Understanding the Dynamics of Biological Systems

Werner Dubitzky  Jennifer Southgate

Hendrik Fuß
Editors

Understanding the Dynamics

of Biological Systems
Lessons Learned from Integrative
Systems Biology

ABC
Editors
Werner Dubitzky Hendrik Fuß
Nano Systems Biology Research Group Nano Systems Biology Research Group
Biomedical Sciences Research Institute Biomedical Sciences Research Institute
University of Ulster University of Ulster
Coleraine BT52 1SA Coleraine BT52 1SA
United Kingdom United Kingdom
[email protected] [email protected]

Jennifer Southgate
Department of Biology
Jack Birch Unit of Molecular Carcinogenesis
University of York
York YO10 5DD
United Kingdom
[email protected]

ISBN 978-1-4419-7963-6 e-ISBN 978-1-4419-7964-3

DOI 10.1007/978-1-4419-7964-3
Springer New York Dordrecht Heidelberg London

c Springer Science+Business Media, LLC 2011

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Systems biology could be defined as the quantitative analysis of the dynamic

interactions among several components of a biological system and aims to under-
stand the behavior of the system as a whole. R&D in systems biology involves the
development and application of systems theory concepts for the study of complex
biological systems through iteration over mathematical modeling and computational
simulation and biological experimentation. Systems biology could be viewed as a
tool to increase understanding of biological systems and to develop more directed
experiments and finally allow predictions.
The field of systems biology arose out of a biological problem which is essen-
tially entailed by the complexity of biological life. It was created because of the
limitations of conventional (reductionistic) biology in the investigation and under-
standing of complex biological phenomena arising from the dynamic interaction
of many biological compounds. At present, a large number of individual genes or
proteins which play key roles in essential physiological processes are known. For
many of these, structural data and detailed mechanistic descriptions at a molecular
level are available. In most cases, however, the individual characterization of these
molecules is not sufficient to fully understand their immediate or their superordi-
nate physiological function. Similarly, large networks of genes, proteins and other
organic molecules have been discovered, mapped and characterized. While underly-
ing mechanisms have been regarded as a promising base for explaining the multitude
of cellular functions and phenomena observed in vivo, there is still a fundamental
gap between the knowledge of a molecular mechanism and the understanding of the
corresponding cellular or higher-level function.
The growing field of systems biology promises to bridge our current gap in un-
derstanding. Systems biology views biological function and macroscopic behavior
as an emergent or supervenient property – i.e., a property that a collection of com-
ponents or complex system possesses but which the individual constituents do not
have. The properties of individual elements, such as proteins, are investigated in the
context of the whole, complex system of interactions. The different spatial and tem-
poral scales involved in biological processes – ranging from the level of molecules
through to organisms and, ultimately, to the level of entire populations or ecosystems
– permit upward and downward causation in complex arrangements of feedback
loops. Systems-level properties arise from interconnected processes on multiple

v
vi Preface

scales of temporal and spatial organization. Understanding such complexity is a

major challenge to the unaided human brain. Thus, using mathematical and compu-
tational models, systems biologists integrate elementary processes of systems into a
coherent description that allows them to predict and characterize the systems-level
properties and behavior of complex biological phenomena.
As the field of systems biology matures, we are beginning to see practical an-
swers to real biological problems. We believe it is now time to step back and review
some of the approaches of systems biology to concrete problems. This volume in-
troduces some of the main methods and techniques of systems biology and assesses
their pros and cons based on concrete case studies. The investigated biological phe-
nomena include tissue organization, hormonal control, bacterial stress response,
tumor growth and cellular metabolism. Each chapter and the book as a whole is
intended to simultaneously serve as design blueprint, user guide, research agenda,
and communication platform.
As design blueprint, the book is intended for biologists, mathematicians and sys-
tems scientists, computer scientists and technology developers, managers, and other
professionals who consider adopting a systems biology approach.
As user guide, this volume addresses the requirements of scientists and re-
searchers to gain an overview and a basic understanding of key systems biology
methodologies and tools. For these users, we seek to explain the key concepts and
assumptions of the various techniques, their conceptual and computational merits
and limitations, and, where possible, give guidelines for choosing the methods and
tools most appropriate to the task at hand. Our emphasis is not on a complete and
intricate formal and technical treatment of the presented methodologies. Instead, we
aim at providing the users with a clear understanding and practical know-how of the
relevant methods in the context of concrete life science problems.
As research agenda, the book is intended for computer and life science students,
teachers, researchers, and managers who seek to understand the state of the art of the
methodologies used in systems biology research and development. To achieve this,
we have attempted to cover a representative range of life science areas and systems
biology methodologies, and we have asked the authors to identify areas in which
gaps in our knowledge demand further research and development.
The book is also intended as a communication platform to bridge the cultural,
conceptual, and technological gap among the key systems biology disciplines of
biology, mathematics, and information technology. To support this goal, we have
asked the contributors to adopt an approach that appeals to audiences from different
backgrounds.
Providing a representative overview of current research, this book aims to illus-
trate the insights gained by adopting a systems biology approach. While systems
biologists typically apply mathematical, statistical, and computational methods,
these insights are presented in the context of current life science research. As a
result, this book is targeted at an interdisciplinary audience comprising life scien-
tists, mathematicians, system and computer researchers, and developers. In pursuing
these goals, the book seeks to bridge the cultural, conceptual, and technological gap
among the key disciplines that contribute to systems biology.
Preface vii

Table 1 Classification of modeling formalisms: examples

Deterministic Stochastic
Continuous ODE, PDE SDE
Discrete Boolean network, cellular automaton Agent-based simulation

In recent years, the increased interest of computer scientists in systems biology

has led to an explosion of novel systems methodologies for modeling, analysis, and
validation, but also for model representation and exchange. In this book, we do not
intend to cover a wide variety of these methods, but we aim to present illustrative
applications of systems biological methods in a representative overview.
In any modeling discipline, modeling formalisms may be classified according to
the type of representation chosen to model time, space, and entities (such as the
cell, proteins, or genes) of the system. These entities or dimensions can be mod-
eled as continuous variables, so that the model can cope with any value within a
meaningful range. Table 1 illustrates this. Continuous means that the model may
output a simulation result at any given time point, t (continuous time) and loca-
tion, x (continuous space), and that the output of the model may assume any value
within a predefined range. In contrast, discrete refers to a modeling strategy that
uses distinct values from a predefined set to represent time, space, and the entities
of the modeled system. The output of a time-discrete model is limited to certain
time intervals; a space-discrete model can explore only certain points in a given
space; and discrete variables express levels or predefined states (on/off, low/high,
cell cycle phase) of the modeled entities. Clearly, any of the combination of discrete
and continuous methods is possible. An agent-based simulation can be backed by a
time-continuous, space-discrete model with agents that are represented using both
continuous and discrete variables. Discrete methodologies sometimes deviate from
the classification shown in Table 1. The most common cases are shown in the table.
Systems biology modeling methodologies may also be divided into deterministic
and stochastic formalisms. Consider a set of interacting cells which behave accord-
ing to certain rules. In reality, the observation of randomly picked single cells may
lead to grossly varying observations; although when looking at a large number of
cells, they all share the same characteristic behavior. Deterministic simulations deal
with this problem by modeling only those characteristics; the stochastic approach,
in contrast, considers a large number of individual simulations and uses statistical
analysis to draw conclusions.
Below we provide a brief overview of the contributed chapters in terms of the
modeling methodology used and the biological problems addressed.
The modeling framework that was probably the first to be adapted for systems
biological modeling – before the term systems biology was even coined – is the
mathematical framework with the longest tradition: differential equations modeling,
or more concrete, ordinary differential equations (ODEs). The ODE methodology
offers a variety of basic, mathematical, and computational tools for modeling, sim-
ulation, and qualitative and quantitative analysis.
viii Preface

Chapter 1 presents two elementary case studies that illustrate ODE-based model
definition as well as timescale analysis and sensitivity analysis. These analysis meth-
ods can be used to extract biologically meaningful information from the model. In
the study, the authors measure the efficiency of the simulated cell’s protein-folding
machinery under various conditions using timescale analysis.
While ODEs offer a general and flexible approach to modeling, this methodology
relies on a qualitatively and quantitatively exact definition of the molecular network
or system to be represented. Chapter 2 illustrates some of the most common mathe-
matical tools in an ODE-based case study relating to folate metabolism.
Chapter 3 presents a delay differential equations (DDE) model of hormonal
control of the menstrual cycle. This study demonstrates that it is sometimes more
interesting to characterize the behavior of a system in relation to its inputs and pa-
rameters, than to just reproduce its outputs using concrete parameter values.
Pharmacokinetic models, most of which are ODE-based, have become an estab-
lished tool in pharmacology. Such models have become an important tool in drug
development to predict the fate of drugs or toxins taken in by the human body.
Chapter 4 introduces this field and highlights the problem of investigating active
transport phenomena.
The studies presented in Chaps. 3 and 4 rely on a reasonably well-established
body of quantitative data. However, in the majority of cases, sufficient amounts of
data are currently not available to systems biologists. The need to abstract from con-
crete sets of parameters has therefore led to the development of different modeling
methods. Piece-wise linear (PL) equations, introduced in Chap. 6, are one example.
Based on ODEs, they divide the entire parameter space into parts that share the same
qualitative behavior. This behavior is approximated using only simple, linear equa-
tions, as opposed to the nonlinear equations that typically arise in complex ODE
systems. This property makes PL models mathematically more tractable.
Flux balance analysis (FBA) is another useful tool in pharmacological applica-
tions of systems biology. An FBA model can predict metabolic activities (fluxes)
under homeostatic conditions. Knowing the relevant metabolites and the stoichiom-
etry of all reactions in the system is sufficient for performing such an analysis. FBA
permits comprehensive studies of qualitative structural changes in the network, such
as deletion of arbitrary genes throughout the genome. Chapter 5 presents an FBA
case study concerned with the metabolism and pathogenicity of Mycobacterium
tuberculosis. The overall goal of the effort is to systematically and efficiently de-
sign anti-tuberculosis drugs. Toward this goal, this chapter also illustrates how other
techniques, besides FBA, can be used. The use of graph-theoretical techniques are
illustrated for analyzing the protein–protein interaction networks, to gain insights
about strategic hub proteins and possible of routes of information flow in triggering
drug resistance. Boolean network modeling, another technique gaining popularity
for studying biological systems, has been used for studying host–pathogen interac-
tions, in this case leading to qualitative understanding of the complex interplay of
the bacterial components with the human immune system.
Another modeling technique which is growing in popularity is the agent-based
model (or individual-based model). Chapter 7 illustrates this methodology with an
Preface ix

application to the problem of bacterial antibiotic resistance. In this model, each cell
is represented as an agent, which moves and interacts with other agents according
to a defined set of rules. The agent paradigm is well suited to investigating the
mechanisms of emergent spatial patterns. This is also discussed in Chap. 8, where an
agent-based model is used to mimic the assembly of microtubules into the mitotic
spindle at cell division.
Since different modeling methodologies are typically suited for different scales
of time and space, it is an appealing proposition to build multi-scale models, where
multiple modeling techniques applied to different aspects of the same biological
problem integrate into a single, integrated model. The agent-based modeling ap-
proach permits the use of arbitrary modeling methods for defining the rule sets by
which the agents are governed. This is illustrated in Chap. 9, where agents are used
to model the behavior of epithelial tissue.
Finally, Chap. 10 uses an entirely different approach to investigate a problem in
synthetic biology. In this discipline, biological molecules are used to engineer func-
tional entities such as logic circuits. In this study, a domain-specific programming
language helps to model and define the behavior of this engineered component.

Coleraine Werner Dubitzky

August, 2010 Jenny Southgate
Hendrik Fuß
Contents

1 Effects of Protein Quality Control Machinery

on Protein Homeostasis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 1
Conner I. Sandefur and Santiago Schnell

2 Metabolic Network Dynamics: Properties and Principles . . . . . . . . . . . . . . . 19

Neema Jamshidi and Bernhard Ø. Palsson

3 A Deterministic, Mathematical Model

for Hormonal Control of the Menstrual Cycle . . . . . . . . . . .. . . . . . . . . . . . . . . . . 39
R. Drew Pasteur and James F. Selgrade

4 Modeling Transport Processes and Their Implications

for Chemical Disposition and Action .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 59
Nick Plant

5 Systems Biology of Tuberculosis: Insights

for Drug Discovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 83
Karthik Raman and Nagasuma Chandra

6 Qualitative Analysis of Genetic Regulatory Networks in

Bacteria . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .111
Valentina Baldazzi, Pedro T. Monteiro, Michel Page,
Delphine Ropers, Johannes Geiselmann, and Hidde de Jong

7 Modeling Antibiotic Resistance in Bacterial Colonies

Using Agent-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .131
James T. Murphy and Ray Walshe

8 Modeling the Spatial Pattern Forming Modules in Mitotic

Spindle Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .155
Chaitanya A. Athale

xi
xii Contents

9 Cell-Centred Modeling of Tissue Behaviour .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . .175

Rod Smallwood

10 Interaction-Based Simulations for Integrative Spatial

Systems Biology .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .195
Antoine Spicher, Olivier Michel, and Jean-Louis Giavitto

Glossary . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .233

Index . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .237
Contributors

Chaitanya A. Athale EMBL, Meyerhofstrasse 1, Heidelberg 69117, Germany

and
IISER Pune, IISER, Central Tower, Sai Trinity Building, Sutarwadi Road, Pashan,
Pune 411021, India, [email protected]; [email protected]
Valentina Baldazzi INRIA Grenoble – Rhône-Alpes, France,
[email protected]
Nagasuma Chandra Bioinformatics Centre, Indian Institute of Science,
Bangalore 560 012, India, [email protected]
Hidde de Jong INRIA Grenoble – Rhône-Alpes, France,
[email protected]
Johannes Geiselmann Université Joseph Fourier, Grenoble, France,
[email protected]
Jean-Louis Giavitto IBISC Lab, FRE 3190 CNRS, Université d’Évry
& Genopole, 523 place des terrasses de l’agora, 91000 Évry, France,
[email protected]
Neema Jamshidi Department of Bioengineering, University of California,
San Diego, 9500 Gilman Drive, La Jolla, CA 92093-0142, USA, [email protected]
Olivier Michel LACL – EA 4219 – Université de Paris 12, Paris EST 61 avenue
du Général de Gaulle, 94010 Créteil Cedex, France,
[email protected]
Pedro T. Monteiro INRIA Grenoble – Rhône-Alpes, France
and
IST/INESC-ID, 9 Rua Alves Redol, 1000-029 Lisbon, Portugal,
[email protected]
James T. Murphy Centre for Scientific Computing and Complex Systems
Modeling, School of Computing, Dublin City University, Dublin 9, Ireland,
[email protected]
Michel Page INRIA Grenoble – Rhône-Alpes, France, [email protected]

xiii
xiv Contributors

Bernhard Ø. Palsson Department of Bioengineering, University of California,

San Diego, 9500 Gilman Drive, La Jolla, CA 92093-0142, USA,
[email protected]
R. Drew Pasteur Department of Mathematics and Computer Science, The College
of Wooster, 1189 Beall Ave., Wooster, OH 44691, USA, [email protected]
Nick Plant Centre for Toxicology, Faculty of Health and Medical Sciences,
University of Surrey, Guildford, Surrey GU2 7XH, UK, [email protected]
Karthik Raman Bioinformatics Centre, Indian Institute of Science, Bangalore
560 012, India, [email protected]
Delphine Ropers INRIA Grenoble – Rhône-Alpes, France,
[email protected]
Conner I. Sandefur Center for Computational Medicine and Bioinformatics,
University of Michigan, 2017 Palmer Commons, 100 Washtenaw Ave, Ann Arbor,
MI 48105, USA, [email protected]
Santiago Schnell Department of Molecular and Integrative Physiology, Center
for Computational Medicine and Bioinformatics, Brehm Center for Type 1
Diabetes Research and Analysis, University of Michigan, 2017 Palmer Commons,
100 Washtenaw Ave, Ann Arbor, MI 48105, USA, [email protected]
James F. Selgrade Department of Mathematics and Biomathematics Program,
North Carolina State University, Raleigh, NC 27695-8205, USA,
[email protected]
Rod Smallwood Department of Computer Science, University of Sheffield,
Regent Court, 211 Portobello, Sheffield S1 4DP, UK, [email protected]
Antoine Spicher LACL – EA 4219 – Université de Paris 12, Paris EST – 61
avenue du Général de Gaulle, 94010 Créteil Cedex, France,
[email protected]
Ray Walshe Centre for Scientific Computing and Complex Systems Modeling,
School of Computing, Dublin City University, Dublin 9, Ireland,
[email protected]
Chapter 1
Effects of Protein Quality Control Machinery
on Protein Homeostasis

Conner I. Sandefur and Santiago Schnell

1.1 Protein Folding is Catalyzed by a Complex Network

of Reactions

A driving force of systems biology is the desire to understand the many interactions
that compose the pathways within a cell. Systems biology is interested in the inter-
actions and emergent properties that result from communication between different
system components. Reducing a system (e.g., a cell) to its parts (e.g., individual
genes and proteins) neglects component interaction and emergent properties. Build-
ing and investigating a complete interaction map provides insight into normal and
diseased individuals that might not be found by traditional methods.
Much of traditional biology has the central dogma of molecular biology at its
basis. This dogma states that DNA is transcribed into RNA which is translated into
protein (Crick 1970), and has guided the study of individual genes and the proteins
they encode. The protein folding network provides an example of how the central
dogma of molecular biology does not explain many of the interactions within cells.
DNA transcription is initiated by proteins and is the first step in protein produc-
tion. For a number of eukaryotic proteins, the process continues with co-translation
through ribosomes into the endoplasmic reticulum (ER). Molecular chaperones and
folding machinery aid in folding protein into its native structure. This native state is
not a random one but is instead the result of both the amino acid sequence and the
complex folding network. These properly folded proteins are transported out of the
ER for further processing.
The path from gene to protein is composed of many different and unknown
interactions between DNA, RNA, proteins, and small molecules. Protein folding
is one network, or subsystem, within the larger system of protein production. A sys-
tems biology approach offers us an opportunity to understand the complicated
network of protein folding and the emergent properties that arise from interacting

C.I. Sandefur ()

Center for Computational Medicine and Bioinformatics, University of Michigan,
2017 Palmer Commons, 100 Washtenaw Ave, Ann Arbor, MI 48105, USA
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 1

Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 1,
c Springer Science+Business Media, LLC 2011
2 C.I. Sandefur and S. Schnell

network components. In this chapter, we explore two models of protein folding and
misfolding to investigate how the protein folding network affects protein home-
ostasis. Using these models, we can identify the protein quality control pathways
regulating folding and offer potential therapeutic targets for protein folding diseases.

1.1.1 Disruptions to the Protein Folding Network are Associated

with Disease

Protein folding is often described by way of a folding energy landscape (Fig. 1.1)
(Chiti and Dobson 2006). The landscape is composed of different conformations
of a given protein each corresponding to a different energy level. The minimum
energy, three-dimensional folded protein structure is termed the “native state” and
for most proteins, is essential for proper function (Alberts et al. 2008). Failure to fold
properly results in misfolded protein conformations. These protein conformations
correspond to energy minima pockets within the folding energy landscape.
Proteins may fail to properly fold through mutations, cellular stress, or stochas-
tic events (Nakatsukasa and Brodsky 2008). A breakdown in the quality of protein
production can lead to the accumulation of toxic levels of misfolded and unfolded
proteins. Improperly folded proteins can form aggregates (Morimoto 2008). When
the level of aggregates reaches a certain concentration threshold, these protein com-
plexes may lead to proteotoxicity.

Fig. 1.1 Different protein conformations have different energies. While the goal is to reach the
lowest energy as a properly folded protein (F), some misfolded proteins (M) are located in energy
minima. Unfolded protein is denoted by U
1 Quality Control on Protein Homeostasis 3

A variety of diseases are the linked to protein misfolding. For example, disruption
of proinsulin folding in “-cells is sufficient to induce diabetes in both humans and
mice (Scheuner and Kaufman 2008). Aggregation due to increased protein mis-
folding is implicated in the neurological diseases Alzheimer’s, Parkinson’s, and
Huntington’s (Soto 2003). The mechanisms behind aggregation of misfolded pro-
teins and how the cell copes with misfolded protein accumulation are unknown.

1.1.2 The ER Functions as a Protein Folding Factory

Despite many technological advances, a complete understanding of the process of

protein folding remains elusive. Proteins fold by transitioning through intermediates
that comprise the folding landscape. However, detecting intermediate structures is
difficult. This is because fast folding intermediates are not easily measured using
current technology (Dobson 2004).
The ER is responsible for the synthesis, folding, assembly, and modification of
one third of the eukaryotic proteome (Kaufman 2004). Most proteins cannot refold
into their native states in the absence of cellular machinery. Protein folding in the ER
is analogous to a factory assembly line with machinery processing proteins into a
final, unique, native conformation. Enzymes and molecular chaperones are a part of
this machinery working along the protein assembly line. Once a protein is properly
folded, it is exported from the ER. If unfolded or misfolded proteins accumulate in
the ER factory above a certain threshold, protein homeostasis is disrupted which can
result in proteotoxicity (Ron and Walter 2007).
Cells have evolved a set of quality control processes that restore protein home-
ostasis. The processes are collectively termed the unfolded protein response (UPR).
The UPR aids in quality control of protein production through three general pro-
cesses. One process of the UPR prevents the influx of new peptides into the ER
(Harding et al. 1999). Halting incoming materials into the factory reduces the bur-
den on the cellular machinery.
The second process of the UPR increases the capacity of the ER-assisted-folding
(ERAF) pathway through upregulation of chaperones and folding catalysts. This
additional machinery aids in efficient processing of proteins within the burdened
factory. Along with assisting in protein folding, chaperones and enzymes also se-
quester polypeptides within the ER. This is done to ensure that the mature folded
proteins meet the factory quality control standards before export (Brodsky 2007).
Third, the UPR invokes the ER-assisted-degradation (ERAD) pathway. Due to
the strict quality control measures of the protein factory, most proteins are near
degradation as they move along the assembly line (Liberek et al. 2008). Chaperones
escort proteins targeted for degradation. The chaperones prevent aggregation by al-
lowing proteins to remain soluble and accessible to retrotranslocation machinery
(Nakatsukasa and Brodsky 2008). After a protein is retrotranslocated to the cytosol
by a retrotranslocon channel, it is degraded by the ubiquitin/proteasome pathway
(Meusser et al. 2005). Enhancement of degradation reduces the assembly line load.
4 C.I. Sandefur and S. Schnell

1.1.3 Mathematical Models of Protein Quality Control Provide

Novel Insights into the Regulation of Protein Assembly

Although great strides have been made in understanding the network of protein
folding, we lack a complete picture of the processes necessary for proteins to prop-
erly fold. We can apply modeling to investigate the mechanisms of protein quality
control and make new experimental predictions. In biochemical processes, mathe-
matical models are generally systems of ordinary equations. Using these models, we
can investigate how varying reaction rates impact relative levels of system compo-
nents through time. Also, we can obtain a dynamical view of the impact of protein
quality control on the synthesis of native protein.
We know that protein folding in the ER involves a quality control mechanism,
but how does this impact the dynamics of the native protein concentration? Ex-
perimental observations of protein quality control show it to be dependent on the
amount of protein within the ER lumen (Ron and Walter 2007). We hypothesize
that this dependence increases the timescale of protein accumulation and depletion
under quality control. We test this hypothesis by comparing two models of protein
folding, one without quality control and the other with.

1.2 Case Studies

In the following case studies, we analyze two models of protein folding. The first
case study is an analysis of a simple model describing protein folding in absence
of the UPR. We follow with a second model describing protein folding regulated
by the UPR. A comparison of the two models serves to illustrate how mathematical
models provide a greater understanding of the dynamics of protein quality control.

1.2.1 Case Study I: Protein Folding Without Quality Control

The experimental measurements obtained from protein folding in vitro led to the
development of the two state model of protein folding. In this model, unfolded pro-
tein spontaneously folds into its native state without intermediates (Anfinsen et al.
1954). This model provides a simple description of protein folding in absence of
quality control machinery.

1.2.1.1 Assumptions

This first model contains three protein conformations: unfolded protein (U), folded
protein (F), and misfolded protein (M) (Fig. 1.2). We are not considering influx or
1 Quality Control on Protein Homeostasis 5

Fig. 1.2 Schematic of protein folding without quality control. The three protein conformations
are represented as follows: unfolded (U), folded (F), and misfolded (M). Folding and misfolding
reaction velocities are first-order with rate constants k1 and k0 , respectively. We assume that folding
and misfolding are irreversible reactions. There is no influx or outflux of protein so the total protein
concentration is conserved

outflux of protein; the system is closed and the total protein concentration is constant
(u C m C f D constant). Note that we denote protein concentrations using lower
case variables.
We model spontaneous folding of unfolded protein at a rate of k1 and misfolding
at a rate of k0 (Anfinsen et al. 1954). In general, chaperones are required for unfold-
ing from a misfolded or folded state (Martin and Hartl 1997). Here, we assume that
both folding and misfolding reactions are irreversible.
Equations (1.1)–(1.3) describe protein folding and misfolding in the absence of
quality control by a linear system of ordinary differential equations.

du
D .k1 C k0 / u (1.1)
dt
dm
D k0 u (1.2)
dt
df
D k1 u: (1.3)
dt
Note that the rate equations describing folded and misfolded protein are both de-
pendent on unfolded protein.

1.2.1.2 Analytical Solution

We can solve this linear model analytically. Setting u0 as the total basal protein
concentration (u.0/ D u0 , m(0) D 0, and f(0) D 0), we find the analytical solution of
our system to be:
u.t/ D u0 e.k0 Ck1 / t (1.4)
6 C.I. Sandefur and S. Schnell

Fig. 1.3 Time course of the

three protein conformation
concentrations in absence of
quality control. We begin
with a basal unfolded protein
(u) concentration, u0 , of
1 M. Misfolded protein
concentration, m, reaches a
k0
maximum of k0 Ck 1
u0 and
folded protein concentration,
f , reaches a maximum of
k1
u . In this figure, k0 =
k0 Ck1 0
0.25 s1 and k1 = 1 s1 . The
timescale for the system is
denoted by 

k0  
m.t/ D u0 1  e.k0 Ck1 / t (1.5)
k0 C k1
k1  
f .t/ D u0 1  e.k0 Ck1 / t : (1.6)
k0 C k1

We can plot the concentrations of the different protein conformations as functions

of time (Fig. 1.3). We begin with some basal unfolded protein concentration (u0 )
which decreases monotonically to zero. Misfolded protein levels increase towards
a maximum misfolded concentration, mmax , while folded protein levels increase to-
wards a maximum folded protein concentration, fmax , where,

k0
mmax D u0 and (1.7)
k0 C k1
k1
fmax D u0 : (1.8)
k0 C k1

1.2.1.3 Timescale Analysis

The timescale is the amount of time required for a significant change in the level of
a protein conformation to occur and can be defined as (Segel 1984):
xmax  xmin
timescale of x(t)  ˇ dx ˇ : (1.9)
ˇ ˇ
dt max
1 Quality Control on Protein Homeostasis 7

Since the rates of formation of folded and misfolded protein depend on unfolded
protein, the two terminal protein conformations are formed under the same timescale
as unfolded protein depletion. The timescale of unfolded protein depletion and mis-
folded and folded protein accumulation is

1
D : (1.10)
k0 C k1

In the initial transient of the folding process, the levels of misfolded and folded
protein increase, as the misfolding and folding reactions compete for the unfolded
protein (Fig. 1.3). Eventually, all of the unfolded protein in the system is either con-
verted to folded or misfolded protein at rates k1 or k0 , respectively. If either rate
is increased, unfolded protein is depleted from the system more quickly. If we in-
crease the rate of folding, k1 , the maximum concentration of folded protein in the
system increases. This also results in a decrease in the timescale of folded protein
accumulation. We observe similar behavior in the misfolded protein levels when the
rate of misfolding is increased.

1.2.1.4 Conclusions for Case Study I

We introduced a simple model of protein folding and misfolding in absence of qual-

ity control. There is one timescale in the system that is dependent on the rates of
folding and misfolding alone. Unfolded protein is depleted from the system on the
same timescale as misfolded and folded protein form. In this linear system, the ex-
act amounts of folded and misfolded protein can be determined at any time point by
knowing the rates of misfolding and folding and the basal unfolded protein concen-
tration. This model is a simplification and does not capture the interactions between
the components of the cellular folding network in the ER. These interactions impact
the overall behavior of the system as we will show in the next subsection.

1.2.2 Case Study II: Protein Folding with Quality Control

In Case Study I, we analyzed a model describing protein folding in the absence of

the UPR. In reality, protein homeostasis within the folding factory of the ER is much
more complicated. In this second case study, we analyze a model of protein folding
regulated by the UPR. We compare the two models to investigate the impact of pro-
tein quality control machinery on protein homeostasis. We also perform a sensitivity
analysis to identify parameters driving folding and misfolded protein accumulation.
We discuss potential therapies for recovering folded protein levels under conditions
promoting the accumulation of misfolded protein, such as those observed in protein
misfolding diseases.
8 C.I. Sandefur and S. Schnell

1.2.2.1 Assumptions

We analyzed a recently formulated model of the UPR in pancreatic “-cells (Fig. 1.4)
(Schnell 2009). This model assesses factory function after activation of the three
responses of the UPR.
As in Case Study I, we assume there is no input of unfolded protein into the
system. Halted protein influx results in a reduction of protein entry into the ER
lumen and is one of the three responses of the UPR (Harding et al. 1999). We begin
with a basal unfolded protein concentration denoted u0 . We also assume that the
rate of protein misfolding (k0 ) follows first-order kinetics and is proportional to the
level of unfolded protein (Nolting 2006). Again, we model protein misfolding as
irreversible (Martin and Hartl 1997).
It has been experimentally demonstrated that complex biochemical processes can
be modeled as single enzyme reactions (Aldridge et al. 2006; Kholodenko 2006;
Wiseman et al. 2007). Using this precedent of describing biochemical processes, two
additional UPR processes were introduced into the quality control model of protein
folding. As discussed above, ERAF and ERAD responses of the UPR are com-
plex pathways comprised of many different components including chaperones and
folding or degradation catalysts. Here, the ERAF response is modeled as a single en-
zyme with unfolded protein as a substrate (see Segel (1984) for details on modeling
enzyme kinetics). The maximum velocity of folding is Vf with a Michelis–Menten
(MM) constant of Kf . This MM constant is representative of the disassociation con-
stant of folding machinery from unfolded protein.
Since a buildup of unfolded and misfolded protein in the ER lumen (which
leads to an activation of the UPR) is assumed, the ERAD degradation machinery is
modeled as responsible for removing both protein conformations (Nakatsukasa and
Brodsky 2008). Therefore, a competition occurs between unfolded and misfolded

Fig. 1.4 Schematic of protein folding with quality control. U is unfolded protein, F is folded pro-
tein, and M is misfolded protein. X is the enzyme representative of the folding machinery. Y is the
enzyme representative of the degradation machinery. The enzyme–substrate complex intermediate
for each pathway is represented by IUX , IUY , and IMY . Misfolding occurs through a first-order
reaction with rate constant k0
1 Quality Control on Protein Homeostasis 9

protein as both are degraded with the same machinery. The maximum velocities
of unfolded and misfolded protein degradation are denoted by Vu and Vm , respec-
tively. Ku corresponds to the disassociation constant of degradation machinery from
unfolded protein. Km corresponds to the disassociation constant of degradation ma-
chinery from misfolded protein. Using the model schematic in Fig. 1.4 and the MM
terms for the ERAF and ERAD process, we write the following system of differen-
tial equations describing protein folding under quality control:

du Vf u Vu u
D k0 u     (1.11)
dt Kf C u Ku 1 C m C u
Km

dm Vm m
D k0 u    (1.12)
dt Km 1 C Kuu C m
df Vf u
D : (1.13)
dt Kf C u

1.2.2.2 Qualitative Dynamical Behavior and Equilibrium Points

Most nonlinear dynamical systems, such as the one described by (1.11)–(1.13),

will not have an analytical solution. There are a variety of techniques useful for
ascertaining the behavior of dynamical systems in this situation. In our analysis, we
find the equilibrium points of the system (Sect. 1.2.2.2), we estimate the timescales
(Sect. 1.2.2.3), and follow with a parametric sensitivity analysis to determine how
the kinetic parameters impact the system (Sect. 1.2.2.4).
In order to find the equilibrium points of a system, we look for situations where
all of the rate equations are equal to zero. In this system, the only equilibrium point
is the trivial one: .u? ; m? / D .0; 0/. Over time, all of the basal unfolded protein
will either fold, misfold, or degrade (Fig. 1.5). Misfolded protein undergoes degra-
dation as well, and therefore, both unfolded and misfolded protein concentrations
are reduced to zero.
The minimum and maximum amounts of unfolded protein are the same across
the two models. We expect the level of unfolded protein in both models to mono-
tonically decrease from u0 to zero. The maximum misfolded protein concentration
is different between the two models. In absence of quality control, the maximum
amount of misfolded protein is only related to the rates of folding and misfolding.
Under quality control, the misfolded protein reaches a maximum level due to mis-
folding but is also undergoing some level of degradation. The misfolded protein is
eventually depleted to a zero concentration by degradation machinery.
10 C.I. Sandefur and S. Schnell

Fig. 1.5 The time course of the unfolded protein (U), misfolded protein (M) and folded protein (F)
concentrations under quality control. The timescales for unfolded protein depletion (u ), misfolded
protein depletion (m ), and folded protein production (f ) are denoted by the vertical lines. The time
course of degraded protein is not represented. The parameter values used were k0 D 0:25 s1 ,
Vf D Vm D 1:0 M s1 , Vu D 0:1 M s1 , Kf D 2:1 M, and Ku D Km D 1:1 M

1.2.2.3 Timescale Analysis

As in Case Study I, we determine the timescale for a process by estimating (1) the
maximum and minimum concentrations of a given protein conformation and (2) the
magnitude of the maximum reaction rate describing the evolution of the protein
conformation over time. However, timescale determinations of non-linear systems is
also a bit of an art. It requires making simplifying assumptions using our biological
intuition about the system (Segel 1972; Segel and Slemrod 1989).
We begin by looking at the timescale for unfolded protein depletion. We know
that the minimum amount of unfolded protein is 0 and the maximum is u0 . At the
beginning of the reaction, the level of misfolded protein is small (m(t  0)  0)
while the level of unfolded protein is near the basal unfolded
ˇ ˇ protein concentration
(u(t  0)  u0 ). We use this information to estimate ˇ du ˇ
dt max
from (1.11) as:
ˇ ˇ  
ˇ du ˇ Vf Vu
ˇ ˇ  u0 k0 C C : (1.14)
ˇ dt ˇ Kf C u0 Ku C u0
max

Applying (1.9) gives the timescale for the unfolded protein depletion:
 1
Vf Vu
u D k0 C C : (1.15)
Kf C u0 Ku C u0

Misfolded protein initially accumulates, reaches a maximum level and then

undergoes depletion. We can also determine the timescale for misfolded protein
1 Quality Control on Protein Homeostasis 11

depletion through some simplifications. We focus on the depletion phase as we are

interested in understanding the behavior of misfolded protein under conditions of
quality control. The depletion phase begins when the concentration of misfolded
protein is at its maximum. We overestimate the maximum amount of misfolded
protein as u0 for the depletion phase. The minimum value of misfolded protein
is zero. During the depletion phase, we assume that the contribution of unfolded
protein is negligible because it has been depleted due to folding, misfolding, or
degradation, so we treat u(t)  0. This allows us to approximate the maximum rate
for misfolded protein from (1.12) as:
ˇ ˇ
ˇ dm ˇ Vm u 0
ˇ ˇ  : (1.16)
ˇ dt ˇ
max K m C u0

We apply (1.9) to determine the timescale for the depletion phase of misfolded pro-
tein as shown:
Km C u0
m D : (1.17)
Vm
In order to restore homeostasis, the degradation machinery works to remove mis-
folded protein from the ER lumen. For example, we see from (1.17) that increasing
the maximum velocity of misfolded degradation reduces the timescale and restores
homeostasis more quickly.
We conclude with an estimation of the folded protein accumulation timescale.
Here, we know that the minimum amount of folded protein is zero (the initial folded
protein concentration). We overestimate the maximum folded protein concentration
by allowing all of the basal unfolded protein to fold (fmax D u0 ). By applying (1.9),
we obtain the timescale for folded protein accumulation from (1.13) as:
Kf C u0
f D : (1.18)
Vf
With these timescale estimations, we proceed with a comparison of two models
to explore how protein quality control machinery impacts the dynamics of protein
homeostasis. The timescale for unfolded protein depletion and misfolded protein
accumulation in absence of quality control is dependent on the rate of misfolding.
Further, the timescale in Case Study I is dependent on k1 . Adding protein quality
control machinery essentially replaces k1 with KfVCu
f
0
C KuVCu
u
0
. We can set up the
following inequality to explore the two timescales as shown:
Vf Vu
k1  C : (1.19)
Kf C u0 Ku C u0
When (1.19) holds, the timescale for unfolded protein depletion in Case Study I
is shorter then the timescale of unfolded protein depletion under protein quality
control. When (1.19) does not hold, the reverse is true.
We can reduce the timescale of unfolded or misfolded protein depletion under
conditions of quality control by increasing the Vf or Vu , for example. By increasing
12 C.I. Sandefur and S. Schnell

Vf through increasing the amount of folding machinery, the timescale for unfolded
protein depletion is reduced in this instance. Therefore, an example of a potential
treatment for protein misfolding disease would entail targeting the maximum veloc-
ity of folding in order to restore protein homeostasis more quickly.
In absence of quality control, the folded protein accumulation timescale in (1.10)
depends on the first-order misfolding and folding rates. Therefore, decreasing the
misfolding rate, for example, would decrease the timescale for folded protein accu-
mulation. It would also lead to an increase in the final, folded protein concentration
in (1.8). In contrast, by our estimation in (1.19), the folded protein accumulation
timescale under quality control does not depend on the misfolding rate. We cannot
use the same argument of decreasing misfolding with the hopes of increasing folded
protein levels when the system is under protein quality control.
In contrast to the timescale of Case Study I shown in (1.10), the timescales of the
model with protein quality control are all dependent on the basal levels of unfolded
protein in the system. Therefore, an increased amount of basal unfolded protein
would lengthen the timescales of folded protein accumulation and unfolded and
misfolded protein depletion under quality control. This is important since physio-
logical changes can increase the demand for folded proteins leading to an influx of
unfolded protein in the ER lumen. For example, increased blood glucose levels lead
to an increased demand for insulin production in order to replace glucose-stimulated
secreted insulin. The ER quality control must allow for proteins to fold while also
managing the influx of new proteins. Cells could upregulate ERAD or ERAF ma-
chinery to manage this influx, but how would this impact folded protein levels? We
use parametric sensitivity analysis to lend insight.

1.2.2.4 Parametric Sensitivity Analysis

Parametric sensitivity analysis allows us to quickly ascertain how the parameters

impact the system. We calculated the relative local sensitivities of each protein con-
formation to the system parameters (Varma et al. 1999). A sensitivity score is a way
to quantify the relationship between system behavior and a parameter. When we
calculate the parametric sensitivity, we are assessing how an input parameter im-
pacts the evolution of the concentration of a protein conformation. The magnitude
of the sensitivity score denotes the strength of the sensitivity. The larger the magni-
tude, the more sensitive the evolution of a protein conformation is to a parameter.
If a parameter increases the accumulation of a protein conformation and diminishes
depletion, the sign of the sensitivity score is positive. On the other hand, if a param-
eter increases the depletion of a protein conformation or diminishes accumulation,
the sensitivity score is negative. If a protein conformation is not sensitive to a given
parameter, the sensitivity score will be zero.
We begin by examining the sensitivity of the unfolded protein concentration for
a given set of input parameters (Fig. 1.6a). All of the sensitivity scores are negative
for unfolded protein. The unfolded protein concentration is most sensitive to k0 ,
followed by Vf , and third, Vu . All three rates reduce the amount of unfolded protein
in the system. The k0 depletes unfolded protein through conversion to misfolded
1 Quality Control on Protein Homeostasis 13

Fig. 1.6 Relative sensitivities

of each protein conformation
for the system parameters
across time. (a) unfolded
protein concentration
sensitivities, S(u;*),
(b) misfolded protein
concentration sensitivities,
S(m;*), and (c) folded protein
concentration sensitivities,
S(f;*). Parameters giving a
relative sensitivity near to
zero for a given protein
conformation were not
included in the graphs.
Sensitivity scores are based
on the following parameter
values: k0 D 0.25 s1 ,
Vf D Vu D Vm D 0.1
M s1 , and Kf D Ku D
Km D 1.1 M

protein. The Vf depletes unfolded protein through conversion to folded protein by

the folding machinery. The Vu depletes unfolded protein levels through degrada-
tion. The concentration of unfolded protein is increasingly sensitive to all three
parameters in the initial transient of the reaction. The sensitivities decrease in magni-
tude as unfolded protein is converted to other protein conformations or is degraded.
Misfolded protein accumulates to a maximum concentration due to conversion
of unfolded protein to misfolded protein by first-order misfolding (accumulation
phase) and then decreases due to degradation through ERAD (depletion phase).
The Vm is the parameter to which misfolded protein accumulation is most sensi-
tive (Fig. 1.6b). This parameter drives misfolded protein degradation. The sensitivity
score for Vm increases in magnitude as the misfolded protein concentration increases
during the accumulation phase. The k0 increases misfolded accumulation in the first
14 C.I. Sandefur and S. Schnell

several time units, but the sensitivity score reduces (as the amount of unfolded pro-
tein available to misfold is depleted). The maximum velocities of Vu and Vf both
reduce misfolded protein accumulation as noted by the negative sensitivity value.
Folding or degradation of unfolded protein both reduce the amount of unfolded pro-
tein available to misfold.
In both case studies, folded protein reaches a maximum concentration, but in the
second model, folded protein accumulation is under quality control. The concentra-
tion of folded protein is most sensitive to Vf . Increasing Vf enhances folded protein
accumulation (Fig. 1.6c). The k0 diminishes the accumulation of folded protein by
reducing the amount of available unfolded protein through misfolding, as denoted
by the negative sensitivity score.
The Vu reduces the amount of unfolded protein available to fold through degra-
dation. Returning to the question of whether upregulating ERAD would impact
folded protein accumulation, we find that it does. If cells increased ERAD machin-
ery, thereby increasing degradation of both misfolded and unfolded protein, reduced
folded protein accumulation would result. Therefore, upregulation of general ERAD
machinery as therapy could ease a burdened ER lumen, but also have the uninten-
tional impact of reducing folded protein assembled in the ER factory.
From the sensitivity analysis, we observe that with quality control, recovery of
folded protein levels may occur in two different ways. One manner of recovery is
through reducing the impact of parameters diminishing folded protein accumulation
(k0 and Vu ). If we decrease k0 or Vu , we increase unfolded protein levels during
the initial transient of the process. A lower k0 also reduces the accumulation of
misfolded protein. We also find that the Vu reduces the accumulation of misfolded
protein.
We can also increase accumulation of folded protein in the UPR model by target-
ing the Vf . Not only does this increase folded protein accumulation, it also dimin-
ishes misfolded protein accumulation. This is in agreement with conclusions made
by Schnell (2009). Our parametric sensitivity analysis also agrees with the Schnell’s
assessments of investigating the ERAD pathway for therapeutic targets in misfolded
diseases. Our analysis highlights the importance of investigating the ERAF pathway
as well. Targeted therapy towards the ERAF pathway may allow for both increased
folded and decreased misfolded protein levels in protein misfolding diseases.

1.2.2.5 Conclusions for Case Study II

The concentration of cellular protein can reach 350 mg/mL, and without quality
control processes, toxic levels of aggregates may result (Dobson 2004). The UPR
works to reduce the build up of unfolded and misfolded proteins within the ER
through attenuating protein synthesis and enhancing protein folding (ERAF) and
degradation (ERAD) of non-native proteins. We used a recently formulated model
of the UPR to explore processes of quality control.
The timescales of the model with quality control were dependent on the con-
centration of basal unfolded protein in the system. We found that the timescale for
1 Quality Control on Protein Homeostasis 15

unfolded protein depletion was dependent on the parameters driving protein folding
as well as unfolded protein degradation. The timescales for misfolded protein
depletion and folded protein accumulation were approximated based on overesti-
mations of protein concentrations.
We also discussed therapeutic strategies based on parametric sensitivity analysis.
There are several ways to increase folded protein accumulation under protein quality
control. We can increase folded protein accumulation through decreasing parameter
values that diminish folded protein accumulation or through increasing parameter
values that enhance folded protein accumulation. Each strategy impacts the unfolded
and misfolded protein conformations differently.
This model is a simplification of the UPR in the ER lumen. An advantage of
this simplification is that this model of protein quality control is relatively tractable.
An important aspect of mathematical modeling is validating predicted system be-
havior. Predictions made through mathematical modeling only help our progress
in science when they are validated by experimental results. For example, one pre-
diction is that increasing Vu would result in reduced levels of misfolded protein as
well as diminished folded protein accumulation. Recall that Vu represents the max-
imum velocity at which unfolded protein degradation may occur. One manner of
increasing this velocity is to increase the amount of degradation machinery in the
ER. Putative proteins involved in ERAD to target for overexpression are OS-9 and
XTP3-B in mammals and YOS9 in yeast (Nakatsukasa and Brodsky 2008). The re-
sults of this overexpression experiment would then be used to further refine the UPR
model. Collaboration with experimentalists is a vital piece in developing mathemat-
ical models that both describe realistic behavior and that can be utilized to make
realistic predictions for potential therapies.

1.3 Lessons Learned

1. Mathematical models are analogous to experimental tools used to test hypothe-

ses. Simple models of protein homeostasis can be represented with first-order
reaction kinetics. The treatment of complex biochemical pathways as single en-
zyme reactions can be used to create a more complete picture of the dynamics of
protein homeostasis in the ER. We can use these models to explore how protein
quality control machinery regulates the dynamics of protein homeostasis.
2. In the absence of protein quality control machinery, protein folding and misfold-
ing rates can be modeled as first-order rate constants. Under these circumstances,
increasing the folding rate results in both increased folded protein levels and de-
creased misfolded protein levels. Decreasing the rate of misfolding in the first
model produces similar results.
3. First-order models (Case Study I) can be solved analytically. In Case Study II,
we estimated timescales for a non-linear system. Scaling of non-linear systems
is a bit of an art. Analysis of these models requires simplifications made using
the biological intuition of the modeler (Segel 1972; Segel and Slemrod 1989).
16 C.I. Sandefur and S. Schnell

4. Incorporating quality control machinery into a model of protein homeostasis im-

pacts the overall behavior of the system. In the absence of quality control, one
timescale describes the model and is based on the rates of misfolding and folding
alone. By contrast, the timescales of the quality control model differed across the
protein conformations. Also, the timescales were dependent on basal unfolded
protein concentration as well as kinetic parameters describing protein misfold-
ing, folding, and degradation.
5. A parametric sensitivity analysis identified how each protein conformation was
impacted by the parameters in the model with quality control. According to our
analysis, two different pathways can be targeted for therapeutic purposes. One
potential therapy to improve folded protein levels involves reducing ERAD ma-
chinery to allow more unfolded protein to fold. Applying this therapy, could lead
to proteotoxicity by allowing an accumulation of unfolded and misfolded pro-
tein within the ER lumen. The other potential therapy would be to target the
ERAF pathway using pharmacological chaperones. Using this latter method, we
decrease the accumulation of misfolded protein as well.

Acknowledgements The authors would like to acknowledge the comments from Marnie Briceno
(University of Washington), Hannah Briolat (University of Michigan), and Michelle Wynn (Uni-
versity of Michigan). This work is based upon research supported by the National Science
Foundation under Grant No. IIS-0852734.

Appendix: Symbols Used in this Chapter

u(t) Unfolded protein concentration at time t

m(t) Misfolded protein concentration at time t
f(t) Folded protein concentration at time t
X Folding machinery
Y Degradation machinery
IUX Unfolded protein – folding machinery complex
IUY Unfolded protein – degradation machinery complex
IMY Misfolded protein – degradation machinery complex
k0 First-order misfolding rate
k1 First-order folding rate
Vf Maximum velocity of folding
Kf MM constant of folding machinery-unfolded protein reaction
Vu Maximum velocity of unfolded protein degradation
Ku MM constant of degradation machinery-unfolded protein reaction
Vm Maximum velocity of misfolded protein degradation
Km MM constant of degradation machinery-misfolded protein reaction
MM Michaelis–Menten
1 Quality Control on Protein Homeostasis 17

References

B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts, and P. Walter. Molecular biology of the cell,
5th edition. Garland Science, San Francisco, 2008
B. B. Aldridge, J. M. Burke, D. A. Lauffenburger, and P. K. Sorger. Physicochemical modeling of
cell signalling pathways. Nature Cell Biology, 8:1195–1203, 2006
C. B. Anfinsen, R. R. Redfield, W. L. Choate, J. Page, and W. R. Carroll. Studies on the gross struc-
ture, cross-linkages, and terminal sequences in ribonuclease. Journal of Biological Chemistry,
207:201–210, 1954
J. L. Brodsky. The protective and destructive roles played by molecular chaperones during ERAD
(endoplasmic-reticulum-associated degradation). Biochemical Journal, 404:353–363, 2007
F. Chiti and C. M. Dobson. Protein misfolding, functional amyloid, and human disease. Annual
Review of Biochemistry, 75:333–366, 2006
F. Crick. Central dogma of molecular biology. Nature, 227:561–563, 1970
C. M. Dobson. Experimental investigation of protein folding and misfolding. Methods, 34:4–14,
2004
H. P. Harding, Y. Zhang, and D. Ron. Protein translation and folding are coupled by an
endoplasmic-reticulum-resident kinase. Nature, 397:271–274, 1999
R. J. Kaufman. Regulation of mRNA translation by protein folding in the endoplasmic reticulum.
Trends in Biochemical Sciences, 29:152–158, 2004
B. N. Kholodenko. Cell-signalling dynamics in time and space. Nature Reviews Molecular Cell
Biology, 7:165–176, 2006
K. Liberek, A. Lewandowska, and S. Zietkiewicz. Chaperones in control of protein disaggregation.
The EMBO Journal, 27:328–335, 2008
J. Martin and F. U. Hartl. Chaperone-assisted protein folding. Current Opinion in Structural
Biology, 7:41–52, 1997
B. Meusser, C. Hirsch, E. Jarosch, and T. Sommer. ERAD: the long road to destruction. Nature
Cell Biology, 7:766–772, 2005
R. I. Morimoto. Proteotoxic stress and inducible chaperone networks in neurodegenerative disease
and aging. Genes and Development, 22:1427–1438, 2008
K. Nakatsukasa and J. L. Brodsky. The recognition and retrotranslocation of misfolded proteins
from the endoplasmic reticulum. Traffic, 9:861–870, 2008
B. Nolting. Protein folding kinetics – Biophysical methods (2nd edition). Springer, Berlin, 2006
D. Ron and P. Walter. Signal integration in the endoplasmic reticulum unfolded protein response.
Nature Reviews Molecular Cell Biology, 8:519–529, 2007
D. Scheuner and R. J. Kaufman. The unfolded protein response: a pathway that links insulin de-
mand with “-cell failure and diabetes. Endocrine Reviews, 29:317–333, 2008
S. Schnell. A model of the unfolded protein response: pancreatic “-cell as a case study. Cellular
Physiology and Biochemistry, 23:233–244, 2009
L. A. Segel. Modeling dynamic phenomena in molecular and cellular biology. Cambridge Univer-
sity Press, Cambridge, 1984
L. A. Segel. Simplification and scaling. SIAM Review, 14:547–571, 1972
L. A. Segel and M. Slemrod. The quasi-steady-state assumption: a case study in perturbation. SIAM
Review, 31:446–477, 1989
C. Soto. Unfolding the role of protein misfolding in neurodegenerative diseases. Nature Reviews
Neuroscience, 4:49–60, 2003
A. Varma, M. Morbidelli, and H. Wu. Parametric sensitivity in chemical systems. Cambridge Uni-
versity Press, Cambridge, 1999
R. L. Wiseman, E. T. Powers, J. N. Buxbaum, J. W. Kelly, and W. E. Balch. An adaptable standard
for protein export from the endoplasmic reticulum. Cell, 131:809–821, 2007
Chapter 2
Metabolic Network Dynamics: Properties
and Principles

Neema Jamshidi and Bernhard Ø. Palsson

2.1 Introduction

Dynamic descriptions of biological processes, especially metabolism, have been of

interest for many years (Segel 1975). The size and complexity of these models,
however, have stagnated for the last 20 years or so, in spite of dramatic improve-
ments in computational capabilities. The development of large-scale kinetic models
(hundreds to thousands of dynamic variables) has been deemed infeasible now for
a number of years. Traditional approaches for parameterization of kinetic models
require time and labor intensive biochemical assays on individual enzymes. This
presents a challenge and a practical limitation to the number of enzymes that can
be described using kinetic rate expressions and hence limits the size of networks
that can be described dynamically. Furthermore, the confines of a microtiter plate or
test tube are often significantly different than those of the intracellular environment.
Hence, even once these measurements are carried out, they may not be relevant since
the conditions were so different than the in vivo environment. Thus the development
of genome-scale kinetic models with this approach has been recognized as infeasi-
ble. However, biology is a technology-driven science and new technologies have
driven the understanding of biology through the ability to make deeper and broader
measurements (e.g., fluxomics and metabolomics). Thus, these new data should
analogously motivate and drive the development of new computational approaches.
Future development of network dynamics in biology, particularly with metabo-
lism, will involve two branches, the construction of dynamic networks and the
subsequent analysis and simulations of the resulting networks. This chapter will
focus on the aspects of the latter; however, the first half will concern basic proper-
ties and features of dynamic networks, which will be relevant for both construction
and analysis of networks. There are a number of reasons for interest in kinetic mod-
els: (1) the ability to make predictions about fluxes as well as concentrations, (2)

N. Jamshidi ()
Department of Bioengineering, University of California, San Diego, 9500 Gilman Drive,
La Jolla, CA 92093-0142, USA
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 19

Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 2,
c Springer Science+Business Media, LLC 2011
20 N. Jamshidi and B.Ø. Palsson

a more direct tie-in with experimental measurements, and (3) the ability to make
a more direct connection with environmental as well as genetic perturbations by
modifying the initial conditions and catalytic and binding constants. The analysis
of kinetic models is a necessarily mathematical and computational topic, and there
is often a tendency to lose the forest for the trees. The first few sections of this
chapter will aim to place the dynamics in a broader context, so that it can be seen
how it relates to steady-state flux-based models. While the principles and equations
described herein will be applicable to most biological networks, our focus will be
metabolism. Furthermore, we will focus on the dynamic hierarchy of metabolic net-
works with the aim of try to dissect and understanding the interactions that occur
between components on different time scales.

2.2 Dynamic Mass Balances and Fundamental Subspaces

Representing biological interactions in terms of mathematical expressions enables

one to be precise and unambiguous about what is being discussed. More importantly,
however, this enables the ability to benefit from underlying mathematical properties
reflected in the equations and the ability to apply physical constraints. For example,
through mathematical representation of a metabolic network, one can enforce mass
and energy conservation and then explore the implications of these constraints.
The dynamic mass balance equations that describe the dynamic states of bio-
chemical reaction networks are (Heinrich et al. 1977; Reich and Selkov 1981):

dx.t/
D S  v.x; k/ (2.1)
dt
in which x is an m-dimensional vector of concentrations of the metabolites in the
network in Rm , v is an n-dimensional vector of reaction fluxes in Rn , k represents
a set of rate parameters, and S is the m  n stoichiometric matrix, containing the
stoichiometric coefficients of reactants and products for each reaction in the net-
work (Palsson 2006). The stoichiometric matrix, S, is a mathematical representation
of a metabolic pathway or network. Each column in the matrix corresponds to an
enzymatic (or nonenzymatic) biochemical conversion which may be reversible or
irreversible. The reaction flux vector contains rate expressions for all of the bio-
chemical conversions described by S.
If we consider a nonzero vector, there can be two types of results when the vec-
tor is multiplied by a matrix: it can result in a null vector (vector with 0 in all of
the entries) or it can be nonzero. Vectors that yield null vectors when multiplied
by a matrix lie in the null space. Pre-multiplication (multiplication from the left-
hand side) of any column vector is actually a mapping from the column to the row
space. Figure 2.1 pictorially illustrates what is described by (2.1). The four sub-
spaces in Fig. 2.1 can be viewed as a 2  2 table with each quadrant defined by
fluxes or concentrations across the top and dynamics or conservation quantities on
the side.
2 Metabolic Network Dynamics: Properties and Principles 21

Fig. 2.1 The fundamental linear sub-spaces and their metabolic network interpretations. The sto-
ichiometric matrix, S, maps v which resides in the row and null spaces to dx/dt , which resides in
the column and left null spaces. The null spaces describe conserved quantities (Palsson 2006); in
the right null space this corresponds to conservation of flux and in the left null space this refers
to conserved moieties within the network. The row and column spaces describe dynamic states.
The row space/right null space and column space/left null space are orthogonal complement pairs,
respectively (Strang 1988)

A complete study of the system properties of (2.1) would result in the characteri-
zation of all four subspaces of S (Strang 1988). The right null and left null spaces of
S have been studied extensively over the past decades (Palsson 2006; Heinrich and
Schuster 1996; Famili and Palsson 2003). The left lower box in Fig. 2.1, for exam-
ple, is the set of flux balances reflecting mass conservation (total mass accumulated
= total mass entering the system – total mass exiting the system). The bounds of
the right null space confine the complete set of allowable steady-state flux distri-
butions by enforcing the principle of mass conservation. This subspace has proven
to be extremely insightful from a biological standpoint and has been studied ex-
tensively during the past decades (Palsson 2006; Heinrich and Schuster 1996). The
left null space contains the time-invariant pools, which reflect conserved moieties
or functional groups of metabolites in a particular system. Although there has been
relatively less investigation into the properties of this subspace in metabolic net-
works, its significance and meaning has been well described (Palsson 2006; Famili
and Palsson 2003).
The row and column spaces are the orthogonal complements to the null spaces,
and while dynamic simulations of metabolism have been carried out since the very
earliest days of the field of biochemistry (Segel 1975), dynamics are rarely dis-
cussed in terms of the subspaces and in relation to their orthogonal complements.
Truly appreciating the general principles and underlying factors of dynamics re-
quires recognition of their role and relationship to different subspaces. For example,
22 N. Jamshidi and B.Ø. Palsson

thinking about dynamics in terms of the row and column spaces (see Fig. 2.1) leads
one to immediately recognize that the null spaces describe “conserved” quantities
(i.e., mass) and the row and column spaces of S describe the driving forces and the
direction of motion for the network variables, respectively.
The topics and challenges of this chapter will focus on the top row of Fig. 2.1;
however, occasional mention to the bottom row will be made, because these sub-
spaces are not independent of one another. For example, if the total NAD moiety in
a network is assumed to be constant (which would be identifiable through analysis
of the left null space), there would be obvious implications in the analysis of the col-
umn space for NAD and NADH (knowing the dynamics of one would immediately
inform the dynamics of the other).

2.2.1 Key Considerations in Networks

In the spirit of appreciating the components needed to construct dynamic net-

works, it is important to be cognizant of the nature of molecular interactions as
well as some basic assumptions that are regularly made when building kinetic
models.
The Michaelis–Menten rate equation is perhaps the most famous and commonly
used rate expression for describing reaction kinetics. The rate expressions have
proven to be extremely useful when the underlying assumptions have not been vi-
olated; however, in vivo conditions such as completely saturated enzymes are not
always met. Hence, these rate expressions cannot be claimed to be valid in gen-
eral. More to the point, the interactions that occur in biological networks, including
macromolecular interactions and enzymatic catalysis, are all fundamentally bilinear
interactions (Fig. 2.2). That is, most reactions involve two molecules combining to
form a third. The general rate law for any of these steps is given by, v D kx1 x2 ,
in which k is a bilinear rate constant, and x1 and x2 are the concentrations of the

Fig. 2.2 Molecular interactions are almost always combinations of bilinear association or
dissociations
2 Metabolic Network Dynamics: Properties and Principles 23

interacting components. Fortunately, due to increased processing speed and memory

in computers, it is possible to begin describing large networks with complex regu-
latory schemes in terms of their bilinear interactions.
There is a long history of investigations into the dynamics and kinetics of
metabolism. During this course, various mathematically driven operations and pro-
cedures have been developed. However, since biology is a technology-driven field,
theories and formalisms are only as useful as their ability to integrate available data
and to make testable predictions. The approach and formalism here, focusing on the
S and G matrices, are predicated on capturing the key biological features of sys-
tems while also enabling the integration of available data types. Along these lines,
the gradient matrix has not been defined in any way; it arises naturally from the lin-
earization of the flux vector comprised of net elementary reaction rates. The gradient
matrix describes the responses of the reactions to be changed in the concentrations.
As will be described later, it is through G that the dual nature of the relationship
between fluxes and concentrations can be developed.
The metabolic dynamics described by (2.1) also assumes that concentrations can
be meaningfully defined (i.e., the number of compounds within the specified vol-
ume) and in the absence of any spatial gradients. There are clearly examples when
these assumptions fail to be satisfied, for example, in transcription of genes when
stochastic effects take place or with excitation–contraction coupling between muscle
contraction and energy metabolism, when temporo-spatial gradients have signifi-
cant effects. We will not address these issues here, but only caution that the modeler
should take heed of the physicochemical environment of the phenomena that being
modeled and to be cognizant of when particular assumptions may or may not be
appropriate.

2.2.2 Properties of Dynamic Systems

Unfortunately, network dynamics are often discussed and viewed with a sense of
“magic”, and an implication that somehow nonlinearity can make something appear
out of nothing. However, if one understands the parts of a model and how they fit
together, the results and predictions will be much more palatable and lead to an
improved understanding of a network model and its behavior rather than increased
confusion. We discuss some key properties of dynamic systems and how they con-
tribute to the properties of networks. There are three matrices that will be of interest
in this chapter: the stoichiometric, gradient, and Jacobian matrices. The stoichiomet-
ric matrix is a mathematical representation of the “links and nodes” of a network.
The columns correspond to the links (or reactions) and the rows correspond to nodes
(or chemical species/metabolites). The gradient matrix represents the dependence of
the links on the nodes (in the linear regime). The Jacobian matrix is used to describe
the overall dynamic relationships in the network; as will be seen however, this ma-
trix can be composed from the stoichiometric and gradient matrices. The key point
here is that the fundamental matrices of interest are the stoichiometric and gradient
matrices, and these are in fact biological data matrices.
24 N. Jamshidi and B.Ø. Palsson

2.2.2.1 Underlying Structure of the Jacobian

Linearization of (2.1) as described in Sect. 2.6.1 results in the ability to define the
Jacobian matrix as a product of the stoichiometric and gradient matrix,

JD SG (2.2)
The gradient matrix can then be factored, such that

JDSKM (2.3)

in which K is an n  n diagonal matrix whose entries are the lengths of the rows of
G with units 1/time.
Hence, these entries are pseudo time constants or characteristic times corre-
sponding to the reactions in the network. Consequently, the rows in G indicate the
direction that each reaction lies. So the K and M matrices describe kinetic and ther-
modynamic driving factors in the network.1 This decomposition into the kinetic
and thermodynamic influences was carried out without any involved mathemati-
cal procedures and has been determined by matrices with biologically meaningful
interpretations.

2.2.2.2 Structural Similarity

Reaction rates are commonly expressed as the net sum of elementary reactions.
When this is done, it follows that S and GT are structurally similar (Jamshidi and
Palsson 2008a) with the corresponding row and column entries have zero or nonzero
values. This similarity underlies the stoichiometric influence in network dynamics.
In spite of these similarities, there are also key differences between these matrices,
which will be touched up on in Sect. 2.4.

2.2.2.3 Flux-Concentration Duality

The first property leads to the ability to define a pair of dual Jacobian matrices. One
for the concentrations,

Jx D S  G (2.4)
and one for the fluxes,
Jv D G  S (2.5)

1
Note that M does not refer to the modal matrix in this chapter.
2 Metabolic Network Dynamics: Properties and Principles 25

The systems described by each of these equations is the same; however, the
independent variables is different, in one the variables are the concentrations and
in the other the variables are the fluxes. Note that to convert from the concentration
Jacobian to the flux Jacobian, not advanced mathematics or decompositions were
used, simply reversing the order of multiplication of two matrices.

2.2.2.4 Hierarchical Dynamics

A key feature of biological networks is the presence of many interactions that occur
on a wide range of different time scales. Analysis of these properties has been ac-
tive for many decades, and there is a rich history in time scale separation and modal
decomposition of metabolic networks (Heinrich et al. 1977; Palsson and Lightfoot
1984; Okino and Mavrovouniotis 1998). This will not be detailed here, suffice it to
say that one approach that has been successfully carried out for biochemical net-
work analysis has been modal decomposition, which involves the diagonalization
of the Jacobian matrix, and the redefinition of concentration variables into “modal”
variables which move on dynamically independent time scales.
One overall goal of dynamic analyses of networks is to simplify network struc-
ture and to determine which interactions are relevant at particular time scales of
interest. This enables one to filter out interactions that are either too fast or too
slow to be of interest and to also characterize the progressive pooling of metabolites
across slower and slower time scales.

2.3 Dual Jacobian Matrices

One consequence of recognizing the gradient matrix is that it leads to the definition
of dual Jacobian matrices and highlights the nature of the relationship between con-
centration and flux dynamics. The duality between fluxes and concentrations results
from the ability to define flux and concentration Jacobian matrices, as mentioned
above. The nature of this relationship while mathematically interesting is also of im-
portance from the biological perspective, and hence we will spend some additional
time discussing the nature of the relationship. Measurements and perturbations are
carried out in terms of concentration variables; however, analysis of the fluxes is
what enables interrogation of the systemic properties of networks.
Each network only has a single stoichiometric matrix, S and gradient matrix, G.
However, biological networks can be analyzed in terms of compound (node) vari-
ables or in terms of flux (link) variables. Thus, there are two Jacobian matrices
describing the same network, Jx D S  G and Jv D G  S, depending on which vari-
ables, concentrations or fluxes, are used as state variables. The former gives a
reaction-centric view of the dynamics, while the latter gives a compound-centric
view. These are complementary views of the same system. The relevance of network
topology in dynamic systems is highlighted by the fact that the Jacobian matrices
26 N. Jamshidi and B.Ø. Palsson

are weighted adjacency matrices containing weighted inner products of the reaction
rows and columns (Jv ) and compound rows and columns (Jx ). Thus, Jv and Jx are
structurally similar to the reaction adjacency matrix and the metabolite adjacency
matrix, respectively (Palsson 2006).
Modal decomposition of the Jacobian has been previously applied for the anal-
ysis of biological networks. We note that the two Jacobian matrices share the
same eigenvalues. The eigenvectors/rows Jx relate to pool formation on vari-
ous time scales (Heinrich et al. 1977; Palsson and Lightfoot 1984), while the
eigenvectors/rows of Jv relate the formation of groups of fluxes that move these
pools (Jamshidi and Palsson 2008a). The key point to appreciate is that both views
describe the same set of network interactions, but in terms of different dynamic
variables; dynamic concentration variables in one case and dynamic flux variables
in the other.

