Professional Documents
Culture Documents
Natural Computing Series
Natural Computing Series
CoDtputation in
Cells and Tissues
Perspectives and Tools of Thought
~ Springer
Editors J. Howard Parish
Ray Paton' School of Biochemestry
and Molecular Biology
Hamid Bolouri
University of Leeds
Institute for Systems Biology Leeds LS2 9JT, UK
Seattle, \VA 98103, USA howard@bmb.ac.uk
hbolouri@systemsbiology.org
Series Editors
Mike Holcombe
G. Rozenberg (Managing Editor)
Department of Computer Science
rozenber@liacs.nl
University of Sheffield
Th. Băck, J.N. Kok, H.P. Spaink
Sheffield SI 4DP, UK
m.holcombe@dcs.shef.ac.uk Leiden Center for Natural Computing
Leiden University
Richard Tateson Niels Bohrweg 1
Future Technologies Group 2333 CA Leiden, The Netherlands
Intelligent Systems Lab
A.E.Eiben
BTexact Technologies
Vrije Universiteit Amsterdam
Ipswich IPS 3RE, UK
The Netherlands
richard. tateson@bt.com
This work is subject to copyright. Ali rights are reserved, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting. reproduction on microfilm or in any other way, and storage in data
banks. Duplication of this publication or parts thereof is permitted only under the provisions
of the German Copyright Law of September 9, 1965, in its current version, and permission for
use must always be obtained from Springer-Verlag Berlin Heidelberg GmbH.
Violations are liable for prosecution under the German Copyright Law.
springeronline.com
© Springer-Verlag Berlin Heidelberg 2004
Originally published by Springer-Verlag Berlin Heidelberg New York in 2004
Softcover reprint of the hardcover 1st edition 2004
The use of general descriptive names, registered names, trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
Cover Design: KiinkelLopka, Werbeagentur, Heidelberg
Typesetting: by the Authors
Production: LE-TEX Jelonek, Schmidt & Vockler GbR, Leipzig
Printed on acid-free paper 4513142/YL - 543210
It is with great sadness that we have to report the sudden death of Dr. Ray Paton,
the principal editor of this volume, just before we went to press. Ray worked tire-
lessly to bring this book to Jruition and it stands as a rich testament to his inspira-
tionalleadership and vision in the field.
Alt of the other editors wish to record our great gratitude to Ray, who was not
only an outstanding scientist but also agreat friend and colleague.
We hope that this book will, in some way, be looked upon as a memorial to Ray's
pioneering work in biologically inspired computing and computational biology.
Preface
Ray Paton'
Biocomputing and Computational Biology Group
Department of Computer Science
May 2004 The University of Liverpool, UK
Contents
Categorical Language and Hierarchical Models for Cell Systems ................ 289
R. Brown et al.
1 Introduction .......................................................................................... 289
2 Category Theory: History and Motivation ........................................... 290
3 Categorical Models for Hierarchical Systems ...................................... 292
4 Conc1usion ............................................................................................ 302
References .................................................................................................... 302
R. Patod
Cells are complex systems. For some people, they are like tiny chemical plants, or
laboratories, or machines. For others they are more like computational devices or
even computational systems. As data continue to be generated about them and
their components and the systems that they make up, new perspectives and models
are needed to deal with the complexity. Cells are more than bags of chemicals just
as macromolecules are more than microscopic billiard balls or strings of beads.
The challenges of an information processing view that complements the more
commonly expressed chemical processing view needs to be taken into account.
The biomolecular, cellular and tissue levels of biologic al organisation have had
considerable inspirational impact on the development of computational models
and systems. Such innovations include neural computing, systolic arrays, genetic
and immune algorithms, cellular automata, artificial tissues, molecular computing,
and protein memories. With the rapid growth in biological knowledge there
remains a vast source of ideas yet to be tapped. These include developments
associated with biomolecular, genomic, enzymic, metabolic and signalling
systems and the various impacts on distributed, adaptive, hybrid and emergent
computation. Many biologists use language that is displaced from computer
sciences not least, program (as in apoptosis), hardware-software, DNA - as data
or program, code, gate, Boolean network, pattern recognition, and so forth.
Indeed, many proteins (such as enzymes and transcription factors) carry out
complex information processing tasks (as individual molecules) including pattern
recognition, switching, logic al decision-making, memory, and signal integration.
This book provides readers with a comprehensive exploration of this subject
from a uniquely multidisciplinary point of view. Contributions from biologists,
computer scientists, engineers and mathematicians are drawn together to provide a
comprehensive picture of biological systems, both as sources for ideas about
computation and as information processing systems. The varieties of perspectives
that are presented provide an integrative view of a complex and evolving field of
knowledge that needs new tools of thought to manage and mobilise the vast
opportunities afforded to the postgenomic and biocomputational sciences.
2 R. Paton
The book begins with several chapters that represent a (bio )mimetic approach to
engineering and computation. Tateson, a biologist by training who now works for
BTexact Technologies, looks at the application of cellular analogies to
telecommunication systems. He argues that in some cases there are sound reasons
for believing that analogies will be helpful. We can identify biological systems,
which resemble in some sense an artificial system, and then use our knowledge
and understanding of the functioning of the biological system to improve or
redesign the artificial system. In other cases there is little more than the assertion
that 'nature knows best' and hence any artefact modelled on nature must be
superior to an artefact for the same purpose devised by human reason alone. This
is not a useful basis for redesigning our artificial systems, and in fact the analogy
with nature is often 'bolted on' to the human-designed system to explain to non-
engineers how it works rather than being genuinely useful at the design stage.
The chapter by Bull and Tomlinson shows how symbiogenetic mechanisms
found at the cellular level can be successfully applied to computational leaming.
Symbiosis is the phenomenon in which organisms of different species live
together in close association, potentially resulting in a raised level of fitness for
one or more of the organisms. Symbiogenesis is the name given to the process by
which symbiotic partners combine and unify - fonning endosymbioses and then
potentially transferring genetic material - giving rise to new morphologies and
physiologies evolutionarily more advanced than their constituents. This process is
known to occur at many levels, from intra-cellular to inter-organism. They use the
abstract NKCS model of coevolution to examine endosymbiosis and its effect on
the evolutionary performance of the entities involved. They suggest the conditions
under which endosymbioses are more likely to occur and why; we find they
emerge between organisms within a window of their respective "chaotic gas
regimes" and hence that the association represents a more stable state for the
partners. This general result is then exploited within a machine leaming
architecture to improve its performance in non-Markov problem domains.
Timmis, Knight, Castro and Hart discuss the growing field of Artificial
Immune Systems (AIS) - that is using the natural immune system as a metaphor
for solving computational problems. The field of AIS is relatively new and draws
upon work done by theoretical immunologists such as Jeme, Perelson, and Varela.
What is of interest to researchers developing AIS is not the modelling of the
immune system, but extracting or gleaning useful mechanisms that can be used as
a metaphor to help solve particular problems. It is quite common to see gross
simplifications of the way the immune system works, but this is not a problem as
it is inspiration computer scientists seek from nature rather than precise
mechanisms. The review is organised in the following manner. First, reasons for
why the immune system has generated such interest and is considered to be a good
metaphor to employ are discussed. This is followed by a simple review of relevant
immunology that creates many of the foundations for work reviewed in this
contribution. Immunology is a vast topic and no effort has been made to cover the
CytoComputational Systems - Perspectives and Tools of Thought 3
whole area. Rather, only those ideas that have proved to be useful to the majority
of research presented in this contribution are explained in some detail. Finally, a
summary of the work presented in this contribution is provided, drawing main
conclusions from the work presented and commenting on the perceived future of
this emerging technology.
Tyrrell considers the analogy between multi-cellular organisms and multi-
processor computers as not too far-fetched, and well worth investigating,
particularly when considering that nature has achieved levels of complexity that
far surpass any man-made computing system. The aspect of biological organisms
on which this chapter is centred is their phenomenal robustness: in the trillions of
cells that make up a human being, faults are rare, and in the majority of cases,
successfully detected and repaired. The Embryonics project (for embryonic
electronics) is inspired by the basic processes of molecular biology and by the
embryonic development of living beings. By adopting certain features of cellular
organisation, and by transposing them to the two-dimensional world of integrated
circuits in silicon, it will be shown that properties unique to the living world, such
as self-replication and self-repair, can also be applied to artificial objects
(integrated circuits). Self-repair allows partial reconstruction in case of a minor
fault, while self-replication allows complete reconstruction of the original device
in cases where a major fault occurs. These two properties are particularly desirable
for complex artificial systems in situations that require improved reliability. To
increase still further the potential reliability of these systems, inspiration has also
been taken from biological immune systems - Immunotronics. The acquired
immune system in humans (and most vertebrates) has a mechanism for error
detection which is simple, effective and adaptable.
The chapter by McNeil and Snowdon serves to provide a further conceptual
bridge between hardware and biology, this time by working with molecules at the
nanoscale. The dawn of nanoscale science can be traced to a now classic talk that
Richard Feynman gave on December 29th, 1959 to the annual meeting of the
American Physical Society at the California Institute of Technology. In this
lecture, Feynman suggested that there exists no fundamental reason to prevent the
controlled manipulation of matler at the scale of individual atoms and molecules.
Twenty one years later, Eigler and co-workers constructed the first man-made
object atom-by-atom with the aid of a scanning tunnelling microscope. Given that
there is "Plenty of room at the bottom" (the title of Feynman's talk), and biological
systems have highly subtle and sophisticated meso- and micro-scale architectures,
the exploitation of this level in medical and computational technologies will
continue to challenge 21 st century biocomputational science.
The kind of approach discussed in the previous chapters has an established
record since the 1940s and the developments in cybernetics, digital electronics,
and general models of computation. Thus we find the developments in digital
models of neurones (the McCulloch-Pitts model), computational models of brains,
and the origins of cellular automata. Many of the ideas that were spawned in the
1940s and 1950s, and many tools of thought for helping scientists and engineers to
organise their knowledge of biological systems can be traced back to this time (as
4 R. Paton
can the revolution that took place in molecular biology). The viewpoint now
moves along towards ways in which the languages of physical science and
mathematics can enhance our appreciation of biological systems.
A common observation made by scientists who are working at multi-
disciplinary interfaces (such as CytoComputational Systems) relates to the
problems encountered by differences in vocabulary, emphasis, modelling
approach, and attitudes to reduction and simplification. The next chapter looks at
some ways displacements of ideas between the disciplines can take place. Nagl,
Parish, Paton, and Wamer consider ways of describing computational topics in
molecular and cellular biology. Methods of classifying DNA, RNA and proteins
are central to current methods for elucidating relationships between sequence,
structure and function. The chapter looks at metaphors for the function of proteins
and points to a unified view of proteins as computational devices capable of
matching patterns as inputs and processing to re suIt in alternative outputs. The
requirement for a systems view of life is also pursued. As such this chapter
provides an immediate bridge between the previous few and next few chapters. In
addition it also anticipates a number of themes emerging in later chapters (such as
Fisher et al. and Wolkenhauer et al.).
Following on from this computational stance, Bolouri and Schilstra provide a
short review of the modelling of Genetic Regulatory Networks (GRNs). GRNs
have a basic requirement to model (at least) some parts of a biological system
using some kind of logical formalism. They represent the set of alI interactions
among genes and their products for determining the temporal and spatial patterns
of expres sion of a set of genes. The origins of modelling the regulation of gene
expres sion go back to the Nobel Prize winning work of Lwoff, Jacob and Monod
on the mechanisms underlying the behaviour of bacterial viruses that switch
between so-called lytic and lysogenic states. The authors briefly discuss some of
the circuit-based approaches to GRNs such as the work of Kauffman, Thomas, and
Shapiro and Adams.
The next two chapters address computational modelling of cells using very
different approaches. The chapter by Gregory, Paton, Saunders and Wu reports on
work concerned with Individual-Based Models (IBMs) of a 'virtual' bacterial cell.
The goal of this project has been to explore the ecological and evolutionary
trajectories of 'artificial bacteria'. This is agreat challenge both to understanding
the cell in sufficient detail and to implementing a system on current computational
architectures. Each bacterium is an independent agent with sufficient genomic and
proteomic equivalents built into the model such that each individual cell can have
up to 120 genes and 50,000 gene products. The chapter reports on the
development of a model that has to incorporate multiple scales in both time and
space.
Feng's chapter develops some mathematical models of stochastic computations
in neurones and neural networks. It is a bridge between the cell-based work
discussed previously, the tissue-based work we look at next, and the bioinspired
approaches that started the book. Feng discusses the developments of his neuronal
decision theoretic approach in relation to the role played by inhibitory inputs. The
CytoComputational Systems - Perspectives and Tools of Thought 5
cellular components execute a leaming rule and the networks that are produced
can be applied to statistical pattern recognition problems.
The next two chapters provide very interesting insights into the workings of
mathematical modellers as their work addresses very specific biological problems.
Monk's chapter looks at his work dealing with spatial patterning in explicitly
cellular environments. Pattern formation in multicellular organisms generally
occurs within populations of cells that are in close contact. It is thus natural and
important to consider models of pattern formation that are constructed using a
spatially discrete cellular structure. Here, the particular case of pattern formation
in cellular systems that depends on contact-dependent (juxtacrine) signalling
between cells is discussed.
At another scale of biological organisation, MacGregor, Leng and Brown
describe a model of the hypothalamic and pituitary components involved in
controlling growth hormone release. Their model has been developed by gathering
and attempting to formali se the experimental data about the system but has been
kept as simple as possible, focusing on the functional rather than mechanical
properties of its components. They show that a relatively simple model can be
capable of producing complex behaviour and accurately reproducing the
behaviour and output of a real brain system.
We now address a collection of modelling approaches that build on a
computational perspective (that is, the underlying metaphor is focused on
computation or information processing rather than the dynamics expressed in
models based on continuous mathematics). Holcombe notes how computational
models have been of interest in biology for many years and have represented a
particular approach to trying to understand biological processes and phenomena
from a systems point of view. One of the most natural and accessible compu-
tational models is the state machine. These come in a variety of types and possess
a variety of properties. This chapter discusses some useful ones and looks at how
machines involving simpler machines can be used to build plausible models of
dynamic, reactive and developing biological systems that exhibit hierarchical
structures and behaviours.
Another computational stance that has been related to coding/decoding issues
concerns the sequence structures and scrambling of genes. Sant and Amos
examine some recent work on models of recombination in Ciliates. They describe
how these single-celled organisms 'compute' by unscrambling their genetic
material. They present a number of models of this process, and discuss their
implications. This could have useful implications for the development of cellular
computers.
In contrast to the previous chapter, which looks at genes and DNA, the chapter
by Fisher, Malcolm and Paton looks at proteins and, specifically, the modelling of
signalling proteins as algebraic agents. Harking back to the previous chapter by
Nagl et al., the authors begin by looking at proteins as computational agents.
Protein information processing networks are discussed, notably secondary
messenger signaling, signalling kinases and phosphatases, scaffold proteins and
6 R. Paton
The CytoCom project was one of a number of networks of scientists funded by the
UK Engineering and Physical Sciences Research Council looking at Emerging
Computing Paradigms. In our case we were using non-neural cells and tissues as
the area of focused study and discussion. We held five workshops between 1999
and 2001, in Liverpool, Leeds, Hertford, Sheffield and London. This book,
together with a number of other publications, further networks and funded
research, were some of the achievements. A less quantifiable achievement was the
general increase in awareness of this level of biologic al organisation as a source of
computational ideas. CytoCom grew from an existing, though loose community of
scientists interested in and attending an international series of workshops called
IPCAT (Information Processing in Cells and Tissues). The first IPCAT was held
CytoComputational Systems - Perspectives and Tools of Thought 7
in Liverpool in 1995 and since then we have had workshops every other year in
Sheffield (1997), Indianapolis (1999) and Leuven (2001). The fifth workshop was
held in Lausanne (2003) and the sixth is planned for York in 2005.
R. Tateson
Abstract. There are many examples of the natural world providing inspiration for
human engineers and designers. Cell biology is one branch of the natural sciences
which has not yet been widely exploited in this way, but which has great potential
for application, particularly in the telecommunications area. The features of cells
map strikingly well on to some of the challenges in current engineering and design
for telecommunications systems. The autonomy, evolution, adaptivity and self-
organisation of cells are alI desirable for the complex, dynamic and geographically
distributed networks we are now constructing and using. Three examples of cur-
rent research illustrate how analogies from cells can lead to radically different
telecommunications systems. Cell fate behaviour in fruitfly cells has inspired a
new, decentralised approach to managing mobile phone networks. Morphogenetic
events in early embryos point to new design methods which add depth to the es-
tablished, and also biologically inspired, techniques of evolutionary optimisation.
Genetic control pathways in bacteria inspire implicit leaming techniques which al-
low individual 'cells' in simulation to discover adaptive behaviour without an ex-
plicit definition of 'fitness'. All of the examples are at the research stage, and will
not be used in real networks until 2004 and beyond. However, they give a glimpse
of the strength of the analogy between biological cells and elements in telecom-
munications systems, and suggest that this will be a productive area of work into
the future.
1 Introduction
drum to invent a telephone transmitter and receiver which relied on the vibration
of a 'membrane' to turn sound to an electric al signal and back again.
Biological analogies are most powerful when we identify a biological system
which resembles in some sense an artificial system, and then use our knowledge
and understanding of the functioning of the biological system to improve or redes-
ign the artificial system. In recent years two areas of biology have proved fertile
for inspirations and analogies in the computation and telecommunications field.
