc. in California has
developed techniques for artificial seeds of celery and alfalfa [177]. Somatic embryos
may also prove useful for long-term storage such as in germplasm banks.
Secondary metabolites are synthesised by somatic embryos; it is possible that
high levels are produced because of the effects of cellular organisation. When
torpedo-stage and older embryos produced in celery suspension cultures were
analysed for production of flavour compounds, higher levels of secondary products
Design of Reactors for Plant Cells and Organs 157

accompanied increased differentiation [178]. The secondary-product composition

of the embryo culture was comparable to that of the intact plant. Synthesis of
cardiac glycosides by embryos of Digitalis lanata has been studied by Luckner
and co-workers [173, 176, 179]. Large numbers of embryoids are formed at low
auxin levels and can be grown as stage-I-globules in nutrient medium. After dilution
and irradiation with light, the embryos develop into stage-II globules and start
to accumulate cardenolides [173]. Cardenolide accumulation is highly sensitive
to photoperiod and energy of irradiation [179].
Globules and developed embryos usually range in size from 0.5 to 1.5 mm and
are well suited for culture in reactors. Airlift and stirred reactors were tested by
Chen et al. [180] for mass-culture of alfalfa embryos; these studies were aimed at
determining which operating parameters affect embryo formation. In the airlift
vessel cell viability remained constant at about 60%; in the stirred reactors viability
declined to less than 30% presumably due to shear effects. However embryogenesis
was inhibited in the airlift vessel, especially at high oxygen levels. A 1-1itre
hanging-stirrer-bar reactor proved to be the most suitable configuration; viability
was not adversely affected and embryo yields were enhanced. Depending on size,
between 30-80% of the potential embryos produced in this reactor germinated
normally on hormone-free medium and developed into plantlets. The resulting
alfalfa plants were uniform and possessed normal morphology.
Development of stage-II globules and late-embryo structures of Digitalis lanata
has been carried out in a 5-1itre internal-loop airlift vessel irradiated by two
high-pressure mercury lamps [181]. Cardenolide content of the embryos depended
on the cell density and developmental stage at the time of inoculation, as well as
on intensity of irradiation and nutrient composition. Concentrations of oxygen
and carbon dioxide in the gas used to agitate the embryo suspension also affected
growth and product synthesis.
Long periods of time are sometimes required for development of embryos from
suspension culture; this limits their application for secondary-metabolite produc-
tion. To alleviate this problem, immobilisation of embryonic cells has been
investigated so that their biosynthetic capacity can be used in reactors for longer
periods of time [182]. Cells of an embryonic strain of alfalfa were immobilised in
polyurethane foam; embryoids were visible after 10d and filled the foam matrix.
The ability of the immobilised embryos to carry out biotransformation reactions
was also verified.

4.3 Shoots, Buds and Plantlets

Like embryo culture, large-scale plantlet culture can be used for either mass
propagation or secondary-metabolite production. Micropropagation of many
plant species by shoot-culture techniques on agar is well established; however
organs can also be formed from suspended plant-cells. Requirements for shake-flask
and bioreactor cultivation of differentiated plants are being investigated.
Conventional shoot propagation is labour intensive, expensive, and requires
thousands of containers to produce a large number of plants. Use of submerged
158 P.M. Doran

culture in reactors to propagate plants is potentially cost-effective, and roots and

shoots can be initiated in the same vessel. Reactor-culture of ornamental plants
has been demonstrated by Takayama and Misawa [183]. Regenerated buds of
Begonia x hiemalis were grown in 3- and 10-1itre air-sparged vessels; a maximum
22-fold increase in biomass was achieved. Growth in both these reactors was lower
than in shake flasks; root growth was also severely inhibited. It was concluded
that the major obstacle to use of reactors for large-scale plant propagation is
injury to shoots caused by vigorous aeration. Damage was reduced when shoots
adhered to the internals of the vessel and were thus protected from shear; this
suggests that some form of immobilisation would be beneficial. Shoots removed
from the medium were very turgid.
Further studies with a 2-1itre rectangular-shaped air-driven reactor were carried
out by Park et al. [184]. With young shoots of Artemisia annua, after 29-d growth
the entire vessel volume was packed with shoots; this represented an 8-fold increase
in biomass. The feasibility of direct root induction was also tested. After transferring
shoots to the reactor containing low-glucose rooting medium, about 20% of the
shoots produced roots in 32d; however the vessel became packed with shoots due
to growth and circulation was impeded.
Alkaloid production in shoot cultures has been reviewed by Heble [185].
Medicinal compounds reported to accumulate at high levels in shoot cultures
include those produced by Atropa belladonna, Rauwolfia serpentina, Catharanthus
roseus, Papaver bracteatum, Cinchona sp., Digitalis sp. and Dioscorea composita.
Reactor culture of alkaloid-producing shoots has been limited. Digitoxin synthesis
in shoot-forming cultures of Digitalis purpurea is described by Hagimori et al.
[186]; shoots were grown in a 3-1itre air-sparged reactor with and without
mechanical agitation. At the start of culture, plantlets circulated freely in the
reactor; as their size increased due to generation of new leaves they floated and
accumulated at the surface and after 15d the vessel was filled with plantlets. This
high biomass density prevented uniform illumination of the culture with the result
that cell digitoxin and chlorophyll contents depended on distance from the reactor
wall. Mechanical agitation at 100 rpm had little effect on behaviour of the culture
compared to when aeration only was used. Aeration rate had a strong influence
on growth and digitoxin formation. Digitoxin content in the reactor-cultured
shoots was a little lower than in shoots grown in shake flasks; light deficiency in
the interior of the vessel was considered the most likely explanation.

5 Photoautotrophic Plant Cells

Photoautotrophic metabolism is not usually expressed in cultured plant cells
because they are grown heterotrophically in dark or dim light conditions in the
presence of an organic carbon source. However, photoautotrophic growth of plant
cells in vitro has been achieved for about 23 different species [187]; these cultures
provide new material for research on photosynthesis and chloroplast development
and are also being considered as a way of increasing secondary metabolite
production in tissue culture. Detailed descriptions of photoautotrophic cultures
Design of Reactors for Plant Cells and Organs 159

and their characteristics can be found in the reviews of Hfisemann [188], Hfisemann
et al. [189] and Widholm [190]. Potential application of green hairy-roots of Tagetes
and Bidens species for secondary metabolite production is discussed by Flores
et al. [191].
In the laboratory, plant cells can be induced to develop photoautotrophic
metabolism in two-tier culture flasks containing K2CO 3 - K H C O 3 buffer which
produces a CO 2 atmosphere [192]. Formation of chlorophyll and fully functional
chloroplasts in cultured plant ceils has been shown [193] to involve:
a) lowering the sugar content and simultaneously increasing the CO 2 partial-
pressure above ambient level;
b) preventing oxygen and ethylene accumulation;
c) providing high light intensities (6000 8000 lux); and
d) maintaining a balanced ratio of auxin/cytokinin in the medium.
When photoautotrophism is achieved, it is usually in sugar-free medium but in the
presence of air containing 1 2% v/v CO2. The reason why most cultured plant-
cells are only capable of sustained photoautotrophic growth at highly elevated
CO 2 partial pressures is still unknown; however, sustained photoautotrophic
growth under ambient CO2 concentrations has been demonstrated for selected cell
lines of Arachis hypogaea, Daucus carota [194] and Gossypiurn hirsutum [190].
Since some enzymes involved in secondary synthesis are associated with
chloroplasts and chloroplast membranes, it is expected that production rates of
these compounds will be enhanced if chloroplasts are present. Comparative
measurements of secondary-metabolite accumulation in heterotrophic, photo-
mixotrophic and photoautotrophic cell have shown that green cell-cultures are
able to express products partially or totally synthesised by chloroplast-localised
enzymes. Systems in which greening enhances secondary production include
Lupinus polyphyllus [195], Morinda lucida [196], Solanum dulcamara [197] and
Solanurn laciniatum [198]. In many cases, however, transition from heterotrophic
to photoautotrophic conditions is not sufficient to induce secondary metabolism
characteristics of leaves, and production of commercially-interesting secondary
metabolites is either unchanged or decreased in photoautotrophic cells. For
example, in Nicotiana tabacum, nicotine and other secondary metabolites occur
mainly in heterotrophic rather than photoautotrophic cells, although N-methyl-
nicotinic acid formation is totally restricted to photoautotrophic cultures [199, 200].
Hagimori et al. [201] found that digitoxin content in undifferentiated photoauto-
trophic Digitalis purpurea cells was the same as in photomixotrophic cultures; in
shoots, photoautotrophism reduced the alkaloid content to 10% of that in
photomixotrophic cultures. Photoautotrophic suspensions of Catharanthus roseus
synthesise very low levels of vindoline and dimeric alkaloids [202]; production of
these compounds is drastically increased when heterotrophic conditions are
restored [189].
For those compounds produced in quantity in photoautotrophic cells, large-scale
production depends on design of suitable reactors. Important considerations for
maintenance of photoautotrophic cultures include provision of elevated levels of
CO2, reduced levels of Oz, and adequate light intensity.
160 P.M. Doran

Continuous culture of photosynthetic plant-cell suspensions was demonstrated

by Dalton [203]. Spinacia oleracea cells were cultured photoautotrophically at
several steady states for more than 7 months in a 1.7-1itre stirred reactor with
dissolved-oxygen control and carbon-dioxide addition. Illumination was provided
with 4 6 fluorescent lights; higher light intensities were obtained using a curved
reflector of aluminium foil. The results from these experiments suggested that light
intensity limited biomass accumulation and biosynthesis. Batch and fed-batch
reactor cultures of green Asparagus officinalis and Ocimum basilicum cells were
subsequently carried out [204].
Photoautotrophic growth of Nicotiana tabacum was achieved in a 5-1itre stirred
reactor with an incident light intensity of 8000 lux and an aeration gas containing
1% CO2, 14% 0 2, and the remainder N2 [205]. Enrichment with 1% CO2 without
a simultaneous reduction in 02 partial pressure did not induce photoautotrophism.
Peel [206] cultured photoautotrophic cells of Asparagus officinalis at steady state
in a small turbidostat reactor for 21d. Hiisemann [207] reports use of a 1.5-1itre
inverted-flask bubble-column reactor for photoautotrophic culture of Cheno-
podium rubrum in the presence of 2% CO2 and 8000 lux external light intensity.
Chlorophyll and protein contents and activities of ribulose bisphosphate carboxy-
lase and phosphoenolpyruvate carboxylase enzymes were reduced compared with
cultures grown in shake flasks. With the same equipment, Hiisemann [208] used
continuous-culture photoautotrophic conditions to show that multiplication and
differentiation of chloroplasts keep pace with cell division. Scale-up of suspended
photoautotrophic Chenopodium rubrum cells in a 17-1itre airlift vessel has been
described by Fischer and Alfermann [209].
Uniform and adequate illumination of cells is an important problem in large-
scale photoautotrophic culture and becomes more difficult in large reactors if an
external light source is used. To overcome these problems, Ohta and Takata [210]
have designed a 10-1itre vessel in which light is transmitted by optical fibres. The
light source is a xenon lamp or solar-ray collector; quartz optical fibres coupled
with polyacrylate resin fibres are inserted into the reactor through protective glass
tubing. This system has been tested with photoheterotrophic cultures of liverwort.
Provided that sufficient light intensity can be transmitted using optical fibres,
devices providing an internal source of light are also likely to be successful with
photosynthesising plant-cell cultures.

6 Conclusions and Future Outlook

The newest developments in plant-cell reactor design are those aimed at culture
of differentiated organs and tissue; this leaves many unsolved problems with reactor
culture of high-density suspended plant-cells. Although air-driven bioreactors work
well at suspended-cell concentrations less than 12 15 kg m-3 dry weight, mixing
becomes limiting at higher biomass densities. Stirred reactors have more potential
for large-scale suspended-cell processes because they are easily able to provide
adequate mixing and mass transfer; however, the shear sensitivity of many plant
cells restricts application of standard configurations. Research into inducing or
Design of Reactors for Plant Cells and Organs 161

selecting for shear tolerance is a worthwhile approach so that the range of

application of agitated systems can be broadened. Alternatively, design of
impellers which maintain high levels of mixing and mass transfer but reduce the
intensity of shear may provide a satisfactory solution.
Design of reactors for immobilised plant-cells and organised tissue is at an early
stage. Many standard reactor configurations appear adequate for immobilised cell
culture. For hairy roots, immobilisation of the biomass to ensure uniform
distribution in the reactor is important; efficient ways of harvesting the cells must
then be found. The general applicability of embryos, shoots and plantlets for
secondary-metabolite production is yet to be demonstrated. However, protection
from shear and provision of adequate light are important considerations in reactor
design.
Engineering analysis of large-scale plant-cell culture must go hand-in-hand with
more fundamental studies of plant-cell metabolism. Requirements for growth of
cells may be very different from those for maximum productivity so that criteria
for reactor design will vary. As an example, aiming for kLaL values which do not
limit growth seems appropriate; however secondary production could be enhanced
under oxygen-limited conditions, or high air flow rates may deplete the culture
of vital components such as carbon dioxide and ethylene. Uncertainty about
culture conditions would be reduced if current studies of pathway regulation in
plant cells can provide a formula for high product yields. In many ways, further
refinement of reactor design for plant-cell culture depends on the outcome of
research into how secondary synthesis can be triggered.

7 Nomenclature
Ad downcomer cross-sectional area
aD interfacial area per unit dispersed volume
aL interracial area per unit volume unaerated liquid
Ar riser cross-sectional area
C oxygen concentration within particle
C* oxygen concentration in liquid in equilibrium with air
Ce oxygen concentration at surface of particle
CL oxygen concentration in liquid
C~ solids concentration
D diffusivity
Db bubble diameter
Dc column diameter
Dd downcomer diameter
De effective diffusion coefficient
Di impeller diameter
DE diffusivity in liquid
Dr riser diameter
Dr tank diameter
fm dimensionless mixing group
162 P.M. Doran

g gravitational acceleration
HD gas-liquid dispersion height
Hd height of the downcomer
HL liquid height
H~ height of the riser
K power-law consistency index
k empirical constant in Eq. (9)
KB frictional loss coefficient for bottom of airlift reactor
kL liquid-film mass transfer coefficient
Km Michaelis-Menten constant
Kx frictional loss coefficient for top of airlift reactor
kw resistance coefficient in Eq. (32)
n power-law flow behaviour index
Ni impeller rotational speed
OTR oxygen transfer rate
P power input
APvd frictional pressure drop in the downcomer
APF~ frictional pressure drop in the riser
qo2 observed specific oxygen consumption rate
R radius of particle
r radial distance measured from centre of particle
Re~ impeller Reynolds number
ro 2 volumetric oxygen consumption rate
tc circulation time
trot mass transfer time constant
tmx mixing time constant
trxn oxygen consumption time constant
UB bubble rise velocity
UG superficial gas velocity
UGr superficial gas velocity in the riser
HL liquid velocity
ULd superficial liquid velocity in the downcomer
ULr superficial liquid velocity in the riser
VL liquid volume
Vmax maximum reaction velocity
X cell concentration
exponent in Eq. (26)
shear rate
~{~v average shear rate
~/max maximum shear rate
local energy dissipation rate
8d gas hold-up in the downcomer
~r gas hold-up in the riser
~T total gas hold-up
q eddy size
L characteristic material time
Design of Reactors for Plant Cells and Organs 163

~[app a p p a r e n t viscosity
gg gas viscosity
gL liquid viscosity
v fluid kinematic viscosity
QI~ liquid density
(y surface tension
z shear stress
to yield stress
~w wall shear stress

8 References
1. Fowler MW (1983) In: Mantell SH, Smith H (eds) Plant biotechnology. Cambridge
University Press, Cambridge, p 3
2. Misawa M (1985) Adv Biochem Eng/Biotechnol 31:59
3. Kurz WGW, Constabel F (1985) CRC Crit Rev Biotechnol 2:105
4. Rhodes MJC, Robins RJ, Hamill J and Parr AJ (1986) NZ J Technol 2:59
5, Kato A, Kawazoe S and Soh Y (1978) J Ferment Technol 56:224
6. Roels JA, van den Berg J and Voncken RM (1974) Biotechnol Bioeng 16:181
7. Charles M (1978) Adv Biochem Eng 8:1
8. Metz B, Kossen NWF and van Suijdam JC (1979) Adv Biochem Eng 11:103
9. Wagner F, Vogelmann H (1977) In: Barz W, Reinhard E, Zenk MH (eds) Plant tissue
culture and its bio-technological applications. Springer, Berlin Heidelberg New York,
p 245
10. Tanaka H (1982) Biotechnol Bioeng 24:425
11. Scragg AH, Allan EJ, Bond PA and Smart NJ (1986) In: Morris P, Scragg AH, Stafford
A and Fowler MW (eds) Secondary metabolism in plant cell cultures. Cambridge
University Press, Cambridge, p 178
12. Vogelmann H (1981) In: Moo-Young M, Robinson CW and Vezina C (eds) Advances
in biotechnology, vol 1. Pergamon, NewYork, p 117
13. Laine J and Kuoppam~iki R (1979) Ind Eng Chem Process Des Dev 18:501
14. Takayama S, Misawa M, Ko K and Misato T (1977) Physiol Plant 41:313
15. Sahai OP and Shuler ML (1982) Can J Bot 60:692
16. Street HE (1973) In: Milborrow BV (ed) Biosynthesis and its control in plants. Academic,
London, p 95
17. Tanaka H, Semba H, Jitsufuchi T and Harada H (1988) Biotechnol Lett 10:485
18. Metzner AB and Otto RE (1957) AIChE J 3:3
19. Hooker BS, Lee JM and An G (1990) Biotechnol Bioeng 35:296
20. Scragg AH, Allan EJ and Leckie F (1988) Enzyme Microb Technol 10:361
21. Hooker BS, Lee JM and An G (1989) Enzyme Microb Technol 11:484
22. Allan EJ, Scragg AH and Pugh K (1988) J Plant Physiol 132:176
23. Meijer J (1990) PhD Thesis, Delft University of Technology, The Netherlands
24. Midler M and Finn RK (1966) Biotechnol Bioeng 8:71
25. Chen H, Wang J-J and Liu Y-G (1990) In: Abstracts VIIth Internat Congr Plant Tiss
Cell Cult, Amsterdam, 24-29 June 1990, p 341
26. Cherry RS and Papoutsakis ET (1986) Bioprocess Eng 1:29
27. Tramper J, Williams JB, Joustra D and Vlak JM (1986) Enzyme Microb Technol 8:33
28. Croughan MS, Sayre ES and Wang DIC (1989) Biotechnol Bioeng 33:862
29. Handa-Corrigan A, Emery AN and Spier RE (1989) Enzyme Microb Technol
11:230
30. Kunas KT and Papoutsakis ET (1990) Biotechnol Bioeng 36:476
31. Murhammer DW and Goochee CF (1990) Biotechnol Prog 6:391
164 P.M. Doran