2.4 Stoichiometry Versus Gradients

Having stepped through the construction and deconstruction of biological networks,

it is hopefully sufficiently impressed upon the reader that network dynamics can be
comprehensively characterized through the definition of two matrices: the stoichio-
metric matrix, S, and the gradient matrix, G. This is a bold statement, and thus it
will be followed by a bold caveat. The gradient matrix is rarely known in general
for any condition; hence, experimental and measurement limitations require that it
be characterized under a limited set of condition(s). Thus, one generally will only
approximate the elements in the matrix and often be restricted to a linearized region
close to a particular steady state.
It is important to recognize that S and G are data matrices, and they are not just
of theoretical relevance but have very practical significance and import for the con-
struction and subsequent analysis of kinetic networks (Jamshidi and Palsson 2008a).
As mentioned above, when a network is described in terms of bilinear net elemen-
tary reactions, which are in general the most appropriate expressions, S and GT have
similar structures. In spite of these structural similarities, there are many important
differences between these matrices, which we mention briefly here.
The stoichiometric matrix describes the chemical transformations and intercon-
versions that occur among compounds in a network, and it is through the S matrix
that mass conservation can be enforced. The gradient matrix, on the other hand,
accounts for the kinetic interactions that occur within a network and is constrained
by thermodynamic bounds. With these differing physical constraints, there is subse-
quently different data types that are used to populate these matrices. Genomic and
bibliomic data are needed to construct S matrices. Alternatively, metabolomic, flux-
omic, thermodynamic (e.g., equilibrium constants), and if possible kinetic data are
needed to define G.
The stoichiometric matrix contains integer entries; hence, it is a “knowable”
matrix with the potential of no error associated with its elements. In contrast, the
2 Metabolic Network Dynamics: Properties and Principles 27

elements of the gradient matrix are non-integer values and are subject to often
significant experimental errors; hence, these entries may often only be known to
an order of magnitude. The values within the gradient matrix may differ by more
than 10 orders of magnitude; hence, G is ill-conditioned and this underlies the stiff-
ness of biological models, which may lead to difficulties when integrating the set of
differential equations. However, it is also this wide range of values that leads to the
characteristic time-scale separation in biological networks.
Thus, given these similarities and differences between S and G, the resulting
biological interpretation of the matrices also differs. Each S matrix is effectively
a genomic representation of a species. Thus, different species will have different
stoichiometric matrices, and changes in the S results from “distal causation” (Mayr
1961). Conversely, the gradient matrix reflects the genetic features of individuals,
and so although a species has a single S matrix, a population will have a large set
of differing G matrices. Thus, the gradient matrix represents individual differences
within populations and results from changes in “proximal causation” (Mayr 1961).

2.5 Example: Folate Metabolism

The remainder of this chapter will investigate some of the dynamic properties in
a dynamic folate metabolic network. As mentioned above, there are many differ-
ent avenues of analysis to focus on in dynamic networks, and the focus here will
be on the metabolite pooling structure within the network on progressively slower
time scales and the effects of environmental perturbations on this pooling. Reed
et al. (2006) carried out multiple studies with folate one-carbon metabolism in
humans with interesting observations with respect to nutrition and genetic variation.
Since this is a validated yet small and relatively simple kinetic model, we focus on
this network to discuss and highlight some of the discussion points earlier in this
chapter.

2.5.1 Constituent Matrices and Subspaces

This network is described by 10 dynamic concentration variables and 20 reaction

fluxes. As reflected by the network map (Fig. 2.3), there are often multiple intercon-
versions between the same metabolites. The stoichiometric matrix and the gradient
matrix appear in Tables 2.1 and 2.2. Note that all of the entries in S are integers,
whereas almost all of the entries in G are real numbers. Since the model was not
constructed strictly from mass action kinetics and various assumptions were made
(e.g., Michaelis–Menten kinetics), the transpose of G and S is not similar in this
case (however, if the relationships were explicitly described using mass action ki-
netics, the similarity between the two matrices would be preserved). Also note that
the methionine input flux (metin) is zero for all of the entries in the gradient matrix.
28 N. Jamshidi and B.Ø. Palsson

Fig. 2.3 A map of the folate one-carbon metabolism network. A map of the folate one-
carbon metabolism network for the model described by Reed et al. (2006). Only the dynamic
metabolite variables have been labeled. Abbreviations: 5MTHF 5-methyltetrahydrofolate,
THF tetrahydrofolate, DHF dihydrofolate, CH2F 5,10-methylenetrahydrofolate, CHF
5,10-methenyltetrahydrofolate, 10FTHF 10-formyltetrahydrofolate, MET methionine, SAM
S-adenosylmethionine, SAH S-adenosylhomocysteine, HCY homocysteine

Table 2.1 The stoichiometric matrix for the folate and methionine cycles metabolism. Only part
of the values of the matrix is shown
1 2 3 4 5 6 7 8 9 10 . . . 20
Vbhmt Vcbs Vdnmt Vgnmt Vmati Vmatiii Vmthfr Vne Vaicart Vdhfr . . . Metin
m5mthf 0 0 0 0 0 0 1 0 0 0 ... 0
thf 0 0 0 0 0 0 0 1 1 1 ... 0
dhf 0 0 0 0 0 0 0 0 0 1 ... 0
ch2f 0 0 0 0 0 0 1 1 0 0 ... 0
chf 0 0 0 0 0 0 0 0 0 0 ... 0
m10fthf 0 0 0 0 0 0 1 0 1 0 ... 0
met 1 0 0 0 1 1 0 0 0 0 ... 1
sam 0 0 1 1 1 1 0 0 0 0 ... 0
sah 0 0 1 1 0 0 0 0 0 0 ... 0
hcy 1 1 0 0 0 0 0 0 0 0 ... 0

This is because this flux was assumed to be constant in the network. As will be seen
later, however (see Sect. 2.5.3), even though metin is a constant, varying this value
can result in changes throughout the gradient matrix.
The rank, as well as the size of the row and column spaces, of the stoichiometric
matrix is 9, and the rank of the gradient matrix is 10. Given the dimensions and rank
of each of these matrices, the size of the null spaces can be calculated (Table 2.3).
The single dimension in the left null space of S reflects conservation of folate within
the network. Since folate is never directly synthesized or degraded in this network,
it appears in the left null space.
As discussed in Sect. 2.4, the entries in the stoichiometric matrix are integers,
whereas those in the gradient matrix are real valued. The condition number for the
2 Metabolic Network Dynamics: Properties and Principles 29

Table 2.2 The gradient matrix for the network with a methionine input flux of 200 M/h. Only
part of the values of the matrix is shown. Note that since the methionine input flux is a constant,
all of its entries in the gradient matrix are 0
m5mthf thf dhf ch2f chf m10fthf met sam sah hcy
1 Vbhmt 0 0 0 0 0 0 0 0.090 0.090 19.31
2 Vcbs 0 0 0 0 0 0 0 0.077 0.077 104.2
3 Vdnmt 0 0 0 0 0 0 0 0.228 1.257 0
4 Vgnmt 67.46 0 0 0 0 0 0 0.294 3.772 0
5 Vmati 0 0 0 0 0 0 0.746 0.230 0 0
6 Vmatiii 0 0 0 0 0 0 1.805 0.885 0 0
7 Vmthfr 0 0 35.49 0 0 0 0.231 0.231 0 0
8 Vne 0 150.0 0 23.20 0 0 0 0 0 0
9 Vaicart 0 0 0 0 0 47.80 0 0 0 0
10 Vdhfr 0 0 8847 0 0 0 0 0 0 0
           
20 Metin 0 0 0 0 0 0 0 0 0 0

Table 2.3 The dimensions of the stoichiometric and gradient matrices

and the sizes of their right and left null subspaces
Left Right
Rows Columns Null Space Null Space
S 10 20 1 11
G 20 10 10 0

system described by theses matrices is approximately 7:8  104 . This is a relatively

large number and reflects the fact that there is a wide range of concentrations for
different metabolites in the network and that some of the biochemical interactions
in the network occur much quicker (or slower) with respect to other reactions within
the network.

2.5.2 Hierarchical Pooling of Metabolites

As discussed in Sect. 2.2.2, a characteristic feature of metabolism, particularly in

higher order organisms, is aggregate pool formation of metabolites when one moves
from very fast to very slow time scales (Palsson and Lightfoot 1984). This concept
is illustrated in Fig. 2.4 for the glycolytic pathway. For this example pooling be-
tween chemical isomers occurs on the earliest time scales (these time scales are all
faster than milliseconds). There are chemical as well as physiological relevance to
pooling of metabolites, and this process occurs in an organized, hierarchical manner.
One challenge in biology is to understand this process, because being able to pool
metabolites of interest on a particular time scale enables modularization and simpli-
fication of an otherwise complex set of interactions. Furthermore, once the network
is modularized, it may be possible to identify metabolites (or summed grouped
30 N. Jamshidi and B.Ø. Palsson

Fig. 2.4 Beginning from the fastest time scale and moving forward. Beginning from the fastest
time scale and moving forward, components that move together on subsequent time scales are
lumped into an aggregate pool variable. A hierarchical reduction of this network is shown in
Fig. 2.6

of metabolites) that reflect different “functional states.” This may help reduce the
number of experimental measurements required to characterize the function of a
particular network.
Aggregate pools of metabolites can be identified using different approaches. If
there are a small number of possible perturbations of interest, then simulation-driven
methods can be used (Kauffman et al. 2002). This approach is limited, however, if
one wants to characterize all possible responses of the system. An alternative is to
adopt an analytical approach through the analysis of the Jacobian around a particular
steady state (Jamshidi and Palsson 2008b). The network can be dynamically decou-
pled, and then any correlations between metabolites (or fluxes) can be assessed on
every single one of the independent time scales. Through calculation of correla-
tions between all of the metabolites (or fluxes) on progressive time scales, removing
the time scales one by one (beginning with the fastest), and recalculating correla-
tions between the components, one can identify the pools that form. This procedure
(see Jamshidi and Palsson (2008b)) was carried out for the folate network with a
methionine input flux of 200 M/min and are visually depicted in Fig. 2.5.
There were seven independent time scales for the network under these condi-
tions, and there is clear separation of the folate carrier branch from the methionine
cycle, although SAM immediately pools with the folate metabolites. Analysis of
the kinetics in the context of the stoichiometric matrix identified that the pooling
of SAM with the folate cycle was not stoichiometric determined (i.e., there are no
reactions that directly involve metabolites from the folate cycle and SAM), but these
were kinetically driven events.
2 Metabolic Network Dynamics: Properties and Principles 31

Fig. 2.5 Pooling among metabolites on progressive time scales. Methionine input flux at
200 M/h. Time scale hierarchy of metabolic pool formation in the human red blood cell. The
lower left triangle indicates the modes after which pooling occurs between the corresponding
metabolites (one being the fastest time scale). The upper right triangle are plots of the slopes
between the two metabolites for the remaining time scales after pool formation (the origin is al-
ways included in these approximations of the slopes), color coded according to the time scale at
which pooling occurs. A correlation cutoff of 0.9 was used for the pooling criteria determination

2.5.3 Environmental Perturbations

A benefit of building a model in silico is the ability to carry out various perturba-
tions and to observe the changes the occur. The methionine input flux is described
by a zero-order rate expression, and as noted in Table 2.2, all of its entries in the
gradient matrix are 0. However, changes in the methionine input will cause the sys-
tem to shift from one steady state to another. This change may result in altered
network dynamics. The methionine input flux was considered at halved as well as
doubled rates. A cursory glance at the numerical entries shows that many of the
values are significantly different under the different conditions. This implies that
different homeostatic states have different dynamic properties and quantitatively
different systemic – in response to perturbations.
One can immediately see differences in the entries of the gradient matrix, as
well as the K and M matrices for these different conditions. The network-wide
changes are more easily highlighted in the tiled pooling arrays of the networks
for methionine input fluxes of 400 and 50 M/h, as shown in Figs. 2.7 and 2.8.
At these alternate methionine input flux states, the pooling among metabolites has
completely changed. Most notably pooling within the folate cycle occurs much later
at the lower methionine input flux rate. At the much higher methionine input flux
rate, we see that there are effectively two time scales in which pooling occurs, the
first time scale (0:5 ms) and the seventh time scale (45 s).
These results highlight not only the importance of environmental conditions in
the analysis of dynamics in metabolic networks but also the potential for different
dynamic properties at different steady states in networks.
32 N. Jamshidi and B.Ø. Palsson

Fig. 2.6 Hierarchical reduction of the network in Fig. 2.3. Progressive pooling of metabolites in
folate and methionine cycles was determined according to Fig. 2.5

Fig. 2.7 Pooling among

metabolites on progressive
time scales. Methionine input
flux at 50 M/h
2 Metabolic Network Dynamics: Properties and Principles 33

Fig. 2.8 Pooling among

metabolites on progressive
time scales. Methionine input
flux at 400 M/h

2.6 Conclusions

To date there has not been a successful, generalized strategy to build genome-scale
kinetic models. This has been principally due to the large number kinetic parameters
required to define the system which is further confounded by the fact that in vitro
measurements of kinetic constants are often not representative of their numerical
values in vivo. These challenges have lead to the infeasibility of achieving cell scale
models using such approaches. Identification of the key structural and dynamic
properties of networks and the inherent relationships between fluxes and concen-
trations will help to achieve dynamic descriptions of genome-scale models. Here,
we showed how the dynamics of a biochemical reaction network can be described
by dual Jacobian matrices, which is enabled by recognition of the fact that dynamic
interactions are constrained by network topology. Fluxes and concentrations are
dual variables in biochemical reaction networks, but they are related via changes in
fluxes and concentrations. These relationships are described by the gradient matrix.
The ability to convert from one set of variables into another is not just of mathemat-
ical interest, but highlights the underlying roots of the relationship between fluxes
and concentrations. Ultimately the characterization of biological systems is to un-
derstand how the system responds to perturbations. To date, dynamic descriptions
of networks have been confined to the column space; however, the relationships
described here allow one to describe the network in terms of column space or
row space variables. This is of particular interest in biological networks, as the
perturbation variables are generally concentration variables in the column and left
null spaces, whereas the response variables are the fluxes, in the row and right
null spaces. Thus, concentration variables perturb a network, and the flux variables
respond to the perturbation and tie the network together. These are complementary
variables that are tied together in the network by the stoichiometric and gradient
matrices.
A key motivation in being able to build larger models is to then analyze, un-
derstand, and hopefully simplify these networks. Herein, we focused on some of
the approaches for simplification of network dynamics in the context of dynamic
34 N. Jamshidi and B.Ø. Palsson

hierarchies. A goal in these efforts is to window in on a time scale of interest and

to determine the simplified, pooled structure of a network. This effectively filters
out processes that occur to slowly or too quickly to be of interest and may highlight
grouped metabolites that can be used as surrogates for network functional states,
such as the redox state or energy charges of a cell.

2.6.1 Future Directions: Constructing Genome-Scale Models

Previously when models of biochemical reactions and networks have been con-
structed, it has been through the statement of assumptions such as quasi-equilibrium
and quasi-steady state, followed by incorporation of data into the models and curve
fitting parameters; thus the statements of assumptions are in effect “preprocessing”
the model. The description and decomposition of models described here are carried
out from a different perspective. The mechanistic, bilinear interactions are repre-
sented in the stoichiometric matrix, the various high-throughput data types (nucleic
acid, protein, and small metabolite concentrations) are incorporated into the model,
and then the decisions are made for assumptions. These assumptions can be varied
and adjusted depending on the question of interest and the time scale(s) of inter-
est. Thus, this is a mechanistic, data-driven approach, in which assumptions are a
“postprocessing” step of model construction. This approach recognizes and app-
reciates the stoichiometric and gradient matrices as the key matrices in building
large-scale networks. There has been some progress in this area for outlining the
approaches to build kinetic models. We have recently developed an approach that is
practical, feasible, and successful in test cases to date (Jamshidi and Palsson 2010).
As progress continues to be made in the “-omics” field, particularly metabolomics,
we anticipate the development of genome-scale kinetic models in the near future.

Appendix: Details About Matrices

Forming the Gradient Matrix

Dynamic analysis of complex systems is normally carried out with the linearization
of the right-hand side of (2.1). Noting that S is a matrix with constant coeffi-
cients, linearization of (2.1) comes down to the Taylor series expansion of reaction
rates v.x/: ˇ ˇ
dv ˇˇ 1 d2 v ˇˇ
v.x/ D v.x0 / C  .x  x0 / C  .x  x0 /2 C ::: (2.6)
dx ˇx0 2 dx2 ˇx0

Neglecting all second order and higher terms yields,

ˇ
dv ˇˇ
v.x/  v.x0 / C  .x  x0 / (2.7)
dx ˇx0
2 Metabolic Network Dynamics: Properties and Principles 35

When the reference state, x0 , is specified as a steady state for the system, then by
definition,
S  v.x0 / D 0 (2.8)
so that the linearized form of (2.1) is,
ˇ
d.x  x0 / dv ˇˇ
DS  .x  x0 / (2.9)
dt dx ˇx0

So quite naturally one can define the gradient matrix, G,

dv
GD (2.10)
dx
We note that this is not an arbitrary or a definition of mathematical convenience,
but simply the result of linearization of fluxes around a specified reference point.
We further note that the gradient matrix is equal to the nonlogarithmic form of the
elasticity matrix in metabolic control analysis (Hatzimanikatis and Bailey 1996).
The stoichiometric matrix has been investigated in detail in the literature (Palsson
2006). Since the gradient matrix has only recently been recognized (Jamshidi and
Palsson 2008a), time will be spent highlighting and contrasting its key features with
the stoichiometric matrix.

The Jacobian Matrix for Concentrations

Specifying the reference point in (2.7), x0 , to be a steady state for the system, (2.1)
becomes:
dx0
D S  G  x0 (2.11)
dt
in which x0 is the deviation variable, .x  x0 /. Jx D S  G is the Jacobian for the sys-
tem of equations describing the concentration variables. Note that this factorization
separates the chemistry that specifies network topology (through S), and the kinet-
ics and thermodynamics that give the driving forces and their time scale of action
(residing in G). These two effects can be effectively separated by scaling the rows
of G to unity as (Jamshidi and Palsson 2008a):

G D Kv  Mv (2.12)

where the rows in Mv represent the direction of the driving forces (the thermody-
namics) in the row space. Mv is a row-normalized gradient matrix, and each row
corresponds to a reaction. The matrix Kv is diagonal. Its elements represent the
time scales on which the thermodynamic force of a reaction acts, of the kinetics.
In this formulation, the rows of the gradient matrix are drivers and the columns of
the stoichiometric matrix define the directions of motion.
36 N. Jamshidi and B.Ø. Palsson

The Jacobian Matrix for Fluxes

The concentrations and fluxes are two sets of variables that characterize the dynamic
state of a network. Either can in principle be used as the set of independent variables
and the other computed as set of dependent variables. Stoichiometric matrices for
biochemical networks are, however, normally rectangular with m < n, and rank,
r < m. S is thus not invertible and (2.1) cannot be directly converted into a system
of dynamic equations in terms of fluxes.
The gradient matrix enables the change of the system of equations from the con-
centration variables to a system of equations in terms of flux variables. Defining the
flux deviation variable, v0 D G  x0 , and premultiplying (2.9) by the gradient matrix
yields:
dv0
D G  S  v0 (2.13)
dt

Thus the Jacobian matrix is Jv D G  S, when treating the fluxes as the indepen-
dent variables. In a similar way as above, we can scale every column in Jv and factor
the gradient matrix as:
G D Mx  Kx (2.14)

yielding Jv D Mx  Kx  S. Here, Mx has the columns of G normalized to unity,

and the diagonal matrix Kx contains the length of these columns, which corre-
spond to compounds. Note that the elements of Mx represent the kinetic potential
of compounds.
Jv is thus reassembled compound by compound, whereas Jx was assembled re-
action by reaction. In this formulation, drivers (the rows of S) are the sums of the
fluxes in and out of a node multiplied by the kinetic potential of the compound.
The directions of motions are given by the columns of Mx , and the elements in the
diagonal matrix Kx determine the weights or influence of the motions. The direc-
tion of a column in Mx designates the kinetically balanced outflow of a compound
from a node, if the concentration of the compound in that node is perturbed from
steady state.

References

I. Famili and B. Ø. Palsson. The convex basis of the left null space of the stoichiometric matrix
leads to the definition of metabolically meaningful pools. Biophys J, 85:16–26, 2003
V. Hatzimanikatis and J. Bailey. MCA has more to say. J Theor Biol, 182:233–242, 1996
H. Heinrich and S. Schuster. The regulation of cellular systems. Springer, Heidelberg, 1996
R. Heinrich, S. M. Rapoport, and T. A. Rapoport. Metabolic regulation and mathematical models.
Prog Biophys Mol Biol, 32:1–82, 1977
N. Jamshidi and B. Ø. Palsson. Formulating genome-scale kinetic models in the post-genome era.
Mol Syst Biol, 4:171, 2008a
2 Metabolic Network Dynamics: Properties and Principles 37

N. Jamshidi and B. Ø. Palsson. Top-down analysis of temporal hierarchy in biochemical reaction

networks. PLoS Comput Biol, 4:e1000177, 2008b
N. Jamshidi and B. Ø. Palsson. Mass action stoichiometric simulation models: Incorporating ki-
netics and regulation into stoichiometric models. Biophys J, 98(2):175–185, 2010
K. J. Kauffman, J. D. Pajerowski, N. Jamshidi, B. Ø. Palsson, and J. S. Edwards. Description
and analysis of metabolic connectivity and dynamics in the human red blood cell. Biophys J,
83:646–662, 2002
E. Mayr. Cause and effect in biology. Science, 124:1501–1506, 1961
M. S. Okino and M. L. Mavrovouniotis. Simplification of mathematical models of chemical reac-
tion systems. Chem Rev, 98:391–408, 1998
B. Ø. Palsson. Systems biology: Determining the capabilities of reconstructed networks.
Cambridge University Press, Cambridge, 2006
B. Ø. Palsson and E. N. Lightfoot. Mathematical modelling of dynamics and control in metabolic
networks. I. On Michaelis-Menten kinetics. J Theor Biol, 111:273–302, 1984
M. C. Reed, H. F. Nijhout, M. L. Neuhouser, J. F. Gregory, B. Shane, S. J. James, A. Boynton,
and C. M. Ulrich. A mathematical model gives insights into nutritional and genetic aspects of
folate-mediated one-carbon metabolism. J Nutr, 136:2653–2661, 2006
J. Reich and E. Selkov. Energy metabolism of the cell: A theoretical treatise. Academic, London,
1981
I. Segel. Enzyme kinetics. Wiley, New York, 1975
G. Strang. Linear algebra and its applications. Harcourt Brace Jovanovich, San Diego, 1988
Chapter 3
A Deterministic, Mathematical Model
for Hormonal Control of the Menstrual Cycle

R. Drew Pasteur and James F. Selgrade

3.1 Introduction and Biological Background

The reproductive endocrine system is one of the most complex systems in the human
body. In women, there are three sources of hormone production: the hypothalamus,
the pituitary gland, and the ovaries. The hormones produced in these three locations
jointly regulate the processes occurring at all three sites and control the processes
surrounding ovulation and menstruation, primarily by feedback from the ovaries to
the pituitary (Speroff et al. 1999). Figure 3.1 summarizes the interactions of these
feedback loops.
Pharmaceutical use of external hormones is common across the life cycle, most
notably for contraception or treatment of abnormal menstrual cycles in younger
women, and later to suppress the undesirable effects of menopause. Hormone-like
substances, particularly xenoestrogens, can also be unintentionally ingested, via
food (Davis et al. 1993) and drinking water (Rudel et al. 1998). Because breast
cancer risk may be related to total lifetime exposure to bioavailable estrogens, there
is rising concern over the long-term dangers of estrogen exposure, regardless of the
source (Davis et al. 1993). A mathematical model can be used to illustrate hormone
levels which prevent ovulation; this is discussed in Sect. 3.5.
While a deterministic mathematical model cannot fully predict the results of ex-
ternal influences on the reproductive endocrine system, it might identify a potential
course of action to be considered in future clinical research studies. Recent mathe-
matical models, such as Clark et al. (2003) and Reinecke and Deuflhard (2007), are
substantially more complex than their predecessors, taking full advantage of modern
high-speed computers.
The availability of separate bioassays for two forms of the ovarian hormone in-
hibin is relatively new, with full-cycle data first appearing in Groome et al. (1996).
Later studies involving inhibin B, such as the one reported in Welt et al. (1999),

R.D. Pasteur ()

Department of Mathematics and Computer Science, The College of Wooster,
1189 Beall Ave., Wooster, OH 44691, USA
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 39

Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 3,
c Springer Science+Business Media, LLC 2011
40 R.D. Pasteur and J.F. Selgrade

menstruation

FSH IhB growth

corpus
P follicle
luteum 4
IhA
FSH
LH
E
Hypothalamus 2
IhA Pituitary
P4 E
2
E2 P
4
LH
corpus LH preovulatory
luteum LH
follicle

ovulation

Fig. 3.1 Phases of the menstrual cycle. The phases displayed indicate interactions among go-
nadotropins follicle-stimulating hormone (FSH) and luteinizing hormone (LH), and ovarian
hormones estradiol (E2), progesterone (P4), inhibin A (IhA), and inhibin B (IhB). A similar figure
was published first in Fields Inst Comm (Selgrade and Schlosser 1999) by the AMS

had larger numbers of subjects and considered women from multiple age groups.
In addition to inhibin B (IhB), the ovaries also produce inhibin A (IhA), estradiol
(E2, a form of estrogen), progesterone (P4), and other hormones. Ovarian hormone
production gradually declines as women approach menopause. For the inhibins,
this decrease starts early, around age 35 (Welt et al. 1999). A more comprehensive
mathematical model has the potential to give insight into hormonal issues in peri-
menopausal women, not just the younger women represented by the current model
(Clark et al. 2003).
Both E2 and P4 are steroid hormones, produced from acetate and choles-
terol via the same chemical pathway as testosterone. The hypothalamus secretes
gonadotropin-releasing hormone (GnRH), which leads to the production of follicle-
stimulating hormone (FSH) and luteinizing hormone (LH), both of which are
gonadotropins, by the pituitary (Speroff et al. 1999).
3 Model for Hormonal Control of the Menstrual Cycle 41

In the absence of exogenous hormones, the bloodstream hormone levels of

women of childbearing age are not at equilibrium, but rather change in a predictable,
periodic manner. Together these dynamic interactions control the menstrual cycle.
While the average human menstrual cycle for a healthy, fertile woman is 28 days,
most women do not have a 28-day cycle, and cycles several days longer or shorter
can be considered normal (Speroff et al. 1999; Vollman 1977; Treloar et al. 1967).
Adolescents in their early teens, as well as women approaching menopause, tend to
have longer, irregular cycles, during their first and last few menstrual years, respec-
tively. Ovulation, which is necessary for fertility, may or may not occur during these
long, irregular cycles (Yen 1999; Vollman 1977; Treloar et al. 1967). The menstrual
cycle is stopped temporarily by oral contraceptives or pregnancy and is permanently
ended naturally at menopause, or sooner via a hysterectomy.
The menstrual cycle can be tracked by noting the dates of the start of menstrua-
tion, vaginal bleeding typically lasting several days. The first day of menstruation is
conventionally considered the first day of a new menstrual cycle. Ovulation occurs
roughly halfway through the cycle, splitting it into two phases; the first half is called
the follicular phase, and the latter half the luteal phase (Fig. 3.1).
At ovulation, a mature follicle ruptures, releasing an egg for possible fertiliza-
tion. The developmental processes leading up to that point last for roughly 85 days
(Oktay et al. 1998); so at any time, there are multiple active follicles at various
stages of maturity, regardless of the phase of the menstrual cycle (Baird 1984).
At puberty, each ovary contains roughly 250,000 immature egg cells called oocytes,
but only a few hundred of these will ever reach full follicular maturity and rupture at
ovulation (Hadley 1992). The others will at some point undergo atresia, a degenera-
tive process; eventually, menopause occurs when no active follicles remain (Speroff
et al. 1999). In a healthy woman not taking hormonal contraceptives, typically only
one follicle develops to the point of ovulation during each menstrual cycle. How-
ever, the incidence of fraternal twins, due to multiple fertilized eggs, demonstrates
the possibility of more than one follicle completing the developmental process dur-
ing a given cycle.
Shortly before menstruation, roughly 2–3 weeks before the next expected ovula-
tion, multiple follicles in each ovary begin an advanced state of development, due
to slowly rising levels of FSH. Each such follicle has the potential of rupturing in
ovulation during the upcoming cycle (Pache et al. 1990). During the week after
menstruation, one follicle begins to grow exponentially, while the others begin to
break down through atresia (Baird 1984). The physiological processes by which
one follicle becomes dominant are not well understood (Zeleznik and Pohl 2006).
FSH levels are elevated throughout the follicular phase, as this is a necessary
condition for the advanced follicular development taking place in the maturing pri-
mary follicle. During the latter week of the follicular phase, this follicle releases E2
in large quantities, to the point that there is an exponential rise in bloodstream levels
of E2 (Baird 1984). The high E2 levels reduce the release of FSH by the pituitary,
which eventually causes a sufficient drop in ovarian FSH levels to accelerate the
demise of less-developed cohort follicles (Speroff et al. 1999). The primary follicle
is unaffected due to internal storage of FSH (Baird 1984).
42 R.D. Pasteur and J.F. Selgrade

Throughout the follicular phase, a large amount of LH is stored in reserve. It is

then emptied into the bloodstream over about 2 days. The resulting high-amplitude
spike in circulating LH levels, peaking at eight times the baseline level, on average
(Welt et al. 1999), ultimately leads to ovulation. Simultaneously, there is a sig-
nificant release of FSH; while less dramatic, it also has physiologic implications
related to ovulation. According to the biological literature, the LH surge is trig-
gered by sustained high concentrations of E2, above a threshold level for at least
36–48 h (Speroff et al. 1999; Young and Jaffe 1976). However, at subthreshold
levels, increased E2 causes a decrease in LH release; so LH does not vary monotoni-
cally with E2. Our model included dual control of LH by E2, with negative feedback
on LH release, but positive feedback on LH synthesis at levels above a threshold (Liu
and Yen 1983). As E2 increases sharply during the late follicular phase, we modeled
the resulting promotion of LH synthesis as being primarily responsible for the LH
surge.
Concentrations of circulating P4 are low throughout the follicular phase, but
rise slowly just before mid-cycle, and likely have a role in the timing of the LH
surge (Speroff et al. 1999). Circulating levels of one more hormone, IhB, also peak
at mid-cycle, possibly due to a substantial release from the ruptured follicle at ovu-
lation, but this is unclear (Speroff et al. 1999; Muttukrishna et al. 2000).
Following ovulation, the luteal phase lasts roughly 2 weeks for a typical healthy
woman. During this phase, the ruptured primary follicle, now called the corpus lu-
teum, releases large amounts of P4, preparing the body for a potential pregnancy
(Baird 1984). Within a few days after ovulation, IhB levels drop substantially, and
IhB has no significant role in regulating FSH during this time. However, IhA con-
centrations, already rising since the mid-follicular phase, peak during the mid-luteal
phase. There is an extended, low-amplitude peak of E2 during the luteal phase; to-
gether, E2 and IhA suppress the production of FSH during the latter half of each
cycle (Welt et al. 1999; Yen 1999).

3.2 The Pituitary and Ovarian Models

In modeling the reproductive endocrine system in women, we began with separate

models of the relevant processes occurring in the pituitary gland and the ovaries.
Pituitary hormone production is pulsatile in nature, in response to pulsatile stimu-
lation by GnRH from the hypothalamus (Speroff et al. 1999). However, following
the models presented by Schlosser and Selgrade (2000), we lumped together effects
of the hypothalamus and pituitary gland, and smoothed out the pulsatile effects by
considering average synthesis, lengthening the time scale. As discussed previously,
when LH and FSH are produced in the anterior pituitary, there is not an immedi-
ate full release of these hormones into the bloodstream upon synthesis. Hence, we
provided for a pituitary reserve called a releasable pool, which is a biologically rea-
sonable assumption based on Speroff et al. (1999).
3 Model for Hormonal Control of the Menstrual Cycle 43

To model the pituitary hormones LH and FSH, we tracked for each the amount
of hormone in the releasable pool, as well as the concentration of hormone in the
bloodstream. This required quantifying the rates of three biological processes for
each hormone: synthesis, release, and clearance. Because synthesis and release rates
depend on the circulating concentrations of the four ovarian hormones, we con-
structed periodic functions of time to represent the concentrations of E2, P4, IhA,
and IhB, fitting the data from Welt et al. (1999). These functions served as inputs
into the systems of differential equations modeling LH and FSH.
Because there are observed delays between various peaks and valleys in the
graphs of ovarian hormones and the expected corresponding peaks and valleys of
graphs of LH and FSH, in Welt et al. (1999) as well as in McLachlan et al. (1990),
we used constant time delays to represent such effects.
Following Schlosser and Selgrade (2000), we modeled LH synthesis, release,
and clearance with the system of differential equations [(3.1) and (3.2)]. The state
variables RPLH and LH represent the amount of LH in the releasable pool and the
LH concentration in the blood, respectively; the parameter v denotes blood volume.
Figure 3.2 is a schematic diagram for the hypothalamus/pituitary system and the
influences on it by the ovaries.

d
RPLH D LH Synthesis.E2; P4/  LH Release.E2; P4; RPLH/
dt
d 1
LH D LH Release.E2; P4; RPLH/  LH Clearance.LH/ (3.1)
dt v
As discussed in the previous section, E2 and P4 have mutually antagonistic ef-
fects on the synthesis and release of LH. Also noteworthy in (3.2) is the numerator
of the LH synthesis term, which includes a rational function called a Hill function.
Physiological research suggests that a sustained high concentration of E2 is needed
in the bloodstream for at least 36–48 h, to trigger the LH surge, which in turn causes
ovulation to occur (Speroff et al. 1999; Young and Jaffe 1976). It is unclear whether

Fig. 3.2 Schematic diagram for the LH system given by (3.1) and (3.2). Arrows represent regula-
tory feedback of a given hormone on a particular physiologic process. Positive and negative signs
indicate promotion and inhibition, respectively, of that process
44 R.D. Pasteur and J.F. Selgrade

this E2 effect controls the synthesis or the release of LH, but our model includes
it in the synthesis term. The parameter KmLH is a threshold value for E2 at which
half-maximal LH synthesis occurs. As the exponent a increases, the gradient for LH
synthesis becomes steep as the E2 concentration is near this threshold level. In our
model, for high values of a, relatively little LH synthesis (above the baseline level
represented by v0;LH ) occurred for E2 levels below KmLH, but production increased
sharply once the concentration of E2 exceeded this value.

.E2.t dE /=KmLH /a

v0;LH C v1;LH  1C.E2.t dE /=KmLH /a
LH Synthesis.E2; P4/ D P4.t dP /
1C KiLH;P
1 C cLH;P  P4.t/
LH Release.E2; P4; RPLH/ D kLH   RPLH.t/
1 C cLH;E  E2.t/
LH Clearance.LH/ D rLH  LH.t/ (3.2)

Of the parameters in (3.1) and (3.2), only two were estimated from the phys-
iological literature. For the clearance rate of LH from the bloodstream, we used
rLH  14 day1 , based on Kohler et al. (1968), and for the volume of circulating
blood, we used v  2:5 L. The delays associated with the effects of E2 and P4 on
LH synthesis, dE and dP , were expected to be on the scale of 0–2 days, and a be-
ginning estimate of the E2 synthesis threshold value KmLH was taken from the
bloodstream concentrations of E2 given in Welt et al. (1999), but none of the other
parameters were known.
The model for FSH in the pituitary, given by (3.3), is similar in form to that of
(3.1), but without a Hill function, because there is no physiological evidence that ei-
ther FSH synthesis or release is as sensitive as LH to small changes in concentration
of any ovarian hormone. Both IhA and IhB inhibit synthesis of FSH, and the release
of FSH is progressively stifled as concentrations of E2 rise. Just as with LH, high
levels of P4 lead to an increase in the relative release rate of FSH. Equations (3.3)
and (3.4) model the synthesis, release, and clearance of FSH (Fig. 3.3).

d
RPFSH D FSH Synthesis.IhA; IhB/  FSH Release.E2; P4; RPFSH/
dt
d 1
FSH D FSH Release.E2; P4; RPFSH/  FSH Clearance.FSH/ (3.3)
dt v

vFSH
FSH Synthesis.IhA; IhB/ D IhA.t dIhA / IhB.t dIhB /
1C KiFSH;IhA C KiFSH;IhB
1 C cFSH;P  P4.t/
FSH Release.E2; P4; RPFSH/ D kFSH   RPFSH.t/
1 C cFSH;E  E2.t/
FSH Clearance.FSH/ D rFSH  FSH.t/ (3.4)
3 Model for Hormonal Control of the Menstrual Cycle 45

Fig. 3.3 Schematic diagram for the FSH system given by (3.3) and (3.4). Arrows represent regula-
tory feedback of a given hormone on a particular physiologic process. Positive and negative signs
indicate promotion and inhibition, respectively, of that process

The parameter v in (3.4) is unchanged from the LH system in (3.2). We had

a physiological estimate that the clearance rate of FSH from the bloodstream is
rFSH  8:21 day1 , based on Coble et al. (1969). Additionally, we expected that the
delays in the effects of the inhibins, dIhA and dIhB , are on the order of 1–3 days, but
all of the other parameters were determined to fit the model to the clinical data best.
In agreement with the previous models of the human menstrual cycle such as
Bogumil et al. (1972), as well as Selgrade and Schlosser (1999), we assumed that
the clearance rates of the ovarian hormones are much faster than those of the go-
nadotropins LH and FSH. With this in mind, we used a different approach to model
the bloodstream concentrations of the ovarian hormones, considering them to be
at quasi-steady state (Keener and Sneyd 2009) because of the difference in time
scales. Selgrade and Schlosser (1999) divided the menstrual cycle into nine stages,
each representing the capacity of the ovary to produce various hormones, with three
stages for the follicular phase, two stages around the time of ovulation, and four
stages for the luteal phase. The circulating concentration of each ovarian hormone
was then modeled as a linear combination of these nine stages as in (3.6) below, plus
a baseline level in some cases. The fast clearance simplifies the model, avoiding the
necessity of tracking synthesis, release, and clearance of each hormone.
We used a similar formulation for the ovarian model, with 12 stages instead of
nine. Because our model includes an additional hormone IhB, which peaks early
in the follicular phase, we needed additional stages that likewise peak very early
in the cycle, to represent preantral follicles. Figure 3.4 is a schematic diagram for
the 12-stage ovarian model. The compartments represent 12 sequential stages of
the menstrual cycle, further breaking down the stages of Fig. 3.1. The mass in a
compartment represents the ability of the ovary to produce particular hormones.
Arrows from one compartment to another indicate transfer of mass, while arrows
from a compartment to itself indicate an exponential growth process within a stage.
Transitions noted with constant coefficients are linear mass transfers; that is, the
transfer rate is proportional to the mass of the stage from which the transfer occurs.
The transfers in the early stages are dependent on LH and/or FSH, as both pituitary
46 R.D. Pasteur and J.F. Selgrade

Fig. 3.4 Schematic diagram for the 12-stage ovarian model. Arrows represent transfer of mass
from 1 stage to another or mass growth within a stage

hormones are important during the follicular phase. Arrows from a stage to itself, in
the 3 and 4 stages, represent exponential growth of the primary follicle, during the
mid-to-late follicular phase.
Also notable is the Hill function which provides the mass flow for the 1 stage,
starting the whole process. The physiological literature suggests that there may be a
critical FSH threshold for production of inhibin B during the luteal-follicular transi-
tion (Welt et al. 1997). For this reason, we have included another Hill function, and
FSH logically is the catalyst. It should be noted that IhB synthesis does not stem
directly from the presence of FSH, but rather follows indirectly, as FSH promotes
follicular development, eventually leading to IhB production.
Each of the 12 stages corresponds to a differential equation, based on the inflow
and outflow of mass at that stage. At any time, the bloodstream concentrations of
E2, P4, IhA, and IhB are determined from the 12 stage values by the use of auxiliary
(3.6), one for each of the four ovarian hormones. Each auxiliary equation consists
of a constant baseline level plus a linear combination of some (but not all) of the
12 stages. Equation (3.5) gives the 12 ordinary differential equations for the ovarian
model, with 21 unknown parameters. Equation (3.6) gives the four auxiliary equa-
tions which determine the ovarian hormone concentrations in terms of the stages,
via 16 additional parameters.

d .FSH=f1 /b
PrA1 D f2   f3  FSH  PrA1
dt 1 C .FSH=f1 /b
d
PrA2 D f3  FSH  PrA1  f4  LHı  PrA2
dt
d
SeF1 D f4  LHı  PrA2 C .c1  FSH  c2  LH˛ /  SeF1
dt
d
SeF2 D c2  LH˛  SeF1 C .c3  LHˇ  c4  LH/  SeF2
dt
d
PrF D c4  LH  SeF2  c5  LH  PrF
dt
3 Model for Hormonal Control of the Menstrual Cycle 47

d
OvF D c5  LH  PrF  c6  OvF
dt
d
Sc1 D c6  OvF  d1  Sc1
dt
d
Sc2 D d1  Sc1  d2  Sc2
dt
d
Lut1 D d2  Sc2  k1  Lut1
dt
d
Lut2 D k1  Lut1  k2  Lut2
dt
d
Lut3 D k2  Lut2  k3  Lut3
dt
d
Lut4 D k3  Lut3  k4  Lut4 (3.5)
dt

E2 D e0 C e1  SeF2 C e2  PrF C e3  Lut4

P4 D p0 C p1  Lut3 C p2  Lut4
IhA D h0 C h1  PrF C h2  Lut2 C h3  Lut3 C h4  Lut4
IhB D j0 C j1  PrA2 C j2  PrF C j3  OvF (3.6)

3.3 Fitting Parameters

We wished to determine the values of the unknown parameters which made the
model best fit the data of Welt et al. (1999). This data set represents the average daily
circulating hormone levels for 23 normally cycling women, all aged between 20 and
35 years old. The data were centered to the day of the mid-cycle LH surge. Because
the parameter-fitting process was daunting due to the large number of parameters,
we began by considering separately the LH system [(3.1) and (3.2)], FSH system
[(3.3) and (3.4)], and ovarian system [(3.5) and (3.6)]. We estimated parameters for
each of these three systems individually, and then used those results as a starting
point for working with a larger, merged model.
The problem this presented is that the three systems are interdependent. The
LH system depends on E2 and P4, the FSH system depends on all four ovarian
hormones, and the ovarian system depends on LH and FSH. With this in mind,
we used explicit functions of time in each system to represent the hormones not
being modeled by that system. We used 28-day periodic functions (to match the as-
sumed period of the data from Welt et al. (1999)) consisting of a baseline constant
term plus one or more peaks generated by negative-exponential terms. Such terms
take the form c  exp..t  a/2 /, where c is the amplitude of the peak and a is
the time (in days) at which the peak occurs. As an example, we show in Fig. 3.5
48 R.D. Pasteur and J.F. Selgrade

Fig. 3.5 Time-dependent input functions for FSH and LH, along with Welt data. Dots represent
the clinical data, while lines indicate the approximating functions

Fig. 3.6 Unmerged LH system model output (line), together with Welt data for LH (dots)

the time-dependent function plots, along with the data, for the two pituitary func-
tions. Additional information regarding the time-dependent functions is included in
Sect. 3.3 of Pasteur (2008).
The LH system has 12 parameters, counting the biological constants rLH and v,
while the FSH system has 10 parameters, including rFSH and the same v. Figures 3.6
and 3.7 show the best fits found for these models. Additional details (includ-
ing lists of the parameter values) are given in Sects. 3.3 and 5.5, respectively, of
Pasteur (2008).
3 Model for Hormonal Control of the Menstrual Cycle 49

Fig. 3.7 Unmerged FSH system model output, together with Welt data for FSH

Fig. 3.8 Output for E2 and P4 from the unmerged ovarian model, with Welt data

In the ovarian system, there are 21 parameters for the 12 differential equations
(3.5) and 16 additional parameters for the 4 auxiliary equations (3.6). Because the
ovarian stages are merely a modeling tool, we include plots only of the auxiliary
equations (Figs. 3.8 and 3.9), which show the best fits for the ovarian hormone lev-
els, computed as linear combinations of the 12 ovarian stages, via (3.6).
After obtaining parameter estimates using these three separate models, the next
step in the modeling process was to merge them, eliminating the need for time-
dependent input functions. Successful completion of this step created a model which
is time-autonomous and which can be validated by other data.
The merged model includes 16 differential equations ((3.1)–(3.5)), some with
constant delays, plus four auxiliary equations (3.6). There are 58 parameters in this
50 R.D. Pasteur and J.F. Selgrade

Fig. 3.9 Output for the inhibins from the unmerged ovarian model, with Welt data

merged model, of which only three could be obtained from the physiological litera-
ture. Finding the best-fit parameter set was a high-dimensional problem in nonlinear
global optimization, because we wished to minimize the total error over the six
hormone outputs, as compared to the Welt data, across several cycles. Further com-
plicating matters, the appropriate initial conditions were known for only 2 of the 16
state variables, because neither the releasable pool holdings nor the 12 ovarian stage
values are readily approximated at the luteal-follicular transition.
A variety of techniques were used in the parameter identification process, most
notably the Nelder–Mead local optimization method, minimizing a weighted total of
the sums of squared errors for the six hormone profile outputs. Sensitivity analysis
(discussed in Sect. 3.4) was helpful in understanding the effects of changing indi-
vidual parameters, as well as in determining which parameters have proportionally
the largest effects on various output measures.
In Fig. 3.10, we show the model output for the optimized parameter set; the as-
sociated values for all 58 parameters are listed in Sect. 5.5 of Pasteur (2008).
These hormone profiles eventually occur for any biologically realistic initial con-
ditions, because there is only one periodic solution and it is an asymptotically
stable (i.e., attracting) solution. The period of the solution is 28.0 days, matching
both the assumed period of the Welt data and the most common menstrual cycle
length (Speroff et al. 1999; Vollman 1977; Treloar et al. 1967). However, the ex-
istence of only one stable periodic solution stands in contrast to the findings of
Clark et al. (2003), in which two stable periodic solutions – one normal and one
anovulatory – were found after fitting a similar model to a different data set, the data
3 Model for Hormonal Control of the Menstrual Cycle 51

Fig. 3.10 Stable normal cycle for the merged model (line), with Welt data (dots)

from McLachlan et al. (1990). The McLachlan data set included only total inhibin
and had a higher mid-cycle E2 peak, perhaps explaining the different qualitative re-
sult. These two solutions could be thought of as representing two different women,
only one of whom has the possibility of sustained abnormal, anovulatory menstrual
cycles, in the absence of external effects.
52 R.D. Pasteur and J.F. Selgrade

3.4 Parameter Sensitivity and Bifurcations

Because of the difficulty of global optimization involving a large number of param-

eters, together with the complicated form of the differential equations, it is very
possible that there are other parameter sets which fit the data as well, or better than
the one that we found. The Welt data set consists of average hormone concentrations
throughout the menstrual cycle, but clearly there are substantial differences among
healthy women. Given that our model is deterministic, individual differences among
women can only be expressed using different parameter sets. With this in mind, we
wished to consider whether changes to individual parameters could lead to quali-
tatively different dynamic behavior. In mathematical terms, we were interested in
bifurcations, particularly those incurred by changes to a single parameter.
To determine which parameters were most likely to cause substantive changes
in the hormone profiles, we computed normalized sensitivity coefficients. These are
quantitative measures of the relative effect of a parameter change on some output
measure X . Hence, for a change in some parameter from p to p.1 C "/, the relative
sensitivity coefficient S.p/ is given by (3.7).
 
1 X.p C "  p/  X.p/
S.p/ D (3.7)
" X.p/

If a large change in a particular parameter is required to bring about a small

change in model output, then the associated parameter has a sensitivity coefficient
with a very small magnitude. On the other hand, if a small parameter change has
substantial impact on the model output, then the coefficient has a large magnitude.
The sign of a sensitivity coefficient is determined by whether increasing a parameter
causes a measured output to increase or decrease. These coefficients are dimension-
less, allowing for straightforward comparisons among all parameters in a model. It
is important to note that sensitivity coefficients are tied to the parameter set at which
they are measured, and thus will vary for different parameter sets.
Given that our model has six hormone profiles as outputs, there are a variety of
measures we could have used in determining sensitivity. Because this process can be
easily automated, instead of settling on any one, we computed sensitivity for each
parameter with respect to many output measures, typically the peak and minimal
levels of each hormone at various points in the cycle, as well as with respect to
the cycle period length. Looking through the resulting tables of data, we identified
which parameters are the most sensitive overall.
Six parameters stand out as the most sensitive, as shown in Table 3.1, and thus
these were the parameters most likely to be associated with qualitative changes in
the dynamics of the model. The most sensitive parameter is ˛, which is a fractional
exponent indicating the strength of the promotion by LH of mass transfer between 2
stages in the ovarian model. Two other sensitive parameters, c1 and c2 , are likewise
involved in early stages of the ovarian model. Matching reasonable intuition, the
threshold values in both Hill functions, KmLH and f1 , also have high-magnitude
sensitivity coefficients. The baseline LH synthesis coefficient v0;LH rounds out this
3 Model for Hormonal Control of the Menstrual Cycle 53

Table 3.1 The most sensitive parameters, and their associated

sensitivity coefficients, with respect to the mid-cycle E2 peak.
Positive coefficients indicate that increasing the parameter leads
to a higher E2 peak, while negative coefficients indicate that a
parameter increase leads to a lower peak concentration of E2
Parameter ˛ KmLH c2 c1 v0;LH f1
Sensitivity 2:37 1:16 1:03 0:87 0:60 0:59

Fig. 3.11 Bifurcation diagram for the parameter ˛, showing the amplitude of hormone change
throughout the cycle, as ˛ is varied. The dot indicates the best fit value ˛ D 0:8001. For ˛ >
0:913, the amplitude is zero, so the system is at equilibrium, i.e., steady-state. This equilibrium is
stable and could represent a woman taking a continuous dose of oral contraceptives (discussed in
Sect. 3.5) as there would also be no LH surge, and thus no ovulation

group of six. For the above list, summarized in Table 3.1, we consider sensitivity
with respect to the preovulatory peak of E2, which triggers the LH surge, leading to
ovulation; however, a variety of other measures could be used, and many of the same
parameters would still top the list. Comprehensive tables of sensitivity coefficients
are included in Appendix B.2 of Pasteur (2008).
In the model of Clark et al. (2003), two stable periodic solutions were observed
for the best-fit parameter set. The implication is that, according to that model, a
woman could have either a normal or an abnormal menstrual cycle, sustained over a
long duration of time, depending on her initial hormone levels. Furthermore, Clark
et al. (2003) showed that a short-term dose of external hormones could cause a
change from an abnormal cycle to a normal one, or the reverse. Similar results, for
a model different than the one presented here, were shown in Chap. 4 of Pasteur
(2008), by changing KmLH, the E2 threshold for LH synthesis.
In contrast, for our model, we observed at most one stable periodic solution,
regardless of varying any single parameter from Table 3.1 within biologically re-
alistic ranges. The existence of a unique stable periodic solution implies that after
any disruptions due to external influences, normal menstrual cycles will eventually
be resumed. Figure 3.11 shows that by altering the value of the most sensitive pa-
rameter ˛, we obtained a different type of change in the model behavior, i.e., the
periodic solution is replaced by a solution in which hormone levels remain constant
over time and ovulation cannot occur.
54 R.D. Pasteur and J.F. Selgrade

3.5 Exogenous Hormone Effects

Ingestion of exogenous reproductive hormones by women is common today,

sometimes intentional (as in pharmaceutical use) and other times not (through
pollutants in drinking water). One of the goals of a model such as the one we
present is to predict the effects of external hormonal influences. Based on standard
contraceptive uses, we know that a sufficiently large daily dose of E2 suppresses
ovulation. In our model, we included exogenous E2 by including one or more ad-
ditional step function terms in the auxiliary (3.6) for E2. As a part of the validation
process, we conducted a dose–response analysis of external estrogen intake. The
mid-cycle LH surge is the key marker of ovulation; if the LH surge is sufficiently
suppressed, ovulation will not occur.
In Figs. 3.12–3.14, we present three cases, involving continuous low, medium,
and high doses, respectively, of external E2. In each case, the solid line represents a
woman treated with external E2, and the dashed line is the untreated control. In the
low-dose case (Fig. 3.12), the bloodstream E2 concentration is raised by 25 pg/ml,
an increase of roughly 10% from the peak level observed in our model. There is
an associated cycle length reduction from 28.0 to 26.5 days, but only a minimal
decrease in the peak amplitude of LH, so ovulation would likely still occur. With
a medium dose (75 pg/ml, roughly 30% above the natural peak), LH is strongly
suppressed, as shown in Fig. 3.13, and E2 concentrations also take a low-amplitude

Fig. 3.12 Effects of a 25 pg/ml dose of external E2. The solid curves show the model output with
treatment, while the broken curves represent a control group from Fig. 3.10
3 Model for Hormonal Control of the Menstrual Cycle 55

Fig. 3.13 LH surge suppressed by a 75 pg/ml dose of external E2

Fig. 3.14 Equilibrium induced by a 125 pg/ml dose of external E2

56 R.D. Pasteur and J.F. Selgrade

profile. With a sufficiently large dose (125 pg/ml, about half as much as the observed
peak in the Welt data), the model predicts a pharmaceutically induced equilibrium,
with constant levels of all hormones, as shown in Fig. 3.14.
All of this behavior matched our expectations based on standard oral contra-
ceptive treatment protocols. Increasingly large amounts of exogenous E2 suppress
the LH surge (and hence, suppress ovulation) to an increasingly large degree. It is
also noteworthy that upon cessation of the treatment, the modeled hormone con-
centrations return to the stable normal cycle of Fig. 3.10 after a few months. These
unsurprising results help validate the model.

3.6 Conclusion

We have presented a model for hormonal control of the female reproductive en-
docrine system which produces, with reasonable accuracy, hormone profiles typi-
cal of a young, healthy woman, throughout several consecutive menstrual cycles.
Unlike other prior models such as Clark et al. (2003), this model takes advantage of
data from relatively new bioassays for two forms of inhibin. Intuitively reasonable
effects on LH have been shown for the treatment with external E2. The use of mod-
ern high-speed computers and specialized software tools has allowed for simulation
and analysis with a complex model.
However, further improvements can be made to the model presented. The FSH
profile in Fig. 3.10 is a far-from-perfect fit, particularly in the week surrounding ovu-
lation in each cycle. Additionally, simulations of pharmaceutical use of external P4
and combinations of E2 and P4, both of which are used clinically for contraception
in some cases (Speroff et al. 1999), did not result in the expected suppression of the
LH surge. These issues demonstrate a need for refitting the parameter set, perhaps
with one or more global optimization techniques, to obtain a model which better re-
flects the biological reality. It is possible that the underlying equations, not just the
parameters, will have to be altered to achieve this goal. It is unknown whether the ef-
fect of E2 on LH is on the synthesis of LH or the release. Our present model assumes
that E2 inhibits LH release but, at high levels, promotes LH synthesis. Future work
will investigate the possibility that the effect on LH release changes from inhibitory
to stimulatory after E2 reaches a threshold level for an extended period of time.
Additionally, a deterministic model is limited in scope, because each woman
has a different hormone profile. There are also fluctuations over time (from one
cycle to the next) in the hormone concentrations of the same woman. Both of these
issues point to the need for a stochastic model. Such a model, based on randomized
parameters, could show individual variations across the population being studied,
and may even allow for simulated clinical trials of hormonal treatment regimens.
3 Model for Hormonal Control of the Menstrual Cycle 57

References

D. T. Baird. The ovary. In C. R. Austin and R. V. Short, editors, Hormonal control of reproduction,
pages 91–114, Cambridge University Press, Cambridge, 1984
R. J. Bogumil, M. Ferin, J. Rootenberg, L. Speroff, and R. L. Vande Wiele. Mathematical studies
of the human menstrual cycle. I: Formulation of a mathematical model. Journal of Clinical
Endocrinology and Metabolism, 35:126–143, 1972
L. H. Clark, P. M. Schlosser, and J. F. Selgrade. Multiple stable periodic solutions in a model
for hormonal control of the menstrual cycle. Bulletin of Mathematical Biology, 65:157–173,
2003
Y. D. Coble, P. O. Kohler, C. M. Cargille, and G. T. Ross. Production rates and metabolic clearance
rates of human follicle-stimulating hormone in premenopausal and postmenopausal women.
The Journal of Clinical Investigation, 48:359–363, 1969
D. L. Davis, H. L. Bradlow, M. Wolff, T. Woodruff, D. G. Hoel, and H. Anton-Culver. Medical
hypothesis: Xenoestrogens as preventable causes of breast cancer. Environmental Health Per-
spectives, 101:372–377, 1993
N. P. Groome, P. J. Illingworth, M. O’Brien, R. Pai, F. E. Rodger, J. P. Mather, and A. S. McNeilly.
Measurement of dimeric inhibin B throughout the human menstrual cycle. Journal of Clinical
Endocrinology and Metabolism, 81:1401–1405, 1996
M. E. Hadley. Endocrinology. Simon and Schuster Co., Englewood Cliffs, third edition, 1992
J. Keener and J. Sneyd. Mathematical physiology I: Cellular physiology. Springer, New York,
second edition, 2009
P. O. Kohler, G. T. Ross, and W. D. Odell. Metabolic clearance and production rates of human
luteinizing hormone in pre- and postmenopausal women. The Journal of Clinical Investigation,
47:38–47, 1968
J. H. Liu and S. S. C. Yen. Induction of midcycle gonadotropin surge by ovarian steroids in
women: A critical evaluation. Journal of Clinical Endocrinology and Metabolism, 57:797–802,
1983
R. I. McLachlan, N. L. Cohen, K. D. Dahl, W. J. Bremner, and M. R. Soules. Serum inhibin
levels during the periovulatory interval in normal women: Relationships with sex steroid and
gonadotropin levels. Clinical Endocrinology, 32:39–48, 1990
S. Muttukrishna, T. Child, G. M. Lockwood, N. P. Groome, D. H. Barlow, and W. L. Ledger. Serum
concentrations of dimeric inhibins, activin A, gonadotropins, and ovarian steroids during the
menstrual cycle in older women. Human Reproduction, 15:549–556, 2000
K. Oktay, H. Newton, J. Mullan, and R. G. Gosden. Development of human primordial follicles to
antral stages in SCID/hpg mice stimulated with follicle stimulating hormone. Human Repro-
duction, 13:1133–1138, 1998
T. D. Pache, J. W. Wladimiroff, F. H. de Jong, W. C. Hop, and B. C. J. M. Fauser. Growth patterns
of nondominant ovarian follicles during the normal menstrual cycle. Fertility and Sterility, 54:
638–642, 1990
R. D. Pasteur. A multiple-inhibin model of the human menstrual cycle. PhD thesis, North Carolina
State University, 2008. URL http://www.lib.ncsu.edu/theses/available/etd-06102008-194807/
I. Reinecke and P. Deuflhard. A complex mathematical model of the human menstrual cycle. Jour-
nal of Theoretical Biology, 247:303–330, 2007
R. A. Rudel, S. J. Melly, P. W. Geno, G. Sun, and J. G. Brody. Identification of alkylphenols and
other estrogenic phenolic compounds in wastewater, septage, and groundwater on Cape Cod,
Massachusetts. Environmental Science and Technology, 32:861–869, 1998
P. M. Schlosser and J. F. Selgrade. A model of gonadotropin regulation during the menstrual cycle
in women: Qualitative features. Environmental Health Perspectives, 108:873–881, 2000
J. F. Selgrade and P. M. Schlosser. A model for the production of ovarian hormones during the
menstrual cycle. Fields Institute Communications, 21:429–446, 1999
L. Speroff, R. H. Glass, and N. G. Kase. Clinical gynecologic endocrinology and infertility.
Lippincott Williams and Wilkins, Philadelphia, sixth edition, 1999
58 R.D. Pasteur and J.F. Selgrade

A. E. Treloar, R. E. Boynton, G. B. Borghild, and B. W. Brown. Variation of the human menstrual

cycle through reproductive life. International Journal of Fertility, 12:77–126, 1967
R. F. Vollman. The menstrual cycle. In E. Friedman, editor, Major problems in obstetrics and
gynecology, pages 1–193, 1977
C. K. Welt, K. A. Martin, A. E. Taylor, G. M. Lambert-Messerlian, W. F. Jr. Crowley, J. A. Smith,
D. A. Schoenfield, and J. E. Hall. Frequency modulation of follicle-stimulating hormone (FSH)
during the luteal-follicular transition: Evidence of FSH control of inhibin B in normal women.
Journal of Clinical Endocrinology and Metabolism, 82:2645–2652, 1997
C. K. Welt, D. J. McNicholl, A. E. Taylor, and J. E. Hall. Female reproductive aging is marked by
decreased secretion of dimeric inhibin. Journal of Clinical Endocrinology and Metabolism, 84:
105–111, 1999
S. S. C. Yen. The human menstrual cycle: Neuroendocrine regulation. In S. S. C. Yen, R. B. Jaffe,
and R. L. Barbieri, editors, Reproductive endocrinology: Physiology, pathophysiology, and
clinical management, pages 191–217, W. B. Saunders Co., Philadelphia, 1999
J. R. Young and R. B. Jaffe. Strength-duration characteristics of estrogen effects on gonadotropin
response to gonadotropin-releasing hormone in women. Journal of Clinical Endocrinology and
Metabolism, 42:432–442, 1976
A. J. Zeleznik and C. R. Pohl. Control of follicular development, corpus luteum function, the
maternal recognition of pregnancy, and the neuroendocrine regulation of the menstrual cycle in
higher primates. In J. D. Neill, editor, Knobil and Neill’s physiology of reproduction. Elsevier,
2006
Chapter 4
Modeling Transport Processes and Their
Implications for Chemical Disposition
and Action

Nick Plant

4.1 Introduction

The body is constantly being exposed to chemicals: This exposure may range from
deliberate (e.g., therapeutic agents, food chemicals) to undesirable (e.g., environ-
mental and food contaminants). The first role of the body must therefore be to
sort these chemicals, allowing ingress of those that are beneficial to the body while
rapidly removing those that could elicit harm. This process must be achieved in a
rapid and efficient manner and be coordinated such that the most efficient response
is elicited for any given chemical exposure. To be able to model such behavior,
and hence predict the outcome of any subsequent exposure, it is first necessary to
understand the basic mechanisms by which the body responds to chemical insult.

4.1.1 The Fate of Chemicals in the Body

The fate of a chemical within the body is determined by the processes of absorption,
distribution, metabolism and excretion (ADME). These four stages control not only
the amount of any given chemical that enters the body but also the rate at which it
is subsequently chemically altered and excreted from the body (Plant 2003). These
four processes are outlined in Fig. 4.1 and detailed in the following sections.

4.1.1.1 Absorption

The body is essentially a series of aqueous environments (cytoplasm) bounded by

lipid (membranes). Hence, to enter, and subsequently distribute, around the body,

N. Plant ()
Centre for Toxicology, Faculty of Health and Medical Sciences, University of Surrey,
Guildford, Surrey GU2 7XH, UK
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 59

Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 4,
c Springer Science+Business Media, LLC 2011
60 N. Plant

Inhalation

Oral

Fig. 4.1 Outline of ADME processes within the body. From the external of the body a chemical
may undergo a number of processes before its final excretion to the external environment. Arrows
indicate potential chemical movement, with thickness indicating propensity for a particular route to
occur. A Absorption, D Distribution, M Metabolism, E Excretion, Cn Other compartments within
the body, GIT Gastrointestinal tract

chemicals must be able to cross lipid membranes. Due to this, those chemicals that
are best absorbed tend to be lipophilic in nature, meaning that they can cross these
membrane barriers by simple passive diffusion. Such a process follows simple first-
order kinetics1 and can be therefore modeled, at the simplest level, by simple mass
action.2 However, for those chemicals that are too hydrophilic to be efficiently ab-
sorbed across lipid membranes two potential entry routes exist. First, the use of
protein “tunnels”, through which these chemicals can diffuse, while having their
hydrophilic nature shielded from the lipid-based membrane. Second, energy can be
expended to move chemicals against a concentration gradient; this energy can be
either in the direct form of hydrolysis of ATP to ADP or indirectly using cotrans-
port or antitransport of chemicals down gradients, thus providing a motive force.
Transport processes that utilize a transmembrane protein during their functioning,
whether it is for facilitated diffusion or active transport, are many and varied in

1
First-order kinetics is defined as any kinetic reaction where the rate of the reaction is not limited
by the level of catalyzing protein, i.e., the protein is not saturated.
2
Mass action refers to the determination of reaction rate as the product of substrate concentration
and the rate constant for the reaction (v D ŒS  k).
4 Modeling Transport Processes 61

nature. These processes act as both influx and efflux transporters, but all essentially
can be defined as saturable processes exhibiting both first- and zero-order kinetics3
(Plant 2003).

4.1.1.2 Distribution

The process of chemical distribution around the body is essentially driven by the
blood flow to any given tissue, although it can be modulated by a number of other
factors. First, nonspecific binding of chemicals to plasma proteins retains them in the
systemic blood supply, delaying both their distribution to organs and their excretion
from the body (Gibson and Skett 2001): Depending upon the nature of this non-
specific binding it can be described as either an equilibrium constant (Keq ), being
the ratio of the forward and reverse rate constants, or have terms describing the sat-
urable nature of the binding. Second, as has been described for the absorption phase
of ADME, the presence of transport proteins in a given tissue can regulate chem-
ical distribution to that tissue, either causing selective uptake or exclusion (Plant
2003). This selective distribution to tissues is a central mechanism used to protect
organs susceptible to toxic damage, such as the brain, and to concentrate chemicals
in metabolically active organs such as the liver to enable their further processing
(Ayrton and Morgan 2001).

4.1.1.3 Metabolism

Alteration of a chemical structure has two major effects: First, as the routes of ex-
cretion are water-based it is often necessary to alter lipophilic chemicals, which
are well absorbed, into more hydrophilic chemicals that can be efficiently excreted.
Second, the act of metabolism alters a chemical structure, often changing its chem-
ical reactivity. Hence, a potential protective mechanism is the rapid deactivation of
chemicals before they can cause toxic damage to a tissue.
Metabolism is classically divided into two phases, catalyzed by distinct groups of
enzymes. Phase I metabolism is predominantly undertaken by the Cytochrome P450
super family of enzymes, while Phase II metabolism is shared by several, smaller,
enzyme families, each of which acts to add a specific chemical moiety to a substrate.
In both phases, the enzymes show marked substrate promiscuity, producing an adap-
tive system that can respond to a wide variety of chemical exposures. In general,
Phase I reactions act to make substrates more chemically reactive (bioactivation),
producing reactive centers for the deactivating reactions catalyzed by Phase II en-
zymes (Gibson and Skett 2001). It should be noted that Phase I and II do not have to
occur sequentially, for if a reactive center already exists within a substrate then this
can negate the requirement for Phase I metabolism, as is seen with acetaminophen

3
Zero-order kinetics is defined as any kinetic reaction where the rate of the reaction is limited by
the level of catalyzing protein, i.e., where the reaction is saturated.
62 N. Plant

(Plant 2003). Both Phase I and Phase II reactions are simple enzyme-catalyzed
reactions and can thus be described using suitable, saturable, kinetic parameters,
such as the Michaelis–Menten equation and its derivatives.

4.1.1.4 Excretion

The final process in the life-cycle of a chemical within the body is its excretion from
the organism. This process predominantly occurs via the kidneys, producing urine,
and liver, producing feces, although other, minor, routes do exist (Gibson and Skett
2001). The process of excretion occurs in a fundamentally similar manner to ab-
sorption, essentially being the reverse, and hence can be modeled as a combination
of passive and active processes.

4.1.2 Chemical and Pathophysiological-Mediated Alterations

in Drug Disposition

One central feature of biological systems is their ability to adapt to the environment
that they exist within. This feature is especially important within ADME processes
as it allows the body to efficiently respond to fluctuations within the environment,
thus bringing the body back to homeostasis. To understand the role of this adap-
tation within biological systems it is important to understand both the concept of
robustness in general and with specific relation to ADME, and these two areas will
be covered below. For the purpose of this text, the term network will be used to
describe the interactions of a number of chemical and biological species, including
drugs, proteins, and genes, within the cellular environment.
Within biological systems there is a requirement for robustness, defined as the
ability of that biological system to continue to carry out its fundamental tasks
(Kitano 2004). Perfect robustness within a network is unlikely to occur unless the
network fulfills some very specific criteria (Shinar et al. 2007), and indeed such per-
fect robustness is probably relatively rare in biological systems. However, the use of
feedback mechanisms allows systems to adapt to alterations in environmental condi-
tions, thus achieving quasirobustness. Within ADME the requirement for robustness
is obvious, as chemical exposure is, by definition, a move away from homeostasis;
without biological robustness body functioning would rapidly breakdown. In addi-
tion, it should be noted that ADME processes also control the fate of endogenous
chemicals within the body; any alteration in ADME processes caused in response
to external chemicals will potentially impact upon these endogenous processes and
this can result in toxicity (Plant 2004). In essence, without robustness the ADME
system that is central to body function would not be able to respond effectively
to chemical exposure from the environment, while still maintaining its endoge-
nous biological roles. This robustness is achieved through three, interconnected
systems. First, promiscuity within the ligand specificity for transporters and drug
4 Modeling Transport Processes 63

metabolizing enzymes ensures that for any given chemical, whether it is endogenous
or exogenous in origin, several complimentary systems can mediate the efficient re-
moval of the stimulating chemical (Plant 2004, 2007; Watkins et al. 2001; Smirlis
et al. 2001). Second, these ADME pathways are subject to complex feedback and
feed-forward loops, coordinating protein expression levels with their requirement to
handle stimulating chemicals (Plant 2007; Pascussi et al. 2000a, b; Aouabdi et al.
2006). Third, activation of ADME pathways is closely coupled with the activation
of protective mechanisms that remove toxic damage elicited by stimulating chemi-
cals should it occur before the chemical can be safely removed (Plant 2003; Roberts
et al. 1997).
In addition to alterations in ADME elicited by chemical exposure, there is also
a well-described alteration in microenvironment caused by pathophysiology, with
the most obvious example being during tumorigenesis. The alterations in gene and
protein expression in tumors have been well described, initially at the level of the
tumor in general (Kim et al. 2009; Sotiriou and Pusztai 2009), providing a frame-
work to both aid in the grading of tumors (Ramaswamy et al. 2001) and predict the
response to therapeutic intervention (Villeneuve et al. 2006). Recent work has fo-
cused on the heterogeneity of expression within single cells within a tumor (Slack
et al. 2008), leading to the potential to model tumor responses at the single cell level.
One concept that is now clear is that a central feature of tumor cells is their altered
ADME processes, leading to altered handling of chemicals compared to normal tis-
sue, which may result in altered therapeutic efficacy. It is vital to understand this
variability in order to be able to optimize therapeutic intervention, either through
the development of novel therapeutics, potentially targeted at fragile nodes within
a network, the modulation of response networks through coadministration of mul-
tiple therapeutics or the optimization of current therapeutic strategies through the
increased understanding of these response networks.