Social insects, particularly ants, have been a rich source of ideas for telecom-
munications networks (reviewed in [1]). The most famous example is the analogy
between foraging behaviour of ants and routing of information packets. The ants
leave pheromone trails which can be followed by the other ants seeking food. On
finding food, ants retum to the nest, laying a trail as they go. In the case where
there are two (or more) alternative routes back to the nest, the shortest route tends
to accumulate pheromone fastest and becomes the preferred route. This is efficient
and hence desirable for the ants and is achieved without any 'bird's-eye view' or
other global map, and without any reasoning or central control. In a communica-
tions network, packets of information are routed from origin to destination by a
number of 'hops' from node to node in the network. Shorter routes are desirable
again for reasons of efficiency. The routing problem is usually solved by provid-
ing each node in the network with a list of destinations so that it 'knows' where to
send any newly arrived packet. The lists are caIculated centrally and disseminated
to the nodes. The ant-inspired alternative is to use the packets themselves to up-
date the lists. Just as the ants leave pheromone trails, the packets leave their influ-
ence on the lists, and tend to reinforce short routes at the expense of long ones.
The other significant area for biologically inspired telecommunications is evo-
lutionary computation. This is a large and active academic and applied field
which, despite the wide variety of specific approaches, shares as its underlying in-
spiration evolution by natural selection [2,3]. If the individuals of a population
produce offspring which inherit some of the parental characteristics, but with
some variation, and if the number of offspring produced by an individual is in
proportion to the 'fitness' of the individual, then as the generations go by the indi-
viduals in the population will tend to become 'fitter'.
In the natural world this is evolution as described by Darwin, and has generated
the vast wealth and range of life on Earth. The individuals are the organisms, be
they bacteria, elephants, redwoods or algae. If the individuals are instead solutions
to some artificial problem we can use the evolutionary process to find better solu-
tions. For example it is possible to represent the telephone wiring for a new hous-
ing development as an 'individual' which can then be given a 'fitness' according
to how much it will cost to build (cheaper is fitter!) and how well it meets the ex-
pected needs of the occupants. With a population of these solution individuals an
evolutionary process can get going. Initially perhaps most of the solutions will be
fairly poor, but some will be less poor than others, and these will have more 'chil-
dren' which resemble them, but with some variation. Over time the solutions im-
prove. There are many complicating factors which must be addressed to make this
kind of 'optimisation' effective, in particular to avoid getting 'stuck' with a popu-
lation of individuals which have improved on their parents but are now trapped in
Cells in Telecommunications 11
a 'local optimum' - a small hillock in the fitness landscape from which they can-
not escape to climb the 'global optimum' mountain.
Artificial evolution has been applied to many optimisation problems, and is
currently used as an effective and time-saving part of the planning process for new
local telephone networks [4].
What is it that leads us to believe cellular biology can also be genuinely useful
as a source of design ideas, rather than a mere post-hoc rationale for our engi-
neered system? The answer can be summed up in two words: 'distribution' and
'autonomy'. These are the key features which will facilitate huge advances in
computing and telecommunications, and these are the key attributes of cellular bi-
ology. If we take a look at the kind of problems which are becoming increasingly
pressing in the telecommunications field we can see that in many ways they match
the general features of cellular systems.
2 Telecommunication Problems
The extraordinarily rapid rise of the power and capability of computers and tele-
communications in the second half of the 20th century is founded on silicon semi-
conductor technology organised in the 'Von Neumann' architecture. Data and a
program are held in memory. According to the sequence of instructions in the pro-
gram, data are brought from the memory to the central processing unit, some logi-
cal function is performed and the results placed in memory. This way of process-
ing information is extremely well suited to general purpose, human engineered
computation. The series of logic al steps can be programmed by one person and
understood by another. Also the continuing advances in silicon chip design have
allowed the central processor, which carries the burden of getting through the (of-
ten very many) logical steps which constitute a program, to keep pace with the re-
quirement for computing power.
This principle of centrali sed information processing is adhered to, for reasons
of engineering tractability, in systems large (such as a mobile phone network) and
small (such as an individual personal computer).
However, there are weaknesses inherent in the conventional approach:
• A centralised design principle. Information must be transported to the central
processor, processed and removed.
• Serial processing. The data are proces sed one after another and the program is
interpreted sequentially.
• Brittle architecture. The hardware and software are not tolerant of errors and
imperfections. Small mistakes or impurities will often lead to system failure.
• Static injlexible solutions. A system or network is often tailored to an assumed
environment (e.g. assumptions about traffic demand in various parts of a net-
work); ifthat situation changes performance can de grade substantially.
The spectacular progress of conventional systems in the late 20th century
means that at the start of the 21st century we have software and hardware of a
12 R. Tateson
3 Features of Cells
Of course there is no such thing as a 'generic' cell. Each real example of a bio-
logical cell is specialised for its functional niche. However, it is possible to look at
cells at a sufficiently abstract le vei to identify some common features shared by ali
cells from E.coli to epithelial cells.
• Evolutionary history
• Division history
• 'Life history'
• Dynamic, metabolic
• Autonomous
• Emergent control
One thing which all natural cells share is an evolutionary lineage. Every one has
an unbroken sequence of ancestors stretching back 3.5 billion years. Needless to
say, evolution by natural selection is a central process for life on Earth and the
concept is essential for understanding contemporary cell biology. The power of
this evolutionary process as a means for achieving functioning systems without
explicit goal-driven design has been a source of inspiration for artificial systems
for several decades.
In addition to the implications of its long evolutionary history for its current form
and range of function, there is the effect of the more recent ancestry. That is: over
time scales which allow no significant evolutionary change there is an influence of
ancestors in terms of choosing which subset of the cell's potential functions is ac-
tually expressed. For example, in a multicellular organism the celilineage during
development has an important impact on the form and function of any individual
cell.
And the finest-grained 'historical' influence on the cell is the impact of events
which have impinged on the cell during its lifetime. Two cells with identical evo-
Cells in Telecommunications 13
lutionary and division lineages may behave in very different ways depending on
the signals they have received. Again a dear example can be taken from a mut-
licellular animal: the fruitfly Drosophila. All the cells in an individual animal
c1early share an evolutionary lineage. If we look at the peripheral nervous system
during development, we can find a single cell which is about to divide. The two
daughter cells c1early share a division lineage, yet one of them will go on to pro-
duce the bristIe which projects through the cutic1e to the 'outside world' while the
other will make the neuron which transmits the sensory information from that bris-
tIe to the central nervous system. There are very many documented examples of
this kind of decision-making process in unicellular and multicellular organisms.
Indeed it is possible to view the development of any organism as a combination of
division lineage and life history of cells.
How does a cell 'know' what it 'should' do next? What is the nature of its internal
state which is influenced by division lineage and life history, and which deter-
mines current cell behaviour and the range of possible reactions to stimuli? At an
instant in time we can in principle take a 'snap shot' of the gene expression state
of a cell. That is to say the degree to which each gene in the genome of the cell is
'turned on' or 'expressed'. We could also make an instantaneous measure of the
level of alI the chemicals of the cellular machinery - the structures, signals and
food of the cell. Together these things might define a cell's state. In practice, even
with modern genomic, proteomic and metabolomic techniques it is not possible to
perform a complete inventory of a cell. In practice people studying cell behaviour
will measure the levels of expression of a carefully selected subset of the cell's
genes, or will measure the concentrations of a subset of the cellular chemicals.
All of these things are dynarnic - the chemical concentrations and the levels of
gene expres sion have not reached a static stable point. In some cases the 'snap
shot' of the cell conceals the fact that the chemical or gene expression being
measured is changing rapidly. In other cases there is a stable level which is main-
tained for a significant time (seconds to years) but this is a dynarnic equilibrium-
chemicals are being broken down and built up. This is the essence of metabolism:
a balance between the catabolic (breaking things down) and the anabolic (building
things up). This 'churning' of cellular components often strikes the human eye as
wasteful but because it gives the cell excellent responsiveness to change, and al-
lows (or indeed enforces) the pooling of cellular resources, it makes an extremely
effective strategy in a world where the raw materials and energy are out there for
whoever can get them.
3.5 Autonomous
The control of cell behaviour is 'emergent' - that is it arises from the interactions
of the many small parts of the ceH to give an outcome which can subsequently be
observed and described at a higher level. This is true over 'design time' (Le. evo-
lutionary time) and 'operational time' (i.e. the life of an individual ceH).
For example we may observe the behaviour of E. coli as it alters its metabolism
to use lactose instead of glucose as its energy source. We can understand the mo-
lecular interactions in terms of their higher level behaviour as a switch, and we
can see the logic of that switch in terms of its action and its benefit to the bacte-
rium. Analysis is made possible by treating the control pathway in isolation from
'irrelevant' context, such as the activity of Other control pathways in the same ceH.
None of these perspectives is available to the bacterium - the pathway has
evolved, and now functions, without any abstraction to higher level function, and
without the ability to ignore its cellular context.
Developing cells have many different types of signal and receptor molecules
which allow them to communicate with neighbouring cells. A given type of signal
molecule will bind specifically to one, or a few, types of receptor. As a conse-
quence of signal binding to receptor, a message is usually passed into the cell on
whose surface the receptor is carried.
One example of such a pair of signal and receptor is Delta (the signal) and
Notch (the receptor). Versions of these molecules are found in many different
animals, from nematodes (c. elegans) to humans (H. sapiens) [6]. They are used
for communication between cells at many stages in development. For example, in
the fruitfly Drosophila, they are needed for correct nervous system formation in
the early embryo and for developing the correct wing shape during pupation.
Fig. 1. The mutual inhibition process. Concentrating on just two adjacent cells, we can see
the 'flip-flop' switch resulting from the molecular interactions. The red arrow represents the
Delta signal. The Notch receptor is shown in green. An inhibitory signal (blue) reduces the
amount of Delta produced in a cell whose Notch receptors have been activated by its neigh-
bours. In this example, panel A shows the initial state in which both cells produce roughly
equal amounts of inhibitory Delta. B the upper cell begins to gain the upper hand. C The
upper cell has decided to make a bristle, and forced its neighbour not to do so
Although there are other molecules which affect the communication between
Delta and Notch, the core of their interaction is simple (Fig. 1). The binding of
Delta to Notch causes a protein to enter the nucIeus of the Notch-carrying cell and
alter the state of expression of the DNA. The effect of the alteration is that the
cell's production of Delta is reduced. Thus its ability to send a signallike the one
it has just received is diminished. Production of Delta is linked to a choice of cell
type. In other words the cells having this conversation are 'contemplating' becom-
ing one sort of cell (for example neural), which correlates with Delta production,
rather than another (for example epidermal). The exact nature of the choice de-
pends on the particular developmental process at hand; the Delta and Notch mole-
cules are not concerned with that, just with ensuring that a choice is made. So, if a
cell perceives a Delta signal from another cell, it makes less Delta itself and hence
becomes both less able to signal Delta to its neighbours, and less likely to choose
the Delta-correlated cell type.
16 R. Tateson
4.1.2 Applicability
The method inspired by development differs markedly from current practice in
frequency allocation. Currently the network operators use centrali sed planning: at
regular intervals a planning committee meets to decide on a new frequency plan
(according to network performance measures since the last such meeting). This
plan is then disseminated to the base stations, and will remain in place until the
next meeting. The current method works best when networks are relatively smaIl
and there is little fluctuation in traffic. The larger the network, and the greater the
fluctuations, the harder it becomes for a centrali sed solution strategy to continue to
de1iver acceptable quaIity of service.
Cells in Telecommunications 17
Fig. 2. A The mutual inhibition process which Iimits the number of cells adopting the neu-
ral 'bristle-making' fate. Many ha ve the potential (grey) but only a few reali se that potential
(black) while inhibiting their neighbours (white). B The pattern of bristIes that results in the
adult fly. CA map of East Anglia in the UK, with the coverage areas of base stations in a
mobile phone network coloured in as if there were only four frequencies to allocate (blue,
red, green, yellow)
In effect the efficiency with which the resource (frequencies) is being used de-
clines because that resource cannot be moved meet demand. As a result the opera-
tor must invest in constructing new base stations (Le. buying more resources) to
maintain quality of service. It is in exactIy these circumstances of dynamic traffic
in a large network that the method inspired by development is likely to deliver
tangible benefits. This work is now being developed under contract to QinetiQ for
the UK Ministry of Defence. It is hoped that it can be usefully applied to manag-
ing frequencies for military radio communications. Battlefield communications
are an excellent example of the kind of network in which this approach delivers
the greatest advantage over centralised control methods. They are complex and
highly dynamic, with even the base stations moving around from minute to min-
ute. In addition the Iines of communication must continue to operate effectively
even when many base stations are no Ion ger functioning.
4.2.1 Simulation
A DBM individual consists of a single celI, with a genome, which is placed at the
centre of a 3D space and alIowed to divide over time to give a cluster of celIs.
Each cell has a small range of behaviours available to it. The genome determines
both the nature of the chemicals which the cell can produce, and the way in which
the cell responds to the chemicals in its immediate environment. During the life-
time of one individual, the chemicals produced by the original cell and all its
daughters can diffuse and react within the 3D space. Hence the individual is con-
stantly altering and responding to its chemical environment. There are several
such individuals in a population. Individuals are evaluated according to a fitness
function based on the positioning of the celIs of the cluster. Fitter individuals are
used as parents to produce genomes for the next generation.
4.2.2 Results
To date DBM as been applied to generating the shapes and differing cell identities
of early embryogenesis. In other words the design requirements are also derived
from developmental biology, rather than an analogous engineering problem. The
authors were interested in generating shapes resembling the blastocyst and seg-
mentation along an axis.
The blastocyst is a holIow ball of celIs produced at a very early stage of verte-
brate development. The infoldings of the blastocyst subsequently produce layers
of cells which differentiate to give the tissue types of the developing animal. It is
challenging to produce a holIow baII from a simulation of ceH division because the
Cells in Telecommunications 19
4.2.3 Applicability
DBM is certainly at an early stage but it has the potential to be very useful as an
automated design tool, most obviously in the telecommunications domain. The
networked nature of many telecommunications problems, ranging from the design
of physical networks to the services provided by those networks, could be weIl
served by the division-based process of DBM. It would be easy, for example, to
imagine the DBM 'cells' as nodes in a network which assume their final position
according to the influences of their neighbours, or even dynamically alter their be-
20 R. Tateson
haviour based on the 'chemical' environment around them. Looking further to the
future, it is possible to envision DBM-type ideas being used to design the autono-
mous software agents which can move across networks providing services to their
'owners'. Adaptivity and specialisation without explicit instructions from the
owner might make such agents far more useful. In summary, DBM has shown that
biological-style morphogenesis can be simulated in a way which preserves the
useful design power of that process and we expect the telecommunications prob-
lems of the future to be of the type which will benefit from this approach.
Fig. 4. 11- and 25-cell individuals evolved for segmentation. In these examples the indi-
viduals have divided into two c\usters. The images are oriented so that the axis along which
they have 'segmented' runs horizontally across the page
4.3 CeliSim
Genome
/ \
10010111]
Fig. 5. The genome strueture of the simulated eel!s. Eaeh cel! has a 400-bit genome. This
genome is divided into ten genes, eaeh with 40-bits. Eaeh gene is divided into five domains,
eaeh with 8-bits. These domains are ealled 'more', 'less', fune', 'modi' and 'act'
4.3.2 Simulation
CellSim simulates the gene expression-based control of cellular behaviour
(Fig. 5). It does not simulate internal cell structure or any limitations on diffusion,
so any 'molecule' can instantaneously interact with any other. CellSim also does
not simulate molecular structure: alI interactions between simulated molecules are
based on 'soft' complementary matching between bit strings (a '1' in string A will
bind to a 'O' at the corresponding position in string B, but 'soft' matching means
22 R. Tateson
that there will be non-zero binding affinity between two strings even if the com-
plementary match is not perfect at alI positions).
The 'more' and 'less' domains determine the 'expression level' of the gene, i.e.
the rate of production from that gene. The 'func' domain dictates the nature of the
product, specifying whether the expressed string will be released from the cell into
the environment, retained within the cell or tethered in the membrane of the cell,
with either its 'modi' or 'act' domain exposed to the external environment. 'func'
also specifies the qualitative effect of binding by another string to the 'modi' do-
main of this string: it can result in an increase or a decrease in the effective activ-
ity of the 'act' domain. 'modi' and 'act' specify the sequence of the product.
The strings already present in the intracellular environment control the expres-
sion levels of the genes by binding to the 'more' and 'less' domains (Fig.6). The
'occupancy' of these domains is the sum of the binding affinities of all existing in-
tracellular strings to the domain. The greater the 'occupancy' of the 'more' do-
main, the greater the expression level. Conversely, the greater the occupancy of
the 'less' domain, the lower the expression level.
Genome
nisali act
Soup Strings nisali act
In addition to binding to the gene domains, the intracellular strings can bind to
each other. The 'act' sequence can bind to the 'modi' sequence of other strings,
and hence alter the effective activity of the string, increasing or decreasing it as
specified by 'func'.
Thus a feedback control network is established within the ceH (Fig. 7), with the
pool of intracellular strings created by, and acting on, the genome. This internal
'genetic regulatory network' is linked to the external environment in two direc-
tions. Firstly strings in the environment can affect the cell by crossing the mem-
brane to bind intraceHular 'modi' domains or by binding to externally facing
'modi' domains. Secondly the ceH may alter the environment by exporting strings
and by exhibiting behaviours such as movement.
Cells in Telecommunications 23
Act domain
Stri
Fig. 7. A Cell in a CellSim world. The long dark strand represents the genome, from which
new strings are synthesized, under the regulation of strings already present in the ceH.