32. J6bses I, Martens D and Tramper J (1991) Biotechnol Bioeng 37:484

33. Bavarian F, Fan LS and Chalmers JJ (1991) Biotechnol Prog 7:140
34. Chalmers JJ and Bavarian F (1991) Biotechnol Prog 7:151
35. Cherry RS and Kwon K-Y (1990) Biotechnol Bioeng 36:563
36. Tanaka H (1981) Biotechnol Bioeng 23:1203
37. Ulbrich B (1986) In: Korhola M, Tuompo H and Kauppinen V (eds) Foundation for
biotechnical and industrial fermentation research 4. Proc 7th Conf on Global
Impacts of Appl Microbiol, p 147
38. Hong YC, Labuza TP and Harlander SK (1989) Biotechnol Prog 5:137
39. Smart NJ and Fowler MW (1984) Appl Biochem Biotechnol 9:209
40. Hegarty PK, Smart NJ, Scragg AH and Fowler MW (1986) J Exp Bot 37:1911
41. Scragg AH, Cresswell RC, Ashton S, York A, Bond PA and Fowler MW (1989) Enzyme
Microb Technol 11 : 329
42. Scragg AH, Ashton S, York A, Bond P, Stepan-Sarkissian G and Grey D (1990) Enzyme
Microb Technol 12:292
43. Goldstein WE, Lasure LL and Ingle MB (1980) In: Staba EJ (ed) Plant tissue culture
as a source of biochemicals. CRC Press, Boca Raton, USA, p 191
44. Tanaka H (1987) Process Biochem 22:106
45. Noguchi M, Matsumoto T, Hirata Y, Yamamoto K, Katsuyama A, Kato A, Azechi
S and Kato K (1977) In: Barz W, Reinhard E and Zenk MH (eds) Plant tissue culture
and its bio-technological applications. Springer, Berlin Heidelberg New York, p 85
46. Curtin ME (1983) Bio/Technol 1:649
47. Fujita Y (1988) In: Bock G and Marsh J (eds) Applications of plant cell and tissue
culture. Ciba foundation syrup 137, Wiley, Chichester, p 228
48. Westphal K (1990) In: Progress in plant cellular and molecular biology. Proc VIIth
Internat Congr Plant Tiss Cell Cult, Amsterdam, 24-29 June 1990, Kluwer Academic,
Dordrecht, p 601
49. Leckie F, Scragg AH and Cliffe KC (1990) In: Progress in plant cellular and molecular
biology. Proc VIIth Internat Congr Plant Tiss Cell Cult, Amsterdam, 24 29 June 1990,
Kluwer Academic, Dordrecht, p 689
50. Chisti MY (1989) Airlift bioreactors, Elsevier, London
51. Nagata S (1975) Mixing: principles and applications, John Wiley, New York
52. Bello RA, Robinson CW and Moo-Young M (1984) Can J Chem Engn 62:573
53. Holland FA and Chapman FS (1966) Liquid mixing and processing in stirred tanks.
Reinhold, New York
54. Metzner AB and Taylor JS (1960) AIChE J 6:109
55. Taguchi H (1971) Adv Biochem Eng 1:1
56. Johnson DN and Hubbard DW (1974) Biotechnol Bioeng 16:1283
57. Norwood KW and Metzner AB (1960) AIChE J 6:432
58. Wang DIC and Fewkes RCJ (1977) In: Development in industrial microbiology, vol
18. Proc 33rd Gen Meet Soc Ind Microbiol, Jekyll Island, Georgia, USA, Aug 14-20
1976, p 39
59. Blakebrough N and Sambamurthy K (1966) Biotechnol Bioeng 8:25
60. Calderbank PH and Moo-Young MB (1959) Trans IChE 37:26
61. Rousseau I and Bu'Lock JD (1980) Biotechnol Lett 2:475
62. Weiland P (1984) Get Chem Eng 7:374
63. Chisti MY and Moo-Young M (1987) Chem Eng Comm 60:195
64. Chisti MY, Halard B and Moo-Young M (1988) Chem Eng Sci 43:451
65. Chakravarty M, Singh HD, Baruah JN and Iyengar MS (1974) Indian Chem
Engr 16:17
66. Popovic M and Robinson CW (1988) Biotechnol Bioeng 32:301
67. Nishikawa M, Kato H and Hashimoto K (1977) Ind Eng Chem Process Des Dev 16:133
68. Allen DG and Robinson CW (1989) Biotechnol Bioeng 34:731
69. Kessell RHJ and Carr AH (1972) J Exp Bot 23:996
70. Payne GF, Shuler ML and Brodelius P (1987) In: Lydersen BK (ed) Large scale cell
culture technology. Hanser, Munich, p 193
Design of Reactors for Plant Cells and Organs 165

71. Kato A, Shimizu Y and Nagai S (1975) J Ferment Technol 53:744

72. Pareilleux A and Vinas R (1983) J Ferment Technol 61:429
73. Mavituna F (1988) In: Pais MSS, Mavituna F, Novais JM (eds) Plant cell biotechnology.
NATO ASI Series H18, Springer, Berlin, Heidelberg, New York, p 1
74. Kobayashi Y, Fukui H and Tabata M (1989) Plant Cell Rep. 8:255
75. van't Riet K (1979) Ind Eng Chem Process Des Dev 18:357
76. Schfigerl K (1981) Adv Biochem Eng 19:71
77. Deindoerfer FH and Gaden EL (1955) Appl Microbiol 3:253
78. Yagi H and Yoshida F (1975) Ind Eng Chem Process Des Dev 14:488
79. Perez JF and Sandall OC (1974) AIChE J 20:770
80. Chisti MY and Moo-Young M (1988) Chem Eng J 38:149
81. Bello RA, Robinson CW and Moo-Young M (1985) Biotechnol Bioeng 28:369
82. Kawase Y and Moo-Young M (1986) Chem Eng Commun 40:67
83. Chisti MY, Fujimoto K and Moo-Young M (1987) In: Ho CS and Oldshue
JY (eds) Biotechnology processes - scale-up and mixing. AIChE, New York, p 72
84. Popovic MK and Robinson CW (1989) AIChE J 35:393
85. Oldshue JY (1966) Biotechnol Bioeng 8:3
86. Wichterle K, Kadlec M, Zak L and Mitschka P (1984) Chem Eng Commun
26:25
87. Henzler H-J (1980) Chem Ing Tech 52:643
88. Tramper J, Joustra D and Vlak JM (1987) In: Webb C and Mavituna F (eds) Plant
and animal cells: process possibilities. Ellis Horwood, Chichester, p 125
89. van't Riet K and Tramper H (1991) Basic bioreactor design. Marcel Dekker, New York
90. Kossen NWF (1984) Proc 3rd Eur Fed Biotechnol Congr, 10-14 September 1984,
Munich, vol 4, Weinheim VCH, p 257
91. Chisti MY and Moo-Young M (1988) Biotechnol Bioeng 31:487
92. Townsley PM, Webster F, Kutney JP, Salisbury P, Hewitt G, Kawamura N, Choi L,
Kurihara T and Jacoli GG (1983) Biotechnol Lett 5:13
93. Drapeau D, Blanch HW and Wilke CR (1986) Biotechnol Bioeng 28:1555
94. Scragg AH and Fowler MW (1985) In: Vasil IK (ed) Cell culture and somatic cell
genetics of plants, vol 2. Academic, Orlando, p 103
95. Wang C-J and Staba EJ (1963) J Pharm Sci 52:1058
96. Bond PA, Hegarty P and Scragg AH (1987) Proc 4th Eur Congr Biotechnol 1987, vol
2. Neijssel OM, van der Meer RR and Luyben KChAM (eds), Elsevier, Amsterdam,
p 440
97. Kawase Y and Moo-Young M (1990) Bioprocess Eng 5:169
98. Bourne JR and Butler H (1965) AIChE-IChE Symp Ser I0:89
99. Ulbrich B, Wiesner W and Arens H (1985) In: Neumann K-H, Barz W and Reinhard
E (eds) Primary and secondary metabolism in plant cell cultures. Springer, Berlin,
Heidelberg, New York, p 293
100. Tanaka H, Nishijima F, Suwa M and Iwamoto T (1983) Biotechnol Bioeng 25:2359
101. Brodelius P, Deus B, Mosbach K and Zenk MH (1979) FEBS Lett 103:93
102. Alfermann AW, Schuller I and Reinhard E (1980) Planta Med 40:218
103. Veliky IA and Jones A (1981) Biotechnol Lett 3:551
104. Furuya T, Yoshikawa T and Taira M (1984) Phytochem 23:999
105. Jirku V, Macek T, Vanek T, Krumphanzl V and Kubanek V (1981) Biotechnol Lett
3 : 447
106. Lindsey K, Yeoman MM, Black GM and Mavituna F (1983) FEBS Lett 155:143
107. Shuler ML, Sahai OP and Hallsby GA (1983) Ann NY Acad Sci 413:373
108. Lindsey K and Yeoman MM (1983) In: Mantell SH and Smith H (eds) Plant
biotechnology. Cambridge University Press, Cambridge, p 39
109. Robins RJ, Parr AJ, Richards SR and Rhodes MJC (1986) In: Morris P, Scragg AH,
Stafford A and Fowler MW (eds) Secondary metabolism in plant cell cultures,
Cambridge University Press, Cambridge, p 162
110. Brodelius P (1984) In: Vasil IK (ed) Cell culture and somatic cell genetics of plants,
vol 1. Academic, New York, p 535
166 P.M. Doran

111. Brodelius P (1986) In: Evans DA, Sharp WR and Ammirato PV (eds) Handbook of
plant cell culture, vol 4. Macmillan, New York, p 287
112. Novais JM (1988) In: Pais MSS, Mavituna F and Novais JM (eds) Plant cell
biotechnology. NATO ASI Set H18, Springer, Berlin, Heidelberg, New York, p 353
113. Rosevear A (1988) In: King RD and Cheetham PSJ (eds) Food biotechnology, vol 2.
Elsevier, London, p 83
114. Hulst AC and Tramper J (1989) Enzyme Microb Technol 11:546
115. Brodelius P and Nilsson K (1980) FEBS Lett 122:312
116. Ketel DH, Hulst AC, Gruppen H, Breteler H and Tramper J (1987) Enzyme Microb
Technol 9:303
117. Hulst AC, Tramper J, Brodelius P, Eijkenboom LJC and Luyben KChAM (1985) J
Chem Tech Biotechnol 35B: 198
118. Ayabe S-I, Iida K and Furuya T (1986) Phytochem 25:2803
119. Prenosil JE, Hegglin M, Baumann TW, Frischknecht PM, Kappeler AW, Brodelius P
and Haldimann D (1987) Enzyme Microb Technol 9:450
120. Holden MA, Hall RD, Lindsey K and Yeoman MM (1987) In: Webb C and Mavituna
F (eds) Plant and animal cells: process possibilities. Ellis Horw0od, Chichester, p 45
121. Ishida BK (1988) Plant Cell Rep 7:270
122. Barnabas NJ and David SB (1988) Biotechnol Lett 10:593
123. Brodelius P and Nilsson K (1983) Eur J Appl Microbiol Biotechnol 17:275
124. Robins RJ and Rhodes MJC (1986) Appl Microbiol Biotechnol 24:35
125. Mavituna F, Wilkinson AK, Williams PD and Park JM (1987) In: Moody GW and
Baker PB (eds) Bioreactors and biotransformations. Elsevier, London, p 26
126. Kim DJ and Chang HN (1990) Biotechnol Bioeng 36:460
127. Ford JR, Lambert AH, Cohen W and Chambers RP (1972) Biotechnol Bioeng Syrup
3: 267
128. Hulst AC, Meyer MMT, Breteler H and Tramper J (1989) Appl Microbiot Biotechnol
30:18
129. Mavituna F, Park JM, Wilkinson AK and Williams PD (1987) In: Webb C and
Mavituna F (eds) Plant and animal cells: process possibilities. Ellis Horwood,
Chichester, p 92
130. Furusaki S, Nozawa T, Isohara T and Furuya T (1988) Appl Microbiol Biotechnol
29:437
131. Pras N, Hesselink PGM, Guikema WM and Malingr6 TM (1989) Biotechnol Bioeng
33:1461
132. Kargi F, Ganapathi B and Maricic K (1990) Biotechnol Prog 6:243
133. Fuller KW (1984) Chem Ind 3 Dec, p 825
134. Dainty AL, Goulding KH, Robinson PK, Simpkins I and Trevan MD (1985) Trends
in Biotechnol 3:59
135. Hoekstra SS, Harkes PAA, Verpoorte R and Libbenga KR (1990) Plant Cell Rep 8: 571
136. Tsoulpha P and Doran PM (1991) J Biotechnol 19:99
137. Fuller KW and Bartlett DJ (1982) PCT Patent Application PCT/GB81/00186
138. Panda AK, Mishra S, Bisaria VS and Bhojwani SS (1989) Enzyme Microb Technol
11:386
139. Kobayashi Y, Fukui H and Tabata M (1987) Plant Cell Rep 6:185
140. Kobayashi Y, Fukui H and Tabata M (1988) Plant Cell Rep 7:249
141. Archambault J, Volesky B and Kurz WGW (1989) Biotechnol Bioeng 33:293
142. Archambault J, Volesky B and Kurz WGW (1990) Biotechnol Bioeng 35:702
143. Kurz WGW, Paiva NL and Tyler RT (1990) In: Progress in plant cellular and molecular
biology. Proc VIIth Internat Congr Plant Tiss Cell Cult, Amsterdam, 24-29 June 1990,
Kluwer Academic, Dordrecht, p 682
144. Tramper J, Wolters I and Verlaan P (1987) In: Laane C, Tramper J and Lilly MD
(eds) Biocatalysis in organic media. Elsevier, Amsterdam, p 311
145. Buitelaar RM, Susaeta I and Tramper J (1990) In: Progress in plant cellular and
molecular biology. Proc VIIth Internat. Congr Plant Tiss Cell Cult, Amsterdam, 24-29
June 1990, Kluwer Academic, Dordrecht, p 694
Design of Reactors for Plant Cells and Organs 167

146. Haldimann D and Brodelius P (1987) Phytochem 26:1431

147. Brodelius P (1984) Ann NY Acad Sci 434:382
148. Vanek T and Macek T (1989) Planta Med 55:680
149. Doran PM and Bailey JE (1986) Biotechnol Bioeng 28:73
150. Bhandary SBR, Collin HA, Thomas E and Street HE (1969) Ann Bot 33:647
151. Tabata M, Yamamoto H, Hiraoka N and Konoshima M (1972) Phytochem 11:949
152. Flores HE (1987) In: LeBaron HM, Mumma RO, Honeycutt RC and Duesing JH
(eds) Biotechnology in agricultural chemistry, ACS Symp Series 334, American
Chemical Society, Washington DC, p 66
153. Aird ELH, Hamill JD and Rhodes MJC (1988) Plant Cell Tiss Organ Cult 15:47
154. Parr AJ and Hamill JD (1987) Phytochem 26:3241
155. Jung G and Tepfer D (1987) Plant Sci 50:145
156. Christen P, Roberts MF, Phillipson JD and Evans WC (1989) Plant Cell Rep 8:75
157. White PR (1934) Plant Physiol 9:585
158. Rhodes MJC, Robins RJ, Hamill JD, Parr AJ and Walton NJ (1987) Internat Assoc
Plant Tiss Cult Newsletter 53, Nov 1, p 2
159. Hamill JD, Parr AJ, Rhodes MJC, Robins RJ and Walton NJ (1987) Bio/Technol 5: 800
160. Doran PM (1989) Aust J Biotechnol 3:270
161. Signs MW and Flores HE (1990) BioEssay 12:7
162. Taya M, Yoyama A, Kondo O, Kobayashi T and Matsui C (1989) J Chem Eng Japan
22:84
163. Sharp JM and Doran PM (1990) J Biotechnol 16:171
164. Rhodes MJC, Hilton M, Parr AJ, Hamill JD and Robins RJ (1986) Biotechnol Lett 8:415
165. Wilson PDG, Hilton MG, Robins RJ and Rhodes MJC (1987) In: Moody GW and
Baker PB (eds) Bioreactors and biotransformations. Elsevier, London, p 38
166. Davioud E, Kan C, Hamon J, Temp6 J and Husson H-P (1989) Phytochem 28:2675
167. Kondo O, Honda H, Taya M and Kobayashi T (1989) Appl Microbiol Biotechnol 32:
291
168. Hilton MG and Rhodes MJC (1990) Appl Microbiol Biotechnol 33:132
169. Wilson PDG, Hilton MG, Meehan PTH, Waspe CR and Rhodes MJC (1990) In:
Progress in plant cellular and molecular biology. Proc VIIth Internat Congr Plant Tiss
Cell Cult, Amsterdam, 24-29 June 1990, Kluwer Academic, Dordrecht, p 700
170. Wilson PDG, Hilton MG, Steer DC, Waspe CR, Robins RJ and Rhodes MJC (1989)
PCT Patent Application PCT/GB89/00510
171. Konar RN, Thomas E and Street HE (1972) Ann Bot 36:249
172. A1-Abta S and Collin HA (1978) New Phytol 80:517
173. Kuberski C, Scheibner H, Steup C, Diettrich B and Luckner M (1984) Phytochem 23:
1407
174. Styer DJ (1985) In: Henke RR, Hughes KW, Constantin MJ and Hollaender A (eds)
Tissue culture in forestry and agriculture. Plenum, New York, p 117
175. Cazzulino DL, Pedersen H, Chin C-K and Styer D (1990) Biotechnol Bioeng 35:781
176. Luckner M and Diettrich B (1988) In: Constabel F and Vasil IK (eds) Cell culture
and somatic ceil genetics of plants, vol 5. Academic, San Diego, p 193
177. Redenbaugh K, Paasch BD, Nichol JW, Kossler ME, Viss PR and Walker KA (1986)
Bio/Technol 4:797
178. A1-Abta S, Galpin IJ and Collin HA (1979) Plant Sci Lett 16:129
179. Scheibner H, Bj6rk L, Schulz U, Diettrich B and Luckner M (1987) J Plant Physiol
130:211
180. Chen THH, Thompson BG and Gerson DF (1987) J Ferment Technol 65:353
181. Greidziak N, Diettrich B and Luckner M (1990) Planta Med 56:175
182. Macek T, Vanek T and Binarova P (1989) Planta Med 55:595
183. Takayama S and Misawa M (1981) Plant Cell Physiol 22:461
184. Park JM, Hu W-S and Staba EJ (1989) Biotechnol Bioeng 34:1209
185. Heble MR (1985) In: Neumann K-H, Barz W and Reinhard E (eds) Primary and
secondary metabolism of plant cell cultures. Springer, Berlin, Heidelberg, New York,
p 281
168 P.M. Doran: Design of Reactors for Plant Cells and Organs