4.1.3 Extrapolation of Data Between Biological Scenarios

For the safe development of novel chemical entities, the optimal usage of existing
therapeutics, and the robust risk assessment of human exposure to potential toxi-
cants, it is important to be able to extrapolate data from one scenario to another. Such
extrapolation may be between normal and pathophysiological conditions, from indi-
vidual to population exposures, including the impact of genetic variability, or from
model species/test systems to the in vivo human situation. Making such decisions
earlier in new chemical entity discovery/development is central to ensuring that the
most likely leads are progressed down the development pipeline, thus enhancing
the possibility of them making it to market. Such decisions are obviously depen-
dent on the ability to make robust extrapolations, and as such the development of
novel in silico models may represent an important driver in the continued optimiza-
tion/understanding of human response/safety to drug treatment. In addition, creation
of such models will provide important insights into the normal physiological pro-
cesses that occur in humans.
64 N. Plant

Given that the chemical industry, both pharmaceutical and industrial, have been
undertaking extrapolation of data between scenarios for many decades, how then
can systems modeling help? To answer this question, it is necessary to understand
the current paradigm on which this extrapolation is based and thus see how this can
be improved and/or modified. The next section will cover the traditional approach
of pharmacokinetics, while Sect. 4.3 will detail a bottom-up systems approach to
ADME modeling.

4.2 Traditional Pharmacokinetic Approaches to Modeling

Drug Disposition

As detailed in the previous section, there is a need to be able to robustly model the
fate of chemicals within the human body. Such mathematical treatment of chem-
ical fate has been undertaken for many decades and is the basis of the science of
pharmacokinetics. In traditional pharmacokinetics a reductionist approach is taken,
whereby several biological reactions are described by only a single term (Tozer and
Rowland 2006). In fact, the underlying principal of traditional pharmacokinetics is
the generation of compartments that represent tissues with similar kinetic profiles,
as opposed to identifying all the individual reactions within each tissue.
Such an approach has the distinct advantage that models can be constructed from
relatively poor data sets, perhaps only containing values for input and/or output lev-
els of the chemical, plus some major transition states in between. Systems such as
SimCYP (Rostami-Hodjegan and Tucker 2007) are able to not only model chemical
fate within a system but also to allow manipulation of individual parameters such
that population responses can be studied. There is, however, an obvious problem
with a reductionist-heavy approach, which is that it may potentially limit the abil-
ity to examine fine detail within the network. For example, the role of individual
parameters, such as the concentration of a single protein, cannot always be studied
as the models are based at a detail level above this. Whereas this may not neces-
sarily be an issue for the study of population-based kinetics, it can have potential
issues under two circumstances. First, extrapolation of data from one species to an-
other may be seriously skewed if there is not a clear relationship between the model
species. For example, many ADME proteins show species-specific characteristics,
with orthologues often showing distinct expression levels and kinetic parameters.
In addition, altered ADME between species also exists, with, for example, mice
and hamsters being highly sensitive to the toxic effects of acetaminophen whereas
rabbits and guinea pigs are relatively insensitive (Bessems and Vermeulen 2001).
Second, as discussed in Sect. 4.1.3 there are known differences between normo-
and patho-physiological tissue states, and hence it is important to understand how
the alteration of these response networks impacts upon biological functioning and
response to therapeutic intervention.
4 Modeling Transport Processes 65

4.3 More Complex Models of Drug Movement Across

Biological Membranes

4.3.1 The Measurement of Chemical Movement Across

Biological Membranes

As described in Sect. 4.1, any chemical that moves across a lipid membrane will do
so through one of two major routes, direct movement through the lipid bilayer or
transfer through an intermediate protein structure (a drug transporter) (Fig. 4.2).
It should be noted that, in reality, the majority of chemicals are transported by
a mixture of passive and active transport processes, and hence both modes will
need to be modeled for any given chemical. In addition, both of these modes of

out

ATP ADP

in
Passive Diffusion Facilitated Transport Active Transport

Fig. 4.2 Major potential modes of transport of chemicals across a biological membrane. Nonpolar
(lipophilic) drugs cross membranes by simple passive diffusion down their concentration gradients.
However, as hydrophilicity increases, protein (T ) is required to act as a pore for the drug, masking
polar groups that would impede transport (facilitated transport). Finally, for transport lacking a
concentration gradient, energy must be supplied, often via the hydrolysis of ATP to ADP (active
transport)
66 N. Plant

transport can occur in either direction across the membrane, although in the case of
active transport this is dependent upon the involvement of separate influx and efflux
transporters.
On the surface it would thus seem relatively simple to model these two modes
of transport, with passive and facilitated diffusion being modeled by simple mass
action kinetics (4.1), whereas active transport can be modeled as an irreversible
Michaelis–Menten style equation (4.2), with the inclusion of a Hill slope if neces-
sary to allow for allosteric interactions (4.3).

ıAout
D k  ŒAout   k  ŒAin  ; (4.1)
ıt
ıAout Vmax  ŒT   ŒAout 
D ; (4.2)
ıt Km C ŒAout 
ıAout Vmax  ŒT   ŒAout h
D ; (4.3)
ıt Km C ŒAout 

where

Ain and Aout are the concentrations of drug in donor and acceptor compartments (M);
k is the rate constant for drug movement, Vmax is the maximal rate of drug transport
across membrane (moles/min/mg protein), Km is the drug concentration required to
achieve a transport rate equal to 1=2 Vmax (M), h is the hill slope, t is the time in
minutes, and ŒT  is the concentration of drug transporter (M).
However, as will be discussed in the following section, such simplistic equations
do not fully encompass the complexity of membrane transport, with refinements
needed to increase the accuracy of modeling. In addition, it is worth noting that
the experimental system used to gain kinetic data can have a major impact on the
in vivo relevance of this data; it is hence worthwhile first addressing the general
assumptions often made during kinetic modeling of transport processes.

4.3.2 General Considerations for Measuring Movement

of Drugs Across Biological Membranes

It is an obvious statement that when kinetic data is derived from in vitro systems for
use in in silico models, it is important that the data accurately reflects the biological
model that is being simulated. However, whereas in vitro systems provide good
surrogates for in vivo, it is important to note the potential differences between the
two scenarios as this may help explain why derived models do not achieve full
predictivity. These differences are well reviewed by Youdim et al. (2003), with a
brief description given below.
4 Modeling Transport Processes 67

4.3.2.1 Simple Versus Complex Measurement Systems

As will be detailed within the next section, transport across biological membranes
can be essentially divided into the passive and active transport components of the
total transport. As such a number of different systems have been developed to mea-
sure these two compartments, including in silico assessment of logP4 or logD5
(Sawada et al. 1999); use of preformed lipid vesicles (Zhou et al. 2009); cell lines
expressing individual transport proteins (Acharya et al. 2006); individual cell lines
(Walle and Walle 1998); coculture systems (Perriere et al. 2007); and in vivo mea-
surements (Vlaming et al. 2009). As these systems range from the highly simplistic,
but easy to handle, to complete in vivo systems where many confounding factors to
accurate measurement exist, then it is important to consider the desired outcomes
before a test system is chosen. The choice of system will impact on not only the
general assumptions outlined below but also the level of detail generated within the
study and the manner in which this needs to be treated.

4.3.2.2 Ionization Status of the Drug

A central role of drug transport processes is the regulation of chemical access from
the environment; as such a major site of drug transport is the gastro-intestinal tract
(GIT). The GIT is a complex tissue, with many functional subdivisions along its
length, essentially producing microenvironments that need to be considered sepa-
rately. For example, there is considerable variance in pH along the length of the tract,
with values varying from alkaline to acidic. In accordance with the Henderson–
Hasselbalch equation (4.4), the ionization state of a chemical is dependent on the
pH of the local environment. Such variation may have a considerable impact on
passive diffusion of chemicals given that the pH partition hypothesis states that only
the nonpolar form of a chemical will cross a lipid membrane by passive diffusion.
The impact of local pH on the permeability of a membrane to a chemical is often
referred to as the passive permeability P .

logŒA C HC 
pH D pKa C ; (4.4)
ŒAH

where

A is the polar drug form concentration (M), HC is the proton concentration (M),
AH is the nonpolar drug form concentration (M), and pKa is log10 (Dissociation
constant for AH).

4
LogP is defined as the log 10 of the Œchemical in lipid (usually octanol)=Œchemical in water at a
specified pH.
5
LogP is defined as the log 10 of the Œchemical in lipid (usually octanol)=Œchemical in water at
pH 7.4.
68 N. Plant

It is important to note that it is the pH of the microenvironment bordering the

membrane that is important, which may not necessarily relate to the pH of the en-
tire lumen. For example, whereas the pH of the stomach is highly acidic (pH  2),
the microenvironment of the membrane tends toward neutral, on average pH 8
(Rechkemmer 1991). This variation in pH can have a large impact on P , as demon-
strated by Palm et al. (1999) who examined the effect of altering the degree of
ionization of alfentanil and cimetidine from 5 to 95% on the passive diffusion of
these chemicals. They demonstrated that transport of the nonpolar form was 150- to
30-fold more rapid than the ionized drug, respectively. In addition, when the frac-
tion of nonpolar drug (fu) was less than 0.1 the contribution of ionization status to
determining P became significant. If one considers the pH values observed in the
gut membrane microenvironment, then fu was consistently less than 0.1, meaning
that ionization status is likely to have a significant impact in the biological systems.
Indeed, Palm et al. (1999) demonstrated that pH variability within the physiological
range altered P by as much as 2.5-fold.

4.3.2.3 Heterogeneity in Drug Dispersion

It is generally presumed that chemical is homogenously distributed throughout the

liquid phase; however, this is certainly not the case close to biological membranes,
as the area either side of the membrane forms an “unstirred water layer” (UWL),
through which the chemical must diffuse prior to entering the membrane. The size
of the UWL can be altered through vigorous stirring of the medium in cell culture
systems, but never totally removed, and in the case of transport assays the UWL may
exceed 1 mm; in comparison, in vivo the UWL is estimated to be 30–100 m thick
within the GIT (Lennernas 2007). Given that most transport assays do not routinely
utilize medium agitation then it can be presumed that the UWL in these situations is
larger than expected in vivo. Such an experimental set-up will, of course, reduce the
robustness of any model based upon in vitro data for predicting in vivo scenarios.

4.3.2.4 Chemical Sequestration

The majority of in vitro techniques for measuring chemical movement across

biological membranes function by measuring the amount of chemical in either com-
partment or in both. However, it should be noted that the ability to cross a membrane
is determined by the free-fraction of chemical, with chemical being potentially se-
questered in to either compartment or even the membrane itself. Nonspecific binding
to extra- or intracellular proteins can be easily measured and incorporated into the
diffusion equation as an extra term. In addition, as these bindings are reversible
then it should be remembered that the total amount of chemical bound will alter
with time as free chemical diffuses from one compartment to the next. With regard
to the extrapolation from in silico models to in vivo then intracellular binding is
assumed to be roughly equal for any given cell type and chemical. However, the
4 Modeling Transport Processes 69

extracellular binding will be determined by the level of binding proteins present in

the serum and medium, and data should be corrected to compensate for inequalities
in this. One easy solution is to ensure the use of cell culture medium that contains
plasma-binding proteins such as albumin, mimicking sequestration in the systemic
circulation.
A more critical component is that most chemicals subject to significant passive
diffusion will, by definition, be lipophilic in nature. As such, there is the potential
for sequestration within the membrane itself. The degree of sequestration is depen-
dent on both the lipophilicity of the chemicals and also the availability of acceptor
chemicals within the receiving volume, which are necessary for the extraction of
the chemical from the membrane. Experimental estimates for the impact of this
retention vary; in Caco-2 cells Wils et al. (1994) estimated retention at only approx-
imately 40% for chemicals with LogD values in the range 3.5–5.2, whereas Sawada
et al. (1999) observed up to 89% retention for highly lipophilic chemicals (logP
1.1–15) being tested in MDCK cells.

4.3.2.5 Physico-Chemical Characteristics of the Chemical

It is generally assumed that within the experimental system chemical characteristics

remain constant throughout the time of the assay. Indeed, this is often a determined
prerequisite for “probe chemicals” used within in vitro test systems (i.e., those
chemicals determined to be satisfactory markers for an individual biological event).
There are actually three assumptions made within this larger assumption: First,
that the chemical is chemically stable over the time of the assay; second, that the
chemical is metabolically stable over the time of the assay; third, that the chemical
maintains the same physico-chemical characteristics over the concentration range
tested within the assay. In general, the first assumption should be relatively robust
unless the assay is undertaken over an extremely extended period of time or the ini-
tial choice of probe chemical was poor. The second assumption is not always valid,
and with the majority of chemicals likely to show some metabolism during the pe-
riod of the assay. Such loss can be dealt with by either measuring parameters that
are resistant to such loss, such as the depletion of chemical from medium where
no metabolic activity is present, or through the incorporation of a term to measure
the rate of loss via metabolism. As an aside, it should be noted that many cell lines
actually have markedly reduced metabolic capacities, compared to in vivo, which
is often considered a limiting factor in their utility, as they are unable to mimic the
metabolism seen in vivo; however, in the case of maintaining probe chemical in-
tegrity this is actually an advantage, as chemical loss through metabolism will be
significantly reduced.
The final assumption often made is that the chemical reacts with its environment
in an identical manner at all concentration used. However, at higher chemical con-
centrations most chemicals will pass a “critical micellular concentration” (CMC),
above which they form micelle structures that have an increased passive perme-
ability through the membrane. Hence, measurements taken above the CMC will be
70 N. Plant

unreliable as the ratio of passive to active transport will be substantially altered. The
solution to this is simple; identify the CMC for the probe chemical before testing
transport rates and do not exceed it within the test system.

4.3.2.6 ATP Usage Within the Test System

Many active transport processes require the hydrolysis of ATP to provide the en-
ergy for transport. Hence, it could be envisaged that ATP levels and kinetics may
influence the overall transport efficiency. In general, however, the input of ATP
into active transport processes is not modeled, with it presumed to be nonlimiting.
Accurate measurements of the kinetics for ATP binding and hydrolysis have yet to
be determined and hence, at present, this assumption must stand. Indeed the avail-
able data suggests that transport data can be fitted without modeling this component,
suggesting that it has a minimal impact on the overall flux.

4.3.3 Measurement of Passive Diffusion

Derivation of a logP or logD for the passive permeability of a chemical is rela-

tively trivial and can be carried out in a simple cell-free system (Wils et al. 1994).
However, it should be noted that this simplistic equation presumes that there is a
nonlimiting surface area through which the chemical can diffuse, and that the perme-
ability of a lipid bilayer remains constant. If either of these assumptions is breached
then it is necessary to include further terms to account for these parameters, derived
from Fick’s first and second laws of diffusion (Fick 1855, (4.5) and (4.6)).

ıø
J D D ; (4.5)
ıx
ıø ı2ø
DD 2 ; (4.6)
ıt ıx
where

J is the the flux through the membrane (Mcm2 s1 ), D is the diffusion coefficient
(cm2  s1 ), ø is the concentration (M), and x is the position (cm).
For the purpose of general modeling of chemical flux, Fick’s laws may be
expressed as shown in (4.7), which can be rewritten as the ordinary differential
equation shown in (4.8).

J D P  ø ; (4.7)
ıAout P A
D ; (4.8)
ıt Vin .Aout  Ain /
4 Modeling Transport Processes 71

where

P is the permeability of the membrane (cm  s1 ), A is the surface area of the mem-
brane (cm2 ), Vin is the volume of acceptor (cm3 ), Ain/out is the concentration of drug
in donor and acceptor compartments (M). It should be noted that the units for the
term PA=V cancel out to produce s1 , which is consistent with this term represent-
ing the rate constant for passive diffusion.
Correction for Multiple Sampling. One potential issue with any system used to pro-
duce kinetic measurements is the requirement for multiple measurements to be taken
over time. This may result in a decrease in the volume of, for example, the acceptor
compartment as samples are taken, and it can be envisaged that this could be-
come significant if extended time or large sampling volumes are used. Two possible
solutions exist for this problem: First, the use of chemicals such as carboxydichlo-
rofluorescein, which can be measured due to their fluorescence, and hence do not
require a reduction in the sampling volume (Howe et al. 2009). Second, Tran et al.
(2004) suggested the inclusion of an additional term into (4.8) to allow automatic
correction for multiple sampling.
Measurement of Passive Diffusion at Different Temperatures. Previously, it has
been assumed that passive diffusion was a temperature-independent phenomenon.
This assumption allows measurement of passive diffusion parameters to be under-
taken at 4ı C, when active transport processes do not occur, thus providing an easy
means to separate passive and active transport components for any given chemical.
However, it is becoming increasingly clear that membrane fluidity, and hence per-
meability, are altered by temperature (Ulrih et al. 2007), and hence measurement of
passive diffusion at 4ı C can potentially be artefactual. For example, Poirier et al.
(2008) demonstrated that the apparent permeability (Papp) of chemicals could be
considerably lower at 4ı C compared to 37ı C, with the Papp for fexofenadine being
16-fold lower when determined at 4ı C compared to 37ı C.
If it is necessary to assess passive diffusion at 37ı C in order to achieve accurate
kinetic values then it is important to be able to distinguish active and passive com-
ponents of transport, both of which can occur at this temperature. One approach is
to undertake the assessment of passive diffusion in artificial membranes, such as
the parallel artificial membrane permeability assay (PAMPA), or through the use of
lipid vesicles, both of which do not contain any active transport proteins. The ad-
vantage of this system is the specificity of the measurement gained, but this comes
at the cost of the use of a more artificial system that may breach some of the general
assumptions outlined in Sect. 4.3.2. An alternative approach is to undertake assays
in cell lines, thus increasing the closeness to in vivo, but with the use of chemical
inhibitors to prevent active transport. The obvious disadvantage of such a system is
that one must know something about the active transport of a chemical before one
knows which inhibitors to use. Fortunately, there are now a number of broad speci-
ficity inhibitors, such as quinidine and verapamil, which allow one to inhibit large
families of transport proteins at one time (Tan et al. 2000).
72 N. Plant

4.3.4 Measurement of Active Transport

The measurement of transporter-mediated movement across biological membranes

is potentially much more complicated than the measurement of the passive diffusion
component. Numerous test systems exist to examine active transport processes in
cells. These can range from whole cells assays undertaken in TranswellsTM , which
measure the total active transport kinetics rather than any single transport protein
(Buesen et al. 2002), through over-expressing cell lines (Hopper-Borge et al. 2004)
to (inside out) lipid-vesicles (Glavinas et al. 2008), the latter two of which allow the
measurement of kinetics for a single transport protein. An advantage of membrane
vesicles is that they can be engineered so that they are either orientated “normally”
or “inside out”, meaning that both influx and efflux proteins can be studied with rel-
ative ease (Glavinas et al. 2008). Such an approach negates the need to preload cells
with chemical for the study of efflux transport, which is an additional complication
that can require optimization.
Regardless of the system of study used, it is usual to fit chemical kinetics to
the Michaelis–Menten (4.2), which has certain limitations. The Michaelis–Menten
analysis was derived from the analysis of soluble enzyme kinetics and hence may
not be suitable for examining the kinetics of membrane-bound transporters. The
largest difference between these two scenarios is the location of chemical prior to
interaction with the enzyme or transporter. In the case of soluble enzymes the chem-
ical will generally be in the aqueous phase (i.e., cytoplasm) and hence can directly
interact with the enzyme. In the case of transporters, however, the chemical may be
present within either the cytoplasm, as occurs with glucose transport, for example,
or embedded within the lipid membrane if it is sufficiently lipophilic, as is the case
for many drug molecules. In both cases diffusion through the cellular environment
will impact upon the rate of association with transporters, but it is logical to hy-
pothesize that this would be of greater impact for diffusion within the membrane
compared to the cytoplasm. One solution is to include a scaling factor, similar to
the Hill function within the transport equation (4.3). This scaling factor does not
necessarily represent multiple binding sites or cooperativity within an individual
protein, as is the case for soluble enzymes, but represents the multiple steps that
interact to provide the total measured kinetics; this scaling factor is often referred
to as the ˇ-coefficient. This solution works well, but has the disadvantage that the
all-encompassing scaling factor hides many of the subtleties involved in transport,
which may be critical in understanding the true biological mechanisms occurring
within the transport process.
Bentz et al. (2005) have attempted to deconvolute the individual steps in drug
transport, measuring the individual kinetic parameters shown in Fig. 4.3 (Tran et al.
2005), and fitting these into a model of chemical transport.
As can be seen from Fig. 4.3, rate constants were derived for each of the stages
involved in drug transport. These included association of the chemical into the lipid
membrane from the aqueous solution, which was chemical specific and highly vari-
able (Kpc); diffusion through the lipid membrane, allowing association with the
transporter, which appears to be both fast and nonchemical specific (k1 =kr ); binding
4 Modeling Transport Processes 73

out

kB
Kpc

k1+kr

k2
ATP ADP ATP ADP

in
Simplistic Active Complex Active
Transport Transport

Fig. 4.3 Simple and complex determinations of active transport kinetics. In the simple mode, the
increase in intracellular drug concentration ŒAin ] is determined by the external drug concentration
ŒAout , transporter concentration ŒT , and efficiency of the transporter toward the drug (Km and
Vmax ). In comparison, the complex model takes into account these three factors plus association
rate of the drug into the lipid membrane (Kpc ), diffusion rate within the membrane (k1 and kr ),
association rate of the drug with the transporter (kB ) and disassociation rate of the drug from the
transporter on the far side of the membrane (k2 )

of the chemical to the transport channel of the transport protein, which is chemical
specific (KB ); and the efflux of the chemical to the other side of the membrane,
again chemical specific (k2 ). So, does this increased detail within the transport
model translate to improved biological knowledge and predictability? Bentz et al.
(2005) examined this by simulating the transport process using physiologically rel-
evant parameter ranges. They were able to demonstrate that modeling of Vmax was
relatively robust, with only a two- to three-fold variance between the modeled and
experimental Vmax values. It should be noted, however, that this conclusion was only
valid when the efflux velocity measured was significantly lower than the ŒAout ,
meaning that V =ŒAout  tended toward zero. Such a scenario was necessary to en-
sure that the passive diffusion component of chemical transport did not become too
large and equated to a transporter occupancy rate of over 80%. This is of interest
because such a rate of occupancy may not always be achieved under experimental
74 N. Plant

conditions. First, systems using over expression plasmids for the transport protein
under examination may have many more copies of the protein expressed than is
seen in vivo and thus achieving the ŒAout  may be difficult. Second, as stated in the
general assumptions for modeling (Sect. 4.3.2), as chemical concentration increases
it is probable that the CMC will be breached and that the dynamics of chemical
interactions with the transporter altered, effectively increasing the input of passive
diffusion into the model. As will be discussed in Sect. 4.4.1, one solution to this
issue is to create separate terms for both the passive and active components of the
transport, which can then be modulated independently.
Whereas Vmax could be reasonably well modeled over the physiological range
using simple equations such as (4.2) and (4.3), the derivation of the overall Km was
not as simple. This is perhaps not surprising given that several of the rate constants
impact upon the Km . Bentz et al. (2005) demonstrated that derivation of biological
Km values from models was subject to considerable inaccuracy, which increased
as Km did. This latter point is especially important when considering the transport
of drug-like molecules, as they are routinely transported by promiscuous transport
proteins with high, micromolar, Km values. Bentz et al. (2005) suggested that the
biological Km could be modeled from the elementary rate constants as shown in
(4.9), recommending that elementary rate constants provide a better estimate of Km
than fitting with the simple Michaelis–Menten equation.

k2 C kr
Kmbiol D ; (4.9)
k1 C Kpc

where

Kmbiol is the “biologically relevant” Km (M); k2 is the disassociation rate constant

for drug from transporter; kr and k1 are the diffusion rate constants for drug within
the membrane, often equal; and Kpc is the association rate constant for drug and
lipid membrane.

4.4 The Integration of Drug Disposition and Drug Fate

into a Predictive Model of the Life Cycle of a Drug
in the Body

The understanding of the kinetics of drug entry into cells is an area that has re-
ceived much attention in the past few years, with an ever increasing understanding
of both the importance and complexity of this process becoming clear (Ayrton and
Morgan 2001). However, this biological understanding is really only the first step-
ping stone to understanding larger biological questions, such as the life cycle of
chemicals within the body. Any chemical, whether it is endogenous or exogenous in
origin, is subject to the processes of ADME, which together determine the longevity
and action of the chemical within the body. As described in Sect. 4.2, the science of
4 Modeling Transport Processes 75

pharmacokinetics has developed specifically to allow these processes to be modeled,

with the ultimate aim of allowing accurate extrapolation and prediction of biological
responses in humans, as detailed in Sect. 4.1.3. One approach for the incorporation
of membrane transport into larger models is the integration of multiple processes
into single formula to cover all aspects of transport. However, as the amount of
information on the mechanics of transports becomes clear then these equations be-
come more complicated and cumbersome.
Whereas the integration of multiple processes into single equations is valid, as
the complexity of the derived equations increases they rapidly become impenetrable
to all but experts, thus diminishing their utility by the general scientific community.
An alternate approach is the use of modeling software, which is similar to path-
way mapping software originally designed to understand DNA microarray transcript
data, providing a user-friendly front-end for complex simulations. Programs such as
CellDesigner6 (Funahashi et al. 2003) provide an easy to use graphical front-end that
allows nonspecialist entry of network maps in Systems Biology Graphical Notation7
(SBGN), an established format that is widely utilized. In addition, whereas the
model is created by a simple point-and-click interface it is encoded by the Systems
Biology Markup Language8 (SBML; Hucka et al. 2003), a universal XML format,
which allows the model to be easily transferred between modeling software. To pro-
vide an example of the utility of such an approach we will examine the modeling of
multiple drug resistance phenotype in the treatment of breast cancer.

4.4.1 Multiple Drug Resistance Phenotype in Cancer Treatment

More than one million women are diagnosed with breast cancer every year, which
represents approximately one quarter of all new cancers in women: Such a rate of
diagnosis represents a lifetime risk of developing breast cancer of one in eight for
women born in the USA, and one in nine for women in the United Kingdom (Coley
2008). Fortunately, there are a number of established therapies for metastatic breast
cancer (MBC), ranging from endocrine-based therapies for hormone-receptor posi-
tive tumors, through anthracyclines and taxanes to the recent development of novel
biologics such as trastuzumab (Herceptin). Despite this range of therapeutics, the
response rate to first-line chemotherapies such as anthracyclines and taxanes is sub-
optimal, being reported as between 30 and 70%, falling to 20–30% for subsequent
treatments with a median duration of response of 6 months. One area that often
limits chronic therapeutic treatment, including anti-cancer chemotherapy, is the de-
velopment of multiple drug resistance (MDR) phenotype, and it has been estimated
that MDR is involved in over 90% of treatment failures for MBC.

6
http://www.celldesigner.org.
7
http://sbgn.org/Main Page.
8
http://sbml.org/Main Page.
76 N. Plant

out in

Fig. 4.4 Mode of action of Taxane-family anticancer drugs. Taxanes can be taken into cells by
either passive diffusion (D) or active transport (ATin ). Once in the cell they can either bind non-
specifically, or specifically to microtubules. In the latter case this elicits a chain of events that leads
to cell death, killing the tumor cell. Taxanes can also be metabolized in the cell by several different
enzymes, and these metabolites removed from the cells by passive diffusion (D) or active transport
(ATout ). Finally, these products can be removed from the body by excretion (E)

The taxane class of anticancer drugs are a front-line treatment for metastatic
breast cancer, with paclitaxel (Taxol) being the class leader (McGrogan et al.
2008). Their mode of action involves high affinity binding to microtubules, cel-
lular structures obligate for biological functioning (Fig. 4.4). Taxol binding prevent
microtubules from growing and shortening, which is essential for their functioning,
and this causes the cell to undergo a series of steps that results, ultimately, in cell
death (Xiao et al. 2006). Prolonged treatment with taxol often results in the pro-
duction of MDR, preventing the drug effectively killing cells. Several mechanisms
have been identified that will contribute toward the development of MDR; increased
metabolism, increased export, altered interaction with target proteins and altered
biological response to that interaction. However, it is not clear how each of these
factors input into the development of MDR, and hence what is the best strategy to
prevent its development: Such a problem is ideally suited to examination using a
systems modeling approach.
Based on the cartoon shown in Fig. 4.4, a comprehensive model of the network
of interactions of taxol with tumor cells was created in CellDesigner. All interac-
tions within the network were derived from published literature, producing a model
with 123 species and 72 reactions. Underlying the graphical front-end, each reac-
tion can be associated with a kinetic law, producing a set of ordinary differential
equations (ODE) to describe the network. By dividing the interaction network into
a series of ODEs it is possible to build the network model from a series of inter-
connected modules; in the example shown, for example, individual modules might
4 Modeling Transport Processes 77

represent cellular influx, metabolism, sequestration, efflux, binding to microtubules

and triggering of programmed cell death. Such an approach is ideal as it reduces the
knowledge base required, both biochemical and mathematical, making the approach
available to a larger number of scientists. In addition, as each reaction is defined by
simple ODEs then as new information becomes available it is easy to replace one,
or a number, or reactions with the new details, thus improving the model.
Once such models have been generated, the next step is to undertake simulations
of the entire network. Such analysis can be undertaken directly in CellDesigner, but
is limited to relatively simple time simulations and parameter scans. A more flex-
ible alternative is to export the model into more powerful analysis software such
as COPASI9 (Hoops et al. 2006), and this is where the use of the universal SBML
format makes such transitions simple. Once in COPASI further analysis options are
available, including steady-state analysis, parameter estimation, network optimiza-
tion, and sensitivity analysis; in addition, both deterministic and stochastic modeling
is available, an important requirement for examination of single cell events. More
experienced modelers may wish to program directly into COPASI, which uses a
tabular format to define reactions, species, etc.; however, the lack of a graphic repre-
senting the overall network can be confusing for inexperienced modelers. Therefore,
it is often easier to build outline models in programs such as CellDesigner due to
their simple model entry system, before porting the model to a comprehensive sim-
ulation tool such as COPASI.
Having transferred the taxol model into COPASI, a first analysis over a time
series shows that, after taxol addition, there is a rapid decrease in the amount of
free microtubules and concomitant increase in taxol-bound microtubules (Fig. 4.5a).
The net effect of this is a decrease in microtubule dynamism, and as can be seen
from panel b, microtubule dynamism is directly related to the rate of apoptosis,

Fig. 4.5 Time series simulation of Taxol in MCF7 breast cancer cells. A model of the interactions
of taxol in MCF7 breast cancer cells was created in CellDesigner, transferred to COPASI and the
network simulate over 200 s, demonstrating a decrease in microtubule dynamism (MT; Panel a).
Decreases in microtubule dynamism can be linked to the rate of programmed cell death, apoptosis
(Panel b)

9
http://www.copasi.org.
78 N. Plant

Fig. 4.6 Time series simulation of Taxol in MCF7 breast cancer cells. A model of the interactions
of taxol in MCF7 breast cancer cells was created in CellDesigner, transferred to COPASI and the
network simulated over 200 s (line). These values were compared to measurements made in vitro
(closed circles) for excreted taxol (Panel a) and production of the 6˛-OH taxol metabolite (Panel b)

programmed cell death, in breast cancer cells (MCF7 cell line; Fig. 4.5b). However,
this desired, pharmacological action is offset by excretion of taxol from the cell by
drug transporters (excreted taxol), and the metabolism of taxol (6˛-OH and 3’-p-OH
taxol).
Having created the model, it is important to show that the simulated species levels
closely match those seen in the biological scenario. Whereas it is not possible to
easily measure all of the parameters, it is important to show that as many as possible
correlate well, as demonstrated in Fig. 4.6.
Following demonstration of correlation between the in silico and in vitro model
systems it is now possible to use the in silico model to examine the underlying de-
sign principles for the network. As an initial step in this process a sensitivity analysis
will reveal those nodes within the network most likely to impact upon MT dy-
namism. As depression of MT dynamism is central to the anticancer effects of taxol,
any node that significantly impacts upon this is a potential mechanism for the devel-
opment of multiple drug resistance phenotype. As can be seen from Fig. 4.7, such
an analysis highlights three factors (microtubule subunit composition, blood flow
to tissue and ABCB1 transporter expression) as being the probable major drivers
for development of multiple drug resistance phenotype. One of these species is the
drug transporter ABCB1, demonstrating the importance of membrane transport in
determining drug disposition and actions.
Having gained this information it is now possible to use molecular techniques to
specifically target these species for intense examination within the in vitro system.
For example, cell lines can be genetically modified to allow alteration of the species
in question, varying their expression in line with that observed in vivo. These mod-
ified cell lines can then be used to confirm the predictions of the in silico model,
as well as acting as test systems to examine how the development of MDR could
be prevented in the clinic. In addition, other biological questions can be answered
4 Modeling Transport Processes 79

Fig. 4.7 Sensitivity analysis. A model of the interactions of taxol in MCF7 breast cancer cells
was created in CellDesigner, and then transferred to COPASI for sensitivity analysis. Sensitivity
of species within the network were determined against their ability to alter the impact of taxol on
microtubule dynamism

surrounding drug treatment of cancer, such as the extrapolation of response to Taxol

in normal and tumor tissue, where protein levels will be different; to extrapolate
between preclinical and clinical scenarios; finally, novel or altered therapeutic treat-
ments can be simulated to assess their potential utility.

4.5 Summary

The cell membrane is an essential barrier for life. It acts not only to contain the cell
contents but also as a barrier to prevent free access of chemicals to the cell interior.
As such, understanding the dynamics of chemical transport across biological mem-
branes is vital to fully understand how the body will respond to chemical exposure.
Recent work has begun to characterize this, perhaps surprisingly complex, mecha-
nism, and is increasing our understanding of the determinants for chemical access
to cell interiors. The integration of these transport mechanisms into larger models
of general cellular response to chemical exposure is currently underway, producing
more refined models of the life cycle and biological impact of chemical exposure.
Such approaches will be vital for the extrapolation of data between different bi-
ological scenarios, such as understanding the differential response of normal and
tumor tissue and the development of multiple drug resistance phenotype, and lead
to improved treatment schedules with increased success rates.
80 N. Plant

References

P. Acharya, T. T. Tran, J. W. Polli, A. Ayrton, H. Ellens, and J. Bentz. P-glycoprotein (P-gp) ex-
pressed in a confluent monolayer of hMDR1-MDCKII cells has more than one efflux pathway
with cooperative binding sites. Biochemistry, 45(51):15505–15519, 2006
S. Aouabdi, G. G. Gibson, and N. Plant. Transcriptional regulation of PXR: Identification of a
PPRE within the proximal promoter responsible for fibrate-mediated transcriptional activation
of PXR. Drug Metabolism and Disposition, 34(1):138–144, 2006
A. Ayrton and P. Morgan. Role of transport proteins in drug absorption, distribution and excretion.
Xenobiotica, 31(8–9):469–497, 2001
J. Bentz, T. T. Tran, J. W. Polli, A. Ayrton, and H. Ellens. The steady-state Michaelis-Menten
analysis of P-glycoprotein mediated transport through a confluent cell monolayer cannot predict
the correct Michaelis constant Km. Pharmaceutical Research, 22(10):1667–1677, 2005
J. G. M. Bessems and N. P. E. Vermeulen. Paracetamol (acetaminophen)-induced toxicity: Molec-
ular and biochemical mechanism, analogues and protective approaches. Critical Reviews in
Toxicology, 31(1):55–138, 2001
R. Buesen, M. Mock, A. Seidel, J. Jacob, and A. Lampen. Interaction between metabolism and
transport of benzo[a]pyrene and its metabolites in enterocytes. Toxicology and Applied Phar-
macology, 183(3):168–178, 2002
H. M. Coley. Mechanisms and strategies to overcome chemotherapy resistance in metastatic breast
cancer. Cancer Treatment Reviews, 34(4):378–390, 2008
A. Fick. Ueber diffusion. Annalen der Physik, 170:59–86, 1855
A. Funahashi, N. Tanimura, M. Morohashi, and H. Kitano. CellDesigner: A process diagram editor
for gene-regulatory and biochemical networks. Biosilico, 1:159–162, 2003
G. G. Gibson and P. Skett. Introduction to drug metabolism. Nelson Thornes, Cheltenham, 3rd
edition, 2001
H. Glavinas, D. Mehn, M. Jani, B. Oosterhuis, K. Heredi-Szabo, and P. Krajcsi. Utilization of
membrane vesicle preparations to study drug-ABC transporter interactions. Expert Opinion in
Drug Metabolism and Toxicology, 4(6):721–732, 2008
S. Hoops, S. Sahle, R. Gauges, C. Lee, J. Pahle, N. Simus, M. Singhal, L. Xu, P. Mendes, and
U. Kummer. COPASI – a COmplex PAthway SImulator. Bioinformatics, 22:3067–3074, 2006
E. Hopper-Borge, Z. S. Chen, I. Shchaveleva, M. G. Belinsky, and G. D. Kruh. Analysis of the
drug resistance profile of multidrug resistance protein 7 (ABCC10): Resistance to docetaxel.
Cancer Research, 64(14):4927–4930, 2004
K. Howe, G. G. Gibson, T. Coleman, and N. Plant. In silico and in vitro modelling of hepato-
cyte drug transport processes: Importance of ABCC2 expression levels in the disposition of
carboxydichlorofluroscein. Drug Metabolism and Disposition, 37(2): 391–399, 2009
M. Hucka, A. Finney, H. M. Sauro, H. Bolouri, J. C. Doyle, H. Kitano, A. P. Arkin, B. J. Bornstein,
D. Bray, A. Cornish-Bowden, A. A. Cuellar, S. Dronov, E. D. Gilles, M. Ginkel, V. Gor,
I. I. Goryanin, W. J. Hedley, T. C. Hodgman, J. H. Hofmeyr, P. J. Hunter, N. S. Juty,
J. L. Kasberger, A. Kremling, U. Kummer, N. Le Novere, L. M. Loew, D. Lucio, P. Mendes,
E. Minch, E. D. Mjolsness, Y. Nakayama, M. R. Nelson, P. F. Nielsen, T. Sakurada, J. C. Schaff,
B. E. Shapiro, T. S. Shimizu, H. D. Spence, J. Stelling, K. Takahashi, M. Tomita, J. Wagner,
and J. Wang. The systems biology markup language (SBML): A medium for representation
and exchange of biochemical network models. Bioinformatics, 19(4):524–531, 2003
S. H. Kim, F. R. Miller, L. Tait, J. Zheng, and R. F. Novak. Proteomic and phosphoproteomic alter-
ations in benign, premalignant and tumor human breast epithelial cells and xenograft lesions:
Biomarkers of progression. International Journal of Cancer, 124(12):2813–2828, 2009
H. Kitano. Biological robustness. Nature Reviews Genetics, 5(11):826–837, 2004
H. Lennernas. Intestinal permeability and its relevance for absorption and elimination. Xenobiotica,
37(10–11):1015–1051, 2007
4 Modeling Transport Processes 81

B. T. McGrogan, B. Gilmartin, D. N. Camey, and A. McCann. Taxanes, microtubules and chemore-

sistant breast cancer. Biochimica Et Biophysica Acta-Reviews on Cancer, 1785(2):96–132,
2008
K. Palm, K. Luthman, J. Ros, J. Grasjo, and P. Artursson. Effect of molecular charge on intestinal
epithelial drug transport: pH-dependent transport of cationic drugs. Journal of Pharmacology
and Experimental Therapeutics, 291(2):435–443, 1999
J.-M. Pascussi, S. Gerbal-Chaloin, J-M. Fabre, P. Maurel, and M-J. Vilarem. Dexamethasone en-
hances constitutive androstane receptor expression in human hepatocytes: Consequences on
cytochrome P450 gene regulation. Molecular Pharmacology, 58(6):1441–1450, 2000a
J.-M. Pascussi, L. Drocourt, J.-M. Fabre, P. Maurel, and M.-J. Vilarem. Dexamethasone induces
pregnane X receptor and retinoid X receptor-a expression in human hepatocytes: Synergistic
increase of CYP3A4 induction by pregnane X receptor. Molecular Pharmacology, 58:361–372,
2000b
N. Perriere, S. Yousif, S. Cazaubon, N. Chaverot, F. Bourasset, S. Cisternino, X. Decleves, S. Hori,
T. Terasaki, M. Deli, J. M. Scherrmann, J. Temsamani, F. Roux, and P. O. Couraud. A functional
in vitro model of rat blood-brain barrier for molecular analysis of efflux transporters. Brain
Research, 1150:1–13, 2007
N. Plant. Molecular toxicology. Advanced Text. BIOS, London, 2003
N. Plant. Interaction networks: Coordinating responses to xenobiotic exposure. Toxicology, 202:
21–32, 2004
N. Plant. The human cytochrome P450 3A sub-family: Transcriptional regulation, inter-individual
variation and interaction networks. Biochimica Biophysica Acta, 1770(3):478–488, 2007
A. Poirier, T. Lave, R. Portmann, M. E. Brun, F. Senner, M. Kansy, H. P. Grimm, and C. Funk.
Design, data analysis, and simulation of in vitro drug transport kinetic experiments using a
mechanistic in vitro model. Drug Metabolism and Disposition, 36(12):2434–2444, 2008
S. Ramaswamy, P. Tamayo, R. Rifkin, S. Mukherjee, C. H. Yeang, M. Angelo, C. Ladd, M. Reich,
E. Latulippe, J. P. Mesirov, T. Poggio, W. Gerald, M. Loda, E. S. Lander, and T. R. Golub. Mul-
ticlass cancer diagnosis using tumor gene expression signatures. Proceedings of the National
Academy of Sciences United States of America, 98(26):15149–15154, 2001
G. Rechkemmer. Transport of weak electrolytes. In M. Field and R. A. Frizzel, editors, Handbook
of physiology, section 6: The gastrointestinal system, pp 371–388. American Physiological So-
ciety, Bethesda, 1991
R. A. Roberts, D. W. Nebert, J. A. Hickman, J. H. Richburg, and T. L. Goldsworthy. Perturbation
of the mitosis/apoptosis balance: A fundamental mechanism in toxicology. Fundamental and
Applied Toxicology, 38(2):107–115, 1997
A. Rostami-Hodjegan and G. T. Tucker. Simulation and prediction of in vivo drug metabolism in
human populations from in vitro data. Nature Reviews Drug Discovery, 6(2):140–148, 2007
G. A. Sawada, C. L. Barsuhn, B. S. Lutzke, M. E. Houghton, G. E. Padbury, N. F. H. Ho, and
T. J. Raub. Increased lipophilicity and subsequent cell partitioning decrease passive tran-
scellular diffusion of novel, highly lipophilic antioxidants. Journal of Pharmacology and
Experimental Therapeutics, 288(3):1317–1326, 1999
G. Shinar, R. Milo, M. R. Martinez, and U. Alon. Input-output robustness in simple bacterial
signaling systems. Proceedings of the National Academy of Sciences of the United States of
America, 104(50):19931–19935, 2007
M. D. Slack, E. D. Martinez, L. F. Wu, and S. J. Altschuler. Characterizing heterogeneous cellular
responses to perturbations. Proceedings of the National Academy of Sciences United States of
America, 105(49):19306–19311, 2008
D. Smirlis, R. Muangmoonchai, M. Edwards, I. R. Phillips, and E. A. Shephard. Orphan recep-
tor promiscuity in the induction of cytochromes P450 by xenobiotics. Journal of Biological
Chemistry, 276(16):12822–12826, 2001
C. Sotiriou and L. Pusztai. Gene-expression signatures in breast cancer. New England Journal of
Medicine, 360(8):790–800, 2009
82 N. Plant

B. Tan, D. Piwnica-Worms, and L. Ratner. Multidrug resistance transporters and modulation. Cur-
rent Opinion in Oncology, 12(5):450–458, 2000
T. N. Tozer and M. Rowland. Introduction to pharmacokinetics and pharmacodynamics: The quan-
titative basis of drug therapy. Lippincott Williams & Wilkins, Philadelphia, 2006
T. T. Tran, A. Mittal, T. Gales, B. Maleeff, T. Aldinger, J. W. Polli, A. Ayrton, H. Ellens, and
J. Bentz. Exact kinetic analysis of passive transport across a polarized confluent MDCK cell
monolayer modeled as a single barrier. Journal of Pharmaceutical Sciences, 93(8):2108–2123,
2004
T. T. Tran, A. Mittal, T. Aldinger, J. W. Polli, A. Ayrton, H. Ellens, and J. Bentz. The elementary
mass action rate constants of P-gp transport for a confluent monolayer of MDCKII-hMDR1
cells. Biophysical Journal, 88(1):715–738, 2005
N. P. Ulrih, U. Adamlje, M. Nemec, and M. Sentjurc. Temperature- and pH-induced structural
changes in the membrane of the hyperthermophilic archaeon Aeropyrum pernix K1. Journal of
Membrane Biology, 219(1–3):1–8, 2007
D. J. Villeneuve, S. L. Hembruff, Z. Veitch, M. Cecchetto, W. A. Dew, and A. M. Parissenti.
cDNA microarray analysis of isogenic paclitaxel- and doxorubicin-resistant breast tumor cell
lines reveals distinct drug-specific genetic signatures of resistance. Breast Cancer Research and
Treatment, 96(1):17–39, 2006
M. L. H. Vlaming, J. S. Lagas, and A. H. Schinkel. Physiological and pharmacological roles of
ABCG2 (BCRP): Recent findings in Abcg2 knockout mice. Advanced Drug Delivery Reviews,
61(1):14–25, 2009
U. K. Walle and T. Walle. Taxol transport by human intestinal epithelial Caco-2 cells. Drug
Metabolism and Disposition, 26(4):343–346, 1998
R. E. Watkins, G. B. Wisely, L. B. Moore, J. L. Collins, M. H. Lambert, S. P. Williams,
T. M. Willson, S. A. Kliewer, and M. R. Redinbo. The human nuclear xenobiotic receptor
PXR: structural determinants of directed promiscuity. Science, 292(5525):2329–2333, 2001
P. Wils, A. Warnery, V. Phungba, S. Legrain, and D. Scherman. High liophilicity decreases drug
transport across intestinal epithelial-cells. Journal of Pharmacology and Experimental Thera-
peutics, 269(2):654–658, 1994
H. Xiao, P. Verdier-Pinard, N. Fernandez-Fuentes, B. Burd, R. Angeletti, A. Fiser, S. B. Horwitz,
and G. A. Orr. Insights into the mechanism of microtubule stabilization by Taxol. Proceedings
of the National Academy of Sciences of the United States of America, 103(27):10166–10173,
2006
K. A. Youdim, A. Avdeef, and N. J. Abbott. In vitro trans-monolayer permeability calculations:
Often forgotten assumptions. Drug Discovery Today, 8(21):997–1003, 2003
Y. Zhou, E. Hopper-Borge, T. Shen, X. C. Huang, Z. Shi, Y. H. Kuang, T. Furukawa, S. Akiyama,
X. X. Peng, C. R. Ashby, X. Chen, G. D. Kruh, and Z. S. Chen. Cepharanthine is a potent re-
versal agent for MRP7(ABCC10)-mediated multidrug resistance. Biochemical Pharmacology,
77(6):993–1001, 2009
Chapter 5
Systems Biology of Tuberculosis: Insights
for Drug Discovery

Karthik Raman and Nagasuma Chandra

5.1 Introduction

It is estimated that about two billion people, equalling one-third of the world’s
total population, are infected with Mycobacterium tuberculosis (Mtb) (World Health
Organisation 2008). There are nearly two million deaths every year, translating to
about four deaths a minute. Tuberculosis (TB) is also the leading killer among HIV-
infected people with weakened immune systems. An additional problem we are
confronted with in the recent years is the emergence of drug resistant varieties of
TB. About 500,000 new multi-drug resistant TB (MDR-TB) cases are estimated to
occur every year (World Health Organisation 2008).
More than 20 drugs and the Bacillus–Calmette–Guerin (BCG) vaccine are avail-
able for the treatment of TB. The existing drugs, although of immense value, have
several shortcomings, the most important of them being the emergence of drug re-
sistance, rendering even the front-line drugs inactive. In addition, drugs such as
rifampicin have high levels of adverse effects making them prone for patient incom-
pliance. Adding to these problems are the vicious interactions between the human
immunodeficiency virus and TB, which lead to further challenges for anti-tubercular
drug discovery (Nunn et al. 2005). For example, protease inhibitors have been
shown to be incompatible with rifampicin-containing anti-TB regimens (Bonora and
Di Perri 2008). The existence of several challenges in tackling TB necessitates the
application of newer techniques to study and understand tubercular infection, as
well as generate methods to counter it.
The genomics and the post-genomics eras, with the parallel advances in high-
throughput experimental methods and screening techniques to analyse whole
genomes and proteomes, are witnessing an explosion in the types and amount
of information available, not only with respect to the genome sequences and protein

K. Raman ()
Bioinformatics Centre, Indian Institute of Science, Bangalore 560 012, India
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 83

Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 5,
c Springer Science+Business Media, LLC 2011
84 K. Raman and N. Chandra

structures but also with respect to gene-expression, regulation and protein–protein

interactions. The availability of such information in publicly accessible databases
and the advances in both computing power and computational methods for data
mining and modeling have led to the emergence of several in silico approaches
to systematically address important questions in biology, with an obvious impact
on drug discovery (Apic et al. 2005; Claus and Underwood 2002). Systems-level
approaches to drug discovery aid at multiple stages in the drug discovery pipeline,
particularly in target identification and in identifying the molecular basis of disease
for rational drug discovery.
Drug discovery in the past has relied heavily on animal models and in vivo stud-
ies. In vitro and biochemical studies have served mainly to back up the findings
and provide mechanistic explanations where possible. It is obvious from these that
the need for considering the system as a whole has always been well recognised.
It can be argued that using a mouse or any typical animal model is also a systems ap-
proach. While that is true in some sense, such approaches deviate significantly from
the current systems practices, since the former is more a “black box”, which enables
a “readout” that is a systems output, but does not tell us why or how such an output
results. The current practices, on the other hand, attempts to reconstruct the system
brick by brick and hence facilitates an understanding of why and how an event takes
place, automatically leading to “what if” type of questions, enabling predictions.
The current approaches also depart from the “spherical cow” abstractions that have
often characterised mathematical modeling (Doyle 2001), by virtue of starting re-
constructions from thousands of fundamental building blocks leading to “realistic”
modeling. Availability of “omics” scale experimental data on various fronts such
as genome sequencing, transcriptional profiles, proteins expressed, lipids, glycans
and the hundreds of metabolites that are interconverted by the molecular metabolic
machinery indeed facilitate such realistic modeling.
This chapter focuses on some of the recent advances in the understanding of
Mtb, from a systems biology perspective and the potential of systems-level analyses
to generate more useful drug targets and a better understanding of the disease and
the pathogen itself. The chapter also discusses possibilities of application of systems
approaches in understanding important issues that arise in drug discovery, such as
interaction between the drug, the target and the system as a whole, possible side
effects and causes of drug toxicity. The complete knowledge of metabolic reactions
in an organism helps to analyse all possible interactions between the drug and the
system and also helps to narrow down possible causes for adverse effects and drug
toxicity. Given the fact that cellular systems are extremely complex, a systematic
analysis of all reactions taking place in a cell across various biochemical pathways
is a challenging task. The following sections illustrate some of the approaches that
have been taken to understand Mtb metabolism, protein–protein interactions, emer-
gence of resistance to anti-TB drugs, as well as the complex interactions of Mtb with
the host immune system.
5 Systems Biology of Tuberculosis 85

5.2 Understanding Mtb: A Parts Catalogue

Systems are composed of individual elements or “parts” that interact in various

ways. In general, the behaviour of a system is quite different from merely the sum
of the interactions of its various parts, applicable even more so for complex bio-
logical systems. As Anderson put it as early as 1972 in his classic paper by the
same title, “More is different” (Anderson 1972), it is not possible to reliably predict
the behaviour of a complex system, despite a good knowledge of the fundamental
laws governing the individual components. Systems biology emphasises the study
of larger systems, in an attempt to better elucidate the complex web of interactions
between various underlying components of biological systems.
Every single cell is made up of a bewildering variety of molecules, macro-
molecules and their complexes. In Mtb, its cell wall itself is an excellent example
of complexity: it is made up of three polymers, arabinogalactan-mycolate (Crick
et al. 2001), covalently linked with peptidoglycan and trehalose dimycolate, which
provide a thick protective layer from general antibiotics and the human immune
system (Takayama et al. 2005). Beneath this surface lies a metabolic network with
about a thousand metabolites ranging from simple carbohydrates to complex lipids
and long-chain fatty acids, facilitated and regulated by an array of proteins.

5.3 Assembling the Parts: Network Reconstruction

Systems biology, being a holistic approach to study biological systems in contrast to

traditional reductionist approaches, involves the synthesis of models of the various
“parts” discussed above into networks depicting metabolism, regulation, signalling
and protein–protein interactions, by a process usually termed as “reconstruction”.
Reconstruction involves the integration of disparate sources of data to create a
representation of the chemical events underlying the different biological networks
(Papin et al. 2005).
Table 5.1 gives a broad overview of the various components of the different
biological networks, as well as the elements involved in their reconstruction and
methods for simulation. Figure 5.1 presents a graphical view of the various levels of
hierarchy at which Mtb can be studied, also indicating the computational methods
that are generally used.

5.3.1 Annotation of Genomes

The sequencing of the entire genome by Cole and co-workers (Cole et al. 1998),
a landmark in TB research, provided the first glimpse of the genomic constitu-
tion leading to deciphering the nearly 4,000 genes in it and their protein products.
This finding has triggered significant downstream research in the area, many of
86 K. Raman and N. Chandra

Table 5.1 Overview of the reconstruction of biological networks. The major network types
and their components, elements of their reconstruction as well as methods for their simulation
are listed here
Signal Transcriptional Protein–protein
Metabolic transduction regulatory interaction
networks networks networks networks
Components Metabolites Proteins Operons Evolutionary/
Proteins Ions Regulons functional/
(Enzymes) structural
Reactions Metabolites Stimulons linkages
between
proteins
Elements of Genomic data Nodes Component data Phylogenetic
reconstruction (annotations) Modules Interaction data profiling
Stoichiometry Motifs Network state Rosetta stone
Gene-protein- Functional data data Gene
reaction Protein–protein Reaction neighbourhood
associations interactions mechanisms Operon
Reaction rates Reaction Kinetic In silico
Cellular mechanisms parameters two-hybrid
constraints Kinetic Causal Experimental
Kinetic parameters parameters relationships methods
Methods for Stoichiometric Boolean Mechanistic Topological
simulation analysis networks modeling analysis
Constraint-based Mechanistic Stoichiometric
methods modeling analysis
Interaction-based Ensemble Boolean
modeling modeling networks
(graphs)
Mechanistic
modeling
(differential
equations)

them at the “omics” scale, such as proteomics, transcriptomics and metabolomics,

through newer technologies. Comprehension of the large oceans of such data
and translation to useful biological insights require an understanding of what the
individual genes and proteins in the genome do. Genome annotation is in fact
a very important and critical requirement to leverage benefit from large-scale
data. Advances in bioinformatics have led to the development of several toolk-
its, which have to a major extent evolved rapidly in the last two decades to meet
the demand stemming from whole genome sequencing. Through bioinformatics
analyses, a significant amount of the biology of the bacillus has been deciphered
through the identification of genes and proteins involved in several functional mod-
ules such as core metabolic pathways, characteristic lipid metabolism, polyketide
5 Systems Biology of Tuberculosis 87

Fig. 5.1 Various levels of hierarchy at which Mtb can be modeled. The various levels of hierarchy
at which Mtb can be modeled, and the experimental data that are available are illustrated. Models
at many of these levels are useful for drug target identification and drug discovery. The lower
pyramid illustrates the different levels of organisation in the cell. While metabolic pathways and
the interactome are more commonly analysed in systems biology studies, the significant interplay
between the levels necessitates the consideration of many other levels, such as the transcriptome
and the metabolome, which present many insights into the complex web of interactions in a cell

and siderophore metabolism, insertion sequences, immunity and pathogenicity

determinants (Cole et al. 1998; Camus et al. 2002). These annotations pave way for
higher order reconstructions of modules in the genome, thus serving as critical step-
ping stones for systems biology. The information contained in the annotated genome
will be more meaningful when ordered into metabolic pathways, regulatory net-
works and signal transduction networks to understand the cellular networks of the
organism. The coding regions or open reading frames in a genome can be identified
by performing similarity searches of the completed genome against databases of
annotated ontologies, which provide initial clues about pathways. Some resources
that help in this initial annotation are Gene Ontology1 and InterPro.2

1
http://www.geneontology.org/.
2
http://www.ebi.ac.uk/interpro/.
88 K. Raman and N. Chandra

5.3.2 Impact of High-Throughput Experiments

The various kinds of downstream “omics” research has led to the generation of
several important resources such as gene essentiality data through the transposon
site hybridisation (TraSH) method (Sassetti et al. 2003), whole-genome expression
profiles facilitated by microarrays and lists of expressed proteins under different
conditions through proteomics experiments, transcriptional profiles through mi-
croarray data and measured metabolite concentrations producing metabolomics
data. These studies can also be carried out as comparative studies in various chemi-
cal environments. For example, the transcriptional response of each gene to a variety
of conditions like genetic perturbation and response to a chemical or drug can be
studied in a single experiment. It is generally believed that genes related in func-
tion (or part of the same pathway) are co-regulated and therefore exhibit similar
expression profiles. The “omics” scale experimental data are an enormously use-
ful resource for building systems level models, since they provide comprehensive
parts lists of various kinds. In addition, the expression profiling data implicitly
capture interactions, dependencies and influences among the various components,
which are manifested in the form of correlated expression patterns. There are also a
large number of important molecular biology and biochemical studies that have pro-
duced data about individual protein molecules, such as protein–protein interactions,
gene knock-outs and site-directed loss-of-function and gain-of-function mutants, to
name a few data types. Information from these are extremely useful in enriching
the bioinformatics-based gene annotations, since they enable incorporation of more
direct functional information.

5.4 Network Modeling and Simulation

5.4.1 Reconstruction of Mtb Metabolism

Metabolic reconstruction is a process through which the various components of

the metabolic network of a biological system, viz. the genes, proteins, reactions
and metabolites that participate in metabolic activity, are identified, categorised and
inter-connected to form a network. Most often, the system is a single cell of interest,
and building on the genomic sequence as a scaffold, reconstructions can incorpo-
rate hundreds of reactions that approximate the entire metabolic activity of a cell.
A comprehensive review of metabolic reconstruction has been published in Reed
et al. (2006a). Metabolic reconstructions fundamentally rely on the availability of
genome sequence and annotations. The reconstructed metabolic networks may be
quite incomplete, if there are a lot of gaps in the annotation of the genome. It is not
uncommon to find many “dead end” reactions in reconstructed metabolic networks –
reactions which produce a metabolite that participates in no further downstream re-
actions, or reactions that consume a metabolite whose precursors are not present
5 Systems Biology of Tuberculosis 89

in the network. With an increase in metabolomics data and improved functional

annotation of genomes, these knowledge gaps are likely to become smaller. The
analysis of reconstructed metabolic networks can also identify metabolic gaps and
predict missing reactions required to reconcile disagreements between reconstructed
metabolic networks and experimental data (Reed et al. 2006b).
For Mtb, the mycolic acid pathway (MAP) has been reconstructed (Raman et al.
2005) and simulated using flux balance analysis (FBA), a constraint-based approach
for analysing metabolic networks. A mathematical abstraction of the biosynthesis of
mycolic acids and the study of the pathway through FBA led to the identification of
key points in the pathway and the delineation of potential drug targets. In 2007, two
genome-scale reconstructions of Mtb were reported (Beste et al. 2007; Jamshidi and
Palsson 2007), with applications in drug target identification, through the analyses of
essential genes and hard-coupled reaction sets. In the following section, we discuss
the basics of FBA, followed by the analysis of the MAP using FBA, as an example
for metabolic reconstruction and simulation, providing insights into the biology of
the pathway. The two genome-scale reconstructions are also briefly discussed.

5.4.1.1 Flux Balance Analysis

One specific example of metabolic modeling using a constraint-based approach is

FBA (Bonarius et al. 1997; Edwards et al. 2002; Kauffman et al. 2003; Raman and
Chandra 2009), which uses linear optimisation to determine the steady-state reac-
tion flux distribution in a metabolic network by maximising an objective function,
such as ATP production or growth rate (Kauffman et al. 2003). FBA has been shown
to be a very useful technique for analysis of metabolic capabilities of cellular sys-
tems (Edwards and Palsson 2000; Förster et al. 2003; Beste et al. 2007; Jamshidi
and Palsson 2007). FBA involves carrying out a steady-state analysis, using the sto-
ichiometric matrix for the system in question. An important assumption is that the
cell performs optimally with respect to a metabolic function, such as maximisa-
tion of biomass production or minimisation of nutrient utilisation, on the premise
that selection pressures during evolution guide systems towards optimality. Once an
objective function is fixed, the system of equations can be solved to obtain a steady-
state flux distribution. This flux distribution is then used to interpret the metabolic
capabilities of the system.
The stoichiometric information on a metabolic system is encoded in a stoichio-
metric matrix, where every metabolite is represented by a row and every reaction by
a column. The entries in each column correspond to the stoichiometric coefficients
of the metabolites (negative for reactants and positive for products) for each reac-
tion. The stoichiometric matrix Smn of m metabolites and n reactions is a sparse
matrix; generally, the entries are integers. The i th row defines the participation or
connectivity of a particular metabolite across all metabolic reactions, and the jth
column provides the stoichiometry of all metabolites in that reaction. The dynamic
mass balance of the metabolic system is described using the stoichiometric matrix,
90 K. Raman and N. Chandra

relating the flux rates of enzymatic reactions, vn1 to time derivatives of metabolite
concentrations, xm1 as
dx
D Sv (5.1)
dt
 >
v D v1 v2 ::: vni b1 b2 ::: bnext (5.2)

where vi signifies the internal fluxes, bi represents the exchange fluxes in the system,
ni is the number of internal metabolites and next is the number of external metabo-
lites in the system. At steady state,

dx
D Sv D 0 (5.3)
dt
Therefore, the required flux distribution belongs to the null space of S. Since
m < n, the system is under-determined and can be solved for v fixing an optimi-
sation criterion, following which the system translates into a linear programming
problem:
min c> v s. t. S  v D 0 (5.4)
v

where c represents the objective function composition, in terms of the fluxes. Fur-
thermore, the lower and upper bounds of the fluxes can be constrained as follows:

0  vi < 1
1 < bi < 1 (5.5)

which necessitates all internal irreversible reactions to have a flux in the positive
direction and allows exchange fluxes to be in either direction. Practically, a finite
upper bound can be imposed, so that the problem does not become unbounded. This
upper bound may also be decided based on the knowledge of cellular physiology.
FBA also has the capabilities to address the effects of gene deletions and other
types of perturbations on the system. Gene deletion studies can be performed by
constraining the reaction flux(es) corresponding to the gene(s) (and therefore, of
their corresponding proteins(s)), to zero. Effects of inhibitors of particular proteins
can also be studied in a similar way by constraining the upper bounds of their fluxes
to any defined fraction of the normal flux, corresponding to the extents of inhibition.

5.4.1.2 Mycolic Acid Pathway

The mycobacterial cell wall is distinctive and is associated with the pathogenicity of
Mtb (Smith 2003; Barry III et al. 1998; Dubnau et al. 2000; Glickman et al. 2000).
The synthesis of mycolic acids, which are long-chain ˛-alkyl-ˇ-hydroxy fatty acids,
the major constituents of this protective layer, has been shown to be critical for the
survival of Mtb (Draper and Daffé 2005).
5 Systems Biology of Tuberculosis 91

A comprehensive stoichiometric model of MAP was built (Raman et al. 2005)

using publicly available databases such as BioCyc and extensive curation of bio-
chemical and genetic data available in literature. The model of the MAP contained
219 reactions and 197 metabolites, mediated through 28 proteins. FBA was per-
formed on the MAP model, which provided insights into the metabolic capabilities
of the pathway. For FBA, the objective function for optimisation was based on the
production of various mycolates in the system, based on their relative ratios in the
mycobacterial cell wall. On solving the optimisation problem, a flux distribution
was obtained (Fig. 5.2).
The strength of many systems-level analyses stems from their abilities to analyse
perturbations to a system; FBA can be readily applied to perform in silico gene
deletions, whereby the effect of deleting one or more genes on the flux distribu-
tion in the system can be predicted. A systematic gene deletion study of the MAP
was also carried out, as well as the inhibition of InhA by isoniazid (Raman et al.
2005). These studies provide clues about proteins essential for the pathway and
hence lead to a rational identification of possible drug targets. Each of the 28 genes
and hence its gene product were systematically deleted from the MAP model, one at
a time, and their effect on the flux distribution was analysed. Figure 5.2b, c are ex-
amples of flux distributions upon gene deletion and correspond to deletion of inhA
and pcaA, respectively. Upon deletion of inhA, which catalyses 21 reactions in the
Fatty Acid Synthase-II system, the fluxes of almost all reactions were seen to be
zero. On the other hand, upon deletion of pcaA, which is involved only in the pro-
duction of ˛-mycolate, the flux pattern remained largely unaltered, except for the
increase in the flux corresponding to cis-methoxy-mycolate production. A flat flux
distribution profile (of near zero) was observed upon deletion of 16 of the genes
(and hence their gene-products) in the MAP model. Some other genes (as in the
case of pcaA), when deleted, did not significantly alter the overall flux distribution,
since cis-methoxy mycolate is produced in increased quantities, to compensate for
the absence of ˛-mycolate. Analysis of the effects of deletion of individual genes
on the flux profiles of the five mycolates provided a handle to define essential and
non-essential genes. Those deletions that resulted in zero or near-zero fluxes of all
the mycolates were considered as essential, and the rest were considered as non-
essential. A good correlation was observed for 19 genes, no experimental data were
available for four genes and disagreement was seen only for five genes. High corre-
lation with experimentally observed data about the essentiality of individual genes
indicates the usefulness of the MAP model and its study using FBA.
Those genes that were classified as essential in the above analysis automatically
form a first list of putative targets for anti-tubercular drugs, since their total inacti-
vation results in loss of production of mycolic acids and hence the viability or the
pathogenicity of the bacillus. However, it was reasoned that an ideal target should
be essential not only in terms of the reaction it can catalyse, but also as the only
protein coded by the genome that can perform the same task. Moreover, an ideal
target should also have no recognisable homologue in the host system, which can
in principle compete with the same drug, leading to unintended/adverse effects in
the host system. Sequence analysis with the Mtb H37Rv and human proteomes was
 b
92
a Without gene deletions inhA deleted
1 1

0.5 0.5

0 0
197
0.03

Flux (relative)
−0.5 −0.5

−1 −1
0 50 100 150 200 250 0 50 100 150 200 250

pcaA deleted 90% inhibition of InhA

c 1 d
0.2

0.5
0.1

0 0
0.03

Flux (relative)
−0.1
−0.5

−0.2
190 195 200 205 210
−1
0 50 100 150 200 250 0 50 100 150 200 250
Reaction Number Reaction Number

Fig. 5.2 Flux distributions obtained from flux balance analysis (FBA) using the mycolic acid pathway (MAP) model. (a) in an unperturbed state, (b) upon
deletion of inhA, (c) upon deletion of pcaA and (d) upon inhibition of InhA. Insets in (a) and (c) refer to enlarged versions of the indicated portions. Note that
the scale for (d) is different
K. Raman and N. Chandra
5 Systems Biology of Tuberculosis 93

therefore carried out for each of the proteins identified as essential from the gene
deletion studies. These studies indicated that, apart from the known InhA, poten-
tial targets for anti-tubercular drug design are AccD3, Fas, FabH, FabD, DesA1
and DesA2.