Strings synthesized from the genome have two 8-bit domains, 'modi' (shown dark) and
'act' (shown Iight). The 'fune' domain of a gene determines both the effect on 'act' if an-
other string binds 'modi', and the position of the string in the eell (free in the ceH, anchored
in the membrane with 'modi' ar 'act' exposed to the exterior, ar secreted from the ceH [not
shown])
divided. Thus cells with some residual chemotactic ability were implicitly re-
warded because they would accumulate food (on average) faster than their com-
petitors. Over several generations chemotactic behaviour was 're-discovered' by
the cells, and as one would expect this behaviour became fixed in the population.
~MOd; Aci
Genc 3 111111 11
4.3.4 App/icability
CellSim is complementary to Design by Morphogenesis (above), concentrating on
the issue of gene expression, while largely ignoring the physical reality of chemi-
cal diffusion. It shares one possible area of application with DBM: software agents
Cells in Telecommunications 25
5 Conclusion
There is great potential for celIular biology to inspire and influence design, engi-
neering and control of complex systems. Telecommunications is a prime area for
the early realisation of this potential because it combines complexity and compu-
tation to give systems which are physically distributed but must respond dynami-
calIy and appropriately to user behaviour. The ability of telecommunications sys-
tems to seamlessly accommodate growth and 'heaI' when damaged is also very
desirable and is well addressed by celI-based approaches.
However, it is not yet possible to point to an example of this thinking which has
been implemented in a real system and demonstrated its effectiveness as a day to
day piece of a functioning telecommunications system. The examples used in this
chapter are alI at the research stage. The fruitfly example is the closest to reali sa-
tion and could conceivably be put into practice within 3 years. The DBM and
CelISim examples are both further from direct application, and currently address
more abstract and general aspects of telecommunications problems. It is most
likely that the cellular thinking in these IaUer two examples will be incorporated
into mainstream telecommunications by offering a new dimension to established
evolutionary computation approaches. In a sense both offer new ways of address-
ing the old problem of escaping local hilIs to climb the optimality mountain.
References
3 Băck, T., Hammel, U., and Schwefel, H-P., 1997. Evolutionary computation:
comments on the history and current state, IEEE Transactions on Evolution-
ary Computation, Val. 1, No. 1
4 Assumu, D. and Mellis, J., GenOSys-automated tool for planning copper
greenfield networks, 1999. British Telecommunications Engineering, vo1.17,
pt.4, pp 281-90
5 Tateson, R., Shackleton, M., Marrow, P., Bonsma, E., Proctor, G., Winter, C.
and Nwana, H., 2000. Nature-inspired computation: towards novel and radi-
cal computting, BT Technology Journal (Millennium Edition) l8... No. 1, pp
73 - 75 plus CD content
6 Artavanis T., S., Matsuno, K. and Fortini, M. E., 1995. Notch signalling. Sci-
ence voI 268 pp 225-232
7 Tateson, R., 1998. Self-organising pattern formation: fmit flies and ceH
phones in A. E. Eiben, T. Back, M. Schoenauer and H-P. Schwefel Proceed-
ings of 5th Int. Conf PPSN, Springer, Berlin. Heidelberg New York, pp 732 -
741
8 Hoile, C. and Tateson, R., 2000. Design by morphogenesis, BT Technology
Journal ~ No. 4, pp 85 - 94
9 Eggenberger, P., Evolving morphologies of simulated 3D organisms based on
differential gene expres sion, 1997. In Proceedings of Fourth European Con-
ference on Artificial Life. pp 205-213. P. Husbands and 1. Harvey (editors),
MIT Press, Cambridge, Mass.
10 Paton, R. (ed.), Computing with biologic al metaphors, 1994. Chapman and
HaU, London
11 Vogel, S., Cats' paws and catapults, 1999. Penguin, London
Symbiogenesis as a Machine Learning
Mechanism
L. Bull, A. Tomlinson
1 Introduction
Symbioses are commonplace in the natural world and it is therefore argued that
the phenomenon is of great evolutionary significance (e.g. [1 D. Symbiogenesis is
the hypothesis that, if the relationship between symbionts evolves in the direction
of increasing dependency, potentially 'a new formation at the level of the organ-
ism arises - a complex form having the attributes of an integrated morphophysi-
ological entity' [2, p.Sl. In a more restricted sense symbiogenesis refers to the
formation of hereditry endosymbioses, the symbiotic association in which partners
exist within a host partner and the relationship is maintained in offspring. Perhaps
the most remarkable example of symbiogenesis may have occured during the evo-
lution of eukaryotes whereby free-living bacteria appear to have become the intra-
cellular organelles (see [2] for an account of the history of this concept). Maynard-
Smith and Szathmary [3] have a1so argued that symbiogenesis gave rise to chro-
mosomes during the evolution of cellular structures.
28 L. Bull, A. Tomlinson
2 Simulated Symbiogenesis
Kauffman and Johnsen [4] introduced the NKCS model to allow the genetics-
based study of various aspects of coevolution. In their model an individual is rep-
resented by a genome of N (binary) genes, each of which depends epistatically
upon K other genes in its genome. Thus increasing K, with respect to N, increases
the epistatic linkage, increasing the ruggedness of the fitness landscapes by in-
creasing the number of fitness peaks, which increases the steepness of the sides of
fitness peaks and decreases their typical heights. 'This decrease reflects conflict-
ing constraints which arise when epistatic linkages increase' [4 p.330]. Each gene
is also said to depend upon C traits in each of the other species S with which it in-
teracts. The adaptive moves by one species may deform the fitness landscape(s) of
Symbiogenesis as a Machine Leaming Mechanism 29
In this paper there are three species (A, B, and E), two of which (A and B) are said
to be in a beneficial symbiotic relationship and have the potential to form an he-
reditary endosymbiosis. The third species (E) is said to represent the other species
with which the symbionts coevolve - their ecologically coupled partners. For A
and B there are three sub-populations (Fig. 1). There are the two sub-populations
for the symbionts living cooperatively and evo1ving separately, each receiving
there own separate measure of fitness, with these being initially set to size P (the
sizes of these two populations are always the same as each other). The other sub-
population is of hereditary endosymbionts, consisting of individuals carrying the
genomes of both symbionts, where they are treated as 'an interspecies supraorgan-
ism' [11] for selection by receiving a combined fitness. The size of this sub-
population is initially set to zero, but during the course of a simulation individuals
can move from the other sub-populations to this, and back, via a megamutation
operator (hence there are always effectively P sets of the two genomes overall).
Over evolutionary time the most appropriate configuration of the two symbionts
of the model - be it an endosymbiosis or to stay as two separate populations of
cooperating species - will emerge rather than being prescribed a priori; the he-
30 L. Bull, A. Tomlinson
A B E A B E
o o
p p A&B
Fig. 1. During evolutionary time the population space of the symbiotic species A and B can
be invaded by their equivalent endosymbiotic association (A&B)
Initially the two symbiotic species exist in separate populations, both of size P,
where they are paired simply by taking the corresponding members from their re-
spective populations (that is speciesA[x] is paired with speciesB[x], where
O.sx<P). A corresponding member of the environmental third species (speci-
esE[x]) is also chosen. Individuals are then evaluated on their given NKCS func-
tion and awarded their own separate measure of fitness. A generational GA is
then TUn on the total population space of the two symbionts. At the end of this
process, and each succeeding evaluation-selection-generation of individuals, a
check is made to see if the symbionts formJdissolve a more intimate association.
The formation of a general symbiosis in nature is termed recognition and 'may in-
clude specific chemi cal interactions, tolerance or suppresion of host defences, and
metabolic, morphological or behavioural interactions' [12 p.S]. For our purposes
the formation of an endosymbiosis is established by simply testing a probability
(megamutation) P" Satisfaction of this probability means both partners form an
hereditary endosymbiosis, where they receive a combined measure of fitness for
selection and where an individual consists of two genomes - the specie sA and
speciesB members. From then on they do not evolve within seperate populations,
but are treated as one by the GA , having been joined via this megamutation op-
erator. The same operator is also applied to the population of hereditary endosym-
bionts, working in reverse, to stop the drift of members away from the population
of separate symbionts; separating is just as likely as joining. Evaluation of the two
endosymbiotic sets of genes is the same as when they were separate; the specie sA
genome is evaluated using the NKCS fitness tables for specie sA and the speciesB
genome with the speciesB tables.
The size of these competing symbiotic sub-populations (endo and separate) is
controlled by a version of the roulette wheel proportional selection used in many
genetic algorithms. Here the total fitness ratings of aII individuals in aII symbiotic
sub-populations are summed at the end of each generation. Then a random number
is chosen between zero and this total fitness. Starting with any given sub-
Symbiogenesis as a Machine Leaming Mechanism 31
population the scores of alI individuals are again summed until this value is equal
to or greater than the random number chosen. The following production of an off-
spring uses gene mutation on the selected single parent. This whole process is re-
peated until P new pairs of individuals have been created, from whichever sub-
populations. It may have been noted that the cooperating separate species receive
individual fitnesses, meaning they are on average half as likely to be chosen com-
pared with the endosymbionts (supraorganisms) through the method of selection.
Therefore we say that the separate individuals' sub-population appears as one,
with the combined fitnesses of speciesA and of speciesB during the summing
processes. Whenever this sub-population is chosen a pair of separate symbionts
are always created, proportionalIy, from their own populations. The environmental
species (E) also evolves asexually within its own population via a generational
GA, using proportional selection and allele mutation. This population allows us to
examine the effect endosymbioses have on their other ecological partners. The
process is repeated for a given number of evolutionary generations.
We now use the model described above to examine the evolutionary perform-
ance of endosymbiosis, with alI results presented being averaged performance
over 20 runs.
2.3 Results
We implement our version of Kauffman and Johnsen's model using three species,
each evolving separately in populations of 100 (P=100) individuals. Three ge-
nome lengths are used throughout, N=8, 12 and 24 - alI experiments are repeated
for the various N. We introduce a second value of interdependence for the symbi-
onts to the third species (Ce); we can alter both the interdependence between the
symbionts (C) and the interdependence both symbionts have with their environ-
ment (Ce). This rate of environmental interdependence is the same for the symbi-
onts whether or not they are joined. The per bit allele mutation rate of the GA (p,,,)
is set at 0.001 and the megamutation operator (p,) is set at 1/(5P) - that is on aver-
age two sub-population members will alter their association once in five genera-
tions (it is suggested that the probability is comparitively low in nature with re-
spect to the gene allele mutations). AlI species have the same K value.
A number of experiments were carried out in which the values of K, C and Ce
were varied. We find that for various combinations of these values different sub-
populations become dominant within the model.
For Ce=1 (Le. low) and any K we find that increasing C increases the popula-
tion percentage of the hereditary endosymbionts, such that when CSK there are
more endosymbionts than separate cooperating symbionts. The greater C is in re-
lation to K, the bigger the difference in population percentage (Figs. 2 and 3).
Conversely when K>C the cooperating separate symbiotic species are dominant
(i.e. >50%), though they never dominate as much as the endosymbionts do. This
shows that, as the degree of dependency increases between symbionts, forming an
hereditary endosymbiosis becomes increasingly beneficial to them. Also note from
Figs. 2 and 3 that the difference between K and C must be large for either associa-
tion to become significantly dominant (>80%). Kauffman and Johnsen [4] state
32 L. Bull, A. Tomlinson
that K=C corresponds to the coevolving partners' optimum edge of chaos regime
- their liquid state. It can be seen from our results that in these models the edge of
chaos has quite a large basin of attraction towards coexistence of both associa-
tions; conditions must stray considerably before the population will diverge sig-
nificantly either way.
al.O
ICI,O
G--i> ......
::~~~I ~
O.OM
lOOO.O
--
i&
EO.O
<0.0
d°o.6:0-~lOOO::c!:-:O~--::«lOO~J>~-IOOI)~,o:--'--::e:c!·
ooo.o
90'-
Fig. 2. Graphs showing that as organisms' attributes K and Care varied the ratio of separate
symbionts to endosymbionts in a population varies. When K > C the separate symbionts
remain dominant. lncreasing C with respect to K increases the population percentage held
by the endosymbionts
Increasing Ce does not appear to alter the general result of greater C giving rise
to more endosymbionts, although from Figs. 3 and 4 it can be seen that an increase
in Ce can allow a higher percentage of endosymbionts to exist for lower C than
before. Conversely for higher C a lower percentage of endosymbionts are found
when Ce is high than when it is low. That is, if the symbionts become increas-
ingly dependent upon other aspects within their environment, forming an endo-
symbiosis is beneficial only up to some point. As this dependence increases (i.e.
when Ce is high), the benefits of joining appear to decrease; if the symbionts' or
endosymbionts' dependence upon other species within their environment in-
creases, the benefits of either association are decreased and the system retums to
Symbiogenesis as a Machine Leaming Mechanism 33
its edge of chaos-like constitution. Burns' [13] suggestion that some endosymbi-
onts no longer experience environmental factors may be a way in which nature has
dealt with this problem, i.e. the effects of Ce are 108t on endosymbionts.
2.4 Discussion
From these simulations it can be seen that as the interdependence (C) between
symbionts increases, hereditary endosymbiosis becomes increasingly dominant
within the overall population; the further the partners are into their chaotic gas re-
gimes, the more likely they are to form an endosymbiosis.
K
3 5 7
I
3 3
C
5 5
C
7 7
9 K
II N=8 Ce=1
N=12Ce=1
135 7
48 41 42 30
3 3 164 59 51 3
C C C 5
5 5 74 63
7 7 71 7
K K K
N=8 Cc=3 N=8 Ce=5 N=8 Ce=7
Fig. 3. Showing the effects of variying K, C and Ce on the population percentage held by
the endosymbiotic association of a given symbiosis
34 L. Bull, A. Tomlinson
IOO.Or;--..-----..---T1i"-..-----,
IIllO 00.0
'-" :: (,0.0
" ro.O .~
"i$ ;;
].
]. ~ -10.0
i5. 40.0
20.0
0 .0 _-~----''--~----I.----'---'
0.0 2000.0 -llXll.O ftXlO.O
;g~ncm'illn
Fig. 4. Showing that as the amount of interdependence with the environment (Ce) is in-
creased, the difference in the two associations' performance decreases
Kauffman and Johnsen [4] investigated the effects interdependence has upon
separate coevolving partners, finding that:
4A: When the partners are least interdependent (C is low) the time taken for the
system as a whole to reach a Nash equilibrium is low, fitness during the period of
pre-Nash oscillations is c10se to the equilibrium fitness and the oscillating fitness
is high.
4B: When the degree of interdependence is increased, eventuallY to being obli-
gate (C=N), they find the reverse is true with an in crease in the time taken for
equilibrium to be reached, with fitness during the oscillations being below the
equilibrium fitness and where this oscillating fitness is low.
The results are attributed to the fact that, as the amount of interdependence is
increased, the effect of each partner on the others' fitness landscape increases; the
higher C, the more landscape movement. This represents an extension to Kauff-
man' s single organism NK model, in which he found that:
4C: Increasing epistasis (K) with respect to N increases the ruggedness of a fit-
ness landscape, but where the height of the optima decreases.
We can use Kauffman's findings to suggest underlying reasons for our result
above, that as interdependence increases between symbionts the relative perform-
ance of an endosymbiotic association increases. Hereditary endosymbionts repre-
sent two or more genomes being carried together effectively as one genome and as
a consequence the partners' inter-genome epistasis becomes intra-genome epista-
sis for a larger genome. The effects of intra-genome epistasis are very different
from inter-genome epistasis, as can be seen from 4A-4C above. That is, the com-
bined landscape of an hereditary endosymbiosis does not move due to the symbi-
onts' interdependence C - but is more rugged, and larger, than that of the individ-
ual partners living separately (4C). There will still be movement from any
environmental interdependences of course, but now their internal epistasis has in-
creased and they have (potentially, depending on Ce) moved into their combined
form's solid state (K>C). This can be seen as a genetics-Ievel analogy of Burns'
Symbiogenesis as a Machine Leaming Mechanism 35
their interdepence are lost and they become a supraorganism with high epistasis. If
this new epistasis is large with respect to the new genome length (K>NI2) it is
more efficient in the long run for the symbionts to stay separated. That is, for
highly dependent symbionts (i.e. those in their chaotic gas regimes), it is better to
stay separated if joining via endosymbiosis will make them a supraorganism we1\
into its solid regime. We would expect to find the same kind of dynamics occur-
ring within our population-based GA models if they were left to run for much
longer (or if we enforce niching or alter the selection pressure). Our results cer-
tainly match those using Kauffman and Johnsen's model over a shorter time scale
or when we fix the existence of both associations .
•.
~ 1200.0
'"
IIXX).!) L-~--.JL-~--.J_~--L_~--'
0.0 1000.0 4OOO.U (i(X'Xl.O WOO.C
Fig. 5. Showing that a limit ex.ists on how far into their respective gas regimes symbionts
can be fore endosymbioses to represent the optimal strategy (results using Kauffman and
Johnsen's basic model)
lkegami and Kaneko [14] were perhaps the first to use inherited genetic linkage in
an artificial system. They introduced a 'genetic fusion' operator to link genomes
within a population using a GA to produce strategies for a simple game, showing
that more successfullarger genomes could emerge over time. Goldberg et al. [15]
have applied the concept at the gene level whereby contiguous genes link to avoid
disruption under crossover, akin to the aforementioned process by which genomes
may have emerged [3]. We now describe the use of symbiogenesis within a sim-
ple machine leaming architecture; the process in inc1uded within an inductive
framework to improve its performance.
Wilson and Goldberg [17] were the first to suggest that rule-linkage may help
LCSs form complex rule structures, what they termed rule corporations. Accord-
ing to Goldberg and Wilson, the rule-base of a 'corporate classifier system' (CCS)
would contain not only single rules, but also clusters of rules. These corporations
would only be reproduced or deleted as a unit, hence synchronisation is assumed,
and formed by a mutation type operator. For reproduction, the fitness of a corpora-
tion would be dependent upon the fitness of its members, possibly the average
strength, such that it would be advantageous for rules to link together rather than
remain single. If average fitness was used to determine the fitness of a corporation
then this may be sufficient to encourage corporate linkage. Given the results in
Sec. 2, the increased stability in evaluation environment can be expected to pro-
mote linkage between highly interdependent rules.