186. Hagimori M, Matsumoto T and Mikami Y (1984) Agric Biol Chem 48:965
187. Hfisemann W, Amino S, Fischer K, Herzbeck H and Callis R (1990) In: Progress in
plant cellular and molecular biology. Proc VIIth Internat Congr Plant Tiss Cell Cult,
Amsterdam, 24-29 June 1990, Kluwer Academic, Dordrecht, p 373
188. Hiisemann W (1985) In: Vasil IK (ed) Cell culture and somatic cell genetics of plants,
vol 2. Academic, Orlando, p 213
189. Hfisemann W, Fischer K, Mittelbach I, Hfibner S, Richter G and Barz W (1989) In:
Kurz WGW (ed) Primary and secondary metabolism of plant cell cultures II. Springer,
Berlin, Heidelberg, New York, p 35
190. Widbolm JM (1989) In: Kurz WGW (ed) Primary and secondary metabolism of plant
cell cultures II. Springer, Berlin, Heidelberg, New York, p 3
191. Flores HE, Hoy MW and Pickard JJ (1987) Trends in biotechnol. 5:64
192. Hfisemann W and Barz W (1977) Physiol Plant 40:77
193. H/isemann W (1988) In: Pais MSS, Mavituna F and Novais JM (eds) Plant cell
biotechnology. NATO ASI Ser H18, Springer, Berlin, Heidelberg, New York, p 179
194. Bender L, Kumar A and Neumann K-H (1985) In: Neumann K-H, Barz W and
Reinhard E (eds) Primary and secondary metabolism of plant cell cultures. Springer,
Berlin, Heidelberg, New York, p 24
195. Wink M and Hartmann T (1980) Planta Med 40:149
196. Igbavboa U, Sieweke H-J, Leistner E, R6wer I, Hfisemann W and Barz W (1985)
Planta 166:537
197. Emke A and Eilert U (1986) Plant Cell Rep 5:31
198. Conner AJ (1987) Phytochem 26:2749
199. Ikemeyer D and Barz W (1989) Plant Cell Rep 8:479
200. Barz W, Ikemeyer D and Beimen A (1989) Planta Med 55:594
201. Hagimori M, Matsumoto T and Mikami Y (1984) Plant Cell Physiol 25:1099
202. Tyler RT, Kurz WGW and Panchuk BD (1986) Plant Cell Rep 3:195
203. Dalton CC (1980) J Exp Bot 31:791
204. Dalton CC and Peel E (1983) Prog Ind Microbiol 17:109
205. Yamada Y, Imaizumi K, Sato F and Yasuda T (1981) Plant Cell Physiol 22:917
206. Peel E (1982) Plant Sci Lett 24:147
207. H/isemann W (1982) Protoplasma 113:214
208. H/isemann W (1983) Plant Cell Rep 2:59
209. Fischer U and Alfermann AW (1989) Planta Med 55:686
210. Ohta Y and Takata T (1990) In: Abstracts VIIth Internat Congr Plant Tiss Cell Cult,
Amsterdam, 24 29 June 1990, p 341
211. Schumpe A and Deckwer W-D (1987) Bioprocess Eng 2:79
212. E1-Temtamy SA, Khalil SA, Nour-E1-Din AA and Gaber A (1984) Appl Microbiol
Biotechnol 19:376
213. Henzler H-J and Kauling J (1985) In: Proc 5th Eur Conf on Mixing, Wiirzburg,
Germany, 10 12 June 1985, p 303
Expert Systems in Bioprocess Control:
Requisite Features

K o n s t a n t i n B. K o n s t a n t i n o v 1., R o b e r t A a r t s 2 a n d T o s h i o m i Y o s h i d a 1

1 International Center of Biotechnology, Faculty of Engineering, Osaka University, Yamada

oka, Suita shi, Osaka 565, Japan
2 VTT, Biotechuical Laboratory, P.O. Box 202, SF-02 151 Espoo, Finland

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
2 Basic Structures, Functions and Implementation Schemes . . . . . . . . . . . . . . . . . . . . . . 170
2.1 Structures of Systems for Expert Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
2.2 Functions of the Knowledge-Based Module of the Control System . . . . . . . . . . . 172
2.3 Implementation Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3 Requisite Features of Expert System Development Tools . . . . . . . . . . . . . . . . . . . . . . 175
3.1 General Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
3.1.1 Real-Time Capabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
3.1.2 Temporal Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.1.3 Integration with External Software Modules . . . . . . . . . . . . . . . . . . . . . . . . . 180
3.1.4 Mechanisms for Knowledge Structuring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
3.1.5 Handling of Various Categories and Levels of Knowledge . . . . . . . . . . . . . 182
3.1.6 Efficient Knowledge Debugging and Integrating . . . . . . . . . . . . . . . . . . . . . . 184
3.1.7 Answering Users' Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
3.1.8 Advanced User Interface with Graphical Capabilities . . . . . . . . . . . . . . . . . 185
3.1.9 Other Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
3.2 Specific Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
3.2.1 Limited "Width" and Enhanced "Depth" . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
3.2.2 Handling of Uncertain, Incomplete and Fuzzy Information . . . . . . . . . . . . 187
3.2.3 Orientation Towards the Processing of Continuous Variables . . . . . . . . . . 187
3.2.4 Capabilities of Intelligent Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
3.2.5 Availability of C o m m o n "Bioprocess" Knowledge Base . . . . . . . . . . . . . . . 188
3.2.6 Low Prototyping Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
3.2.7 Low Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
5 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190

The development of intelligent control in biotechnical processes, which only a decade ago
was considered an exciting, but obscure vision, has today become an area of intensive and
realistic research. One of the keys to success in this field is the selection of an adequate
software tool for building the intelligent system. The ideal tool must possess a large set of
features, which reflect both the real-time nature of the control problem, and the peculiarities
of the biotechnical system itself. Here, some of these features are introduced. In addition,
the main concepts and trends in the field of expert control in biotechnical processes are
discussed.

* To whom all correspondence should be addressed. Present address: Department of

Chemical Engineering, University of Delaware, DE 19716, USA.

Advancesin BiochemicalEngineering
Biotechnology,Vol. 48
Managing Editor: A. Fiechter
9 Springer-VerlagBerlinHeidelberg 1993
170 K.B. Konstantinov, R. Aarts and T. Yoshida

1 Introduction
Because of the focus on formal analysis and synthesis, many other aspects of
control system design have largely been disregarded in various application areas
[1]. This has been the case with bioprocesses control, even though the limitations
of the traditional control approaches in the field of biological systems have long
been perceived. Nowadays, it is becoming clear that conventional control theory
alone cannot provide the required platform for building high-performance systems
for the control of bioprocesses. Consequently, a trend towards exploitation of new
methods capable of manipulating and utilizing informal knowledge of the biological
plant has emerged. Recent achievements in expert system (ES) technology have
already stimulated research aimed at its application in the field of bioprocesses
control [2].
However, when turning to the ES approach, the bioprocess engineer is likely
to encounter a number of unexpected difficulties, which arise from the still limited
experience in the field of expert control, the fundamental differences from traditional
control methodology, and the development of the software system itself. One of
the tasks which must be solved at an early stage of a project is the selection of
an appropriate expert system development tool (ESDT), also called an "expert
shell". This is crucial to the overall design process; an unsuitable tool may
predestine the research to fruitless effort for years, while selection of the right one
provides a reliable basis for rapid success. Unfortunately, the significance of this
problem is still not sufficiently realized. There is a tendency for people to think
of ESs as one thing, and to consider that any ESDT will suit their particular case
[3]. In fact, current ES technology is leading to the creation of hundreds of products,
which vary tremendously in size, purpose, capability and price. Since a commercial
ESDT specialized for control of bioprocesses has not yet been produced, developers
should make their choice only after careful screening of the tools currently available
on the market. The product selected must provide a large set of features, many
of which are non-trivial and are not supported by the more common ESDTs.
These reflect both the real-time nature of the control problem, and the peculiarities
of the biotechnical system itself.
The principal purpose o f this paper is to introduce a set of ESDT features which
are required in building intelligent systems for the control of bioprocesses. In
addition, the basic structure, functions, and current limitations of ES technology
in the context of problems encountered in the control of bioprocesses, are discussed.

2 Basic Structures, Functions and Implementation Schemes

2.1 Structures of Systems for Expert Control

There are many ways to incorporate expert knowledge into a process control
system. Nevertheless, almost all possible cases can be reduced to two basic schemes
for expert control, referred to as direct and indirect [1]. Direct systems are those
Expert Systems in Bioprocess Control: Requisite Features 171

Control system
On-off 1
I ~t~ntrr~llArJ
Ysp _u_> Biotechnical
process
Y
>

Fig. 1. Structure of a system for direct expert control. The knowledge-based (KB) controller
works at the level of the standard controllers, such as PID or "on-off'

whose knowledge-based modules are involved in the control loop (Fig. 1). These
modules operate at the level of standard P I D controllers, and are useful in realizing
more complicated, nonlinear control algorithms. Such modules are known mostly
as fuzzy controllers, because they are often based on fuzzy logic. Recent
achievements in neural networks provide another alternative for the design of
such types of controllers. Generally, the direct expert control scheme, though
useful in some cases, is limited to local and low-level problems. This means that
although its applicability to the control of bioprocesses is possible, it is of limited
significance.
The structure of a system for indirect expert control (also called "supervisory")
is shown in Fig. 2 [4-6]. This is composed of two hierarchical levels providing
clear distribution of the system functions. The standard set of control tasks, such
as measurement, filtration, data acquisition, control, etc., are entrusted to the
lower level, which represents just a conventional control system. Such systems

U IKn~ >
~I supervisory | . Y
| system J =
[__ -r

Conventional 1 Y
control -
system l" O
_J
/
Fig. 2. Structure of a system for indirect

u[ Biotechnical
process
(supervisory) expert control. The knowled-
ge-based module is not involved in the
low-level control loop; instead it supervises
the lower control level by issuing high level
commands
172 K.B. Konstantinov, R. Aarts and T. Yoshida

are included in almost all modern biotechnical equipment, either in the form of
hardware controllers, Or as computer-implemented algorithms. The higher level
represents a knowledge-based superstructure to the conventional control part. In
contrast to the direct scheme, the knowledge-based module, which is indeed a
complete ES, is not explicitly involved in the low-level control (although exceptions
are sometimes possible); it does not generate signals to the control plant, but
helps the lower level to perform its job better. To this end, the higher level issues
supervisory commands which tell the lower level when to do what [7]. As the
techniques for design of low level controllers are relatively well established, the
synthesis of the conventional part of the system is not likely to cause special
difficulties. Thus, the success of the overall system development is dependent
mainly on the design of the knowledge-based module using an appropriate ESDT.
The indirect expert control concept offers new possibilities for development of
high-performance intelligent systems for the control of bioprocesses. This structure
flexibly combines the advantages of the traditional approach with those of ES
technology. It allows enhancement of the control system by the capability of
intelligent decision-making based on informal interpretation of the complex
behavior of the living system. If properly implemented, this structure will be
capable of covering various control problems which usually remain outside the
scope of conventional systems. The subsequent discussion will focus on systems
for indirect expert control.

2.2 Functions of the Knowledge-Based Module of the Control System

The features of the ESDT necessary for building the knowledge-based module are
correlated to the functions which this module is expected to perform in real-time.
Although considerable differences are possible according to the application, these
functions can be summarized into four main groups:
- Identification of the state of the cell population. This includes continuous
on-line evaluation of the physiological state of the cell population, informal
interpretation of cell behavior, prediction of future states, detection and diagnosis
of expected (e.g. stage transfers) or unexpected (e.g. deviations from normal
behavior) physiological phenomena, and others. Undoubtedly, this is the most
important, difficult and advanced function of the higher system level [8, 9].
- Identification of the state of the process equipment. The major problem here
is the detection and diagnosis of instrumental failures, such as troubles with sensors
or actuators. It has been pointed out that this function must be explicitly separated
from the previous one [10]. Although this may be not easy in case of biological
plants, it will contribute to the clear modularization of the knowledge base of the ES.
- Supervision of the conventional control part. Identification of the state of the
cell population and the process equipment are passive procedures creating
declarative information on the plant. To become useful, this must be mapped into
high level commands for supervision and synchronization of the work of the
conventional control part. The major purpose is to achieve intelligent handling
and control of the physiological phenomena during the process. Typical supervisory
Expert Systems in Bioprocess Control: Requisite Features 173

commands are activation/deactivation of control loops (changes of the control

strategy), modification of control parameters, changes of setpoints, etc. Particular
attention should be paid to the strategy-switching capability, which is needed to
respond to structural transformations in biological systems [11, 12]. In the case
of drastic physiological transfers, the knowledge-based part may need to react
by dynamic replanning of the process. Such advanced capability cannot be achieved
by the enumerated supervisory command; it requires additional internal seg-
mentalization of the knowledge-based module, and establishment of hierarchical
relationships between its parts [13].
Supervisions of the system by high level commands is necessary to handle
detected instrumental failures. The purpose is to compensate as much as possible
for any eventual damage, and enable the system to maintain safe operation under
abnormal conditions, that is, to ensure robust fault-tolerant behavior.
A similar function of growing importance is the intelligent monitoring of the
biotechnical plant. This is necessary because of the complexity of modern
measurement equipment, which requires special supervision. There is probably no
better recourse than to entrust this task to the knowledge-based module, thus
providing coherent functioning of all the system components.
- Advanced communication with the user. The knowledge-based module should
be able to represent all the information about the process, as well as to explain
its decisions and activities in a convenient form for the user. Indeed, this function
is mandatory for every ES, but the real-time constraints impose some more specific
requirements.

2.3 Implementation Schemes

According to the method of coupling the knowledge-based part and the conven-
tional control part, two types of implementation scheme are possible:
- Interface scheme. According to this, the knowledge-based part resides on a
dedicated computer, interfaced to the conventional control part (Fig. 3). Both parts
are software and hardware independent of each other. Typically, the higher level
is purchased (as an ESDT), developed and added separately, after the conventional
part has been set in operation.
Today, due to lack of other realistic alternatives, the interfaced scheme is the
standard solution of the bioprocess control problem [4, 14, 15]. This is because
the ESDTs available so far are large programs designed with the interfaced
structure in mind, meant to run on stand-alone computers. Furthermore, modern
bioreactors are equipped with computerized low-level controllers which perform
some of the tasks of the conventional control module, and are easy to link to the ES.
Though popular, an interfaced architecture has several shortcomings: the
software of both system parts is provided by different suppliers, which hampers
integration; communication between computers may cause problems; the informa-
tion storage and the man-machine interface of the system is redundant [10, 16].
- Embedded scheme. In this case, the knowledge-based module is embedded
into the real-time control environment, co-existing with the conventional control
174 K.B. Konstantinov, R. Aarts and T, Yoshida

~Knowledge-based~| "~
supervisory ~/ --~
system~/

Conventional I -~
control ~ y $
system o~
,_1

Fig. 3. Interfaced scheme. The knowledge-ba-

sed supervisory module is implemented as an
u J Biotechnical independent expert system on a separated
-~ process computer interfaced to the lower level control
system

jf
7
/ / now[edge-base
/ >[ supervisory |4
i I ] modu{e I
f

|~Integrated (hybrid)l| I I . . . .
/I contro[systern 1/ ~---1 t'~
t,.. . . . ~2~ ~l j I
~ 1 INIl ~ I.~l---~- - - ~ -

Biotechnical
process I

Fig. 4. Embedded scheme. The knowledge-based supervisory module coexists with the
conventional control system on the same computer. The knowledge-based module is
embedded into the real-time control environment
Expert Systems in Bioprocess Control: Requisite Features 175

algorithms on the same computer (Fig. 4). The integration between both parts is
much more efficient, and the problems of the interfaced scheme are overcome [17].
Although the embedded scheme is considered to be more advanced, its implementa-
tion is still difficult because the commercial ESDTs available today are not designed
for such integration [18].

3 Requisite Features of Expert System Development Tools

Some of the features which an ESDT used to control bioprocesses should possess
are common to all ESDTs designed for real-time applications, while others are
more specific, and peculiar to bioprocesses. Below, the general features are described
first.