5.4.1.3 Genome-Scale Metabolic Models

Two genome-scale metabolic models have been reported for Mtb, viz. GSMN-TB
(Beste et al. 2007) and iNJ661 (Jamshidi and Palsson 2007). These models are based
on a careful reconstruction of Mtb metabolism, based on publicly available pathway
databases and a survey of literature. The models consist of a stoichiometric matrix
representing the metabolism of Mtb, the connections between the genes, proteins
and reactions, as well as the various constraints on the cellular fluxes. For FBA,
an objective function is used, based on the knowledge of mycobacterial biomass
composition. Genome-scale metabolic models are particularly useful to understand
the metabolic capabilities of organisms and for the study of gene deletions in silico,
to identify critical points in the metabolism, which may be potential drug targets.
GSMN-TB. The genome-scale metabolic network (GSMN) of Mtb comprises 849
unique reactions involving 739 metabolites and 726 genes (Beste et al. 2007). The
constraint-based metabolic model was calibrated by growing M. bovis BCG in con-
tinuous culture and the measurement of steady state growth parameters. FBA was
used to calculate substrate consumption rates, which were in good agreement with
experimental measurements. The objective function for FBA was the maximisa-
tion of biomass production; the biomass composition was based on the components
necessary for growth in vitro (Beste et al. 2007). FBA-based in silico gene deletions
were also reported, with a prediction accuracy of 78%. The model predicts that about
34% of the genes in the model are essential for growth in the minimal Middlebrook
7H10 media. FBA of the model also correctly predicts the essentiality for growth
of known drug targets such as inhA, embAB, ddlA and alr. The model was able to
correctly predict increased isocitrate lyase activity in slow-growing cells. The model
demonstrates the predictive power of FBA-based metabolic models, which can be
used to generate a number of hypotheses that may be verified experimentally. Thus,
such metabolic models of pathogenic organisms provide valuable insights into the
biology of the organism, paving the way for new strategies to counter disease.
iNJ661. Another genome-scale metabolic model for Mtb has been reported by Pals-
son and co-workers (Jamshidi and Palsson 2007). The model was used to analyse
the growth of the bacterium on various in silico media. Growth rates consistent with
experimental data were observed in varying media conditions. The agreement of
gene essentiality predictions with experimental data was about 55%; this is due to
the variability of gene expression under different conditions and the incomplete-
ness of biological knowledge. Furthermore, hard-coupled reaction sets, which are
groups of reactions that are forced to operate in unison due to the constraints in the
network (arising due to mass balance and connectivity), were identified, which have
application in the identification of drug targets.
94 K. Raman and N. Chandra

Although the curation of such metabolic models is an extremely tedious pro-

cess, the models are quite versatile and find use in understanding the metabolism of
pathogens, and consequently in drug target identification. The importance of con-
sidering metabolism in drug design has been emphasised earlier (Cornish-Bowden
and Cárdenas 2003), particularly since a key activity of drugs is to alter metabolism;
many known drug targets are enzymes or receptors. Such genome-scale reconstruc-
tion studies also have their limitations. It is often difficult to accurately determine
the biomass composition, which is fundamental to FBA and predictions of gene
essentiality. This is particularly true in case of pathogenic organisms such as Mtb,
where it is particularly difficult to perform experiments and determine the biomass
constituents/various growth parameters; practically, it is possible to estimate such
parameters from experiments with avirulent strains, such as M. bovis, which may
not truly reflect the behaviour of Mtb. Furthermore, as discussed earlier, the quality
of the metabolic reconstructions is limited by the availability of genome annota-
tions. The predictions of gene essentiality also suffer from this incompleteness, as
also from the incomplete definition of the biomass function, presence of unknown
isozymes for a given reaction (another dimension of network incompleteness) and a
failure to consider the build-up of toxic intermediates. A detailed discussion of the
factors underlying incorrect in silico predictions of essential metabolic genes has
been presented elsewhere (Becker and Palsson 2008). Notwithstanding these limi-
tations, models and the simulation methodologies currently adopted still capture the
metabolic structure in the cells fairly accurately and will no doubt serve as a frame-
work to integrate newer information, as it becomes available and thus further refine
the models and pose a variety of questions that may be addressed with increased
confidence.

5.4.2 Transcriptional Analysis

The advances in microarray technology has enabled the genome-scale analysis of

mRNA expression profiles in various organisms, including Mtb. A detailed review of
genome-scale expression analyses of Mtb has been reported elsewhere (Waddell and
Butcher 2007). A comprehensive study of the differential transcriptional response
of Mtb to drugs and growth-inhibitory conditions has been reported earlier (Boshoff
et al. 2004). A total of 430 microarray profiles were generated, which were then
clustered. Agents of known mechanism of action were clustered together, while the
mechanism of action of unknown agents could also be predicted (Boshoff et al.
2004). The fine clustering of genes provides insights into the metabolic response of
Mtb to drug-induced stress, presenting a rational basis for the selection of critical
metabolic targets for new anti-mycobacterials.
In another study, the response of Mtb to minimal inhibitory concentrations of six
anti-microbials was determined, using microarray analysis to elucidate mechanisms
of innate resistance in Mtb (Waddell et al. 2004). A common response to drug ex-
posure which overlapped with a number of other mycobacterial stress responses,
5 Systems Biology of Tuberculosis 95

as well as compound-specific responses were distinguished, including a number of

putative transcriptional regulators and translocation-related genes. These genes may
be implicated in the intrinsic resistance of Mtb to drugs.
It has also become possible to perform large-scale studies of gene essential-
ity (Sassetti et al. 2003), using microarrays. The method known as transposon site
hybridisation mutagenesis uses microarrays to map sites of transposon insertions.
DNA from a transposon library is isolated, and labelled probes are synthesised
from promoters within the transposon. Immediately after mutagenesis, each mutant
contains a single transposon insertion, and the library contains mutations in each
gene in the genome. After a growth phase, mutants harbouring insertions in genes
that are required for survival are lost from the library. A TraSH “insertion probe”
is generated from the selected library, comprising only those sequences comple-
mentary to genes that contain insertions in the selected library. A genomic probe
comprising randomly labelled chromosomal DNA will hybridise to every gene rep-
resented on the array. Spots that hybridise to the genomic probe, but not to the
insertion probe, represent genes that are required for mycobacterial growth (Sassetti
et al. 2003).
The analysis of the bacterial transcriptional response to infection can elucidate
the physiological state of the infecting bacteria, bacterial mechanisms to counter
infection, as well as the micro-environments encountered by the bacteria during the
course of infection (Waddell et al. 2007). Whole-genome transcriptional profiling
of both host and pathogen in ex vivo, animal model and human disease contexts
have been reviewed in Waddell et al. (2007). More recently, an RNA amplification
strategy that has potential to throw light on host–pathogen interactions has also been
reported (Waddell et al. 2008). The expression of Mtb genes in macrophages has also
been studied (Schnappinger et al. 2003) by analysing RNA isolated from infected
murine macrophages using microarrays. In the macrophages, 454 induced genes
and 147 repressed Mtb genes were identified (compared to broth cultures), termed
as the “differential intraphagosomal transcriptome”. Integrating such genome-scale
transcriptional analyses, which provide a wealth of data, can aid in improving the
understanding of TB disease progression.

5.4.2.1 Transcriptional Regulatory Networks in Mtb

Balázsi and co-workers have reported a large transcriptional regulatory networks

(TRN) in Mtb characterising the temporal response of this network during adapta-
tion to stationary phase and hypoxia, using published microarray data (Balázsi et al.
2008). The TRN principally consists of gene regulatory interactions from literature
as well MtbRegList (Jacques et al. 2005). The network was further expanded based
on orthology with Escherichia coli. All Mtb operons was also incorporated, based
on the assumption that transcription factor (TF) binding to the promoter region af-
fects the expression of all genes within an operon. The TRN comprises 783 nodes
corresponding to Mtb genes and their protein products, with 937 links correspond-
ing to 45 TFs directly regulating the expression of target genes. Significantly, 29 of
96 K. Raman and N. Chandra

these 45 TFs (auto-)regulate their own expression. Gene pairs such as Rv2358-furB,
Rv1404-Rv1931c and mprA-sigE participate in two-gene feedback loops. A distinct
set of transcriptional sub-networks affected early and late during adaptation to hy-
poxia and stationary phase were identified, illustrating a progressive shift of modular
network response to growth arrest. Most of the sub-networks were affected in both
conditions, suggesting that a general condition-independent repertoire of transcrip-
tional modules is used in Mtb growth arrest (Balázsi et al. 2008).
Studies such as this hold the key to unravelling the mechanisms of mycobacte-
rial persistence, which is a critical problem in mycobacterial infection, where the
bacteria enter a non-replicating state, insensitive to anti-mycobacterial drugs.

5.4.3 Analysis of the Mtb Interactome

Protein–protein interactions are extremely important in orchestrating the events in

a cell. They form the basis for several signal transduction pathways in the cell, as
well as various transcriptional regulatory networks. The need to understand protein
structure and function has been a critical driving force for biological research in the
recent decades.
Genome-wide functional linkages between proteins can be inferred from high
throughput experimentation or from computational analyses. Eisenberg and co-
workers have reported genome-wide functional linkages in Mtb (Strong et al. 2003),
inferred by computational methods based on genomic context, such as the Rosetta
Stone, which is based on domain fusion (Marcotte et al. 1999), Phylogenetic profile,
based on co-occurrence of proteins across genomes (Pellegrini et al. 1999), Operon
and Conserved Gene Neighbour, based on the proximity of genes on the chromo-
some across several genomes (Dandekar et al. 1998). By clustering proteins with
similar functional linkage profiles, it is possible to infer the function of uncharac-
terised proteins and identify functionally linked gene clusters across the proteome.
This study once again demonstrates the utility of a genome-scale analyses vis-à-vis
analyses of individual protein interactions/functional linkages. Such protein–protein
interaction maps also find utility in drug target identification (Verkhedkar et al. 2007;
Raman et al. 2008) and in the analysis of resistance pathways (Raman and Chandra
2008), as will be discussed in later sections. Various concepts from graph theory
have been applied to study biological networks (Barabási and Oltvai 2004). Many of
the highly connected proteins in protein interaction networks, referred to as “hubs”,
have been shown to be critical for cellular function (Jeong et al. 2001); such hub
proteins also represent potential drug targets (Verkhedkar et al. 2007).
Although such protein–protein functional linkages enable a wide variety of
analyses, in some ways they represent an over-simplified static view of the dy-
namic interactions in the cell. Presently, it is difficult to estimate the parame-
ters governing these interactions, such as association and dissociation constants,
which govern many aspects of cellular function, particularly signalling. High-
throughput studies for identifying protein–protein interactions, such as the yeast
5 Systems Biology of Tuberculosis 97

two-hybrid assay (Fields and Song 1989), despite their numerous advantages and
versatility, produce a number of false positives and false negatives, although these
are being addressed a number of recent advances. A comprehensive overview of
the systems biology applications and limitations of the yeast two-hybrid assay
is presented elsewhere (Brückner et al. 2009). Computational methods to predict
functional linkages also suffer from false positives and negatives, but these can
be addressed by considering consensus predictions from multiple methods. The
STRING database (Von Mering et al. 2007) considers predictions based on mul-
tiple methods as well as experimental data and assigns a confidence score to each
interaction. With a refinement of both computational and experimental techniques
to delineate protein–protein interactions, the quality of constructed interactomes
is likely to improve significantly in the future, enabling analyses with greater
confidence.

5.5 Target Identification

Drug discovery has itself witnessed a paradigm shift from the traditional medicinal
chemistry-based ligand-oriented drug discovery approaches to rational drug target
identification and target-driven lead discovery, by targeting the molecular mecha-
nisms of disease. Traditionally, targets have been identified through knowledge of
the function of individual protein molecules, where their function has been well
characterised. Potential targets thus identified are generally taken through a val-
idation process involving whole-cell or animal experiments, gene knock-outs or
site-directed mutagenesis that lead to loss-of-function phenotypes. Target valida-
tion is one of the critical steps in drug discovery, where a lot of time and money is
spent in the pharmaceutical industry. The need for systematic and large-scale val-
idation in the post-genomic era has led to the use of computational methods for
validation (Raman et al. 2007).
A number of studies have been carried out using various experimental meth-
ods to identify drug targets in Mtb (Mdluli and Spigelman 2006). Attempts have
also been made for the same purpose, based on sequence comparisons of metabolic
enzymes (Anishetty et al. 2005), and using various features such as metabolic
choke-points at the systems-level (Hasan et al. 2006). The wealth of information
available from the genome sequence, as well as metabolic and protein interaction
networks can be analysed to identify potential drug targets in Mtb. In this section,
we discuss how systems biology concepts and understanding the microbe as a whole
open up new opportunities for computational target identification.

5.5.1 Multi-Level Target Identification Pipeline: TargetTB

It is now well established that better insights into biological systems may be ob-
tained by considering large-scale systems-level models, since biological systems
98 K. Raman and N. Chandra

are complex networks of many processes. The conventional method of focussing

on a single protein at a time, however important the protein may be, would mean
losing perspective of its larger context and hence may not provide the right answers,
especially in drug discovery. Broader insights about the appropriateness of a po-
tential target can be obtained by considering pathways and whole-system models
relevant to that disease. For example, an enzyme that may be identified as a good
target for a particular disease may not actually be critical or essential, when viewed
in the context of the entire metabolism in the cell. Analysing systems-level models
can help in assessing criticality of the individual proteins by studying any alternate
pathways and mechanisms that may naturally exist to compensate for the absence
of that protein.
An integrated analysis of Mtb at various levels – metabolic reactions, protein–
protein interactions, protein sequences and structure – can provide a more rational
handle to identify drug targets. Illustrating this, a comprehensive in silico target
identification pipeline for Mtb, targetTB (Raman et al. 2008), has been reported,
which can also be used as a general framework for in silico target identification.
The analyses are focused at a systems-level, based on network analyses and FBA.
The pipeline incorporates (a) a network analysis of the protein–protein interactome,
(b) an FBA of the reactome, (c) experimentally derived phenotype essentiality data,
(d) sequence analyses and (e) a structural assessment of targetability using novel
algorithms. Using FBA and network analysis, proteins critical for survival of Mtb
are first identified, followed by comparative genomics with the host, finally incor-
porating a novel structural analysis of the binding sites to assess the feasibility of
a protein as a target. Further pruning of the chosen targets was done based on (f)
analysis of expression of suggested target proteins, based on available expression
data and (g) non-similarity to gut flora proteins as well as (h) “anti-targets” in the
host, leading to the identification of 451 high-confidence targets. Through phyloge-
netic profiling against 228 pathogen genomes, shortlisted targets have been further
explored to identify broad-spectrum antibiotic targets, while also identifying those
specific to TB. Targets that address (i) mycobacterial persistence and (j) drug resis-
tance mechanisms are also analysed.
Besides essentiality to the pathogen, an ideal target should have several other
properties such as non-similarity with human proteins whose inhibition could lead
to potential adverse drug effects, an aspect that has been analysed at multiple levels
in this study (see Fig. 5.3).
The simplest level of course is to check for sequence similarity of the target be-
ing queried with all the proteins in the human proteome. However, sequence filtering
while important cannot be the sole criteria for identifying high quality targets, since
two proteins that are considerably dissimilar in their sequences could have very
similar binding sites (Ramachandraiah and Chandra 2000; Vinod et al. 2006). Thus,
while sequence similarity very often leads to structural and hence functional simi-
larity, it is not a necessary condition for two proteins to have similar ligand-binding
profiles.
5 Systems Biology of Tuberculosis 99

Fig. 5.3 The targetTB target identification pipeline. The flowchart depicts the order in which the
entire proteome of Mtb is considered and analysed at different layers. “A” refers to the systems level
studies, which includes A1, for network analysis of the interactome; A2, for FBA of the reactome;
and A3, for genome-scale essentiality data determined experimentally as reported by Sassetti et al.
(2003). Those proteins that passed these filters are indicated as “A”, and combined with the results
of sequence analysis (A), to derive those that passed both filters (depicted as “A&B”). These were
then taken through Filter C, referring to the structural assessment filter, yielding the list of 622
proteins as the D-List (A&B&C). Further steps of filtering are indicated in the smaller funnel as
E (expression under various conditions), F (non-similarity to anti-targets) and G (non-similarity to
gut flora proteins). Those proteins that pass all the six levels of filtering (indicated as D&E&F&G)
form the H-List comprising 451 targets. Additional filters I, J and K used for analysing the H-List
are also indicated

Genome-scale structural assessment. In the process of target identification, the crit-

ical aspect of a good target is to have a binding site in the target protein that is
sufficiently different from that of any host protein. It is important to consider speci-
ficity at the binding site, hence the molecular recognition level, since a given drug
should be available in intended quantities to the desired target. At the same time,
a given drug should ideally not exhibit unintended recognition by some other host
protein, so that adverse effects due to unanticipated functional manipulation of other
host proteins can be avoided. For this purpose, it is important to study the possible
binding profile of a given drug to all those proteins to which it is likely to be ex-
posed. Towards this goal, possible pockets in the set of Mtb and human structures
100 K. Raman and N. Chandra

were first identified, using PocketDepth, a validated algorithm that was recently
developed (Kalidas and Chandra 2008). All such putative pockets were tested for
certain criteria such as size and volume, retaining only those that were likely to
bind to small molecules. The filtered pockets from preliminarily short-listed targets
from Mtb were then screened for similarity against pockets from the human pro-
teins, which involved over 245 million comparisons, using PocketMatch, another
recently developed site-matching algorithm (Yeturu and Chandra 2008). From this,
145 putative targets were eliminated due to high similarity with one or more human
proteins. Some examples of molecules that have failed at this stage are DdlA, GyrB,
AftA and AlrA. It must be noted that some of these were ranked as high priority
targets by other studies that did not consider the structural aspect explicitly, again
emphasising the need for structural level analysis. Eliminating those proteins with
high similarity to proteins in the gut flora also helps in ultimately reducing the risk
of side effects.
The last stages of filtering and post-identification analyses resulted in identify-
ing two categories of targets: broad-spectrum targets and Mtb-specific targets. It
is necessary to identify targets in both the categories, since they are required in
different situations. Mtb-specific targets are believed to be safer since they would
not lead to many organisms developing resistance against the drugs of such tar-
gets. Broad-spectrum targets, on the other hand, would be extremely useful when
multiple infections co-exist or in some cases where a specific diagnosis is not pos-
sible. A comprehensive phylogenetic analysis of the short-listed targets against 228
different pathogenic genomes has been carried out, to identify broad-spectrum tar-
gets. Identification of pathways and proteins involved in generating drug resistance
and then targeting them simultaneously as co-targets along with the primary broad-
spectrum targets would reduce the risk of drug resistance significantly, making many
more molecules accessible for therapeutic intervention.

5.5.1.1 Importance of Systems-Based Approaches

The pipeline described shows how systems biology methods can be used to obtain
significant insights into essentiality, identifying possible lists of essential proteins
and of course understanding reasons for their essentiality as well. The study de-
scribed here demonstrates the usefulness of such insights in target identification for
tuberculosis and how they can be integrated along with other canonical lines of in-
vestigation such as sequence and structural analyses of the individual molecules.
The pipeline developed provides rational schema for drug target identification that
are likely to have high rates of success, which is expected to save enormous amounts
of money, resources and time in the drug discovery process. A thorough comparison
with previously suggested targets in the literature demonstrates the usefulness of the
integrated approach used in the study, highlighting the importance of systems-level
analyses in particular (Raman et al. 2008). The method has the potential to be used
as a general strategy for target identification and validation and hence significantly
impacts most drug discovery programmes.
5 Systems Biology of Tuberculosis 101

5.5.2 Disruption of Metabolism

It has been said that drug design has often not included the idea that what cells do
is metabolism, and a major thing drugs are supposed to do is to alter metabolism:
of the 500 well-known targets, 30% are enzymes and 45% are receptors (Cornish-
Bowden and Cárdenas 2003). Therefore, it is quite important to consider metabolism
during drug design and drug target identification. Given this, disrupting mycobac-
terial metabolism to the point of destruction would be a more useful approach
than to consider one metabolic target at a time. A recent study highlights the use
of a protein–protein influence network derived from metabolic linkages to iden-
tify combinations of proteins, which when simultaneously inhibited can together
disrupt bacterial metabolism to a significant extent, thus ensuring bacterial clear-
ance (Raman et al. 2009). An FBA of these identified combinations indicate that
metabolism has indeed been disrupted. With key proteins in the network, multiply
targeted, the chances of recovery by the bacilli and emergence of drug resistance
would also be hampered in a major way. Targeting multiple points in a metabolic
pathway can be a useful strategy in drug design in general, perhaps explaining why
combination therapy is popular. A report in literature highlights the possibility of
“natural” crude drugs acting on multiple targets with multiple mechanisms, attribut-
ing their success to this plurality (Csermely et al. 2005).

5.5.3 Tackling Resistance in Mtb

A major problem with the current chemotherapeutic agents for TB is the emergence
of drug resistance. Although several approaches have been explored to counter re-
sistance, there has been limited success due to a lack of understanding of how
resistance emerges in bacteria upon drug treatment.
A proteome-scale network of protein–protein associations in Mtb has been used,
to discover possible pathways that may be responsible for generating drug resis-
tance (Raman and Chandra 2008). The protein–protein interactome of Mtb enables
a novel formulation of the problem of drug resistance and forms a first step to-
wards countering drug resistance at the drug discovery stage itself. In particular, the
questions such as: (a) how does the information flow from the drug target to the re-
sistance machinery, and (b) how do targets differ in their propensities for triggering
resistance, can be addressed.
A genome scale protein–protein interaction network for Mtb H37Rv was derived
from the STRING database (Von Mering et al. 2007). A set of proteins involved
in both intrinsic and extrinsic drug resistance mechanisms were identified from lit-
erature. Shortest paths from different drug targets to the set of resistance proteins
in the protein–protein interactome were computed, to derive a sub-network rele-
vant to study emergence of drug resistance. The shortest paths were then scored
and ranked based on (a) drug-induced gene upregulation data, from microarray ex-
periments reported in literature, for the individual nodes and (b) edge-hubness, a
102 K. Raman and N. Chandra

network parameter that signifies centrality of a given edge in the network. High-
scoring paths, which contain “central” proteins up-regulated on exposure drugs,
indicate most plausible pathways for the emergence of drug resistance. Different
targets appear to have different propensities for four drug resistance mechanisms,
giving rise to a very important direction to explore in drug discovery.
The study leads to the identification of possible pathways for drug resistance, pro-
viding novel insights into the problem of resistance. A new concept of “co-targets”,
to counter resistance by simultaneously inhibiting a protein responsible for resis-
tance, along with the intended target of the drug, has been proposed to counter
mycobacterial drug resistance. RecA, Rv0823c, Rv0892 and DnaE1 were among
the best examples of co-targets for combating drug resistance in TB (Raman and
Chandra 2008). This approach is also inherently generic, likely to significantly im-
pact drug discovery.

5.6 Interface with the Host: Modeling Host–Pathogen

Interactions

The establishment of any infection is contingent upon the interplay of virulence

mechanisms of a pathogenic organism, the defence mechanisms of the host as
well as the counter-defence of either organism. A comprehensive understanding
of the mechanisms of host–pathogen interactions can aid in the identification of
the critical points for countering infection. Although a comprehensive mechanistic
model of host–pathogen systems is still not available, several approaches have been
undertaken to identify and model host–pathogen interactions. These range from sim-
pler models for the prediction of protein–protein interactions between the host and
pathogen, to complex models for the signal transduction networks and Boolean net-
work models of immunological components of the interplay of various mechanisms
of attack and defence in the host and pathogen.

5.6.1 Response Networks

Response network analysis involves the analysis of experimental data such as gene
expression profiles, in the context of biological networks. Superposing network in-
formation with experimental data, networks representing the best system response
according to the tested experimental conditions are identified (Forst 2006). Siegel
and co-workers integrate expression data with molecular interaction data to iden-
tify active sub-networks, or, connected regions of the network showing significant
changes in expression (Ideker et al. 2002). Forst and co-workers also explore the dif-
ferential network expression during response of Mtb to stress (induced by hydrogen
peroxide) and drugs such as isoniazid, using concepts from graph theory (Cabusora
et al. 2005). The expression data of known stress responders and DNA repair genes
in Mtb were used to construct a generic stress response sub-network. This was then
5 Systems Biology of Tuberculosis 103

compared to similar networks constructed from data obtained from subjecting Mtb
to various drugs; this analysis helps to distinguish between generic stress response
and specific drug response, which can be exploited in drug discovery. With a growth
in microarray data, it is possible to extend these ideas to the host–pathogen interac-
tome networks. Genes that are selectively expressed during infection may be more
likely to be involved in virulence.

5.6.2 Mechanistic Models of Immune System Dynamics

Kirschner and co-workers have worked on several mathematical models for the in-
teraction of Mtb with the human immune system (Wigginton and Kirschner 2001;
Marino and Kirschner 2004; Segovia-Juarez et al. 2004; Marino et al. 2007a).
These mathematical models use differential equations encapsulating the interactions
between various host cells, cytokines and the pathogen. Comprehensive reviews
of mathematical models of Mtb infection and its interactions with the human im-
mune system have been published elsewhere (Kirschner and Marino 2005; Young
et al. 2008). A virtual model of the immune response to Mtb that characterises
the cytokine and cellular network that is operational during TB infection has been
reported (Wigginton and Kirschner 2001). The dynamics of the various model
components such as macrophage and cytokines are described using differential
equations. Using this model, the parameters governing the behaviour of the sys-
tem towards the different outcomes have been identified. The study concludes that
factors affecting macrophage functions (such as activation, infection and bacterici-
dal capabilities) as well as effector T cell functions (cytotoxicity from CD4C T cells
as well as other cells such as CD8C T cells) must achieve a balance to control
infection. Virtual deletion and depletion experiments have also been performed to
confirm these hypotheses. The model has been further extended to a two compart-
mental model capturing the important processes of cellular activation and priming
that occur between the lung and the nearest draining lymph node. The model is able
to reproduce typical disease progression scenarios including primary infection, la-
tency or clearance (Marino and Kirschner 2004). Agent-based models for simulating
granuloma formation have also been reported (Segovia-Juarez et al. 2004).
Marino and Kirschner have developed a model which shows that delays in ei-
ther dendritic cell migration to the draining lymph node or T-cell trafficking to
the site of infection can alter the outcome of Mtb infection (Marino et al. 2004).
A mathematical model of immune response to Mtb in the lungs, exploring the role
CD8C T cells, has also been developed (Sud et al. 2006). Ray and Kirschner have
also developed a mathematical model comprising several differential equations de-
scribing macrophage biochemical processes based on three functional modules, viz.
activation, killing and iron regulation (Ray and Kirschner 2006). They suppose the
requirement of multiple activation signals for the macrophage to overcome the qui-
escent state. While the innate immune response develops first occurring on the order
of minutes and hours, adaptive immunity follows later occurring on the order of days
104 K. Raman and N. Chandra

or weeks. Each has an inherent delay in their development, and this timing may be
crucial in determining success or failure in clearing the pathogen. A general model
of the twofold immune response, specifically to intracellular bacterial pathogens,
incorporating mathematical delays for both innate and adaptive immune response
has been developed (Beretta et al. 2007).
The role of tumour necrosis factor (TNF-˛) in protection against the tubercle
bacillus in both active and latent infection has also been modeled, providing in-
sights into the role of TNF-˛ in TB pathology and control (Marino et al. 2007b).
The model consists of non-linear differential equations describing the dynamics of
macrophage, T cells, cytokines and bacteria. The effect of TNF-˛ and IFN- sig-
nalling on activation of the macrophage during Mtb infection has also been analysed
using a mathematical model (Ray et al. 2008). Each component of the model, such
as TNF-˛, IFN- and nitric oxide (NO), is represented as a continuous entity in
an ordinary differential equation. Using the model, it has been shown that negative
feedback from production of nitric oxide, the key mediator of mycobacterial killing,
which typically optimises macrophage responses to activating stimuli, may reduce
effective killing of Mtb.

5.6.3 Boolean Modeling of Mtb-Human Interactions

The roots of Boolean network modeling may be traced to as early as 1969, when
Kaufmann described the use of such models for studying cellular control pro-
cesses (Kauffman 1969). Another insightful exposition of Boolean network theory
for modeling genetic circuits was given later by Thomas (1973). Boolean network
models have been used successfully to predict the expression pattern of the segment
polarity genes in Drosophila melanogaster (Albert and Othmer 2003). Brahmachari
and co-workers have applied Boolean network modeling to analyse a neurotransmit-
ter pathway implicated in schizophrenia (Gupta et al. 2007). Albert and co-workers
have applied Boolean networks for the modeling of host–pathogen interactions in
Bordetella (Thakar et al. 2007).
Boolean network models are composed of various nodes, representing important
components or processes in the system. The state of each node in the network can
be either “on” (true) or “off” (false), a qualitative description of the concentration or
activity. Boolean network representations involve transfer functions that encode the
interactions between the various states. Transfer functions define a discrete dynamic
system, using logical constructs such as “AND”, “OR” and “NOT”. For example,
activations can be represented by an “OR” operator, while an inhibition can be en-
coded for, by an “AND NOT” operator. When more than one of the components
need to be present concurrently, to cause an activation, an “AND” can be used. Each
iteration of simulation determines the evolution of the state of nodes.
One of our latest studies involves building a multi-level model of host–pathogen
interactions in TB, based on an extensive survey of various experiments reported in
literature, accounting for the innate and adaptive immune responses of the host, as
5 Systems Biology of Tuberculosis 105

well as the various defence mechanisms of the pathogen (Raman 2008; Raman et al.
2010). The complex regulation by the various cytokines present in the cell has also
been encoded in the model. The model contains 75 nodes, about one-fourth of them
relating to bacterial components, the rest being components of the human immune
system. Boolean transfer functions describe the relationships between the nodes.
For example, the state of activated dendritic cells in the system could be described
as follows:
Activated Dendritic cells* D (Dendritic cells and Bacteria) or Activated phagocy-
tic cells or (Dendritic cells and Bacteria and
(Th1RC or Th2RC)).
This is based on the knowledge that immature DCs, upon stimulation by bacteria,
get activated and mature in the lymph nodes; the activation of dendritic cells may
also be aided by activated phagocytic cells and cytokines produced in T helper cells
(Th1 or Th2).
Virtual deletion experiments have been performed, where one or more
components of the system are removed and the response of the system to this
perturbation is analysed. Disabling processes such as phagocytosis and phagolyso-
some fusion or important cytokines such as TNF-˛ and IFN- greatly impaired
bacterial clearance, while removing cytokines such as IL-10 alongside bacterial
defence proteins such as SapM greatly favoured bacterial clearance. The propensity
of the tubercle bacillus to persist is highlighted in the simulations. Studies of this
nature are useful to identify key points in the human immune response as well as the
components critical for the elimination of bacteria. An overall understanding of the
interplay of the various mechanisms in host–pathogen interaction lays an excellent
foundation for tackling the disease.

5.7 Future Perspectives

Systems biology signals a departure from the now common view in drug discov-
ery of “single target, one drug, lone therapeutic indication”. Targeting a broader
range of related biological structures should result in compounds that have common
structural and functional properties, and common mechanisms of action, ultimately
creating the potential for the application of a therapeutic to multiple diseases by tar-
geting common pathways implicated in pathogenesis (Davidov et al. 2003). The cul-
mination of systems modeling lies in the modeling of complete systems, accounting
for all component reactions, the localisation of these components and their interac-
tions. The interaction between these organelles or compartments and the interface
with the physical world, in terms of external temperature, pH and other effects be-
comes more relevant in the highest levels of biological organisation. Computational
models of human physiology come into play both to relate to whole animal models
used in traditional pharmacology and more importantly to build integrated data-
driven models that can be refined to mimic the human physiology more closely.
The Physiome project (Hunter and Borg 2003) (http://www.physiome.org/) is one
106 K. Raman and N. Chandra

such effort aimed at describing the human organism quantitatively, to understand

key elements of physiology and pathophysiology. The salient features of the project
are the databasing of physiological, pharmacological and pathological information
on humans and other organisms and integration through computational modeling.
Molecular level understanding of the processes involved in the pharmacokinetics,
bio-availability and toxicity is still very poor. With the current rate of advances in
systems biology, we can also expect significant enhancements in pathway models,
process models and indeed in entire system models, both in terms of mathemati-
cally representing such complex phenomena as well as in terms of mimicking and
simulating the biological events.
While several advances have been made in modeling the host–pathogen interplay,
there still remains a lot to be explored. Accurate mechanistic models of host–
pathogen systems can give reliable insights into complicated phenomena. However,
such models are often limited in their scope. On the other hand, systems-level
models give a much better holistic view of the interplay, at the expense of some ac-
curacy. Large-scale systems-level models of host–pathogen interactions, integrating
information from various levels of abstraction, would be of immense use in under-
standing processes of infection and developing strategies for combating disease. The
future of host–pathogen systems modeling holds promise for uncovering the molec-
ular bases of disease and consequently aids in the discovery of novel therapies.
We can also envisage that the use of pharmacogenomics and tailor-made medicines
could be distinct possibilities in the near future.
Experimentation and computational modeling must be used in complement, each
deriving benefits from the other. Computational modeling can be used to generate
novel hypotheses, which can then be used to guide experimentation. Experimental
verification or validation of a model can render it much more useful, as more reli-
able predictions can be made, on the strength of its proven validity. Systems biology
approaches are also likely to impact development of molecular level pharmacoki-
netic and pharmacodynamic models for individual drugs, to provide comprehensive
profiles of drug actions. Development of comprehensive systems-level models that
encode most of the features of a system will enable a better understanding of drug
toxicity and hence eliminate poor candidates early in the discovery pipeline. Insights
that systems-level models can ultimately translate into more rational and person-
alised therapeutic intervention strategies in clinical practice. Thus, the stage is set
for the integration and application of skills from mathematics, computer science and
engineering disciplines, to address complex problems in biology and drug discovery,
in a big way.

References

R. Albert and H. G. Othmer. The topology of the regulatory interactions predicts the expression
pattern of the segment polarity genes in Drosophila melanogaster. J Theor Biol, 223(1):1–18,
2003
P. W. Anderson. More is different. Science, 177(4047):393–396, 1972
5 Systems Biology of Tuberculosis 107

S. Anishetty, M. Pulimi, and G. Pennathur. Potential drug targets in Mycobacterium tuberculosis

through metabolic pathway analysis. Comput Biol Chem, 29(5):368–378, 2005
G. Apic, T. Ignjatovic, S. Boyer, and R. B. Russell. Illuminating drug discovery with biological
pathways. FEBS Lett, 579(8):1872–1877, 2005
G. Balázsi, A. P. Heath, L. Shi, and M. L. Gennaro. The temporal response of the Mycobacterium
tuberculosis gene regulatory network during growth arrest. Mol Syst Biol, 4:225, 2008
A-L. Barabási and Z. N. Oltvai. Network biology: Understanding the cell’s functional organization.
Nat Rev Genet, 5(2):101–113, 2004
C. E. Barry III, R. E. Lee, K. Mdluli, A. E. Simpson, B. G. Schroeder, R. A. Slayden, and
Y. Yuan. Mycolic acids: Structure, biosynthesis and physiological functions. Prog Lipid Res,
37:143–179, 1998
S. A. Becker and B. Ø. Palsson. Three factors underlying incorrect in silico predictions of essential
metabolic genes. BMC Syst Biol, 2:14, 2008
E. Beretta, M. Carletti, D. E. Kirschner, and S. Marino. Stability analysis of a mathematical model
of the immune response with delays, In: Mathematics for life science and medicine, pages
177–206. Springer, Berlin, 2007
D. J. V. Beste, T. Hooper, G. Stewart, B. Bonde, C. Avignone-Rossa, M. E. Bushell, P. Wheeler,
S. Klamt, A. M. Kierzek, and J. McFadden. GSMN-TB: A web-based genome-scale network
model of Mycobacterium tuberculosis metabolism. Genome Biol, 8:R89, 2007
H. P. J. Bonarius, G. Schmid, and J. Tramper. Flux analysis of underdetermined metabolic net-
works: The quest for the missing constraints. Trends Biotechnol, 15(8):308–314, 1997
S. Bonora and G. Di Perri. Interactions between antiretroviral agents and those used to treat tu-
berculosis: Clinical pharmacology of antiretroviral drugs. Curr Opin HIV & AIDS, 3:306–312,
2008
H. I. Boshoff, T. G. Myers, B. R. Copp, M. R. McNeil, M. Wilson, and C. E. Barry III. The
transcriptional responses of Mycobacterium tuberculosis to inhibitors of metabolism: Novel
insights into drug mechanisms of action. J Biol Chem, 279(38):40174–40184, 2004
A. Brückner, C. Polge, N. Lentze, D. Auerbach, and U. Schlattner. Yeast two-hybrid, a powerful
tool for systems biology. Int J Mol Sci, 10(6):2763–2788, 2009
L. Cabusora, E. Sutton, A. Fulmer, and C. V. Forst. Differential network expression during drug
and stress response. Bioinformatics, 21(12):2898–2905, 2005
J-C. Camus, M. J. Pryor, C. Medigue, and S. T. Cole. Re-annotation of the genome sequence of
Mycobacterium tuberculosis H37Rv. Microbiology, 148(10):2967–2973, 2002
B. L. Claus and D. J. Underwood. Discovery informatics: Its evolving role in drug discovery. Drug
Discov Today, 7:957–966, 2002
S. T. Cole, R. Brosch, J. Parkhill, T. Garnier, C. Churcher, D. Harris, S. V. Gordon, K. Eiglmeier,
S. Gas, C. E. Barry III, F. Tekaia, K. Badcock, D. Basham, D. Brown, T. Chillingworth,
R. Connor, R. Davies, K. Devlin, T. Feltwell, S. Gentles, N. Hamlin, S. Holroyd, T. Hornsby,
K. Jagels, A. Krogh, J. McLean, S. Moule, L. Murphy, K. Oliver, J. Osborne, M. A. Quail,
M.-A. Rajandream, J. Rogerand, S. Rutter, K. Seeger, J. Skelton, R. Squares, S. Squares,
J. E. Sulston, K. Taylor, S. Whitehead, and B. G. Barrell. Deciphering the biology of My-
cobacterium tuberculosis from the complete genome sequence. Nature, 393:537–544, 1998
A. Cornish-Bowden and M. L. Cárdenas. Metabolic analysis in drug design. C R Biol, 326(5):
509–515, 2003
D. C. Crick, S. Mahapatra, and P. J. Brennan. Biosynthesis of the arabinogalactan-peptidoglycan
complex of Mycobacterium tuberculosis. Glycobiology, 11:107R–118R, 2001
P. Csermely, V. Ágoston, and S. Pongor. The efficiency of multi-target drugs: The network ap-
proach might help drug design. Trends Pharmacol Sci, 26:178–182, 2005
T. Dandekar, B. Snel, M. A. Huynen, and P. Bork. Conservation of gene order: A fingerprint of
proteins that physically interact. Trends Biochem Sci, 23(9):324–328, 1998
E. J. Davidov, J. M. Holland, E. W. Marple, and S. Naylor. Advancing drug discovery through
systems biology. Drug Discov Today, 8(4):175–183, 2003
J. Doyle. Computational biology. Beyond the spherical cow. Nature, 411(6834):151–152, 2001
108 K. Raman and N. Chandra

P. Draper and M. Daffé. The cell envelope of Mycobacterium tuberculosis with special reference to
the capsule and outer permeability barrier. In: Stewart T. Cole, Kathleen D. Eisenach, David N.
McMurray, and William R. Jacobs Jr., editors, Tuberculosis and the tubercle bacillus, pages
261–273. American Society of Microbiology Press, 2005
E. Dubnau, J. Chan, C. Raynaud, V. P. Mohan, M. A. Lanéelle, K. Yu, A. Quémard, I. Smith, and
M. Daffé. Oxygenated mycolic acids are necessary for virulence of Mycobacterium tuberculo-
sis in mice. Mol Microbiol, 36(3):630–637, 2000
J. S. Edwards and B. Ø. Palsson. The Escherichia coli MG1655 in silico metabolic genotype: Its
definition, characteristics, and capabilities. Proc Natl Acad Sci USA, 97(10):5528–5533, 2000
J. S. Edwards, M. W. Covert, and B. Ø. Palsson. Metabolic modelling of microbes: The flux-
balance approach. Environ Microbiol, 4(3):133–133, 2002
S. Fields and O. Song. A novel genetic system to detect protein-protein interactions. Nature, 340
(6230):245–246, 1989
C. V. Forst. Host-pathogen systems biology. Drug Discov Today, 11(5–6):220–227, 2006
J. Förster, I. Famili, P. Fu, B. Ø. Palsson, and J. Nielsen. Genome-scale reconstruction of the
Saccharomyces cerevisiae metabolic network. Genome Res, 13(2):244–253, 2003
M. S. Glickman, J. S. Cox, and W. R. Jacobs Jr. A novel mycolic acid cyclopropane synthetase is
required for cording, persistence, and virulence of Mycobacterium tuberculosis. Mol Cell, 5(4):
717–727, 2000
S. Gupta, S. S. Bisht, R. Kukreti, S. Jain, and S. K. Brahmachari. Boolean network analysis of a
neurotransmitter signaling pathway. J Theor Biol, 244(3):463–469, Feb 2007
S. Hasan, S. Daugelat, P. S. Rao, and M. Schreiber. Prioritizing genomic drug targets in pathogens:
Application to Mycobacterium tuberculosis. PLoS Comput Biol, 2(6):e61, 2006
P. J. Hunter and T. K. Borg. Integration from proteins to organs: The Physiome project. Nat Rev
Mol Cell Biol, 4(3):237–243, 2003
T. Ideker, O. Ozier, B. Schwikowski, and A. F. Siegel. Discovering regulatory and signalling cir-
cuits in molecular interaction networks. Bioinformatics, 18 Suppl 1:S233–S240, 2002
P-E. Jacques, A. L. Gervais, M. Cantin, J-F. Lucier, G. Dallaire, G. Drouin, L. Gaudreau, J. Goulet,
and J. Brzezinski. MtbRegList, a database dedicated to the analysis of transcriptional regulation
in Mycobacterium tuberculosis. Bioinformatics, 21(10):2563–2565, 2005
N. Jamshidi and B. Ø. Palsson. Investigating the metabolic capabilities of Mycobacterium tubercu-
losis H37Rv using the in silico strain iNJ661 and proposing alternative drug targets. BMC Syst
Biol, 1:26, 2007
H. Jeong, S. P. Mason, A-L. Barabási, and Z. N. Oltvai. Lethality and centrality in protein networks.
Nature, 411(6833):41–42, 2001
Y. Kalidas and N. Chandra. Pocketdepth: A new depth based algorithm for identification of ligand
binding sites in proteins. J Struct Biol, 161(1):31–42, 2008
K. J. Kauffman, P. Prakash, and J. S. Edwards. Advances in flux balance analysis. Curr Opin
Biotechnol, 14(5):491–496, 2003
S. A. Kauffman. Metabolic stability and epigenesis in randomly constructed genetic nets. J Theor
Biol, 22(3):437–467, 1969
D. Kirschner and S. Marino. Mycobacterium tuberculosis as viewed through a computer. Trends
Microbiol, 13(5):206–211, 2005
E. M. Marcotte, M. Pellegrini, H-L. Ng, D. W. Rice, T. O. Yeates, and D. Eisenberg. Detecting
protein function and protein-protein interactions from genome sequences. Science, 285(5428):
751–753, 1999
S. Marino and D. E. Kirschner. The human immune response to Mycobacterium tuberculosis in
lung and lymph node. J Theor Biol, 227(4):463–486, 2004
S. Marino, S. Pawar, C. L. Fuller, T. A. Reinhart, J. L. Flynn, and D. E. Kirschner. Dendritic
cell trafficking and antigen presentation in the human immune response to Mycobacterium
tuberculosis. J Immunol, 173(1):494–506, 2004
S. Marino, E. Beretta, and D. E. Kirschner. The role of delays in innate and adaptive immunity to
intracellular bacterial infection. Math Biosci Eng, 4(2):261–288, 2007a
5 Systems Biology of Tuberculosis 109

S. Marino, D. Sud, H. Plessner, L. P. Lin, J. Chan, J. L. Flynn, and D. E. Kirschner. Differences in

reactivation of tuberculosis induced from anti-TNF treatments are based on bioavailability in
granulomatous tissue. PLoS Comput Biol, 3(10):1909–1924, 2007b
K. Mdluli and M. Spigelman. Novel targets for tuberculosis drug discovery. Curr Opin Pharmacol,
6(5):459–467, 2006
P. Nunn, B. Williams, K. Floyd, C. Dye, G. Elzinga, and M. Raviglione. Tuberculosis control in
the era of HIV. Nat Rev Immunol, 5(10):819–826, 2005
J. A. Papin, T. Hunter, B. Ø. Palsson, and S. Subramaniam. Reconstruction of cellular signalling
networks and analysis of their properties. Nat Rev Mol Cell Biol, 6(2):99–111, 2005
M. Pellegrini, E. M. Marcotte, M. J. Thompson, D. Eisenberg, and T. O. Yeates. Assigning protein
functions by comparative genome analysis: Protein phylogenetic profiles. Proc Natl Acad Sci
USA, 96(8):4285–4288, 1999
G. Ramachandraiah and N. Chandra. Sequence and structural determinants of mannose recogni-
tion. Proteins, 39(4):358–364, 2000
K. Raman. Systems-level modelling and simulation of Mycobacterium tuberculosis: Insights for
drug discovery. PhD thesis, Indian Institute of Science, Bangalore, 2008
K. Raman and N. Chandra. Mycobacterium tuberculosis interactome analysis unravels potential
pathways to drug resistance. BMC Microbiol, 8:234, 2008
K. Raman and N. Chandra. Flux balance analysis of biological systems: Applications and chal-
lenges. Brief Bioinform, 10(4):435–449, 2009
K. Raman, P. Rajagopalan, and N. Chandra. Flux balance analysis of mycolic acid pathway: Tar-
gets for anti-tubercular drugs. PLoS Comput Biol, 1(5):e46, 2005
K. Raman, Y. Kalidas, and N. Chandra. targetTB: A target identification pipeline for Mycobac-
terium tuberculosis through an interactome, reactome and genome-scale structural analysis.
BMC Syst Biol, 2(1):109, 2008
K. Raman, R. Vashisht, and N. Chandra. Strategies for efficient disruption of metabolism in My-
cobacterium tuberculosis from network analysis. Mol Biosyst, 5:1740–1751, 2009
K. Raman, A. G. Bhat, and N. Chandra. A systems perspective of hostpathogen interactions: Pre-
dicting disease outcome in tuberculosis. Mol Biosyst, 6:516–530, 2010
K. Raman, Y. Kalidas, and N. Chandra. Model-driven drug discovery: Principles and practices,
Biological database modeling, pages 163–188. Artech House, New York, 2007
J. C. J. Ray and D. E. Kirschner. Requirement for multiple activation signals by anti-inflammatory
feedback in macrophages. J Theor Biol, 241(2):276–294, 2006
J. C. J. Ray, J. Wang, J. Chan, and D. E. Kirschner. The timing of TNF and IFN- signaling affects
macrophage activation strategies during Mycobacterium tuberculosis infection. J Theor Biol,
252(1):24–38, 2008
J. L. Reed, I. Famili, I Thiele, and B. Ø. Palsson. Towards multidimensional genome annotation.
Nat Rev Genet, 7(2):130–141, 2006a
J. L. Reed, T. R. Patel, K. H. Chen, A. R. Joyce, M. K. Applebee, D. D. Herring, O. T. Bui,
E. M. Knight, S. S. Fong, and B. Ø. Palsson. Systems approach to refining genome annotation.
Proc Natl Acad Sci USA, 103(46):17480–17484, Nov 2006b
C. M. Sassetti, D. M. Boyd, and E. J. Rubin. Genes required for mycobacterial growth defined by
high density mutagenesis. Mol Microbiol, 48(1):77–84, 2003
D. Schnappinger, S. Ehrt, M. I. Voskuil, Y. Liu, J. A. Mangan, I. M. Monahan, G. Dolganov,
B. Efron, P. D. Butcher, C. Nathan, and G. K. Schoolnik. Transcriptional adaptation of My-
cobacterium tuberculosis within macrophages: Insights into the phagosomal environment.
J Exp Med, 198(5):693–704, 2003
J. L. Segovia-Juarez, S. Ganguli, and D. E. Kirschner. Identifying control mechanisms of granu-
loma formation during M. tuberculosis infection using an agent-based model. J Theor Biol, 231
(3):357–376, 2004
I. Smith. Mycobacterium tuberculosis pathogenesis and molecular determinants of virulence. Clin
Microbiol Rev, 16(3):463–496, 2003
110 K. Raman and N. Chandra

M. Strong, T. G. Graeber, M. Beeby, M. Pellegrini, M. J. Thompson, T. O. Yeates, and

D. Eisenberg. Visualization and interpretation of protein networks in Mycobacterium tubercu-
losis based on hierarchical clustering of genome-wide functional linkage maps. Nucleic Acids
Res, 31(24):7099–7109, 2003
D. Sud, C. Bigbee, J. L. Flynn, and D. E. Kirschner. Contribution of CD8C T cells to control of
Mycobacterium tuberculosis infection. J Immunol, 176(7):4296–4314, 2006
K. Takayama, C. Wang, and G. S. Besra. Pathway to synthesis and processing of mycolic acids in
Mycobacterium tuberculosis. Clin Microbiol Rev, 18:81–101, 2005
J. Thakar, M. Pilione, G. Kirimanjeswara, E. T. Harvill, and R. Albert. Modeling systems-level
regulation of host immune responses. PLoS Comput Biol, 3(6):e109, 2007
T. Thomas. Boolean formalization of genetic control circuits. J Theor Biol, 42(3):563–585, 1973
K. D. Verkhedkar, K. Raman, N. Chandra, and S. Vishveshwara. Metabolome based reaction
graphs of M. tuberculosis and M. leprae: A comparative network analysis. PLoS One, 2(9):
e881, 2007
P. K. Vinod, B. Konkimalla, and N. Chandra. In-silico pharmacodynamics: Correlation of adverse
effects of H2-antihistamines with histamine N-methyl transferase binding potential. Appl Bioin-
form, 5(3):141–150, 2006
C. Von Mering, L. J. Jensen, M. Kuhn, S. Chaffron, T. Doerks, B. Krüger, B. Snel, and P. Bork.
STRING 7 – recent developments in the integration and prediction of protein interactions.
Nucleic Acids Res, 35(Database issue):358–362, 2007
J. S. Waddell and P. D. Butcher. Microarray analysis of whole genome expression of intracellular
Mycobacterium tuberculosis. Curr Mol Med, 7(3):287–296, 2007
S. J. Waddell, R. A. Stabler, K. Laing, L. Kremer, R. C. Reynolds, and G. S. Besra. The use of
microarray analysis to determine the gene expression profiles of Mycobacterium tuberculosis
in response to anti-bacterial compounds. Tuberculosis (Edinb), 84(3–4):263–274, 2004
S. J. Waddell, P. D. Butcher, and N. G. Stoker. Rna profiling in host-pathogen interactions. Curr
Opin Microbiol, 10(3):297–302, 2007
S. J. Waddell, K. Laing, C. Senner, and P. D. Butcher. Microarray analysis of defined Mycobac-
terium tuberculosis populations using rna amplification strategies. BMC Genomics, 9:94, 2008
J. E. Wigginton and D. E. Kirschner. A model to predict cell-mediated immune regulatory
mechanisms during human infection with Mycobacterium tuberculosis. J Immunol, 166(3):
1951–1967, 2001
World Health Organisation. Global tuberculosis control: Surveillance, planning, financing: WHO
report 2008. World Health Organisation, 2008 ISBN 978-9241563543
K. Yeturu and N. Chandra. PocketMatch: A new algorithm to compare binding sites in protein
structures. BMC Bioinform, 9:543, 2008
D. Young, J. Stark, and D. E. Kirschner. Systems biology of persistent infection: Tuberculosis as a
case study. Nat Rev Microbiol, 6(7):520–528, 2008
Chapter 6
Qualitative Analysis of Genetic Regulatory
Networks in Bacteria

Valentina Baldazzi, Pedro T. Monteiro, Michel Page, Delphine Ropers,

Johannes Geiselmann, and Hidde de Jong

6.1 Introduction

The functions of living organisms are controlled on the molecular level by networks
of biochemical reactions involving genes, mRNAs, proteins, metabolites, and sig-
naling molecules. The elucidation of the structure of these networks has much
progressed thanks to decades of work in genetics, molecular biology, and biochem-
istry, including the development of high-throughput experimental techniques. Most
of the time, however, it is not well understood how the dynamics of the networks
emerge from the reactions between the variety of its individual components. This
has called forth an increasing interest in the mathematical modeling of complex
cellular processes, in the context of a broader movement called systems biology
(Bettenbrock et al. 2005; Chen et al. 2004; Klipp et al. 2005; Leloup and Goldbeter
2003; Schoeberl et al. 2002).
In theory, it is possible to write down kinetic models of biochemical networks,
and study these by means of classical analysis and simulation tools. In practice, this
is not easy to achieve though, as the values of kinetic parameters are often only con-
strained to within a range spanning several orders of magnitude for most systems
of biological interest. Moreover, the models consist of a large number of variables,
are strongly non-linear, and include different timescales, which makes them difficult
to handle both mathematically and computationally. This has motivated the use of
approximations reducing the size and complexity of the models. Various approxi-
mations have been proposed in the literature, tailored to typical response functions
and timescale hierarchies found in genetic or metabolic regulation (de Jong and
Ropers 2006; Heijnen 2005; Heinrich and Schuster 1996; Millat et al. 2007; Okino
and Mavrovouniotis 1998; Papin et al. 2004; Pecou 2005; Roussel and Fraser 2001;
Savageau 2001; Thomas and Kaufman 2001). The approximations typically reduce
the dimension of the state and parameter space, and they simplify the mathematical
form of the equations.

V. Baldazzi ()
INRIA Grenoble – Rhône-Alpes, France
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 111
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 6,
c Springer Science+Business Media, LLC 2011
112 V. Baldazzi et al.

In this chapter, we discuss a class of so-called piecewise-linear (PL) models of

genetic regulatory networks, based on the use of step-function approximations of
the sigmoidal response functions involved in gene regulation. The PL models, origi-
nally introduced by Glass and Kauffman (1973) provide a coarse-grained picture of
the dynamics of genetic regulatory networks. They associate a protein concentration
variable to each of the genes in the network, and capture the switch-like character
of gene regulation by means of step functions that change their value at a threshold
concentration of the proteins. The advantage of using PL models is that the qual-
itative dynamics of the high-dimensional systems are relatively simple to analyze,
using an ordering of parameters rather than exact numerical values (Batt et al. 2008;
de Jong et al. 2004b). This makes the PL models a valuable tool for the analysis
of genetic regulatory networks, as demonstrated by several examples in bacteria
and higher organisms (Chaves et al. 2006; de Jong et al. 2004a; Halász et al. 2007;
Ropers et al. 2006; Sepulchre et al. 2007; Usseglio Viretta and Fussenegger 2004).
We will introduce the PL models in the context of the network of global regula-
tors controlling the carbon starvation response in the enterobacterium Escherichia
coli. In order to survive, E. coli cells constantly have to adapt their functioning to the
availability of carbon sources, essential for growth. The adaptation involves multiple
levels of regulation, from metabolic fluxes and enzyme activity to gene regulation
(Gutierrez-Rı́os et al. 2007; Hardiman et al. 2007; Kremling et al. 2009). In this
chapter, we focus in particular on the role of the global regulators of transcription,
such as CRP, Fis, DNA supercoiling, and RpoS. These global regulators form the
backbone of the network coordinating the long-term response of E. coli cells to
starvation conditions.
In Sect. 6.2 we briefly review the carbon starvation response in E. coli and the role
of the global regulators. Section 6.3 describes how we can systematically reduce
a classical kinetic model of the network of global regulators to a PL model. The
different mathematical and computational techniques available for the analysis of
PL models are discussed in Sects. 6.4 and 6.5, and illustrated on the E. coli model.
A particularly interesting property of the PL models is that they allow the parameter
space to be partitioned into regions with the same qualitative dynamics, using sets
of simple inequality constraints between parameters. This property is exploited in
Sect. 6.6 for the analysis of incompletely specified models. Section 6.7 discusses
the strengths and limitations of the PL models, and their relation to other qualitative
models, such as Boolean networks.

6.2 Carbon Starvation in E. coli

Under favorable environmental conditions, bacterial cells quickly grow and divide,
leading to an exponential increase of their biomass, called exponential phase. Upon
a variety of stress conditions, like the depletion of carbon sources, the bacteria
abandon exponential phase and enter a state in which cells stop dividing, capital-
izing upon the few available resources to maintain the basic metabolic functions
6 Analysis of Genetic Regulatory Networks 113

necessary for survival. This so-called stationary phase is rapidly reversed and
fast growth restored once the environmental conditions become favorable again
(Huisman et al. 1996).
Glucose is the preferred carbon source of E. coli. The adaptation of the bacteria
to the depletion of glucose from the growth medium is under the control of a large
and complex network of biochemical reactions involving genes, mRNAs, proteins,
metabolites, and signaling molecules. The role of the metabolic and signaling net-
works in the adaptation of E. coli to carbon source starvation have been extensively
studied (e.g., Bettenbrock et al. 2006; Chassagnole et al. 2002; Rohwer et al. 2000),
but much less has been done at the level of gene expression. In particular, it is not
well understood how the backbone of the network, formed by the global regulators
of transcription, coordinates the cascades of molecular events driving the growth
arrest of E. coli cells starved for glucose. These transcription factors respond di-
rectly or indirectly to glucose depletion, by controlling in a combinatorial fashion
the expression of a large number of genes involved in cellular adaptation and sur-
vival. In addition, they control each other’s expression, thus giving rise to a complex
regulatory network.
Figure 6.1 shows the network of key global regulators involved in the control
of the carbon starvation response. It includes well-known pleiotropic transcription
regulators, like the histone-like protein Fis, the catabolic repressor cAMPCRP (re-
sulting from the expression of genes crp and cya, and the activation of Cya by carbon
depletion), and the general stress response factor RpoS or  S (whose stability is
regulated by RssB). Changes in DNA topology and its dependence on the relative

Fig. 6.1 Network of global regulators involved in the carbon starvation response in E. coli (Ropers
et al., 2011, 2006). The graphical conventions (Kohn 2001) are explained in the legend
114 V. Baldazzi et al.

expression level of the genes gyrA, gyrB, gyrI, and topA are also considered, as the
three-dimensional structure of DNA modulates the transcription of a large number
of genes. Finally, stable RNAs expressed from the rrn operons are considered as
their amount provides a reliable indicator of the growth rate of the cell, being high
during an exponential phase and low during a stationary phase.
The synthesis and degradation of proteins and stable RNAs, and their regulation
by the global regulators, are central processes in the system of Fig. 6.1. However, the
network also involves other types of biochemical reactions, such as the formation of
protein complexes (GyrABGyrI), the modification of proteins by small molecules
(cAMPCRP), and enzymatic reactions (the synthesis of cAMP by Cya).

6.3 Modeling and Model Reduction

Due to the size and complexity of the network, the dynamics of the carbon starva-
tion response is difficult to predict intuitively and a mathematical model may be a
useful tool to clarify the global behavior of the system. Modeling may also allow
the formulation of hypotheses about missing components of the system, opening the
way to further experimental investigations.
A wide variety of modeling formalisms are available to describe networks of
biochemical reactions. The most common approach is based on ordinary differential
equations (ODEs) and describes the rate of change of the concentrations of proteins,
RNAs, metabolites, and other molecular species in the network. ODE models have
a solid foundation in the kinetic theory of biochemical reactions (Cornish-Bowden
1995; Heinrich and Schuster 1996), but their application requires knowledge of the
precise molecular mechanisms involved and quantitative information on parame-
ter values. For many systems, like the network of global regulators in E. coli, this
level of knowledge is not available. We therefore propose a model-reduction strat-
egy based on quasi-steady-state (QSS) and PL approximations, in order to obtain
models that are easier to handle mathematically and computationally.
We start from a detailed nonlinear ODE model that we build following standard
approaches from biochemistry. Figure 6.1a–b shows a small part of the ODE model,
concerning the activation of the transcription factor CRP, which we will use as ex-
ample network throughout the chapter (see Ropers et al. (2011) for the complete
model). Depending on their type, the reactions are assumed to follow mass-action,
Michaelis–Menten, or Hill kinetics. The resulting ODE system is highly nonlinear
and depends on a large number of parameters, whose values that are mostly un-
known within a range of several orders of magnitude.
A first reduction step is motivated by the fact that the processes described by
the model occur on widely differing timescales. Even in the absence of precise
information on the parameter values, we can to a first approximation distinguish
two different timescales. The first is a fast timescale, corresponding to reactions
like the formation of the complexes cAMPCRP and the metabolic reactions re-
sponsible for the synthesis, export, and degradation of cAMP. The second is a
6 Analysis of Genetic Regulatory Networks 115

slow timescale corresponding to the synthesis and degradation of the proteins Cya
and CRP. Based on timescale separation, the original model can be rewritten into
two distinct subsystems, corresponding to slow processes (protein synthesis and
degradation) and fast processes (complex formation and enzymatic reactions). The
fast and the slow processes are described by so-called fast and slow variables,
respectively.
To reduce the size and complexity of the nonlinear model, we apply the QSS
assumption (Heinrich and Schuster 1996). The QSS assumption is based on the
hypothesis that the fast variables instantaneously adapt to changes in the slow vari-
ables. This means that, after an initial transient, the dynamics of the fast system
can be well approximated by an algebraic function of the slow variables. The al-
gebraic function replaces the ODEs for the fast subsystem. Figure 6.1c compares
the temporal evolution of the concentration of cAMPCRP complex that is pre-
dicted by the nonlinear and QSS models, showing a good agreement between the
two solutions.
The QSS model thus obtained (Fig. 6.1d) is still nonlinear, but its dimension
is reduced to the slow variables Cya and CRP only, whose dynamics completely
defines the changes in the amount of cAMPCRP complex (the fast variable).
As a further approximation we simplify the mathematical form of the nonlinear-
ities of the system. This can be obtained by using PL differential equations. This
simplification is motivated by the fact that at the slow timescale of interest the reg-
ulation of gene expression gives rise to multivariate sigmoidal response functions,
which are conveniently approximated by algebraic expressions (e.g., sums or prod-
ucts) of step functions. Following the QSS approximation, the control exerted by
cAMPCRP on the synthesis of CRP protein can be rewritten as a function of Cya
and CRP concentrations (Fig. 6.1e). Gene expression appears to be maximal when
both Cya and CRP concentration are at high levels, within their physiological range,
and rapidly reduces for smaller concentrations of these proteins. Such a behavior can
be well approximated by a product of two-step functions, s C .xy ; y1 / and s C .xc ; c1 /,
where xy and xc represent the concentrations of Cya and CRP proteins, respectively,
and y1 and c1 the threshold concentrations of these proteins. The step-function
expressions are equivalent to logical functions, and account for the combinatorial
control of gene expression by regulatory proteins (Glass and Kauffman 1973; Mestl
et al. 1995; Thomas and d’Ari 1990). The PL model which includes these step func-
tions is shown in Fig. 6.1f. It states that, in the presence of a carbon starvation signal,
the gene crp will be expressed at its maximal rate if Cya and CRP protein concen-
trations are simultaneously larger than their threshold value, respectively y1 and c1 .
When this condition is not fulfilled, the synthesis of CRP protein is simply reduced
to its basal rate.
The above reduction strategy has been applied to the whole carbon starvation net-
work in Fig. 6.1, resulting in a PL model with 9 variables and 50 parameters (Ropers
et al. 2011). The quality of the QSS and PL approximations has been systematically
tested, using an ensemble approach and appropriate distance and correlation mea-
sures to compare the numerical solutions of the different models at each reduction
step (Ropers et al. 2011). The results show that, in comparison with conventional
116 V. Baldazzi et al.

b
xP y D y1 C y2 hC .xcm ; cm
2
; mcm /  y xy
C.k1 C k2 h .us ; s ; ms // xyp
xP c D c1 C c2 hC .xcm ; cm
1
; mcm /  c xc
Ck4 xcm  k4 xc xm
xP yp D k1 xy xp .k1 Ck2 hC .us ; s ; ms /Cy / xyp
xP cm D k4 xc xm  .k4 C c / xcm
xP m D k2 hC .us ; s ; ms / xyp C k4 xcm  k3 xm
k4 xc xm

x 10−6 d
c 1.8
xP y D y1 C y2  y xy
1.6
xP c D c1 C c2 hC .xcm ; cm
1
; mcm /  c xc
1.4 hC .us ; s ; ms / xc xy
xcm D
K4 K3 C hC .us ; s ; ms / xy
1.2 k3
x c∼m (M)

K3 D
k2
1 k4
K4 D
k4
0.8

0.6

0.4

0.2
0 100 200 300 400 500
Time (min)

f
xP y D y1 C y2  y xy
xP c D c1 Cc2 s C .xc ; c1 / s C .xy ; y1 / s C .us ; s /  c xc
6 Analysis of Genetic Regulatory Networks 117

nonlinear models, the PL approximations generally preserve the dynamics of the

carbon starvation response network, encouraging the use of PL models in situations
where the reference timescale is that of protein synthesis and degradation.

6.4 Qualitative Analysis of Dynamics

The dynamical properties of PL models have been the subject of active research
for more than three decades (Batt et al. 2007; Edwards 2000; Ghosh and Tomlin
2004; Glass and Kauffman 1973; Gouzé and Sari 2002; Mestl et al. 1995; Plahte
and Kjóglum 2005). The models have favourable mathematical properties due to
the use of step functions in the right-hand side of the differential equations. The
thresholds of the concentration variables define a subdivision of the state space, the
set of possible states of the system, into hyperrectangular regions.
In every region not located on a threshold plane, the step functions evaluate to
0 or 1, and the PL model reduces to an analytically solvable system of differential
equations. For instance, in the region 0 < xy < y1 and 0 < xc < c1 (named D1 in
Fig. 6.2), our example model (Fig. 6.1f) simply reduces to

xPy D ky1 C ky2  y xy

xPc D kc1  c xc ; (6.1)

as both s C .xc ; c1 / and s C .xy ; y1 / evaluate to 1 in D1 . It can be shown that all so-
lutions of (6.1) monotonically converge towards the point ..ky1 C ky2 /=y ; kc1 =c /.
The point ..ky1 C ky2 /=y ; kc1 =c / is a called a focal point of the system. Setting
further .ky1 C ky2 /=y > y1 and kc1 =c > c1 , as explained below, it immediately fol-
lows that Cya and CRP concentrations increase everywhere in region D1 . The above
example holds more generally: in every region D the original PL model simplifies

J
Fig. 6.1 Nonlinear ODE model that we build following standard approaches from biochemistry.
(a) Activation network. (b) Detailed ODE model for the activation network. xy , xc , xyp , xcm ,
and xm denote the concentrations of free Cya, free CRP, CyaATP, cAMPCRP, and free cAMP,
respectively, while us denotes the external glucose concentration. The total concentrations of Cya,
CRP, and cAMP are referred to as xy , xc , and xm , respectively. hC denotes a positive Hill function:
m
hC .x; ; m/ D xmxC m . (c) nonlinear and QSS solutions. The blue curve represents a solution
for the concentration variable xcm in the nonlinear model. The red curve is the corresponding
solution for xcm in the QSS model. After an initial transient the nonlinear solution rapidly re-
laxes to the QSS solution. (d) QSS model for the activation network. The model approximates the
nonlinear model by coupling the fast variable xcm to the slow variables xy and xc . (e) Plot of
hC .xcm ; cm ; mcm / as a function of xy and xc . The sigmoidal surface is approximated by the
product of step functions s C .xy ; y / s C .xc ; c / s C .us ; s /, with c , y , and s threshold values. s C
denotes a positive-step function: s C .x;  / D 1 if x >  , and 0 if x <  . (f) PL model for the
activation network with the step D function approximation of hC .xcm ; cm ; mcm /
118 V. Baldazzi et al.