Holland [18] has presented the Echo system, an artificial ecosystem simulation
in which agents move from site to site, interacting with each other and local re-
sources. The agents in the system are given the ability to increase in size and com-
plexity as individual agents join 'complex aggregates' or merge together to form
'macro-agents'. The proposed approach to this is to give each simple agent a long
chain of 'tag' strings in addition to its basic chromosome. Some of these strings
will remain dormant in the single agent and will only be activated if the agent
joins to form a higher level structure (triggered by a pattern matching procedure).
This suggests one possible approach to the implementation of corporations within
a classifier system based on the idea of dormant linkage templates.
In ZCS, a ruIe consists of a condition, an action, and also a strength value. In
this work an implementation of a corporate classifier system has been facilitated
by adding a few more parameters.
If corporations are viewed as chains of rules, then a rule can at most be directly
linked to only two other rules. If this approach is taken then each rule will require
two link parameters ('link forward' and 'link back') that when active reference
other rules within a corporation. These links will be initialised as inactive but
when two rules are selected for joining, then one of each rule's links ('link for-
ward' for one rule, 'link back' for the other) will be set to reference the other rule.
This concept of rule-linkage is analogous to the gene-linkage employed by the
afore-mentioned work in GAs. Here, linkage between genes is used to encourage
the formation and propagation of good building blocks in the population of a GA.
In a corporate classifier system rule-linkage is used to encourage associations be-
tween rules through the formation of inter-dependent rule chains.
In addition to this each rule also contains a 'corporate size' parameter and a
'corporate id.' parameter included to facilitate subsequent processing. InitialIy size
is set to 1 and corporate ido is left inactive. Within corporations, alI rules will hold
the same values for size and corporate id, and these are set during the formation of
the corporation, either through 'corporate joining' or through the action of cross-
over by the GA. The classifier system keeps a record of how many corporations
have been formed and this is used to determine the ido reference for each new cor-
poration.
Symbiogenesis as a Machine Leaming Mechanism 39
corporation X, and likewise copies of rules 6 and 7 from corporation Y. The copy
of rule 1 is called rule 1', and those of rules 6 and 7 are called rules 6' and 7' re-
spectively. Corporation Z produced by this corporate crossover operation contains
the foHowing rules: [r1', r8, r6', r7']. In this way the offspring rule, rule 8, is linked
back to the facsimile of rule 1 (rule 1') and linked forward to the facsimile of rule
6 (rule 6').
Crossover point
Corp' X Parent I
Corp' Y 4 7 Parent 2
Each additional rule that is reproduced by crossover donates half of its strength
to its offspring as above for reproduction without crossover. Mutation is applied
only to the new rule derived from crossover (i.e. rule 8 in the example).
The basic ZCS model was modified to act as a prototype corporate classifier
system (ZCCS). Modifications were implemented as described in the last section
and alI other system parameters were maintained as in Wilson's original experi-
ments.
ZCCS was tested in the same environment as Wilson's original ZCS experi-
ment, W oods 1. A record was kept of system performance for each trial and also
the mean number of corporations active during each tria!.
3.3 Woods 1
* F -Food
0- Tree
O O F
* -Animat
O O O
O O O
Fig. 7. Woods I
On each time step the animat receives a message from the environment which
describes the surrounding eight ceHs. The message is encoded as a 16-bit binary
string with two bits representing each of the eight ceHs. A blank ceH is represented
by 00, food (F) by 11 and trees (O) by 10 (01 has no meaning). The message is or-
dered with the ceH direct1y above the animat represented by the first bit-pair, and
then proceeding clockwise around the animat.
The trial is repeated 10,000 times and a record is kept of a moving average
(over the previous 50 trials) of how many steps it takes for the animat to move into
a food ceH on each tria\. If the animat moved randomly then its performance
would balance out to about 27 steps per tria\. Optimum performance in Woods 1 is
1.7 steps.
In this initial test there is no discemible difference between the performance of
ZCCS and ZCS (Fig. 8 ). The number of corporations in ZCCS rose from O to 40
in 100 trials and then c\imbed slowly to 80 by the end of the run.
Hence this experiment has demonstrated that it is possible to implement a cor-
porate classifier system as proposed by Wilson and Goldberg. The corporate clas-
sifier system used for the experiment can be considered merely a template design
kept as minimal as possible using the simplistic symbiogenesis process explored
in Sec.2. However, as noted in Sec. 1, there are many ways in which system de-
sign could be expanded or modified to achieve more directed gains drawing
closely on the natural phenomenon. Some of these are now considered.
Maynard-Smith and Szathmary [1] note that during the emergence of 'proto-ceHs'
genetic linkage between 'naked' replicators is not sufficient for more complex
evolutionary structures to emerge. That is, a linked set of cooperating entities at a
given physical location are still open to exploitation from neighbours which are
not linked to them and hence do not necessarily share in their evolutionary future.
Hence, whilst pas si ve localisation can initiate cooperative relationships, only
through the active formation of a protective membrane can a symbiogenetic entity
perpetuate itself effectively. We now consider this within our corporate classifier
system by first including an analogue to spatial location and then a membrane to
exclude cheats.
42 L. Bull, A. Tomlinson
10 r------------,
8
~ zCS
cr-t'JZCCS
O L-----------~
O 2000 4000 6000 8000 10000
Frial.
Fig. 8. ZCCS v ZCS in Woods I
Whilst the position of the rules within the rule-base of an LCS is not significant
to their use, they do have logical relationships with each other and so corporations
are now encouraged to encapsulate chains of inference. Corporate links here take
on a temporal connotation and imply that the rules within a corporation are placed
so as to fire in succession and thus to map a proposed plan of action during the so-
lution of a multiple time step problem.
This is achieved by making linkage a niche operation, or more precisely a
cross-niche operation. Coupling occurs between subsequent match sets. This
means that on time step t there is a possibility that a rule that matches the current
message from the environment may link to a rule which matched the stimulus at
time tol . This encourages the structuring of meaningful sequences of rules.
To be selected for coupling, a rule must be in the current match set (termed [M])
and its appropriate link must be inactive. Coupling occurs over two time-steps. On
the first, a mie in [M] is selected probablistically (roulette wheel, based on
strength) from those with an inactive 'link-forward' , on the second, a rule in the
new match set is selected again probablistically from those with an inactive 'link-
back'. Rules already in a corporation are not allowed to join to rules within their
own corporation. In ali environments used during testing, the system was reset af-
ter receipt of a reward from the environment on some time step. Corporate Iinks
are not allowed to form between this reward time step and the first one of the fol-
lowing trial as this 'move' does not represent any form of causal transition under
the control of the system.
To further maintain temporal integrity amongst corporate rule-strings the GA is
adjusted to operate within match sets. The idea of a niche GA operating in the
match set was suggested by Booker [20] to introduce mating restrictions and thus
to assist the GA in producing more meaningful offspring as like breeds with like.
In CCS the adjustment is made so that if corporations are selected for crossover
then the resultant corporation should still represent a meaningful series of re-
sponses to experienced stimuli.
Symbiogenesis as a Machine Leaming Mechanism 43
The modified CCS model was initially tested in the previously used environment,
Woods 1. Fig. 9 show graphs of the average steps taken to reach food over ten
runs. The system performed well, reaching an average of about 2.2 steps to food
over 10,000 runs. The optimum perforrnance in Woods 1 is 1.7, and ZCS achieved
an average of about 3 steps to food. In these tests the ZCS GA was modified to
operate in the match set and the rule replacement rate was increased to one rule
per time step on average, to facilitate a fair comparison. The modified ZCS
achieved a rate of 2.6 steps to food; in this simple Markov environment CCS can
be seen to provide minimal benefits.
An alternative test environment is now presented which allows for a clearer dif-
ferentiation between competing systems' capabilities. The new test is a simple
variable multi-step environment. On each of N time steps the system is presented
with a stimulus and must select one of A actions, where A is a variable integer
value which defines the breadth of a maze. N is the number of states or nodes to a
reward and thus defines the maze depth. After N steps, the system receives a re-
ward from the environment and a new task then begins. The size of the reward de-
pends on which route the system chooses and so over time the system leams the
optimum reward yielding route through the maze. There is, however, more than
one maze. There can be up to Mz different mazes. The system is informed which
particular maze it is being presented with only on the first time step of each trial.
On all subsequent steps the stimulus is representative only of the current time step
in the trial. Hence these tasks falI into the non-Markov category. The maze is se-
lected randomly at the start of each trial.
Symbiogenesis as a Machine Learning Mechanism 45
10 r-------------------------------,
8 --z S
G---E) S
6
"O
o
<2
8
'"c.. 4
2:l
(/J
Tri al
Fig. 9. CCS performance in Woods l
Fig. 10 illustrates a simple task of this type with A set to 2, N set to 2 and Mz
set to 2. The environmental stimulus at each time step is also included. In this ex-
ample, a reward of 1,000 is awarded for one route on each map. AII other routes
receive a reward of O.
In the example in Fig. 10 the message length L is set to 3. With L set to 3 there
are eight possible stimuli and so for a two-map problem the maximum depth will
be 7, as the first time step (ts,,) takes two stimuli. Similarly with L set to 3 a four-
maze problem may have a maximum depth of 5, etc.
Clearly ZCS will be unable to master more than a single map due to the sensory
ambiguity after the first time step, but ees should be able to tackle multiple maze
trials. In the task depicted in Fig. 10 some form of linkage is necessary for the sys-
tem to be able to determine the appropriate move on the second timestep (ts,). In
maze one the correct action is O and in maze two it is 1 (the stimulus, however, is
the same, i.e. 001).
ees should be able to overcome the ambiguities present in the delayed reward
environments. Also presented are plots of zes performance for comparison. Gen-
eral system parameters are the same as for the tests in the Woods environment for
ali systems.
For the purposes of these tests the mazes are kept relatively simple. A is set to 2
for ali tests so on each time step the system merely has to make al-bit binary de-
cision. L is set to 3 and the systems are tested on tasks consisting of two and four
mazes of depths of 2 and 3 (figs. Il, 12 and 13). AlI parameters are set as in pre-
vious tests with a coupling rate of 0.25 and a base GA activation rate of also of
46 L. Bull, A. Tomlinson
0.25 for CCS. These graphs show the average scores of ten mns for each system
over 10,000 trials, again with a SO-point moving average filter applied to smooth
the curves. In figs. 8-10, task 4:3 for example represents a task consisting of four
mazes of depth three.
I Environmenlal Messagcs
000 001 I II 00 I
Immediately it can be seen that ZCS is not really equipped to tackle these prob-
lems. If Mz is set to 1 then ZCS can operate and is able to leam a single maze but
with four mazes ZCS will be intentionally correct on average about one in four
times at best (i.e. when the map it has leamt is presented). It can be seen that as N
is increased the system becomes increasingIy unable to locate the reward. At the
end of a IO,OOO-trial run on task 2:2 (see Fig. 10) the CCS ruIe-base was examined
and below are presented two typically observed corporations.
IIXK) r----------------,
xex)
21KI
JlXMI . . . - - -- - - - - - -- - - - - - ,
"o
u
ti)
2(XI
1000 , . . . . - - - - - - - - - - - - - - - ,
--z S
~ CCS
00
~ 600
o
u
ti) 400
200
Corporation 3846 responds to maze 1 (Fig. 10). At time O rule 349 responds to
the presented stimulus (000) and proposes aetion O. If rule 349 wins the auction
then at time 1 rule 134 will, as it matehes the new stimulus (001), automatieally be
in control of the produetion system. Rule 134 matehes the stimulus and then pro-
poses the correct aetion for maze L at time 1. This is a strong eorporation and
many eopies of it ean be seen in the rule-base.
Corporation 3931 responds to maze 2 (Fig. 10). Rule 202 matehes the stimulus
at time O and proposes the eorreet aetion. On the subsequent time step rule 328
matehes the new stimulus and again proposes the eorreet aetion. This is also a
strong eorporation, eopies of whieh ean be seen in the rule-base. In CCS these two
eorporations alone are suffieient to solve task 2:2 where as ZCS is unable to taekle
the ambiguity present on time step 1.
48 L. Bull, A. Tomlinson
Table 1.
The average number of corporations formed in CCS after 10,000 trials in task
2:2 is just under 200 and these are alI of size 2. Virtually every rule in the popula-
tion (size 400) has linked to a partner to form a two-member corporation. When N
is increased to 3 corporations of size 2 form initialIy (about 80 by trial 300) and
corporations of size 3 form more slowly (50 by trial 300). By trial 2000 there are
only two or three corporations of size 2, but just under 130 of size 3. Most rules in
the population (about 390 of 400 rules) therefore now belong to three-member
corporations.
CCS offers a much improved performance on alI tasks, but the results indicate
that the discovery components of both systems found their performance impeded
as N increased.
The above results show that adding the process of symbiogenesis to the simple
ZCS can give improvements in performance in complex environments which con-
tain ambiguous sensory inputs.
4 Conclusion
Symbiogenesis appears to have been the major factor in two important steps of
cell evolution - the evolution of chromosomes and eukaryotes. In this paper we
have used a tuneable abstract model of multi-species evolution to examine the
conditions under which the main phase of symbiogenesis, hereditary endosymbio-
ses, can occur.
We viewed the formation of an hereditary endosymbiosis between two separate
species as a megamutation, that is a macro-Ievel evolutionary phenomenon. It was
seen, perhaps not unsurprisingly, that, as the amount of interdependence between
the two species increased, the chances of the endosymbiosis becoming a more ad-
vantagous association also increased. We found that, for any amount of intra-
genome epistasis, roughly an equivalent amount of inter-species epistasis was re-
quired for the endosymbiosis to equally share a fixed population space, with it
dominating thereafter.
This general process was then added to a machine leaming architecture and
shown to improve performance in more complex environments containing sensory
Symbiogenesis as a Machine Leaming Mechanism 49
References
1. Timmis, T. Knight
Computing Laboratory, University of Kent, Canterbury, UK
1.Timmis@ukc.ac.uk
L. N. de Castro
School of Electrical Engineering and Computing,
State University of Campinas, BraziI
E. Rart
School of Computing, Napier University, Edinburgh, Scotland, UK
1 Introduction
This contribution examines the growing field of artificial immune systems (AIS).
AIS can be defined as computational systems inspired by theoretical immunology
and observed immune functions, principles and models, which are applied to prob-
lem solving [1]. The field of AIS is relativity new and draws upon work done by
many theoretical immunologists (e.g., Jerne [2], Perelson [3], and Bersini and
Varela [4] to name a few). What is of interest to researchers developing AIS is not
the modelling of the immune system, but extracting or gleaning of useful mecha-
nisms that can be used as metaphors or inspiration to help in the development of
(computational) tools for solving particular problems. Within biologically inspired
computing, it is quite common to see gross simplifications of the biologic al sys-
52 J.Timmis et al.
tems, on which the artificial systems are based: AIS is no exception. However, it
should be remembered that, although a good understanding of the biologic al sys-
tem is essential in this domain, it is inspiration from nature that is sought, rather
than the creation of accurate models.
Through reading the literature, it can be observed that AIS have been applied to
a wide range of application domains. Some of the first work in applying immune
system metaphors was undertaken in the area of fault diagnosis [5]. Later work
applied immune system metaphors to the field of computer security and virus de-
tection [6], which seemed to act as a catalyst for further investigation of the im-
mune system as a metaphor in many areas. However, as yet, there seems to be no
niche area for AIS. Some people have commented that this may be a weakness, or
a gap in the field [7] and that there needs to be a serious undertaking to find such a
niche area and this will in turn go to strengthen the area. It has also be argued that
AIS are incredibly flexible, as are many biologically inspired techniques, suitable
for a number of applications and can be thought of as a novel soft computing
paradigm, suitable for integration with many more traditional techniques [8].
The growing interest in AIS is reflected in the growing number of special ses-
sions and invited tracks at a number of well-established international conferences,
such as the IEEE SMC and GECCO conferences. The frrst international confer-
ence on artificial immune systems (ICARIS) took place at the University of Kent
at Canterbury (UKC) in September 2002. Its great success in terms of organisation
and quality of papers presented motivated the second ICARIS to be held in Edin-
burgh in September 2003.
This chapter is organised in the following manner. First, reasons are given why
the immune system has generated such interest within the computing and engi-
neering community. This is fOllowed by a simple review of relevant immunology
that has served as a foundation for much of the work reviewed in this contribution.
Immunology is a vast topic and no effort has been made to cover the whole area:
suitable citations are provided in the text to further direct the reader. The area of
AIS is then presented, in terms of a general framework proposed in [1]. A review
of AIS applications is then presented, however, providing a general overview of a
number of different application areas. Final1y, comments on the perceived future
of this emerging technology are then presented.
When considered from a computational point of view, the immune system can be
considered to be a rich source of inspiration as it displays leaming, adaptability, is
self-organising, highly distributed and displays a memory. There are many reasons
why the immune system is ofinterest to computing [1,9]; these can be summarised
as fOllOWS:
• Recognition: The immune system has the ability to recognise, identify and re-
spond to a vast number of different patterns. Additionally, the immune system
can differentiate between malfunctioning self cells and harmful non-self cells,
therefore maintaining some sense of self.
An Overview of Artificial Immune Systems 53
• Feature extraction: Through the use of Antigen Presenting Cells (APC) the
immune system has the ability to extract features of the antigen by filtering mo-
lecular noise from disease causing agents called an antigen, before being pre-
sented to other immune cells, including the lymphocytes.