3.1 General Features

3.1.1 Real-Time Capabilities

The real-time capabilities of the ESDT are critically important in process control.
They have considerable impact on all other system features. For this reason ESDTs
are classified into two groups: real-time and non real-time. In fact, the term
"real-time" does not represent a single feature, but a complex of several
characteristics:
- Fast response time. The speed of execution has always been subject to careful
consideration in computer control systems. This is particularly important for ESs,
which due to the processing of symbolic information, are known to be slow and
clumsy. The speed of the conventional ES is estimated as two to three orders of
magnitude lower than that required for real-time work [19].
The response time required in bioprocesses control is highly dependent on the
particular application. Generally, the speed limitations are not so rigid as in the
control of other technological processes, and will vary from few seconds (to handle
fast phenomena, e.g. DO jumps, instrumental failures) to several hours (to interpret
gradual physiological transfers). If the response time of an ESDT is not known,
it can be practically estimated using some simple procedures as described in the
literature [20].
Typically, high speed ESDTs are developed not in a symbolic language such
as Lisp or Prolog, but in a fast conventional language such as C. Considerable
reduction of the response time is achieved by using an ESDT which is based on
compilation of the knowledge bases from a textual into a more efficient compressed
form [3, 19]. However, as this procedure reduces the possibilities for on-line
knowledge editing, and ESDT combining both compilation and interpretation of
the knowledge base (called "incremental compiling" [21, 22]) appears to be the
best solution.
- Activation of rules according to the time. This means that the system should
be able to interpret instructions of the type
176 K.B. Konstantinov, R. Aarts and T. Yoshida

Run Rule 15 every 30 s

i.e. allow the attachment of a specific activation interval to every rule [23].
Furthermore, it is desirable to be able to change dynamically these time intervals
from withing the rule conclusions, for example 1

If (Process stage is State 2) and (DO is V E R Y L O W )

Then (Run Rule 15 every 5 s)

This can be useful in controlling bioprocesses because of their variable

dynamics. Time-driven rule activation also provides temporal grouping of rules,
which reduces the load of the interference mechanism; at any moment, only those
rules whose time interval expires need to be processed.
- Acting within a time limit (guaranteed response time). Without such a
mechanism the response time of the system is difficult to predict. It depends on
the size of the knowledge base, the number of concurrently running tasks, and
the availability of the data needed for the inference. If, for example, the data
from the auto-analyzer are not yet available, but the decision cannot be postponed
further, the system must be able to infer the best possible decision at the current
moment [24]. To this end, the interference mechanism should be provided with a
time-out mechanism which will interrupt the normal reasoning process when the
permitted time-interval expires, and will force urgent, probably sub-optimal,
decision making [18, 25].
- Cyclic continuous operation. Contrary to conventional ESs, which terminate
after completing the decision, the ESs used in process control should be able to
work continuously, in a never-terminating cycle [26]. Although this requirement
seems to be simple, it is not always satisfied. A problem arises because many of
the ESDTs are developed in symbolic language, usually Lisp. To prevent
accumulative memory consumption during a long run, this language is provided
with a so called "garbage collection" mechanism, which is automatically activated
in unpredictable time instances. When running, the garbage collector interrupts
for seconds, or even for minutes, all other operations, including the inference
mechanism [22], which is absolutely inadmissible in a real-time environment.
In some new Lisp compilers the traditional "mark and sweep" garbage collection
mechanism is replaced by a more advanced "incremental" type. An ESDT based
on such a compiler would not have a problem with the interruption of continuous
work.
There are several other characteristics, extensively discussed in the literature,
which are needed to complete the real-time profile of the ESDT [18, 23, 24]. Any
of them may be more or less important, dependent on the application. Unfor-
tunately, most of the ESDTs available today do not support such features, and
are not designed for real-time applications [19, 27]. There are, however, a few

1 For simplicity, the exemplary rules are written in natural language. In real cases, they
should conform to the ESDT syntax.
Expert Systems in Bioprocess Control: Requisite Features 177

products which do provide some real-time capabilities [28], and it is on these that
the attention of the developed should be entirely focused.
It is worth noting, that to a great extent the real-time capabilities of the ESDT
are an inheritance from those of the operating system; if the operating system
is not real-time, the ESDT will lack real-time features too. Therefore, tools
designed for non-real time operating systems (e.g. DOS), are basically in-
appropriate for control applications. From this viewpoint, the most suitable
ESDTs are those working in real-time UNIX (or UNIX-like) environments [29].

3.1.2 Temporal Reasoning

This is the capability of the ES to reason about discrete or continuous events and
phenomena, accounting for the time dimension [30]. Temporal reasoning is closely
related to the real-time features of the system. In order to emphasize the importance
of temporal reasoning, it is discussed separately. Temporal reasoning has different
aspects, some of which can be extremely useful in bioprocesses control:
- Reasoning about the history of continuous process variables [30]. The ES
must have the ability to detect patterns, e.g. constraint violations, trends, or shapes,
in specified pieces of historical data. This is illustrated by the template rule

If (Time-period Variable Descriptor)

Then ( . . . )

where Time-period defines a period from the process history (often the most recent
one) either explicitly, or in respect to a certain event in the past, Variable is a
particular process variable, and Descriptor represents the expected pattern. The rule

If (SGR is NEGATIVE) and (Since entering Phase 3 RQ > 0.9)

Then (Deactivate the glucose feeding algorithm)

illustrates a simple form of reasoning in respect to the current value of the specific
growth rate SGR and the recent history of the respiration quotient RQ. A slightly
more complicated case is represented by the rule

If (The SGR has been DECREASING for more than 1 h)

Then (The process is entering the LATE GROWTH STAGE)

whose condition considers the recent trend of the SGR. There are many cases
when the information from a biotechnical system can be adequately interpreted
by accounting for the trends of variables. However, this is not always sufficient
to analyze the behavior of bioprocesses in which the temporal shapes of the
variables provide indispensable information about the underlying physiological
phenomena. To achieve better control, these shapes must also be properly
interpreted. This would be quite probable because the decisions of a human
operator are based both on the current values and the recent historical profiles
of the process variables. The rule
178 K.B. Konstantinov, R. Aarts and T. Yoshida

0.10 I
i

I
0.08

0.06
r I
I
0.04 I
n.. I

I
0.02 I
I. At
I
I i
I i
t o- 2 t o- 1 to
Time (h)

Fig. 5. Characteristic time profile of the Ra/g indicating intensive excretion of acetic acid
during cultivation of Escherichia coli

If (During the last 2 h R,/o has INCREASED CONCA VEL I1)

Then (Report." Intensive excretion of acetic acid)

represents such a case. It accounts for the shape of the variable Ra/g (ratio of the
ammonia feed rate to the glucose uptake rate), and will execute its conclusion if
this shape matches the descriptor "'INCREASED CONCAVELY" (Fig. 5). The
shape descriptors may be different, e.g. "PASSED 0 VER MAXIMUM", "STOP-
PED INCREASING", "STOPPED DECREASING", "DECREASING MONO-
TONICALLY", "OSCILLATING", possibly stored in a user-extendible library
of shapes which the ES can recognize [10, 30, 31, 32, 33].
Another useful form of reasoning about the history of variables utilizes the
data over the given time period for calculation of certain statistical markers, and
decision making based on their values [16]:

if (During the last 5 min the OD STANDARD DEVIATION is outside the

interval [0.1, 2.0])
Then (Report." Problem with the OD sensor)

The corresponding patterns are shown in Fig. 6.

- Reasoning about past events and phenomena. Except for continuous process
variables, reasoning in respect to time can be applied to some events and
phenomena which have been registered by the ES in the past, and whose occurrence
may affect the decision in the current situation. The following example illustrates
a simple case:
Expert Systems in Bioprocess Control: Requisite Features 179

100

\
9..-. 50
a
o
At

a to- 5 to
Time (rain)
100[ I
I
I
I

Fig. 6a, b. Two different patterns in-

dicating failure of the OD sensor: a
~t the normal noise level suddenly drops
i-
to zero; b the noise level increase
abruptly
t o- 5 to
b Time (min)

If (The time after beginning of continuousfeeding >5 min ) and(DO is HIGH)

Then (Increase the GFR by A GFR)

where GFR is the glucose feed rate (see Fig. 7). To process this rule, the system
should remember when the continuous glucose feeding began, i.e. to "time-stamp"

~.50- o100I
I
I
I
/DO I GFR

g2s
,,

I
I
O- 01 I
t1- 5 t1
Time (min)
Fig. 7. Typical time profiles of the DO and the GFR upon transfer from batch to continuous
glucose-limited cultivation. If the initial GFR is not sufficient, DO will remain high, which
can be used to correct the GFR
180 K.B. Konstantinov, R. Aarts and T. Yoshida

this event. Such time-stamping of all detected events is another typical feature of
the real-time ES, which always stores facts in the knowledge-bases together with
the time of their appearance.
A slightly different form of temporal reasoning is represented by the introduction
of lower and upper time bounds (permitted interval) for expected events. The
system must then check whether or not the occurrence of the event fulfills these
time constraints, for example

If (Transfer Stage 1 --, Stage 2 is outside the interval [20-26] h)

Then (Report: The transfer is time-incompatible)

According to Lun and MacLeod [27], temporal reasoning mechanisms should

support the construction of more abstract relationships, using time predicates
such as BEFORE, AFTER, and DURING. This will facilitate reasoning about
the order of the events in time.
Although temporal reasoning functions appear natural and useful, their
realization raises many problems which have not yet been solved completely, even
in specialized real-time ESs [1].

3.1.3 Integration with External Software Modules

Figure 2 shows that the knowledge-based part of the control system is not an
isolated, independent module. To perform its function, it needs to exchange
information with the external environment. This means that the ESDT must be
provided with powerful communication mechanisms.
First, it should be integrated with the external software modules performing
the conventional control functions. As it is virtually impossible to provide the
ESDT originally with all algorithms needed in specific applications, the capability
of linking with user-defined algorithms is an absolute requirement [23]. It has
been reported that the power of the advanced shell G2, applied to bioprocess
control, can be greatly enhanced by interfacing to user-defined algorithmic
procedures for modeling, parameter identification, estimation of unmeasurable
variables, prediction, signal processing, on-line optimization, statistical data-
processing, and modeling of non-linear relationships by neural networks [14]. The
integration should be understood either as incorporation of external software
modules into the body of the ES, or as an establishment of communication channels
between the ES and these modules.
Second, the ES should be able to communicate with hardware devices using,
for example, serial lines. In other cases it may be necessary to retrieve or send
data directly to the low-level process interface, e.g. AD/DAC and DI/DO. Since
the communication with such devices is realized by software, the problem is again
reduced to integration of the ES with the user's programs.
Most of today's ESs are not designed to communicate with conventional
real-time software, but only with the operator through the keyboard [19], This
limitation is considered to be among the most serious reasons still preventing ES
technology from being adopted by many industries [34, 35]. Some new products
do, however, provide consistent integration capabilities using two basic mech-
Expert Systems in Bioprocess Control: Requisite Features 181

r Rules for identification of 1

g the cultivation stage
(always active)
w

0 Rules for detection, diagnosis and handling of

,v, stage-specific phenomena:
Rules active I I Rules active .~ [ Rules active
in Stage 1 in Stage 2 in Stagen

Rules for detection,

diagnosis and handling
of random phenomena Fig. 8. Typical structurization of the
(always active) knowledge base in the case of batch or
fed-batch cultivation

anisms: fixed-protocol messages, which support communication between the

ES and the external modules, and "software hooks", which are used for
incorporation of the user's code directly into the ES. The last one is much more
efficient, but unsafe because errors in the user's program may cause the ES to
crash. Deep fusion at the level of knowledge- and data-bases, resulting in a so
called "fully integrated" ES [16], is at present impossible.

3.1.4 Mechanisms for Knowledge Structuring

The knowledge of a bioprocess is not a homogeneous, plain bank of information.
It is highly structured according to certain functional criteria, or in respect to
time. For example, in the case of fed-batch cultivations, different chunks of
knowledge are required in each of its stages (Fig. 8), i.e. at any moment, only a
specific group(s) of rules should work. Application of rules which do not belong
to the current context might be undesirable, even dangerous. Similar considerations
hold true for continuous cultivations; however, the structuring will not have a
chronological character, but the groups will correspond to possible process
situations. To manipulate knowledge dynamically, the ESDT must provide
mechanisms for activation or deactivation of a particular group(s) of rules. It is
especially useful to do this from within the rule conclusions, for example:

(RQ is LO W) and (SGR is LO W) and (SGR is DECREASING)

Then (Report." Transfer to Stage_2)
Then (Activate Rule_group_2)
Then (Deactivate Rule group_l)

The explicit grouping of rules represents the simplest method of knowledge

structuring. Although in many cases it will work successfully, high-performance
ESDTs offer more sophisticated mechanisms for determining the currently active
182 K.B. Konstantinov, R. Aarts and T. Yoshida

set of rules. They are usually based on some associative criteria, for example
all rules which are related to a particular class of objects or category of problems
[24]. The capabilities of driving the inference engine to narrow its scope of
interest, and to use only part of the knowledge, are sometimes jointly referred to
as "focus-of-attention" feature.
Another consistent way to structure knowledge is to assign priorities to the
rules. The priority represents the importance of the particular rule; a rule will be
checked only after all other rules with higher priority have been processed. This
mechanism is useful because, by assigning different priorities to different groups
of rules, one part of the knowledge base can be forced to work before another [36].
Apart from the logical clarity, there are two important advantages of knowledge
structuring. First, there is a dramatic improvement in the performance speed of
the system due to the reduced number of rules which require processing at any
moment. Second, knowledge structuring simplifies the debugging of the knowledge
base.

3.1.5 Handling of Various Categories and Levels of Knowledge

A fundamental feature of ESDTs is their flexibility in knowledge representation.
This is crucial in the field of bioprocesses control, where various categories of
knowledge are involved, ranging from superficial experiences to fundamental
analytical models [37, 38]. Usually, the development of the control system is the
final task of multi-phase research, including genetic, biochemical, microbial and
behavioral study of organisms of interest. To take advantage of the information
accumulated at each of these stages, the ESDT must provide rich capabilities for
the representation and handling of diverse types of knowledge.
Although the ES approach is intuitively considered as heuristical, heuristics
usually represent just a part of the knowledge available. The ES capabilities should
not be limited only to representation of knowledge of this type. Knowledge of
the bioprocess may be available in the form of analytical models (differential and
difference equations, material and energy balances, kinetic models, stochiometric
equations), qualitative models, or as fundamental information about cell genetics
and metabolism. Such knowledge is represented in different forms, and the ES
should be able to handle any of them equally well. This implies a simple rationale:
when developing a new ES, the knowledge available should be analyzed first, the
involved knowledge categories clarified, and based on this, the ESDT selected [39].
In addition to form, knowledge of the process differs also in depth. ESs are
often criticized as being shallow models of their application domain, in the sense
that they draw conclusions directly f r o m the observed phenomena, without
penetrating the underlying mechanisms. Recently, it has become evident that ES
performance can be much improved by enrichment of this shallow knowledge by
deeper knowledge, which provides consistent interpretation of the events, based
on the underlying mechanisms, structures and dependencies [40, 41]. Often, deep
knowledge is represented in the form of mathematical models. These can be very
useful in solving tasks for the detection of sensor failures, deviations from expected
behavior, estimation of inaccessible variables, forecasting, planning, hypotheses
validation and knowledge-testing simulation [42]. From this viewpoint, ES
Expert Systems in Bioprocess Control: Requisite Features 183

technology should not be considered as a contradiction to modeling techniques;

instead, the significance of the analytical models in ESs is expected to become
more profound [43].
It would be incorrect to consider that deep knowledge can be represented only
by mathematical models. It may exist in other forms, such as qualitative models
[40, 44, 45], as linguistic models using English-like expressions [46], or even as set
of rules. As our knowledge of the process represents a continuum, rather than
clearly segregated layers, it is impossible to define a boundary between shallow
and deep knowledge; knowledge considered as deep for one purpose, appears to
be shallow for another. It is up to the system developer to decide the most
appropriate depth of description in the particular application, but in any case the
ES should not be the limiting factor in knowledge base development.
An important point to raise is that, in many cases, shallow knowledge is sufficient
to solve various control problems. Let us consider the rule

If (R~/gis HIGH)
Then (Reduce GFR by A GFR)

which illustrates how to detect and how to prevent the excretion of acetic acid
in a glucose-limited Escherichia coli cultivation [47]. This rule is a typical example
of shallow knowledge. Nevertheless, it can work (and it really does!) quite
satisfactorily. It represents, however, a '"shortcut" through knowledge, which
eliminates all intermediate steps, forming the causal link between the condition
and the conclusion. The underlying physiological phenomena remain hidden, and
the observed fact is mapped directly into the control decision. However, although
very fast and memory-efficient, such rules reduce the capabilities of the ES. They
treat the plant as a black-box, and restrict any possibility of explaining the observed
phenomena. The eliminated intermediate conclusions will not be available in the
knowledge base, and cannot be used by other rules in the system.
To describe the phenomena in more detail, the above rule can be expanded
into a chain of rules. These will show that when the glucose feeding is high the
cell oxidative capacity is exceeded, and cells start to excrete acetic acid, which
results in a raise in the Ra/g value. For more precise quantitative description of
the phenomenon, a mathematical model can be added to the rules. Undoubtedly,
the detailization may continue, involving more and more deep knowledge, down
to the level of enzymes and pathways. The question is whether or not this is useful
or efficient. Generally, excessive overloading of the ES with ballast knowledge is
also harmful. As the control problem itself is usually formulated at a rather high
level of abstraction, extremely detailed description can hardly be useful. Fur-
thermore, the vast amount of knowledge will put a serious load on the computer,
causing an unacceptable slow-down in the real-time environment. On the other
hand, in some cases inclusion of deep knowledge may cause some heuristics to
become redundant, resulting in a net reduction of the size of the knowledge base
[44]. Therefore, the depth of the knowledge must be well balanced with the control
purposes, the desired explanation capabilities, and the time and memory con-
straints.
184 K.B. Konstantinov, R. Aarts and T. Yoshida

3.1.6 Efficient Knowledge Debugging and Integrating

Building of a new knowledge base is an incremental and repetitive task, A few
rules should first be defined and tested, and then a few more added. After every
update, knowledge must be validated again. This is a mandatory step, since initially
the knowledge will probably contain a couple of logical bugs and inaccuracies.
To facilitate their localization and elimination, the ESDT should provide proper
mechanisms, generally referred to as knowledge debuggers.
ESDTs usually include capabilities for off-line debugging of the knowledge base,
such as the planting of breakpoints into the rules and step-wise tracing of the
inference process. More advanced systems provide special mechanisms called
knowledge-integrity checkers, which scan automatically the rule set for any logical
inconsistencies, such as unreachable clauses, dead-end clauses, cyclic clauses,
redundant rules, and subsuming rules [48].
Before real operation, the knowledge base should be tested also in realistic
real-time conditions. This is usually achieved by some kind of simulation. The
simplest way is the replacement of the sensor inputs by real cultivation data,
collected in previous experiments. However, not all situations which may arise in
practice can be checked in this way. As it is impossible to sacrifice a whole process
to deliberately provoke the desired situation (sensor failures, nutrient limitation,
contamination), this must be achieved by computer simulation. For such purposes
the most advanced ESDTs are equipped with special simulators. This facility is
considered to be one of the most powerful mechanisms for knowledge debugging
[391.
Another important ESDT feature is the editing of old and integrating of new
knowledge during on-line operation [49]. As the knowledge validation problem
cannot be solved 100% off-line, errors are often discovered during the real run.
These have to be resolved, without stoppage of the system, by on-line knowledge
editing. Similarly, during cultivation, new phenomena or events may be observed.
Instead of waiting until the end of the cultivation, it might be better to enrich the
knowledge base by the new rule immediately. Similar conclusions can be drawn
for editing and integrating other types of knowledge, e.g. analytical models.