Fig. 6.2 State-space and state transition for PL model. (a) State-space diagram for the PL model
of the subsystem controlling CRP activation (Fig. 6.1). The parameter values are assumed to satisfy
the following inequalities: c1 =c > c1 , . c1 C c2 /=c > c1 , and . y1 C y2 /=y > y1 . The arrows
indicate the direction of the vector field in each of the regions of the state space, while the solid line
represents a solution trajectory evolving towards a stable steady state in the region D22 . (b) State
transition graph representing the qualitative dynamics of the PL model. The state D 1 corresponds
to the region with the same name in part (a) of the figure, and satisfies the atomic properties
0  xc < c , 0  xy < y , dxc =dt > 0, and dxy =dt > 0. The labeled path (D1 , D6 , D11 ,
D12 , D17 , D22 / in the graph corresponds to the solution trajectory in (a). (c) Visualization of the
concentration bounds and signs of derivatives for each of the variables along the path in (b). It
shows that temporal-logic property (6.2) in the main text is satisfied by the graph. That is, the
graph contains a path in which the CRP concentration is increasing before becoming steady

into set of linear differential equations such that the systems locally behaves in a
qualitatively homogeneous way, all solutions monotonically converging towards a
focal point given by the ratio of (sum of) synthesis rate constants and (sum of) degra-
dation rate constants. These results can be generalized to the case of regions located
on a threshold plane (Batt et al. 2007; Gouzé and Sari 2002). Figure 6.2a shows the
subdivision of the state space for the PL model of the simple network controlling
the activation of CRP, as well as the dynamics in each of the regions.
The fact that the system behaves in a qualitatively homogeneous way in every
region motivates a discrete representation of the dynamics of the PL models, by
means of a so-called state transition graph. The states in the graph correspond to the
regions in the state space, while the transitions arise from solutions that enter one
region from another. Each transitions thus corresponds to a discrete event, namely
the crossing of a threshold by one or more concentration variables, possibly entailing
a change in the derivative (trend) of these variables.
6 Analysis of Genetic Regulatory Networks 119

The generation of the state transition graph from a PL model has been imple-
mented in the computer tool Genetic Network Analyzer (GNA) (Batt et al. 2005;
de Jong and Page 2008; de Jong et al. 2003), explicitly developed for the simula-
tion and the analysis of PL models. The state transition graph associated with the
example network is shown in Fig. 6.2b. The paths in the graphs denote sequences
of qualitative events, notably threshold crossings of the variables and changes in the
sign of the derivatives, as illustrated in panel (c) of Fig. 6.2.
Interestingly, it can be shown that the state transition graph, and thus the quali-
tative dynamics of the system, are completely determined by inequality constraints
defining the ordering between the threshold parameters i of a variable xi and the
values of the focal points for that variable. The definition of the inequality con-
straints between parameters can generally be inferred from available data in the
experimental literature or by intuitive reasoning, even in the absence of quantitative
information on parameter values. For example, the position of the focal point con-
centration . y1 C y2 /=y for protein Cya can be deduced just noticing that the ratio
. y1 C y2 /=y defines the maximum steady-state concentration that Cya can reach.
In order for Cya to have a regulatory effect on CRP synthesis, it is thus natural
to assume that . y1 C y2 /=y > y1 , otherwise the regulation of crp expression
would never be functional. GNA exploits the inequality constraints to symbolically
compute the attractors of the model and the states that are reachable from given
initial conditions for the concentration variables. This process is called qualitative
simulation.
We carry out a qualitative simulation of the carbon starvation network in Fig. 6.1.
Given the PL model of the E.coli network obtained in the previous section, we
define a set of inequality constraints on the parameters, as illustrated in Fig. 6.2
for the CRP activation subnetwork. The full specification of the PL model requires
more than 50 inequality constraints (Ropers et al. 2011). The PL model has been
used to address the question whether the interactions between the global regulators
in Fig. 6.1 are sufficient to explain the growth adaptation of E. coli. We analyze the
attractors of the system and simulate the response of E. coli to a depletion or sudden
availability of glucose.
Attractor analysis identifies the presence of two stable steady states. The first
steady state, characterized by the presence of the carbon starvation signal and by a
low level of stable RNAs, corresponds to stationary-phase conditions, whereas the
second steady state is representative of exponential phase, with a high level of stable
RNAs and no carbon starvation. Depending on the presence or absence of glucose
in the growth medium, the bacteria reach one of the attractors and are in exponential
or stationary phase, respectively. In order to better investigate the dynamics of the
transitions between growth phases, we simulate the qualitative behavior of the net-
work, starting from the steady state corresponding to exponential phase (stationary
phase) and perturbing the system by switching on (off) the carbon starvation signal.
An example of a path in the state transition graph produced by the qualitative sim-
ulation of the entry into stationary phase is shown in Fig. 6.3. The path traces how
the concentrations of the global regulators Fis and CRP evolve during the transition
from exponential to stationary phase. In response to a lack of glucose, the first key
120 V. Baldazzi et al.

Fig. 6.3 Path in the state transition graph. Path in the state transition graph produced by a quali-
tative simulation of the entry into stationary phase following carbon depletion, using the PL model
of the network in Fig. 6.1. The evolution of Cya (xy ), Fis (xf ), Crp (xc ), GyrAB (xg ), TopA (xt )
and the stable RNAs (xn ) is shown

event is a decrease in the concentration of Fis protein, followed by an progressive

increase of the level of CRP, strengthening the negative effect on fis expression. The
decrease in Fis levels ultimately leads to a downregulation of the level of stable
RNAs, witnessing the growth arrest as the cells enter stationary phase.
The predicted evolution of the Fis concentration is in agreement with exper-
imental data showing a 50-fold decrease in protein levels when E. coli cells
enter stationary phase (Ali Azam et al. 1999; Ball et al. 1992; Pratt et al. 1997).
Unfortunately, similar measurement are not available for CRP, but some experimen-
tal observations tend to confirm the model predictions: a low protein concentration
has been measured in the presence of glucose, whereas CRP is shown to accumulate
when glucose is absent (Ishizuka et al. 1994).
The question can be raised how general the above conclusions are. The simula-
tion results shown in Fig. 6.3 represent just one of the possible qualitative behaviors
of the system starting from the initial conditions, among a large number of paths
in the state-transition graph. Indeed, the entire state space for the PL model of the
network in Fig. 6.1 consists of the order of 1010 states whereas the subset reachable
from a particular growth phase of the bacteria contains approximately 103 states.
The size of the graph makes a general statement on the system dynamics difficult to
obtain by manual inspection of the individual simulations.
6 Analysis of Genetic Regulatory Networks 121

6.5 Formal Verification of Network Properties

The formal verification field provides powerful methods to deal with the analysis of
large models of cellular interaction networks by specifying dynamical properties of
interest, as statements in a formal language called temporal logic (Antoniotti et al.
2003; Batt et al. 2005; Bernot et al. 2004; Calder et al. 2005; Chabrier-Rivier et al.
2004; Fisher et al. 2007). Efficient, so-called model-checking, algorithms exist to
determine whether these statements are satisfied by the model, without the need of
explicitly checking all the paths in the graph (Clarke et al. 1999).
Temporal logic queries are meant to capture patterns in the temporal evolution of
the system dynamics, like the relation and the ordering between qualitative events
(e.g., increasing/decreasing of a protein concentration) or the reachability of an at-
tractor of the system (a steady state or a limit cycle). For example, given the model
for CRP activation, we can ask whether there exists a path in which the concentra-
tion of CRP increases before becoming steady. Such a property can be specified in
temporal logic by the following formula

EF .dxc =dt > 0 ^ EF .dxc =dt D 0// (6.2)

that is satisfied by the state transition graph in Fig. 6.2, as witnessed by the path
in panel (c). The definition of specific queries can be difficult for nonexpert users
though, due to the existence of a variety of temporal logic operators that can be ap-
plied alone or in combination, with sometimes subtle differences in the meaning of
the resulting formulas. Alternatively, high-level query templates can be defined that
link the intuitive description of biological properties to temporal logic (Monteiro
et al. 2008).
We apply model-checking tools to the analysis of the state transition graphs
produced by qualitative simulation of the carbon starvation response in E. coli, in-
vestigating the role of the mutual inhibition between the genes fis and crp. When
found in isolation, a mutual-inhibition motif has been shown to lead to bistability
(Gardner et al. 2000): in particular, it excludes the simultaneous presence of high
concentrations of the proteins CRP and Fis, as shown in the simulation in Fig. 6.3.
The question can be asked if this is always true, i.e., if this motif maintains its
functionality when embedded in a large network like the one in Fig. 6.1. If this is
the case, it should be impossible for the concentrations of Fis and CRP to be si-
multaneously high or low at steady state. Reformulate into a statement in temporal
logic, in particular the logic CTL, following (Monteiro et al. 2008), this property
reads:

: EF .high.xc / ^ high.xf / ^ isSteadyState/

^ : EF .low.xc / ^ low.xf / ^ isSteadyState/: (6.3)

The property states that, starting from the initial conditions, there does not exist
a future state (:EF ) where xc and xf simultaneously have a high value and the
122 V. Baldazzi et al.

system is in steady state (high.xc / ^ high.xf / ^ isSteadyState), and similarly, there

does not exist a future state where xc and xf simultaneously have a low value and the
system is in steady state (low.xc / ^ low.xf / ^ isSteadyState). The predicates high
and low are defined in terms of inequality constraints. For instance, high.xf / can
be defined as xf > f4 for a high threshold f4 of Fis in the PL model.
Formal verification using the open-source model-checking tool NuSMV (Cimatti
et al. 2002), in combination with GNA, reveals that the property holds true for all
possible paths in the state-transition graph. The positive loop motif involving fis and
crp thus remains functional inside the large network of Fig. 6.1, with the conse-
quent rearrangements of gene expression levels following the transition from a state
with high Fis and low CRP concentrations (characteristic of the exponential growth
phase) to a situation with low Fis and high CRP (typical of the stationary growth
phase).
The carbon starvation response network also contains a negative feedback loop,
involving the genes gyrAB, topA, and fis (Fig. 6.1). GyrAB and TopA are responsi-
ble for the control of the intracellular level of DNA supercoiling. GyrAB is a gyrase
protein which supercoils the DNA structure, whereas the topoisomerase TopA re-
laxes it. An increase of the DNA supercoiling level stimulates expression of Fis,
which in turn decreases the supercoiling level, by stimulating topA expression and
inhibiting gyrAB expression. This negative loop allows the bacteria to rapidly adjust
protein concentrations and resume growth once nutrients become available again.
Qualitative simulations have shown the emergence of damped oscillations of Fis
and GyrAB concentrations in response to a sudden carbon upshift. Is this property
bound to occur following a carbon upshift? We used formal verification methods
to check whether the carbon upshift is a sufficient condition for the occurrence of
damped oscillations (Monteiro et al. 2008), i.e.:

AG ..us < s / ) AF .isOscillatoryState.xf ; xg //: (6.4)

The predicate isOscillatoryState.xf ; xg / denotes that a state is part of a cycle in

the graph in which the concentrations of Fis and GyrAB oscillate.
The model checker returned true for the query: the model thus predicts that
cells necessarily resume growth through damped oscillations after a carbon upshift.
However, no experimental data are currently available to confirm or disconfirm this
prediction.

6.6 Model Completion

As explained in Sect. 6.4, the state-transition graph defining the qualitative dynam-
ics is completely determined by the inequality constraints on the parameters. Each
set of inequality constraints defines a region in the parameter space in which the
system has the same qualitative dynamics (Batt et al. 2008). In case it is not possi-
ble to completely specify the ordering of the parameters, formal verification can
6 Analysis of Genetic Regulatory Networks 123

Parameter space ODE Parameter space PL

Model reduction

P̂
P

Model completion

Property is
true

Model checking

P̂
P̂
Property is
Parameter space PL Parameter space PL
false
+ (additional) inequality constraints + inequality constraints

Fig. 6.4 Overall scheme for model completion by means of formal verification tools. A nonlinear
ODE model with parameter space P is reduced to a PL model with parameter space PO . This
parameter space can be partitioned into regions with the same qualitative dynamics, represented by
the state transition graph, using inequality constraints on the parameter values. By testing properties
that have to be satisfied by the model, certain regions of the parameter space can be eliminated,
thus further constraining the parameter values

provide a method to discriminate between alternative hypotheses (Bernot et al.

2004). Different sets of inequality constraints can be enumerated and the dynam-
ics of the corresponding PL models tested against known biological properties by
means of model checking. Sets of inequality constraints that are proven to be incon-
sistent with the properties are rejected and further analysis can be performed on the
biologically-meaningful models left. This model completion approach allows the
exhaustive exploration of the parameter space of the PL models (Fig. 6.4).
As an example, we investigate the role played by the general stress factor RpoS
in the control of the DNA supercoiling level and in the growth adaptation of E. coli
cells following carbon depletion. The global regulatory network in Fig. 6.1 extends
a previously published PL model (Ropers et al. 2006), which failed to predict the
observed decrease of the DNA supercoiling level during the transition from the ex-
ponential to the stationary growth phase. We therefore refined the description of the
control of DNA supercoiling level by including in the model the general stress fac-
tor RpoS and its regulators, since a rpoS deletion mutant has been shown deficient
in the regulation of plasmid topology in stationary phase (Reyes-Dominguez et al.
2003). How does RpoS control the DNA supercoiling level during the adaptation
of E. coli cells to carbon depletion? And is the entry into stationary phase always
preceded by an accumulation of RpoS?
124 V. Baldazzi et al.

Model completion was carried out by taking a partially specified PL model of

the network in Fig. 6.1. The model does not constrain the ordering of the threshold
values at which Fis, GyrAB, GyrI, and TopA control the DNA supercoiling level
and at which Fis stimulates the expression of the stable RNAs. This yields a total
of 1296 different PL models, for which we first test the ability to capture the steady
states corresponding to stationary and exponential phase. All the PL models have
two stable steady states, one corresponding to stationary-phase conditions, with the
carbon starvation signal present and a predicted low level of stable RNAs, indicat-
ing the absence of cellular growth. In the other steady state the signal is switched
off, but in many cases the state is not representative for exponential-phase condi-
tions since a low level of stable RNAs is predicted. Because we are interested only
in biologically meaningful models, we have studied which sets of inequality con-
straints confer a good prediction of the exponential-phase conditions. It appears that
one simple inequality is sufficient to discriminate between the models. In particular,
the threshold value at which Fis protein inhibits the expression of gyrAB should be
higher than the concentration needed to activate the production of stable RNAs, to
allow the accumulation of the latter in presence of glucose. This relative ordering
of Fis binding affinities for the promoter regions of the stable RNAs and the gy-
rase has never been described in the experimental literature and thus constitutes an
interesting prediction for further investigation.
The addition of the inequality constraint between the above-mentioned thresh-
olds for Fis reduces the admissible sets of inequality constraints to 258. For all 258
PL models, we have run a qualitative simulation of the response to carbon starvation
in exponential phase. All the models tested predict that the accumulation of RpoS
to high levels is essential for the downregulation of the DNA supercoiling level, as
tested by means of model checking. RpoS may thus be responsible for the relaxation
of DNA topology at the entry into stationary phase. On the contrary, in none of the
258 models, the accumulation of RpoS to high levels or a low DNA supercoiling
level are necessary for the cell’s growth arrest. Only the accumulation of the com-
plex cAMPCRP is essential for this process, confirming the key role played by the
mutual inhibition of fis and crp.
These results thus allow us to clarify our picture of the carbon starvation response
by assigning a role to the individual global regulator in the network. A striking result
is that the 258 PL models predict essentially the same effect of RpoS accumulation
on the DNA supercoiling level and the downregulation of the stable RNAs. This
suggests that these properties are robust for a large range of parameter values con-
trolling the DNA supercoiling level.

6.7 Conclusions

The modeling of the regulatory networks controlling the response of bacteria to ex-
ternal perturbations leads to large and complex systems of nonlinear ODEs. These
models are difficult to study in the absence of quantitative information on parameter
6 Analysis of Genetic Regulatory Networks 125

values, which has motivated the use of various approximations adapted to typical
response functions and timescale hierarchies found in genetic and metabolic reg-
ulation. In this chapter, we have reviewed one such model reduction approach,
based on QSS and PL approximations. Under the condition that the dominant
timescale of interest is that of protein synthesis and degradation, which involves
sigmoidal response functions that can be suitably approximated by step functions;
these approximations are expected to give rise to simplified models of complex net-
works with dynamics close to that of classical ODE models (Ropers et al. (2011),
see also Chaves et al. (2006), Davidich and Bornholdt (2008)).
The PL models obtained after model reduction have been well-studied in the
literature, following their original introduction by Glass and Kauffman (Batt et al.
2007; Edwards 2000; Ghosh and Tomlin 2004; Glass and Kauffman 1973; Gouzé
and Sari 2002; Mestl et al. 1995; Plahte and Kjóglum 2005). The qualitative dy-
namics of the PL models can be represented by a state-transition graph, consisting
of states and transitions between states. A major result is that this graph is invariant
for large sets of parameter values, defined by inequality constraints that can be in-
ferred from the experimental literature. Moreover, the state transition graph can be
computed from the inequality constraints by means of simple, symbolic rules. In or-
der to support its application to large and complex genetic regulatory networks, the
analysis of the PL models has been implemented in the computer tool GNA. As the
graphs become too large to be analyzed by hand, GNA can be coupled with model-
checking tools for the automatic verification of dynamic properties of the network.
This approach can be particularly useful for the analysis of incompletely specified
models, as summarized in Fig. 6.4.
The use of PL models is justified by the intuition that, to a first approximation,
genes can be considered logical switches that transform continuous inputs – i.e.,
the concentration of regulatory proteins – into discrete outputs – i.e., the activation
state of the genes (Sugita 1963; Yuh et al. 1998). Instead of developing this intuition
for models with continuous time and concentration variables, one could also de-
cide to employ discrete models. The major example of this approach is the Boolean
network formalism developed by Kauffman, Thomas, and others (Kauffman 1969,
1993; Thomas 1973; Thomas and d’Ari 1990). The application of Boolean net-
works rests on the assumption that a gene is either active or inactive, and that genes
change their activation state synchronously. For the purpose of modeling actual ge-
netic regulatory networks, these assumptions are usually too strong. In response to
this problem, more general formalisms with multivalued activation states and asyn-
chronous transitions have been proposed and successfully applied to the analysis of
complex developmental regulatory networks (Gonzalez et al. 2006; Mendoza et al.
1999; Sánchez and Thieffry 2003). The advantage of Boolean networks and their
generalizations is that they provide a convenient way to express the logic of gene
expression regulation. However, they have difficulty in treating dynamic properties
of genetic regulatory networks taking place at the threshold of activation or inacti-
vation of a gene, where steady states may be located (Casey et al. 2006).
PL models may not be appropriate when the systems under study do not in-
volve sigmoidal response functions, and are therefore not well approximated by
126 V. Baldazzi et al.

step functions. This is particularly so for metabolism and cell signaling, where
Michaelis-Menten and mass-action kinetics are common. Other types of approxi-
mations may apply in these cases though, such as lin-log models (Heijnen 2005),
power-law models (Savageau 2001), and piecewise multi-affine models (Belta and
Habets 2006). Even when step-functions approximations are appropriate, they may
not be sufficiently precise. The methods described in this chapter strive at capturing
the qualitative dynamics of the networks, but when quantitative precision is sought,
it may be necessary to use less drastic approximations of the sigmoidal response
functions.
In this chapter, we have reviewed the use of PL models in the context of the mod-
eling of the network of global regulators involved in the carbon starvation response
in E. coli. Qualitative simulation has been used to obtain predictions of the behavior
of networks which is currently not yet well understood by biologists. While some
of the predictions help clarifying the role of particular regulatory mechanisms (the
mutual inhibition of fis and crp), others concern phenomena that have not yet been
experimentally investigated (the occurrence of damped oscillations after a nutrient
upshift and the relative ordering of Fis binding affinities for different promoter re-
gions). The latter two predictions are particularly interesting from a biological point
of view, because they generate new questions and suggest further experiments.
The basic motivation for the use of the PL models is the current absence of
precise and quantitative information on kinetic parameters. The advantage of the
reduction of ODE models to PL models is that they allow a quick scan of the quali-
tative dynamics of the system, without numerical information on parameter values.
This provides a first insight into the global dynamics of the system, which is interest-
ing in itself but also yields structural and parametric constraints that may guide the
development of more detailed kinetic models. In particular, it may orient the difficult
problems of system identification and parameter estimation from experimental data
(Bettenbrock et al. 2006; Gardner et al. 2003; Kuepfer et al. 2007; Quach et al. 2007;
Moles et al. 2003; Porreca et al. 2008; Ronen et al. 2002; van Riel and Sontag 2006;
Vilela et al. 2008; Westra et al. 2007; Zwolak et al. 2005).

Acknowledgment VB, DR, JG, HdJ were supported by the European commission under project
EC-MOAN (FP6-2005-NEST-PATH-COM/043235). PM was partially supported by FCT program
(PhD grant SFRH/BD/32965/2006 to PTM) and PDTC program (project PTDC/EIA/71587/2006).

References

T. Ali Azam, A. Iwata, A. Nishimura, S. Ueda, and A. Ishihama. Growth phase-dependent variation
in protein composition of the Escherichia coli nucleoid. J. Bacteriol., 181 (20): 6361–6370,
1999.
M. Antoniotti, A. Policriti, N. Ugel, and B. Mishra. Model building and model checking for bio-
chemical processes. Cell Biochem. Biophys., 38 (3): 271–286, 2003.
C. A. Ball, R. Osuna, K. C. Ferguson, and R. C. Johnson. Dramatic changes in Fis levels upon
nutrient upshift in Escherichia coli. J. Bacteriol., 174 (24): 8043–8056, 1992.
6 Analysis of Genetic Regulatory Networks 127

G. Batt, D. Ropers, H. de Jong, J. Geiselmann, R. Mateescu, M. Page, and D. Schneider. Valida-

tion of qualitative models of genetic regulatory networks by model checking: Analysis of the
nutritional stress response in Escherichia coli. Bioinformatics, 21 (Suppl 1): i19–i28, 2005.
G. Batt, B. Yordanov, R. Weiss, and C. Belta. Robustness analysis and tuning of synthetic gene
networks. Bioinformatics, 23 (18): 2415–2422, 2007.
G. Batt, H. de Jong, M. Page, and J. Geiselmann. Symbolic reachability analysis of genetic regu-
latory networks using discrete abstractions. Automatica, 44 (4): 982–989, 2008.
C. Belta and L. C. G. J. M. Habets. Controlling a class of nonlinear systems on rectangles. IEEE
Trans. Autom. Control, 51 (11): 1749–1759, 2006.
G. Bernot, J.-P. Comet, A. Richard, and J. Guespin. Application of formal methods to biological
regulatory networks: Extending Thomas’ asynchronous logical approach with temporal logic.
J. Theor. Biol., 229 (3): 339–348, 2004.
K. Bettenbrock, S. Fischer, A. Kremling, K. Jahreis, T. Sauter, and E. D. Gilles. A quantitative
approach to catabolite repression in Escherichia coli. J. Biol. Chem., 281 (5): 2578–2584, 2005.
K. Bettenbrock, S. Fischer, A. Kremling, K. Jahreis, T. Sauter, and E. D. Gilles. A quantitative
approach to catabolite repression in Escherichia coli. J. Biol. Chem., 281 (5): 2578–2584, 2006.
M. Calder, V. Vyshemirsky, D. Gilbert, and R. Orton. Analysis of signalling pathways using the
PRISM model checker. In G. Plotkin, editor, Proc. of CMSB, pages 179–190, Edinburgh, Scot-
land, 2005.
R. Casey, H. de Jong, and J.-L. Gouzé. Piecewise-linear models of genetic regulatory networks:
Equilibria and their stability. J. Math. Biol., 52 (1): 27–56, 2006.
N. Chabrier-Rivier, M. Chiaverini, V. Danos, F. Fages, and V. Schächter. Modeling and querying
biomolecular interaction networks. Theor. Comput. Sci., 325 (1): 25–44, 2004.
C. Chassagnole, N. Noisommit-Rizzi, J. W. Schmid, K. Mauch, and M. Reuss. Dynamic modeling
of the central carbon metabolism of Escherichia coli. Biotechnol. Bioeng., 79 (1): 53–73, 2002.
M. Chaves, E. D. Sontag, and R. Albert. Methods of robustness analysis for Boolean models of
gene control networks. IET Syst. Biol., 153 (4): 154–167, 2006.
K. C. Chen, L. Calzone, A. Csikasz-Nagy, F. R. Cross, B. Novak, and J. J. Tyson. Integrative
analysis of cell cycle control in budding yeast. Mol. Biol. Cell, 15 (8): 3841–3862, 2004.
A. Cimatti, E. Clarke, E. Giunchiglia, F. Giunchiglia, M. Pistore, M. Roveri, R. Sebastiani, and
A. Tacchella. NuSMV2: An OpenSource tool for symbolic model checking. In D. Brinksma
and K. G. Larsen, editors, 14th International Conference on Computer Aided Verification (CAV
2002), volume 2404 of LNCS, pages 359–364. Springer, Berlin, 2002.
E. M. Clarke, O. Grumberg, and D. A. Peled. Model Checking. MIT, Boston, MA, 1999.
A. Cornish-Bowden. Fundamentals of Enzyme Kinetics. Portland Press, London, revised edition,
1995.
M. Davidich and S. Bornholdt. The transition from differential equations to Boolean networks:
A case study in simplifying a regulatory network model. J. Theor. Biol., 255: 269–277, 2008.
H. de Jong and M. Page. Search for steady states of piecewise-linear differential equation models
of genetic regulatory networks. ACM/IEEE Trans. Comput. Biol. Bioinform., 5 (2): 208–222,
2008.
H. de Jong and D. Ropers. Strategies for dealing with incomplete information in the modeling of
molecular interaction networks. Brief. Bioinform., 7 (4): 354–363, 2006.
H. de Jong, J. Geiselmann, C. Hernandez, and M. Page. Genetic Network Analyzer: Qualitative
simulation of genetic regulatory networks. Bioinformatics, 19: 336–344, 2003.
H. de Jong, J. Geiselmann, G. Batt, C. Hernandez, and M. Page. Qualitative simulation of the
initiation of sporulation in B. subtilis. B. Math. Biol., 66 (2): 261–299, 2004a.
H. de Jong, J.-L. Gouzé, C. Hernandez, M. Page, T. Sari, and J. Geiselmann. Qualitative simulation
of genetic regulatory networks using piecewise-linear models. B. Math. Biol., 66 (2): 301–340,
2004b.
R. Edwards. Analysis of continuous-time switching networks. Phys. D, 146 (1–4): 165–199, 2000.
J. Fisher, N. Piterman, A. Hajnal, and T. A. Henzinger. Predictive modeling of signaling crosstalk
during C. elegans vulval development. PLoS Comput. Biol., 3 (5): e92, 2007.
128 V. Baldazzi et al.

T. S. Gardner, C. R. Cantor, and J. J. Collins. Construction of a genetic toggle switch in escherichia

coli. Nature, 403 (6767): 339–342, 2000.
T. S. Gardner, D. di Bernardo, D. Lorenz, and J. J. Collins. Inferring genetic networks and identi-
fying compound mode of action via expression profiling. Science, 301 (5629): 102–105, 2003.
R. Ghosh and C. J. Tomlin. Symbolic reachable set computation of piecewise affine hybrid au-
tomata and its application to biological modelling: Delta-Notch protein signalling. IET Syst.
Biol., 1 (1): 170–183, 2004.
L. Glass and S. A. Kauffman. The logical analysis of continuous non-linear biochemical control
networks. J. Theor. Biol., 39 (1): 103–129, 1973.
A. Gonzalez Gonzalez, A. Naldi, L. Sánchez, D. Thieffry, and C. Chaouiya. GINsim: a software
suit for the qualitative modelling, simulation and analysis of regulatory networks. Biosystems,
84 (2): 91–100, 2006.
J.-L. Gouzé and T. Sari. A class of piecewise linear differential equations arising in biological
models. Dyn. Syst., 17 (4): 299–316, 2002.
R. M. Gutierrez-Rı́os, J. A. Freyre-Gonzalez, O. Resendis, J. Collado-Vides, M. Saier, and
G. Gosset. Identification of regulatory network topological units coordinating the genome-wide
transcriptional response to glucose in Escherichia coli. BMC Microbiol., 7: 53–53, 2007.
A. Halász, V. Kumar, M. Imielinski, C. Belta, O. Sokolsky, S. Pathak, and H. Rubin. Analysis of
lactose metabolism in E. coli using reachability analysis of hybrid systems. IET Syst. Biol., 1
(2): 130–48, 2007.
T. Hardiman, K. Lemuth, M. A. Keller, M. Reuss, and M. Siemann-Herzberg. Topology of
the global regulatory network of carbon limitation in Escherichia coli. J. Biotechnol., 132:
359–374, 2007.
J. J. Heijnen. Approximative kinetic formats used in metabolic network modeling. Biotechnol.
Bioeng., 91 (5): 534–545, 2005.
R. Heinrich and S. Schuster. The regulation of cellular systems. Chapman & Hall, New York, 1996.
G. W. Huisman, M. M. Siegele D. A., Zambrano, and Kolter R. Morphological and physiological
changes during stationary phase. In F. C. Neidhardt, R. Curtiss III, J. L. Ingraham, E. C. C.
Lin, K. B. Low, B. Magasanik, W. S. Reznikoff, M. Riley, M. Schaechter, and H. E. Umbarger,
editors, Escherichia coli and Salmonella: Cellular and Molecular Biology, pages 1672–1682.
ASM, Washington D.C., 1996.
H. Ishizuka, A. Hanamura, T. Inada, and H. Aiba. Mechanism of the down-regulation of cAMP
receptor protein by glucose in Escherichia coli: role of autoregulation of the crp gene. EMBO
J., 13 (13): 3077–3082, 1994.
S. A. Kauffman. The origins of order: Self-organization and selection in evolution. Oxford
University Press, New York, 1993.
S. A. Kauffman. Metabolic stability and epigenesis in randomly constructed genetic nets. J. Theor.
Biol., 22 (3): 437–467, 1969.
E. Klipp, B. Nordlander, R. Krüger, P. Gennemark, and S. Hohmann. Integrative model of the
response of yeast to osmotic shock. Nat. Biotechnol., 23 (8): 975–982, 2005.
K. W. Kohn. Molecular interaction maps as information organizers and simulation guides. Chaos,
11 (1): 84–97, 2001.
A. Kremling, S. Kremling, and K. Bettenbrock. Catabolite repression in Escherichia coli- a com-
parison of modelling approaches. FEBS J., 276: 594–602, 2009.
L. Kuepfer, M. Peter, U. Sauer, and J. Stelling. Ensemble modeling for analysis of cell signaling
dynamics. Nat. Biotechnol., 25 (9): 1001–1006, 2007.
J.-C. Leloup and A. Goldbeter. Toward a detailed computational model for the mammalian circa-
dian clock. Proc. Nat. Acad. Sci. USA, 100 (12): 7051–7056, 2003.
L. Mendoza, D. Thieffry, and E. R. Alvarez-Buylla. Genetic control of flower morphogenesis in
Arabidopsis thaliana: A logical analysis. Bioinformatics, 15 (7–8): 593–606, 1999.
T. Mestl, E. Plahte, and S. W. Omholt. A mathematical framework for describing and analysing
gene regulatory networks. J. Theor. Biol., 176 (2): 291–300, 1995.
T. Millat, E. Bullinger, J. Rohwer, and O. Wolkenhauer. Approximations and their consequences
for dynamic modelling of signal transduction pathways. Math. Biosci., 207 (1): 40–57, 2007.
6 Analysis of Genetic Regulatory Networks 129

C. G. Moles, P. Mendes, and J. R. Banga. Parameter estimation in biochemical pathways:

A comparison of global optimization methods. Genome Res., 13 (11): 2467–2474, 2003.
P. T. Monteiro, D Ropers, R Mateescu, A. T. Freitas, and H. de Jong. Temporal logic patterns
for querying dynamic models of cellular interaction networks. Bioinformatics, 24: i227–i233,
2008.
M. S. Okino and M. L. Mavrovouniotis. Simplification of mathematical models of chemical reac-
tion systems. Chem. Rev., 98 (2): 391–408, 1998.
J. A. Papin, J. Stelling, N. D. Price, S. Klamt, S. Schuster, and B. O. Palsson. Comparison of
network-based pathway analysis methods. Trends Biotechnol., 22 (8): 400–405, 2004.
E. Pecou. Splitting the dynamics of large biochemical interaction networks. J. Theor. Biol., 232
(3): 375–384, 2005.
E. Plahte and S. Kjóglum. Analysis and generic properties of gene regulatory networks with graded
response functions. Phys. D, 201 (1): 150–176, 2005.
R. Porreca, S. Drulhe, H. de Jong, and G. Ferrari-Trecate. Structural identification of piecewise-
linear models of genetic regulatory networks. J. Comput. Biol., 15 (10): 1365–1380, 2008.
T. S. Pratt, T. Steiner, L. S. Feldman, K. A. Walker, and R. Osuna. Deletion analysis of the fis
promoter region in Escherichia coli: antagonistic effects of integration host factor and Fis. J.
Bacteriol., 179 (20): 6367–6377, 1997.
M. Quach, N. Brunel, and F. d’Alché Buc. Estimating parameters and hidden variables in non-
linear state-space models based on ODEs for biological networks inference. Bioinformatics, 23
(23): 3209–3216, 2007.
Y. Reyes-Dominguez, G. Contreras-Ferrat, J. Ramirez-Santos, J. Membrillo-Hernandez, and M. C.
Gomez-Eichelmann. Plasmid DNA supercoiling and gyrase activity in Escherichia coli wild-
type and rpoS stationary-phase cells. J. Bacteriol., 185 (3): 1097–1100, 2003.
J. M. Rohwer, N. D. Meadow, S. Roseman, H. V. Westerhoff, and P. W. Postma. Understanding
glucose transport by the bacterial phosphoenolpyruvate:glycose phosphotransferase system on
the basis of kinetic measurements in vitro. J. Biol. Chem., 275 (45): 34909–34921, 2000.
M. Ronen, R. Rosenberg, B. I. Shraiman, and U. Alon. Assigning numbers to the arrows: Parame-
terizing a gene regulation network by using accurate expression kinetics. Proc. Natl. Acad. Sci.
USA, 99 (16): 10555–10560, 2002.
D. Ropers, H. de Jong, M. Page, D. Schneider, and J. Geiselmann. Qualitative simulation of the
carbon starvation response in Escherichia coli. Biosystems, 84 (2): 124–152, 2006.
D. Ropers, V. Baldazzi, and H. de Jong. Model reduction using piecewise-linear approximations
preserves dynamic properties of the carbon starvation response in Escherichia coli. IEEE/ACM
Transactions on Computational Biology and Bioinformatics, 8 (1): 166–181, 2011.
M. R. Roussel and S. J. Fraser. Invariant manifold methods for metabolic model reduction. Chaos,
11 (1): 196–206, 2001.
L. Sánchez and D. Thieffry. Segmenting the fly embryo: A logical analysis of the pair-rule cross-
regulatory module. J. Theor. Biol., 224 (4): 517–537, 2003.
M. A. Savageau. Design principles for elementary gene circuits: Elements, methods, and examples.
Chaos, 11 (1): 142–159, 2001.
B. Schoeberl, C. Eichler-Jonsson, E.-D. Gilles, and G. Mller. Computational modeling of the dy-
namics of the MAP kinase cascade activated by surface and internalized EGF receptors. Nat.
Biotechnol., 20 (4): 370–375, 2002.
J.-A. Sepulchre, S. Reverchon, and W. Nasser. Modeling the onset of virulence in a pectinolytic
bacterium. J. Theor. Biol., 44 (2): 239–257, 2007.
M. Sugita. Functional analysis of chemical systems in vivo using a logical circuit equivalent: II.
The idea of a molecular automaton. J. Theor. Biol., 4: 179–192, 1963.
R. Thomas. Boolean formalization of genetic control circuits. J. Theor. Biol., 42 (3): 563–585,
1973.
R. Thomas and R. d’Ari. Biological feedback. CRC, Boca Raton, FL, 1990.
R. Thomas and M. Kaufman. Multistationarity, the basis of cell differentiation and memory:
II. Logical analysis of regulatory networks in terms of feedback circuits. Chaos, 11 (1):
180–195, 2001.
130 V. Baldazzi et al.

A. Usseglio Viretta and M. Fussenegger. Modeling the quorum sensing regulatory network of
human-pathogenic Pseudomonas aeruginosa. Biotechnol. Prog., 20 (3): 670–678, 2004.
N. A. W. van Riel and E. D. Sontag. Parameter estimation in models combining signal transduction
and metabolic pathways: The dependent input approach. IET Syst. Biol., 153 (4): 263–274,
2006.
M. Vilela, I. Chou, S. Vinga, A. Vasconcelos, E. Voit, and J. Almeida. Parameter optimization in
S-system models. BMC Syst. Biol., 2: 35, 2008.
R. L. Westra, G. Hollanders, G. J. Bex, M. Gyssens, and K. Tuyls. The identification of dynamic
gene-protein networks. In K. Tuyls, R. Westra, Y. Saeys, and A. Nowé, editors, Proc. KDECB
2006, volume 4366 of LNCS, pages 157–170. Springer, Berlin, 2007.
C.-H. Yuh, H. Bolouri, and E. H. Davidson. Genomic cis-regulatory logic: Experimental and com-
putational analysis of a sea urchin gene. Science, 279: 1896–1902, 1998.
J. W. Zwolak, J. J. Tyson, and L. T. Watson. Parameter estimation for a mathematical model of the
cell cycle in frog eggs. J. Comput. Biol., 12 (1): 48–63, 2005.
Chapter 7
Modeling Antibiotic Resistance in Bacterial
Colonies Using Agent-Based Approach

James T. Murphy and Ray Walshe

7.1 Introduction

Multi-drug resistance in Staphylococcus aureus bacteria has become a major health

care challenge in recent decades. Infections with S. aureus had a mortality rate of
over 80% before the introduction of antibiotics in the early 1940s (Skinner and
Keefer 1941). The first ˇ-lactam antibiotic, penicillin, was introduced into clinical
use during the early 1940s and dramatically reduced the mortality rate associated
with these infections (Chain et al. 1993). However, the widespread use of ˇ-lactam
antibiotics led to the rapid expansion of resistant strains of bacteria. As a result,
today greater than 95% of all S. aureus isolates have been found to possess resistance
to penicillin, and methicillin resistance is estimated to be in 40–60% of clinical
isolates in the USA and the UK (Levy and Marshall 2004; Neu 1992).
Advances in cell and molecular biology techniques in the latter half of the twen-
tieth century have led to rapid increases in our knowledge about the basic cellular
processes involved in antibiotic resistance. This development has allowed a more
fine-grained approach to be taken in investigating the spread of resistance in popu-
lations of bacteria. However, the overall population response to antibiotic treatment
is often a function of a diverse range of interacting components. A sound theoretical
understanding of the systems of interactions taking place (e.g. between antibiotic
molecules and cell surface proteins) is required to develop strategies to minimise
the spread of antibiotic resistance. The rapid development in the areas of pharma-
cokinetics and pharmacodynamics studies in recent years is a response to this need
to understand the complex dynamics that contribute to the bacterial response to drug
treatment (Ambrose et al. 2007).
There has also been a strong development of studies into the population and
growth dynamics of bacterial populations using various computational modeling ap-
proaches. The most common of these are the mathematical models which describe

J.T. Murphy ()

Centre for Scientific Computing and Complex Systems Modeling, School of Computing,
Dublin City University, Dublin 9, Ireland
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 131
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 7,
c Springer Science+Business Media, LLC 2011
132 J.T. Murphy and R. Walshe

the population as a whole using state variables. These have been important for
developing insights into parameters at the population level that influence the de-
velopment of the colony (Lacasta et al. 1999). They allow an integrated view of
colony development to be realised, and to identify key determinants of population
growth and development.
However, models that use global parameters for a population imply that they are
in a homogeneous, mixed environment. In nature, bacteria often form highly hetero-
geneous colonies where there can be significant localised variations in the chemical
environment such as nutrient availability, temperature, ion concentrations and pH
(Devine 2004). A different approach, which allows the heterogeneity of popula-
tions to be explicitly modeled, is the agent-based (or individual-based) modeling
approach. For this, the individual bacterial cells represent the fundamental units of
the simulation. This approach where the parameters are defined for the bacterial
cells rather than for the population as a whole is called a “bottom-up” approach.
The properties of a colony emerge from the set of interactions of a population of
heterogeneous bacterial agents.
This approach is what was taken in our research to model the interactions of bac-
terial cells with antibiotic molecules. An agent-based model was developed, called
the Micro-Gen Bacteria Simulator, which simulates the life cycle of bacteria grow-
ing in culture and their interactions with various molecules including antibiotics
(Walshe 2006; Murphy and Walshe 2006; Murphy et al. 2007). The model engine
was designed to be adaptable to represent different species and strains of bacteria
using basic cellular information. For our research, a model of methicillin-resistant
Staphylococcus aureus (MRSA) bacteria was developed, since they are currently of
great clinical significance and there is a wealth of biological information available
about them.
The model incorporates representations of the two main antibiotic resistance
mechanisms characteristic of MRSA (ˇ-lactamase enzymes and PBP2a, see below).
It includes a quantitative model of the interactions between ˇ-lactam antibiotics
and MRSA bacteria using kinetic rules for these reactions derived from experi-
mental studies (Zygmunt et al. 1992; Lu et al. 1999). By simulating the individual
molecular interactions of antibiotic molecules and bacteria, and scaling this up to
large population sizes using the agent-based approach, Micro-Gen can be used to
explore the emergent dynamics that contribute to antibiotic resistance in bacterial
populations.

7.1.1 MRSA Antibiotic Resistance Mechanisms

Antibiotic resistance refers to the ability of a microorganism to resist the effects of

an antibiotic. Examples of mechanisms of antibiotic resistance include the synthe-
sis of antibiotic-degrading enzymes (e.g. ˇ-lactamase) and modifications to drug
targets such as the penicillin-binding proteins (PBPs) in bacterial cell membranes.
There is a complex range of interacting factors that influence the development and
7 Modeling Antibiotic Resistance in Bacterial Colonies 133

spread of antibiotic resistance in pathogenic bacteria. Even though the appearance

of resistant strains of bacteria is thought to be an ancient evolutionary event, the
fitness cost associated with resistance mechanisms limited their proliferation before
the introduction of antibiotics (Wright 2007). The widespread use of antibiotics has
resulted in a significant positive selective pressure for resistant strains, particularly
in the clinical setting (Levy and Marshall 2004).
There are several resistance mechanisms that have been observed in bacteria in
response to antibiotic exposure. These include mechanisms to limit the uptake of
antibiotic molecules into the bacterial cells. For example, the deletion of porin pro-
teins in gram-negative bacteria to block the passage of antibiotic molecules through
the bacterium’s outer membrane, and the activation of efflux exporter proteins to
“pump out” the antibiotic (Walsh 2000; Fisher et al. 2005). Perhaps the most clin-
ically significant resistance mechanisms, with respect to ˇ-lactam antibiotics, in
gram-positive pathogens such as S. aureus bacteria are the expression of enzymes
called ˇ-lactamases, and alterations to the molecular targets (PBPs) of the antibi-
otics (Fig. 7.1).
Penicillin resistance in S. aureus is mediated by the production and release of
an enzyme, called ˇ-lactamase, which hydrolytically cleaves the ˇ-lactam ring
structure present in penicillin and other ˇ-lactam antibiotics. ˇ-lactamase was first
discovered in Escherichia coli bacteria in 1940, and ˇ-lactamase-expressing S. au-
reus bacteria were also isolated soon afterwards (Abraham and Chain 1940; Bondi

Fig. 7.1 Diagram of the two principal antibiotic resistance mechanisms observed in MRSA bac-
teria. (a) Expression of alternate form of penicillin-binding protein PBP2, called PBP2a, with
reduced binding affinity for antibiotic. (b) Production and release of ˇ-lactamase enzyme which
cleaves and inactivates antibiotic molecules
134 J.T. Murphy and R. Walshe

and Dietz 1945; Kirby 1944). ˇ-lactamases are thought to have evolved long before
the clinical introduction of ˇ-lactam antibiotics. However, they only became broadly
distributed across many bacterial species under selective pressure from widespread
antibiotic use (Fisher et al. 2005).
The interaction between ˇ-lactams in the environment and a cell surface signal-
transducer protein, BlaR1, triggers the process whereby expression of ˇ-lactamase
occurs (Fuda et al. 2005; Lewis et al. 1999). Most of the ˇ-lactamase enzyme pro-
duced by the cell is secreted into the extracellular milieu, while some remains bound
to the cytoplasmic membrane of the cell (Nielsen and Lampen 1982). Re-repression
of ˇ-lactamase expression occurs when the antibiotic concentration in the environ-
ment decreases and BlaR1 is no longer auto-activated (Zhang et al. 2001).
MRSA strains were first discovered soon after the introduction of methicillin
in 1959 (Fuda et al. 2005). Methicillin had been introduced to treat infections of
penicillin-resistant S. aureus which had become a significant health concern at that
stage. MRSA strains were isolated as early as 1961, and they have steadily in-
creased in frequency since then in response to selective pressure (Eriksen 1961).
MRSA bacteria contain a gene called mecA, which encodes a penicillin-binding
protein, PBP2a, which circumvents the mode of action of ˇ-lactam antibiotics. Nor-
mal S. aureus cells produce four types of membrane-bound transpeptidase proteins
called penicillin-binding proteins (PBPs 1–4), which assemble and regulate the fi-
nal stages of cell wall biosynthesis. The mode of action of the ˇ-lactam antibiotics
involves the acylation of a catalytic residue in the transpeptidase active site of PBPs
which results in the inhibition of their cell-wall cross-linking function, a crucial step
in cell wall assembly during bacterial cell division.
Therefore, when the bacterial cell is unable to correctly assemble the cell wall
due to binding of the PBPs by antibiotic, this can result in inhibition of cell division
(bacteriostatic effect) or even cell death (bactericidal effect). However, the prod-
uct of the mecA gene, PBP2a, does not bind the ˇ-lactam moiety readily because
the approach to the active site is sterically encumbered. When an MRSA organ-
ism is exposed to ˇ-lactam antibiotics, PBP2a confers resistance by supplementing
its transpeptidase activity (cell-wall cross-linking) to the transglycosylase function
of native PBPs during cell wall synthesis (Fuda et al. 2005). As a result, cell wall
assembly can continue to occur even in the presence of antibiotic.

7.1.2 Overview of Modeling Approaches

The power of computational modeling approaches lies not so much in their abil-
ity to make predictions (since some degree of experimental validation is needed to
confirm any predictions) but rather in their ability to give new insights into the un-
derlying mechanistic basis for the observed biological phenomena. A computational
model was defined by Volker Grimm as a “purposeful representation” of an entity
or system whose “purpose is to capture the essence of a problem and explore differ-
ent solutions of it” (Grimm 1999). The most important role of a model is therefore
7 Modeling Antibiotic Resistance in Bacterial Colonies 135

to aid in our understanding of a particular process. From this perspective, all the
different modeling approaches share the same principal aim, though they may differ
in the assumptions and tools that are used.
The most common mathematical approaches to modeling bacterial population
growth have been the use of ordinary differential equations (ODEs) and partial
differential equations (PDEs). ODE models are often used in systems biology be-
cause they are computationally efficient and mathematically robust, and can be used
to develop an integrated view of biological systems. Extensions to the basic ODE
methods have also been developed over the years such as stochastic ODEs and com-
partmentalised ODE models. PDEs are used to model processes that have spatial as
well temporal dependencies.
Mathematical population models use global parameters or state variables to de-
scribe the growth and development of a bacterial colony as a unit (Grimson and
Barker 1994; Lacasta et al. 1999). These “top-down” approaches have the advantage
that they are computationally efficient and less parameter-rich than more low-level
approaches. However, with state variable approaches it is sometimes difficult to
trace back the system behaviour to the behaviour of the individual agents. For exam-
ple, this approach does not explicitly explain the underlying factors that lead to the
population exhibiting a particular growth rate or carrying capacity (Grimm 1999).
However, it is important because it provides an appropriate integrated view of the
population behaviour.
Another modeling technique commonly used in systems biology of bacterial
colonies is the theory of cellular automata (CA) (Ben-Jacob et al. 1994). This
approach is a powerful tool for representing both temporal and spatiotemporal
processes in biological systems. In CA models, the environment of the model is
represented by a discrete lattice/grid where the states of the components evolve syn-
chronously in discrete time steps according to a set of rules. The CA simulation
Conways Game of Life was one of the first examples of computer applications in
biology (Gardner 1970). This model consisted of randomly placed cells on a square
lattice and simulated birth, death and interactions according to pairwise interaction
rules which used Boolean logic conditions. CA methods have continued to be devel-
oped since then and applied to a diverse range of problems in computational systems
biology (Materi and Wishart 2007).

7.1.3 Agent-Based Modeling Approach

An alternative approach to modeling bacterial growth and development, and the

main subject of this chapter, is the agent-based (or individual-based) modeling ap-
proach (Ginovart et al. 2002; Kreft et al. 1998). The distinguishing characteristic of
the agent-based approach is that the properties of the individual cells, rather than
the colony as a whole or a subgroup of it, are modeled. This “bottom-up” approach
allows a finer-grained analysis than the other techniques, connecting local changes
at the cellular level to the overall patterns of population growth. The agent-based
136 J.T. Murphy and R. Walshe

approach shares some of the strengths of the cellular automata modeling approach,
in which it is able to explicitly model both temporal and spatiotemporal processes.
For this reason, it is particularly amenable to modeling processes such as chemo-
taxis, diffusion and pattern formation in bacterial colonies. However, it represents
an even finer-grained approach than CA in which the individual biological entities
being modeled are explicitly represented by unique software agents.
Complex agent-based systems consist of many similar and simple components.
Often the system as a whole has a complex behaviour that is more than the sum
of its constituent parts. Agents can be used to conceptualise and implement such a
software modeling application. An agent-oriented approach lends itself naturally to
the complexity of modeling bacterial interactions. For the purposes of this simula-
tion, the definition of an agent will be adopted as given by Jennings et al. (1998)
in which an agent is described as a “computer based system, situated in some envi-
ronment which is capable of flexible, autonomous action to meet its goals”. Within
this definition, there are three fundamental concepts – situatedness, autonomy and
flexibility.
“Situatedness” in this context means that the agent is in some environment from
which it receives sensory input and is capable of modifying the environment. So, for
example, in terms of a bacterial simulation, a bacterium in an agar plate senses nutri-
ent and can consume it. “Autonomy” in this context implies that the system should
be able to act independently from human intervention. The system has local control
over its own state, i.e. once it is set up and running no intervention is required. “Flex-
ibility” means that the system should be re-active, pro-active and social. Re-active:
the system should respond in a timely fashion to changes within the environment.
Pro-active: the system should bring about changes in the environment to meet its
design objectives. Social: the system should interact, collaborate or compete, where
necessary to complete its design objectives.
The theory of autonomous agents is a useful approach for the modeling of bac-
terial cell colonies as it allows large-scale population models to be derived from
simple rules dictating the growth and interactions of the individual bacterial cells
of the population. The model produces global information about population growth
in different environmental conditions using basic information about the cell biology
of bacteria. The agent-based approach was chosen to explicitly model the hetero-
geneity in environmental conditions, for example between the interior and exterior
of a colony, and between individual bacterial cells. This would not be so amenable
using another approach such as a simpler mass action model. In complex microbial
communities, there can be highly heterogeneous localised niches where the chem-
istry varies significantly over small distances, and the agent-based approach allows
us to take this into account. It could also be used to model other mechanisms such
as phase and antigenic variation in bacteria that result in heterogenic phenotypes
within a clonal population (van der Woude and Baumler 2004).
Some of the challenges of this approach include the fact that it can sometimes
require more parameters than a state variable approach since the individual entities
are explicitly modeled and it also may become susceptible to empirical knowledge
(Ambrose et al. 2007). In other words, it is sometimes difficult to obtain accurate
7 Modeling Antibiotic Resistance in Bacterial Colonies 137

parameters for individual cells. However, advances in experimental techniques

related to studying individual cells will help in building more accurate models at
the cellular level (Elfwing et al. 2004; Niven et al. 2006). Using appropriate aggre-
gation of parameters, and cognisant of its limitations, the agent-based approach can
be used as a powerful tool for tracing back system behaviour to that of its individual
components.
The agent-based approach and higher-level mathematical approaches are not mu-
tually exclusive but rather it is envisaged that they should complement each other in
studies of population dynamics. Mathematical models allow theories at the systems
level to be developed by providing a general conceptual framework for the popu-
lation as a whole. Meanwhile, the agent-based approach allows the overall system
properties to be related back to the individuals of the population. Agent-based mod-
els can be more computationally intensive than high-level mathematical approaches
such as ODEs, because each individual of a population is explicitly modeled. How-
ever, in cases where the population expresses a high degree of heterogeneity, both
spatially and between individuals, the agent-based approach can represent a pow-
erful tool for exploring how this heterogeneity contributes to the overall system
dynamics.

7.2 Micro-Gen Bacterial Simulator

In this section, there is a detailed overview of the main structure and components
of the agent-based software model Micro-Gen, developed to model the growth and
interactions of bacterial cells with antibiotics in vitro (Murphy and Walshe 2006;
Murphy et al. 2007). It was developed from existing work in the laboratory carried
out previously on an agent-based model called the Bacteria–Antibiotic Interaction
Tool (BAIT) (Walshe 2006). BAIT consisted of a simple model of bacterial growth
and interactions with antibiotic molecules in a discrete two-dimensional grid envi-
ronment using the Java programming language. It demonstrated the feasibility of
this approach for examining the individual dynamics of antibiotic molecules inter-
acting with bacterial cells in culture. Micro-Gen represents a significant expansion
and re-design of the original BAIT tool, to build a more realistic representation of
bacterial growth and development in culture and a quantitative model of their inter-
actions with antibiotics.
Micro-Gen is coded in the CCC object-oriented programming language. The
individual microorganisms are represented by software agents which store physical
traits of the bacterial cells as well as behavioural rules associated with them. The
modular nature of the program means that functionalities/characteristics specific
to particular bacterial species can be readily incorporated into the basic cellular
model. A key component of the model is the ability to quantitatively model an-
tibiotic molecules and their interactions with the bacterial cells. These interactions
between the antibiotic molecules and targets in the bacterial cell are governed by de-
fined kinetic parameters specific to the type of antibiotic and bacterial strain being
138 J.T. Murphy and R. Walshe

modeled. This allows a quantitative model of antibiotic interactions with bacteria to

be built up and their pharmacokinetic properties to be investigated.
The model incorporates two important antibiotic resistance mechanisms used by
bacterial cells against antimicrobial agents which form the cornerstone of the an-
tibiotic arsenal: special enzymes released by bacteria, called ˇ-lactamases, which
degrade the antibiotic molecules, and reduced binding affinities between the antibi-
otics and receptors in the bacterial cells (see Sect. 7.1.1). These antibiotic resistance
mechanisms are of great clinical concern as their development and spread across
many species of bacteria has led to the erosion of the efficacy of many commonly
prescribed antibiotics, in particular penicillin and its derivatives.

7.2.1 Environment

The simulated culture environment consists of a discrete, two-dimensional grid con-

taining diffusible elements such as antibiotics, nutrients (glucose) and ˇ-lactamase
enzymes (Fig. 7.2). The environment was restricted to a two-dimensional plane to
minimise the computational burden of the program. Each discrete grid position is
referred to as a “patch” (to differentiate it from a bacterial “cell”) and contains
variables to record the levels of the various molecular components in it. It also in-
cludes pointers to bacterial agents that currently occupy the patch. A discretised
implementation of Fick’s first law of diffusion is used to calculate the movement of
molecules between adjacent patches down local concentration gradients (Ginovart
et al. 2002). The amount of substance exchanged between two adjacent patches

Fig. 7.2 Overview of two-dimensional grid environment in Micro-Gen. Each grid element is
referred to as a “patch” and contains various simulation components including bacterial cells,
ˇ-lactamase enzymes, antibiotic molecules and nutrients
7 Modeling Antibiotic Resistance in Bacterial Colonies 139

a b
i=1 i=2 i=3 i=1 i=2 i=3

i=5 i=5

i=4 [bio]i i=6 i=4 [bio]i i=6

p(xi)= 1.0 − p(xi)= 1.0 −
9 9
Σ[bio]
n=1
n Σ[bio]
n=1
n

i=7 i=8 i=9 i=7 i=8 i=9

Fig. 7.3 Diffusion and overcrowding algorithms. (a) Diffusion algorithm (Fick’s First Law of
diffusion) applied to molecules. D diffusion coefficient, Mol concentration difference. (b) Over-
crowding algorithm applied to bacterial cells. p(xi) probability of bacterial agent moving to patch
i, [bio]i bacterial biomass in patch i

is calculated as the concentration difference (Mol) multiplied by a user-defined

diffusion coefficient, D (Fig. 7.3a). p
When patches are diagonally adjacent to one
another, a diffusion rate modifier (1/ 2) is applied.

7.2.2 Bacterial Agents

The bacterial agents are autonomous entities with a set of behavioural rules that
determine how they interact with the environment and parameters associated with
them recording details such as their energy state (or nutrient reserve). In order to
keep the model simple, the internal subcellular processes of the bacteria are not
explicitly modeled. This minimises the number of parameters associated with each
agent, thus optimising the performance of the program and avoiding the problems
associated with a parameter-rich model (Ambrose et al. 2007). The main input pa-
rameters for the bacterial agents, with some sample values for S. aureus, are listed
in Table 7.1.
The bacterial cells are assumed to be simple, independent entities that pas-
sively absorb nutrients from the environment and grow and divide asexually as
they accumulate more nutrients. The following sections detail a number of pa-
rameters associated with the bacterial agents. These parameters are necessarily an
abstract representation of the complex underlying mechanisms that contribute to
the bacterial cell behaviour. For example, the complex nutrient uptake processes in
a cell are reduced to a simple “nutrient intake” parameter. By doing this, we lose
140 J.T. Murphy and R. Walshe

Table 7.1 Input parameters Input parameter Input value

for bacterial agents in
Biomass threshold for division 10,000
Micro-Gen model
Nutrient Intake (b.u. loop1 ) 10.0
Survival cost (b.u. loop1 ) 0.2
Stationary phase relative metabolic rate 0.2
Lag phase length (min) 66
ˇ-lactamase production rate (M s1 ) 3:28  107
ˇ-lactamase production cost (b.u.) 0.1
Antibiotic intake (M) 6:0  108
Kinetic parameters (k2 , Kd , kcat , KM ), see Table 7.2
Sample values for Staphylococcus aureus species
included b.u. D biomass units

some of the insight that may be gained into the subtle subcellular dynamics of the
nutrient uptake mechanism. The emphasis during the design phase was to minimise
the number of parameters and aggregate where possible. This results in a simple,
lean model that is memory-efficient and thus can be scaled up to very large popula-
tion sizes even on limited computing hardware.

7.2.2.1 Growth Parameters

The biomass of the cell is represented in the simulation by simulation units called
“biomass units”. Bacterial agents increase their biomass by absorbing nutrient from
the immediate environment. A “nutrient intake” parameter specifies the rate of nu-
trient absorption by the cell. There is also a “survival cost” associated with normal
metabolic activities of the cell, and this is subtracted from the cell biomass in each
time step. Reproduction is triggered when the cellular biomass increases beyond a
certain threshold for division (“biomass threshold for division”). The cell divides
into two identical daughter cells, approximately half the size of the original cell, in
a process known as binary fission.
In order to estimate the nutrient intake rate and biomass threshold for division, the
model was fitted to an experimentally determined bacterial doubling time of 29 min
(generation time of S. aureus strain BB255) (Ender et al. 2004). The survival cost
parameter influences the length of the stationary phase of the growth cycle. This is
modifiable by the user to represent different stationary phase lengths.
A higher survival cost results in a shorter stationary phase because cells enter the
death phase more quickly. A survival cost of 2% of the rate of nutrient intake was
chosen for the test simulations recorded here. This represents a level which does not
limit the exponential phase of bacterial growth. However, in nature this would vary
considerably between different strains, and for a more detailed quantitative repre-
sentation of the growth curve of a particular strain this would need to be estimated
from experimental studies.
Another parameter associated with the bacterial agents is the “stationary phase
relative metabolic rate”. This parameter is included to take into account the state
7 Modeling Antibiotic Resistance in Bacterial Colonies 141

of reduced metabolic activity that bacterial cells enter when they are subjected to
severe stress such as nutrient deprivation. The “survival cost” parameter, mentioned
above, is reduced to the proportion specified. The principal effect of this parameter
is to extend the length of the stationary phase before the bacteria enter the death
phase. It represents the bacterium’s ability to preserve itself in hostile, nutrient-
deprived conditions by shutting down non-essential metabolic activities. A sample
value of 0.2 is used in our test simulations for illustrative purposes; however, as
with the “survival cost” parameter, this would need to be experimentally estimated
to give a quantitatively accurate representation of the length of the stationary phase.
The “lag phase length” parameter determines the length of time it takes for the
bacteria to adapt to their new environment at the start of the simulation. During this
phase, bacterial cells adapt to their conditions by synthesising the required cellular
components to process the nutrients in their new environment. Their rate of nutri-
ent intake increases until the maximum normal intake rate is achieved. There is a
random element introduced by the fact that the bacteria are initialised with different
internal nutrient levels at the start of the simulation.

7.2.2.2 Antibiotic Resistance Mechanisms

The exposure of bacteria to ˇ-lactam antibiotics triggers the synthesis and release
of the ˇ-lactamase enzyme into the extracellular milieu (Fig. 7.4). The ˇ-lactamase
production rate is estimated by varying it over a range of values and calculating the
minimum inhibitory concentration (MIC) of penicillin G at each value (using exper-
imentally determined kinetic parameters for penicillin G). The MIC is the minimum
concentration of an antibiotic that results in complete inhibition of bacterial growth
in vitro.
The ˇ-lactamase production rate is calibrated so that the simulated MIC equals
the experimentally determined MIC for penicillin G versus the particular bacterial

Cell wall
Extracellular milieu

Bound
antibiotic

S. aureus
(inside cell)

Cytoplasmic
membrane

β-lactamase enzymes

Fig. 7.4 Diagram of release of ˇ-lactamase enzymes from Staphylococcus aureus cell. Production
of ˇ-lactamase is induced by binding of ˇ-lactam antibiotics to a cell surface signal transducer
protein (BlaR1). Most of the ˇ-lactamase enzyme is secreted into the extracellular milieu, while
some remains bound to the cytoplasmic membrane of the cell
142 J.T. Murphy and R. Walshe

strain being modeled. For the test simulations of MRSA that we have carried out,
Type A and Type C ˇ-lactamase-producing MRSA strains were modeled. The ex-
perimentally determined MIC results of Norris et al. (1994) for penicillin G (Type
A MIC D 72.1 g/ml; Type C D 47.9 g/ml) were used to estimate the ˇ-lactamase
production rates in the simulations.
Table 7.1 also lists some parameters for the interactions between the bacterial
agent and antibiotic molecules. The “antibiotic intake” parameter determines the
rate at which free antibiotic is depleted in the patch by absorption across the cell
wall of the bacteria. There are also two kinetic parameters (k2 , Kd ) which deter-
mine the rate at which the antibiotic molecules bind to PBPs in the cell membrane.
Two kinetic parameters (kcat , KM ) are included to describe the interactions between
antibiotic molecules and ˇ-lactamase enzymes in the environment. Values for these
kinetic parameters were taken directly from experimental literature. A detailed ex-
planation of these kinetic parameters is included in Sects. 7.2.3 and 7.2.4.

7.2.2.3 Overcrowding Algorithm

In the case of S. aureus bacteria, which are non-motile, an overcrowding algorithm

is applied to take into account the physical size constraints of a single patch in the
environment. The area of each patch is configured to represent approximately 1 m2
of medium. An overcrowding algorithm is applied when more bacteria occupy the
patch than can be physically accommodated. For example, the estimated cell diam-
eter of a newly divided S. aureus cell is 0.5 m (Giesbrecht et al. 1998). Therefore,
when more than four such cells occupy a single patch, the overcrowding algorithm
is applied (Fig. 7.3b). The probability, p(xi ), of a bacterial cell in an overcrowded
patch being moved to an adjacent patch i is inversely proportional to the relative
bacterial biomass in the adjacent patch. The direction a cell is moved is determined
by sampling from the resultant probability distribution of the surrounding patches.