• Diversity: There are two major processes involved in the generation and main-
tenance of diversity in the immune system. The frrst is the generation of recep-
tor molecules through the recombination of gene segments from gene libraries.
By recombining genes from a finite set, the immune system is capable of gen-
erating an almost infinite number of varying types of receptors, thus endowing
the immune system with a large coverage of the universe of antigens. The sec-
ond process, whieh assists with diversity in the immune system, is known as
somatic hypermutation. Immune cells reproduce themselves in response to in-
vading antigens. During reproduction, they are subjected to a somatie mutation
process with high rates that allow the creation of novel patterns of receptors
molecules, thus increasing the diversity of the immune receptors [10].
• Leaming: The mechanism of somatic hypermutation followed by a strong se-
lective pressure also allows the immune system to fine-tune its response to an
invading pathogen; a process termed affinity maturation [11]. Mfinity matura-
tion guarantees that the immune system becomes increasingly better at the task
of recognising patterns. The immune network theory is another powerful exam-
ple of leaming in the immune system. It suggests that the immune system has a
dynamic set of mutually recognising cells and molecules, and the presence of
an invading antigen causes a perturbation in this network. As a result, the dy-
narnic immune network, whieh presents an intrinsic steady state in the absence
of antigens, has to self-organise its pattern of behaviour again, so as to accom-
modate the disturbance [7]. Therefore, invading antigens require the immune
network to adapt itself to this new element.
• Memory: After an immune response to a given antigen, some sets of cells and
molecules are endowed with increased life spans in order to provide faster and
more powerful immune responses to future infections by the same or similar
antigens. This process, known as the maturation of the immune response, al-
lows the maintenance of those cells and molecules successful at recognising an-
tigens. This is the major principle behind vaccination procedures in medicine
and immunotherapy. A weakened or dead sample of an antigen (e.g., a virus) is
inoculated into an individual so as to promote an immune response (with no
disease symptoms) in order to generate memory cells and molecules to that an-
tigen.
• Distributed detection: There is inherent distribution within the immune system.
There is no one point of overall control; each immune ceH is specifically stimu-
lated and responds to new antigens that can invade the organism in any loca-
tion.
• Self-regulation: Immune systems dynamics are such that the immune system
population is controlled by local interactions and not by a central point of con-
trol. After a disease has been successfuHy combated by the immune system, it
returns to its normal steady state, until it is needed in response to another anti-
gen. The immune network theory explicitly accounts for this type of self-
regulatory mechanism.
54 J.Tirnmis et al.
• Metadynamics: The immune system is constantly creating new cells and mole-
cules, and eliminating those that are too old or are not being of great use.
Metadynamics is the name given to this continuous production, recruitment and
death of immune cells and molecules [12] .
• lmmune Network: In 1974 Jerne[2] proposed the immune network theory as an
alternative to explain how the immune system works. Re suggested that the
immune system is a dynamic system whose cells and molecules are capable of
recognising each other, thus forming an internal network of communication
within the organism. This network provides the basis for immunological mem-
ory to be achieved, via a self-supporting and self-organising network.
The remainder of this chapter outlines some of the salient features of the im-
mune system that have been employed in the development of AIS. Attention is
then drawn to significant applications of the immune system as a metaphor for
computational systems.
The vertebrate immune system is composed of diverse sets of cells and molecules
that work in collaboration with other bodily systems in order to maintain a steady
state within the host. A role of the immune system is to protect our bodies from in-
fectious agents such as viruses, bacteria, fungi and other parasites. On the surface
of these agents are antigens that allow the identification of the invading agents
(pathogens) by the immune cells and molecules, thus provoking an immune re-
sponse. There are two basic types of immunity, innate and adaptive. Innate immu-
nity [13] is not directed towards specific invaders into the body, but against any
pathogens that enter the body. The innate immune system plays a vital role in the
initiation and regulation of immune responses, including adaptive immune re-
sponses. Specialised cells of the innate immune system evolved so as to recognise
and bind to common molecular patterns found only in microorganisms, but the in-
nate immune system is by no means a complete solution to protecting the body.
Adaptive or acquired immunity [14], however, allows the immune system to
launch an attack against any invader that the innate system cannot remove. The
adaptive system is directed against specific invaders, and is modified by exposure
to such invaders. The adaptive immune system mainly consists of lymphocytes,
which are white blood cells, more specifically B-cells and T-cells. These cells aid
in the process of recognising and destroying specific substances. Any substance
that is capable of generating such a response from the lymphocytes is called an an-
tigen or immunogen. Antigens are not the invading microorganisms themselves;
they are substances such as toxins or enzymes in the microorganisms that the im-
mune system considers foreign. Adaptive immune responses are normally directed
against the antigen that provoked them and are said to be antigen specific. The
immune system generalizes by virtue of the presence of the same antigens in more
than one infectious agent. Many immunisations exploit this by presenting the im-
mune system with an innocuous organism, which carries antigens present in more
An Overview of Artificial Immune Systems 55
dangerous organisms. Thus the immune system leams to react to a particular pat-
tern of antigen.
The immune system is said to be adaptive, in that when an adaptive immune re-
sponse is elicited B-cells undergo cloning in an attempt to produce sufficient anti-
bodies to remove the infectious agent [2,15]. When cloning, B-cells undergo a
stochastic process of somatic hypermutation [10] where an attempt is made by the
immune system to generate a wider antibody repertoire so as to be able to remove
the infectious agent from the body and prepare the body for infection from a simi-
lar but different infection at some point in the future.
After the primary immune response, when the immune system first encounters
a foreign substance and the substance has been removed from the system, a certain
quantity of B-cells remain in the immune system and acts as an immunological
memory [2,16]. This is to allow for the immune system to launch a faster and
stronger attack against the infecting agent, called the secondary immune response.
A primary response [17] is provoked when the immune system encounters an an-
tigen for the first time. A number of antibodies will be produced by the immune
system in response to the infection, which will help to eliminate the antigen from
the body. However, after a period of days the levels of antibody begin to degrade,
unti! the time when the antigen is encountered again. This secondary immune re-
sponse is said to be specific to the antigen that first initiated the immune response
and causes a very rapid growth in the quantity of B-cells and antibodies. This sec-
ond, faster response is attributed to memory cells remaining in the immune system,
so that when the antigen, or similar antigen, is encountered, a new immunity does
not need to be built up, it is already there. This means that the body is ready to
combat any re-infection Fig. 1 illustrates this process.
The amount of antibody is increased by the immune system generating a mas-
si ve number of B-cells through a process called clonal selection [16], which is
now discussed in relation to the B-cell in the immune system.
The B-cell is an integral part of the immune system. Through a process of recogni-
tion and stimulation, B-cells will clone and mutate to produce a diverse set of an-
tibodies in an attempt to remove the infection from the body [18]. The antibodies
are specific proteins that recognise and bind to another protein. The production
and binding of antibodies is usually a way of signalling other cells to kill, ingest or
remove the bound substance [19]. Each antibody has two paratopes and two epi-
topes that are the specialised parts of the antibody that identify other molecules
[20]. Binding between antigens and antibodies is governed by how well the para-
topes on the antibody matches the epitope of the antigen; the closer this match, the
stronger the bind. Although it is the antibodies that surround the B-cell, which are
responsible for recognising and attaching to antigen invaders, it is the B-cell itself
that has one of the most important roles in the immune system.
56 J.Timmis et al.
Responseto
Lag Ag"
Time
Fig. 1. Primary and secondary immune response. Agl infects the system and a lag occurs
before a primary immune response is initiated. The host is then re-infected with Ag 1 and a
different antigen Ag,. A fast secondary response is elicited against Ag 1, whilst a primary re-
sponse is initiated against Ag,. At some point in the future, the host is then infected with
Agl" which is a slight variation on Ag 1• Due to the generalist capability of the immune sys-
tem, a secondary response is elicited against the antigen. (Redrawn from [1])
This is not the full story, as B-cells are also affected by helper T-celIs during
the immune response [21]. T-cell paratopes are different from those on B-cells in
that they recognise fragments of antigens that have been combined with molecules
found on the surfaces of the other cells. These molecules are called MHC mole-
cules (Major Histocompatibility Complex). As T-cells circulate through the body
they scan the surfaces of body cells for the presence of foreign antigens that have
been picked up by the MHC molecules. This function is sometimes called immune
surveillance. These helper T-cells when bound to an antigen secrete interleukines
that act on B-cells helping to stimulate them.
It is possible to identify two main philosophical avenues that try to explain how
immune memory is acquired and maintained [22-35]: (1) clonal expansion and se-
lection and (2) immune network.
Throughout the lifetime of an individual, it is expected to encounter a given an-
tigen repeatedly. The initial exposure to an antigen that stimulates an adaptive
immune response is handled by a spectrum of small clones of B-cells, each pro-
ducing antibodies of different affinity. The effectiveness of the immune response
to secondary encounters is considerably enhanced by storing some high affinity
antibody producing cells from the first infection, named memory cells, so as to
form a large initial clone for subsequent encounters. Thus memory, in the context
of secondary immune responses, is a clonal property [26].
Another theory that has been used in AIS for inspiration is the theory first pro-
posed by Jeme[2] and reviewed by Perelson [3] called the immune network the-
An Overview of Artificial Immune Systems 57
ory. This theory states that B-cells co-stimulate each other via portions of their re-
ceptor molecules (idiotopes) in such a way as to mimic antigens. An idiotope is
made up of amino acids within the variable region of an antibody or T -cell. A
network of B-cells is thus formed and highly stimulated B-cells survive and less
stimulated B-cells are removed from the system. It is further proposed that this
network yields useful topological information about the relationship between anti-
gens. For these reasons, this section focuses on this theory.
---------~ supreSSK>n
st iln \.11 ntion
Farmer el al. [27] created a simplistic model to simulate the immune system.
The model ignored the effect of T-cells and of macrophages in an attempt to cap-
ture the essential characteristics of the immune network. Central to their work was
the ca\culation of the dynamics of B-cell population related to a B-cell's stimula-
tion level. The authors proposed a simple equation that they consider takes into
account the three main contributing factors to B-cell stimulation level, these are:
(i) the contribution of the antigen bind ing (ii) the contribution of neighbouring B-
cells and (iii) the suppression of neighbouring B-cells. The rate of change of anti-
body concentration is given by [27:]
dx
- ' =C
["m
N . xx . -k,"mxx
N M
. + "m . xy .
]
dt L...) ., ' ) L... ' ·f ') L... ).' , ) (1)
}=, }= 1 } =1
where the first term represents the stimulation of the paratope of an antibody type i
by the epitope of an antibody j . The second term represents the suppression of an
antibody of type i when its epitope is recognised by the paratope of type j . The pa-
rameter c is a rate constant that depends on the number of collisions per unit time
and the rate of antibody production stimulated by a colIision. Constant k, repre-
sents a possible inequality between stimulation and suppression.
The stimulation of B-cell c10ning and mutation were included in the model to
create a diverse set of B-cells. The amount by which any one B-cell c10ned was in
relation to how stimulated the B-cell was. The more stimulated a B-cell, the more
c10nes it produced. Three mutation mechanisms were introduced on the strings:
crossover, inversion and point mutation. Crossover is the interchanging of two
points on two different strings, inversion is the simple inverting of the value of the
bit in a string, a O to a 1 and vice versa and point mutation is the random changing
of a bit in a given string.
An Overview of Artificial Immune Systems 59
Coutinho [29] first postulated the idea of repertoire completeness. He stated that,
if the immune systems antibody repertoire is complete, that is, presents receptor
molecules capable of recognizing any molecular shape, then antibodies with im-
munogenic idiotopes can be recognised by other antibodies, and therefore an idio-
typic network would be created.
However, in order to understand completeness, it is first necessary to under-
stand the concept of shape space. Shape space has been an important mechanism
to create and represent abstract models of immune cells and molecules [1]. The
basic idea is that aII the features of a receptor molecule necessary to characterise
its binding region with an antigen are called its generali sed shape. The generali sed
shape of any receptor molecule can be represented by an attribute string of a given
length L in a generic L-dimensional space, called shape space.
To illustrate this idea, consider abi-dimensional space as illustrated in Fig. 3.
The set of aH possible shapes lie within a finite volume V in this bi-dimensional
shape space. The antibodies are represented by the letter A (black dots) and the an-
tigens are depicted by the x. Each antibody (A) can recognise a given number of
antigens within an affinity threshold f: and therefore can recognise a volume (V)
of antigens (x) in shape space. Therefore, a finite set of antibodies appropriately
placed in the shape space and with appropriate affinity thresholds are sufficient to
cover the whole shape space; thus being capable of recognisng any molecular
shape that can be presented to the immune system.
It has been proposed that the immune network can be thought of as being cogni-
tive [12] and exhibits leaming capabilities. The authors proposed four reasons as
to why they consider immune systems to be cognitive: (i) they can recognise mo-
lecular shapes; (ii) they remember history of encounters; (iii) they define the
boundaries of self, and (iv) they can make inferences about antigenic pattems they
have yet to encounter. Taking these points, the paper explores cognitive mecha-
nisms of the immune system and proposes that the immune network can be
thought of as a cognitive network, in a similar way to a neural network.
60 lTimmis et al.
The work suggests that the immune network is capable of producing dynamic
patterns of activity over the entire network and that there is a self-regulatory
mechanism working that helps to maintain this network structure. These emerging
patterns within the immune network are characterised by varying numbers of B-
cells that when in a response to an antigen undergo clonal selection. The authors
use the term metadynamics of the immune system; see also [30). This can essen-
tially be taken to mean the continuaI productinn and death of immune cells and
molecules. A large variety of new B-celIs will be produced, but not alI will be a
useful addition to the immune system and many will never enter into the dynamics
of the immune system (interact with other B-celIs in the network) and will eventu-
ally die. The authors produced a simple model using these ideas and found that
there are oscillations in many of the variables within their system, in particular the
number of B-cells that are produced. There would often be rapid production of B-
cells, followed by a sharp decline in number, which the authors argue, is what you
expect to see in the natural immune system. Coupled with this oscillatory pattern,
the authors observed that a certain core and stable network structure does emerge
over time. This structure emerges due to a topological self-organisation within the
network, with the resulting network acting to record the history of encounters with
antigens. Therefore, the authors concluded that the immune system is an excellent
system for learning about new items and can support a memory of encounters by
the use of complex pattern matching and a self-organising network structure, and
can thus be thought of as being cognitive.
An Overview of Artificial Irnrnune Systems 61
There is other research that goes to support the ideas presented above. Bersini
and Varela [4] implemented the model proposed by Varela et al, [12] and sug-
gested that mechanisms such as immune memory, adaptability and the immune
system's ability to perform distributed processing could be of potential use to en-
gineering problem solving, in particular adaptive control [31] and computational
problem solving.
Following their earlier work [4] in Bersini and Valera [30] provides an effec-
ti ve summary of work done on exploring the dynamics and metadynamics of the
immune system. They c1aim that the metadynamics of the immune system allows
the identity of the immune system to be preserved over time, but still allows itself
to adapt to new situations. Simulations of an immune network confirmed this. The
reader is also directed to [7] where further arguments for this position are pro-
posed.
As a way to model the immune system metadynamics the authors proposed the
use of the immune recruitment mechanism (lRM). The IRM is a mechanism by
which the best new cells and molecules in the system are incorporated into the
network. This can be translated as saying that one should only incorporate the best
new items that are produced into the network. Therefore the selection of new
items is based on the state of the surrounding network: any other items that are
produced are lost. This gives rise to the metadynamical system that is believed to
occur in the vertebrate immune system. In this paper, the authors proposed seven
general principles that can be extracted from the immune system and applied to
creating a controlling system for the area of adaptive control, but they hope, to
other fields as well. These principles are:
• Principle J: The control of any process is distributed araund many operators in
a network structure. This allows for the development of a self-organising sys-
tem that can display emerging properties.
• Principle 2: The controller should maintain the viability of the process being
controlled. This is keeping the system within certain limits and preventing the
system from being driven in one particular way.
• Principle 3: While there may be perturbations that can affect the process, the
controller leams to maintain the viability of the process through adaptation.
This leaming and adaptation requires two kinds of plasticity: a parametric plas-
ticity, which keeps a constant population of operators in the process, but modi-
fies parameters associated with them; and a structural plasticity which is based
on the recruitment mechanism which can modify the current population of op-
erators.
• Principle 4: The leaming and adaptation are achieved by using a reinforcement
mechanism between operators. Operators interact to support common opera-
tions or controls.
• Principle 5: The dynamics and metadynamics of the system can be affected by
the sensitivity of the network.
• Principle 6: The immune recruitment mechanism can be considered to be a
stand-alone optimisation algorithm.
• Principle 7: The controller retains a population-based memory, which can
maintain a stable level in a changing environment.
62 lTimmis et al.
The authors suggest that these principles, while being very general, could prove
useful to many domains of leaming, engineering control and so ono Indeed, in their
paper they present a way of applying these general principles to the areas of adap-
tive control and to the creation of other immune-inspired algorithms.
When antibodies on a B-cell bind with an antigen, the B-cell becomes activated
and begins to proliferate. New B-cell clones are produced that are an exact copy of
the parent B-cell, but then undergo somatic hypermutation [Il] and produce anti-
bodies that are specific to the invading antigen. The clon al selection principle [15]
is the term used to describe the basic properties of an adaptive immune response to
an antigenic stimulus and is an alternative view to the position presented in the
previous section. It establishes the idea that only those cells capable of recogniz-
ing an antigenic stimulus will proliferate, thus being selected against those that do
not. Clonal selection operates on both T-cells and B-cells.