3.1.7 Answering User's Questions

The complexity of bioprocesses makes the understanding of the accompanying
phenomena by non-experienced operators difficult. Consequently, the activities of
the ES will remain obscure for the novice, unless the logic of its operations is
represented in a transparent form. This is achieved by an ES explanation facility,
which allows the user to ask questions during operation; the ES will give answers
based on the knowledge available and the current context. Consistent on-line
explanation would contribute to the understanding of the process and the
underlying control concepts. It would also increase the user's confidence in the
ES, and may serve as a knowledge testing tool.
So far, the explanation facilities of ESDTs are rather restricted [50]. They are
reduced to interpretation of a limited number of questions, typically "Why do
you need this information?" and "How did you arrive at that decision?". These
Expert Systems in Bioprocess Control: Requisite Features 185

questions are easy to reply to because the answer is constructed in a straightforward

manner, simply by looking one step up or down in the inference chain [22].
More sophisticated systems can answer questions such as "What are you doing
now?", "What will happen i f . . . ? " , "What do you know so far about ...?", etc.
Related features are the interactive explanation, and explanation at different levels
of complexity according to the user's competence.
It has been shown that from an informational viewpoint the knowledge required
for consistent explanation is more than that needed for the work of the ES [51].
Today's ESs try to explain their decisions and actions, i.e. their knowledge,
by means of that knowledge itself. Generally, for intelligent explanation, deeper
knowledge than that used in problem-solving, is required. However, enrichment
of the knowledge base by such supplemental information just for explanation
purposes will always slow down the inference process. A possible solution is
the development of a separate knowledge base especially for explanations.
In many ESDTs, this feature, though not supported, is emulated in a purely
mechanistic way. It allows the developer to attach explanatory texts to the
rules, which will appear on the screen as answers to a user's questions under
particular circumstances. This mechanism is primitive, but it can be useful in many
cases.
It is expected that improvement of the explanation facilities will be possible by
narrowing the ES application area [50]. For example, in the future, ESDTs
specialized for bioprocesses control that answer questions such as "How has the
current physiological state been reached?", "What is your prediction about the
next state?", "Explain the growth profile", "Why is OUR decreasing?", might be
designed.

3.1.8 Advanced User Interface with Graphical Capabilities

This feature is relatively well developed today. The main difficulties ensue from
the necessity to combine simplicity of the man-machine interface with the advanced
performance of the ES. In bioprocess control, it is extremely important to maintain
simplicity because potential users tend to lack specialized computer skills.
Fortunately, modern software techniques for graphical representation, menu-
driven dialogs, icon definitions, simplified data input using a mouse, separation
of the developer interface from the end-user interface, and others, provide a fairly
good balance between simplicity and sophistication.
A major role in the user interface is played by the graphical capabilities of the
ESDT. Since in bioprocesses the main part of the information is represented by
continuous variables, graphical features are of utmost importance [52]. Today,
these are supported by almost all advanced ESDTs, though some companies are
selling the graphics packages as a separate product. Apart from the representation
of process variables, modern systems use graphics for the development and
representation of schemes, tables and diagrams. These help in providing more
realistic visualization o f the situation in the control plant [24], or in representing
the rule base as an easy-to-understand network [36, 37].
186 K.B. Konstantinov, R. Aarts and T. Yoshida

3.1.9 Other Features

ES technology is advancing rapidly, and modern tools are enhanced by additional
features. Among them, the capability of object-oriented representation, which is
regarded as being very natural for biotechnological applications, deserves special
attention [53, 54]. However, this feature is provided in different products at
different levels. Some ESDTs, which originally lacked object-oriented representa-
tion, have recently been modified to include it. The resulting "sewn on" capabilities
are usually restricted, because the internal ESDT structure, created initially without
an object-oriented approach in mind, cannot be completely reorganized to handle
the new paradigm [36].
The neural network approach, which has recently proven useful in modeling
systems of unknown structure, including bioprocesses [14, 55], is another issue
concerning ESDTs. It has been shown that ESs can incorporate neural networks
to enhance their problem-solving power [56]. This capability should be also
considered when selecting an ESDT.
Learning is a feature which allows progressive improvement of the system
performance. It is inherent to the neural network approach, but in the ES field
is still a subject of theoretical research [1]. Nevertheless, some positive results of the
application of learning algorithms in bioprocess control have already been
published [8].
ES progress in the field of the man-machine interface is being stimulated both
by developments in theory and improvements in hardware. Some ESDTs are
already equipped with natural language processors; further enhancements in the
areas of video and audio "multimedia" capabilities are also expected [57].
There are many other features of ESDTs described in the specialized literature,
which could be more or less important in bioprocess control. It is worth noting
that together with their useful capabilities, much attention is often given to some
"fancy" features. The serious designer should always take care to differentiate
between really feasible issues, and such unnecessary elaborations [58, 59].

3.2 Specific Features

In addition to the general features discussed above, there are some issues related
more specifically to the control of bioprocesses. They reflect the unique charac-
teristics of biological systems, which differ from those of conventional control
plants, and influence the selection of the ESDT.

3.2.1 Limited "'Width" and Enhanced "Depth"

Unlike in other fields, in bioprocesses control, the set of the on-line measured
variables is quite limited. To minimize the risk of contamination, the number of
sensors in large-scale systems is even fewer than in small-scale laboratory
bioreactors. This implies that the capabilities of scanning tens of thousands of
inputs and maintaining huge data bases, provided by some commercial ESDTs
[28, 60, 61], are not necessary. Also, it is not expected that the size of the ES rule
Expert Systems in Bioprocess Control: Requisite Features 187

base will be large. The number of rules will probably range between several dozen
and a few hundred [42, 62, 63]. Thus, it is not likely that ESs for the control of
bioprocesses will grow too much in width.
The reduced size of the knowledge base does not, however, mean simplicity.
To interpret intricate and vague situations, the form of rules and the structure of
the knowledge base will be complex. Consequently, a compact and flexible ESDT,
providing a rich set of functional capabilities (referred to as the "depht" of the
ES) will best suit the problem of interest.

3.2.2 Handling of Uncertain, Incomplete, and Fuzzy Information

The problem of the amount and quality of the information available on-line in
biotechnical systems is probably more serious than in any other control plant.
Many of the physiological phenomena are poorly understood, large regions of
the physiological state space remain unknown, while the information about others
is not crisp and quantitative, but fuzzy, with a qualitative character. In addition,
due to the lack of sensors for biochemical variables, it is impossible to supply the
control system with all the information necessary on-line (note that even an
adequate cell concentration sensor is not yet available !).
This imposes a serious requirement on an ESDT used for bioprocesses control:
it must be able to handle and utilize such uncertain, incomplete and fuzzy
information, to make decisions in conditions of concurrent solutions, and to act
adequately in "unknown" situations. Indeed, the problem is related to one of the
main tasks of the higher level of the control system (Fig. 2), which must resolve
process uncertainties, creating a simple deterministic environment for the function-
ing of the lower system level. Such capabilities are usually implemented by
mechanisms for handling fuzzy sets, qualitative knowledge, certainty factors, and
confidence levels, which are provided by some of the ESDTs available today.
The processing of rules of different types (crisp or fuzzy) requires dynamic
alteration of the method for calculating the certainties of their conclusions from
the certainties of the facts in the condition. The ESDT should provide such
flexibility, possibly by assigning the preferred method to every rule as an argument.
In the author's experience, such a feature is of fundamental importance in the
handling of complex and uncertain phenomena [8, 17], but it is not supported by
commercial ESDTs.

3.2.3 Orientation Towards the Processing of Continuous Variables

Almost 100% of the information available on-line is in the form of continuous
process variables with complex dynamic behavior. Because of the lack of sensors,
the control system must "squeeze" as much information as possible from the
available variables and their histories. Therefore, capabilities of advanced signal
processing, both deterministic and stochastic, are highly desirable.

3.2.4 Capabilities of Intelligent Monitoring

Because of the complicated dynamics of bioprocesses and the growing complexity
of recent measurement equipment, intelligent supervision is needed not only for
the control procedures, but also for the measurement tasks. According to the
188 K.B. Konstantinov, R. Aarts and T. Yoshida

situation, various modifications of the measurement procedures might be required.

Examples are a change of the sampling intervals according to the dynamics of
the plant, activation/deactivation of the measurement or estimation of variables
valid only in particular time-intervals, periodical on-line sensor calibration, range
switching, supervision of the work of advanced analytical devices (mass spectrome-
ters, flow injection- and other analyzers, auto samplers, etc.), and control of some
specific measurement procedures which require test actions. A simple example is
the on-line calibration of exhaust gas analyzers. This procedure causes a jump in
the CO2 and O2 concentrations, and disturbs all related variables (OUR, CER,
RQ, etc.) which might be used for control purposes. To prevent incorrect decisions
and actions, the ES should suspend the updating of these variables during the
calibration period.
The ESDT must be flexible enough to accommodate knowledge for supervision
of the measurement procedures. Enrichment of the knowledge base with such
rules will provide the measurement with a kind of intelligence too, and will
contribute to the synchronization and coordination of the ES activities.

3.2.5 Availability of Common "Bioprocess" Knowledge Bases

General ESDTs come with empty knowledge bases which must be filled by the
knowledge engineer, or by the user himself. The knowledge bases of specialized
ESDTs might be originally loaded with a certain amount of fundamental
information, valid in most of the applications in particular field. It is believed that
such organization will be useful in bioprocess control [14]. Some basic parts of
the knowledge of biological systems concerning metabolism, population behavior,
or other aspects of bioengineering, can be permanently fixed into the ESDT. In
many cases this would shorten the development of the control system, and make
the knowledge of the plant more comprehensive. Unfortunately, such a feature is
not yet available. In future, it might be provided by an ESDT designed especially
for the control of bioprocesses.

3.2.6 Low Prototyping Time

Modern biotechnology is an area of rapid technological evolution, in which the
development of a new process, or the improvement of an old one, does not usually
take long. It would be impractical if the effort required for building the control
system were too time consuming, thus delaying the application period. Therefore,
the time needed for ES development is another criterion which should be carefully
considered. A terse and flexible, yet powerful ESDT is favored, whose application
requires minimal study and programing effort. This is especially important for
biotechnology teams, which normally do not include software specialists.
From a software viewpoint, there are several ways to create an ES. It can be
made completely by oneself using a proper programing language, or the system
can be constructed from an ESDT. According to the complexity of the task, the
time for development of an ES from an ESDT usually varies from three months
to about one year [58], which is generally acceptable for biotechnology applications.
However, the time required for building a system using conventional language is
estimated to be several times longer than using an ESDT [21]. The conclusion is
Expert Systems in Bioprocess Control: Requisite Features 189

straightforward: low prototyping time can be achieved only by using an ESDT.

Indeed, this is the common practice in all ES application fields, with over 90% of
the systems being based on ESDTs [58].

3.2.7 Low Price

Most of today's biotechnology laboratories are equipped with a number of
bioreactors, often used in parallel for development of different processes.
Depending on the circumstances, several of these reactors may need monitoring
and control by an advanced computer system. Practically, it is possible to use a
single ES for control of more than one process, but as strain characteristics,
cultivation conditions, logic of operation and control objectives might vary
tremendously, such integration would be artificial, resulting in a series of problems.
In such a case, better solution is to provide each bioreactor with an independent
control system. This is, however, possible only if the corresponding hardware and
software
are inexpensive.
Fortunately, the characteristics of the bioprocesses allow the application of
non-expensive hardware, such as 32-bit PCs. As almost all ESDTs are available
in PC versions [58], there is a common trend today towards the application of
PCs in expert technology. However, the prices of the real-time ESDTs themselves
are prohibitively high, ranging from several thousand dollars for systems with
moderate capabilities, to above 50000 dollars for top-class ESDTs. Undoubtedly,
the software price is another serious restriction, still preventing ES control
technology from becoming daily biotechnological practice.

4 Conclusions

Today, there is no single ESDT which combines all the features discussed above,
and since such a perfect tool may not be available soon, the developer should
select the most convenient product from those currently on the marked. The most
appropriate are ESDTs with build-in real-time capabilities, such as NEXPERT
OBJECT, ART, KEE, and particularly those designed for process monitoring and
control, e.g. ESCORT, R'TIME, PICON, MUSE, COMDALE and G2 (the last
being considered to be the most advanced one [16, 28, 64]). Due to their flexibility,
these ESDTs can be customized to a particular application, though such adaption
will not satisfy all specific requirements. Nevertheless, the resulting ES will be
capable of covering a large portion of the problems arising in the control of
bioprocesses.
Certainly, the ideal ESDT in this field would be far more specialized. The lack
of such a product from the market extends the time and effort needed for the
creation of control systems considerably. Since large-scale use of ES technology
in the area of bioprocess control seems inevitable, the development and com-
mercialization of such an ESDT would be an important contribution to the
biotechnology community.
190 K.B. Konstantinov, R. Aarts and T. Yoshida

5 References

1. Verbruggen HB, ~str6m KJ (1989) Proceed IFAC Workshop on AI in Real-Time

Control, Shenyang, PRC, p 1
2. Halme A, Visala A (eds) (1989) Proceed International Workshop on Expert Systems in
Biotechnology, Helsinki, Finland
3. Shirley RS, (1987) IEEE Control Systems Magazine 7:11
4. Cooney CL, O'Connor GM, Sanches-Riera F (1988) Proceed Int Biotechnol Symp,
Paris, p 563
5. Cardello RJ, San K (1989) Proc. American Control Conferencel Pittsburgh, Pennsylvania,
2411
6. Konstantinov KB, Yoshida T (1992) Biotechnol Bioeng 39:479
7. Smuts WB, MacLeod IM (1989) Proceed IFAC Workshop on AI in Real-Time Control,
Shenyang, PRC, p 31
8. Konstantinov KB, Yoshida T (1989) Biotechnol Bioeng 33:1145
9. Halme A (1988) Proceed IFAC Symp on Computer Applications in Fermentation
Technology, Cambridge, UK, p 159
10. Lukas MB, Oyen RA, Keyes MA, Kaya A (1989) Proceed IFAC Workshop on AI in
Real-Time Control, Shenyang, PRC, p 79
11. Konstantinov KB, Yoshida T (1990) J Ferm Bioeng 70:48
12. Bakay A, Madaraasz L, Hinsenkamp A, Papp Z, Dobrowiecki T (1990) Preprints l l t h
IFAC World Congress, Tallin, USSR, 7:264
13. Saridis GN (1989) Atomatica 25:461
14. Aynsley M, Peel D, Morris A (1989) Proceed American Control Conference, Pittsburgh,
Pennsylvania, p 2239
15. Aynslay M, Hofland AG, Montague GA, Peel D, Morris AJ (1990) Proc American
Control Conference, San Diego, California, p 1992
16. Arzen KE (1989) Proceed American Control Conference, Pittsburgh, Pennsylvania,
p 2233
17. Konstantinov KB, Yoshida T (1991) Proceed IFAC Workshop on Computer Software
Structures Integrating AI/KB Systems in Process Control, Bergen, Norway, p 198
18. Beck T, Lauber J (1990) Preprints l l t h IFAC World Congress, Tallin, USSR, 7:158
19. Laffey TJ, Cox PA, Schmidt JL, Kao SM, Read JY (1988) AI Magazine 9:27
20. Press L (1989) IEEE Expert 4 (Spring), p 37
21. Gevarter, WB (1987) Computer, 20 (May), p 24
22. Winston PH, Horn BK (1989) Lisp, Addison-Wesley, Massachusetts
23. Laffey TJ (1991) Byte (January), p 259
24. Moore R, Rosenof H, Stanley G (1990) Preprints Of 1lth IFAC World Congress, Tallin,
USSR, 7:234
25. Wright ML, Green MW, Fiegl G, Cross P (1986) IEEE Software (March), p 16.
26. Ennis RL, Griesmer JH, Hong SJ, Karnaugh M, Kastner JK, Klein DA, Milliken KR,
Schor MI, Van Woerkom HM (1986) IBM J Res Develop 30 (January), p 14
27. Lun V, MacLeod IM (1989) Proceed IFAC Workshop on AI in Real-Time Control,
Shenyang, PRC, p 25
28. Coyle F (1990) IEEE Expert 5 (October), p 12
29. James JR, Suski GJ (1988) Proceed IEEE Conf on Decision and Control, Austin, Texas,
p 580
30. Konstantinov KB, Yoshida T (1992) AIChE J. 38:1703
31. De Feyter AR (1989) Proceed IFAC Conf Advanced information processing in Automatic
Control, Nancy, France, p 231
32. Konstantinov KB, Yoshida T (1991) IEEE Trans Syst, Man, Cybern SMC-21:908
33. Konstantinov KB, Yoshida T (1992) Proceed IFAC Symp on On-Line Fault Detection
and Supervision in the Chemical Process Industries, Newark, Delaware, p 93
34. Rao M, Cruz R, Yang T, Jiang T, Wang S, Kim I (1989) Proc American Control
Conference, Pittsburgh, Pennsylvania, p 2418
Expert Systems in Bioprocess Control: Requisite Features 191