7.2.3 Antibiotics

The antibiotic level in each patch of the environment is stored as a variable which is
subject to diffusion between patches according to Fick’s First Law of diffusion (see
Sect. 7.2.1). There is also a half-life associated with the antibiotic, derived from the
biological literature, which determines the rate at which the molecule degrades over
time (Wishart et al. 2006). Bacterial agents that are in the same patch as the antibi-
otic will absorb it according to their antibiotic intake rate. The ˇ-lactam antibiotics
inhibit bacterial cell division by binding to proteins in the cell membrane called
PBPs which are necessary for cell division (Fig. 7.5). If a significant proportion of
PBPs in the cell are inactivated, the bacteria will be unable to reproduce and cell
death may occur (Giesbrecht et al. 1998).
7 Modeling Antibiotic Resistance in Bacterial Colonies 143

Fig. 7.5 Diagram of a ˇ-lactam antibiotic binding to penicillin-binding protein (PBP) in bacterial
cell membrane. PBP is an essential component for correct cell wall synthesis during bacterial
cell division. However, binding and acylation of the PBP by antibiotic result in inhibition of this
function

The simulator includes a quantitative model to estimate the rate of binding of a

particular antibiotic to PBPs in the cell membrane, using experimentally calculated
kinetic parameters for the reaction. The reaction is a pre-steady state reaction, and
therefore the kinetic parameters k2 (rate of inactivation of PBP2a) and Kd (dissocia-
tion constant) are used to describe it. The ratio of these values (k2 /Kd ), or the second
order rate constant, is a common measure of the antibiotic efficacy at inhibiting PBP
function. The proportion of PBP that is inactivated per second, ka (the apparent first
order rate constant), at a given drug concentration can be calculated as a function of
these parameters (7.1) (Chambers et al. 1994).

k2 ŒAb
ka D (7.1)
Kd C ŒAb

The interactions between the ˇ-lactam antibiotic molecules and PBP2a are ex-
plicitly represented in the simulation, but not the interactions with the other PBPs
present in the bacterial cell membrane (PBPs 1 – 4). This is sufficient for represent-
ing MRSA because the limiting reaction for antibiotic efficacy is that with PBP2a,
which has a binding affinity for ˇ-lactams that is significantly reduced compared to
the other PBPs. This minimises the level of complexity caused by introducing more
empirical parameters into the model.
Values for the kinetic parameters, k2 and K d , of PBP2a were obtained from ex-
perimental studies in the biological literature (Table 7.2) (Fuda et al. 2004; Graves-
Woodward and Pratt 1998; Lu et al. 1999). It is possible to estimate the proportion
of PBP2a deactivated by antibiotic each time step using these parameters.
144 J.T. Murphy and R. Walshe

Table 7.2 Experimentally determined kinetic parameters for ˇ-lactam antibiotics versus
Type A and Type C ˇ-lactamase-producing MRSA
Type A MRSA Type C MRSA
Parameter Pen G Amp Ceph Pen G Amp Ceph
kcat (s1 ) 171.0 308.0 0.015 210.0 355.0 0.095
K M (M) 51.1 255.0 6.8 55.9 122.0 5.2
k2 (s1 ) 0.18500 0.00470 0.00115 0.18500 0.00470 0.00115
K d (M) 15,400 495 586 15,400 495 586
Pen G Penicillin G, Amp Ampicillin, Ceph Cephalothin Values derived from biological
literature

7.2.4 ˇ-Lactamase Enzymes

Each bacterial agent has a true/false flag for ˇ-lactamase expression associated with
it. When bacterial agents occupy a patch where antibiotic is present, ˇ-lactamase
expression is induced (i.e. the flag is changed to true) and there is an exponential
increase in the ˇ-lactamase production rate until the maximum rate is reached after
approximately 80 min (Lewis et al. 1999). This represents the activation of gene
expression mediated by antibiotic binding to the signal-transducer protein BlaR1
and the ensuing time lag during which ˇ-lactamase synthesis is initiated.
The ˇ-lactamase which is then released into the local patch is subject to dif-
fusion according to Fick’s First Law of diffusion (Sect. 7.2.1). Also, like the
antibiotics, the enzyme has a half-life parameter determining its rate of degrada-
tion over time in the environment. If there is no longer any antibiotic left in the
bacterium’s patch, re-repression of ˇ-lactamase expression occurs and in the model
this is represented by the software flag for expression by the bacterial agent be-
ing changed to false (corresponding to the BlaR1 protein in nature no longer being
auto-activated).
In order to achieve a quantitative representation of the reactions between
ˇ-lactamase enzymes and antibiotic molecules, information from kinetic stud-
ies is used. The reaction is described using Michaelis–Menten kinetics, with the
reaction rate, V, calculated as the rate at which antibiotic is cleaved (or de-activated)
by the enzyme per second (7.2):

kcat ŒEt ŒAb

V D (7.2)
KM C ŒAb

There are two principal kinetic parameters used as input to calculate the reaction
rate: the turnover rate, kcat , and the Michaelis constant, KM . The ratio of these values
(kcat /KM ) is often used as a measure of enzyme efficiency (Zygmunt et al. 1992).
As with the previous kinetic parameters, values for these were derived from the
biological literature (Table 7.2) (Zygmunt et al. 1992). [Ab] and [E]t refer to the
concentrations of antibiotic ˇ-lactamase enzyme (sum of both free and occupied
enzyme) in the local patch, respectively.
7 Modeling Antibiotic Resistance in Bacterial Colonies 145

7.2.5 Program Flow Structure

The structure of the program consists of an initialisation stage during which the
main components of the environment, the patches and the bacterial agents are ini-
tialised. This is followed by the principal program loop during which the main tasks
of the simulation are carried out in successive discrete time steps until certain exit
conditions (e.g. no bacteria left) are reached (Fig. 7.6). The loop is configured to
represent approximately 2 s of real-time, although this is modifiable by the user to
apply a different temporal granularity. Time-dependent input parameters, such as
the nutrient intake rate or antibiotic kinetic parameters, are then configured to the
specified timescale.
There are six stages in the program loop during which the activities of the simu-
lation are carried out. Stage 1 (diffusion) involves updating the levels of the various
molecular components (nutrients, antibiotics and enzymes) in the environment. The
movement of molecules between patches is calculated by applying the algorithm for
Fick’s First Law of diffusion.
Stage 2 (static time) consists of several housekeeping processes. These include
subtracting a survival cost from the bacterium’s energy stock, representing energy
expended on basic metabolic processes in the cell. If ˇ-lactamase gene expression is
active, then the enzyme is released into the environment, with an associated energy
cost. There is also an optional graphical output that the program can produce to
display the simulated culture environment and the levels of various constituents in
it (Fig. 7.7).
A movement/overcrowding algorithm can be applied in stage 3 of the loop to
update the positions of the bacterial agents. This comes into effect when the density

Start

Initialise Parameters Input File

Initialise Environment Initialise Bacterial Fabric

6. Cell
1. Diffusion
Division
Output
File
Main
2. Static
5. Feeding Program
Time
Loop Graphic
Display

4. Interact 3. Move
Fig. 7.6 Diagram of program
flow structure in Micro-Gen Exit Conditions?
Bacteria Simulator with No
Yes
principal stages of program
loop labelled End
146 J.T. Murphy and R. Walshe

Fig. 7.7 Screenshot of Micro-Gen simulation running in parallel on four computing nodes. Circu-
lar shaped MRSA bacterial colonies can be observed growing on a background of nutrient medium.
Contours produced by different shades of grey represent nutrient gradient (lighter colour equals
higher nutrient concentration)

of cells in a patch exceeds the size limit of the patch, and thus there is a possibility
of one or more of the cells being moved into an adjacent patch (see Sect. 7.2.2.3).
Stage 4 is when the algorithms defining the interactions between antibiotics and
bacterial cells and/or ˇ-lactamase enzymes are applied. Kinetic parameters derived
from biological studies are used to determine the reaction rates (7.1), (7.2). Stage 5
of the loop (feeding) is when the bacterial agents can absorb nutrient from the en-
vironment and increase their internal energy stocks. In the case of the simulations
carried out for this study, the rate of intake is that which will result in a generation
time of 29 min (Ender et al. 2004).
The final stage of the loop is when binary fission occurs, whereby the bacterial
agents reproduce asexually to produce two identical daughter cells. This can occur
if the bacterial cell’s energy stock has exceeded a certain threshold for division, and
the level of antibiotic damage (measured by the proportion of inactivated PBP) is
below a critical level.

7.2.6 Parallelisation

The agent-based approach to modeling bacterial colonies means that the compu-
tational resources increase proportionately with the number of agents required.
Therefore, to scale simulations up to represent very large population sizes, it is
important to be able to take advantage of modern high performance comput-
ing resources. The size of the bacterial populations in nature range from 106 to
1010 cells per millilitre of seawater with even higher concentration in individuals
7 Modeling Antibiotic Resistance in Bacterial Colonies 147

Fig. 7.8 Diagram of communication between adjacent nodes at overlapping boundary conditions
when Micro-Gen is run in a parallel configuration

suffering from acute infections (Guan and Kamino 2001). Therefore, Micro-Gen
was designed to take advantage of parallel computing resources and achieve an effi-
cient scale up to hundreds and even potentially thousands of computing processors.
It does this by dividing the simulated environment equally among all available
computing nodes so that each processor is responsible for computing only a subsec-
tion of the overall population (Fig. 7.7). In cases where bacteria and/or molecules
move across the boundaries separating two nodes, a communication strategy is used
whereby overlapping boundary conditions exist between nodes where the elements
are exchanged (Fig. 7.8). They are sent to the receiving node across the available
network connection using a communication protocol known as the message passing
interface (MPI).

7.3 Simulations of Bacteria–Antibiotic Interactions

The complex relationships between the biomolecular/kinetic properties of drug

compounds and emergent pharmacodynamic parameters, such as the MIC, are an
important area to study to develop optimal drug treatment regimens and for better
rational drug design strategies (Regoes et al. 2004). It provides a basis for under-
standing the dynamics involved in the development of antibiotic resistance and thus
can lead to better strategies for limiting its expansion. Micro-Gen represents a good
theoretical framework for analysing the in vitro dynamics of antibiotics interact-
ing with bacteria and for connecting knowledge from low-level biochemical studies
with emergent properties of the population. Furthermore, future work will involve
extending the model to represent the more complex dynamics found in the in vivo
clinical setting as well.
148 J.T. Murphy and R. Walshe

Fig. 7.9 Effect of antibiotic exposure on simulated growth curve of MRSA bacteria in nutrient-
limited culture conditions. Cephalothin antibiotic (103.1 g/ml) added after 3.5 h of incubation,
during the exponential phase of colony growth. This results in inhibition of colony growth for a
finite period of time (inhibition time) until depletion of antibiotic according to its natural half-life,
or hydrolysis by enzymes such as ˇ-lactamases, allows growth to resume

In order to demonstrate the capabilities of the Micro-Gen model, some sample

results are included in this section from simulations of MRSA bacteria growing in
batch culture. Figure 7.9 shows the simulated growth curve of a sample MRSA
bacterial colony outputted by Micro-Gen. The control culture of MRSA, where
no antibiotic is added, follows the characteristic standard growth curve of bacte-
ria grown in nutrient-limited culture conditions (lag, log and stationary phases –
death phase not shown). A second curve shows the effect of adding an inhibitory
dose of antibiotic (103.1 g/ml of cephalothin), after 3.5 h of incubation, in the
development of the colony.
The addition of an inhibitory dose of antibiotic during the exponential phase of
growth causes inhibition of growth for a finite period of time. The length of time
bacterial growth is inhibited is important as it determines the recommended dosage
regimen for an antibiotic. Growth must be inhibited for a long enough period of
time to cover the gap between successive doses of medication to effectively block
bacterial expansion, so that the immune response can remove the infection. Factors
such as the half-life of the antibiotic and the action of bacterial enzymes, such as
ˇ-lactamases which degrade the antibiotic molecules, influence the length of the
inhibition time.
Some test simulations were carried out to validate the model and verify that
the algorithm representing bacteria–antibiotic interactions reproduces the effects
observed in real-life experiments (Murphy et al. 2008). In order to do this, pa-
rameters from the biological literature applicable to three types of MRSA bacteria
were used for the test simulations. These types of MRSA are differentiated by their
7 Modeling Antibiotic Resistance in Bacterial Colonies 149

ˇ-lactamase status: Type A MRSA and Type C MRSA are named because they
produce ˇ-lactamase enzymes of these respective types with different kinetic pa-
rameters associated with them. The third type tested was a ˇ-lactamase-negative
strain which was included as a control.
Type A and Type C ˇ-lactamase enzymes are distinguished by their kinetic pa-
rameters (kcat /KM ), and values for these were derived from experimental literature
(see Table 7.2). They were chosen for this study because they are the most common
types of ˇ-lactamase found in MRSA bacteria. A study by Norris et al. (1994) found
that among 50 ˇ-lactamase-producing MRSA isolates taken from nine locations
across the USA, 80% expressed Type A ˇ-lactamase and the remainder expressed
Type C. Type B and Type D ˇ-lactamases are thought to be less common among
MRSA strains.
The MIC was calculated from the model for a number of common ˇ-lactam
antibiotics against the MRSA strains, and compared with the real-world results. The
MICs were estimated from the model in an analogous way to the broth dilution test
carried out in experimental studies: a series of simulations were performed with
stepwise increases of the concentration of antibiotic in each run. The recorded MIC
is the minimum concentration of antibiotic that resulted in complete inhibition of
bacterial growth for a pre-determined length of time.
Figure 7.10 shows the predicted MIC values from Micro-Gen for three com-
mon ˇ-lactam antibiotics (penicillin G, ampicillin and cephalothin) compared

Fig. 7.10 Predicted (Micro-Gen) versus experimentally determined MICs (

g/ml) of penicillin G,
ampicillin and cephalothin antibiotics against three different types of MRSA. Experimentally de-
termined MICs for ˇ-lactamase-positive (Types A and C) strains are from Norris et al. (1994).
Experimentally determined MICs for ˇ-lactamase-negative strains are from Malouin et al. (2003).
Predicted MICs are derived from triplicate simulations with the geometric mean MIC ˙ SEM
(g/ml) displayed. (A) Type A ˇ-lactamase producing MRSA, (C) Type C ˇ-lactamase producing
MRSA, () ˇ-lactamase-negative MRSA
150 J.T. Murphy and R. Walshe

with results from the experimental studies published in the scientific literature
(Murphy et al. 2008; Malouin et al. 2003; Norris et al. 1994). There is clearly a
very good correspondence between the predicted values from Micro-Gen and the
real-world situation, despite the fact that parameters from different experimental
studies carried out in different laboratories were used. This indicates that despite
the minor differences that arise from varying experimental conditions, the over-
all MICs of the antibiotics are primarily determined by their low-level kinetic
parameters.
The comparison between the predicted and experimental results is a useful ini-
tial step for validating the model. However, this only confirms existing knowledge
about these particular antibiotics. The true power of the model exists in being able
to deconstruct these values and relate them back to the low-level molecular and
cellular parameters of the individual bacterial cells. Through varying the individual
parameters, one can gain a better understanding of the root causes of the observed
drug resistance. For example, resistance of MRSA to cephalothin is mediated by the
antibiotic’s inability to bind efficiently to the PBP2a protein on the bacterial cell sur-
face (k2 /Kd D 1.96 M1 s1 ). By contrast, penicillin G binds more readily to PBP2a
(k2 /Kd D 12.0 M1 s1 ), but its higher susceptibility to cleavage by ˇ-lactamase
enzyme negates this advantage. This is a simple example of how resistance can be
the result of different underlying mechanisms and the dynamic interplay between
them. The modeling approach allows us to investigate a wide range of possible sce-
narios to investigate the underlying mechanistic basis for the resistance.
These results were obtained without attempting to “fit” the kinetic input param-
eters to the MIC results. The only parameter, which influences the kinetics of the
bacterial agents, that required to be fitted was the ˇ-lactamase production rate. How-
ever, for the results of the ˇ-lactamase-negative strain in Fig. 7.10, this fitting step
did not have to be carried out. Therefore, the predicted MICs were purely an emer-
gent property of the inputted kinetic parameters for the PBP2a-antibiotic binding
reaction. There was still a close quantitative agreement between the experimental
and simulation results even in this scenario.
We have carried out extensive research and sensitivity analyses to examine the
impacts of the various molecular/environmental factors that lead to the observed re-
sponses to antibiotic treatment. These have included explorations of the parameter
space with particular emphasis on the kinetic parameters associated with the antibi-
otic molecules. A detailed discussion of these studies is beyond the scope of this
chapter; however, for further information the reader is directed towards a couple of
significant publications on the subject (Murphy et al. 2008, 2009). In general, the
reaction of a bacterial community to antibiotic treatment is sensitive to environmen-
tal conditions such as the rate of diffusion, temperature, pH, etc. These factors can
play an important role in contributing to the differences observed between results
in a laboratory setting and the actual in vivo response to treatment. Simulations are
useful for examining situations not readily reproduced in the experimental setting
due to logistical constraints. For example, the model can be used to examine evo-
lutionary pathways and assess the fitness advantages/disadvantages associated with,
for example, changes in the reaction profiles of the bacteria–antibiotic interactions.
7 Modeling Antibiotic Resistance in Bacterial Colonies 151

Knowledge from these studies would be useful in rational drug design to assess the
potential for antibiotic resistance to develop rapidly following introduction of a drug
into clinical use.

7.4 Conclusions and Future Work

An integrated systems-level understanding of the complex dynamics that lead to the

rapid development and spread of antibiotic resistance within bacterial populations is
required to meet the growing demands of treating patients with multi-drug resistant
bacterial infections. In order to meet this challenge, scientists will need to approach
the problem from various angles in both the theoretical and experimental domains.
In terms of theoretical approaches to understanding the system dynamics of resis-
tance development, there are a number of techniques available, some of which were
mentioned at the start of this chapter, such as ODEs and CA. However, the agent-
based approach discussed here is a particularly important tool for investigating the
relationship between the individual molecular components and bacterial cells of the
system and the overall treatment outcome.
There is much scope for improvement and development of the agent-based ap-
proach as it is only relatively recently that it has begun to gain traction as an
important addition to the bacterial population modeling toolset. Its potential applica-
tions are extremely broad, and it represents a great means for investigating complex
situations such as the diverse, heterogeneous bacterial ecosystems present in the
intestinal tract or the network of genetic exchanges arising from horizontal gene
transfer in bacterial populations (Sorensen et al. 2005). The modular nature of the
agent-based approach and its “bottom-up” approach to modeling systems means that
seemingly intractable problems may be broken down to their simplest, basic units
and new insights gained into the emergence of complexity from their interactions.
Future work will include using the model to investigate the system dynamics of
combination therapy where multiple classes of antibiotic are applied simultaneously
to treat infections. It can also be used to test hypothetical scenarios by varying the
parameters of existing antibiotics to explore how potential novel compounds might
act. The agent-based approach is also suitable for modeling the evolution of an-
tibiotic resistance over time by incorporating genetic components into the bacterial
agents. This would allow the examination of both the temporal and spatial dynamics
of antibiotic resistance development in a population exposed to antibiotics.
Other important developments of the model include expanding the environment
to represent three-dimensional space to model more complex spatially structured
microbial communities such as biofilms. Biofilms are complex aggregations of mi-
crobial cells that are characterised by their complex cellular interactions, genetic
diversity and structural heterogeneity. They are characterised by highly heteroge-
neous localised niches where the chemistry varies dramatically over small distances.
The agent-based approach is a powerful tool for modeling the interactions within a
heterogeneous environment such as this.
152 J.T. Murphy and R. Walshe

Acknowledgements The authors would like to thank Marc Devocelle from the Royal College of
Surgeons in Ireland, who collaborated on the biological aspects of the Micro-Gen research project.
The authors would also like to acknowledge the contributions of Mathieu Joubert, Grainne Kerr,
Chris Pender, Marian Duggan and Ronan Winters who developed the original BAIT software tool
under the supervision of Ray Walshe in Dublin City University.

References

E. P. Abraham and E. Chain. An enzyme from bacteria able to destroy penicillin. Nature, 146:837,
1940
P. G. Ambrose, S. M. Bhavnani, C. M. Rubino, A. Louie, T. Gumbo, A. Forrest, and G. L. Drusano.
Pharmacokinetics-pharmacodynamics of antimicrobial therapy: it’s not just for mice anymore.
Clinical Infectious Diseases, 44(1):79–86, 2007
E. Ben-Jacob, O. Schochet, A. Tenenbaum, I. Cohen, A. Czirok, and T. Vicsek. Generic modelling
of cooperative growth patterns in bacterial colonies. Nature, 368(6466):46–49, 1994
A. Bondi and C. C. Dietz. Penicillin resistant staphylococci. Proceedings of the Society for Exper-
imental Biology and Medicine, 60(1):55–58, 1945
E. Chain, H. W. Florey, M. B. Adelaide, A. D. Gardner, N. G. Heatley, M. A. Jennings, J. Orr-
Ewing, and A. G. Sanders. Penicillin as a chemotherapeutic agent. 1940. Clinical Orthopaedics
and Related Research, 295:3–7, 1993
H. F. Chambers, M. J. Sachdeva, and C. J. Hackbarth. Kinetics of penicillin binding to penicillin-
binding proteins of Staphylococcus aureus. The Biochemical Journal, 301(Pt 1):139–144, 1994
D. A. Devine. Cationic antimicrobial peptides in regulation of commensal and pathogenic micro-
bial populations. Mammalian Host Defense Peptides, pages 9–39, 2004
A. Elfwing, Y. LeMarc, J. Baranyi, and A. Ballagi. Observing growth and division of large num-
bers of individual bacteria by image analysis. Applied and Environmental Microbiology, 70(2):
675–678, 2004
M. Ender, N. McCallum, R. Adhikari, and B. Berger-Bachi. Fitness cost of sccmec and methicillin
resistance levels in Staphylococcus aureus. Antimicrobial Agents and Chemotherapy, 48(6):
2295–2297, 2004
K. R. Eriksen. Celbenin-resistant staphylococci. Ugeskrift for laeger, 123:384–386, 1961
J. F. Fisher, S. O. Meroueh, and S. Mobashery. Bacterial resistance to beta-lactam antibiotics:
compelling opportunism, compelling opportunity. Chemical Reviews, 105(2):395–424, 2005
C. Fuda, M. Suvorov, S. B. Vakulenko, and S. Mobashery. The basis for resistance to beta-lactam
antibiotics by penicillin-binding protein 2a of methicillin-resistant Staphylococcus aureus. The
Journal of Biological Chemistry, 279(39):40802–40806, 2004
C. C. Fuda, J. F. Fisher, and S. Mobashery. Beta-lactam resistance in Staphylococcus aureus:
the adaptive resistance of a plastic genome. Cellular and Molecular Life Sciences, 62(22):
2617–2633, 2005
M. Gardner. Mathematical games: The fantastic combinations of john conways new solitaire game
life. Scientific American, 223(4):120–123, 1970
P. Giesbrecht, T. Kersten, H. Maidhof, and J. Wecke. Staphylococcal cell wall: morphogenesis and
fatal variations in the presence of penicillin. Microbiology and Molecular Biology Reviews, 62
(4):1371–1414, 1998
M. Ginovart, D. Lopez, and J. Valls. Indisim, an individual-based discrete simulation model to
study bacterial cultures. Journal of Theoretical Biology, 214(2):305–319, 2002
K. Graves-Woodward and R. F. Pratt. Reaction of soluble penicillin-binding protein 2a of
methicillin-resistant Staphylococcus aureus with beta-lactams and acyclic substrates: kinetics
in homogeneous solution. The Biochemical Journal, 332(3):755–761, 1998
V. Grimm. Ten years of individual-based modelling in ecology: what have we learned and what
could we learn in the future? Ecological modelling, 115(2–3):129–148, 1999
7 Modeling Antibiotic Resistance in Bacterial Colonies 153

M. J. Grimson and G. C. Barker. Continuum model for the spatiotemporal growth of bacterial
colonies. Physical Review. E, Statistical Physics, Plasmas, Fluids, and Related Interdisci-
plinary Topics, 49(2):1680–1684, 1994
L. L. Guan and K. Kamino. Bacterial response to siderophore and quorum-sensing chemical signals
in the seawater microbial community. BMC Microbiology, 1:27, 2001
N. R. Jennings, K. Sycara, and M. J. Wooldridge. A roadmap of agent research and development.
Autonomous Agents and Multi-Agent Systems, 1(1):7–38, 1998
W. M. M. Kirby. Extraction of a highly potent penicillin inactivator from penicillin resistant staphy-
lococci. Science, 99(2579):452–453, 1944
J. U. Kreft, G. Booth, and J. W. Wimpenny. Bacsim, a simulator for individual-based modelling of
bacterial colony growth. Microbiology, 144(12):3275–3287, 1998
A. M. Lacasta, I. R. Cantalapiedra, C. E. Auguet, A. Penaranda, and L. Ramirez-Piscina. Mod-
eling of spatiotemporal patterns in bacterial colonies. Physical Review. E, Statistical Physics,
Plasmas, Fluids, and Related Interdisciplinary Topics, 59(6):7036–7041, 1999
S. B. Levy and B. Marshall. Antibacterial resistance worldwide: causes, challenges and responses.
Nature Medicine, 10(12 Suppl):S122–129, 2004
R. A. Lewis, S. P. Curnock, and K. G. Dyke. Proteolytic cleavage of the repressor (blai) of beta-
lactamase synthesis in Staphylococcus aureus. FEMS Microbiology Letters, 178(2):271–275,
1999
W. P. Lu, Y. Sun, M. D. Bauer, S. Paule, P. M. Koenigs, and W. G. Kraft. Penicillin-binding
protein 2a from methicillin-resistant Staphylococcus aureus: kinetic characterization of its
interactions with beta-lactams using electrospray mass spectrometry. Biochemistry, 38(20):
6537–6546, 1999
F. Malouin, J. Blais, S. Chamberland, M. Hoang, C. Park, C. Chan, K. Mathias, S. Hakem,
K. Dupree, E. Liu, T. Nguyen, and M. N. Dudley. Rwj-54428 (mc-02,479), a new cephalosporin
with high affinity for penicillin-binding proteins, including pbp 2a, and stability to staphylo-
coccal beta-lactamases. Antimicrobial Agents and Chemotherapy, 47(2):658–664, 2003
W. Materi and D. S. Wishart. Computational systems biology in drug discovery and development:
methods and applications. Drug Discovery Today, 12(7–8):295–303, 2007
J. T. Murphy and R. Walshe. Micro-gen: an agent-based model of bacteria-antibiotic interactions
in batch culture. In Proceedings of the 20th Annual European Simulation and Modelling Con-
ference, pages 239–242. Eurosis-ETI, October 23–25, 2006
J. T. Murphy, R. Walshe, and Marc Devocelle. Agent-based model of methicillin-resistant Staphy-
lococcus aureus and antibiotics in batch culture. In Proceedings of 21st Annual European
Simulation and Modelling Conference, pages 409–414. Eurosis-ETI, October 20–22, 2007
J. T. Murphy, R. Walshe, and M. Devocelle. A computational model of antibiotic-resistance mecha-
nisms in methicillin-resistant Staphylococcus aureus (MRSA). Journal of Theoretical Biology,
254(2):284–293, 2008
J. T. Murphy, R. Walshe, and M. Devocelle. Modelling the population dynamics of antibiotic-
resistant bacteria: An agent-based approach. International Journal of Modern Physics C, 20
(3):435–457, 2009
H. C. Neu. The crisis in antibiotic resistance. Science, 257(5073):1064–1073, 1992
J. B. Nielsen and J. O. Lampen. Membrane-bound penicillinases in gram-positive bacteria. The
Journal of Biological Chemistry, 257(8):4490–4495, 1982
G. W. Niven, T. Fuks, J. S. Morton, S. A. Rua, and B. M. Mackey. A novel method for measuring
lag times in division of individual bacterial cells using image analysis. Journal of Microbiolog-
ical Methods, 65(2):311–317, 2006
S. R. Norris, C. W. Stratton, and D. S. Kernodle. Production of a and c variants of staphylococcal
beta-lactamase by methicillin-resistant strains of Staphylococcus aureus. Antimicrobial Agents
and Chemotherapy, 38(7):1649–1650, 1994
R. R. Regoes, C. Wiuff, R. M. Zappala, K. N. Garner, F. Baquero, and B. R. Levin. Pharmaco-
dynamic functions: a multiparameter approach to the design of antibiotic treatment regimens.
Antimicrobial Agents and Chemotherapy, 48(10):3670–3676, 2004
154 J.T. Murphy and R. Walshe

D. Skinner and C. S. Keefer. Significance of bacteremia caused by Staphylococcus aureus. Archives

of Internal Medicine, 68:851–875, 1941
S. J. Sorensen, M. Bailey, L. H. Hansen, N. Kroer, and S. Wuertz. Studying plasmid horizontal
transfer in situ: a critical review. Nature Reviews Microbiology, 3(9):700–710, 2005
M. W. van der Woude and A. J. Baumler. Phase and antigenic variation in bacteria. Clinical micro-
biology reviews, 17(3):581–611, 2004
C. Walsh. Molecular mechanisms that confer antibacterial drug resistance. Nature, 406(6797):
775–781, 2000
R. Walshe. Modelling bacterial growth patterns in the presence of antibiotic. In Proceedings of
the 11th IEEE International Conference on Engineering of Complex Computer Systems, pages
177–188, Washington, DC, USA, 15–17 August 2006. IEEE Computer Society
D. S. Wishart, C. Knox, A. C. Guo, S. Shrivastava, M. Hassanali, P. Stothard, Z. Chang, and
J. Woolsey. Drugbank: a comprehensive resource for in silico drug discovery and exploration.
Nucleic Acids Research, 34(Database issue):D668–672, 2006
G. D. Wright. The antibiotic resistome: the nexus of chemical and genetic diversity. Nature Reviews
Microbiology, 5(3):175–186, 2007
H. Z. Zhang, C. J. Hackbarth, K. M. Chansky, and H. F. Chambers. A proteolytic transmem-
brane signaling pathway and resistance to beta-lactams in staphylococci. Science, 291(5510):
1962–1965, 2001
D. J. Zygmunt, C. W. Stratton, and D. S. Kernodle. Characterization of four beta-lactamases pro-
duced by Staphylococcus aureus. Antimicrobial Agents and Chemotherapy, 36(2):440–445,
1992
Chapter 8
Modeling the Spatial Pattern Forming Modules
in Mitotic Spindle Assembly

Chaitanya A. Athale

8.1 Introduction

The phyllotaxy of leaves, the scaling of animal limbs and organs, the cellular
patterns in tissue, the shapes of individual cells, the form of subcellular struc-
tures like spindles, cilia and flagella as well as polarization of signalling molecules
all encompass biological pattern formation. Many of these are caused by underly-
ing processes that are in common with physical processes that generate nonliving
patterns – crystals, faults, river-networks and chemical oscillators. This process in-
volves both self-assembly and self-organization. While self-assembly is driven to
energy minima attained by interacting components and often has rather determinis-
tic ends, self-organization is driven often by local interactions with feedback which
lead to unpredictable outcomes.
The most prominent example of modeling biological spatial pattern formation
was the use of reaction-diffusion models by Alan Turing to explain the appearance
of periodic concentrations of a chemical that would generate periodic morpho-
genetic patterns. Using a combination of a slow-diffusing activator and a fast
diffusing inhibitor, Turing could show that given an initial random distribution of
both species, the system within a range of reaction parameters would produce a
steady-state periodic pattern. This counter-intuitive finding about the emergence of
order from a random distribution was explained by the mathematical properties of
such a system. At the moment experimental evidence of similar Turing patterns has
been demonstrated in few systems like the Escherichia coli oscillating Min pro-
teins which help find the centre of the bacterium during division (Meinhardt and
de Boer 2001). This model however has more components and different interac-
tions from a minimal activator substrate-depletion system which could theoretically
produce the same patterns. Some of the complexity of the network is necessary to

C.A. Athale ()

EMBL, Meyerhofstrasse 1, Heidelberg 69117, Germany
and
IISER Pune, IISER, Central Tower, Sai Trinity Building, Sutarwadi Road,
Pashan, Pune 411021, India
e-mail: [email protected]; [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 155
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 8,
c Springer Science+Business Media, LLC 2011
156 C.A. Athale

robustly suppress certain modes of the system, demonstrating that pattern forming
reaction-diffusion systems in biology do not necessarily show the minimal activator-
inhibitor system although similar principles of diffusion driven instability might
govern them.
Spindle assembly occurs in eukaryotic cells in the M-phase of cell division. In the
stage between prophase and metaphase, simultaneously the chromatin condenses,
the nuclear envelope ruptures, a pair of centrosomes move to opposite ends of the
nucleus and microtubule fibers nucleated from the two poles attach to chromosomes
as well as the opposite poles to form a mitotic spindle. In some cells (e.g., yeast)
spindle assembly is not accompanied by nuclear envelope breakdown, while in oth-
ers such as starfish oocytes, congression of chromosomes is driven by acto-myosin
dependent chromosome motility to the metaphase plate. In meiosis for instance, the
spindle lack specific centrosomes, instead the poles self-organize by a combina-
tion of chromosomal nucleation and minus-end directed transport of microtubules.
Taken together the mitotic spindle is a machine that is responsible for segregation
of the chromosomes and its assembly is of fundamental importance for inheritance,
cancer and a fundamental understanding of biological self-replication. Its assem-
bly can be seen as a process of pattern formation involving different modules and
processes. Although the discovery of the mitotic spindle is over 120 years old dat-
ing to work by Flemming (1882), and many of the motor proteins (as reviewed in
Walczak et al. (1998)) and the network of regulators (reviewed in Akhmanova and
Steinmetz (2008)) have been found, we still do not understand how these parts func-
tion together to produce the complex machinery of the spindle. Such understanding
will not just have consequences for the field of cell division but might also act as a
template for solving other problems in pattern-forming systems at the cellular scale.
Applying a systems biology approach that combines experiment reconciled to
theory promises to reveal some of the aspects of this complex pattern forming sys-
tem. Modeling has been successfully applied to understanding the assembly of the
mitotic spindle. These models have approached this complex problem by reducing
the spindle assembly process into modules. These modules can be classified as:
 Microtubule dynamics
 Microtubule-motor interactions
 Chromosome dynamics
 Reaction-diffusion gradients of microtubule dynamics regulation
Spindle-assembly is a complex combination of various processes like reaction-
diffusion, mechanical contact, nonlinear interactions and spatial-pattern formation.
In order to understand it, even as a model, it is necessary to study the mod-
ules that govern it. In recent work, this author and co-workers have taken such
an approach, separating the effect of the reaction-diffusion components from the
mechanical in spindle assembly, and in the process discovered novel subcellular
gradient shapes (Athale et al. 2008).
Experiments of spindle assembly have been performed in several model systems.
The minimal system is the commonly used biochemically pure in vitro system with
microtubules, motors, microtubule associated proteins (MAPs) and DNA. It has
8 Spatial Modeling of Mitotic Spindle Assembly 157

not yet been possible to assemble a fully functional spindle in such a reduced
system, though aspects like the localization of proteins to microtubule tips have
been recently successfully reproduced (Bieling et al. 2007). The use of cytoplasmic
extracts from the eggs of the African clawed frog Xenopus laevis with fluores-
cently labelled tubulin and DNA to examine the mitotic spindle (Fig. 8.1a, b) has

b
Microtubules Centrosome

DNA Motors

Metaphase Plate Axis

d
Microtubules
Centrosome

Fig. 8.1 Fluorescence microscopy and schematic view of spindle. (a) A fluorescence microscopy
image of the same spindle assembled in Xenopus egg extracts and fixed on a slide with the Cy3
(a cyanine fluorescent dye) tubulin labelled microtubules (left) and Hoechst labelled sperm DNA
(right) imaged on a wide-field microscope. (b) A schematic view of the spindle depicts the major
components: centrosomes, microtubules chromosomal DNA and microtubule-dependent motors.
(c) The centrosome when placed in mitotic Xenopus egg extract with fluorescent tubulin nucleates
a star-like distribution of microtubules, referred to as the centrosomal aster. (d) A schematic view
displays the two major components of an aster, the centrosome which nucleates microtubules and
the radiating microtubules
158 C.A. Athale

become widespread. Simply placing one of the components, the centrosome in the
extract also produces dynamic microtubule asters (Fig. 8.1c, d). Spindle assembly
has also been demonstrated using plasmid DNA in the Xenopus egg-extract. This
DNA lacks any chromosome-specific information and even centrosomes are not
required to assemble the spindle (Heald et al. 1996). Spindle assembly appears
thus to be self-organized and has many redundant pathways. Another commonly
used system to study spindle assembly is fission yeast Schizosaccharomyces pombe
that combines easy genetic and mechanical manipulation of the spindle. A more
complex context for spindle assembly is the Drosophila syncytial blastoderm stage
embryo (cycles 10–13) with up to 103 nuclei undergoing synchronized mitoses.
How these modules of spindle assembly have been used to study spindle assembly
by experimental and theoretical approaches will be further elaborated upon in the
following sections.

8.2 Microtubule Dynamics

The microtubule cytoskeleton is a critical structural component in the process of

assembling the mitotic spindle. This protein forms an oriented polymer showing a
dynamic plus-end that are mostly found to grow, while minus-ends are those that
mostly disassemble by losing subunits (Fig. 8.2a, b). The assembly of the spin-
dle occurs typically within 20–30 min (Wollman et al. 2005) and requires dynamic
changes in its components. This is seen in the case of the microtubule cytoskeleton
when it transitions from its interphasic to mitotic state. The interphasic microtubules
extend from the centrosome to cell periphery and show little change in length. Mi-
totic microtubules in contrast are shorter, more dynamic and appearing to randomly
switch between growing and shrinking states (Verde et al. 1992). What triggers this
transition from interphasic to mitotic microtubules dynamics is still a subject of ac-
tive research. Experimental findings suggest that discrete phosphorylation states of
the network of MAPS may play a role in this transition (Niethammer et al. 2007).
Microtubules can be broadly defined to have two main aspects – nucleation and
polymerization.

8.2.1 Nucleation

The nucleation of microtubules in vivo occurs most commonly from centrosomes

which act as templates for tubules to form. Verde et al. (1992) modeled the centroso-
mal nucleation rate as a constant. However, a later study modeled the nucleation of
microtubules from centrosomes using a mean field approach. The study demon-
strated the existence of two regimes – one that is nucleation “site limited” for
small number of nucleation sites and another that is “diffusion limited” for large
numbers of nucleation sites (Dogterom et al. 1995). The diffusion limitation arises
8 Spatial Modeling of Mitotic Spindle Assembly 159

from a large numbers of nucleation sites that consume GTP-tubulin locally around
centrosomes due to polymerization. Such physical mechanisms are in contrast to
conventionally accepted biochemical regulation of nucleation (Raynaud-Messina
and Merdes 2007). In addition plant cells and meiotic spindles do not require the
presence of centrosomes to nucleate microtubules. Instead other nucleation centres
at the cortex (Ehrhardt 2008) and chromosomes (Gruss et al. 2001) locally nucleate
microtubules. This module of non-centrosomal nucleation has been used to model
spindle assembly by in a “slide and cluster” model by combining motor proteins
with microtubule nucleation on existing microtubules (Burbank et al. 2007; Clausen
and Ribbeck 2007).

8.2.2 Polymerization

Microtubule polymerization proceeds by random transition of between growing and

shrinking states that has been defined by Mitchison and Kirschner (1984) as dynamic
instability. They further proposed that the experimental dynamics of mitotic micro-
tubules indicate out of equilibrium oscillations due to stabilization/destabilization
in the presence/absence of a GTP capped state. Simultaneously a general model
of a two-state polymer was developed that could explain the trends in experiment
(Hill 1984). Monte Carlo simulations of this two-state polymer model were used
to explore the parameter space that reproduced the experimental microtubule length
change (Chen and Hill 1985). The transition of a growing polymer to shrinking state
was defined as catastrophe and the transition from shrinking to growing defined as
rescue. More recent work on the transition of microtubule dynamics between inter-
phase and mitosis modeled microtubule dynamics of polymerization by dynamical
probabilistic equations (Verde et al. 1992). The model parameters are – frequencies
of rescue (fres ) and catastrophe (fcat ), and velocities of growth (vg ) and shrinkage
(vs ). Catastrophe refers to the transition from growing to shrinking state and rescue
is the transition from shrinking to growing state. The average length (hLi) of the
polymer thus depends on the flux and is given by:
vg vs
hLi D : (8.1)
vs fcat  vg fres
Such a model simplifies the known structural components of microtubule fibers
that consist of 12–14 protofilaments of dimeric ˛ and ˇ tubulin organized in a spiral
(Fig. 8.2a). Additionally one end of this tube grows faster than the other – referred
to as the plus-end. The plus-end is thought originate in the GTP bound state of the
dimers. In contrast, the end of the tubule that is more likely to shrink is referred to as
the minus-end. However, the dynamic instability models ignore all these details, in-
stead treating the tubule as a polymer with subunits that are being added or removed
at certain velocities (Fig. 8.2b). The power of this model however lies in its ability
to precisely predict the length distributions of experimental microtubules, based on
these simple parameters (Verde et al. 1992).
160 C.A. Athale

a b

αβ Tubulin dimer
Rescue Catastrophe

Minus End Plus End

25 nm

Fig. 8.2 The dynamics of microtubules. (a) The dynamics of microtubules can be described at
a molecular level with a model of dimer addition and removal. The GTP-cap model proposes
that non-dynamic forms of the tubulin tube are GDP bound while those at the ends form a
GTP-bound cap. (b) A more abstract treatment often used in Monte Carlo simulations of micro-
tubule dynamics is of microtubules as a long polymer chain with subunits falling off and being
added. The process of transition from growing to shrinking is called catastrophe and shrinking
to growing rescue. The fibers transition in this model spontaneously at experimentally measured
frequencies

Polymerizing microtubules can generate forces as has been seen in experiments

with centrosomes placed in geometrically restricted chambers – the centrosomes
reach the geometric centres of the chambers and oscillate (Holy et al. 1997). Exper-
imental measurement has shown that the microtubule dynamics are affected by the
force acting on them, i.e., if a microtubule comes into contact with an immovable
obstacle, that microtubule will undergo shrinkage (Janson et al. 2003). A further
study quantified the forces of microtubules in bundles by optical tweezer measure-
ments. It demonstrated that force generation and single microtubule catastrophes
could be coupled, leading to a model of microtubule force generation and oscilla-
tion of lengths of bundles (Laan et al. 2008).
Increasing length of microtubules appears to decrease rescue and increase catas-
trophe frequency of microtubules nucleated from centrosomes (Dogterom et al.
1996a). Such a mechanism has been found to be necessary to add to simulations
of S. pombe microtubules in order to match the simulated organization with that
experimentally measured (Foethke et al. 2009).

8.3 Microtubule-Motor Interactions

The spatial organization of microtubules in the spindle assembly process is primarily

driven in vivo by energy dependent molecular motor proteins. The microtubules
have an orientation of plus- and minus-ends. Similarly, the movement of motors
is seen to fall into two classes of motors that move to either plus- or minus-ends.
Typically these interactions have been studied in scales ranging from single motor-
microtubule activities to the role of collective behaviour.
8 Spatial Modeling of Mitotic Spindle Assembly 161

8.3.1 Microtubule Gliding Assays

This assay is used typically to asses the kinetics of movement of microtubules by

motors that are bound to a surface. A model of such “gliding assays” has been
developed to understand the dependencies between the motors and microtubules
that lead to the statistical properties of motion (Duke et al. 1995). Furthermore,
defects in the motor density that lead to spiralling of the microtubule on the surface
has been modeled to estimate the forces exerted by the motor proteins (Bourdieu
et al. 1995). More recently, the collective behaviour of motors has been shown to
depend on the rigidity of the motor in the case of kinesin where a shortened version
of the motor decreases microtubule motility with increasing density of the motor
(Bieling et al. 2008). Thus, loose coupling appears to be essential for cooperative
activity of motor driven microtubule movement.

8.3.2 Motor Mechanics

Microtubule dependent motors are classified as either being plus- or minus-ended,

depending on the direction of the microtubule they preferentially move towards.
Most kinesins are plus-end directed, while most dyneins are minus-end directed.
The step sizes and force produced by single motors has been estimated in the past
using single molecule optical tweezer experiments for kinesin (Svoboda et al. 1993;
Visscher et al. 1999). How these motor mechanics affect spindle assembly is more
complex. One of the explanations is that a force-balance between inward and out-
ward forces determines the length of the spindle as demonstrated by studies in
Drosophila embryos by live imaging and modeling (Cytrynbaum et al. 2005). The
inward forces are derived from minus-end directed motors (e.g., dynein) as demon-
strated in the centering of microtubules in fish melanocyte fragments (Cytrynbaum
et al. 2003). Experimental testing of these predictions demonstrated the need to in-
clude nuclear stretching in the generation of inward force (Cytrynbaum et al. 2005).
A more minimal motor-microtubule modeling approach explored the steady state
effects of motors on two centrosomal asters and demonstrated sliding forces could
generate antiparallel overlaps and result in bipolar structures resembling spindles
(Nedelec 2002). However, the hypothesized coupled dimeric plus- and minus-ended
motor complexes are yet to be experimentally discovered.

8.3.3 Microtubule-Motor Patterns

Looking at collective behaviour, when mixtures of multi-headed kinesin were mixed

with randomly nucleated microtubules in vitro, asters form in unconstrained and
spatially restricted chambers by multiple pathways (Nedelec et al. 1997). This can
also be simulated based on basic principles as seen in Fig. 8.3. Subsequently a model
162 C.A. Athale

Fig. 8.3 Spontaneous microtubule pattern formation. Spontaneous patterns formed in a rigid circle
in a 2D simulation of randomly distributed microtubules .n D 60/ are organized by plus-ended
dimeric kinesin motors .n D 3;000/ into a radial centered array. The grey dots are motors which are
either freely diffusing or can bind to and act on microtubules with a rate of attachment of 10 s1
and detachment of 0.5 s1 . (Simulations performed using Cytosim a CCC library described in
Nedelec and Foethke (2007))

solution of a convection–diffusion equation1 was developed, showing distribution

of molecular motor densities on centrosomal asters can be described by continu-
ously varying exponents (Nedelec et al. 2001). Performing stochastic simulations
and experimental measurements on fluorescently tagged kinesin molecules in the
presence of centrosomal asters both produced a similar outcome. When micro-
tubules were mixed with oligomeric motors of opposite directionality – ncd (non
claret disjunction, a minus-end kinesin motor protein) and kinesin-5 a plus-end
kinesin – microtubules demonstrated spontaneous emergence of diverse patterns
– spirals, whorls and asters. A theoretical model predicted these patterns, thus
demonstrating that such patterns are the result of fundamental physical principles
(Surrey et al. 2001). The role of microtubule-dynein interactions to produce an aster
was explored using computer simulations applied to melanocyte granule aggrega-
tion (Cytrynbaum et al. 2004). A more general approach to phase transitions and
ordering of microtubule filaments of motor systems has been taken by modeling
cooperative behaviour of microtubule in motility assays using Langevin dynamics
(Kraikivski et al. 2006). Such simulations use stochastic components to represent
omitted variables, in order to enable abstract models. The typical Langevin equa-
tion of motion in such models as implemented in Cytosim (Nedelec and Foethke
2007) reads:
dx D
F .x; t/dt C dB.t/; (8.2)
where F .x; t/ represents the forces acting on the vector of points x at time t, B.t/
are the random molecular collisions leading to Brownian motion and
contains
the mobility coefficient parameters. Most recently, a model of plus- and minus-
end motor mechanics was combined with a model of short microtubules that are

1
Convection refers to directional movement of molecular motors, diffusion refers to random move-
ment due to diffusion.
8 Spatial Modeling of Mitotic Spindle Assembly 163

nucleated from chromosomes as well as on pre-existing microtubules. Such a system

was shown to self-organize to a bipolar spindle structure in a model referred to as
“slide and cluster” (Burbank et al. 2007). These approaches of modeling the patterns
formed by microtubule motor systems appear to produce structures qualitatively
similar to mitotic spindles and are represent a promising step towards a theory of
spindle assembly.

8.4 Chromosome Dynamics

8.4.1 Search and Capture

During prometaphase, microtubules are more dynamic than in interphase, and are
apparently efficient in finding chromosomes – a process termed as “search and
capture”. The time from prometaphase to metaphase plate formation is typically
< 30 min. The initial event of spindle assembly that involves both centrosomes
and chromosomal DNA is thought to be result of randomly growing and shrink-
ing microtubules (search), which might get immobilized on chromosome (capture)
(Fig. 8.4 inset). This “search and capture” has to occur in a short time window with

Random
Step
80 Random Search and Capture

60 Random
Capture time (t c) (min)

DNA-Aster distance Biased Search and Capture

40
Step gradient

20

0 10 20 30 40 50 60 70
DNA-Aster distance (μm)

Fig. 8.4 Search and capture. Search and capture is the process by which a microtubules first en-
counters a chromosomal patch a fixed distance away – and the time taken is capture time (tc ).
A simulation of microtubules nucleated from a centrosomal aster perform random search and cap-
ture when the process purely depends on stochastic fluctuations in microtubule length (left inset),
while in biased search and capture a gradient around the chromosomes preferentially increases
microtubules growth locally (right inset). tc in random search and capture remains biologically
realistic (<10 min) even when the distance between aster and chromatin is <45 m, while tc for
random search and capture is already >10 min for distances of >25 m
164 C.A. Athale

a high degree of accuracy since any mistake might result in loss of chromosomes
or missegregation. Models that proposed the optimal values of dynamic instabil-
ity parameters which would allow for reliable search and capture for small cells
were proposed (Hill 1985; Holy and Leibler 1994). However, for cells larger
than 25 m sizes the process of “random” search and capture appears to be in-
sufficient (Fig. 8.4). These calculations are made considering the encounter of a
single-microtubule with a single chromosomal patch of 5 m diameter. However,
in a model that considers all 46 human chromosomes in a 3D environment in a
typical cell, tc is limited by the delay in those chromosomes capturing at least
one microtubule that did not initially capture a microtubule. This delay increases
the waiting time before anaphase to 1 h. To overcome this, short range gradients
of microtubule regulators (range 5 m) have been proposed that can change the
capture time to more realistic values (Wollman et al. 2008). These gradients are
based on experimentally measured chromosomal gradients of RanGTP2 (Kalab et al.
2002) and complexes (Kalab et al. 2006; Caudron et al. 2005). Most recently this
author has shown with co-workers how a step-gradient of stabilization maintains
capture times under 10 min for distances between chromosomal DNA and cen-
trosome of up to 45 m (Athale et al. 2008) (Fig. 8.4). In starfish oocytes where
distances between chromosomes and centrosomes can exceed 45 m, experimen-
tal work has demonstrated the role of a contractile, the acto-myosin network, that
initially causes congression of chromosomes (Lenart et al. 2005). Thus such a sim-
ple model of “search and capture” has provided insights into a module of spindle
assembly, that has contributed to our understanding of the more complex in vivo
situation.

8.4.2 Metaphase Plate Formation

Chromosomes form the structure which is to be separated by the spindle appara-

tus. The process of attachment of each kinetochore of a pair of chromosomes such
that the kinetochores lie equidistant from both poles at the metaphase plate in a
spindle is referred to as congression. Mechanisms where chromosomes play ei-
ther a passive or an active role in chromosome congression have been proposed.
A physical model proposed that interactions of astral microtubules with the cell
surface produces a balance of forces that leads to spindle placement at positions
of stable equilibrium where forces with chromosomes simply reacting to this bal-
ance of forces (Bjerknes 1986). As an alternative, microtubule polymerization and
depolymerization at the kinetochore was proposed to be transduced into chromo-
some movement (Hill 1985). A qualitative model of chromosome displacement in
prometaphase proposed chromosomes move in two states: (a) energy driven active
motion, and (b) neutral diffusion driven kinetochore movements, and a stochas-

2
Ran D Ras-related nuclear protein.
8 Spatial Modeling of Mitotic Spindle Assembly 165

tic switching between these two states (Khodjakov et al. 1999). Such a simple
model provided good agreement with the data that quantified the time dynamics
of chromosome movement. An improved mechanical model explicitly considered
the polar ejection forces (forces due to astral microtubules) and balance of kine-
tochore microtubule forces from the spindle poles to generate the same behaviour
without making assumptions about stochastic switching of states, relying instead
on measured microtubule dynamics (Joglekar and Hunt 2002). When chromosomes
congress at metaphase, the question arises if this compaction is random or follows
any specific order. An experiment that used bleach marks in fluorescenctly labelled
histone proteins to mark the orientation of chromatin DNA in live dividing cells was
designed to test if the axes parallel and perpendicular to the metaphase plate axis
(Fig. 8.1b) were maintained after anaphase, i.e., daughter nuclei formation. Only a
model that assumes the maintenance of global spatial order of chromosomes (treated
as rigid spheres) in the formation of the metaphase plate can explain the why the
bleach mark perpendicular to the metaphase plate is maintained from a prophase
nucleus into its daughter nuclei (Gerlich et al. 2003). Models at various level of
description are thus beginning to shed light on the genetically most important com-
ponents of the spindle assembly process.

8.5 Reaction-Diffusion Gradients of Microtubule

Dynamics Regulation

Very early on, experimental work by Bataillon (1912) showed that on injection of
somatic nuclei into frog eggs they adopted the same mitotic state as the egg. Many
workers elaborated on this to develop the concept of a global cytoplasmic state.
However, in Xenopus egg extracts, mitotic spindles were shown to assemble in the
absence of centrosomes and kinetochores (Heald et al. 1996), leading to the question
whether the structures were all being affected by physical contact with chromatin or
if chromatin was modifying the local state of cytoplasm. Through experimentation
it could be shown that local modification of microtubules by gradients of protein
phosphorylation was the most likely mechanism (Karsenti et al. 1984; Dogterom
et al. 1996b). Some of the regulatory gradients determining spindle assembly have
been modeled and are summarized here.

8.5.1 Stathmin

Initially, a known regulator of microtubule dynamics, stathmin which induces mi-

crotubule catastrophes, was shown on hyperphosphorylation to suppress catastrophe
frequencies of microtubule dynamics (Andersen et al. 1997). Eventually, an elegant
FRET sensor was developed that allowed the direct visualization spatial patterns
of stathmin phosphorylation and it was indeed found to localize in a gradient
166 C.A. Athale

around chromosomes in spindle assembly (Niethammer et al. 2004). This gradient

was modeled using an analytical solution of a the reaction-diffusion system of
phosphorylation due to cytoplasmic phosphatases and a cell membrane bound ki-
nase (Brown and Kholodenko 1999). The model predicted a gradient length of
4–8 m, corresponding well with the experimental extent of the gradient.

8.5.2 RanGTP Nucleation and Stabilization Gradients

Gradients of protein phosphorylation have been measured and modeled in the in-
terphasic nucleo-cytoplasmic transport machinery where the nuclear localization
signal (NLS) protein shuttling is governed by nuclear enriched RanGTP and cy-
toplasmic RanGDP (Fig. 8.6a) due to preferential localization of the kinase on
chromatin, and the phosphatase in the bulk cytoplasm (Mattaj and Englmeier 1998;
Gorlich and Kutay 1999; Gorlich et al. 2003). The cargo protein to be targeted to
the nucleus is bound non-covalently to a class of carrier proteins, the importins. The
release of the cargo protein from importins depends on binding with Ran in its GTP
bound form (RanGTP).
In mitosis, in the absence of a nuclear membrane, the RanGTP system with ki-
nase on chromatin and the phosphatase RanGAP (Ran GTPase activating protein)
in the cytoplasmic bulk produces a short range gradient of phosphorylation which
could be measured (Kalab et al. 2006). The induction of nucleation and stabilization
of microtubules by chromatin requires RanGTP. One of the nucleation factors has
been found to be RanGTP dependent TPX2 (TPX2 D targeting protein for XKLP2)
(Gruss et al. 2001, 2002) that is thought to be released in a short-range gradient
around chromosomes. The range of the RanGTP-Importin complex was measured
to estimate the range of such Importin-released factors and a value of approximately
7 m was arrived at (Caudron et al. 2005; Kalab et al. 2006). This gradient range
is also in agreement with calculations using the reaction rates and assumed diffu-
sion coefficients in the analytical solution for sub-cellular phosphorylation gradients
(Brown and Kholodenko 1999) (Fig. 8.6b).

8.5.3 Long-Range Stabilization Gradients

The measured gradients of RanGTP and its complex with Importin were mea-
sured to be relatively short-range (5–10 m), as also expected from calculations
(Fig. 8.6). However, experiments had shown that centrosomal microtubules display
polarized growth due to long-range stabilization, originating from chromosomes in a
RanGTP dependent manner. This gradient has a longer range, in the order of 20 m
(Carazo-Salas and Karsenti 2003). In order to explain this discrepancy a model was
developed by the author (Athale et al. 2008) of the dynamic instability of centro-
somal microtubules using a stochastic simulation of finite numbers of microtubules
8 Spatial Modeling of Mitotic Spindle Assembly 167

b 8μm 90 8μm
90

180 0 180 0

270 270

Fig. 8.5 Simulation of a radial aster steady state mean distribution of astral microtubules. (a) The
simulation of a radial aster in the presence of a circular chromatin structure. The dark dots at the
microtubule ends indicate those microtubule plus-ends that are growing, and the light-grey dots
indicate those that are shrinking. (b) The steady state mean distribution of astral microtubules is
symmetric for the case without a stabilization gradient (left) and asymmetric in the presence of a
gradient (right)

growing under the influence of a gradient of stabilization (Fig. 8.5). The effect of the
gradient was based on experimental values of RanGTP modification of microtubule
dynamics (Table 8.1). Its shape was derived from assumptions about the gradient
forming reactions. By comparing the simulation of aster asymmetry with published
and fresh experimental data, we were surprised to find that expected models of
phosphorylation–dephosphorylation gradients (Brown and Kholodenko 1999) did
not work, and a long range step-like gradient needed to be invoked. How such a
step-like gradient shape could be generated in bulk cytoplasm however remained to
be answered. We approached the problem by adding a hypothetical reaction mod-
ule to the existing RanGTP reaction-diffusion network and solving the system as a
partial differential equation (PDE) in one spatial dimension, assuming a spherical
geometry around chromatin. This module included a substrate W which is initially
168 C.A. Athale

a
RanBP1 RanGTP-Importinβ Hypothetical
E2
Importinβ
RanGTP-RanBP1
W Wp
RanGAP

E1-Importinβ E1

RanGDP RanGTP

DNA RCC1

Chromatin RanGTP-Importinβ

Cytoplasm
b
RanGTP (μM)

1.0 0.05
RanGTP
E1
E1 (μM)

0.5 0.03

0.0 0.00

0 10 20
Linear Hill-cooperativity Positive feedback Zero-order
1
Wp (μM)

0.5

0
0 25 50 0 25 100 0 25 100 0 25 100
Distance from chromosome (μm)
Fig. 8.6 Reaction-network of the chromatin mediated components and radial gradient from the
chromatin surface into the cytoplasm. (a) The reaction-network of the chromatin mediated compo-
nents that generate a RanGTP protein gradient around chromosomes. The addition of a hypothetical
reaction network was simulated to explain experimental data. (b) The radial gradient from the
chromatin surface into the cytoplasm produced by RanGTP and E1 are short range, while the hypo-
thetical phosphorylated substrate Wp can form a long-range gradient for some reaction topologies
(adapted from the author’s work in (Athale et al. 2008))

unphosphorylated but can be phosphorylated by a kinase E1 to Wp (Fig. 8.6a).

The phosphorylated form can in turn be dephosphorylated by a phosphatase E2,
forming a cyclic reaction network. The E1 was assumed to be sequestered by
importins and released in the presence of RanGTP, just like other NLS factors
(parameters used in Table 8.1). Testing four different kinetic regimes, linear, Hill-
cooperativity, positive-feedback and zero-order ultrasensitivity, we could show that
the zero-order ultrasensitive network was most capable of producing the effects
that agreed qualitatively with the experimental data of microtubule asymmetry near
chromatin (Fig. 8.6b). Recently a RanGTP dependent factor that stabilizes micro-
tubules has been identified to be a kinase CDK11 (Yokoyama et al. 2008). It is quite
likely that this might be the identity of the hypothesized E1 kinase, demonstrating
the power of predictive modeling to design experiments and gain insights.
8 Spatial Modeling of Mitotic Spindle Assembly 169

Table 8.1 Parameters for the stabilization of microtubules by the RanGTP gradient generating
system
Microtubule dynamic instability parameters
fcat Mitotic 0.0498/s –a
Chromosome stabilized 0.03 s1 –a
fres Mitotic 0.0048/s –a
Chromosome stabilized 0.012 s1 –a
vg Mitotic and stabilized 0.196 m=s –a
vs Mitotic and stabilized 0.325 m=s –a
Reaction diffusion parameters of hypothesized reaction systemb
Diffusion Initial
Coefficient concentration
Species Localization (m2 =s) (M)
E1 (Kinase) Cytoplasmic 10 1
E2 (Phosphatase) Cytoplasmic 10 0.1
Wt (Total substrate) Cytoplasmic 10 1
Wp (Phosphorylated substrate) Cytoplasmic 10 0
WE1 (Substrate-kinase complex) Cytoplasmic 5 0
WpE2 (Product-phosphatase complex) Cytoplasmic 5 0
a
Wilde et al. (2001), Carazo-Salas et al. (2001)
b
Athale et al. (2008)

8.6 Outlook

The preceding sections have summarized how a problem of cellular pattern forma-
tion, namely the assembly of a spindle machinery has been tackled using methods
ranging from microscopy, molecular biophysics, surface chemistry, partial differen-
tial equations, monte-carlo simulations and molecular biology. All the findings have
been built gradually into theories of modules of the system. Often older theories
have also been modified by new experimental findings and experimental techniques,
as in the example of the “random search and capture” model being replaced by a
“biased search and capture” model in spindle assembly. As is often true of models,
the simplifications however ignore the obvious details such as search and capture not
being the only mechanism necessary for spindle assembly, as demonstrated in exper-
iments where spindles assemble in the absence of centrosomes (Heald et al. 1996).
Thus, we find also new principles of pattern formation being discovered during
this process of iterative modeling and experimental comparison. An example is the
case of measurements of aster anisotropy in experiments (Carazo-Salas and Karsenti
2003; Dogterom et al. 1996a) which suggested a RanGTP dependent chromosomal
stabilization gradient. It was initially assumed the RanGTP dependent release of a
factor would be sufficient to produce this effect of stabilization based on measure-
ments and models of the gradient forming reactions (Caudron et al. 2005; Wollman
et al. 2005). However, it was only after a hybrid simulation of the reaction-diffusion
dynamics and the stochastic microtubule growth dynamics were modeled that it
became apparent that this model was not sufficient to explain the experimental data,
170 C.A. Athale

leading to both new experimental design to confirm this, as well as a new model to
propose mechanisms that might fulfil the criteria (Athale et al. 2008).
Taken together, the spindle assembly process represents a paradigm for cellular
pattern formation, since it involves all the common processes that influence cellular
patterns – chemical reaction-diffusion, cytoskeletal mechanics, motor driven motil-
ity, genetic regulation and self-organization. Additionally many of the molecular
actors have been identified by a combination of genetics, biochemistry and recon-
stitution of pure proteins to mimic the in vivo condition. The question still remains
whether an overarching model can be found to describe the assembly and function-
ing of this complex molecular machine.
In future such an approach of simulating modules of the spindle assembly system
will probably intensify driven by better and more quantitative experimental tech-
niques. This will not just make the models more valid and better tested, but also
open the way to modeling that involves combinations of different modules. Even-
tually it can be envisaged that the ambitious goal of modeling the whole structure
might become more realistic. At that point predictions could then become multi-
scale, stretching from molecular interaction dynamics to functional properties. The
integration of modeling and experimentation using quantitative biophysical tools
will thus pave the way to a systems-level understanding of not just spindle assembly,
but possibly other cellular processes that use similar modules such as cell polariza-
tion, migration and differentiation.

Acknowledgements C.A.A. was funded by BioMS and hosted in the lab of Eric Karsenti at
EMBL Heidelberg. The author is are grateful to Eric Karsenti for discussions and Francois Nedelec
for providing access to the CCC tool Cytosim for simulations of microtubule and motor dynamics.

References

A. Akhmanova and M. O. Steinmetz. Tracking the ends: a dynamic protein network controls the
fate of microtubule tips. Nat Rev Mol Cell Biol, 9:309–322, 2008
S. S. Andersen, A. J. Ashford, R. Tournebize, O. Gavet, A. Sobel, A. A. Hyman, and E. Karsenti.
Mitotic chromatin regulates phosphorylation of Stathmin/Op18. Nature, 389:640–643, 1997
C. A. Athale, A. Dinarina, M. Mora-Coral, C. Pugieux, F. Nedelec, and E. Karsenti. Regu-
lation of microtubule dynamics by reaction cascades around chromosomes. Science, 322:
1243–1247, 2008
E. Bataillon. La parthenognse des Amphibiens et la fcondation chimique de Loeb. Ann Sci Nat
Zool, 16:249–307, 1912
P. Bieling, L. Laan, H. Schek, E. L. Munteanu, L. Sandblad, M. Dogterom, D. Brunner, and
T. Surrey. Reconstitution of a microtubule plus-end tracking system in vitro. Nature, 450:1100–
1105, 2007
P. Bieling, I. A. Telley, J. Piehler, and T. Surrey. Processive kinesins require loose mechanical
coupling for efficient collective motility. EMBO Rep, 9:1121–1127, 2008
M. Bjerknes. Physical theory of the orientation of astral mitotic spindles. Science, 234:1413–
1416, 1986
L. Bourdieu, T. Duke, M. B. Elowitz, D. A. Winkelmann, S. Leibler, and A. Libchaber. Spiral
defects in motility assays: A measure of motor protein force. Phys Rev Lett, 75:176–179, 1995
8 Spatial Modeling of Mitotic Spindle Assembly 171

G. C. Brown and B. N. Kholodenko. Spatial gradients of cellular phospho-proteins. FEBS Lett,

457:452–454, 1999
K. S. Burbank, T. J. Mitchison, and D. S. Fisher. Slide-and-cluster models for spindle assembly.
Curr Biol, 17:1373–1383, 2007
R. E. Carazo-Salas and E. Karsenti. Long-range communication between chromatin and micro-
tubules in Xenopus egg extracts. Curr Biol, 13:1728–1733, 2003
R. E. Carazo-Salas, O. J. Gruss, I. W. Mattaj, and E. Karsenti. Ran-GTP coordinates regulation
of microtubule nucleation and dynamics during mitotic-spindle assembly. Nat Cell Biol, 3:
228–234, 2001
M. Caudron, G. Bunt, P. Bastiaens, and E. Karsenti. Spatial coordination of spindle assembly by
chromosome-mediated signaling gradients. Science, 309:1373–1376, 2005
Y. Chen and T. L. Hill. Theoretical treatment of microtubules disappearing in solution. Proc Natl
Acad Sci USA, 82:4127–4131, 1985
T. Clausen and K. Ribbeck. Self-organization of anastral spindles by synergy of dynamic instabil-
ity, autocatalytic microtubule production, and a spatial signaling gradient. PLoS ONE, 2:e244,
2007
E. N. Cytrynbaum, J. M. Scholey, and A. Mogilner. A force balance model of early spindle pole
separation in Drosophila embryos. Biophys J, 84:757–769, 2003
E. N. Cytrynbaum, V. Rodionov, and A. Mogilner. Computational model of dynein-dependent self-
organization of microtubule asters. J Cell Sci, 117:1381–1397, 2004
E. N. Cytrynbaum, P. Sommi, I. Brust-Mascher, J. M. Scholey, and A. Mogilner. Early spindle
assembly in Drosophila embryos: role of a force balance involving cytoskeletal dynamics and
nuclear mechanics. Mol Biol Cell, 16:4967–4981, 2005
M. Dogterom, A. C. Maggs, and S. Leibler. Diffusion and formation of microtubule asters: physical
processes versus biochemical regulation. Proc Natl Acad Sci USA, 92:6683–6688, 1995
M. Dogterom, M. A. Félix, C. C. Guet, and S. Leibler. Influence of M-phase chromatin on the
anisotropy of microtubule asters. J Cell Biol, 133:125–140, 1996a
M. Dogterom, M. A. Félix, C. C. Guet, and S. Leibler. Influence of M-phase chromatin on the
anisotropy of microtubule asters. J Cell Biol, 133:125–140, 1996b
T. Duke, T. E. Holy, and S. Leibler. “Gliding assays” for motor proteins: A theoretical analysis.
Phys Rev Lett, 74:330–333, 1995
D. W. Ehrhardt. Straighten up and fly right: microtubule dynamics and organization of non-
centrosomal arrays in higher plants. Curr Opin Cell Biol, 20:107–116, 2008
W. Flemming. Zellsubstanz, Kern und Zellteilung. Verlag Vogel, Leipzig, 1882
D. Foethke, T. Makushok, D. Brunner, and F. Ndlec. Force- and length-dependent catastrophe
activities explain interphase microtubule organization in fission yeast. Mol Syst Biol, 5:241,
2009
D. Gerlich, J. Beaudouin, B. Kalbfuss, N. Daigle, R. Eils, and J. Ellenberg. Global chromosome
positions are transmitted through mitosis in mammalian cells. Cell, 112:751–764, 2003
D. Gorlich and U. Kutay. Transport between the cell nucleus and the cytoplasm. Annu Rev Cell
Dev Biol, 15:607–660, 1999
D. Gorlich, M. J. Seewald, and K. Ribbeck. Characterization of Ran-driven cargo transport and
the RanGTPase system by kinetic measurements and computer simulation. EMBO J, 22:1088–
1100, 2003
O. J. Gruss, R. E. Carazo-Salas, C. A. Schatz, G. Guarguaglini, J. Kast, M. Wilm, N. Le Bot,
I. Vernos, E. Karsenti, and I. W. Mattaj. Ran induces spindle assembly by reversing the in-
hibitory effect of importin alpha on TPX2 activity. Cell, 104:83–93, 2001
O. J. Gruss, M. Wittmann, H. Yokoyama, R. Pepperkok, T. Kufer, H. Sillj, E. Karsenti, I. W. Mattaj,
and I. Vernos. Chromosome-induced microtubule assembly mediated by TPX2 is required for
spindle formation in HeLa cells. Nat Cell Biol, 4:871–879, 2002
R. Heald, R. Tournebize, T. Blank, R. Sandaltzopoulos, P. Becker, A. Hyman, and E. Karsenti.
Self-organization of microtubules into bipolar spindles around artificial chromosomes in Xeno-
pus egg extracts. Nature, 382:420–425, 1996
172 C.A. Athale

T. L. Hill. Introductory analysis of the GTP-cap phase-change kinetics at the end of a microtubule.
Proc Natl Acad Sci USA, 81:6728–6732, 1984
T. L. Hill. Theoretical problems related to the attachment of microtubules to kinetochores. Proc
Natl Acad Sci USA, 82:4404–4408, 1985
T. E. Holy and S. Leibler. Dynamic instability of microtubules as an efficient way to search in
space. Proc Natl Acad Sci USA, 91:5682–5685, 1994
T. E. Holy, M. Dogterom, B. Yurke, and S. Leibler. Assembly and positioning of microtubule asters
in microfabricated chambers. Proc Natl Acad Sci USA, 94:6228–6231, 1997
M. E. Janson, M. E. de Dood, and M. Dogterom. Dynamic instability of microtubules is regulated
by force. J Cell Biol, 161:1029–1034, 2003
A. P. Joglekar and A. J. Hunt. A simple, mechanistic model for directional instability during mitotic
chromosome movements. Biophys J, 83:42–58, 2002
P. Kalab, K. Weis, and R. Heald. Visualization of a Ran-GTP gradient in interphase and mitotic
Xenopus egg extracts. Science, 295:2452–2456, 2002
P. Kalab, A. Pralle, E. Y. Isacoff, R. Heald, and K. Weis. Analysis of a RanGTP-regulated gradient
in mitotic somatic cells. Nature, 440:697–701, 2006
E. Karsenti, J. Newport, R. Hubble, and M. Kirschner. Interconversion of metaphase and interphase
microtubule arrays, as studied by the injection of centrosomes and nuclei into Xenopus eggs.
J Cell Biol, 98:1730–1745, 1984
A. Khodjakov, I. S. Gabashvili, and C. L. Rieder. “Dumb” versus “smart” kinetochore models for
chromosome congression during mitosis in vertebrate somatic cells. Cell Motil Cytoskeleton,
43:179–185, 1999
P. Kraikivski, R. Lipowsky, and J. Kierfeld. Enhanced ordering of interacting filaments by molec-
ular motors. Phys Rev Lett, 96:258103, 2006
L. Laan, J. Husson, E. L. Munteanu, J. W. Kerssemakers, and M. Dogterom. Force-generation and
dynamic instability of microtubule bundles. Proc Natl Acad Sci USA, 105:8920–8925, 2008
P. Lenart, C. P. Bacher, N. Daigle, A. R. Hand, R. Eils, M. Terasaki, and J. Ellenberg. A contractile
nuclear actin network drives chromosome congression in oocytes. Nature, 436:812–818, 2005
I. W. Mattaj and L. Englmeier. Nucleocytoplasmic transport: the soluble phase. Annu Rev Biochem,
67:265–306, 1998
H. Meinhardt and P. A. J. de Boer. Pattern formation in Escherichia coli: A model for the pole-to-
pole oscillations of Min proteins and the localization of the division site. Proc Natl Acad Sci
USA, 98:14202–14207, 2001
T. Mitchison and M. Kirschner. Dynamic instability of microtubule growth. Nature, 312:237–242,
1984
F. Nedelec. Computer simulations reveal motor properties generating stable antiparallel micro-
tubule interactions. J Cell Biol, 158:1005–1015, 2002
F. Nedelec, T. Surrey, and A. C. Maggs. Dynamic concentration of motors in microtubule arrays.
Phys Rev Lett, 86:3192–3195, 2001
F. J. Nedelec, T. Surrey, A. C. Maggs, and S. Leibler. Self-organization of microtubules and motors.
Nature, 389:305–308, 1997
Francois Nedelec and Dietrich Foethke. Collective langevin dynamics of flexible cytoskeletal
fibers. New Journal of Physics, 9(11):427, 2007. URL http://stacks.iop.org/1367-2630/9/427
P. Niethammer, P. Bastiaens, and E. Karsenti. Stathmin-tubulin interaction gradients in motile and
mitotic cells. Science, 303:1862–1866, 2004
P. Niethammer, I. Kronja, S. Kandels-Lewis, S. Rybina, P. Bastiaens, and E. Karsenti. Discrete
states of a protein interaction network govern interphase and mitotic microtubule dynamics.
PLoS Biol, 5:e29, 2007
B. Raynaud-Messina and A. Merdes. Gamma-tubulin complexes and microtubule organization.
Curr Opin Cell Biol, 19:24–30, 2007
T. Surrey, F. Nedelec, S. Leibler, and E. Karsenti. Physical properties determining self-organization
of motors and microtubules. Science, 292(5519):1167–71, 2001
K. Svoboda, C. F. Schmidt, B. J. Schnapp, and S. M. Block. Direct observation of kinesin stepping
by optical trapping interferometry. Nature, 365:721–727, 1993
8 Spatial Modeling of Mitotic Spindle Assembly 173

F. Verde, M. Dogterom, E. Stelzer, E. Karsenti, and S. Leibler. Control of microtubule dynamics

and length by cyclin a- and cyclin b-dependent kinases in xenopus egg extracts. J Cell Biol,
118(5):1097–108, 1992
K. Visscher, M. J. Schnitzer, and S. M. Block. Single kinesin molecules studied with a molecular
force clamp. Nature, 400:184–189, 1999
C. E. Walczak, I. Vernos, T. J. Mitchison, E. Karsenti, and R. Heald. A model for the proposed roles
of different microtubule-based motor proteins in establishing spindle bipolarity. Curr Biol, 8:
903–913, 1998
A. Wilde, S. B. Lizarraga, L. Zhang, C. Wiese, N. R. Gliksman, C. E. Walczak, and Y. Zheng.
Ran stimulates spindle assembly by altering microtubule dynamics and the balance of motor
activities. Nat Cell Biol, 3:221–227, 2001
R. Wollman, E. N. Cytrynbaum, J. T. Jones, T. Meyer, J. M. Scholey, and A. Mogilner. Efficient
chromosome capture requires a bias in the ’search-and-capture’ process during mitotic-spindle
assembly. Curr Biol, 15:828–832, 2005
R. Wollman, G. Civelekoglu-Scholey, J. M. Scholey, and A. Mogilner. Reverse engineering of
force integration during mitosis in the Drosophila embryo. Mol Syst Biol, 4:195, 2008
H. Yokoyama, O. J. Gruss, S. Rybina, M. Caudron, M. Schelder, M. Wilm, I. W. Mattaj, and
E. Karsenti. Cdk11 is a RanGTP-dependent microtubule stabilization factor that regulates spin-
dle assembly rate. J Cell Biol, 180:867–875, 2008
Chapter 9
Cell-Centred Modeling of Tissue Behaviour

Rod Smallwood

9.1 Introduction: Towards a Virtual Cell Biology

The nature of the problem to be solved determines the modeling paradigm to use.
The problem of interest is the development of normal structure and function, and the
mechanisms that control homeostasis, in epithelial tissues, and the behaviour when
the tissue is damaged (wound healing) or the normal homeostatic mechanisms are
deranged (development of cancer). These processes are the result of the interaction
of individual cells whose behaviour depends on internal and external information.
The internal information store is the genetic material, and the external information
is chemical and physical signals. This is clearly a cell-centred description of tissue
behaviour, and equally clearly, an individual-based modeling paradigm, in which
the cell is the individual, is appropriate. Which particular individual-based model
to choose is less obvious, but in this case serendipity and logic led to the choice of
Eilenberg’s X-machine, which has sufficient power to enable multi-scale and multi-
paradigm modeling. Modeling at the cell level has in general concentrated on either
the biochemical aspects or the physical aspects, and not on the combination of the
two, which is essential for understanding cellular behaviour. The choice of epithelial
tissues was largely pragmatic – two excellent in vitro models of tissue were avail-
able, which was grown for human implantation (urothelial monolayers and skin);
and epithelial tissues are about as simple a structure as one can find in biology –
a very small number of cell types, no connective tissue, no nerve endings or blood
vessels – and they have important barrier functions and are the source of all carci-
nomas. The long-term aim is to use the epithelial cell model as a starting point for
a virtual stem cell, and to enable the expression of biological problems within a vir-
tual cell biology which cell biologists can use for experimental work in parallel with
wet biology. Ultimately, I think it is essential that the interface through the computer
scientist or the physical scientist to the virtual cell biology is removed – the cell bi-

R. Smallwood ()
Department of Computer Science, University of Sheffield, Regent Court,
211 Portobello, Sheffield S1 4DP, UK
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 175
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 9,
c Springer Science+Business Media, LLC 2011
176 R. Smallwood

ologist should be able to perform and interpret virtual experiments just as they now
perform and interpret wet biology experiments. This will require a sophisticated
interface, the ability to build cells with appropriate functions in a modular manner
and the ability to specify experimental conditions that are translated into starting
and boundary conditions for the virtual biology. The aim is not to simulate biology,
but to make testable predictions outside the parameter space of the input data. The
reader is referred to Smallwood (2009) for a general review of epithelial tissue mod-
eling, to Walker et al. (2004a, b, 2006a, b) and to Sun et al. (2007, 2008) for detailed
descriptions of the models.