The B-cells, in addition to proliferating or differentiating into plasma cells, can
differentiate into long-lived B memory cells. Memory cells circulate through the
blood, lymph and tissues, probably not manufacturing antibodies [32]. However,
when exposed to a second antigenic stimulus they comrnence differentiating into
large lymphocytes capable ofproducing high affinity antibody.
The immune system is said to be complete: it has the ability to recognise alI anti-
gens. Antibodies and T-cell receptors produced by the lymphocytes can recognise
anY foreign (or self) molecule. Antibody molecules have idiotopes and it folIows
from the idea of completeness that these will be recognised by other antibody
molecules.
Therefore, alI molecules (shapes) can be recognised including our own, which
are also seen as antigens, or self-antigens. For the immune system to function
properly, it needs to be able to distinguish between the molecules of our own celIs
(self) and foreign molecules (non-self), which are a priori indistinguishable [33].
If the immune system is not capable of performing this distinction, then an im-
mune response will be triggered against the self-antigens, causing autoimmune
diseases.
An encounter between an antibody and an antigen does not inevitably result in
activation of the lymphocyte. It is possible that the encounter could actually cause
the death of the lymphocyte. In order for this to happen, there must be some form
of negative selection that prevents self-specific lymphocytes from becoming
prevalent.
those naive T-cells that recognise the self-molecules. This means that only T-cells
that do not recognise self-molecules are allowed to survive and become functional
T-cells.
4.1 Summary
Immunology is a vast topic; therefore; this chapter has introduced only those areas
of immunology that are pertinent to this contribution. Through a process of match-
ing between antibodies and antigens and the production of B-cells through clonaI
selection [15] and somatic hypermutation [10], an immune response can be elic-
ited against an invading antigen so that it is removed from the system. In order to
remember which antigens the immune system has encountered, some form of im-
munological memory must be present; this can be explained in part through theo-
ries such as the clonaI selection theory or the more controversiaI immune network
theories. Clearly, the immune system is performing a very important role within
the body. The sheer complexity of the system is staggering, and current immunol-
ogy only knows part of the story. Through complex interactions, the immune sys-
tem protects our bodies from infection, interacts with other bodily systems to
maintain a steady state (homeostasis). The focus of this chapter has been more on
the immune network theory. This is not to lend more weight to that particular view
point of the immune system, it has merely been presented in more depth to pro-
vide the reader with a deeper insight into one of the many complex ideas within
immunology, that have helped computer scientists and engineers over the years.
This area will now be examined in more detail.
66 J.Timmis et al.
This section introduces the reader to the field of Artificial Immune Systems (AIS).
There have been a number of attempts over the years to try and define exactly
what is an AIS. For example, [18] defined AIS to be 'a computational system
based upon metaphors of the natural immune system' and [9] defined them to be
'intelligent methodologies inspired by the immune system toward real-world prob-
lem solving'. Feeling that neither of these definitions was complete, the most re-
cent definition is taken from de Castro and Timmis[l], where they define AIS to
be 'adaptive systems, inspired by theoretical immunology and observed immune
junctions, principles and models, which are applied to problem solving'. In this
latest definition, a more complete view of what AIS are has been captured: the fact
they are inspired by the immune system, but the inspiration is not restricted to
purely theoretical immunology, but also 'wet lab' type immunology, the systems
are adaptive which means they must demonstrate some element of adaptability
and are not restricted to pieces of software but could equally be implemented on
hardware and that there is some form of application ultimately in mind - this al-
lows for the distinction between the creation of pure models of the immune system
(which indeed are useful for AIS, as has been discussed).
This section presents an overview of many different applications of AIS that
can be seen in the literature. No attempt has been made on an exhaustive survey;
for this the readers are directed to de Castro and Timmis[I], Chap. 4, where such
an exhaustive review is presented. The aim of this section is to merely illustrate
the wide applicability of AIS. Very recently, de Castro and Timmis[l] have pro-
posed the idea of a framework for AIS which consists of basic components and
processes, from which it is possible to both describe and build AIS. This frame-
work is now presented - due to the fact that it was only recently proposed, how-
ever, the framework has not been used in this artic1e when describing AIS litera-
ture published before the existence of this framework.
In an attempt to create a common basis for AIS, de Castro and Timmis[1] pro-
posed the idea of a framework for AIS. The authors argued the case for proposing
such as framework from the standpoint that in the case of other biologically in-
spired approaches, such as artificial neural networks (ANN) and evolutionary al-
gorithms, such a basic idea exists and helps considerably with the understanding
and construction of such systems. For example, de Castro and Timmis[l] consider
a set of artificial neurons, which can be arranged together so as to form an artifi-
cial neural network. In order to acquire knowledge, these neural networks undergo
an adaptive process, known as leaming or training, which alters (some of) the pa-
rameters within the network. Therefore, the authors argued that in a simplified
form, a framework to design an ANN is composed of a set of artificial neurons, a
pattern of interconnection for these neurons, and a leaming algorithm. Similarly,
the authors argued that in evolutionary algorithms, there is a set of 'artificial
chromosomes' representing a population of individuals that iteratively suffer a
An Overview of Artificial Immune Systems 67
Immune Aigorithms
Representation
Application Domain
corporates interactions between free antibodies, B-cells, and memory cells, using
the clonal selection processes as the core element of the algorithm. This three-
layered approach consists of a free-antibody layer, B-cell layer and a memory
layer. The free antibody layer provides a general search aud pattern recognition
function. The B-cell layer provides a more refined pattern recognition function,
with the memory layer providing a stable memory structure that is no longer influ-
enced by strong evolutionary pressure. Central to the algorithm is feedback that
occurs between B-cells and is part of the secondary immune response in the algo-
rithm. Novel data are incorporated into the B-cell layer and are given a chance to
thrive, thus providing a primary immune response. Initial testing of this algorithm
has shown good performance at static clustering.
a reduced number of points; thus reducing the complexity of the data. Work by de
Castro and von Zuben [55] explores the possibility of using immunological meta-
phors for Boolean competitive networks.
as the recognition radius of the location, and alI data recognised are said to lie
within the access circle of the location. In the case of storing binary data, this dis-
tance is simply interpreted as Hamming Distance. Each location also has an asso-
ciated set of counters, one for each bit in its length, which it uses to 'vote' on
whether a bit recalled from the memory should be set to 1 or O. An item of data is
stored in the memory by distributing it to every location which recognises it - if
recognition occurs, then the counters at the recognising locations are updated by
either incrementing the counter by 1 if the bit being stored is 1, or decrementing
the counter by 1 if the bit being stored is O. To recall data from the memory, alllo-
cations, that recognise an address from which recall is being attempted vote by
summing their counters at each bit position; a positive sum results in the recalled
bit being set to 1, a negative sum in the bit being set to O. This results in a mem-
ory, which is particularly robust to noisy data due to its distributed nature and in-
exact method of storing data.
These properties make it an ideal candidate as a basis for building an immune
system based model for addressing clustering problems in large, dynamic data-
bases. For example, we can consider each physicallocation along with its recogni-
tion radius to define a cluster of data; the location itself can be considered to be a
concise representation or description of that cluster, and the recognition radius
specifies the size of the cluster. Clusters can overlap - indeed, it is this precisely
this property that allows all data to be recognised with high precision whilst main-
taining a relatively low number of clusters. This has a direct parallel in the bio-
logical immune system in which antibodies exhibit cross-reactivity. If no overlap
were allowed in an SDM, then a large number of locations would be required to
cluster the data, the system would become overly specific, and hence general
trends in the data would be lost. The analogy between the immune system and the
SDM class of associative memories is detailed in Table 1. , taken from Smith et al.
[16].
In its original form however, the SDM is a static form of memory, and is built
on several assumptions that make it unsuitable to use directly as a model for data
clustering. In brief, these assumptions are that the addresses of the hard locations
are randomly chosen and fixed from the start, and that the recognition radii of
each address are equal and constant. Hart and Ross first addressed these problems
in a system named COSDM (Hart and Ross, 2001) in which they adapted a co-
74 J.Timmis et al.
Database
lnput antigen
data
antibody
SOSDM
Fig. 5. Diagrammatic representation of the SOSDM model
5.3 Robotics
Attempts have been made to apply the immune network idea to control large
populations of robots to have some form of self-organising group behaviour. Work
by Mitsumoto et al. [68] attempts to create a group of robots, which behave in a
self-organising manner, to search for food without any global control mechanism.
Central to their idea is the interaction between robots at the local level. The au-
An Overview of Artificial Immune Systems 77
thors use three main immunological metaphors. The first is B-cells, where a robot
represents a B-cell and each robot has a particular strategy on how to find food.
The second is the immune network, allowing for interaction between robots. The
third is the calculation of B-cell stimulation, where the more the robot is stimu-
lated, then the better its strategy is considered to be. In order to calculate B-cell
(robot) stimulation a modified version of Eq. (1) is used, where the robot is stimu-
lated and suppressed by neighbouring robots and stimulated by the outside envi-
ronment. Each robot carries a record of its degree of success in collecting food,
while neighbouring robots compare their success and strategies and stimulate and
suppress each other accordingly. If a robot's stimulation level is considered low,
then the strategy is considered too weak and,losing that strategy, randomly selects
another. If the robot is weB stimulated, the strategy is considered to be good and is
preserved. Over time the robots interact and successfully achieve the food collec-
tion. The authors claim good results on their test data, but indicate the need for
further research and testing.
This work is advanced by Mitsumoto et al. [69]. where similar techniques were
applied to create a group of robots to interact and achieve the transportation of
multiple objects to multiple locations. The algorithm is very similar to the first: the
B-cell is represented by a robot, the work to be done by the robots being analo-
gous to antigens, and communication between robots is achieved via the network.
The idea of B-cell cloning is also introduced into the algorithm, which is used to
represent messages to other robots. Here, a robot is stimulated by interaction be-
tween other neighbouring robots and the work environment. If a robot is achieving
the work, then it receives more stimulation. If that robot becomes well stimulated,
it produces clone B-cells that contain information about the work it is doing, since
it is considered to be good work. Other robots in the network then match these
and, if they share similar work, they become stimulated and produce other similar
work B-cells. If they do not match well, the robot will attempt to adapt its work to
the most common work strategy it encounters. Both this interaction and passing of
messages enables a group behaviour to emerge that can solve the transportation
problem. It was also shown by the authors that this is successful if the work re-
mains static or if the work requirement changes over time.
In very similar work by Lee et al. [70], the immune network metaphor is ap-
plied to creating swarm strategies for mobile robots. However, this work is virtu-
ally identical to that presented above. The authors do extend the concept in (Lee et
al, 1999) who introduce the metaphor of the T-cell into the algorithm. They pro-
pose a modified version of Eq. (1) with the addition of the T-cell metaphor. How-
ever, the authors fail to include the results of using the modified equation in their
simulation results, presenting instead results of only using the equation without the
T-cell interaction.
Work by Watanabe et al. [71] and Kondo et al. [72] attempts to create a mecha-
nism by which a single, self-sufficient autonomous robot, the immunoid, can per-
form the task of collecting various amounts of garbage from a constant1y changing
environment. The environment for the immunoid consists of garbage to be col-
lected, and a home base consisting of a wastebasket and a battery charger. The au-
thors use the metaphors of antibodies, which are potential behaviours of the im-
munoid, antigens, which are the environmenta1 inputs such as existence of
78 J.Timmis el al.
garbage, wall and home bases and the immune network, which is used to support
good behaviours of the immunoid. In order for the immunoid to make the best
strategy decision, the immunoid detects antigens and matches the content of the
antigen with a selection of aH the antibodies that it possesses. For example, the
immunoid may have antibodies that are suitable for when a wall is met head-on
and therefore needs to turn right. Each antibody of the immunoid records its con-
centration level, which is calculated using Eq. (1). A number of antigens (envi-
ronmental inputs) are detected and the levels of antibodies are calculated and the
antibody with the highest concentration is selected as the appropriate behaviour to
employ. In experimental results, the authors prepared 24 antibodies for the immu-
noid (potential behaviours) and observed good results. The authors then extended
this work. This was an attempt to create more emergent behaviour within the net-
work of robots [71] by the introduction of genetic operators.
The field of diagnosis is a vast field driven by the requirement to accurately pre-
dict or recover from faults occurring in plant. One approach to detect abnormal
sensors within a system [73] has been to use the combination of Leaming Vector
Quantization (LVQ) [74] and the immune network metaphor. The idea behind the
system is to use LVQ to determine a correlation between two sensors from their
outputs when they work properly, and then use an immune network to test sensors
using extracted correlations. Within the system, each sensor corresponds to a B-
ceH and sensors test one another' s outputs to see whether or not they are normal.
Each sensor calculates a value based on an adapted version of Equation 1 where
the inputs to the equation are reliability of the sensor, rather than similarity to the
neighbour. A sensor that has a low value is considered to be fau1ty and can there-
fore be flagged for needing repair. Using this method has the advantage of having
no overall control mechanism for checking for faulty sensors; they can detect for
themselves when they are faulty. Simulations of their system showed the potential
for good diagnostic results, and the paper points the way forward for more re-
search and actual application to real plants.
Aiso in the field of diagnosis, there has been an interest in creating other dis-
tributed diagnostic systems. Initial work by Ishida et al. [5, 75] proposed a parallel
distributed diagnostic algorithm. However, the authors likened their algorithm to
that of an immune network, due to its distributed operation, and the systems emer-
gent co-operative behaviour between sensors. This work was then continued by
Ishida et al. [76,77] and active diagnostic mechanism [78]. The work in [78]
builds on foundations laid in the others so will be briefly examined here.
Active diagnosis continually monitors for consistency between the current
states of the system with respect to the normal state. The authors argue that the
immune system metaphor is a suitable idea for creating an effective active diag-
nostic system. Central to their idea is the immune network theory, where each sen-
sor can be equated with a B-cell [73]. Sensors are connected via a network (the
immune network), with each sensor maintaining a record of sensory reliability,
which is continuaHy changed over time - creating a dynarnic system. Sensors in
An Overview of Artificial Immune Systems 79
the network can test each other for reliability, but where this work differs from the
above is the way in which the reliability of each sensor is calculated. This will not
be explored here. The key features of the immune system that is used by this work
are distributed agents that interact with each other in parallel (each agent only re-
acting on its own knowledge and not via a central controller), and the creation of
memory of the sensor state formed via a network.
Hardware fault tolerant systems seek to provide a high degree of reliability and
flexibility even in the presence of errors within the system. The said system must
be protected from a variety of potential faults, manifesting in such forms as per-
manent stuck at faults or intermittent faults.
Bradley and Tyrrel 1[79] proposed what they called Immunotronics (immu-
nological electronics) in order to implement a finite state machine based counter
using imrnune principles. Their proposed system relied upon the negative selec-
tion algorithm that is responsible for creating a set of tolerance conditions to
monitor changes in hardware states. They employed a binary Hamming shape-
space to represent the tolerance conditions.
Recent work Timmis et al. [80] discusses important issues when considering
the design of immune inspired fault tolerant embedded systems. The authors high-
light that one advantage of using a technique based on AIS in comparison to tradi-
tional fault tolerant approaches, is the possibility to exploit the evolutionary prop-
erty of the immune system. While conventional fault tolerant techniques generate
static detectors that have to be updated offline, AIS-based techniques will enable
the development of adaptable fault tolerant systems, in' which error detectors may
evolve during runtime. This feature will increase the availability of embedded sys-
tems since acceptable variations of non-erroneous states can be integrated to the
self-system. For example, external factors (e.g. temperature) induce changes that
might have significant effects on the system functionalities, while internal changes
(e.g. component replacement) could give rise to variability in self that must be no-
ticed. The authors also argue that AIS techniques, however, pose some challenges.
One of them is the need to ensure that the detectors generated fully cover the non-
self space (i.e. the erroneous states). This is determined by the mode of detector
generation, which in turn affects the resulting detector set as well as the speed of
the operation. However, the distribution of the self-data can be exploited to en-
hance the process. Other metaphors of the immune system are also tagged as po-
tential avenues for research in this area such as the adaptability feature, which is
inherent in the immune network metadynamics [7].
5.5 Optimisation
In order to address the issue of designing a Genetic Algorithm (GA) with im-
proved convergence characteristics, particularly in the field of design constraints,
Hajela et al. [81] proposed a GA simulation of the immune system. The motiva-
tion for their work stems from the fact that genetic algorithms, when applied to
design constraints, have been found to be very sensitive to the choice of algorithm
parameters, which can ultimately affect the convergence rate of the algorithm. The
authors use the idea of antibody-antigen binding to define a complex matching
80 J.Timmis et al.
5.6 Scheduling
The problem of protecting computers from viruses, unauthorised users, etc. consti-
tutes a rich field of research for artificial immune systems. The existence of a
natural immune system to fight against biological microorganisms like viruses and
bacteria is probably the most appealing source of inspiration for the development
of an artificial immune system to combat computer viruses and network intruders.
6 Summary
Using the immune system as inspiration has proved very useful when trying to ad-
dress many computational problems. The immune system is a remarkable leaming
system. Through the use of B-cells and T-cells the immune system can launch an
attack against invading antigens and remove them from the system. This is
achieved through a process of B-cell stimulation followed by cloning and muta-
tion of new antibodies. This diversity that is generated allows the immune system
to be adaptive to new, slightly different infections. The immune system is able to
retain information about antigens; so that next time the body is infected a quicker,
84 J.Timmis et al.
For each of the many contributions discussed in Sec. 5 it would be possible to talk
at length regarding possible extensions and improvements. Instead, we will keep
our discussion about future trends on AIS at a high level.