35. Haase VH (1990) Preprints l l t h IFAC World Congress, Tallin, USSR, 1:141
36. Payne EC, McArthur RC (1990) Developing expert systems, Wiley, New York
37. Aarts, RJ, Suviranta A, Rauman-Aalto P, Linko P (1989) Food Biotechnology 4:301
38. MontellanoR, Bernier M, CheryA, FarzaM(1990)Preprints llthIFACWorldCongress,
Tallin, USSR, 7:54
39. Johannsen G, Alty JL (1991) Automatica 27:97
40. Kuipers BJ (1986) Artificial Intelligence 29:289
41. Price C, Lee M (1988) Proceed 12th IMACS World Congress, Paris, p 85
42. Liibbert A, Hitzman B, Kracke-Helm H, Sch/igerl K (1988) Proceed IFAC Syrup on
Computer Applications in Fermentation Technology, Cambridge, UK, p 297
43. Voss H (1988) Proceed IFAC Workshop on AI in Real-Time Control, Clyne Castle,
UK, p l
44. Aarts RJ (1989) Proceed International Workshop on Expert Systems in Biotechnology,
Helsinki, Finland, p 108
45. Beaumont SF, Leyval L, Gentil S (1989) Proceed IFAC Conf Advanced Information
Processing in Automatic Control, Nancy, France, p 181
46. Smith JM, Stutely R (1988) SGML: The users guide, Ellis Horwood, Chichester
47. Konstantinov KB, Nishio N, Seki T, Yoshida T (1991) J Ferment Bioeng 71:350
48. Cragun BJ, Stendel HJ (1987) Int J Man-Mach Stud 26:633
49. Walker A (1986) IBM J Res Develop 30 (January), p 2
50. Wick MR, Slagle JR (1989) IEEE Expert 4 (Spring), p 26
51. Chandrasekaran B, Mittal S (1983) Int J Man-Machine Studies (November), p 425
52. Stephanopoulos G (1987) Chemical Engineering Progress 83:44
53. Stephanopoulos G, Stephanopoulos G (1986) Trends in Biotechnology 4:241
54. Linko S, Aarts R (1990) Abstracts of the European Congress in Biotechnology,
Copenhagen, p 326
55. Thibault J, Van Breusegem V, Chwruy A (1990) Biotechnol Bioeng 36:1041
56. Rich E (1990) IEEE Expert 5:5
57. Bielawski L, Lewland R (1991) Intelligent System Design: Integrating Hypermedia and
Expert Systems Technologies, Wiley. New York
58. San Giovanni J, Romans H (1987) Chem Eng Progress 83:52
59. Bristol EH (1987) Proceed IEEE Conf on Decision and Control, Los Angeles, p 1948
60. Moore RL (1985) Control Engineering (April), p 118
61. Satch PA, Paterson AM, Turner MH (1986) Expert Systems 3:22
62. Chen Q, Wang S, Wang J (1988) Proceed IFAC Symp on Computer Applications in
Fermentation Technology, Cambridge, UK, p 253
63. Xu E, Xu G, Zhang S (1989) Proceed IFAC Workshop on A1 in Real-Time Control,
Shenyang, PRC, p 97
64. Eldeib HK (1989) Proc American Control Conference, Pittsburgh, Pennsylvania, p 2637
Author Index Volumes 1-48

Aarts, R. see Konstantinov, K. B, Vol. 48, p. 169

Acosta Jr., D. see Smith, R. V. Vol. 5, p. 69
Acton, R. T., Lynn, J. D.: Description and Operation of a Large-Scale Mammalian Cell, Suspension
Culture Facility. Vol. 7, p. 85
Agarwal, G. P.: Glycerol. Vol. 41, p. 77
Agrawal, P., Lira, H. C.: Analysis of Various Control Schemes for Continuous Bioreactors. Vol. 30,
p. 6l
Aiba, S.: Growth Kinetics of Photosynthesis Microorganisms. Vol. 23, p. 85
Aiba, A., Nagatani, M.: Separation of Cells from Culture Media. Vol. 1, p. 31
Aiba, S. see Sudo, R. Vol. 29, p. 117
Aiby, S., Okabe, M.: A Complementary Approach to Scale-Up. Vol. 7, p. 111
Alfermann, A. W. see Reinhard, E. Vol. 16, p. 49
Anderson, L. A., Phillipson, J. D., Roberts, M. F.: Biosynthesis of Secondary Products by Cell
Cultures of Higher Plants. Vol. 31, p. 1
Arnaud, A. see Jallageas, J.-C. Vol 14, p. 1
Arora, H. L., see Carioca, J. O. B. Vol. 20, p. 153
Asher, Z. see Kosaric, N. Vol. 32, p. 25
Atkinson, B., Daoud, I. S.: Microbial Flocs and Flocculation. Vol. 4, p. 41
Atkinson, B., Fowler, H. W.: The Significance of Microbial Film in Fermenters. Vol. 3, p. 221
Aynsley, M., Hofland, A., Morris, A. J., Montague, G. A.: Artifical Intelligence and the Supervision of
Bioprocesses. Vol. 48, p. 1

Bagnarelli , P., Clementi, M.: Serum-Free Growth of Human Hepatoma Cells. Vol. 34, p. 85
Bailey, J. E.: Host-Vector Interactions in Escherichia coli. Vol. 48, p. 29
Barford, J.-P., see Harbour, C. Vol. 37, p. 1
Barker, A. A., Somers, P. J.: Biotechnology of Immobilized Multienzyme Systems. Vol. 10, p. 27
Barnes, T. J. see Karube, 1. Vol. 46, p. 63
Beardmore, D. H. see Fan, L. T. Vol. 14, p. 101
Bedetti, C., Cantafora, A.: Extraction and Purification of Arachidonic Acid Metabolites from Cell
Cultures. Vol. 35, p. 47
Belfort, G. see Heath, C. Vol. 34, p. 1
Belfort, G. see Heath, C. A. Vol. 47, p. 45
Beker, M. J., Rapoport, A. J.: Conservation of Yeasts by Dehydration. VO1. 35, p. 127
Bell, D. J., Hoare, M., Dunnill, P.: The Formation of Protein Precipitates and their Centrifugal
Recovery. Vol. 26, p. 1
Berlin, J., Sasse, F.: Selection and Screening Techniques for Plant Cell Cultures. Vol. 31, p. 99
Binder, H. see Wiesmann, U. Vol. 24, p. 119
Bjare, M.: Serum-Free Cultivation of Lymphoid Cells. Vol. 34, p. 95
Blanch, H. W., Dunn, I. J.: Modelling and Simulation in Biotechnical Engineering. Vol. 3, p. 127
Blanch. H. W., see Moo-Young, M. Vol. 19, p. 1
Blanch, H. W., see Maiorelle, B. Vol. 20, p. 43
Blaszczyk, R. see Kosaric, N. Vol. 42, p. 27
Blenke, H. see Seipenbusch, R. Vol. 15, p. 1
Blenke, H.: Loop Reactors. Vol. 13, p. 121
Bliem, R. F., Konopitzky, H. W., Katinger, H. W. D.: Industrial Animal Cell Reactor Systems:
Aspects of Selection and Evaluation. Vol. 44, p. 1
194 Author Index Volumes 1 48

Blumauerov~, M. see Hostalek, Z. Vol. 3, p. 13

B?)hme, P. see Kopperscht/iger, G. Vol. 25, p. 101
de Boer, L., Dijkhuizen, L.: Microbial and Enzymatic Processes for L-Phenylalanine Production
Vol. 41, p. 1
Boesten, W. H. J. see Kamphuis, J. Vol. 42, p. 133
Bottino, P. J. see Gamborg. O. L. Vol. 19, p. 239
Bowers, L. D., Carr, P. W.: Immobilized Enzymes in Analytical Chemistry. Vol. 15, p. 89
Brandl, H., Gross; R. A., Lenz, R, W., Fuller, R. C.: Plastics from Bacteria: Poly([3-
Hydroxyalkanoates) as Natural, Biocompatible, and Biodegradable Polyesters. Vol. 41, p. 77
Brauer, H.: Power Consumption in Aerated Stirred Tank Reactor Systems. Vol. 13, p. 87
Brodelius, P.: Industrial Applications of Immobilized Biocatalysts. Vol. 10, p. 75
Brosseau, J. D. see Zajic, J. E. Vol. 9, p. 57
Broxterman, Q. B. see Kamphius, J. Vol. 42, p. 133
Bryant, J.: The Characterization of Mixing in Fermenters. VOI. 5, p. 101
Buehholz, K.: Reaction Engineering Parameters for Immobilized Biocatalysts. Vol. 24, p. 39
Brfickmann, A. F., Carrea, G.: Synthesis and Application of Water-Soluble Macromolecular
Derivatives of the Redox Coenzymes NAD (H), NADP (H) and FAD. Vol. 39, p. 97
Bungay, H. R.: Biochemical Engineering for Fuel Production in Used States. Vol. 20, p. 1
Butler, M.: Growth Limitations in Microcarries Culture. Vol. 34, p. 57

Cantafora, A. see Bedetti, C. Vol. 35, p. 47

Carrea, G. see Bfickmann, A. F. Vol. 39, p. 97
Chan, Y. K. see Schneider, H. Vol. 27, p. 57
Chang, H. N., Furusaki, S.: Membrane Bioreactors: Present and Prospects. Vol. 44, p. 27
Carioca, J. O. B., Arora, H. L., Khan, A. S.: Biomass Conversion Program in Brazil. Vol. 20, p. 153
Carr, P. W. see Bowers, L. D. Vol. 15, p. 89
Chang, M. M., Chou, T. Y. C., Tsao, G. T.: Structure, Pretreatment, and Hydrolysis of Cellulose. Vol.
20, p. 15
Charles, M.: Technical Aspects of the Rheological Properties of Microbial Cultures. Vol. 8, sp. 1
Chen, L. F., see Gong, Ch-S. Vol. 20, p. 93
Chou, T_ Y. C., see Chang, M. M. Vol. 20, p. 15
Cibo-Geiby/Lepetit.: Seminar of Tropics of Fermentation Microbiology. Vol. 3, p. 1
Claus, R. see Haferburg, D. Vot. 33, p. 53
Clementi, P. see Bagnarelli, P. Vol. 34, p. 85
Cagoli, A., Tschopp, A.: Biotechnology in Space Laboratories. Vol. 22, p. 1
Coolbear, T., Daniel, R. M., Morgan, H. W.: The Enzymes from Extreme Thermophiles: Bacterial
Sources, Thermostabilities and Industrial Relevance. Vol. 45, p. 57
Cooney, C. L. see Koplove, H. M. Vol. 12, p. 1
Cooney, C. L. see Koplove, R. L. Vol. 43, p. 11
Costentino, G. P. see Kosaric, N. Vol. 32, p. 1

Dabora, R. L., Cooney, C. L.: Intracellular Lytic Enzyme Systems and Their Use for Disruption of
Eseherichia colL Vol. 43, p. 1
Daniel, R. M., see Coolbear, T. Vol. 45, p. 57
Daoud, L S. see Atkinson, B. Vol. 4, p. 41
Dos, K. see Ghose, T. K. Vol. 1, p. 55
Davis, P. J. see Smith, R. V. Vol. 14, p. 61
Deckwer, W.-D. see Schumpe, A. Vol. 24, p. 1
Demain, A. L.: Overproduction of Microbial Metabolites and Enzymes due to Alteration of
Regulation. Vol. 1, p. 113
Diamond, A. D., Hsu, J. T.: Aqueous Two- phase Systems for Biomoleculare Separation, Vol. 47, p. 89
Dijkhuizen, L. see de Bes, L. Vol. 41, p. 1
Doelle, H. W., Ewings, K. N., Hollywood, N. W.: Regulation of Glucose Metabolism in Bacterial
Systems. Vol. 23, p. 1
Doelle, H. W. see Johns, M. R. Vol. 44, p. 97
Doran, P. M.: Design of Reactors for Plant Cells and Organs. Vol. 48, p. 115
Dubois, D. see Miller, O. A. Vol. 39, p. 73
Dunn, I. J. see Blanch, H. W. Vol. 3, p. 127
Dunn, I. J. see Heinzle, E. Vol. 48, p. 79
Dunnill, P. see Bell, D. J. Vol. 26, p. 1
Author Index Volumes 1 48 195

Duvnjak, Z. see Kosaric, N. Vol. 20, p. 119

Duvnjak, Z. see Kosaric, N. Vol. 32, p. 1

Eckenfelder Jr., W. W., Goodman, B. L., Englande, A. J.: Scale-Up of Biological Wastewater
Treatment Reactors. Vol. 46, p. 145
Economidis, I.: An Overview of the Biotechnology research Activities in the European Community.
Vol. 46, p. 225
Einsele, A., Fiechter, A.: Liquid and Solid Hydrocarbons. Vol. 1, p. 169
Electricwala, A. see Griffiths, J. B. Vol. 34, p. 147
Enari, T. M., Markkanen, P.: Production of Cellulolytic Enzymes by Fungi. Vol. 5, p. 1
Enatsu, T., Shinmyo, A.: In Vitro Synthesis of Enzymes. Physiological Aspects of Microbial Enzyme
Production Vol. 9, p. ! 11
Endo, I., Nagamune, T., Tachikava, S., Inaba, H.: A Tubular Bioreactor for High Density
Cultivation of Microorganisms. Vol. 42, p. 1
Endo, I.: A Human Genome Analysis System (HUGA-1) Developed in Japan. Vol. 46, p. 103
Enfors, S-O., Hellbust, H., K6hler, K., Strandberg, L., Veide, A.: Impact of Genetic Engineering on
Downstream Processing of Proteins Produced in E. coli. Vol. 43, p. 31
Engels, J., Uhlmann, E.: Gene Synthesis. Vol 37, p. 73
Englande, A. J. see Eckenfelder Jr., W. W. Vol. 2, p. 145
Eriksson, K. E.: Swedish Developments in Biotechnology Based on Lignocellulose Materials. Vol.
20, p. 193
Eriksson, K.-E. L. see Vallander L. Vol. 42, p. 63
Esser, K.: Some Aspects of Basic Genetic Research on Fungi and Their Practical implications. Vol. 3, p.
69
Esser, K., Lang-Hinrichs, Ch.: Molecular Cloning in Heterologous Systems, Vol. 26, p. 143
Ewings, K. N. see Doelle, H. W. Vol. 33, p. t

Faith, W. T., Neubeck, C. E., Reese, E. T.: Production and Application of Enzymes. Vol. 1, p. 77
Fan, L. S. see Lee, Y. H. Vol. 17, p. 131
Fan, L. T., Lee, Y.-H., Beardmore, D. H.: Major Chemical and Physical Features of Cellulosic
Materials as Substrates for Enzymatic Hydrolysis. Vol. 14, p. 101
Fan, L. T., Lee, Y.-H., Charppuray, M. M.: The Nature of Lignocellulosics and Their Pretreatments
for Enzymatic Hydrolysis. Vol. 23, p. 155
Fan, L. T. see Lee, Y.-H. Vol. 17, p. 101 and p. 131
Faust, U., Sittig, W.: Methanol as Carbon Source for Biomass Production in a Loop Reactor.
Vol. 17, p. 63
Fiechter, A.: Physical and Chemical Parameters of Microbial Growth. Vol. 30, p. 7
Fiechter, A., Gm/inder, F. K.: Metabolic Control of Glucose Degradation in Yeast and Tumor Cells.
Vol. 39, p. 1
Fiechter, A. see Einsele, A. Vol. 1, p. 169
Fiechter, A. see Janshekar, H. Vol. 27, p. 119
Fiechter, A., Sonnleitner, B.: Impacts of Automated Bioprocess Systems on Modern Biological
Research. Vol. 46, p. 143
Finocchiaro, T., Olson, N. F., Richardson, T.: Use of Immobilized Lactase in Milk Systems. Vol. 15,
p. 71
Flaschel, E. see Wandrey, C. Vol. 12, p. 147
Flaschel, E., Wandrey Ch., Kula, M.-R.: Ultrafiltration for the Seperation of Biocatalysts. Vol. 26,
p. 73
Flickinger, M. C. see Gong, Ch.-S. Vol. 20, p. 93
Fowler, H. W. see Atkinson, B. Vol. 3, p. 221
Friehs, K., Reardon, K. F.: Parameters Influencing the Productivity of Recombinant E. coli
Cultivations. Vol. 48, p. 53
Fukui, S., Tanaka, A.: Application of Biocatalysts Immobilized by Prepolymer Methods. Vol. 29, p. 1
Fukui, S., Tanaka, A.: Metabolism of Alkanes by Yeasts. Vol. 19, p. 217
Fukui, S., Tanaka, A.: Production of Useful Compounds from Alkane Media in Japan, Vol. 17, p. 1
Fuller, R. C. see Brandl. H. Vol. 41, p. 77
Furusaki, S. see Chang, H. N. Vol. 44, p. 27
Furusaki, S., Seki, M.: Use and Engineering Aspects of Immobilized Cells in Biotechnology. Vol. 46,
p. 161
196 Author Index Volumes 1-48

Galzy, P. see Jallageas, J.-C. Vol. 14, p. 1

Gamborg, O. L., Bottino, P. J.: Protoplasts in Genetic Modifications of Plants. Vol. 19, p. 239
Gaudy Jr., A. F., Gaudy, E. T.: Mixed Microbial Populations. Vol. 2, p. 97
Gaudy, E. T. see Gaudy Jr., A. F. Vol. 2, p. 97
Gharpuray, M. M. see Fan, L. T. Vol. 23, p. 155
Ghose, T. K., Das, K.: A Simplified Kinetic Approach to Cellulose-Cellulase System. Vol. 1, p. 55
Ghose, T. K.: Cellulase Biosynthesis and Hydrolysis of Cellulosic substances. Vol. 6, p. 39
Glatz, C. E. see Niederauer, M.Q. Vol. 47, p. 159
Glumoff, V. see Reiser J. Vol. 43, p. 75
Gmfinder, F. K. see Fiechter, A. Vol. 39, p. 1
Gogotov, I. N. see Kondratieva, E. N. Vol. 28, p. 139
Gomez, R. F.: Nucleic Acid Damage in Thermal Inactivation of Vegetative Microorganisms. Vol. 5,
p. 49
Gong, Ch. S. see Mc Cracken, L. D. Vol. 27, p. 33
Gong, Ch. S., Chen, L. F., Tsao, G. T., Flickinger, M. G.: Conversion of Hemicellulose Carbo-
hydrates, Vol. 20, p. 93
Goodman, B. L. see Eckenfelder Jr., W. W. Vol. 2, p. 145
Grampp, G. E., Stephanopoulos, G. N., Sambanis, A.: Use of Regulated Secretion of Protein
Production from Animal Cells: An Overview. Vol. 46, p. 35
Graves, D. J., Wu, Y.-T.: The Rational Design of Affinity Chromatography Separtation Processes.
Vol. 12, p. 219
Greenfield, P. F. see Johns, M. R. Vol. 44, p. 97
Griffiths, J. B., Electricwala, A.: Production of Tissue Plasminogen Activators from Animal Cells.
Vol. 34, p. 147
Gross, R. A. see Brandl. H. Vol. 41, p. 77
Gu, T. see Truei, Y.-H. Vol. 47, p. 1
Gutschick, V. P.: Energetics of Microbial Fixation of Dinitrogen. Vol. 21, p. 109