9.2 Can Computation Cope with Cellular Complexity?

The first problem that becomes apparent when discussing a cell-based model of tis-
sue which incorporates molecular detail is the complexity – the probability space
provided by 30,000 genes and 105 proteins in the 1013 cells in the human body,
combined with the range of spatial and temporal scales – from 1012 m (atom) to
100 m (organism), and 109 s (molecular interaction) to 109 s (lifetime). Feytmans
et al. (2005) have shown that, if 100 genes are required to code for each func-
tion, there are 10289 possible combinations of 30,000 genes, and adding one more
gene adds 10287 new functions (there are about 1080 atoms in the universe). Clearly
some constraints have to be applied, and the constraints are evolutionary – biology
is actually highly constrained and conservative (which is why Arabidopsis and C.
elegans are useful model organisms). Despite the uncountable size of the combi-
natorial space, there are only about 200 different types of cells in mammals, and
only 9 body plans in the Animalia. Similarly, despite the enormous numbers of
proteins involved in signalling, there are only a handful of generic types of sig-
nalling – transport through ion channels and gap junctions, receptor–ligand binding,
mechanotransduction, etc. This suggests that one strategy for reducing the apparent
complexity would be to commence at a functional level, and introduce component
(mechanistic) detail only where required for the problem in hand – an integrative,
rather than reductionist, approach. Multiscale modeling can be used to introduce de-
tails only where required (and conversely, can be used to abstract away detail where
it is not required). This also implies the use of the largest length and time scales
which still give a reasonable representation of the underlying reality – an example
is the use of large heterogeneous finite elements to provide mechanical details of
the ventricles in cardiac models, in which only about 100 elements are required,
but have properties that have been determined from sub-micrometer structural mod-
els. It also appears to be possible, with cellular-level models, to get representative
behaviour from significantly fewer cells that are present in the biological model.
A skin wound containing O (106 ) epithelial cells will not heal without grafting, but
similar growth behaviour can be seen with O (104 ) cells in a computational model.
This implies that the length scale at which cellular-level behaviour can be abstracted
away to continuum-based tissue level behaviour may be as small as 1 mm.
9 Cell-Centred Modeling of Tissue Behaviour 177

9.2.1 Being Generic: Function Versus Detail

The signal transduction knowledge environment (see footnote 4) lists 2,000

molecules involved in signal transduction, and >100 different pathways. At a
generic level, these could be classified in a small number of groups:
 Movement of ions between adjacent cells via gap junctions.
 Binding of membrane-inserted proteins on the same cell or neighbouring cells.
 Binding of ligands diffusing through the extracellular space with membrane-
inserted receptors.
 Mechanotransduction: mechanical to chemical transduction at focal adhesions;
stretch-sensitive ion channels.
 ...

The description of these signalling processes does not require details of the pro-
teins involved and their relationships, unless this is needed to explore how, for
instance, the availability of a particular protein in the signalling chain (as a result of,
e.g. sequestering of the protein in a store as a result of some other reaction which is
not part of the signalling chain) affects the system level performance. Section 9.8.1
describes a model of a signalling chain which includes sequestration of one of the
proteins, and Sect. 9.8.2 describes diffusion of ligands with receptor binding and
trafficking.

9.3 Cells and Computation

The use of an individual-based approach leads to a 1:1 mapping between the bi-
ological cell and the cell’s computational representation as a finite state machine
(which I will refer to as an agent). Each agent has a local frame of reference – the
individual agent behaviour is the result of internal events (cell cycle, growth, di-
vision, etc.) which are mediated by the external environment (physical interaction
with other agents and the substrate; mechanical and chemical signalling). The agents
have a physical location defined by the centroid of the cell on a continuous scale;
a size, shape and mechanical properties; and exert and respond to forces resulting
from growth and the formation of cell–cell and cell–substrate bonds. The physical
behaviour could be embedded within the function set of the agent, but we have cho-
sen to use a physical model which is separate from, but exchanges data with, the
agent model. Extracellular signalling is of two generic types – communication with
adjacent cells (gap junctions, binding of membrane-inserted ligands and receptors)
which corresponds to local message passing between agents; and diffusible agents
such as growth factors which provide a non-local signalling mechanism and require
a diffusion model. Activities such as receptor trafficking need additional functions.
Internal signalling, gene activation, etc. can be handled by individual or differen-
tial equation-based approaches as appropriate. The physical and chemical domains
are linked by mechano-transduction, in which internal signalling is mediated by
178 R. Smallwood

Fig. 9.1 Internal and external relationships of a cell/agent

externally applied forces. Although this chapter is primarily about epithelial cells,
in a more general case, the behaviour could include electrical excitability and active
length and shape change, electrical to mechanical coupling, and also mechano-
electrical coupling (stretch-sensitive ion channels). This clearly requires complex,
multi-scale models. The cellular functions are represented in Fig. 9.1, which is the
basis for developing the set of states and functions that comprise the X-machine.

9.4 Developing a Multi-Scale Model

In principle, an individual-based model can be used at any level. The logic at the
cellular level is that the transition from cellular behaviour to tissue behaviour is an
emergent property of cellular interaction; so the modeling paradigm has to be able to
reproduce this behaviour. At a higher level than a few tens of thousands of cells, one
can argue that an ensemble average of the emergent properties results in a quasi-
deterministic (or at least stochastic) process that one can more efficiently model
using differential equations or continuum methods. This argument holds equally
at lower levels than the cell – if more than a few thousand individual molecules
are involved, the emergent detail can be abstracted away to an equation. There is
an implicit assumption in so dealing with subcellular chemical events, which is
that we are dealing with a well-stirred solution. This is clearly not the case within
the cell, which is full of mechanisms for the local transport and sequestering of
9 Cell-Centred Modeling of Tissue Behaviour 179

molecules, but pragmatism and computational load have so far dictated that the local
inhomogeneity is generally ignored. As the cellular mechanisms are evolved mech-
anisms, they presumably occupy at least a local optimum in a fitness landscape, and
therefore their particular form is important; but I am not aware of any comprehen-
sive attempt to relate subcellular morphology to function. In practice, it is clearly
impossible for any one research group to build de novo a model of any reasonably
sized biological system, and existing models and modeling paradigms will have to
be used. The linking of different modeling paradigms will be discussed later.

9.5 The Agent Basis: The Communicating-Stream X-Machine

Individual-based models have been little used in engineering contexts (the major use
has been in ecology – Grimm (1999) provides a critical review of individual-based
modeling in ecology); so a familiarity with them cannot be assumed in the way it
can be for differential equation or finite element-based models. I have elsewhere
reviewed computational models of epithelial tissue (Smallwood 2009), so will limit
my remarks to the choice of the X-machine as the formalism for the software agents.
The requirements are as follows:
 The modeling paradigm has to be robust, as the goal is to develop computational
models which are able to predict the effect of intervention in clinical problems.
 A 1:1 mapping between cells and agents was considered desirable; so the agent
had to be capable of handling cellular complexity.
 Building models de novo of the whole of cell and tissue biology is clearly not
feasible, so that linking or importing models of specific processes which had
been built using different modeling paradigms was essential.
 The physical environment is an important determinant of cellular function; so
linking or incorporating physical solvers was essential.
 The problems of state explosion had to be avoided.
 The number of cells in the smallest diameter non-healing skin wound is O (106 );
so it had to be possible to model this number of cells, which implies that paral-
lelisation was essential.
The only candidate which met all the requirements was the X-machine, which
was introduced by Eilenberg (1974). A good introduction and bibliography are pro-
vided by Stannett (2005). The communicating-stream X-machine (Fig. 9.2) is an
extension of the basic deterministic stream X-machine (Kefalas et al. 2003), which
is formally defined as an 8-tuple:

M D .˙; ; Q, M; ˚; F; q0 ; m0 /

where
 ˙ and  are the input and output finite alphabets, respectively.
 Q is the finite set of states.
180 R. Smallwood

Fig. 9.2 The communicating X-machine [after Kefalas et al. (2003)]. S i are the states,  i the
functions operating on inputs  and memory m

 M is the (possibly) infinite set called memory.

 ˚ is the type of the machine M, a finite set of partial functions  that map an input
and a memory state to an output and a new memory state  W ˙  M !   M.
 F is the next state partial function that, given a state and a function from the
type ˚, denotes the next state. F is often described as a transition state diagram.
F W Q  ˚ ! Q.
 q0 and m0 are the initial state and memory, respectively.

The addition of communication adds an input stream  and an output stream  .

The key points are as follows:
 The X-machine has been shown to be Turing-complete; so can compute anything
that is computable.
 It has a memory that restrains state explosion, and is the equivalent of the genetic
database which informs cellular function.
 It can bi-directionally communicate with other X-machines.
 The internal functions can represent processes of any level of complexity.
 The memory contains a complete listing of the state of the X-machine at each
time step, so can be used as the interface to other modeling paradigms such as a
finite element solver for resolving the forces between cells.
The X-machine can be related to the cellular functions: the memory is re-
lated to gene expression; the input and output streams are signals between cell
and environment (e.g. cell–cell signalling or the import/export of a growth factor
molecule); the states Q could be, for instance, stages in the cell cycle; and the func-
tions  define the transitions between states. These relationships are summarised
in Fig. 9.3.
The classic method of communication (Kefalas et al. 2003) is to write outputs
to rows of a communication matrix and read inputs from columns. This rapidly
becomes impracticable in a cell model. A few mm3 of tissue contains O (106 )
cells, giving an O (1012 ) communication matrix. A simple search for neighbour-
ing cells (as each cell communicates only with neighbours) is of O (n2 ), so equally
9 Cell-Centred Modeling of Tissue Behaviour 181

Fig. 9.3 Input and output relationships for the X-machine

intractable. There are many more efficient solutions. For spatially localised agents
such as cells, the cell coordinates can be used as an index into a 3D matrix of O (n),
and the search problem is then trivial.

9.6 Biology, Physics, Chemistry and Computation

If the strategy for coping with cellular complexity is to start at the functional rather
than the component level, a catalogue of essential and desirable biological functions
is required. The over-arching aim is to develop an understanding of the develop-
ment of normal structure and function in tissues from a cellular perspective, for
which a representation of the cell cycle is essential (Fig. 9.4). As a minimum this
has to include growth, the check points which control progression around the cell
cycle, entry and exit from the G0 state (cessation and resumption of cell growth as
a normal part of homeostasis, and progression to malignancy), cell division and dif-
ferentiation. The behaviour of the individual cell is a function of its environment,
both chemical and physical.
Externally, the physical environment involves the formation of bonds with neigh-
bouring cells and the substrate or extracellular matrix, and responding to forces
applied to the cell (mechano-transduction at focal adhesions; stretch-sensitive ion
channels). The chemical environment includes receptor–ligand binding at the cell
membrane (the ligands may be diffusing in the extracellular space or localised in
the membrane of a neighbouring cell), and all the mechanisms for transport across
the cell membrane (active and passive ion channels; importing and exporting re-
ceptors). Ion channels may communicate with neighbouring cells (gap junctions)
or with the extracellular space. There are models of the internal biochemistry of
182 R. Smallwood

Fig. 9.4 A grossly simplified representation of the cell cycle

individual cells that require a super-computer; so abstraction is clearly required for

a multicellular model to be computationally tractable (the abstraction may of course
be informed by the output of these detailed models). Internally, at least the sig-
nalling pathways that are implicated in the problem being studied will be required,
and other mechanisms (e.g. calcium stores) may also be required. At cell division,
the cell becomes polarised – it rounds up, becomes axi-symmetric and divides along
a plane perpendicular to the axis, which may be important in both normal and ab-
normal development. Compartmentalising cellular behaviour into biology, physics
and chemistry does not appear to me to be particularly productive, if it were not
for the fact that biologists and biochemists have largely ignored the physical mech-
anisms, without which nothing moves, so nothing happens! Physical concepts are
essential at the cell level and above to describe movement, the effect of cell bonds
and growth on the rearrangement of cells, the forces generated by contracting cells,
and mechano-transduction – the exquisite sensitivity of cells to applied forces. They
may also be necessary at a subcellular level if this detail is required to determine
the physical properties of cells or for force generation, and at membrane level for
stretch-sensitive ion channels. Chemical concepts are mainly required at subcellu-
lar or membrane level, and there are many tools from systems biology to handle
cell chemistry. The use of stochastic differential equations (Burrage et al. 2004)
or individual-based models (Pogson et al. 2008) might be desirable to handle the
relatively small numbers of molecules and localised reactions.

9.6.1 Forces on Cells

Smallwood (2009) provides a comprehensive review of computational modeling of

epithelial tissues, with particular reference to physical models; so I will confine
9 Cell-Centred Modeling of Tissue Behaviour 183

myself to more general remarks about the physical environment of the cell. My
thesis is that the physical and chemical environments of the cell are of equal impor-
tance for an understanding of cellular behaviour; so any “realistic” representation
of cellular behaviour has to include physics, biochemistry and the coupling be-
tween physics and biochemistry (mechano-transduction). The current situation is
that physical models are considerably less well developed than biochemical models.
The only area in which there has been any significant modeling of the effect of force
on cellular behaviour is in mechano-electrical feedback in cardiac myocytes (Kohl
and Noble 2008). The effect of individual cell behaviour (growth, division, apopto-
sis, motility) on tissue behaviour and the effect of tissue-level strains on individual
cell behaviour have hardly been touched upon. We are working on linking individ-
ual to continuum behaviour within continuum mechanics, image analysis, signal
processing and system identification (CMISS).1
An excellent starting point for understanding the influence of subcellular compo-
nents on the mechanics of a single cell is provided by Boal (2002). However, most of
the cellular details have to be abstracted away to achieve a computationally tractable
model containing millions of cells. The most extreme abstraction is to represent the
physical cells as quasi-incompressible spheres, possibly subject to Hertzian contact
mechanics, which are linked by a set of springs representing bonds and cell growth.
There is a probability that cells will form bonds with each other (or the substrate)
when they are sufficiently close. The simplest representation of the resulting force is
a spring joining the cell centres, with a tension proportional to the separation of the
cell membranes. If two cells are initially in contact, and one or both increase in size,
the resulting repulsive force can again be represented by a spring joining the cell
centres, with a negative tension proportional to the overlap of the membranes [the
cells clearly cannot occupy the same physical space, but the separation of growth
time steps (agent model) and force resolution time steps (physical model) leads to
the concept of “cell overlap” being resolved by the physical model]. The detail of
this simple model is provided by Adra et al. (2010) and is able to provide a plausi-
ble representation of cell movement during growth conditions which influence the
formation of cell–cell bonds (Walker et al. 2004a, b, 2006a).

9.7 Hierarchy in Computational Models

The essence of modeling is abstraction – we would learn no more from a complete

model of a cell than we would from the cell itself, and the cell computes in real
time, not orders of magnitude slower. The need for abstraction is evident from a
consideration of length and times scales – from 109 m (molecule) to 100 m (human
organism) and 109 s (molecular interaction) to 109 s (human lifetime) involving 105

1
An interactive computer program for Continuum Mechanics, Image analysis, Signal processing
and System identification (CMISS), http://www.cmiss.org.
184 R. Smallwood

Fig. 9.5 Scale separation regions

proteins. Tools are available for modeling at all of these length and time scales, but
linking them to form multi-scale models is in its infancy. One promising approach is
to decompose the multi-scale problem into a set of single-scale models, and then use
a scale-separation map to envisage the data and control flows between the single-
scale models (Hoekstra et al. 2007). For instance, the cell cycle is on a spatial scale
from individual cell components (1 m) to cell size (10 m), and a temporal
scale from hours to days, whereas cell signalling events are on a smaller length
scale (<m) and shorter time scale (ms to hours). The relationship between scales
is illustrated in Fig. 9.5.
Process A has length and time scales that place it in region 0 in the map. What
is the relationship with another process B? If B also resides in region 0, there is
no separation in either length or time, and a single model must be built which in-
cludes both processes. In the other cases (regions 1, 2, 3.1, 3.2) separate models
are possible. In region 1, the spatial scales are the same but time scales are sepa-
rated, and in region 2 the time scales are the same but the spatial scales separated.
Regions 3.1 and 3.2 are separated both temporally and spatially. In 3.1, fast events
on a small spatial scale are coupled to slow events on a longer scale (e.g. cell sig-
nalling coupled to the cell cycle); and in 3.2 slow events on a small spatial scale
are coupled to fast events on a longer scale (e.g. cellular response to flow-induced
shear stress). The scale separation map is a graph on which the vertices are the
separated processes/models, and the edges are the information and control conduits
between the processes. This is illustrated for the example of the vascular response to
the emplacement of a coronary artery stent by Evans et al. (2008). One immediate
consequence of coupling processes with differing length/time scales is that some
means has to be found to reduce the amount of information that is passed from the
finer to the coarser representation. One approach to abstracting information from the
molecular level upwards (Fig. 9.6) is to inform generalised models from the output
9 Cell-Centred Modeling of Tissue Behaviour 185

Fig. 9.6 Reducing complexity by linking individual-based and continuum models

of individual-based models – an individual-based model of molecular interaction

could inform a differential equation model of cell signalling which was incorpo-
rated into individual cell models, which in turn informs a finite element model of
the tissue.
A second approach, which we have adopted partly for pragmatic reasons (min-
imisation of effort, sharing models) and partly as this approach enables a individual
cell model to be incorporated into existing physiome models such as cardiac mod-
els, is to confine the individual-based models to the cell level and link to other
model paradigms (in which the required models have already been developed) at
other levels. In terms of the X-machine paradigm, subcellular models (e.g. intra-
cellular signalling or electrical excitation) are incorporated as members of the class
of functions ˚. Supracellular models (e.g. diffusion through the extracellular space
or resolution of the forces acting on the cells) can be called at the end of each
X-machine time step. It is usual for there to be many time steps or iterations of
the linked models for each agent time step – for the EGF signalling model consid-
ered later, the diffusion time step is about four orders of magnitude shorter than the
agent time step. Information exchange between agent and supracellular models is
186 R. Smallwood

done by reading from and writing to the X-machine memory. As far as possible,
we would want to use existing models of signalling, biochemical networks and
electrical activation, as building all the required models from scratch is not feasi-
ble. There are several hundred curated models available on the websites of CellML
(Cell Markup Language2), SBML (Systems Biology Markup Language3) and STKE
(Signal Transduction Knowledge Environment4). The key to importing these mod-
els is to incorporate an open source solver as a function call. This has been built
for COPASI (Complex Pathway Simulator5 ), which can be used to develop ordinary
differential equation models of biochemical reactions, and can import SBML and
CellML models (Adra et al. 2010; Sun et al. 2010). We are currently developing a
similar function call to J-Sim6 which will enable the import and solution of partial
differential equation models such as CellML models of electrical activation. So far, I
have made the implicit assumption that the resolution of the forces between the cells
is done globally, i.e. the size, position, number and type of bonds, and physical prop-
erties of each cell are passed to a global solver that resolves the forces throughout the
cell mass, and returns the updated parameters to the agent model. It is also possible
to resolve the forces on a local scale. Each cell only experiences the forces imposed
by its nearest neighbours; so force resolution could be a function call within each
cell. A small scale test of this with 2,000 cells in a monolayer (Hose, personal com-
munication) demonstrated that two passes over the whole cell mass would resolve
the forces. For the particular physical model used, the local solution was about a fac-
tor of ten slower than the global solution. However, the local solution scales linearly
whereas the global solution scales as n2 ; so for a typical simulation with 104 cells
the local solution would be much more efficient. An additional advantage is that,
as it is embedded in the X-machine, it is inherently parallel. Nevertheless, we have
chosen to not pursue this route because we see considerable advantages in being
able to populate the properties of finite element representations of tissue properties
in an existing package from the cellular level, and are pursuing X-machine to fi-
nite element integration in CMISS (see footnote 1), which is becoming a de facto
standard for use in Physiome projects.

9.8 Examples at Molecular and Cell Level

As examples, I will discuss three projects. Several papers have been published on
these three examples; so I will not describe the models or results in detail, but will
use them to illustrate the more general discussion above. The three examples are

2
http://www.cellml.org.
3
http://sbml.org.
4
http://stke.sciencemag.org/.
5
http://www.copasi.org.
6
http://j-sim.org.
9 Cell-Centred Modeling of Tissue Behaviour 187

NF- B signalling (individual-based modeling of molecular interaction); the growth

of a monolayer of epithelial cells which includes signalling, diffusion and the res-
olution of physical forces; and the growth of multilayered epithelial cells and the
generation and healing of a wound.

9.8.1 NF-B Signalling

This is, by a considerable margin, the simplest of the examples, as the molecules do
not have any internal mechanisms and their physical interactions are ignored. Part of
the signalling chain was modeled (Pogson et al. 2006). NF- B (a protein) is held in-
active in the cytoplasm by an inhibitor I B – the two molecules are bound together
to form a complex. If signalling is initiated from outside the cell (there are three
mechanisms to do this), an enzyme IKK (I B-kinase) starts to degrade I B, releas-
ing NF- B, which can then be transported to the nucleus. Within the nucleus, NF- B
activates genes which control the production of I B. The I B is returned to the cy-
toplasm and inactivates NF- B. This is clearly a feedback system with a time delay,
and therefore has the potential to oscillate, as has been demonstrated (Nelson et al.
2004). The starting point for the individual-based model is the individual molecule
(various receptors were also modeled, but the principles are the same and so they
will be ignored).
If we start with a simple chemical reaction:

ACB •C

we can write down the differential equations and solve them. In the individual-based
model, we populate the reaction vessel with a known number of molecules A and B
which perform random walks (Brownian motion) throughout the vessel, interact and
have a probability of forming the product C. Similarly, C has a probability of dis-
associating. However, if our molecules are point objects, they have an infinitesimal
chance of interacting; so we assign a pseudo-volume to the molecules which de-
fines the interaction probability (analogous to the collision cross-section in nuclear
physics). The pseudo-volume is clearly related to the forward reaction constant k1 ,
and it is shown by Pogson et al. (2006) that the radius of the reaction volume is
given by:
r
3 3kt
rD
4 103 L

where k is the reaction rate, t the time step and L the Avogadro number. The data
against which the model was compared were derived from measurements on single
cells in which the proteins of interest had been labelled with fluorescent tags, so
that the dynamics could be quantified. The model behaviour was incorrect when
188 R. Smallwood

the experimental ratio of NF- B to I B was used. It was suggested that there was
a mechanism for sequestering the inhibitor I B within the cytoplasm, and that this
could be binding to the cytoskeleton. A rudimentary cytoskeleton was added to the
model together with a mechanism for reversibly binding I B and actin (Pogson
et al. 2008). The model then gave the correct dynamics with the correct total I B to
NF- B ratio, and the proportion of bound I B was subsequently confirmed by wet
biology experiment.
This model begins to address the issue of locality in biological models by spa-
tially locating all of the agents. In particular, TIR receptors and nuclear importing
and exporting receptors were localised in two concentric spherical shells repre-
senting the cell membrane and the nuclear membranes, respectively. It has been
demonstrated (unpublished work) that transforming the spherical cell to a more real-
istic shape (determined from the confocal microscopy) does not alter the dynamics.
The actin cytoskeleton was modeled as a random mesh of filaments, and all of the
proteins were localised in space. This only partially addresses localisation – reac-
tions involving receptors and cytoskeleton could only take place at membrane and
cytoskeleton positions, but the reactions in the cytoplasm and nucleus took place in
what was effectively a well-stirred but dilute solution. The individual-based model
appears to be inherently robust, in which it is impossible to achieve non-physical
results such as negative concentrations, which can be achieved by an inappropriate
choice of solver for differential equations (one could argue that this demonstrates
a lack of competence!). However, although localisation would appear to be a com-
pelling argument for the use of an individual-based model of cell signalling, and the
biological confirmation of the sequestration of I B supports this, I am not aware of
any demonstration that an approach that allows localisation of reactions has yielded
important new results.

9.8.2 Urothelium Monolayer Growth

The urothelium is the epithelial tissue which lines the urinary bladder. In vitro, the
cells are grown in an environment which results in a contiguous monolayer of cells.
There is only one type of cell. The initial model (Walker et al. 2004a) explored the
growth of cells with differing concentrations of external calcium – a low calcium
regime in which the binding protein, E-cadherin, is not expressed, so that the cells
do not form bonds with each other; and a physiological calcium regime in which E-
cadherin is expressed so that the cells do form bonds. The individual cells possessed
a cell cycle, increased in size throughout the cell cycle, and a simple physical model
was included. If, during the time step of one agent, the growth of a cell resulted in
its overlapping with another cell, a repulsive force was generated proportional to
the amount of overlap. If a cell formed a bond with another cell (the probability of
bonding depended on the distance between cells, and followed a sigmoid curve), an
attractive force is generated proportional to the distance apart from the cells. The
9 Cell-Centred Modeling of Tissue Behaviour 189

size, position and number of bonds for each cell was exported to an external solver,
the forces resolved and the new size, and position of the cells was passed back to
the agent model.
The growth curves generated by this model were similar to the in vitro growth
curves, and the final cell density also varied with exogenous calcium in a simi-
lar manner to the in vitro cell density. The model did not correctly reproduce the
differences in growth rate with different levels of exogenous calcium, and it was
postulated that this was a result of the effect of epidermal growth factor (EGF),
which was not included in the model. Cells which are growing both produce their
own EGFR-binding ligands and respond to exogenous EGF, which diffuses through
the extracellular space, generating variations in concentration which depend on lo-
cal cell activity. Incorporation of EGFR signalling therefore requires the addition of
two new models: a partial differential equation model of diffusion of EGF through
the extracellular space; and an ordinary differential equation model of the binding
of ligands to receptors at the cell surface, and the turnover of the receptors. There
are thus two global models – diffusion and force – and two local models – agent
and receptor binding and trafficking. The resulting model predicts the role of EGF
in cell growth in vitro (Walker et al. 2006a).
The case for the use of individual-based models of cells is far more compelling
than for molecules, and there are many examples of cellular automata and agent-
based models, and the influential Cellular Potts model is also an individual-based
model (see Smallwood (2009) for details). The development of structure and func-
tion in tissues is an emergent property of the behaviour of individual cells, and it is
difficult to see how emergence could be captured other than by considering the cells
as individuals. The division essentially is between processes or scales in which cell
growth, division and death are important, and those in which they are not. Examples
of the former are developmental biology, wound healing and the mechanisms which
regulate cell number. An example of the latter is the coupled electrical and me-
chanical activity of the ventricular wall, where the effect of changes at a cellular or
subcellular level (e.g. mutations in ion channels) is important, but can be considered
as affecting tissue properties and not the cellular interaction per se.

9.8.3 Epidermis Multilayer Growth

The epidermis is the outer layer of the skin and contains three different cell
types – keratinocytes, fibroblasts and melanocytes. We have developed computa-
tional models of monocultures of keratinocytes and co-cultures of keratinocytes and
fibroblasts, and have compared the behaviour of the computational models with in
vitro monocultures and co-cultures (Sun et al. 2007, 2008). A major interest in the
case of skin is the response of the cells to wounding. In vivo, a skin wound which
is more than about 2 cm diameter will not heal. In vitro, the model to explore this
is a scratch wound – cells are grown to confluence, a pipette is drawn across the
190 R. Smallwood

dish to remove a strip of cells, and the resulting behaviour of the cells is monitored.
Transforming growth factor (TGF)-ˇ is involved in the control of differentiation and
proliferation in most cells, and has other functions as well, and there are contradic-
tory results in the literature. In order to explore these effects, a TGF-ˇ signalling
model has been included in the epidermal model using the Copasi function call in
the X-machine. The resulting computational model has been used to grow the cells
to produce a full-thickness, properly structured and differentiated epidermis, a sec-
tion of the cells and the basement laminar have been deleted to emulate the process
of producing a scratch wound, and the re-epithelialisation of the wound has been
followed. A narrow wound will heal, giving a properly structured tissue, but a wide
wound fails to heal (Adra et al. 2010; Sun et al. 2010).
The skin model introduces more than one family of complex agents to represent
the different cell types (the agents in the signalling model are simple, as they are
the lowest level in the hierarchy). The agents of course have the same structure,
but the memory, functions and states may differ for different types of cells. This is
an important point – the agent model of the cell is generic and could therefore in
principle represent any type of cell, including totipotent or pluripotent stem cells.
The challenge is to develop the agent model of the cell in such a way that it retains
this versatility.

9.9 A Framework for Multi-Scale Modeling

A software environment, flexible large-scale agent modeling environment (FLAME7 ),

has been developed for the building and operation of individual-based models, with
the ability to link to other modeling paradigms, and is freely available to academic
users from the FLAME website (see footnote 7). In addition to cell modeling, the
environment has been used for social insect modeling and also for financial model-
ing. The models are specified in a markup language, XXML (X-machine Markup
Language). The functions are written in C, and the initial conditions are also an
XML file, which is the initial memory content of the X-machines. At each step, the
memory is written to an XML file, and the series of XML files contain a complete
history of the model run (Fig. 9.7). A parser is available to generate C code from the
XMML model description, the functions file and library files. As MPI is used for
the message passing between agents, the realisation is inherently parallel and can
be compiled for Windows and Linux environments, to run on stand-alone, vector,
parallel and grid machines. More information, documentation and references are
available on the FLAME website.

7
http://www.flame.ac.uk.
9 Cell-Centred Modeling of Tissue Behaviour 191

Fig. 9.7 Flexible large-scale agent modeling environment (FLAME)

9.10 Describing Individual-Based Models

A central feature of the scientific method is that if work cannot be repeated, it is

not accepted. The implication is that the data that inform experimental work, and
the tools which are used, should be freely available. This is certainly not currently
the case for many of the complex computational models used in biology, and the
difficulty of adequately describing complex models is a significant barrier to mak-
ing information more freely available. Both CellML and SBML have published
standards for model description in machine-readable formats, and maintain a large
repository of curated models (i.e. models in which the equations and data have been
checked, so that the model will run in a suitable software environment). There is
also a web-based tool (SBML2LaTeX) available to produce human-readable text
from machine-readable SBML. The situation with individual-based models is less
advanced, with no widely agreed model description language and no public repos-
itory for models. Grimm (2006) have suggested a standard protocol for describing
individual-based and agent-based models which they have named the Overview, De-
sign Concepts, Details (ODD) protocol. While this is an excellent beginning, it has
been formulated for ecological models, which in general appear to be very much
simpler than biological models, are not multi-scale and multi-paradigm, and do not
use a platform-independent description language. If we examine the model struc-
tures described above, a model description would require the following:
 A natural language description of the model [as used for the SBML repository
or described by Grimm (2006)]. This should also include the data and control
conduits to models which are lower and higher in the hierarchy, and between
local and global models at a single level.
192 R. Smallwood

 A machine-readable description of the core agent model(s) – the de facto

standard being a mark-up language.
 A set of initial conditions, also machine-readable.
 A set of functions (currently, for our X-machine models, these are in C, but
should be in a platform-independent machine-readable format, e.g. MathML).
 A description of the interface(s) between agent model(s) and other models, in-
cluding data exchange requirements and timing details.
 A machine-readable description of ordinary differential equation models (SBML
or CellML) and partial differential equation models (CellML).
 A description of other model types, e.g. diffusion models and finite element
models.

9.11 Visualisation and Graphical Output

Cell biology is a very visual discipline, with images ranging from conventional
optical microscopy to electron microscopy, and a host of more recent imaging tech-
niques such as confocal microscopy, aided by stains in the case of fixed (dead) tissue
and fluorescent proteins for living tissues. Quantification includes techniques such
as Western blots for dead tissue and measuring fluorescence levels for living tissue.
Measurements are often made on a large number of cells to achieve the requisite
sensitivity, which can mask considerable individual variation. The goal has to be to
obtain time series from many individual cells (so that parameter variability can be
ascertained), with multiple data points so that cellular dynamics can be compared to
model dynamics. The primary target for modellers is to produce outputs that can be
directly related to measures that can be made on wet biology, e.g. comparison with
images or the concentrations and locations of proteins acquired from images. It is
too often assumed that, because the output from the computational model looks like
a biological image, the processes that produced the output must also be the same.
It is a necessary, but not sufficient, condition that the outputs look the same, but
it is also necessary to demonstrate that the processes are the same as well, which
requires time series data.

9.12 Repeatability, Sensitivity Analysis and Validation

A physicist would consider most biological data to be of fairly dubious reliability

– parameters have often been measured in different species and the values may
vary by an order of magnitude, some parameters may be impossible to measure
directly so that their magnitude has to be inferred from indirect measurements, and
some may not be known at all so that an estimate (an educated guess!) has to be
made. In this situation, a sensitivity analysis is essential, but for complex models
the combinatorial explosion means that a complete analysis is impractical. There is
9 Cell-Centred Modeling of Tissue Behaviour 193

a limited literature on the validation of complex models, and the area needs more
work if computational models are to be routinely used to predict the outcomes of
drug trials or other interventions in disease processes. Sornette et al. (2007) provides
a general framework for model validation with examples taken from non-biological
fields, and Marino et al. (2008) have recently written a comprehensive account of
techniques for validating computational models.

9.13 Lessons Learned

 Multi-scale, multi-paradigm modeling is possible.

 Logic-based (i.e. rule-based) agents are computationally lightweight and very
large numbers can be modeled (equivalent to macro-scale pieces of tissue).
 Mathematical functions or more complex models will dominate the computa-
tional load.
 Encapsulation of complex models within higher level models is possible.
 Resolving forces between cells can be done, and tools to inform the constitutive
equation of finite element models from individual cell models are being devel-
oped.
 There are serious validation issues which need to be addressed.
 The need for validation (in particular, for time series and robust individual cell
data) is changing biological experimentation.

Acknowledgements The opinions expressed are my own, but have benefitted from discussions
with and work by Jenny Southgate, Mike Holcombe, Sheila Mac Neil, Dawn Walker, Simon Coak-
ley, Mark Pogson, Sun Tao, Nik Georgopoulos, Phil McMinn, Salem Adra, Des Ryan, Goodarz
Kodabakshi, Rod Hose and Pat Lawford, all of whom I wish to acknowledge and thank.

References

S. Adra, T. Sun, S. MacNeil, M. Holcombe, and R. Smallwood. Development of a three dimen-

sional multiscale computational model of the human epidermis. PLoS ONE, 5(1):e8511, 2010.
D. Boal. Mechanics of the Cell. Cambridge University Press, Cambridge, 2002.
K. Burrage, T. Tian, and P. Burrage. A multi-scaled approach for simulating chemical reaction
systems. Progress in Biophysics and Molecular Biology, 85:217–234, 2004.
S. Eilenberg. Automata, Languages and Machines, Vol. A. Academic, London, 1974.
D. J. W. Evans, P. V. Lawford, J. Gunn, D. C. Walker, D. R. Hose, R. H. Smallwood, B. Chopard,
M. Krafczyk, J. Bernsdorf, and A. Hoekstra. The application of multiscale modelling to the
process of development and prevention of stenosis in a stented coronary artery. Philosophical
Transactions of the Royal Society A, 366:3343–3360, 2008.
E. Feytmans, D. Noble, and M. Peitsch. Genome size and numbers of biological functions. Trans-
actions on Computational Systems Biology, 1:44–49, 2005.
V. Grimm. A standard protocol for describing individual-based and agent-based models. Ecologi-
cal Modelling, 198:115–126, 2006.
194 R. Smallwood

V. Grimm. Ten years of individual-based modelling in ecology: What have we learned and what
could we learn in the future? Ecological Modelling, 115:129–148, 1999.
A. Hoekstra, E. Lorenz, J. Falcone, and B. Chopard. Towards a complex automata framework for
multi-scale modeling: Formalism and the scale separation map. In ICCS 2007, Part I, Lecture
Notes in Computer Science, Springer, volume 4487, pages 922–930, 2007.
P. Kefalas, G. Eleftherakis, and E. Kehris. Communicating X-machines: from theory to prac-
tice. In Advances in Informatics, Lecture Notes in Computer Science (ed. Y. Manolopoulos,
S. Evripidou, and A. Kakas.), Springer, volume 2563, pages 316–335, 2003.
P. Kohl and D. Noble. 97 (2008) 159-162 editorial life and mechanosensitivity. Progress in Bio-
physics and Molecular Biology, 97:159–162, 2008.
S. Marino, I. B. Hogue, C. J. Ray, and D. E. Kirschner. A methodology for performing global
uncertainty and sensitivity analysis in systems biology. Journal of Theoretical Biology, 254:
178–196, 2008.
D. E. Nelson, A. E. C. Ihekwaba, M. Elliott, J. R. Johnson, C. A. Gibney, B. E. Foreman, G. Nelson,
V. See, C. A. Horton, D. G. Spiller, S. W. Edwards, H. P. McDowell, J. F. Unitt, E. Sullivan,
R. Grimley, N. Benson, D. Broomhead, D. B. Kell, and M. R. H. White. Oscillations in NF-kB
Signaling Control the Dynamics of Gene Expression. Science, 306:704–708, 2004.
M. Pogson, R. Smallwood, E. Qwarnstrom, and M. Holcombe. Formal agent-based modelling of
intracellular chemical interactions. Biosystems, 85:37–45, 2006.
M. Pogson, M. Holcombe, R. H. Smallwood, and E. Qwarnstrom. Introducing spatial informa-
tion into predictive nf-kb modelling – an agent-based approach. PLoS ONE 3(6): e2367.
doi:10.1371/journal.pone.0002367, 3(6), 2008.
R. H. Smallwood. Computational modelling of epithelial tissues. Wiley Interdisciplinary Reviews
Systems Biology, URL http://www3.interscience.wiley.com/journal/122305009/abstract, 2009.
D. Sornette, A. B. Davis, K. Ide, K. R. Vixle, V. Pisarenko, and J. R. Kamm. Algorithm for model
validation: Theory and applications. PNAS, 104:6562–6567, 2007.
M. Stannett. Theory of x-machines. http://x-machines.com/, 2005. URL http://x-machines.com/.
T. Sun, P. McMinn, S. Coakley, M. Holcombe, R. H. Smallwood, and S. MacNeil. An integrated
systems biology approach to understanding the rules of keratinocyte colony formation. Journal
of the Royal Society Interface, 4:1077–1092, 2007.
T. Sun, P. McMinn, M. Holcombe, R. Smallwood, and S. Macneil. Agent-based modeling helps
in understanding the rules by which fibroblasts support keratinocyte colony formation. PLoS
ONE, 3, 2008. doi: e2129doi:10.1371/journal.pone.0002129.
T. Sun, S. Adra, R. Smallwood, M. Holcombe, and S. Macneil. Exploring the hypotheses of the
actions of TGF-“1 in epidermal wound healing using a 3D computational multiscale model of
the human epidermis. PLoS ONE, 4(12):e8515, 2010.
D. C. Walker, J. S. Southgate, G. Hill, M. Holcombe, D. R. Hose, S. M. Wood, S. MacNeil, and
R. H. Smallwood. The Epitheliome: modelling the social behavior of cells. BioSystems, 76:
89–100, 2004a.
D. C. Walker, G. Hill, S. M. Wood, R. H. Smallwood, and J. S. Southgate. Agent-based modelling
of wounded epithelial cell monolayers. IEEE Transactions on Nanobioscience, 3:153–163,
2004b.
D. C. Walker, S. Wood, J. S. Southgate, M. Holcombe, and R. H. Smallwood. An integrated agent-
mathematical model of the effect of intercellular signaling via the epidermal growth factor
receptor on cell proliferation. Journal of Theoretical Biology, 242:774–789, 2006a.
D. C. Walker, T. Sun, S. MacNeil, and R. H. Smallwood. Modeling the effect of exogenous calcium
on keratinocyte and HaCat cell proliferation and differentiation using an agent-based computa-
tional paradigm. Tissue Engineering, 12:2301–2309, 2006b.
Chapter 10
Interaction-Based Simulations for Integrative
Spatial Systems Biology

Antoine Spicher, Olivier Michel, and Jean-Louis Giavitto

10.1 Introduction

It was Fermi et al. (1965) who proposed that computers, instead of simply
performing standard calculus, could be used to study and test a physical idea.
This was the introduction, in 1955, of the idea of numerical experiments, also called
in silico experiments by biologists.
This epistemological and sociological change had far reaching consequences,
providing to systems biology a unique tool in the investigation of biological phe-
nomena. Computer modeling and simulation give to the biologist an access to
“experimental results” that cannot be provided by direct experiments because of
practical, economical or ethical reasons. However, as biologists realize the limita-
tions of informal, intuitive analysis of complex systems (McAdams and Shapiro
1995; Von Dassow et al. 2000), the computer is no longer used only to perform a
computation that cannot be done analytically or by hand: its is also used to check
and compare theoretical models, to systematically investigate the consequences of
an hypothesis, to explore the possible range of the parameters, and to record, an-
alyze, control and summarize some elements of the (possibly non-deterministic)
behavior of a complex biological system.
Within biology, systems biology is a particularly demanding application domain
since it requires to integrate several models coming from unrelated area of sci-
ence like mechanics, chemistry, etc. The computer modeling and simulation of such
systems require the coupling of several model fragments specifying deterministic
or stochastic interactions between the system’s entities to represent continuous or
discrete evolution. For instance, the modeling of the growth of the meristem at
a cellular level (Barbier de Reuille et al. 2006a) requires the coupling of molec-
ular mechanisms (e.g., chemical reaction, diffusion, active transport), mechanical
stresses, developmental changes, and genetic regulation.

A. Spicher ()
LACL – EA 4219 – Université de Paris 12, Paris EST – 61 avenue du Général de Gaulle,
94010 Créteil Cedex, France
e-mail: [email protected]

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 195
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3 10,
c Springer Science+Business Media, LLC 2011
196 A. Spicher et al.

Computer science has developed (or appropriated) many languages and tools to
help build models of real-world processes and to relate different models that operate
on different levels of abstraction and various spatial and time scales. In this chap-
ter, we advocate the use of a rule-based framework based on spatial interactions
as a unifying framework for the concise and expressive simulation of a broad class
of biological systems. We will address related issues such as: Can the same frame-
work be used to model deterministic and stochastic systems? Do we need different
frameworks for the expression of continuous and discrete systems? Could the same
approach allow the natural and concise expression of various theoretical approaches
(for the purpose of simulation)? An answer to such questions cannot be derived the-
oretically, but convincing elements can be provided through paradigmatic examples.
This chapter is then organized as follows.
Section 10.2 discusses some of the requirements of systems biology models,
the growing role of agent-based models and the current focus put on the notion
of interaction. We emphasize also the need to handle explicit spatial relationships.
Section 10.3 presents MGS, a rule-based, spatial interaction-oriented, exper-
imental programming language dedicated to the simulation of a broad class of
biological systems.
Section 10.4 introduces the running example we use to illustrate the versatility of
the rule-based approach: a synthetic multicellular bacteria or SMB. This example
comes from a project presented at the International Genetically Engineered Ma-
chine Competition (iGEM)1 contest in synthetic biology. SMB combines diffusion,
genetic regulation, and signalling in a population.
Section 10.5 illustrates the use of the MGS rule-based approach with the devel-
opment of several models of the SMB. Each model focuses on a specific time scale
using a dedicated theoretical framework. We show how the MGS approach, em-
phasizing the notion of spatial interactions, is able to express concisely in the same
unified and uniform simulation framework, stochastic and deterministic models, and
discrete and continuous ones.
A short presentation of the perspectives and challenges opened by this work con-
cludes this chapter.

10.2 Computer Modeling and Simulation in Integrative

and Spatial Systems Biology

In this section, we sketch several approaches in the modeling of biological systems.

We propose to base a unifying simulation framework on the spatial organization of
the interaction between the entities that compose the system. An experimental pro-
gramming language based on this idea is proposed in the next section and illustrated
by several examples in the second part of this chapter.

1
The SMB: Synthetic Multicellular Bacterium (iGEM’07) Paris Team Web site: http://parts.mit.
edu/igem07/index.php/Paris.
10 Interaction-Based Simulations for Systems Biology 197

10.2.1 Dynamical Systems in Systems Biology

Biological processes are often modeled as dynamical systems (Smith 1999). At any
point of time, a dynamical system is characterized by a set of state variables. The
evolution of the state over time is specified through a transition function which
determines the next state of the system (over some time increment) as a function of
its previous state and, possibly, the values of external variables (input to the system).
The evolution function can be generalized to an evolution relation to handle non-
deterministic (e.g., stochastic) evolution.
Various mathematical framework with diverse properties can be considered to
formalize a dynamical system. For instance, state variables may take values from
a continuous or discrete domain. Likewise, time may advance continuously or in
discrete steps. Some examples of dynamical systems characterized by different com-
binations of these features are listed in Table 10.1. Other combinations exist and are
not listed: the disintegration of a radio-active atom is a continuous-time Markov
process with discrete state for instance.
These various formalisms can be applied to the same system to capture differ-
ent aspects of the system’s evolution. For example, the same reaction–diffusion
process (Turing 1952) in a tissue can be modeled in continuous space by partial
differential equations (PDE) or in a discrete space by a system of coupled ordi-
nary differential equations (ODE), where the state variables are the concentration of
morphogens in each cell (Turing did both in his seminal paper). Reaction–diffusion
processes can be also modeled by iterated mapping, sometimes called “continu-
ous automata”, a variant of von Neumann’s cellular automata (CA) (Von Neumann
1966) where a cell is described by real-valued local concentrations (Turk 1991).
And totally discrete (space, time, and state) models of reaction–diffusion have
also been proposed, for instance, in Greenberg and Hastings (1978).

10.2.1.1 The Need of a Unifying Simulation Language

The previous example shows that a simulation workbench for integrative biology
cannot support a unique theoretical framework. In addition, even confronted to the
development of one specific simulation, the programmer must cope with the wide

Table 10.1 Formalisms used to specify dynamical systems.

Some formalisms used to specify dynamical systems accord-
ing to the discrete or continuous nature of time, space, and
state variables. The “space” row is explained in Sect. 10.2.2.2.
(C: continuous D: discrete)
Discrete Iterated Finite
or continuous PDE ODE mappings automata
State C C C D
Time C C D D
Space C D D D
198 A. Spicher et al.

variety of biological entities (genes, proteins, membranes, cells, tissue, etc.). They
cannot be described in a unique formalism, and yet they must be placed in a single
simulation framework. This is also the case in multi-scale modeling where models of
the same system at different scales can have fundamentally different characteristics
(e.g., deterministic vs. stochastic).
These observations do not imply that only a general programming language can
be used for the implementation of simulations in systems biology. As a matter of
fact, the notion of dynamical system is a very general one, but nevertheless, it may
receive some specific support that motivates the development of domain-specific
languages (DSLs). DSLs offer, through appropriate notations and abstractions,
expressive power focused on, and usually restricted to, a particular problem domain.
We believe that it is possible to provide abstractions and notations generic enough
to encompass and unify the variety of formalisms needed in systems biology. Such
DSLs will support the expressive representation of various kind of states, time, and
evolution functions as well as the building of coupled heterogeneous models such as
discrete/continuous or stochastic/deterministic dynamical models. The programmer
will be able to express the various models in a concise and expressive way, making
easier to debug, tune, and evolve the simulations. Such DSL makes also possible to
relate models through their implementation.

10.2.1.2 State and Evolution Function in Systems Biology

From the previous presentation, it is obvious that a DSL dedicated to simulation

in systems biology must support in one way or the other the notions of state and
evolution function. However, these two simple notions must be looked in a fresh way
in the context of systems biology. Indeed, most of the systems considered in biology
consist of populations of interacting entities. A good example is a biological cell
modeled by a system of molecules that react and interact to form (other) molecules
and molecular machines.
It is customary to abstract over these entities and use state variables to denote
macroscopic observables or population level properties like a global concentration
or a temperature.2 It is assumed that as the population size increases, the behavior
of the biological system is asymptotic to that of this state-variable model.
The terms aggregate or mean-field are sometimes used to qualify this approach.
It allows a concise expression of the model, and despite severe limitations, mean-
field approximations have been standard methodology for modeling populations of
interacting entities, especially for large and homogeneous populations. One reason
is that no viable alternative existed before the widespread availability of inexpensive
computing power.

2
Relying on a mean-field approach where the idea is to replace all interactions to any entity with
an average interaction, reducing any multiple entities problem into an effective one-entity problem.
10 Interaction-Based Simulations for Systems Biology 199

However, aggregate models rely on two assumptions that must be seriously scru-
tinized in systems biology:
1. The state space can be described a priori and remains fixed.
2. The global evolution function can be defined explicitly.
We examine these two assumptions in the remainder of this section.

10.2.1.3 Dynamical Systems with a Dynamical Structure

Very often the state space of the considered biological process cannot be described
a priori. The reason is that the structure of the biological system and therefore its
description (by a set of state variables) may itself vary over time, as pointed out
by Giavitto et al. (2002a). An example is given by the development of an embryo.
Initially, the state of the system is described solely by the chemical state of the egg.
After several divisions, the state of the embryo is given not only by the chemical
state of each cell but also by their spatial arrangement. The number of cells, their
spatial organization, and their interactions evolve constantly in the course of the
development and is not handled by one static structure. It means that the phase
space used to characterize the structure of the state of the system at each time step
must be computed jointly with the running state of the system.
The dynamicity of the structure of a biological system have been repeatedly
emphasized, and several formalisms have been proposed to specify both the evo-
lution of states and the evolution of the structure. Examples include: the concept
of (hyper)-cycle introduced by Eigen and Schuster in the study of auto-catalytic
networks (Eigen and Schuster 1979), the notion of autopoietic systems formulated
by Varela et al. (1974), Luisi (2003), the variable structure system theory devel-
oped in control (Itkis 1976), or the concept of organization introduced by Fontana
and Buss (1994) to formalize and study the emergence of self-maintained functional
structures in a range of chemical reactions.
We call such systems dynamical systems with a dynamical structure3 or (DS)2
in short (Giavitto and Michel 2002c, 2003; Giavitto 2003). Biological examples
include the production of molecules and their dynamic association with multimolec-
ular complexes (Fontana 1992) or the birth and death of cells with their mechanical
constraints and signaling relations within a developing organism (morphogenesis).

10.2.1.4 Local Interactions

A consequence of a dynamical structure is that a global evolution function cannot

be specified. As a matter of fact, if the set of variables that describe the system

3
Bailly and Longo (2006) recognize the importance of this class of dynamical systems and call
it “dynamicité auto-constituante” (which could be translated to “self-producing dynamicity”), a
distinctive feature of living organisms.
200 A. Spicher et al.

cannot be known in advance, it is impossible to specify a global evolution function

(a dynamical structure is not mandatory to prevent the explicit definition of a global
evolution function).
This does not mean that the (global) evolution function does not exist: it simply
cannot be defined explicitly. This is the case when the individual (local) interactions
between the system’s entities are well characterized, but the corresponding global
evolution function cannot be deduced. The macroscopic (global) evolution of the
system must be computed as the “integration” of all the various local and dynamic
interactions between entities.

10.2.2 Individual-Based Models and Their Simulations

Individual-based models (Lynch 2008), also called agent-based models, propose

an alternative approach to mean-field approximation. Such models describe a sys-
tem from the perspective of its constituent units and focus on the representation
of the evolution of each individuals that appears in the system. As a consequence,
they tackle more easily the enormous modeling difficulties raised by the dynamical
structure of biological systems.
Individual-based approaches attract a renewed interest and become viable al-
ternatives because of the increasing availability of inexpensive computing power.
However, they have their own drawbacks. Their mathematical analysis appears
to be at least as difficult as analysing aggregate variables models, and the sim-
ulation remains the main tool to study the system’s evolution and for reaching
conclusions.
Thus, beyond aggregate models, a simulation language dedicated to systems bi-
ology must be able to implement individual-based models.

10.2.2.1 Multi-Agent Implementation

Multi-agent systems (MAS) (Woolridge and Wooldridge 2001) are often advocated
as the tool of choice for the implementation of individual-based models (Spicher
et al. 2009). A MAS is a collection of autonomous decision-making entities called
agents. Each agent individually assesses its situation and makes decisions on the
basis of a set of rules.
It is easy to use an agent in a MAS to represent the state of an entity part of the
modeled system. The global state of the system is then the set of the state of each
agent that composes the system.
However, MAS provide no support for the notion of interaction. Several entities
are engaged simultaneously in an interaction while agents are supposed to evolve
autonomously. Admittedly, in determining its evolution, an agent takes into account
its neighbors. But it cannot take into account, for example, the evolution of its neigh-
bors, which can be problematic.
10 Interaction-Based Simulations for Systems Biology 201

A good illustration is given by a simple model of growth sometimes called the

Eden model (Eden 1961). It has been used since the 1960s as a model for tumor
growth. In this model, a space is partitioned in empty or occupied cells. At each
step, occupied cells with an empty neighbor are selected, and the corresponding
empty cell is made occupied. An exclusion principle prevents two occupied cells to
invade the same empty cell.
This specification of the local evolution of the system defines the interaction
between an occupied and an empty cell. It is difficult to turn this specification into a
simple rule for one cell evolution because an empty cell can query its neighborhood
to find if they are occupied cells, but they cannot know which or even if an occupied
cell will invade it. Conversely, when an occupied cell decides to invade an empty
one, it cannot determine whether another occupied cell makes the same decision at
the same time.

10.2.2.2 The Spatial Structure of Interactions

Usually, only physically close entities interact because information exchange ul-
timately has a local character (e.g., transport of signaling molecules between
neighboring cells). Thus the possible interactions of the entities in the system reflect
the underlying physical space. The other way round, we can say that the spatial
organization of the entities composing the system organizes also their interactions.
In Table 10.1, we have introduced an additional criterion, space, to categorize
dynamical systems formalisms following the discrete or continuous setting used
for the spatial organization of their entities. For instance, in a cellular automaton,
entities called “cells” are organized in a regular lattice. In PDE, fields are localized
in a continuous space and can extend over an entire subspace.
It is interesting to examine the case of aggregate models. For example, the aggre-
gate model of a chemical reaction supposes that the chemical solution is well stirred
and abstracts a population of molecules by a set of concentration variables. In this
case, there is no need to record the position and the velocity of each molecules in
the continuous underlying physical space: it is as if each molecule could interact
with any other. The possibility of chemical reactions is then only constrained by
the “compatibility” of the reactants and is better described by a discrete structure,
the molecular interaction network. Even if this network represents functional con-
straints rather than constraints from the physical underlying space, it is obtained by
“erasing” the localization of the molecules keeping only the possibility of interac-
tions between different species.
Taking the previous discussions seriously pushes to make a switch from state and
evolution function to individuals and interactions. In the next section, we will see
how topological notions used to describe neighborhood relationships can be used to
support the description of interactions in a programming language.
202 A. Spicher et al.

10.3 The MGS Domain-Specific Programming Language

The previous section points out several difficulties raised by the computer modeling
and the simulation of biological processes in the context of integrative and spa-
tial systems biology: the necessity to accommodate a wide range of mathematical
formalisms (from continuous to discrete and from deterministic to stochastic), the
dynamical structure of the system and the specification of the dynamics through
local, spatially organized interactions.
To face these difficulties and to ease the building of a simulation program, we
advocate a language-based approach through a domain-specific language (DSL)
dedicated to the simulation in systems biology. DSLs are programming languages
for solving problems in a particular domain. To this end, they provide abstractions
and notations for the domain at hand. They are more attractive for programming
in the dedicated domain than general-purpose languages because of easier pro-
gramming, systematic reuse, better productivity, reliability, maintainability, and
flexibility. Moreover, DSLs are usually small and more often declarative than imper-
ative. Declarative programming focuses on what should be computed instead of how
it must be done. Objects and constructions are close to the mathematical standards
which enable an easier mathematical reasoning on programs. Thus, a declarative
program is an executable specification not burdened by the implementation details
and is close to the mathematical model.
In the rest of this section, we present such a DSL: the MGS modeling lan-
guage.4 MGS allows a clear and concise specification of processes through spatial
interactions. In MGS, the state of a dynamical system is specified using an orig-
inal and generic data structure: the topological collection (Giavitto and Michel
2002a). Topological collections are based on the topological relations between the
interacting subparts of the system. Furthermore, the specification of the evolution
law, through local interactions, is simplified by the definition of transformations.
Transformations are functions defined by a set of rules. Topological collections and
transformation are handled in a declarative style.
The notions of topological collection and transformation subsume several mod-
els of computation inspired by biological systems or used in their modeling and
in their simulation like CA, Lindenmayer systems (used in growth plant modeling)
or P systems (advocated for the modeling of compartmentalized molecular interac-
tion networks). The MGS approach is illustrated in Sect. 10.5. The benefits of the
MGS approach have been demonstrated through a number of complex applications
in system biology (see Sect. 10.6).

10.3.1 Topological Collection

One of the key features of the MGS language is its ability to describe and manipu-
late a collection of entities structured by a neighborhood relationship. Such device,

4
The Website of the project is http://mgs.spatial-computing.org.
10 Interaction-Based Simulations for Systems Biology 203

called a topological collection, is used to represent the state and the organization
of a biological systems: the elements of the collection are the components of the
system, and the topology of the collection sets the potential interactions (i.e., two
elements in the collection may interact only if they are neighbor).
Intuitively a topological collection generalizes the notion of field widely used in
physics: each collection is build on an underlying space by associating some value
with each position in this space. Positions can be points, but also more generally
lines, surfaces, volumes, etc. The value associated with a surface may represent a
flux, the value associated with a volume may represent a concentration, etc.
Topological collections can also be thought as a generalization of the notion of
array where the index of an element is replaced by a position in the underlying
space (Giavitto and Michel 2002a). This view subsumes a large family of important
data-structures used in simulations. For instance, a labeled graph is a special case
of a topological collection where the positions are the nodes of the graph and the
neighborhood relationships are given by the edges of the graph.
Technically, the formalization of topological collections relies on the notion of
chain complex (Munkres 1984) defined in algebraic topology and has been thor-
oughly studied in previous work of the authors (Giavitto and Michel 2002c; Giavitto
and Spicher 2008b).
Several neighborhood relationships are expressible in MGS. In the rest of
this chapter, we will mainly use records, multiset, group-based fields (GBF), and
Delaunay collections. These collections represent several important families of
interactions and we will show that they are homogeneously handled in MGS. In
addition, MGS allows heterogeneous collections (the elements of a collection can
have different types) and the arbitrary nesting of collections (i.e., an element of a
collection can itself be a collection). These features greatly facilitate the develop-
ment of models and their simulations.

10.3.2 Transformation

Topological collections represent an adequate medium to specify the interactions

between the elements of a biological system. In MGS, the specification of a
transformation T :
trans T = f ...  => f .; : : : /; ... g
corresponds to the definition of a set of rules, where the left-hand side  is a pattern,
matching for a subcollection, and the right-hand side f .; : : : / is an expression that
evaluates a new subcollection that will be inserted in place of the matched one. The
notion of subcollection depends on the neighborhood relationships of the collection:
a subcollection is a connected subset of a collection and two elements are connected
if they are neighbors.
204 A. Spicher et al.

Intuitively, a rule represents a possible (local) evolution of a (sub)system. The

pattern in the left-hand side of the rule represents a potential configuration, and the
expression in the right-hand side computes the local evolution of this configuration.
A very simple transformation is given by:
trans simpleT = f 0 => 1; g
This transformation is composed of a single rule which replaces the value 0 in the
collection by the value 1. There are two important points to note.
First, this transformation may be applied to any kind of collections. Such a trans-
formation is called polytypic (Jansson and Jeuring 1997). Polytypic transformations
encapsulate an abstract process that can be reused in a variety of situations. For
example, MGS is expressive enough to allow the definition of a generic diffusion
process that can be used on any kind of collections (Giavitto and Spicher 2008b).
Second, if the transformation simpleT defines the replacement of 0 by 1, it
does not specify which 0 s must be replaced. If there are several occurrences of 0 in
the collection, do we have to replace all of them, some of them, or just one of them?
In the two latter cases, how are the occurrences chosen? These choices are under the
control of a rule application strategy. The application of the transformation T on a
topological collection e using a strategy St is written as:
T [strategy = St](e)
In the current implementation of MGS, all available strategies are built-in (but
the functional composition of the transformations allows a certain flexibility for
specific requirements). In the following, we will use two of them: the Gillespie
strategy based on the stochastic simulation algorithm proposed by Gillespie to simu-
late chemical reactions (Gillespie 1977) (see Sect. 10.5.3) and the maximal-parallel
strategy widely used in the context of L-systems (Lindenmayer 1968a) and P sys-
tems (Păun 2001). In the maximal-parallel strategy, which is the default strategy, a
maximal set of non-intersecting occurrences of the pattern are simultaneously re-
placed by the right-hand side of the rule. When several such sets exist, one of them
is non-deterministically chosen.