Although several papers have been discussed proposing the use AIS to solve
problems in many areas of research, few of those have attempted to present a for-
mal artificial immune system (e.g., [19, 43, 98]. In reviewing alI these works on
An Overview of Artificial Immune Systems 85
AIS, it becomes clear that the area is lacking the proposal and development of a
general framework in which to design artificial immune systems. Observing other
comparable computational intelligence (or soft computing) paradigms, such as ar-
tificial neural networks, evolutionary computation and fuzzy systems, it is clear
that there is the presence of well-described sets of components and/or mechanisms
with which to design such algorithms. To this end, a framework has been pro-
posed [1], although much work remains to be done on this framework in terms of
formalisation from a mathematical viewpoint and augmentation in terms of new
shapes spaces and development of new algorithms which have been inspired by
other areas of immunology, as yet unexplored by computer scientists and engi-
neers.
This leads to an interesting avenue of research. To date, the concentration has
mainly been on the more basic immunology, such as simple antibodies, clonal se-
lection and so forth. Recent work by Aickelin and Cayzer, [108] postulated the use
of the Danger Theory [109] for AIS. This is an interesting idea, sadly beyond the
scope of this chapter. However, it is quite possible the danger theory has much to
offer AIS in terms of a paradigm shift in thinking - as yet unexplored. This shift is
away from the idea that the immune system controls the response, typically
adopted in AIS, to the idea that the tissue initialises and controls the response: in
some way contextualises the response. This could provide a powerful metaphor,
especially in terms of data mining in dynamic environments, where the context of
what you might want to learn may change over time. Indeed, this debate was
opened further by Bersini, [7], who argued that the danger theory idea is nothing
more than a change of terminology from the idea of self/non-self to the idea that
something is dangerous or not dangerous. It will be interesting to observe if this
debate grows pace.
The debate also has to be had: is there a single immune algorithm? To date, this
has not been addressed and is stiH an open question. One observation that is made
regarding AIS is that there are so many different algorithms around, it is so com-
plicated and you never know which one to use. In answer to this, you may state
that either more rigour and analysis has to be applied to the current algorithms to
identify exactly their suitability for problems and therefore, what AIS will offer is
a very rich suit of effective and well-understood algorithms. Altematively, you
could pursue a single unified algorithm: but then would that either enhance or re-
strict the field of AIS? It may enhance it in the fact that there then exits a single
commonly understood algorithm - so people then know what you mean when you
say an immune algorithm, or it may re strict it in the sense that the immune system
is such a complex system, why try and limit that to one simple algorithm - why
not exploit the many complexities therein?
It would certainly seem that there are many challenges ahead. Immunology has
a great deal to teach people involved with computational problems: we have only
just scratched the surface. A greater interaction between biology and computer
science is needed if we are to fully exploit the richness of this marvellous biologi-
cal system.
86 J.Timmis et al.
References
De Castro, L.N and Timmis, J (2002). Artificial Immune Systems: A New Computa-
tional Intelligence Approach. Springer, Berlin, Heidelberg, New York. ISBN 1-
85233-594-7
2 Jerne, N. (1974). Towards a network theory of the immune system. Annals of Immu-
nology (Inst.Pasteur). 12SC. pp. 373-389.
3 Perelson, A. S. (1989). Immune Network Theory, Imm. Rev., 110, pp. 5-36.
4 Bersini, H and Varela, F. (1990). Hints for adaptive problem solving gleaned from
immune networks. Parallel Problem Solving from Nature, 1" Workshop. PPSW1.
Dortmund, Germany. Springer, Berlin, Heidelberg, New York, pp. 343-354.
5 Ishida, Y. (1990). Fully Distributed Diagnosis by PDP Leaming Algorithm: Towards
Immune Network PDP Model. Proc. of the IEEE International Joint Conference on
Neural Networks. San Diego, USA, pp. 777-782.
6 Forrest, S, Perelson, A, Allen, L and Cherukuri, R (1994). Self-Nonself Discrimina-
tion in a Computer. Proc. of IEEE Symposium on Research in Security and Privacy.
Oakland, USA, pp. 202-212.
7 Bersini, H. (2002). ATribute to ..... In proceedings of 1" International Conference on
Artificial Immune Systems (ICARIS). Timmis J. and Bentley P. (Eds.) pp. 107-112.
8 De Castro, L.N and Timmis, J. (2003). Artificial Immune Systems as a Novel Soft
Computing Paradigm. Soft Computing.
9 Dasgupta, D (1998b). An overview of artificial immune systems. Artificial Immune
Systems and Their Applications. pp. 3-19. Springer, Berlin, Heidelberg, New York
10 Kepler, T and Perelson, A. (1993). Somatic Hypermutation in B-cells : An Optimal
Control Treatment. Journal of Theoretical Biology. 164. pp. 37-64.
11 Berek, C. & Ziegner, M. (1993). The Maturation of the Immune Response, Immunol.
Today, 14(8), pp. 400-402.
12 Varela, F, Coutinho, A, Dupire, B and Vaz, N. (1988). Cognitive Networks : Immune
and Neural and Otherwise. TheoreticalImmunology: Part Two, SFI Studies in the Sci-
ences ofComplexity, 2, pp.359-371
13 Janeway, C. (1993). Life, Death and the Immune System. Scientific American Special
Issue. How the immune system recognises invaders, pp. 27-36.
14 Roitt, I. (1997). Essential Immunology: 9''' Edition. Chap. Specific Acquired Immu-
nity, pp. 22-39. Pub. Blackwell Science, Oxford
15 Bumet, F. M. (1959). The Clonal Selection Theory of Acquired lmmunity, Cambridge
University Press., Cambridge.
16 Smith, D. J., S. Forrest & A. S. Perelson (1998). Immunological Memory is Associa-
tive. In Artificiallmmune Systems and their Applications. Ed. D. Dasgupta. Springer,
Berlin Heidelberg New York
17 Tizzard, 1. (1988a). lmmunology: An lntroduction. 2m1 edition. Chap. The response of
B-cells to Antigen. Pp. 199-223. Saunders College.
18 Timmis, J. (2000). Artificial Immune Systems: A novel data analysis technique in-
spired by the immune network theory. Ph.D. Thesis. University of Wales, Aberyst-
wyth.2000.
19 De Castro, L. N. & Von Zuben, F. J. (1999). Artificial Immune Systems: Part 1- Ba-
sic Theory and Applications, Technical Report - RT DCA 01/99, p. 95
An Overview of Artificial Immune Systems 87
43 Hunt, J, King, C and Cooke, D. (1996). Immunising Against Fraud. Proc. Knowledge
Discovery and Data Mining. IEE Colloquiurn. IEE., pp. 3845.
44 Hunt, J, Timmis, J, Cooke, D, Neal, M and King, C. (1998). nSYs: Development of
an Artificial Immune System for real world applications. In Artificial Immune Systems
and theory Applications. Ed. D. Dasgupta. pp. 157-186.
45 Neal, M, Hunt, J and Timmis, J. (1998). Augmenting an artificial immune network.
Proc. ofthe IEEE SMC, San Diego, Calif., pp. 3821-3826.
46 Timmis, J, Neal, M and Hunt, J. (2000). An Artificial Immune System for Data
Analysis. Biosystems. 55(1/3), pp. 143-150
47 Fisher, R (1936). The use of multiple measurements in taxonomic problems. Annual
Eugenics. 7, ll. pp. 179-188
48 Kohonen, T. (1997a). Selj-Organising Maps. 200 Edition.
49 Timmis, J, Neal, M and Hunt, J. (1999). Data Analysis with Artificial Immune Sys-
tems and Cluster Analysis and Kohonen Networks: Some Comparisons. Proceedings
ofthe IEEE SMC, Tokyo, Japan. pp. 922-927.
50 Timmis, J and Neal, M.(2oo1) A Resource Limited Artificial Immune System for
Data Analysis. Knowledge Based Systems, 14(34):121-130, June 2001.
51 Timmis, J (2001). aiVIS: Artificial Immune Network Visualisation. EuroGraphics
UK 2001 Conference Proceedings, pp. 61-69, University College London, April2oo1.
52 Knight, T and Timmis, J. (2001). In N Cercone, T Lin, and Xindon Wu, editors, IEEE
International Conference on Data Mining, pp. 297-304, San Jose, Calif. December
2001. IEEE, New York
53 Neal, M. (2002). An Artificial Immune System for Continuous Analysis of Time-
Varying Data. In 1" International Conference on Artificial Immune Systems (ICARIS),
pages 75-86, Canterbury, UK.
54 Knight, T and Timmis, J. (2002). A Multi-Layered Immune Inspired Approach to
Data Mining. Recent Advances in Soft Computing, Nottingham, UK. 2002
55 De Castro, L. N., & Von Zuben, F. J., (2001). The Construction of a Boolean Com-
petitive Neural Network Using Ideas From Immunology, submitted.
56 Slavov, V & Nikoleav, N (1998). Immune network dynamics for inductive problem
solving. Lecture Notes in Computer Science, 1498, pp. 712-721. Springer, Berlin,
Heidelberg, New York.
57 Hart, E & Ross, P. (2001). Clustering Moving Data with a Modified Immune Algo-
rithm. EvoWorkshops 2001 - Real World Applications of Evolutionary Computing.
58 Hart, E & Ross, P. (2002a). Exploiting the Analogy between Immunology and Sparse
Distributed Memories. Proc. of 1CARIS-2002, pp. 49-58.
59 Hart, E. (2002b) Immunology as a Metaphor for Computationaiinformation Process-
ing: Fact of Fiction ? PhD thesis. University of Edinburgh.
60 Kanerva, P. (1998) Sparse Distributed Memory. MIT Press, Cambridge, Mass.
61 Potter, M.A. & De Jong, K.A (2000) Cooperative coevolution: An architecture for
evolving co adapted subcomponents. Evolutionary Computation, 8(1):1--29.
62 Carter, J.H. (2000). The Immune System as a Model for Pattern Recognition and
Classification. Journal of the American Medical Informatics Assocation.711. pp. 28-
41.
63 Wettschereck, D. Aha, D.W, and Mohri, T. 1997. A review and empirical evaluation
of feature weighting methods for a class of lazy leaming algorithms. Artificial Intelli-
gence Review. 11:273-314.
64 Gennari, J.H. Langley, P and Fisher, D. (1989). Models of information concept forma-
tion. Artificial1ntelligence; 40: 11-61.
65 Watkins, A. (2001). A resource limited artificial immune classifier. MS Thesis. Mis-
sissippi State University. Miss.
An Overview of Artificiallmmune Systems 89
66 De Castro, L. N. & Von Zuben, F. J. (2oo0c). The Clonal Selection Algorithm with
Engineering Applications, Proc. of GECCO '00 - Workshop Proceedings, pp. 36-37.
67 Watkins, A and Timmis, J. (2002). Artificiallmmune Recognition Systems (AIRS):
Revisions and Refinements. In Proceedings of the 1" International Conference on Ar-
tificiallmmune Systems. pages 173-181, University of Kent at Canterbury, September.
68 Mitsumoto, N, Fukuda, T and Idogaki, T. (1996). Self-Organising Multiple Robotic
System. Proceedings of IEEE International Conference on Robotics and Automation.
pp. 1614-1619. Minneapolis, USA. IEEE.
69 Mitsumoto, N, Fukuda, T, Arai, F & Ishihara, H (1997). Control of distributed
autonomous robotic system based on the biologically inspired immunological archi-
tecture. Proceedings of IEEE International Conference on Robotics and Automation.
pp. 3551-3556. Albuquerque, N.M .. IEEE, New York
70 Lee, Dong-Wook and Sim, Kwee-Bo. (1997). Artificial immune network based co-
operative control in collective autonomous mobile robots. Proc. of IEEE International
Workshop on Robot and Human Communication. Sendai, Japan. IEEE, New York,
pp.58-63.
71 Watanabe, Y, Ishiguro and Uchikawa, Y. (1998). Decentralised behaviour arbitration
mechanism for autonomous mobile robots using immune network. In Artificial Im-
mune Systems and their applications. Ed. D. Dasgupta. pp. 187-209. Springer, Berlin
Heidelberg New York
72 Kondo, T, Ishiguro, A, Watanabe, Y and Uchikawa, Y. (1998). Evolutionary con-
struction of an immune network based behaviour arbitration mechanism for autono-
mous mobile robots. Electrical Engineering in Japan. 123/3. pp. 1-10
73 Kayama, M, Sugita, Y, Morooka, Y & Fukuodka, S. (1995). Distributed diagnosis
system combining the immune network and learning vector Quantization, pp. 1531-
1536 of Proc. IEEE 21" International Conference on Industrial Electronics and Con-
trol and Instrumentation, Orlando, USA.79.
74 Kohonen, T. (1997b). Selj-Organising Maps. 2"d Edition. Chap. Leaming Vector
Quantization. pp. 203-217. Springer, Berlin Heidelberg New York
75 Ishida, Y & Mizessyn, F. (1992). Learning algorithms on immune network model:
application to sensor diagnosis. Proc. International Joint Conference on Neural Net-
works, Beijing, China, pp. 33-38.
76 Ishida, Y (1996). Distributed and autonomous sensing based on immune network.
Proc. ofArtificial Life and Robotics. Beppu. AAAI Press, pp. 214-217.
77 Ishida, Y & Tokimasa, T. (1996). Diagnosis by a dynamic network inspired by im-
mune network. Proc. World Congress of Neural Networks, San Diego, Calif. pp. 508-
511.
78 Ishida, Y. (1997). Active Diagnosis by Self-Organisation : An approach by the im-
mune network metaphor. Proceedings ofthe International Joint Conference on Artifi-
cial Intelligence. pp. 1084-1089. Nagoya, Japan.
79 Bradly, D. W. & Tyrrell, A. M. (2oo0a), Immunotronics: harware Fault Tolerance In-
spired by the Immune System, Lecture Notes in Computer Science, 1801, ppll-20.
80 Timmis, J, de Lemos, R, Ayara, M and Duncan R. (2002) Towards Immune Inspired
Fault Tolerance in Embedded Systems. To appear in the Proceedings of International
Conference on Neural Information Processing. Singapore. November 2002.
81 Hajela, P., Yoo, J. & Lee, J. (1997). GA Based Simulation of Immune Networks -
Applications in Structural Optimization, Journal of Engineering Optimization.
90 lTimmis et al.
82 Toma, N, Endo, S & Yamada, K (1999). Immune algorithm with immune network
and MHC for adaptive problem solving. Proc. lEEE SMC. Tokyo, Japan, IV, pp. 271-
276.
83 Mori, K, Tsukiyama, M and Fukuda, T. (1996). Multi-optimisation by immune algo-
rithm with diversity and leaming. Proc. ofthe lEEE SMC, pp. 118-123.
84 Mori, K, Tsukiyama, M and Fukuda, T (1998). Application of an immune algorithm
to multi-optimisation problems. Electrical Engineering in Japan. 122/2. pp. 30-37
85 Fukuda, T, Mori, K and Tsukiyama, M. (1998). Parallel Search for Multi-Modal
Function Optimisation with Diversity and Leaming of Immune Algorithm. Artificial
lmmune Systems and Their Applications. pp. 210-220. Springer, Berlin Heidelberg
New York
86 Mori, K, Tsukiyama, M and Fukuda, T. (1994). Immune Algorithm and Its Applica-
tion to Factory Load Dispatching Planning. pp. 1343-1346 of Proc. Japan-USA Sym-
posium on Flexible Automation.
87 Chun, J, Kim, M & Jun, H. (1997). Shape Optimisation of Electromagnetic Devices
Using Immune Algorithms. lEEE Transactions on Magnetics, 33,(2).
88 Hart, E. Ross, P. and Nelson, T (1998). Producing robust schedules via an artificial
immune system. Prac. oflEEE CEC'98, pp. 464-469. IEEE.
89 Hart, E. & Ross, P. (1999a). The Evolution and Analysis of a Potential Antibody Li-
brary for Use in Job-Shop Scheduling, In New ldeas in Optimisation, D. Come, M.
Dorigo & F. Glover (Eds.), McGraw HiU, London, pp. 185-202.
90 Hart, E. & Ross, P. (1999b). An Immune System Approach to Scheduling in Chang-
ing Environments", Proc. ofGECCO'99, pp. 1559-1566.
91 Dasgupta, D (1999). Immunity based intrusion detection systems: A general frarne-
work. Proceedings of the 22nd National lnformation Systems Security Conference
(N1SSC). Pp. 147-159
92 D'haeseleer, P, Forrest, S and Helman, P (1996). An Immunological Approach To
Change Detection: Algorithm and Analysis and Implications. Proceedings ofthe 1996
lEEE Symposium on Computer Security and Privacy. pp. 110-119
93 Forrest, S, Hofmeyr & Somayaji, A & Longstaff, T. (1996). A sense of self for UNIX
processes. Proc. lEEE Symposium on Research in Security and Privacy. Oakland,
USA, pp. 120-128.
94 Forrest, S, Hofmeyr, S and Somayaji, A (1997). Computer Immunology. Communica-
tions ofthe ACM. 40/10. pp. 88-96
95 Somayaji, A., Hofmeyr, S. A. & Forrest, S. (1997), Principles of a Computer Immune
System, Proc. ofthe new Security Paradigms Workshop, pp. 75-81.
96 Hofmeyr, S, Forrest, S & Somayaji, A. (1998). Intrusion detection using a sequence of
system calls. Joumal of Computer Security, 6, pp. 151-180.
97 Hofmeyr, S and Forrest, S (1999). Immunity by Design: An artificial immune system.
Proc. ofGECCO'99, Pub. Morgan-Kaufman. pp. 1289-1296
98 Hofmeyr, S.A. and Forrest, S. (2000). Architecture for an Artificial Immune System.
Evolutionary Computation 7(1):45-68.