Haferberg, D., Hommel, R., Claus, R., Kleber, H.-P.: Extraceltular Microbial Lipids as Bio-
surfactants. Vol. 33, p. 53
HahIbrock, K., Schr6der, J., Vieregge, J.: Enzyme Regulation in Parsley and Soybean Cell Cultures,
Vol. 18, p. 39
I-Iakanson, H. see Mattiasson, B. Vol. 46, p. 81
Haltmeier, Th.: Biomass Utilization in Switzerland. Vol. 20, p. 189
Hampel, W,: Application of Microcomputers in the Study of Microbial Processes. Vol. 13, p. 1
I-Iarbour, C., Barford, J.-P., Low, K>S.: Process Development for Hybridoma Cells. Vol. 37, p. 1
Harder, A., Roels, J. A.: Application of Simple Structures Models in Bioengineering. Vol. 21, p. 55
Hardman, N.: Recent Developments in Enzyme and Microbial Biotechnology - Strategies in
Bioprocess Design. Vol. 40, p. 1
Harrison, D. E. F., Topiwala, H. H.: Transient and Oscillatory States of Continuous Culture. Vol. 3,
p. 167
Heath, C., Belfort, C.: Immobilization of Suspended Mammalian Cells: Analysis of Hollow Fiber
and Microcapsute Bioreactors. Vol. 34, p. 1
Heath, C. A., Belfort, G.: Synthetic Membranes in Biotechnology: Realities and Possibilities. Vol. 47,
p. 45
Hedman, P. see Janson, J.-C. Vol. 25, p. 43
Heinzle, E.: Mass Spectrometry for On-line Monitoring of Biotechnological Processes. Vo. 35, p. 1
Heinzle, E., Dunn, I. J., Ryhiner, G. B.: Modeling and Control for Anaerobic Wastewater Treatment.
Vol. 48, p. 79
Hellebust, H. see Enfors, S.-O. Vol. 43, p. 31
Hermes, H. F. M. see kamphuis, J. Vol. 42, p. 133
Hesolt, H.: Genetics and Genetic Engineering of the Industrial Yeast Yarrowia lipolytica. Vol. 43,
p. 43
Ho, Ch., Smith, M. D., Shanahan, J. F.: Carbon Dioxide Transfer in Biochemical Reactors. Vol. 35,
p. 83
Hoare, M. see Bell, D. J. Vol. 26, p. 1
Hi~ke, H. see Syldatk, Ch. Vol. 41, p. 29
Hofland, A. see Aynsley, M. Vol. 48, p. 1
Hofmann, E. see Kopperschl/iger, G. Vol. 25, p. 101
Hallb, J. see Nyeste, L. Vol. 26, p. 175
Author Index Volumes 1-48 197

Hollywood, N. W. see Doelle, H. W. Vol. 23, p. 1

Hommel, R. see Haferberg, D. Vol. 33, p. 53
Ho~ihlek, Z., Blumauerov/L, M., Vanek, Z.: Genetic Problems of the Biosynthesis of Tetracycline
Antibiotics. Vol. 3, p. 13
Hsu, J. T. see Diamond, A. D. Vol. 47, p. 89
Hu, G, Y. see Wang, P. J. Vol. t8, p. 61
Humphrey, A. E., see Rolz, G. E. Vol. 21, p. 1
Hustedt, H. see Kula, M.-R. Vol. 24, p. 73

Imanaka, T.: Application of Recombinant DNA Technology to the Production of Useful Bio-
materials. Vol. 33, p. 1
lnaba, H. see Endo, I. Vol. 42, p. 1
Inculet, I. I. see Zajic, J. E. Vol. 22, p. 51

Jack, T.-R., Zajic, J. E.: The Immobilization of Whole Cells. Vol. 5, p. 125
Jallageas, J.-C., Arnaud, A., Galzy, P.: Bioconversions of Nitriles and Their Applications. Vol. 14,
p. 1
Jang, C. M., Tsao, G. T.: Packed-Bed Adsorption Theories and Their Applications to Affinity
Chromatography. Vol. 25, p. 1
Jang, C. M., Tsao, G. T.: Affinity Chromatography. Vol. 25, p. 19
Jansen, N. B., Tsao, G. T.: Bioconversion of Pentose to 2,3-Butanediol by Klebsiella pneumonia.
Vol. 27, p. 85
Janshekar, H., Fiechter, A.: Lignin Biosynthesis, Application, and Biodegradation. Vol. 27, p. 119
Janson, J.-C., Hedman, P.: Large-Scale Chromatography of Proteins. Vol. 25, p. 43
Jeffries, Th, W.: Utilization of Xylose by Bacteria. Yeasts, and Fungi, Vol. 27, p. 1
Jiu, J.: Microbial Reactions in Prostaglandin Chemistry, Vol. 17, p. 37
Johns, M. R., Greenfield, P. F., Doelle, H. W.: Byproducts from Zymomonas mobilis. Vol. 44, p. 97

Kiilin, M. see Reiser, J. Vol. 43, p. 75

Kamihara, T., Nakamura, I.: Regulation of Respiration and its Related Metabolism by Vitamin B~
and Vitamin B 6 in Saccharmyces Yeasts. Vol. 99, p. 35
Kamphuis, J., Boesten, W. H. J., Broxterman, Q. B. Hermes, H. F. M., Van Balken, J. A. M., Meijer,
E. M., Schoemaker, H. E.: New Developments in the Chemo-Enzymatic Production of Amino
Acids. Vol. 42, p. 133
Karim, M. N., Rivera, S. L.: Artifical Neural Networks in Bioprocess State Estimation. Vol. 46, p. 1
Karube, I., Takeuchi, T., Barnes. T. J.: Biotechnological Reduction of Co s Emissions. Vol. 46, p. 63
Katinger, H. W. D. see Bliem, R. F. Vol. 44, p. 1
Keenan J. D. see Shieh, W. K. Vol. 33, p. 131
Khan, A. S., see Carioca, J. O. B. Vol. 20, p. 153
Kimura, A.: Application of recDNA Techniques to the Production of ATP and Glutathione by the
"Syntechno System". Vol. 33, p. 29
King, C.-K. see Wang, S. S. Vol. 12, p. 119
King, P. J.: Plant Tissue Culture and the Cell Cycle. Vol. 18, p. !
Kjaergaard, L.: The Redox Potential: its Use and Control in Biotechnology. Vol. 7, p. 131
Kleber, H.-P. see Haferburg, D. Vol. 33, p. 53
Kleinstreuer, C., Poweigha, T.: Modeling and Simulation of Bioreactor Process Dynamics. Vol. 30,
p. 91
Kochba, J. see Spiegel-Roy, P. Vol. 16, p. 27
Kbhler, K. see Enfors, S.-O. Vol. 43, p. 31
Kondratieva, E. N., Gogotov, I. N.: Production of Molecular Hydrogen in Microorganism. Vol. 28,
p. 139
Konopitzky, K. see Bliem, R. F. Vol. 44, p. 1
Konstantinov, K. B., Yoshida, T., Aarts, R.: Expert Systems in Bioprocess Control: Requisite
Features. Vol. 48, p. 169
Koplowe, H. M., Cooney, C. L.: Enzyme Production During Transient Growth. VoI. 12, p. 1
Kopperschlhger, G., B6hme, H.-J., Hofmann, E.: Cibacron Blue F3G-A and Related dyes as Ligands
in Affinity Chromatography. Vol. 25, p. 101
Kosaric, N., Asher, Y.: The Utilization of Cheese Whey and its Components. Vol. 32, p. 25
Kosaric, N., Duvnjak, Z., Stewart, G. G.: Fuel Enthanol from Biomass Production, Economics, and
Energy. Vol. 20, p. 119
198 Author Index Volumes 1-48

Kosaric, N. see Magee, R. J. Vol. 32, p. 61

Kosaric, N., Wieczorek, A., Cosentino, G. P., Duvnjak, Z.: Industrial Processing and Products from
the Jerusalem Artichoke. Vol. 32, p. 1
Kosaric, N., Zajic, J. E.: Microbial Oxidation of Methane and Methanol. Vol. 3, p. 89
Kosaric, N. see Zajic, K. E. Vol. 9, p. 57
Kosaric, N. see Turcotte, G. Vol. 40, p. 73
Kosaric, N., Blaszczyk, R.: Microbial Aggregates in Anaerobic Wastewater Treatment. Vol. 42, p. 27
Kossen, N. I4/. F. see Metz, B. Vol. 11, p. 103
Kren, V.: Bioconversions of Ergot Alkaloids. Vol. 44, p. 123
Kristapsons, M. Z. see Viesturs, U. Vol. 21. p. 169
Kroner, K. H. see Kula, M.-R. VoL 24, p. 73
Kubicek, C. P.: The Cellulase Proteins of Trichoderma reesei: Structure, Multiplicity, Mode of
Action and Regulation of Formation. VoL 45, p. 1
Kula, M.-R. see Elaschel, E. Vol. 26, p. 73
Kula, M.-R., Kroner, K. H., Hustedt, H.: Purification of Enzymes by Liquid-Liquid Extraction.
Vol. 24, p. 73
Kumar, P. K. R. see Singh, A. Vol. 45, p. 29
Kurtzmann, C. P.: Biotechnology and Physiology of the D-Xylose Degrading Yeast Pachysolen
tannophilus. Vol. 27, p. 73

L/iufer, A. see Syldatk, Ch. Vol. 41, p. 29

Lafferty, R. M. see Schlegel, H. G. Vol. p. 143
Lambe, C. A. see Rosevear, A. Vol. 31, p. 37
Lang-Hinrichs, Ch. see Esser, K. Vol. 26, p. 143
Lee, K. J. see Rogers, P. L. Vol. 23, p. 37
Lee, Y.-H. see Fan, L. T. Vol. 14, p. 101
Lee, Y.-H. see Fan, L. T. Vol. 23, p. 155
Lee, Y.-H., Fan, L. T., Fan, L. S.: Kinetics of Hydrolysis of Insolube Cellulose by Cellulase, Vol. 17,
p. 131
Lee, K-H., Fan, L. T.: Properties and Mode of Action of Cellulase, Vol. 17, p. 101
Lee, Y.-H., Tsao, G. T.: Dissolved Oxygen Electrodes. Vol. 13, p. 35
Lehmann, J. see Schfigerl, K. Vol. 8, p. 63
Lenz, R. W. see Brandl, H. Vol. 41, p. 77
Levitans, E. S, see Viesturs, U. Vol. 21, p. 169
Liefke, E. see Onken, U. Vol. 40, p. 137
Lillehoj, E. P., Malik, V. S.: Protein Purification. Vol. 40, p. 19
Lim, H. C. see Agrawal, P. Vol. 30, p. 61
Lira, H. C. see Parulekar, S. J. Vol. 32, p. 207
Linko, M.: An Evaluation of Enzymatic Hydrolysis of Cellulosic Materials. Vol. 5, p. 25
Linko, M.: Biomass Conversion Program in Finland, Vol. 20, p. 163
Liras, P. see Martin, J. F. Vol. 39, p. 153
Low, K.-S., see Harbour, C. Vol. 37, p. 1
Liicke, J. see Schtigerl, K. Vol. 7, p. 1
Lficke, J. see Schtigerl, K. Vol. 8, p. 63
Luong, J. H. T., Volesky, B.: Heat Evolution During the Microbial Process Estimation, Measure-
ment, and Application. gol. 28, p. 1
Luong, J. H. T., Nguyen, A. L.: Novel Separations Based on Affinity Interactions. Vol. 47, p. 137
Luttman, R., Munack, A., Thoma, M.: Mathematical Modelling, Parameter Identification and
Adaptive Control of Single Cell Protein Processes in Tower Loop Bioreactors. Vol. 32, p. 95
Lynd, L. R.: Production of Ethanol from Lianocellulosic Materials Using Thermophilic Bacteria:
Critical Evaluation of Potential and Review. Vol. 38, p. 1
Lynn, J. D. see Acton, R. T. Vol. 7, p. 85

MacLeod, A. J.: The Use of Plasma Protein Fractions as Medium Supplements for Animal Cell
Culture. Vol. 37, p. 41
Magee, R. J., Kosaric, N.: Bioconversion of Hemicellulosics. Vol. 32, p. 61
Maiorella, B., Wilke, Ch. R., Blanch, H. W.: Alcohol Production and Recovery. Vol. 20, p. 43
Mizlek, I.: Present State and Perspectives of Biochemical Engineering: Vol. 3, p. 279
Maleszka, R. See Schneider, H. Vol. 27, p. 57
Author Index Volumes 1-48 199

Malik, V. S. see Liltehoj, E. P. Vol. 40, p. 19

Mandels, M.: The Culture of Plant Cells. Vol. 2, p. 201
Mandels, M. see Reese, E. T. Vol. 2, p. 181
Mangold, H. K. see Radwan, S. S. Vol. 16, p. 109
Marison, I. W. See yon Stockar, U. Vol. 40, p. 93
Markkanen, P. see Enari, T. M. Vol. 5, p. 1
Marsili-Libelli, St.: Modelling, Identification and Control of the Activated Sludge Process. Vol. 38,
p. 89
Martin, J.-F.: Control of Antibiotic Synthesis by Phosphate. Vol. 6, p. 105
Martin, J.-F., Liras, P.: Enzymes Involved in Penicillin, Cephalosporin and Cephamycin Biosyn-
thesis: Vol. 39, p. 153
Matrin, P. see Zajic, J. E. Vol. 22, p. 51
Mattiasson, B., Hakanson, H.: Immunochemically Based Assays for Process Control. Vol. 46, p. 81
McCracken, L. D., Gong, Ch.-Sh.: D-Xylose Metabolism by Mutant Strain of Candida sp. Vol. 27, p.
33
Miller, O. A., Menozzi, F. D., Dubois, D.: Microbeads and Anchorage- Dependent Eukaryotic Cells:
The Beginning of a New Era in Biotechnology. Vol. 39, p. 73
Meijer, E. M. see Kamphuis, J. Vol. 42, p. 133
Menozzi, F. D. see Miller, O. A. Vol. 39, p. 73
Merchuk, J. C.: Shear Effects of Suspended Cells. Vol. 44, p. 65
Messing, R. A.: Carriers for Immobilized Biologically Active Systems. Vol. 10, p. 51
Metz, B., Kossen, N. W. F., van Suijidam, J. C.: The Rheology of Mould Suspensions. Vot. 11, p. 103
Misawa, M.: Production of Useful Plant Metabolites. Vol. 31, p. 59
Miura, Y.: Submerged Aerobic Fermentation. Vol. 4, p. 3
Miura, Y.: Mechanism of Liquid Hydrocarbon Uptake by Microorganisms and Growth Kinetics
Vol. 9, p. 31
Montague, G. A. see Aynsley, M. Vol. 48, p. 1
Moo- Young, M., Blanch, H. W.: Design of Biochemical Reactors Mass Transfer Criteria for Simple
and Complex Systems. Vol. 19, p. 1
Moo-Young, M. see Scharer, J. M. Vol. 11, p. 85
Morandi, M., Valeri, A.: Industrial Scale Production of J3-Interferon. Vol. 37, p. 57
Morgan, H. W. see Coolbear. T. Vol. 45, p. 57
Morris, A. J. see Aynsley, M. Vol. 48, p. 1
Miiller, R. see Syldatk, Ch. Vol. 41, p. 29
Munack, A. see Luttmann, R. Vol. 32, p. 95

Nagai, S.: Mass and energy Balances for Microbial Growth Kinetics. Vol. 11, p. 49
Nagamune, T. see Endo, I. Vol. 42, p. 1
Nagatani, M. see Aiba S. Vol. 1, p. 31
Nakajima, H. see Tanaka, A. Vol. 42, p. 97
Nakamura, I. see Kamihara, T. Vo[. 29, p. 35
Neubeck, C. E. see Faith, W. T. Vol. 1, p. 77
Neirinck, L. see Schneider, H. Vol. 27, p. 57
Nguyen, A. L. see Luong, J. H. T. Vol. 47, p. 137
Niederauer, M. Q., Glatz, C. E.: Selective Precipitation. Vol. 47, p. 159
Nielsen, J.: Modelling the Growth of Filamentous Fungi. Vol. 46, p. 187
Nyeste, L., P6cs, M., Sevella, B., Hol16, J.: Production of L-Tryptophan by Microbial Processes,
Vol. 26, p. 175
Nyiri, L. K.: Application of Computers in Biochemical Engineering. Vol. 2, p. 49

Ochsner, U. see Reiser, J. Vol. 43, p. 75

O'Driscoll, K. F.: Gel Entrapped Enzymes. Vol. 4, p. 155
Oels, U. see Schfigerl, K. Vol. 7, p. 1
Ohshima, T., Soda, K.: Biochemistry and Biotechnology of Amino Acid Dehydrogenases. Vol. 42, p.
187
Okabe, M. see Aiba S. Vol. 7, p. 111
OIson, N. F. see Finocchiaro, T. Vol. 15, p. 71
Onken, U., Liefke, E.: Effect of total and partial Pressure (Oxygen and Carbon Dioxide) on Aerobic
Microbial processes. Vol. 40, p. 137
200 Author Index Volumes 1 48

Pace, G. W., Righelato, C. R.: Production of Extracellular Microbial. Vol. 15, p. 41

Parisi, F.: Energy Balances for Ethanol as a Fuel. Vol. 28, p. 41
Parisi, F.: Advances in Lignocellulosic Hydrolysis and in the Utillization of the Hydrolyzates.
Vol. 38, p. 53
Parulekar, S. J., Lim, H. C.: Modelling, Optimization and Control of Semi-Batch Bio-reactors.
Vol. 32, p. 207
Pbcs, M. see Nyeste, L. Vol. 26, p. 175
Phillipson, J. D. see Anderson, L. A. Vol. 31, p. 1
Pitcher Jr., W. H.: Design and Operation of Immobilized Enzyme Reactors. Vol. 10, p. 1
Potgieter, H. J.: Biomass Conversion Program in South Africa. Vol. 20, p. 181
Poweigha, T. see Kleinstreuer, C. Vol. 30, p. 91
Primrose, S. B.: Controlling Bacteriophage Infections in Industrial Bioprocesses. Vol. 43, p. 1
Prokop, A., Rosenberg, M. Z.: Bioreactor for Mammalian Cell Culture. Vol. 39, p. 29