10.3.3 Two Models of Diffusion

We illustrate the notion of transformation with the simulation of a paradigmatic

diffusion process. Diffusion is defined in a continuous setting in one dimension by
the following equation:
@u @2 u
DD 2
@t @x
where D is the diffusion coefficient, u is the concentration of the diffusing sub-
stance, and x is the position.
Below, we describe two different approaches to simulate the diffusion of a chem-
ical on a one-dimensional rod.
10 Interaction-Based Simulations for Systems Biology 205

10.3.3.1 The Numerical Resolution of the Continuous Model

This example shows MGS’ ability to handle a continuous model. By their nature,
computer simulations operate in discrete time. Models initially formulated in terms
of continuous time and space must therefore be discretized. Using a simple finite
difference method, the previous equation is discretized as:
X
u.i; t C dt/ D u.i; t/ C h .u.j; t/  u.i; t//
j

where
u.i; t/ represents the concentration at time t of the i th element of the discretized
rod, and j ranges over the neighbor of i . Parameter h depends on the discretization
and on the diffusion constant D.
This computation can be programmed in MGS by the following transformation:
trans diffuse[h] = f
u => u + h * (neighborsfold(+,0,u)
- u * neighborsize(u))
g
where
u is a pattern variable that matches any element in a collection, the expres-
sion neighborsfold(op,e,u) uses operation op to combine the values of the
neighbors of u starting from the initial value e, and neighborsize(u) returns
the number of neighbors of u. An additional parameter is provided between brackets
after the name of the transformation and corresponds to the parameter h.
It is straightforward to extend this process to a surface or a volume instead of a
1D rod. More elaborate discretization schemes are also handled similarly: for ex-
ample, we give in the annex of this chapter the MGS program corresponding to the
implementation of the Range–Kutta methods.

10.3.3.2 The Discrete Stochastic Evolution of a Diffusing Particle

Now, we want to take the same system but we focus on the level of the molecules.
The rod is discretized as a sequence of small boxes, indexed by a natural integer,
each containing zero or many molecules. At each time step, a molecule can choose
to stay in the same box, or to jump to a neighboring box, with the same probability p
(whose value depends on the time and space discretization). The state of a molecule
is the index of the box where it resides. The entire state of the system is then repre-
sented as a multiset of indices.
A multiset is a generalization of a set (Banâtre et al. 2006): the same element
can appear multiple times in a multiset. In a multiset, each element is neighbor
of any others. Thus, a multiset is a good idealization of a well stirred “chemical
206 A. Spicher et al.

Fig. 10.1 Evolution of a chemical diffusing in a 1D rod. Evolution modeled as a continuous pro-
cess (a) or as a discrete stochastic one (b). Intuitively, the left figure is the limit of the right figure
when the number of boxes in the rod and the number of particles grow to infinity

soup” (Giavitto et al. 2004). In our example, if there are m molecules in the box
numbered n, then there is m occurrences of the integer n in the multiset.
The evolution of the system can then be specified as a transformation with three
rules:
trans diffuse[p] = f
q =f P = (1 - 2*p) g=> q
q =f P = p g=> q + 1
q =f P = p g=> q - 1
g
The arrow construction =f...g=> is used to specify the specific parameters of a
rule. Here we give a value to the parameter P used in the probabilistic application
of the rule. In this strategy, a matched pattern is replaced by the right-hand side of
the rule only with a probability P. Additional rules (not shown here) are provided to
deal with boundary conditions.
Figure 10.1 illustrates the iteration of the continuous and stochastic transforma-
tions. In the initial state, all particles are randomly distributed in the middle third of
the rod.

10.4 A Synthetic Multicellular Bacterium

In the forthcoming sections, we propose to illustrate the expressiveness brought by

the MGS language for the modeling, at various spatial and time scales, of the same
biological process: a synthetic multicellular bacterium (SMB) built during the 2007
iGEM competition (footnote 1) by the French team in Paris.
We start by a short presentation of synthetic biology, the iGEM competition and
then we describe the SMB project of the Paris team.
10 Interaction-Based Simulations for Systems Biology 207

10.4.1 Synthetic Biology

Synthetic biology is an emergent field which proposes an engineering point of view

on biology. It aims at building new biological systems by assembling standard low-
level components called BioBricks (Knight 2006). These components, designed in
the projects presented for the iGEM contest, are described and stored in an ontology
hosted by the MIT.5 They are pieces of DNA used to build biological functions (as,
for example, a logical gate) and integrable within existing genomes. For example, a
brick activating the production of a chemical species in the presence of a sufficient
concentration of molecules of types A and B can be interpreted as a function calcu-
lating the conjunction of the chemical signals associated with the species A and B.
The basic principles of construction of the biological components, establish-
ing the biosynthetic methodology, were elaborated at the MIT at the turn of the
twenty-first century. They rely on classical engineering strategies: standardiza-
tion, decoupling, and abstraction (Endy 2005). The purpose of standardization is
twofold: to ensure compatibility between the bricks and to allow the development of
generic and normalized building protocols (i.e., functioning for all bricks) econom-
ically accessible and easily implementable. Decoupling is a strategy that separates
complicated problems into simpler ones. For instance, the separation of the various
functions of a synthetic system allows the modularization of the system, the reuse of
its parts, the independent evolution of each of them, etc. The separation of the phases
of design and implementation reduce and eliminate the dependence between the de-
sign of a gene regulatory network and the effective building of a strand of DNA.
Finally, an abstraction hierarchy supports the engineering of integrated genetic sys-
tems by hiding information and managing complexity through relevant levels of
expression: from DNA nucleotides to parts, devices, and (complete biological) sys-
tems. Abstraction levels limit the exchange of information across levels and allow
individuals to work at any level without regard for the details that define other levels.

10.4.2 The International Genetically Engineered Machine

Competition

The iGEM is a competition launched by the MIT in 2003. More than 110 teams
coming from all ever the world participated in the 2009 issue. The competition
is aimed at undergraduate students who are given the opportunity to manipulate
complex molecular biology processes made simple by the synthetic biology princi-
ples. During a 3 months time period, students mentored by post-graduate students
and researchers design, model, and assemble BioBricks to produce new biological

5
The BioBricks are available in the Registry of Standard Biological Parts at the following Web
site http://partsregistry.org/Main Page.
208 A. Spicher et al.

functions integrated into living systems. At the end of the competition, all teams
gather at the MIT in the first weekend of November during the Jamboree where
their projects are being evaluated.
In 2007, a French team supervised by A. Lindner and S. Bottani participated in
the competition and was ranked first in the “foundational research” category for their
Synthetic Multicellular Bacterium project. MGS was used to produce most of the
simulations needed to validate the design (one simulation was done in MATLAB).
In Sect. 10.5, we present several simulations that are inspired or extend the initial
SMB simulations.

10.4.3 Objectives of the SMB Project

The objective of the SMB project is the design of a synthetic multicellular bac-
terium. This organism was thought as a tool that would allow the expression of
a lethal or dangerous transgenic gene in the Escherichia coli bacterium without
disturbing the development of its biomass. The main difficulty was to install a
mechanism of irreversible bacterial differentiation which makes possible to express
the transgene only in a part of the population unable to reproduce. The two lines,
germinal (not differentiated) and somatic (differentiated and unable to reproduce),
are interdependent and then constitute a multicellular organization (hence the name
“multicellular bacterium”). In order to ensure that the ratio between the two popula-
tions makes it possible for the system to grow, the sterile somatic cells are designed
to provide to the germinal cells a molecule essential to their reproduction: DAP (di-
aminopimelate). Figure 10.2 sketches the general principle of the project. Additional
information is available through the iGEM Paris Team Web site (footnote 1).
The design of this organization asked for the development of two distinct biolog-
ical functionalities, one for the cellular differentiation and the other for the feeding
of DAP to the germinal cells. The study of this system was at the same time the-
oretical and practical. Although the biological implementation of the system could
not be entirely carried out by lack of time, the students at iGEM Paris provided
experimental evidences and theoretical proofs that the SMB organism was viable.

differentiates into

reproduces cannot reproduce

Germline Soma
the organism the organism

is required for

Fig. 10.2 The SMB is composed of two cell types: germ cells (G) and somatic (S) cells. G cells are
able to live by producing two different types of cells: G cells and S cells. S cells are derived from
G cells by an irreversible differentiation step, exhibiting a new function required for the survival of
the G cells. S cells cannot reproduce. This dependency between G and S cells defines the organism
10 Interaction-Based Simulations for Systems Biology 209

constitutive ftsK endogeneous

dapAp promoter promoter
Germline
Cre LOX T ftsK T LOX dapA T
DNA
ftsK endogeneous
constitutive promoter
dapAp promoter
Soma T ftsK T

DNA
Cre dapA T +
LOX
Plasmid

Fig. 10.3 Gene regulatory networks of the germinal and somatic cells. Gene regulatory networks
describing the feeding device (light gray) and differentiation device (dotted box). Cre, dapA, and
ftsK are genes, LOX is a recombination site, and T are terminators

10.4.4 The Paris Team Proposal

To implement this functionality into the E. coli bacterium, the Paris team has
proposed an original construction. The gene regulatory networks of the proposal
is described in Fig. 10.3.
Two functions are described: a feeding device based on the production of DAP
molecules (light gray) and a differentiation device based on a classical Cre/LOX
recombination scheme (dotted box).
In the germline G, there is a natural expression of ftsK. This gene is essential for
replication. The protein product of gene dapA is DAP. This protein diffuses in the
environment and is rapidly degraded. However, in the germline, the dapA gene is
not active since it lacks a promoter to initiate its transcription and G is auxotrophic
in DAP.
The promoter dapAp is sensitive to DAP concentration. Located before the gene
Cre, it allows to adjust the production of Cre to the presence of DAP in the environ-
ment. The production of Cre initiates the recombination/differentiation process.
After recombination, the genomic reassembly leads, by the excision of the parts
between the two LOX recombination sites, to the cell of type S and a plasmid
that is rapidly degraded. In the feeding device S, dapA is under the control of its
constitutive promoter and can be expressed. The synthesized DAP diffuses in the
environment allowing to reach G cells. Lacking the ftsK gene, S cells are sterile and
eventually die.

10.5 Modeling in MGS

In this section, we illustrate the expressive power of MGS through four examples
derived from the SMB. These four examples have been chosen to illustrate the
MGS concepts on individual-based models as well as aggregated models, and on
spatialized as well as non-spatialized models (see Table 10.2).
210 A. Spicher et al.

Table 10.2 Aggregated/individual-based and spatial/non-spatial models.

Aggregated models vs. individual-based models and spatialized vs. non-
spatialized models in the SMB simulation examples
Aggregated Individual-based
Non-spatialized ODE Stochastic simulation
model (Sect. 10.5.1) à la Gillespie (Sect. 10.5.3)
Spatialized Discrete diffusion of Cell–cell dynamical
model DAP (Sect. 10.5.2) Interaction (Sect. 10.5.4)

10.5.1 Solving Differential Equations

This first modeling of SMB is a kind of proof of concept based on the study of a
differential equations system. We propose here a rule-based expression of this model
with two simple resolution schemes: the Euler and Runge–Kutta methods.

10.5.1.1 The SMB Proof of Concept

The very design of SMB is based on the composition of a feeding device together
with a differentiation device. We wonder here whether this architecture could reach
homeostasis, no matter how these devices are implemented. So a minimal model is
required to give such a proof of concept of the design.
To answer this fundamental question, the Paris team proposed a theoretical study
of the population dynamics based on a classical differential equations model. Let
ŒG, ŒS , and ŒD denote the concentration of germinal cells, somatic cells, and
DAP molecules in a well-mixed solution. Their dynamics are captured by the three
following equations:

dŒG ŒDn
D ˛1 ŒG  ˛2 ŒG  ˛3 ŒG (10.1)
dt ŒDn C k n
dŒS 
D ˛2 ŒG  ˛4 ŒS  (10.2)
dt
dŒD
D ˛5 ŒS   ˛6 ŒD (10.3)
dt

They give the time variation of each concentration as functions of ŒG, ŒS , and
ŒD. Parameter ˛1 denotes the growth rate of germ cells, parameter ˛2 denotes the
differentiation rate, parameter ˛3 denotes the death rate of germ cells, parameter ˛4
denotes the death rate of somatic cells, parameter ˛5 denotes the production rate
of DAP by the somatic population, and parameter ˛6 denotes the degradation rate
10 Interaction-Based Simulations for Systems Biology 211

of DAP. In this model, the differentiation device is parameterized by ˛2 , and the

feeding device is captured by parameters ˛5 for the DAP production and ˛1 that is
weighted by a Michaelis–Menten function representing the dependence of germinal
cells growth to the DAP concentration.

10.5.1.2 Analysis of the ODE Model

In general, such models based on differential equations are not easily investigated.
The parameters are often numerous and qualitative analyses are difficult. In our
case, parameters ˛5 and ˛6 can be dropped assuming that the DAP concentration is
stabilized (i.e., when ŒD remains constant and (10.3) vanishes). This simplification
of the model allows to stress out two main population behaviors. Indeed, it reveals
a nontrivial fixed point .ŒG0 ; ŒS 0 / that is unstable:
 For greater values of cell concentrations, an exponential growth is observed.
 For lower values of cell concentrations, both populations collapse to reach the
second and trivial fixed point .0; 0/.
But is this result relevant? In other words, is the DAP stabilization assumption
realistic? Should the production of DAP fluctuate, the previous sketch does not give
any information on the viability of the SMB. In the following sections, we propose
to focus on this question relying on different characterizations of the dynamics using
numerical simulations.

10.5.1.3 A Numerical Solution of Differential Equations

By their nature, simulations operate in discrete time. Models initially formulated in

terms of continuous time must therefore be discretized. Strategies for discretizing
time in a manner leading to efficient simulations have extensively been stud-
ied. Here we use as an example a straighforward and very simple approach, the
Euler method. This method particularly fits well the simulation of problems of the
form:
dX.t/
D f .X.t// X.0/ D X0
dt
where
X.t/ is a vector of values representing the state of the system at a given time
t, and X0 is the initial state. The function f computes the variation of each co-
ordinate of X at a given time t. As far as our problem is concerned, one has the
state X D .ŒG; ŒS ; ŒD/ and function f corresponds to the three equations (10.1),
(10.2), and (10.3).
212 A. Spicher et al.

The Euler method computes a sequence of vectors Xn , where X0 D X0 at

the initial time and the generic term is given by the first two terms of the Taylor
expansion:
XnC1 D Xn C tf .Xn /
where
t denotes the simulation time step.
We start the MGS expression of this computation by representing the state of
the system in terms of topological collection. We use here a record. A record is one
of the simplest collection consisting of two or more values so that each component
(called a field or member of the record) can be accessed through a symbolic name.
Each value in the record is “isolated” and has no neighbor. Hence, there is no direct
interaction between the elements of a record. Elements of a record are given between
braces.
The record used here has three members describing the concentrations ŒG; ŒS 
and ŒD with a value of type float (a real number):
record State = f G:float, S:float, D:float g
The variation of each concentration can be computed from such a state. The follow-
ing function implements this procedure according to (10.1), (10.2), and (10.3):
fun Variation[a1,a2,a3,a4,a5,a6,k,n](X) = f
X.Dn
G = ( *a1 - a2 - a3)*X.G,
X.Dn C kn
S = a2*X.G - a4*X.S,
D = a5*X.S - a6*X.D
g
Parameters ai , k, and n are given in brackets. Parameters in brackets are optional
arguments. Note that the function Variation returns a State. It allows collec-
tions X and Variation(X) to be of the same type, and then to share the same
set of positions (here fields G, S, and D). This property eases the computation.
For example, while the concentration ŒG is obtained at position G of collection
dŒG
X (by the expression X.G), its variation is at the same position G of collection
dt
Variation(X) (corresponding to the expression Variation(X).G).
Finally, one step of the Euler method can be expressed by a transformation to be
applied on a collection of type State:
trans Euler[dt=t,f] = f
x => let dx = f(self).(ˆx) in x + dt * dx
g
In this transformation, the unique rule specifies how each element x of the collection
has to be updated by computing its variation dx. This variation is taken at ˆx (i.e.
the position of x) of the collection and is computed by the function f (a parameter
of the transformation) applied on self. The identifier self always refers to the
collection which the transformation is applied to. In our example, the actual value
10 Interaction-Based Simulations for Systems Biology 213

of f will be the previous function Variation. The whole trajectory is obtained

by iterating the application of transformation Euler on an initial condition.
The reader is invited to note that transformation Euler is fully independent
from the specification of State and Variation, and can be used as a generic
implementation of the Euler method in many different contexts. Moreover, whereas
the Euler method is sufficient for the simulations described below, we would like to
underline that other integration methods can also be straightforwardly implemented
in MGS. The implementation of the Runge–Kutta method is elaborated in the Ap-
pendix of this chapter.

10.5.1.4 Interpretation of the Simulations’ Results

Numerical approaches suffer from the lack of knowledge regarding the values of
parameters. Hence, we cannot rely on any quantitative information on the system.
Nevertheless, experience and classical examples give us sufficient information to de-
termine a range of possible parameters. For the sake of the simplicity, we arbitrarily
set them to the intervals given in Table 10.3.
Our objective was to observe all the possible behaviors of the system for different
settings of parameters (chosen in the parameters space defined by Table 10.3) and
starting from a common initial state.
The protocol of our study has consisted in running 10 000 simulations of the
model. Each simulation has consisted in computing the Euler trajectory of the sys-
tem over 11 000 iterations with a time step equal to 0:01 (i.e., 110 arbitrary units of
simulation time) starting from an initial state where only germinal cells are present
with a very high concentration of DAP. In each run, parameters were randomly cho-
sen according to the intervals given in Table 10.3.
Euler[dt=0.001,f=Variation]
(f G = 100, S = 0, DAP = 10000 g)
Results are given in Fig. 10.4. Only three clearly distinguished behaviors are ob-
served and coincide with the dynamics provided by the qualitative analysis:
1. Population collapse (see Fig. 10.4a).
2. Exponential growth of the population (see Fig. 10.4b).
3. The unstable fixed point (see Fig. 10.4c).
In all behaviors, the system starts by consuming DAP molecules to replicate. Once
DAP concentration falls below a certain level, differentiated cells start to appear and
initiate the production of DAP.

Table 10.3 Intervals of the Parameter Range Parameter Range

parameters for the
˛1 [0,2] ˛5 [0,1000]
ODE-based model
˛2 [0,1] ˛6 [0,1]
˛3 [0,1] n 2
˛4 ˛4 D ˛3 h 100
214 A. Spicher et al.

a b
10000 1e+06 [G]
[G] [S]
[S] [D]
[D]

Concentrations
[D]/([G]+[S])
Concentrations

[D]/([G]+[S])

1e−06 0.1
0 Time 100 0 Time 100

c
100000
Concentrations

[G]
[S]
[D]
[D]/([G]+[S])
0.1
0 Time 100

Fig. 10.4 Results of simulations of the ODE-based model. Figure 10.4a–c illustrates the three
observed behaviors (resp. population collapse, exponential growth, and unclassified behavior).
Figure 10.4d–h gives on stacked histograms the distribution of the 10 000 simulation runs for
each parameter (resp. ˛1 , ˛2 , ˛3 , ˛5 , ˛6 ) with populations collapse in medium gray, exponential
growth in dark grey, and unclassified behavior in light gray
10 Interaction-Based Simulations for Systems Biology 215

In order to understand what characteristics prevent an exponential growth, the

simulations have been classified according to the three behaviors, and their distri-
butions have been analyzed (see Figs. 10.4d–h). Each histogram shows the behavior
of the system following one parameter and irrespectively of the other ones. Popula-
tion collapse occurs when germinal replication rate is low (either because of a small
growth rate ˛1 or because of a high death rate ˛3 , see Figs. 10.4d,f) or when the
differentiation rate is too low (see Fig. 10.4e). The last two figures (Figs. 10.4g–h)
show that the system is not perturbed by the behavior of DAP production or degra-
dation. This explains why no additional behaviors are observed compared to the
qualitative analysis. In fact, the dotted lightgray curves of Figs. 10.4a–c (that cor-
ŒD
respond to the ratio ŒGCŒS ) show that the normalized DAP concentration remains
 
ŒD
constant after a transient phase; the assumption dtd ŒGCŒS D 0 seems appropriate
for a qualitative analysis.
The conclusion of this study is twofold:
1. The choice of DAP as the main molecule to design the feeding device is good if
the auxotroph germ line is robust and grows well in the presence of DAP.
2. The differentiation device is required to be efficient (a reversible differentiation
should be prohibited).
We aimed here at illustrating how MGS can be used as a prototyping tool for provid-
ing quick results and orienting further investigations. This preliminary work could
be improved by taking into account more realistic parameter ranges provided by the
litterature. The use of data analysis techniques could also give a better knowledge
(e.g., parameters correlations) on the results of the simulations.

10.5.2 Cellular Automata

In this second modeling, we focus on the effects on the SMB due to the spatial orga-
nization of the SMB. We propose to design a cellular automaton and to implement
it in MGS.

10.5.2.1 The Spatial Organization of the SMB

The ODE-based model proposed in Sect. 10.5.1 considers the SMB as a molecu-
lar solution of three different species uniformly distributed in space. In reality, the
system consists of two populations of cells that will be organized in space. Such
an organization may induce heterogeneity in the cells distribution leading to some
spatial artefacts not taken into account by the ODE model (Durrett and Levin 1994).
Some interesting spatial self-organizations could even be observed: for instance,
one can easily imagine that the SMB collapses at some locations while it grows
216 A. Spicher et al.

exponentially at others making some patterns appear at the population scale (Shnerb
et al. 2000). As a consequence, one has to investigate whether space matters or not
in the SMB development.

10.5.2.2 A Discrete Spatial Framework

Different formalisms allow to take space into account. A first direction consists in
extending the ODE of Sect. 10.5.1 by considering the spatial distribution of con-
centrations (i.e., ŒG, ŒS , and ŒD would depend on time but also on space). This
extension would introduce in the formula the use of two additional terms to deal
with the spatial diffusion of concentration of bacteria and DAP molecules. These
modifications make the original ODE system become a PDE system and increase
the paramater space. The associated phase space becomes more difficult to study.
CA and multi-agents systems (MAS) are another class of formalisms that explic-
itly consider spatial organization. Both rely on a individual-based point of view. We
focus here on a CA approach.
A cellular automaton is a regular lattice of places called “cells”, where each cell
is characterized by a state taken from a finite set. The global evolution of the CA
consists in applying synchronously, on each cell, a local evolution function that
computes the new state of the cell as a function of its current state and of the states
of the cells in its neighborhood.
The Paris team proposed a CA to study the relation between DAP diffusion and
differentiation. Their model is based on states encoded as real numbers to represent
the concentration of DAP in each CA cell, and the use of non-deterministic rules
using random number generators. On the contrary, we propose a totally determin-
istic CA with discrete states and very simple rules. More precisely, we consider a
superposition of two CA: the one deals with the DAP diffusions process while the
other takes into account the differentiation of the cells.
DAP Diffusion CA. Contrarily to the simulations of diffusion given in Sect. 10.3.3,
we aim at specifying a phenomenological diffusion in a CA, that is the propagation
of an information (e.g., “contains some DAP”) from a source CA cell to its neigh-
borhood. This behavior corresponds to a classical propagation rule (Wolfram 1986)
where a cell becomes activated if one of its neighbors is activated. In order to limit
the radius of the propagation, the following rule may be used:

xt C1 D max.0; yt1  1; : : : ; ytn  1/ (10.4)

where
x denotes a cell of the CA, xt its state at time t, and y 1 , . . . , y n the neighbors of x.
Here states are encoded by integers that are gradually decremented from the source:
0 means no activation and n > 0 means that the activation propagates with a radius
n around the cell. Some evaporation may be introduced to deal with the removal of
10 Interaction-Based Simulations for Systems Biology 217

Fig. 10.5 Results of simulations of the CA model. Top line shows the propagation of DAP around
an isolated source with radius 5: from left to right, initial state, state after 1 iteration, steady state. A
light gray cell means no DAP, the gray scale represents the concentration of DAP and the black cell
is the source. Bottom line shows the evolution of the CA defined by (10.6), on a 40  40 hexagonal
grid only filled by germinal cells with a randomly chosen ratio: from left to right, initial state, state
after 13 iterations, state at fixed point. Germinal cells figure in dark gray and somatic cells in light
gray

the source. A source maintains its state to a specific integer value denoted by NR .
Figure 10.5a–c shows the discrete diffusion around an isolated source for NR D 5
on an hexagonal grid.
Differentiation CA. This CA focuses on the bacterial layer. Each cell of the CA
represents a part of the whole population. Under some conditions on the DAP level,
a CA cell progressively goes from a majority of germinal bacteria to a majority of
somatic bacteria. We abstract the ratio between the two populations in a CA cell by
an integer between 0 and NP : 0 means that only somatic bacteria are present and
NP only germinal bacteria are present.
The dynamics of a CA cell is as follows: if there is enough DAP, the number
of germinal bacteria becomes maximal. Otherwise, this number decreases at each
time step. Finally, if no germinal bacterium remains, the ratio is locked to 0. Equa-
tion (10.5) summarizes this behavior:
8
< NP if there is enough DAP
ut C1 D 0 if ut D 0 (10.5)
:
ut  1 otherwise

where
u denotes a CA cell and ut its state at time t.
218 A. Spicher et al.

Coupling both CAs. Equations (10.4) and (10.5) are combined to define the final
CA. Let ct D .ut ; xt / denotes the state of a CA cell c a time t. The local evolution
function is given by:

.0; NR / if xt D 0
ct C1 D (10.6)
.ut C1 ; max.0; xt C1  NC // otherwise

where
xt C1 and ut C1 are given by (10.4) and (10.5), and NC represents the DAP con-
sumption of germinal bacteria. Finally, we consider that there is not enough DAP
when xt C1  NC is negative and no DAP source is available in its neighborhood.

10.5.2.3 MGS Expression of a Cellular Automaton

MGS allows an easy specification of CA. We use the group-based field topolog-
ical collections (GBF) to represent regular and uniform lattice, as used in CA or
for the numerical solutions of PDEs. The neighborhood relationships of a GBF are
described in terms of a mathematical group, the group of elementary displacements
in the lattice (Giavitto and Michel 2001; Giavitto et al. 2002b). The corresponding
space can be visualized as a graph, the Cayley graphs of the presentation of the
group (Giavitto and Michel 2002b). This abstract approach enables the easy speci-
fication and a uniform handling of a large family of circular and screwed, bounded
or unbounded grids in any dimension.
For the purpose of this example, each GBF position is labelled by an MGS record
of type fx:int, u:intg representing the state c t . Each GBF position has six
neighbors achieving an hexagonal grid.
The dynamics of (10.6) are implemented as follows:
trans SMB CA = f
c / c.u == 0 => f x = NR , u = 0 g
c / NoPrs(c) => f x = Diff(c)-NC , u = NP g
c / c.u == 1 => f x = NR , u = 0 g
c => f x = 0, u = c.u-1 g
g

where
NoPrs(c) computes whether there is not enough DAP on cell c and Diff(c)
computes the diffusion on cell c with respect to (10.4). Note that the order of the
rules matters: for instance, the matching of a cell by the third rule implies that it
cannot be matched by the first two ones. This transformation is applied using the
standard maximal parallel strategy of MGS.
10 Interaction-Based Simulations for Systems Biology 219

10.5.2.4 Interpretation of the Simulations’ Results

Figure 10.5d–f shows how differentiation appears in a population of germinal cells.

As the CA transition function is deterministic, symmetry is broken in the initial state
(otherwise all cells would exhibit the same behavior): we have chosen to start the
simulation with cells of the form f x = 0, u = 1+random(NR )g, that is to
say germinal cells with a ratio uniformly chosen in Œ1; NR , and no DAP.
No symmetrical pattern appears during the simulation, whatever the parameters
NR , NP , and NC are. The distribution of the differentiation cells follows the ratio
distribution chosen at the initial state. An equivalent behavior is observed when
the symmetry is broken by randomly initializing the field x. The uniformity of the
dynamics supports the assumption of a well-mixed solution used in Sect. 10.5.1 and
confirms the previous result.
We have shown with this model how a rule-based programming style fits well
the specification of CA. Not more than ten lines are required to describe it in MGS.
Moreover, thanks to the polytypic feature of MGS, the specification of the topology
is totally decoupled from the definition of the dynamics; transformation SMB AC
could be applied on any kind of topological collection, and more especially on any
kind of grids and neighborhoods (like square grids with von Neuman or Moore
neighborhoods, toric grids, etc.). More specialized CA tools are often ad hoc and do
not exhibit so much flexibility and genericity in the expression of models.

10.5.3 Stochastic Simulations

The two previous approaches provide results at the level of the synthetic device.
In this section, we study the construction of these devices in terms of biological
parts and synthetic construction as described in Sect. 10.4.1. More specifically, we
propose a stochastic model of the SMB at the level of one bacterium together with
its simulations using the exact stochastic simulation algorithm defined by Gillespie
(1977).

10.5.3.1 Robustness Analysis of the SMB Design

The characterization of a synthetic device depends on its implementation. We aim

at checking if the behavior of the Paris team construction respects the main objec-
tive of the SMB. More precisely, we focus on the noise sensitivity and the relation
between parts parameters (like the rate of DNA Cre/LOX recombining) and the de-
vices parameters (such as the differentiation rate).
A common way of modeling gene regulation is to consider the regulatory net-
work as a set of biochemical reactions. The set of chemical interactions induced by
the Paris team construction (see Fig. 10.3) are abstracted by the following reactions:
220 A. Spicher et al.

C0
CreGGGGGGGA : (10.7)

C1
DAPGGGGGGGA : (10.8)

C2
GGGGG D Cre
DCre C DAP FGGGGGB (10.9)
C2

C3
DCre GGGGGGGA DCre C Cre (10.10)

C4
D Cre GGGGGGGA D Cre C Cre (10.11)

2C5
DG C CreGGGGGGGA D G (10.12)

C5
D G C CreGGGGGGGA DS (10.13)

C6
DS GGGGGGGA DS C DAP (10.14)

Cex
DAP FGGGGGGB
GGGGGG DAPex (10.15)
Cim

These reactions involve two kinds of chemical species: the DAP and Cre molecules,
and the DNA constructions of Fig. 10.3 abstracted by:
 DCre , D
Cre : Differentiation-free part of the construction composed of promoter
dapAp and the coding region for Cre. The two symbols represent, respectively,
the activated (no DAP repression on dapAp) and inhibited (DAP binds dapAp)
state of the promoter.
 DG , D G , DS : Part of the DNA modified by the Cre/LOX recombination mecha-
nism, DG before and DS after recombination. D G corresponds to an intermediate
state where only one LOX site is bound by Cre.
Reactions (10.7) and (10.8) describe the natural degradation of molecules Cre and
DAP. Reactions (10.9)–(10.11) express the behavior of the promoter dapAp: inhibi-
tion/activation by DAP and production of Cre (Reactions (10.10) and (10.11) differ
in their reaction constants: C4 C3 ). Reactions (10.12) and (10.13) specify the
two steps of a Cre/LOX recombination: DG ! D G ! DS . The regulation induced
by DG (expression of the gene ftsK) is not considered in this model. On the contrary,
the behavior of DS is specified by (10.14), corresponding to a constitutive produc-
tion of DAP. The last reaction (10.15) expresses importation and exportation of DAP
from the extracellular environment, where DAPex denotes the external occurrences
of DAP molecules.
10 Interaction-Based Simulations for Systems Biology 221

10.5.3.2 Stochastic Modeling for Sensitivity to Noise Analysis

The Paris iGEM team has chosen to investigate this kind of molecular model us-
ing a differential equation approach based on the mass action law. Thanks to this
study, they provided a set of optimized parameters for an exponential growth of the
SMB. Nevertheless, such results may be biased since the differential approach (1)
relies on a global homogeneous assumption and (2) does not take noise into account.
Since the number of molecules involved in gene regulation is in general very low, a
stochastic approach may provide complementary result on noise sensitivity.
A usual abstraction in the simulation of biochemical systems consists in consid-
ering the system (here the bacterium) as an homogeneous chemical solution where
the reactions of the model are taking place. Gillespie (1977) has proposed an al-
gorithm for producing the trajectories of such a chemical system by computing the
next reaction and the elapsed time since last reaction occurred. Let
be a chemi-
cal reaction, the probability that
takes place during an infinitesimal time step is
proportional to:
 c , the stochastic reaction constant6 of reaction
.
 h , the number of distinct molecular combinations that can activate reaction
.
 d, the length of the time interval.

Gillespie proved that the probability P .;

/d that the next reaction will be of type

and will occur in the time interval .t C ; t C  C d/ is:

P .;
/d D a ea0  d
where P
a D h c is called the propensity of reaction
, and a0 D  a is the com-
bined propensity of all reactions.
This probability leads to the first straightforward Gillespie’s algorithm called
the first reaction method. It consists in choosing an elapsed time  for each
reaction
according to the probability P .;
/. The reaction with the lowest
elapsed time is selected and applied on the system, making its state evolve. A
new probability distribution is then computed for this new state, and the process is
iterated.

10.5.3.3 Gillespie-Based Simulations in MGS

Here, we consider the bacterium as a well-mixed chemical solution. It can be repre-

sented by a multiset, that is a topological collection, where any element may interact
with all the others (Fisher et al. 2000). The simulation of “real” chemical reactions

6
Evaluating the stochastic constants is one of the key issues in stochastic simulations of biochem-
ical reactions. The interested reader should refer to De Cock et al. (2003) and Zhang et al. (2003)
for the description of two experiences in that field.
222 A. Spicher et al.

requires a strategy for multiset rules application in accordance with the reactions
kinetics. The MGS language provides such a strategy based on Gillespie’s algo-
rithm. We propose to use this strategy for simulating the previous set of chemical
reactions.
As said above, the state of the bacterium is represented by a multiset of values.
The considered molecules are abstractly represented using MGS symbols denoted
by back-quoted identifiers. For example, the MGS symbol ‘Cre corresponds to one
molecule of Cre. Thus, each chemical reaction is translated into a transformation
rule (or two if the reaction is reversible) characterized by an arrow parameter C
representing the stochastic constant of the reaction. For example, the reversible re-
action (10.9) can be straightforwardly translated to the two following MGS rules:
‘dCrA, ‘DAP =f C = C2 g=> ‘dCrI and
‘dCrI =f C = C2 g=> ‘dCrA, ‘DAP
Consequently, the whole dynamics is captured by the following set of rules in trans-
formation SMB STO:
trans SMB STO[DAPEx] = f
‘Cre =f C = C0 g=> .
‘DAP =f C = C1 g=> .
‘dCrA, ‘DAP =f C = C2 g=> ‘dCrI
‘dCrI =f C = C2 g=> ‘dCrA, ‘DAP
‘dCrA =f C = C3 g=> ‘dCrA, ‘Cre
‘dCrI =f C = C4 g=> ‘dCrI, ‘Cre
‘dG1, ‘Cre =f C = 2*C5 g=> ‘dG2
‘dG2, ‘Cre =f C = C5 g=> ‘dS
‘dS =f C = C6 g=> ‘dS, ‘DAP
‘DAP =f C = Cex g=> (DAPEx++; .)
. =f A = DAPEx*Cim g=> (DAPEx--; ‘DAP)
g

In the last two rules, the external DAPex molecules are specified by the counter
DAPEx given as an optional parameter. This counter is incremented (resp. decre-
mented) when a DAP molecule is imported (resp. exported).
The last rule of transformation SMB STO explicitly computes the propensity A
instead of using the usual parameter C. This feature allows a fine control of the
Gillespie application strategy when required.

10.5.3.4 Interpretation of the Simulations’ Results

A simulation is run by calling the transformation SMB STO using Gillespie’s

strategy:
SMB STO[strategy=‘Gillespie, DAPEx=1000]
(‘dCrA::‘dG1::():bag)
10 Interaction-Based Simulations for Systems Biology 223

a b
1000
External DAP
DAP 1000
Number of molecules

Cre

Number of molecules
External DAP
DAP

1 100
0 100 0 100
Time Time
c d
2500 90000

Mean number of iterations

Mean differentiation time

to differentiation

0 0
0 30 0 30
Number of external DAP molecules Number of external DAP molecules

Fig. 10.6 Results of simulations of the stochastic-based model. On top, two examples of stochastic
simulations when external DAP (in solid black) remains constant (Fig. 10.6b) or not (Fig. 10.6a).
Internal DAP is drawn in light gray and Cre in dotted line. Figure 10.6c shows the mean simula-
tion time to differentiation with a constant external DAP concentration for different values of that
concentration. Vertical bars correspond to the standard deviation. Figure 10.6d shows the mean
computation time to differentiation

The initial state is specified by two molecules, namely DCre and DG , added (with the
insertion operator ::) to an empty multiset (denoted by ():bag in the MGS syn-
tax). This state represents a germinal cell. External DAP is specified by initializing
the counter DAPEx to 1000 molecules.
Top line of Fig. 10.6 gives two different runs of the simulation corresponding
to the evolution of Cre, DAP, and DAPex populations over 100 arbitrary units of
time (AUT).
Figure 10.6a shows the classical behavior of a germinal cell: the DAPex is im-
ported from the outside until no DAP remains in the system (this process takes 50
arbitrary units of time). During this first part of the simulation, the expression of Cre
is repressed by the over representation of DAP. After that, the repression decreases
during 20 AUT until some Cre molecules are produced. At time 80 AUT, DAP
molecules appear which means that the differentiation occurred. On the contrary,
Fig. 10.6b shows the evolution of a germinal cell when DAPex remains constant
(i.e., expressions DAPEx++ and DAPEx-- are removed from the specification of
SMB STO). DAP molecules are exchanged between the cell and its environment un-
til an equilibrium is reached. While any differentiation occurs during this simulation,
the germinal cell will differentiate since parameter C4 is not null.
224 A. Spicher et al.

We propose to use the stochastic model to evaluate the differentiation rate of

the SMB. More specifically, we focus on the mean simulation time required for
a germinal cell to differentiate while DAPex is constant. Results are given on the
bottom line of Fig. 10.6. The protocol of this experiment consists in running 1000
simulations for each value of DAPex 2 Œ0; 30 and starting with the same initial state.
Each simulation stops when the differentiation occurs (i.e., when ‘dS appears in the
collection). Mean times and associated standard deviation are plotted in Fig. 10.6c.
Surprisingly both quantities seem to behave linearly with DAPex .
One has to pay attention to the fact that such simulations are costly in computing
time. Figure 10.6d gives the mean computation time of the simulation, showing
that it increases more than linearly with the value of DAPex . Actually, Gillespie’s
algorithm, in its original definition, only allows a small number of molecules.
As a conclusion, one can establish that the differentiation rate is easily related
to the quantity of DAP released in the environment by somatic cells. Such a re-
sult is meaningful as it relates quantities of different scales: the population and
cellular scale of the differentiation and the genetic and molecular scale of DAP
concentration.
Our stochastic simulations exhibit low concentrations of chemicals within a cell
(e.g., there are less than ten molecules of Cre after the differentiation occurs). This
result questions the pertinence of ODE models such as the one proposed by the Paris
team (footnote 1).

10.5.4 Integrative Modeling

So far, we have considered classical ways of modeling and simulating a biological

process at a given level of description. In this last model, we aim at simulating the
entire population of bacteria with an explicit representation of cells in a 2D space.7
In addition, we want to integrate in the model physical and biological behaviors.
Our proposition is based on the specification of cell–cell dynamical interactions
and the computation of the neighborhood of the cells using an implicit Delaunay
triangulation.

10.5.4.1 Description of the Model

In our proposition, bacteria are represented by circles localized in the 2D Euclidean

space, with a radius depending on their size. Bacteria push away each other and con-
sequently change their position in space and their immediate neighborhood. Thus,
this neighborhood is required to be dynamically computed according to the spatial

7
The third dimension is not considered as the SMB is supposed to grow in the plane of a Petri
dish, for example.
10 Interaction-Based Simulations for Systems Biology 225

coordinates of their associated circles. This approach has already been successfully
used in systems biology for the modeling of cell population (Gibson et al. 2006;
Barbier de Reuille et al. 2006a).
The modeling of SMB is organized into two coupled models: a mechanical model
and a biological model. A cell is encoded by a record which includes the posi-
tion, the radius, the local DAP concentration, the differentiation state (germinal or
somatic), etc. Cells are elements of a Delaunay topological collection. This kind of
collection computes implicitly and transparently the Delaunay triangulation of a set
of entities embedded in Rn (Aurenhammer 1991).
Mechanical model. The mechanical model consists of a mass/spring system. Bac-
teria are considered as punctual masses localized at the center of their associated
circle; the presence of a spring between two masses depends on the neighborhood
computed by the Delaunay triangulation. The mechanical effects of the growth of
the bacteria is captured by the elongation of the springs rest lengths. Thus, each
cell computes its acceleration by summing all mechanical forces induced by its
incident springs, and consequently moves in space. This is done by the transforma-
tion Meca.8 Meca sums the forces applied on each cell using a neighorsfold
expression. Then, the Euler transformation (see Sect. 10.5.1) is used twice to in-
tegrate during a time step 1 t acceleration into velocity and velocity into new
positions.
Biological model. The cellular model is an extension of the CA given in
Sect. 10.5.2. The discrete DAP diffusion is replaced by the classical continuous
model given in Sect. 10.3.3, and a stochastic differentiation is used instead of the
notion of populations ratio. New rules are added to deal with cellular growth, di-
vision, and death: in the presence of DAP, G cells grow by increasing their radius.
When the G cell radius reaches a threshold, the cell divides. S cells keep on growing
and then die when another threshold is reached. The corresponding transformation
is called Cell8 and computes the cellular evolution during a period 2 t.

10.5.4.2 Integration of the Two Models

As classical functions, transformations can be arbitrarily composed. This is the key

to the coupling of the two models. The iteration of a function can be specified by the
MGS option iter. It allows to deal with different time scales: assuming that the
mechanical process is faster than the cellular process (i.e., 2 t > 1 t), the whole
model is captured by the following evolution function:
fun SMB DEL(state) =
2 t
Cell[dt=2t](Meca[dt=1 t, iter= ](state))
1 t

8
The whole MGS program of the simulation is available at http://mgs.spatial-computing.org/
integrative biology.tgz.
226 A. Spicher et al.

where
option dt corresponds to the time step used in transformations Meca and Cell.
2 t
Here transformation Meca is applied times for only one application of Cell.
1 t

10.5.4.3 Interpretation of the Simulations’ Results

The protocol of the proposed simulation consists in iterating 10 000 times the
function SMB DEL starting from a small initial population of 25 germinal cells with
a deficit in DAP. Screenshots of the simulation are shown in Fig. 10.7. The vizual-
ization of the evolution exhibits two interesting phenomena.
On small population size (i.e., at the beginning of the simulation), the ratio of
the two cell lines fluctuates. Figure 10.7b presents a state of the population where
most of the cells are germinal. On Fig. 10.7c , S cells predominate. The oscillations
are due to delays between configurations where G cells are well fed (many S cells
are present) and configurations of DAP starvation (not enough S cells are present).
The fluctuations decrease as the population size increases, and the ratio globally
stabilizes as predicted by the ODE-based model (see Fig. 10.4). Indeed, the spatial
distribution blurs the synchronization between cells.
The population tends to spatially organize in the following way: uniformly dis-
tributed small clusters of G cells surrounded by somatic ones. Clusters remain
small-sized because when their size increase (by G cell divisions), the interior cells
lack DAP and differentiate. This dynamics is reminiscent of the behavior exhibited
by the CA model of Sect. 10.5.2.

Fig. 10.7 Results of the integrative model. Germinal cells are in dark gray and somatic cells in
light gray. Figure 10.7a–c correspond to an initial population and its evolutions at time 43AUT and
62AUT. Figure 10.7d–f focus on a cluster of germinal cells surrounded by somatic cells. See text
for explanation
10 Interaction-Based Simulations for Systems Biology 227

10.6 Related Work, Conclusions, and Perspectives

In this short conclusion section, we present related work, close to our approach,
used for the modeling and simulation of biological systems. Then, we conclude by
assessing the importance of using a single and coherent domain-specific language
for the modeling, at various spatial scales, of biological systems.

10.6.1 Related Work

In this chapter, we have illustrated the use of MGS, an experimental program-

ming language that investigates the use of topological collections and transfor-
mations in the simulation of complex biological systems. Based on the notion of
spatial interaction, MGS provides a unified simulation framework encompassing
discrete/continuous and deterministic/stochastic formalisms. Even though MGS is
a prototype in constant evolution, MGS’ concepts have been validated on numerous
applications: the crawling of the sperm cell of Ascaris suum (Spicher and Michel
2005), a simplified version of neurulation (Spicher and Michel 2007), the growth
of the meristem at a cellular level (Barbier de Reuille et al. 2006a), molecular self-
assembly (Giavitto and Spicher 2008a), the modeling of the paradigmatic phage
lambda genetic switch (Michel et al. 2009), etc.
As a side-effect, changing the topology of a collection makes it possible for MGS
to emulate some well-known computational models. Transformation on multiset is
reminiscent of multiset-rewriting (Banâtre and LeMetayer 1993). Nesting multiset
lead to P systems (Păun 2001), a new distributed parallel computing model based
on the notion of a membrane structure. P systems are advocated for the modeling of
compartmentalized molecular interaction networks. Lindenmayer systems (Linden-
mayer 1968a), which loosely correspond to transformations on sequences or string
rewriting, have long and successfully been used in the modeling of (DS)2 in the do-
main of plant growth.
And, as shown in Sect. 10.5.2, it is straightforward to express CA in MGS.

10.6.2 Conclusions and Perspectives

Work in systems biology generally puts a considerable emphasis on the end result,
the model of a complex biological system, and neglects so far the problem of build-
ing this model. The construction of a model and its use, e.g., for simulation, is a
difficult task, and it often requires the combination of many formalisms and comple-
mentary approaches. We also want to stress the importance of dynamical structures
in biological systems. This kind of dynamical systems is very challenging to model
and simulate. New programming concepts must be developed to ease their modeling
and simulation.
228 A. Spicher et al.

Using MGS’ topological collections and transformations allowed us to model

our problem in the same formal framework: a mean-field approach using ODE for a
quick proof of concept of the synthetic construction (Sect. 10.5.1), a discrete spatial
model using CAs on various topologies allowing a finer analysis of the relations
between diffusion and differentiation (Sect. 10.5.2), a robustness analysis on noise
sensitivity on the SMB parts (Cre/LOX recombination), and devices (recombina-
tion rate) parameters using a stochastic modeling (Sect. 10.5.3). The last model
integrates physical and biological constraints in a 2D model, making it possible to
analyze the effects of the spatial distribution on the various possible configurations
(Sect. 10.5.4).
The perspectives opened by this work are numerous. At the language level, the
study of the topological collections concepts must continue with a finer study of
transformation kinds. A lesson learned from the use of MGS by biologists is the
needs of user-friendly interfaces and of graphical tools for the analysis of the sim-
ulation’s results. Another direction of developments is the coupling of MGS with
various databases and repositories to retrieve parameters or to store and reuse frag-
ments of models.
The development of MGS is part of a “grand challenge” aimed at the develop-
ment of a methodological and technological framework that, once established, will
enable the sharing of models, the analysis, and the derivation of predictive hypoth-
esis from them.

Acknowledgements The authors would like to thank the reviewers for their valuable comments
on a first version of this chapter.
We gratefully acknowledge all the people who contributed to make the first French participation
in iGEM in 2007 a success: the students, D. Bikard, F. Caffin, N. Chiaruttini, T. Clozel, D. Guegan,
T. Landrain, D. Puyraimond, A. Rizk, E. Shotar, G. Vieira, the instructors, F. Delaplace, S. Bottani,
A. Jaramillo, A. Lindner, V. Schächter; the advisors, F. Le Fevre, M. Suarez, S. Smidtas, A. Spicher,
and P. Tortosa.
Further acknowledgments are also due to J. Cohen, B. Calvez, F. Thonnerieux, C. Kodrnja, and
F. Letierce who have contributed in various ways to the MGS project.
This research is supported in part by the the University of Évry, the University of Paris-Est,
the CNRS, GENOPOLE-Évry, the Institute for Complex Systems in Paris-Ile de France, the ANR
white project AutoChem and the French working group GDR GPL/LTP.

Appendix: An MGS Implementation of the Runge–Kutta

Methods

The Runge–Kutta methods is a famous and widely used family of integration

scheme for solving problems of the form:

dX.t/
D f .X.t/; t/ X.0/ D X0 (10.16)
dt
They are based on the techniques developed by C. Runge and M.W. Kutta at the
beginning of the twentieth century.
10 Interaction-Based Simulations for Systems Biology 229

In the following, we propose an MGS implementation of the classical explicit

fourth order Runge–Kutta method (RK4). This example can be extended to most of
other Runge–Kutta methods. As the Euler method, the RK4 allows the computation
of a sequence of vectors Xn with a more complex scheme reducing errors of approx-
imations. Starting from an initial state X0 at time t0 , a step a the RK4 is given by
t
XnC1 D Xn C .k1 C 2k2 C 2k3 C k4 / tnC1 D tn C t (10.17)
6
with
 
tk1 t
k1 D f .Xn ; tn / k2 D f Xn C ; tn C
2 2
 
tk2 t
k3 D f Xn C ; tn C k4 D f .Xn C tk3 ; tn C t/ (10.18)
2 2

The MGS implementation can be divided into 3 steps:

t
1. The computation of expressions Xn C ck where c is a coefficient (either or
2
t) and k takes the value ki .
2. The computation of XnC1 knowing the ki .
3. The composition of the two first steps.
Steps (1) and (2) are straightforward and can be implemented as follows:
trans RK4a[c,k] = f x => x + c*k.(ˆx) g
trans RK4b[dt,k1,k2,k3,k4] = f
dt
x => x + ( k1.(ˆx) + 2*k2.(ˆx) + 2*k3.(ˆx)
6
+ k4.(ˆx) )
g
These transformations apply the required computations on each coordinate of X.
The final step consists in implementing the function RK4:
fun RK4[dt,f](X,t) = (
let k1 = f(X,t) in
dt dt
let k2 = f(RK4a[c= ,k=k1](X),t+ ) in
2 2
dt dt
let k3 = f(RK4a[c= ,k=k2](X),t+ ) in
2 2
let k4 = f(RK4a[c=dt,k=k3](X),t+dt) in
(RK4b[dt=dt,k1=k1,k2=k2,k3=k3,k4=k4](X),t+dt)
)
230 A. Spicher et al.

References

F. Aurenhammer. Voronoi diagrams–A survey of a fundamental geometric data structure. ACM

Comput Surv, 23(3):345–405, 1991
F. Bailly and G. Longo. Mathmatiques et sciences de la nature. Hermann, Paris, 2006
J.-P. Banâtre and D. LeMetayer. Programming by multiset transformation. Comm ACM, 36(1):98,
1993
J.-P. Banâtre, P. Fradet, and Y. Radenac. Generalised multisets for chemical programming. Math
Struct Comput Sci, 16(4):557–580, 2006
P. Barbier de Reuille, I. Bohn-Courseau, K. Ljung, H. Morin, N. Carraro, C. Godin, and J. Traas.
Computer simulations reveal novel properties of the cell-cell signaling network at the shoot
apex in Arabidopsis. Proc Natl Acad Sc USA, 103(5):1627–1632, 2006a
K. De Cock, X. Zhang, M. F. Bugallo, and P. M. Djuric. Stochastic simulation and parameter
estimation of first order chemical reactions. In 12th European Signal Processing Conference
(EUSIPCO’04), 2003
R. Durrett and S. Levin. The importance of being discrete (and spatial). Theor Popul Biol, 46(3):
363–394, 1994
M. Eden. A two-dimensional growth process. In Proceedings of Fourth Berkeley Symposium on
Mathematics, Statistics, and Probability, Vol. 4, pages 223–239, 1961
M. Eigen and P. Schuster. The hypercycle: A principle of natural self-organization. Springer,
Berlin, 1979
D. Endy. Foundations for engineering biology. Nature, 438:449–453, 2005
E. Fermi, J. Pasta, and S. Ulam. Studies of nonlinear problems, LASL Report LA-1940 (5). Tech-
nical report, 1965. Reprinted in the collected work of E. Fermi, Vol. 2, pp. 977–988, 1965
M. Fisher, G. Malcolm, and R. Paton. Spatio-logical processes in intracellular signalling. BioSys-
tems, 55:83–92, 2000
W. Fontana. Algorithmic chemistry. In C. G. Langton, C. Taylor, J. D. Farmer, and S. Rasmussen,
editors, Proceedings of the Workshop on Artificial Life (ALIFE’90), Vol. 5, pages 159–210,
1992
W. Fontana and L. W. Buss. “The arrival of the fittest”: Toward a theory of biological organization.
Bull Math Biol, 1994
J.-L. Giavitto. Topological collections, transformations and their application to the modeling and
the simulation of dynamical systems. In Rewriting Techniques and Applications (RTA’03),
LNCS 2706, pages 208–233, 2003
J.-L. Giavitto and O. Michel. Declarative definition of group indexed data structures and approx-
imation of their domains. In Proceedings of the 3rd ACM SIGPLAN International Conference
on Principles and Practice of Declarative Programming (PPDP’01), pages 150–161, 2001
J.-L. Giavitto and O. Michel. Data structure as topological spaces. In Proceedings of the 3nd In-
ternational Conference on Unconventional Models of Computation UMC’02, Vol. LNCS 2509,
pages 137–150, 2002a
J.-L. Giavitto and O. Michel. Pattern-matching and rewriting rules for group indexed data struc-
tures. ACM SIGPLAN Not, 37(12):76–87, 2002b
J.-L. Giavitto and O. Michel. The topological structures of membrane computing. Fundam Inform,
49:107–129, 2002c
J.-L. Giavitto and O. Michel. Modeling the topological organization of cellular processes. BioSys-
tems, 70(2):149–163, 2003
J.-L. Giavitto and A. Spicher. Simulation of self-assembly processes using abstract reduction sys-
tems. In Systems self-assembly: Sultidisciplinary snapshots, pages 199–223. Elsevier, 2008a
J.-L. Giavitto and A. Spicher. Topological rewriting and the geometrization of programming. Phys-
ica D, 237:1302–1314, 2008b
J.-L. Giavitto, C. Godin, O. Michel, and P. Prusinkiewicz. Modelling and simulation of biological
processes in the context of genomics. In Computational Models for Integrative and Develop-
mental Biology, 2002a
10 Interaction-Based Simulations for Systems Biology 231

J.-L. Giavitto, O. Michel, and J. Cohen. Pattern-matching and rewriting rules for group indexed
data structures. In ACM Sigplan Workshop (RULE’02), pages 55–66, 2002b
J.-L. Giavitto, G. Malcolm, and O. Michel. Rewriting systems and the modelling of biological
systems. Comp Funct Genomics, 5:95–99, 2004
M. C. Gibson, A. B. Patel, R. Nagpal, and N. Perrimon. The emergence of geometric order in
proliferating metazoan epithelia. Nature, 442:1038–1041, 2006
D. T. Gillespie. Exact stochastic simulation of coupled chemical reactions. J Phys Chem, 81(25):
2340–2361, 1977
J. Greenberg and S. Hastings. Spatial patterns for discrete models of diffusion in excitable media.
SIAM J Appl Math, pages 515–523, 1978
Y. Itkis. Control systems of variable structure. Wiley, New York, 1976
P. Jansson and J. Jeuring. PolyP – A polytypic programming language extension. In Principles of
programming languages, pages 470–482. ACM, 1997
T. Knight. Idempotent vector design for standard assembly of biobricks, 2006. MIT Synthetic
Biology Working Group
A. Lindenmayer. Mathematical models for cellular interaction in development, Parts I and II.
J Theor Biol, 18:280–315, 1968a
P. L. Luisi. Autopoiesis: A review and a reappraisal. Naturwissenschaften, 90:49–59, 2003
J. Lynch. Logical characterization of individual-based models. In 23rd Annual IEEE Symposium
on Logic in Computer Science (LICS’08), volume n, pages 379–390, 2008
H. McAdams and L. Shapiro. Circuit simulation of genetic networks. Science, 269(5224):650,
1995
O. Michel, A. Spicher, and J.-L Giavitto. Rule-based programming for integrative biological mod-
eling – Application to the modeling of the  phage genetic switch. Nat Comput, 8(4):865–889,
2009
J. Munkres. Elements of algebraic topology. Addison-Wesley, Reading, MA, 1984
G. Păun. From cells to computers: Computing with membranes (P systems). Biosystems, 59(3):
139–158, 2001
N.M. Shnerb, Y. Louzoun, E. Bettelheim, and S. Solomon. The importance of being discrete: Life
always wins on the surface. PNAS, 97(19):10322–10324, 2000
J. Smith. Shaping life: Genes, embryos and evolution. Yale University Press, New Haven, 1999
A. Spicher and O. Michel. Using rewriting techniques in the simulation of dynamical systems:
Application to the modeling of sperm crawling. In Fifth International Conference on Compu-
tational Science (ICCS’05), Part I’, Vol. 3514 of LNCS, pages 820–827, 2005
A. Spicher and O. Michel. Declarative modeling of a neurulation-like process. BioSystems, 87
(2–3):281–288, 2007.
A. Spicher, N. Fats, and O. Simonin. From reactive multi-agents models to cellular automata. In
International Conference on Agents and Artificial Intelligence, pages 422–429, 2009
A.M. Turing. The chemical basis of morphogenesis, Series B: Biological Sciences. Phil Trans R
Soc Lond, 237:37–72, 1952
G. Turk. Generating textures for arbitrary surfaces using reaction-diffusion. In T.W. Sederberg,
editor, Computer Graphics (SIGGRAPH ’91 Proceedings), pages 289–298, 1991
F.J. Varela, H.R. Maturana, and R. Uribe. Autopoiesis: The organization of living systems, its
characterization and a model. BioSystems, 5:187–196, 1974
G. Von Dassow, E. Meir, E. M. Munro, and G. Odell. The segment polarity network is a robust
developmental module. Nature, 406(6792):188–192, 2000
J. Von Neumann. Theory of self-reproducing automata. University of Illinois Press, Urbana, 1966
S. Wolfram. Theory and applications of cellular automata. World Scientific Publication,
Singapore, 1986
M. Woolridge and M. Wooldridge. Introduction to multiagent systems. Wiley, New York, 2001
X. Zhang, K. De Cock, M.F. Bugallo, and P.M. Djuric. Stochastic simulation and parameter esti-
mation of enzyme reaction models. In IEEE Workshop on Statistical Signal Processing, 2003
Glossary

Antibiotic resistance The ability of a microorganism to resist the effects of an

antibiotic. Examples of mechanisms of antibiotic resistance include the synthesis of
antibiotic-degrading enzymes (e.g., ˇ-lactamase), and modifications to drug targets
such as the penicillin-binding proteins (PBPs) in bacterial cell membranes.
BAIT Bacteria-Antibiotic Interaction Tool (BAIT) is a precursor of Micro-Gen,
which implements a simple model of bacterial growth and interactions with antibi-
otic molecules in a discrete, 2D environment.
Bibliome A collection composed of the primary and review literature, in addition
to textbooks on a particular topic.
Column space For a matrix S and vector v, any vector that satisfies the relationship
S  v D b. When b D dx=dt, the time derivative of the concentration vector, then
the column space describes metabolite dynamics.
Delay differential equations with discrete time delays Ordinary differential
equations where the derivatives of the state variables at time t depend on the state
variables at time t and possibly at a finite number of discrete times less than t. Such
equations are functional differential equations of retarded type.
Dynamic instability The existence of two phases of growing and shrinking of mi-
crotubules in a population with rare transitions between these phases is referred to
as dynamic instability.
Dynamical system with a dynamical structure A kind of dynamical system that
requires its phase space, used to characterize the structure of the state of the system
at each time step, to be computed jointly with the running state of the system.
Endocrine hormone A chemical substance produced by a ductless gland of the
body, secreted into the blood steam and transported via the blood to affect other bod-
ily organs having cell receptors specific to that hormone. For example, the pituitary
gland secretes follicle stimulating hormone which promotes the growth of follicles
within the ovaries.
Global regulator A transcription factor or another regulatory protein capable of
controlling the expression of a large number of others genes.

W. Dubitzky et al. (eds.), Understanding the Dynamics of Biological Systems: Lessons 233
Learned from Integrative Systems Biology, DOI 10.1007/978-1-4419-7964-3,
c Springer Science+Business Media, LLC 2011
234 Glossary

Genetic regulatory network A network of genes, RNAs, proteins, metabolites,

and their mutual regulatory interactions.
Genome-scale The characterization of a of biological function and components on
spanning the genome of the respective organism, i.e., incorporation/consideration of
all known associated components encoded in the organisms genome.
Hill function In biochemistry, the binding of a ligand to a macromolecule is re-
ferred to as cooperative binding. The Hill function (or Hill equation) is used to
describe this effect. It is defined as y D KŒxh =.1 C KŒxh /, where y, the frac-
tional saturation, is the fraction of the total number of binding sites occupied by the
ligand, Œx is the free (unbound) ligand concentration, K is a constant, and h is the
Hill coefficient.
Integrative spatial systems biology An emergent field in systems biology that
deals with the necessary integration of spatial properties into integrative biology.
Jacobian matrix The Jacobian matrix refers to a matrix of all first-order partial
derivatives of a vector-valued function. The Jacobian matrix represents the best lin-
ear approximation to a differentiable function near a given point.
(Left) null space For a matrix S and vector v, any vector that satisfies the rela-
tionship ST  v D 0, in which ST is the transpose of the matrix, is said to lie in
the left null space of S. This corresponds to conserved chemical moieties (i.e., com-
pounds or chemical groups that are neither consumed nor produced in the network)
in metabolic networks.
Law of mass action In chemistry, the law of mass action states that the rate of
a chemical reaction is directly proportional to the molecular concentrations of the
reacting substances. The law of mass action covers the equilibrium as well kinetic
aspects (reaction rates) of chemical reactions.
MGS A domain-specific language aimed at the modeling and simulation of dynam-
ical systems with a dynamical structure. MGS relies on the notions of local interac-
tions, topological collections and their transformation in a declarative framework.
Menstrual cycle The reproductive cycle of a human female. The menstrual cy-
cle describes the monthly changes which occur in a woman of reproductive age
needed for the creation of an embryo. Under the control of endocrine hormones,
follicles develop in the ovaries and then the dominant follicle releases its egg (ovu-
lation) and becomes the corpus luteum which prepares the body for fertilization and
implantation.
MRSA Methicillin-resistant Staphylococcus aureus (MRSA) is a multi-drug resis-
tant form of S. aureus which was first isolated in 1961. Resistance conferred by
expression of penicillin-binding protein 2a which has reduced binding to ˇ-lactam
antibiotics.
Model checking Algorithms that test the truth value of properties expressed in
temporal logic on a state transition graph.
Glossary 235

Model reduction The approximation of a model of a complex (non-linear) dynam-

ical systems, with the aim of obtaining a simplified model that is easier to analyze
but preserves essential properties of the original model.
(Right) null space For a matrix S and vector v, any vector that satisfies the rela-
tionship S  v D 0, is said to lie in the right null space of S. This corresponds to
steady state flux distributions in metabolic networks.
Row space For a matrix S and vector v, any vector that satisfies the relationship
ST  v D b. When b D dx=dt, the time derivative of the concentration vector, then
the column space describes flux dynamics.
Ordinary differential equation In chemical kinetic theory, the interactions
between species are commonly expressed using ordinary differential equations
(ODEs). An ODE is a relation that contains functions of only one independent vari-
able (typically t), and one or more of its derivatives with respect to that variable. The
order of an ODE is determined by the highest derivative it contains (for example,
a first-order ODE involves only the first derivative of the function). The equation
5x.t/ C x.t/
P D 17 is an example of a first-order ODE involving the independent
P
variable t, a function of this variable, x.t/, and a derivative of this function, x.t/.
Since a derivative specifies a rate of change, such an equation states how a function
changes but does not specify the function itself. Given sufficient initial conditions,
various methods are available to determine the unknown function. The difference
between ordinary differential equations and partial differential equations is that
partial differential equations involve partial derivatives of several variables.
Partial differential equation Is similar to an ordinary differential equation except
that it involves functions with more than one independent variable.
Piecewise-linear model A set of linear differential equation models, each of which
is defined on a part of the state space.
Ran A Ras related GTPase that can be in a GTP (guanosine tri-phosphate) or GDP
(guanosine di-phosphate) bound state. It is implicated in nuclear import of proteins
and cell division.
RanGAP Ran GTPase activating protein. It accelerates the hydrolysis of GTP
by Ran.
RCC1 Regulator of chromatin condensation 1 (RCC1) is a nuclear protein. Also
referred to as RanGEF (Ran guanine nucleotide exchange factor). It is imported to
the nucleus and binds to chromatin.
Search and capture Microtubules radially nucleated from a centrosome search
cellular space till they encounter the centromere of a chromosome and are captured.
This is process is a model of events that lead to successful chromosome-microtubule
linkage, critical in cell division and maintenance of ploidy.
Sensitivity analysis An important tool to study the dependence of systems on their
parameters. Sensitivity analysis helps to identify those parameters that have signif-
236 Glossary

icant impact on the system output and capture the essential characteristics of the
system. Sensitivity analysis is particularly useful for complex biological networks
with a large number of variables and parameters.
State transition graph A directed graph representing the behavior of a dynamic
system in computer science. The nodes of the graph correspond to the states of
the system, whereas the edges account for transitions from one state to another.
More specifically, in the qualitative modeling of genetic regulatory networks using
piecewise-linear models, the nodes describe a qualitative state of a network, consist-
ing of the region in the state space in which the system resides, and the signs of the
derivatives of the concentration variables in this region.
Temporal logic A formalism for describing sequences of transitions between states
of a system, where the notion of time-order is introduced through the use of temporal
operators.
Topological collection A collection of entities structured by a neighborhood rela-
tionships and handled as a whole. The fundamental data structure of MGS for the
representation of the state of a dynamical system.
Transformations A set of rules acting on a topological collection. Used in MGS
for the specification of the evolution function of a dynamical system.
Zero-order ultrasensitivity The chemical reactions in which the product forma-
tion is ultrasensitive and independent of substrate concentration since the enzyme is
working at saturation (i.e. it depends on the zeroth power of substrate concentration)
are seen to display zero-order ultrasensitivity.
ˇ-lactamase Enzyme produced by bacterial cells which degrades ˇ-lactam antibi-
otic molecules by cleaving their ˇ-lactam ring structure.
Index

A Global regulators of transcription, 112–114,

ADME, 59 119, 120, 123
Analysis Gradient matrix, 23–26, 28, 34–36
tool
Genetic Network Analyzer, 119
Antibiotic resistance, 132, 133, 138, 141, 147 H
Hill function, 44, 46, 53, 66, 72, 117

B
I
Bifurcation, 52
Inhibin, 40, 45, 46, 51, 56
Bilinear interaction, 23, 34
Interaction
Biological network reconstruction, 85, 89
antibiotic molecules – cell surface, 131
Boolean network, vii, viii, 85, 102, 104, 112,
antibiotics-bacteria, 132, 137, 142
125, 135
bacterial, 136
cell-cell, 179, 224
drug-cancer, 78
C
drug-system, 84
CellDesigner, 78
drug-target, 84
Cellular automaton, 135
host-pathogen, 95, 102–105
Contraceptives, 41
human-virus, 83
COPASI, 78
kinetic, 26
protein-protein, 2, 84, 88, 96, 98
D spatial, 196, 202, 227
Drug resistance, 75 Interactome, 96, 97, 101, 103

J
E Jacobian matrix, 23, 25, 35, 36
E. coli, 112
Estradiol, 40
K
Kinetics
F Hill, 114
Flux balance analysis, viii, 20, 21, 89 Michaelis-Menten, 114
Flux-concentration duality, 24, 26
Folate, viii, 27
M
Mass action, 136, 221
G kinetics, 28, 66
Genetic regulatory network, 112, 125 law, 221

237
238 Index

Menstrual cycle, viii, 39, 41, 45, 51, 53 P

Methionine, 28, 30, 31 Pituitary gland, 39, 42
Michaelis-Menten
constant, 8
equation, 62, 74
Microtubule, ix, 76–78, 156, 158, 160, 161, S
165, 168 Sensitivity analysis, viii, 7, 9, 12, 14, 15, 50,
Model 77, 78, 193
agent-based, 132, 136, 137, 147, 151
Sensitivity coefficient, 52, 53
kinetic, 19
Simulation
qualitative, 112
Model checking, 123, 124 bacteria-antibiotic interactions, 147
Modeling State transition graph, 117, 118, 120, 121, 123
bottom-up, 64, 132 Stoichiometric matrix, 20, 23, 26, 28, 30, 35,
multi-scale, 190 89, 93
tool
CellDesigner, 77
COPASI, 77
MRSA, 132, 134, 142, 143 T
Mycobacterium tuberculosis, 83 Target identification, 2, 16, 84, 87, 89, 91, 93,
94, 96–98, 106
O Temporal logic, 121
Ovulation, 39, 41, 42, 45, 53, 54, 56 Tuberculosis, 83, 100