Quicker, G. see Schumpe, A. Vol. 24, p. 1

Radlett, P. J.: The Use Baby Hamster Kidney (BHK) Suspension Cells for the Production of Foot
and Mouth Disease Vaccines. Vol. 34, p. 129
Radwan, S. S., Mangold, H. K.: Biochemistry of Lipids in Plant Cell Cultures. Vol. 16, p. 109
Ramasubramanyan, K., Venkatasubramanian, K.: Large-Scale Animal Cell Cultures: Design and
Operational Considerations. Vol. 42, p. 13
Ramakrishna, D.: Statistical Models of Cell Populations. Vol. 11, p. 1
Rapoport, A. J. see Beker, M. J. Vol. 35, p. 127
Reardon, K. F. see Friehs, K. Vol. 48, p. 53
Reese, E. T. see Faith, W. T. Vol. 1, p. 77
Reese, E. T., Mandels, M., Weiss, A. H.: Cellulose as a Novel Energy Source. Vol. 2, p. 181
Rehh~ek, Z.: Ergot Alkaloids and Their Biosynthesis. Vol. 14, p. 33
Rehm, H.-J., Reiff, I.: Mechanism and Occurrence of Microbial Oxidation of Long-Chain Alkanes,
Vol. 19, p. 175
Reiff, I. see Rehm, H.-J. Vol. 19, p. 175
Reinhard, E., Alfermann, A. W.: Biotransformation by Plant Cell Cultures. Vol. 16, p. 49
Reiser, J., Glumoff, V., K/ilin, M., Ochsner, U.: Transfer and Expression of Heterologous Genes in
Yeasts Other Than Saccharomyces cerevisiae. Vol. 43, p. 75
Reuveny, S. see Shahar, A. Vol. 34, p. 33
Richardson, T. see Finocchiaro, T. Vol. 15, p. 71
Righelato, R. C. see Pace, G. W. VoL 15, p. 41
Rivera, S. L. see Karim, M. N. Vol. 46, p. 1
Roberts, M. F. see Anderson, L. A. Vol. 31, p. 1
Roels, J. A. see Harder, A. Vol. 21, p. 55
Rogers, P. L.:. Computation in Biochemical Engineering. Vol. 4, p. 125
Rogers, P. L., Lee, K. L., Skotnicki, M. L., Tribe, D. E.: Ethanol Production by Zymomonas Mobilis.
Vol. 23, p. 37
Rolz, C., Humphrey, A.: Microbial Biomass from Renewables: Review of Alternatives. Vol. 21, p. 1
Rosazza, J. P. see Smith, R. V. Vol. 5, p. 69
Rosenberg, M. Z. see Prokop, A. Vol. 39, p. 29
Rosevear, A., Lambe, C. A.: Immobilized Plant Cells. Vol. 31, p. 37
Ryhiner, G. B. see Heinzle, E. Vol. 48, p. 79

Sahm, H.: Anaerobic Wastewater Treatment. Vol. 29, p. 83

Sahm, H.: Metabolism of Methanol by Yeasts. Vol. 6, 77
Sahmn, H.: Biomass Conversion Program of West Germany. Vol. 20, p. 173
Sambanis, A. see Grampp, G. E. Vol. 46, p. 35
Sasse, F. see Berlin, J. Vol. 31, p. 99
Scharer, J. M., Moo-Young, M.: Methane Generation by Anaerobic Digestion of Cellulose-
Containing Wastes. Vol. 11, p. 85
Schlegel, H. G., Lafferty, R. M.: The Production of Biomass from Hydrogen and Carbon Dioxide
Vol. 1, p. 143
Schmid, R. D.: Stabilized Soluble Enzymes. Vol. 12, p. 41
Author Index Volumes 1-48 201

Schneider, H., Maleszka, R., Neirinck, L., Veliky, I. A., Chan, Y. K., Wang, P. Y.: Ethanol Production
from D-Xylose and Several Other Carbohydrates by Pachysolen tannophiluys. Vol. 27, p. 57
Shoemaker, H. E. see Kamphuis, J. Vol. 42, p. 133
Schrbder, J. see Hahlbrock, K. Vol. 18, p. 39
Schumpe, A., Quicker, G., Deckwer, W. D.: Gas Solubilities in Microbial Culture Meida. Vol. 24, p. l
SchiigerI, K.: Oxygen Transfer Into Highly Viscous Media. Vol. 19, p. 71
Schfigerl, K.: Characterization and Performance of Single- and Multistage Tower Reactors with
Outer Loop for Cell Mass Production, Vol. 22, p. 93
Schfigerl, K., Oels, U., L/icke, J.: Bubble Column Bioreactors. Vol., p. 1
Schiigerl, K., Lficke, J., Lehmann, J., Wagner, F.: Application of Tower Bioreactors in Cell Mass
Production. Vol. 8, p. 63
Schiigerl, K. see Singh, A. Voi. 45, p. 29
Schwab, H.: Strain Improvement in Industrial Microorganisms by Recombinant DNA
Techniques.Vol. 37, p. 129
Seipenbusch, R., Blenke, H.: The Loop Reactor for Cultivating Yeast on n-Praffin Substrate Vol. 15,
p. 1
Seki, M. see Furusaki, S. Vol. 46, p. 161
Sevella, B. see Nyeste, L. Vol. 26, p. 17
Shahar, A., Reuveny, S,: Nerve and Muscle Cells on Microcarriers Culture. Voi. 34, p. 33
Shanahan, J.-F. see Ho, Ch. S. Vol. 35, p. 83
Shieh, W. K., Keenan, J. D.: Fluidized Bed Biofilm Reactor for Wastewater Treatment. Vol. 33,
p. 131
Shimizu, S. see Yaman~, T. Vol. 30, p. 147
Shinmyo, A. see Enatsu, T. Vol. 9, p. 111
Shioya, S.: Optimization and Control in Fed-Batch Bioreactors. Vol. 46, p. 111
Singh, A., Kumar, P. K. R., Sch/igerl, K.: Bioconversion of Cellulosic Materials to Ethanol by
Filamentous Fungi. Vol. 45, p. 29
Sitti 9, W., see Faust, U. Vol. 17, p. 63
Skotnicki, M. L. see Rogers, P. L. Vol. 23, p. 37
Smith, M. D. see Ho, Ch. S. Vol. 35, p. 83
Smith, R. V., Acosta Jr., D., Rosazza, J. P.: Cellular and Microbial Models in the Investigation of
Mammalian Metabolism of Xenobiotics. Vol. 5, p. 69
Smith, R. V., Davis, P. J.: Induction of Xenobiotic Monooxygenases. Vol. 14, p. 61
Soda, K. see Yonaha, K. Vol. 33, p. 95
Soda, K. see Ohshima, T. Vol. 42, p. 187
Solomon, B.: Starch Hydrolysis by Immobilized Enzymes. Industrial Application. Vol. 10, p. 131
Somers, P. J. see Barker, S. A. Vol. 10, p. 27
Sonnleitner, B. Biotechnology of Thermophilic Bacteria: Growth, Products, and Application.
Vol. 28, p. 69
Sonnleitner, B. see Fiechter, A. Vol. 46, p. 143
Spiegel-Roy, P., Kochba, J.: Embryogenesis in Citrus Tissue Cultures. Vol. 16, p. 27
Spier, R. E.: Rficent Developments in the Large Scale Cultivation of Animal Cells in Monolayers.
Vol. 14, p. 119
Stephanopoulos, G. N. see Grampp, G.E. Vol. 46, p. 35
Stewart, G. G. Kosaric, N. Vol. 20, p. 119
yon Stockar, U., Marison, I. W.: The Use of Calorimetry in Biotechnology. Vol. 40, p. 93
Stohs, S. J.: Metabolism of Steroids in Plant Tissue Cultures. Vol. 16, p. 85
Strandberg, L. see Enfors, S.-O. Vol. 43, p. 31
Sudo, R., Aiby, S.: Role and Function of Protozoa in the Biological Treatment of Poluted Waters.
Vol. 29, p. 117
Suijidam, van, J. C. see Metz, N. W. Vol. 11, p. 103
Sureau, P.: Rabies Vaccine Production in Animal Cell Cultures. Vol. 34, p. 111
Syldatk, Ch., L/iufer, A., Mfiller, R., H6ke, H.: Production of Optically Pure D- and L-m-Amino
Acids by Bioconversion of D,L-5-Monosubstuted Hydantoin Derivatives. Vol. 41, p. 29
Szezesny, T. see Volesky, B. Vol. 27, p. 101

Tachikawa, S. see Endo, I. Vol. 42, p. 1

Taguchi, H.: The Nature of Fermentation Fluids. Vol. 1, p. 1
Takeuchi, 72 see Karube, I. Vol. 46, p. 63
202 Author Index Volumes l 48

Tanaka, A. see Fukui, S. Vol. 17, p. 1 and Vol. 19, p. 217

Tanaka, A. see Fukui. S. Vol. 29, p. 1
Tanaka, A., Nakajima, H.: Application of Immobilized Growing Cells. Vol. 42, p. 97
Taylor, D C. see Weber N. Vol. 45, p. 99
Thoma, M. see Luttmann, R. Vol. 32, p. 95
Topiwala, H. H. see Harrison, D. E. F. Vol. 3, p. 167
Torma, A. E.: The Role of Thiobacillus Ferrooxidans in Hydrometallurgical Processes. Vol. 6, p. 1
Tran Than Van, K.: Control of Morphogenesis or What Shapes a Group of Cells? Vol. 18, p. 151
Tribe, D. E. see Rogers, P. L. Vol. 23, p. 37
Truei, Y.-H., Gu T., Tsai, G.-J., Tsao, G. T.: Large-Scale Gradient Elution Chromatography. Vol. 47,
p. 1
Tsai, G.-J. see Truei, Y.-H. Vol. 47, p. 1
Tsao, G. T. see Lee, Y. H. Vol. 13, p. 35
Tsao, G. T. see Chang, M. M. Vol. 20, p. 93
Tsao, G. T. see Jang, C.-M. Vol. 25, p. 1
Tsao, G. T. see Jang, C.-M. Vol. 25, p. 19
Tsao, G. T. see Jansen, N. B. Vol. 27, p. 85
Tsao, G. T. see Truei. Y.-H. Vol. 47, p. 1
Tschopp, A. see Cogoli, A. Vol. 22, p. 1
Turcotte, G., Kosaric0 N.: Lipid Biosynthesis in Oleaginous Yeasts. Vol. 40, p. 73

Uhlmann, E. see Engels, J. Vol. 37, p. 73

Underhill, E. W. see Weber, N. Vol. 45, p. 99
Ursprung, H.: Biotechnology: The New Change for Industry. Vol. 30, p. 3

Valeri, A. see Morandi, M. Vol. 37, p. 57

Van Balken, J. A. M. see Kamphuis, J. Vot. 42, p. 133
Vallander, L., Eriksson, K.-E. L.: Production of Ethanol from Lignocellulosic Materials: State of the
Art. Vol. 42, p. 63
Vanek, Z. see Hostalek, Z. Vol. 3, p. 13
Veicle, A. see Enfors, S.-O. Vol. 43, p. 31
Veliky, I. A. see Schneider, H. Vol. 27, p. 57
Venkatasubramanian, K. see Ramasubramanyan K. Vol. 42, p. 13
Vieregge, L. see Hahlbrock, K. Vol. 18, p. 39
Viesturs U. E., Kristapsons, M. Z., Levitans, E. S,, Foam in Microbiological Processes. Vol. 21,
p. 169
Volesky, B., Szczesny, T.: Bacterial Conversion of Pentose Sugars tO Acetone and Butanol. Vol. 27,
p. 101
VoIesky, B. see Luong, J. H. T. Vol. 28, p. 1

Wagner, F. see Schfigerl, K. Vol. 8, p. 63

Wandrey, Ch., Flaschel, E.: Process Development and Economic Aspects in Enzyme Engineering
Acylase L-Methionine System. Vol. 12, p. 147
Wandrey, Ch. see Flaschel, E. Vol. 26, p. 73
Wang, P. J., Hy, C. J.: Regeneration of Virus-Free Plants Through in Vitro Culture. Vol. 18, p. 61
Wang, P. Y. see Schneider, H. Vol. 27, p. 57
Wang, S. S., King, C.-K.: The Use of Coenzymes in Biochemical Reactors. Vol. 12, p. 119
Weber, N., Taylor, D. C.., Underhill, E. W.: Biosynthesis of Storage Lipids in Plant Cell and Embryo
Cultures. Vol. 45, p. 99
Weiss, A. H. se Reese, E. T., Vol. 2, p. 181
Wieezorek, A. see Kosaric, N. Vol. 32, p. 1
Wilke, Ch. R., see Maiorella, B. Vol. 20, p. 43
Wilson, G.: Continuous Culture of Plant Cells Using the Chemostat Principle. Vol. 16, p. 1
Wingard Jr., L. B.: Enzyme Engineering Vol. 2, p. 1
Wiesmann, U., Binder, H.: Biomass Separation from Liquids by Sedimentation and Centrifugation.
Vol. 24, p. 119
Withers, L. A.: Low Temperature Storage of Plant Tissue Cultures. Vol. 18, p. 101
Wu, Y.-T. see Graves, D. L. Vol. 12, p. 219
Author Index Volumes 1~48 203

Yamada, Y.: Photosynthetic Potential of Plant Cell Cultures. Vol. 31, p. 89

Yaman~, T., Shimizu, S.: Fed-batch Techniques in Microbial Processes. Vol. 30, p. 147
Yarovenko, V. L.: Theory and Practice of Continuous Cultivation of Microorganisms in Industrial
Alcoholic Processes. Vol. 9, p. 1
Yonaha, K., Soda, K.: Applications of Stereoselectivity of Enzymes: Synthesis of Optically Active
Amino Acids and c~-Hydroxy Acids, and Stereospecific Isotope- Labeling of Amino Acids, Amines
and Coenzymes. Vol. 33, p. 95
Yoshida, 7: see Konstantinov, K. B. Vol. 48, p. 169

Zajic, J. E. see Kosaric, N. Vol. 3, p. 89

Zajic, J. E. see Jack, T. R. Vol. 5, p. 125
Zajic, J. E., Kosaric, N., Brosseau, J. D.: Microbial Production of Hydrogen. Vol. 9, p. 57
Zq]ic, J. E., Inculet, I. I., Martin, P.: Basic Concept in Microbial Aerosols. Vol. 22, p. 51
Zlokarnik, M.: Sorption Characteristics for Gas-Liquid Contacting in Mixing Vessels. Vol. 8, p. 133
Zlokarnik, M.: Scale-Up of Surface Aerators for Waste Water Treatment. Vol. 11, p. 157
Subject Index

Acetic acid 84, 86, 87, 90 Fatty acids, volatile 107

Acetogenesis 84, 85 Flow cytometry 33, 34
Acidogenesis 84 Focus-of-attention mechanism 182
Adaptive control 106 Fuzzy logic 171, 187
Aggregate formation, plant cell 120-121
Airlift reactor 124-146 Garbage collection, incremental 176
Algorithms, control 171, 173 Gas production rate 98
, external, user-defined 180
Ammonia 91, 92 Hairy-root culture, reactors 154 156
Anaerobic contact 101 Heat shock response 35
Artificial neural networks, applications 20-24 Host-vector interactions 29, 33, 47
, architecture 19 - system 32
Hydrogen sulfide 92
Bacillus subtilis 69 Immobilized plant cells 147-154
Bingham fluid 118 Inclusion bodies 45, 57, 58, 65 68, 71
Biofilm 86, 94 Inference chain 183, 185
Biomass yield, 83, 89 Intelligent monitoring 173, 187
Bioscan, real-time knowledge based system Interface, developer/end-user 173, t85
7-16
Butyric acid 84, 86, 90 Kalman filter 100
Knowledge base, debugging 182, 184
- _ editing 184
C, computer language 175
- -, scheduling 8-9
Cell disruption 67, 68 -, size 187
Chaperone 40, 66
-

-, structuring 172, 181, 187

Computer cell 48
-

[3-Lactamase 39, 40, 42, 64, 69, 70

DOS 177 Lactic acid 87
Lisp 175, 176
E. coli 31, 33-46, 49 51, 178, 183 Membrane energetics 42
, mRNA 58, 60 63 Methane, production 85, 97, 98
-, periplasmic space 63-65, 67 71 Methanobacterium 86
, phosphatase 69, 70 Methanogenesis 85, 86
-, plasmid 55 59, 64, 68, 71 73 Monod-type kinetics 83-84, 89 91
-, promoter 55, 58, 60, 61, 66, 68 Multimedia 186
-, ribosomal binding site 62 Mycelial process 5 6
-, transcription 60 63, 68
Embryo culture, artificial seed 156 Neural networks 171, 180, /86
Enterobacteria 86 -, artificial 19-24
ES (expert system) 170
ESDT (expert system development tool) 170, Off-gas 100
188, 189 Operating systems, DOS/UNIX/real-time 177
Ethanol 86 Organic acids 85, 86, 99, 100
Expert control 171, 172 Overpopulation lethality 41
- system/shell 170 Oxygen limitation 44, 45
 206 Subject Index

Penicillin process 5 RNA polymerase 34, 35, 38, 49, 50

Phage 32, 38 Rotating-drum reactor, hairy-root culture t55
- T7 polymerase 38 --, plant-cell culture 147
Phosphatase 69, 70
Phytochemicals, embryo culture 156 157 Self-immobilized plant cells 151
PID 99-105 Shikonin 117, 123
Plant cells, aggregate formation 120 121 Shoot culture 157 158
- -, immobilized 147-154 Sludge retention time 88
- - , photoautotrophic, reactors 159 160 Staphylococcus aureus 69
- ,self-immobilized 151 Streptomyces lividans 69
- -, shear damage 121 Stress response 35, 37, 38, 41
- - , shear sensitivity 121-123, 144 Sulfur 89
- -, suspension, non-Newtonian fluid 118-119 Supervisory commands 172
Plasmid 32-36, 38M5, 47-51 S ynthrophobacter 86
Population balance 51 Syntrophomonas 86
Porin 68 System architecture, embedded/interfaced 173
Power-law fluid 118-119, 129
Prolog 175 Time constant, analysis for plant-cell culture
Promoter 34, 35, 38, 40, 41, 45, 49, 50 137 142, 144-145
Propanol 86 Toxicity, anaerobic wastewater 83, 86, 99, 100
Propionic acid 84, 87, 90 Transcription 34, 35, 38, 39, 45
- bacteria 86 Turnover 40, 41
Protease 37, 41, 46, 63-65, 68, 71
Proteolysis 63, 64, 68
UNIX t77
R B S 62
Real-time knowledge based systems 3, 6, 7, Viscosity, apparent, pseudoplastic fluids 127,
10 16 131, 134
Replication 32-34, 40 Vitreoscilla hemoglobin 45
Rheology, plant-cell suspension 118-120
Ribosomal RNA 35, 38 Wastewater, anaerobic 83ff
Ribosome 35, 38, 39, 48, 49
-, specialized 38 Yield stress, plant-cell suspension 119