99 Kim, J and Bentley, P. (1998). The human immune system and network intrusion de-
tection. Proc. of 7th European Congress on lntelligent Techniques - Soft Computing.
Aachan and Germany
100 Kim, J. & Bentley, P. (1999), Negative Selection and Niching by an Artificial Irn-
mune System for Network Intrusion Detection, Proc. of GECCO'99, pp. 149-158.
An Overview of Artificial Immune Systems 91
101 Dasgupta, D. (2000). An Immune Agent Architecture for Intrusion Detection, Proc. of
GECCO'OO, Workshop on Artificial Immune Systems and Their Applications, pp ..
102 Kephart, J. O. (1994). A Biologically Inspired Immune System for Computers, R. A.
Brooks & P. Maes (Eds.), Artificial Life IV Proceedings of the Fourth International
Workshop on the Synthesis and Simulation of Living Systems, MIT Press, Cambridge,
Mass., pp. 130-139.
103 Kephart, J. O., Sorkin, G. B. & Swimmer, M. (1997), An Immune System for Cyber-
space, Prac. of the IEEE SMC'97, pp. 879-884.
104 Kephart, J. Sorkin, B. Swimmer, M and White, S. (1998). Blueprint for a computer
immune system. In Artificial Immune Systems and their Applications. Ed. D. Das-
gupta. pp. 242-260. Springer, Berlin Heidelberg New York
105 Marmelstein, M, Veldhuizen & Lamont, G. (1998). A Distributed Architecture for an
Adaptive Computer Virus System. Proc. of the IEEE SMC, San Diego, Calif.m. pp.
3838-3843.
106 Lamont, G. B., Marmelstein, R. E. & Van Veldhuizen D. A. (1999), A Distributed
Architecture for a Self-Adaptive Computer Virus Immune System, New Ideas in Op-
timisation, D. Corne, M. Dorigo & F. Glover (Eds.), McGraw Hill, London, pp. 167-
183.
107 Harmer, P.K. and lamont, G.B. (2000). An Agent Based Architecture for a Computer
Virus Immune System. In Proceedings of Artificial Immune Systems Workshops. pp.
45-46. GECCO 2000, Las Vegas, USA.
108 Acklien, U and Cayzer, S. (2002). The Danger Theory and its Application to Artificial
Immune Systems. Proceedings of the r International Conference on Artificial Im-
mune Systems (ICARIS). pp. 141-148. Canterbury, UK.
109 Matzinger, (. (1994a). Tolerance, Danger and the Extended Family. Annual Review of
Immunology. 12:991-1045.
IlO Warrender, C, Forrest, S & Pearhmutter, B. (1999). Detecting intrusions using system
calIs: Alternative data models. Proc. of Symposium on Security and privacy. IEEE,
New York, pp. 133-145.
Embryonics and Immunotronics:
Biologically Inspired Computer Science Systems
A. Tyrrell
Abstract. This first part of this article details and expands the work on embryon-
ics, a recently proposed fault-tolerant cellular architecture with reconfiguration
properties inspired by the ontogenetic development of multicellular systems. The
design of a selector-based embryonic cell and its applications are presented. The
second part of this article describes a novel approach to hardware fault tolerance
that takes inspiration from the human immune system as a method of fault detec-
tion. The human immune system is a remarkable system of interacting cells and
organs that protects the body from invasion and maintain reliable operation even
in the presence of invading bacteria Of viruses. Here we seek to address the field
of electronic hardware fault tolerance from an immunological perspective with the
aim of showing how novel methods based upon the operation of the immune sys-
tem can both complement and create new approaches to the development of reli-
able hardware systems. The final part of the article suggests a combined architec-
ture that would have the characteristics and advantages of both Embryonics and
immunotronics.
1 Introduction
genome contains the ensemble of the genetic inheritance of the individual and, at
the same time, the instructions for both the construction and the operation of the
organism. The parallel execution of 60 trillion genomes in as many cells occurs
ceaselessly from the conception to the death of the individual. This process is re-
markable for its complexity and its precision. Moreover, it relies on completely
discrete information: the structure of DNA (the chemi cal substrate of the genome)
is a sequence offour bases, usually designated with the letters A (adenine), C (cy-
tosine), G (guanine), and T (thymine).
The analogy between multicellular organisms and multiprocessor computers is
therefore not too far fetched, and well worth investigating, particularly when con-
sidering that nature has achieved levels of complexity that far surpass any man-
made computing system. The aspect of biological organisms on which this chapter
is centred is their phenomenal robustness: in the trillions of cells that make up a
human being, faults are rare, and in the majority of cases, successfully detected
and repaired. This level of reliability is remarkable, and relies on very complex
mechanisms that are difficult to translate directly into silicon (e.g. biology is typi-
cally dealing with 3D structures, it also has almost infinite resources at hand).
Nevertheless, it will be seen that, by drawing inspiration from the overall structure
of biologic al organisms, architectures can be developed that are inherent1y fault
tolerant.
The embryonics project (for embryonic electronics) is inspired by the basic
processes of molecular biology and by the embryonic development of living be-
ings [1]. By adopting certain features of cellular organisation, and by transposing
them to the two-dimensional world of integrated circuits in silicon, it wiIl be
shown that properties unique to the living world, such as self-replication and self-
repair, can also be applied to artificial objects (integrated circuits). Self-repair al-
lows partial reconstruction in case of a minor fault, while self-replication allows
complete reconstruction of the original device in cases where a major fault occurs.
These two properties are particularly desirable for complex artificial systems in
situations that require improved reliability, such as [2]:
• Applications which require very high levels of reliability, such as avionics or
medical electronics
• Applications designed for hostile environments, such as space, where the in-
creased radiation levels reduce the reliability of components
• Applications which exploit the latest technological advances, and notably the
drastic device shrinking, low power supply levels, and increasing operating
speeds, which accompany the technological evolution to deeper sub-micron
levels and significantly reduce the noise margins and increase the soft-error
rates
To increase still further the potential reliability of these systems, inspiration has
also been taken from biologic al immune systems - immunotronics. The acquired
immune system in humans (and most vertebrates) has a mechanism for error de-
tection, which is simple, effective and adaptable. The second part of this chapter
will introduce immunotronic ideas and suggest ways they might be incorporated
into an embryonic architecture
Embryonics and lmmunotronics 95
2 An Overview of Embryonics
In any living being, every one of its constituent cells interprets the DNA strand al-
located in its nucleus to produce the proteins needed for the survival of the organ-
ism, independently of the particular function it performs. Which part or parts of
the DNA are interpreted will depend on the physicallocation of the cell with re-
spect to its neighbours.
The aim of Embryonics is to transport these basic properties to the two-dimen-
sional world of cellular arrays using specifically designed FPGAs as building
blocks. In any embryonic system, every one of its FPGA-based celIs interprets a
configuration register allocated in its memory, independently of the particular
logic function it performs. Which configuration register is interpreted will depend
on the coordinates of the cell determined by those of its neighbours. Embryonic
cellular arrays share the following properties with their biological counterparts [3-
5]:
At start up all the celIs are identical, i.e. either they have their memory initialised
to a pre-defmed value or the content of the memory is of no relevance, Fig. 1. A
mother cell (the zygote), arbitrarily defined as having the coordinates 0,0, propa-
gates the genome to the neighbouring (daughter) cells to the north and east. The
process continues until alI the cells in the arrays have a copy of the genome. Each
cell will be different to the others because every one will execute one gene accord-
ing to its coordinates, Fig. 2.
Each gene is part of the global program of the cell, the genome, and the execution
of a particular gene depends only on the position of the cell within the array. In
other words, eacb cell bas a copy of the genome (set or configuration registers al-
located in its memory), and extracts and executes the gene which configures it. In
this sense, each cell is universal, i.e. it can perform the function of any other celI,
given the proper set of coordinates.
96 A. Tyrrell
'/"III!
, , , r"
'1fII" '"
" li/III'
11".'/ I"I'/r
Hence, rather than specifying a program, it specifies that a ceH should have the
functionality of a particular logic function and also the relevant connectivity of a
ceH. Another ceH, having different coordinates, wiII interpret a different part of
the genome and hence wiH have a different functionality and different connec-
tivity.
/ w
00
Memory Router
Configurati
Registers 10
II
12
20 110
21
22
Logic
block
Each part of this basic ceH will be described in more detaillater in this chapter. This ar-
chitecture presents the foHowing advantages
The foHowing sections describe in detail each of the constituent blocks of the em-
bryonic ceH.
Embryonics and Immunotronics 99
4.1 Memory
Each cell must have enough memory to contain a copy of the configuration regis-
ter of aII the cells in the array (or at least the celIs within the same column as it).
Therefore, every cell will be able to replace any other in case of failure . For our
cell a 256x20 memory was chosen. With this memory size one can build arrays of
up to 16x 16 celIs. While the size of applications might be limited by memory re-
sources, the partitioning of the cellular system, and limitation on the universality
of cells allow implementation to be achieved without any detriment to the sys-
tem's ability to reconfigure on fault.
On each cell, the outputs of the memory will only be used intemally to config-
ure the programmable logic, therefore, a serial-input parallel-output memory was
chosen to avoid the necessity of a 20-bit bus interconnecting cells. Figure 4 shows
the memory system of
In19:0
8-bit
256x 20
counter Control of
8-bit Memory
cIk Logic 8lock
2-1 Addr7:0
MUX Outl9:0 and VO router
Addre
Fig. 4. 810ck diagram of the memory system located in each ceH one ceH
Setting-up an embryonic array implies two phases: one on which the coordi-
nates are caIculated and the genome is downloaded, and a second on which the ar-
ray performs the desired function. These are called the configuration and opera-
tional phases, respectively.
During configuration, memory addresses are taken from the 8-bit counter,
which is incremented every time a 20-bit string is loaded. When aII configuration
registers have been shifted in, the mode signal changes and the caIculated coordi-
nates select the appropriate configuration register for each ceH.
100 A. Tyrrell
This block calculates the coordinates of the ceH from the coordinates of either its
south or east neighbour. Figure 5 shows the basic architecture of this block. Re-
member it is the coordinates of a particular ceH that differentiates it from other
cells in the system and defines which part of the genome it should interpret to de-
fine its functionality.
X,Y
s
~~~~~ E
X,Y East X,Y
lihl liiJ tii] [ii] ~ E
To calculate its co-ordinates, each ceH receives the co-ordinates generated at its
east and south neighbours and increments either the row or column value of one of
them depending on the state of a selection signal. The resulting values are then
presented to the north and west neighbours to enable their positions in the array to
be calculated (Fig. 5a). Figure 5b illustrates the address generation process on a
4x4 array.
Once coordinates have been calculated and stored in a register, they can be
used to select the corresponding configuration register from the memory. In this
system, each coordinate is 4 bits wide. To construct an address, both X (row) and
Y (column) coordinates are appended, with X as the most significant nibble and Y
as the least significant one.
O_ AO
l_ AI
SIN _ Al
ErN _ A3
A4
WIN - AS
Q- z
SIBUS - A6
SOBUS - A7
OUT
REG-- --'
EIN
HN-..-~
WOUT
SIN
EOUT
EOBUS
EIBUS
L O - SO WlN
_ SI
Ll _ S2
L2 EBUS - - - - '
°
L3:0 and R3:0 select one out of eight possible inputs on their respective multi-
plexers. It is possible to select or 1 as the signal to be propagated in order to fa-
cilitate the implementation. The REG bit will determine if the output is combina-
tional or sequential. The selection input for the main multiplexer element (marked
with a star on Fig. 6), can also be selected from two of the signals controlled by
the IlO router. The value of EBUS determines whether EOBUS Of EIBUS will se-
lect the block's output. If the value of PRESET is 1, then the registered output is
set to 1. If it is 0, then the registered output will become the output of the main
multiplexer element. WV controls the fIow of information in the horizon-
taI/vertical direction. If WV is 1, then the south input SIN is propagated through
both EOUT and WOUT outputs. If it is 0, then EOUT and WOUT propagate the
inputs in WIN and EIN respectively.
In a conventional cellular array the output generated by a particular ceH can only
be propagated to the nearest neighbours. In a embryonic array the IlO router pro-
vides additional paths aHowing information to propagate not only to the nearest
neighbours, but also to more distant ceHs. Figure 7 shows the mechanisms for in-
formation to be routed in this block.
102 A. Tyrrell
,/ ;=il
NO
N [i=il
[O
wi=il
WO
si=il
so
NOBU NIBUS ,/
I
I:
SI SO SO SO
I : I
SI SO
D CN DECE DECW DECS
Z3 Z2 ZI Z2 ZI Z2 ZI Z3 Z2 ZI
WIBU ;OBUS
NOUT
WOBU
IlO -IDU NIDUS
ElDUS
Router WIBUS
smus
Close inspection of Fig. 7b iIIustrates the various paths any input could folIow.
This is achieved by using tristate buffers to connect four inputs to a single output
line. Tristate buffers are controlled using selection Iines generated by 2-4 decod-
ers. In Fig. 7b, labels in bold represent the selection bits stored in the configura-
tion registers. NOUT is the output coming from the corresponding functional
block.
The smartest reconfiguration strategy is no good unless there is fast and accurate er-
ror detection system that wilI signal the start of the reconfiguration process. When
a fault is self-detected by any of the celIs the process of reconfiguration begins.
References on diagnostic techniques and BIST logic can be found in [7] and [8].
Error detection in this particular case is achieved with the use of duplicated
hardware and a simple compare and validate procedure. That is, the major func-
tional blocks are duplicated and fed the same inputs. The output from any func-
tional block (e.g. memory, address generation, and logic block) and their dupli-
cates are compared, if the same operation continues as normal, if a difference is
identified an error is signaled and a reconfiguration 'cycle' is initiated.
Any ceH detecting a self-failure issues a non-OK signal to aH its neighbours,
and they propagate this signal along the row and/or column of the affected ceH
(whether it is propagated to row or column neighbours will depend on the recon-
figuration strategy: if we are to reconfigure by column then the whole row must be
made transparent, by row then the column mist become transparent). When a cell
receives a non-OK signal from one of its neighbours, it becomes transparent for
the address-calculation process, i.e. the· ceH transmits the addresses received from
its neighbours without modification. Unaffected cells recalculate their coordinates
and consequently select a new configuration register. Figure 8 presents an embry-
onic array when aII its cells are fault-free (a) and when one of the cells has failed
(b). Notice that when reconfiguration occurs, coordinates are changed folIowing a
diagonal direction.
Embryonics and Immunotronics 103
a) b)
5 Examples
C
a) b)
Fig. 9. a Binary decision diagram for a three-input voter. b Implementation using multi-
plexers
WIN EOUT
EIBUS EOBUS
SIN STBUS
Fig. 10. Configuration of cells to propagate signals diagonally
Fig. 11 shows the final implementation of the voter. Bold arrows indicate the
flow of information. Note that propagation of the output signal also follows the
diagonal direction. The highlighted cells were programmed to implement the cir-
cuit. One advantage of this system over conventional implementations is that the
output could be routed through several cells so that the value of f is presented at
more than one output pin simultaneously, and because cells can perform self-
diagnostics, a correct output could always be selected.
A major improvement in the workings of an embryonic architecture would be
to increase the ability of the cells to self-diagnose errors before reconfiguration.
One possible way to do this might be to look towards biology again and consider
Embryonics and Immunotronics 105
how immune systems work and whether such inspiration might again be mapped
into the world of electronics. The next section in this chapter explores this in some
detail.
6 Immunotronics
Ensuring the reliability of computing and electronic systems has always been a
challenge. As the complexity of systems increases, the inclusion of reliability
measures becomes progressively more complex and often a necessity for VLSI
circuits where a single error could potentially render an entire system useless.
B C
7 Reliability Engineering
Reducing the failure probability and increasing reliability have been a goal of
electronic systems designers ever since the first components were developed. No
matter how much care is taken designing and building an electronic system,
sooner or later an individual component will faiI. For systems operating in remote
environments such as space applications, the effect of a single failure could results
in a multi-million pound installation being rendered useless. With safety critic al
systems such as aircraft, the effects are even more severe. Reliability techniques
need to be implemented in these applications and many more. The development of
fault-tolerant techniques was driven by the need for ultra-high availability, re-
duced maintenance costs, and long life applications to ensure systems can con-
tinue to function in spite of faults occurring. The implementation of a fault-
tolerant mechanism requires four stages [Il]:
• Detection of the errar
• Confinement of the errar, to prevent propagation through the system
• Error recovery, to remove the error from the system
• Fault treatment and continued system service to repair and return the system to
normal operation
The later three stages were to a greater or les ser extent considered in the first
part of this chapter; we will now deal with the detection of errors.
Any digital systems can be analysed by modelling it as a finite state machine
(FSM) with its associated state table description. In principle any sequential digi-
tal system can be modelled as FSM, or a set of interconnecting FSMs. An exam-
ple is shown in Fig.12, which shows normal states and transitions and samples of
those that could potentially lead to a system failure. The FSM is therefore an ideal
representation for deve10ping a hardware immune system.
Here we concentrate on an FSM representation as these devices are used
throughout alI stages of a sequential system design, are a source of comparison
with reliability engineering research, and are also used in hardware design pack-
ages to permit direct instantiation of their design as a complete system netlist.
What follows could, however, be used with other representations and design
methods of digital design.
Faults are represented and analysed through the use of fault models at both the
gate and functionallevel within an electronic system [12]. Gate level fault models
describe the effect of an error in terms of individual logic gates and their connec-
tions. Functional fault models check the entire function of a system at once, under
the premise that, if the functionality is correct, then the system under test is fault
free. The work presented here concentrates on the development of a novel func-
tional approach to error detection. By modelling the faults of a sequential circuit
through an analysis of the state table (that describes the functionality of the cir-
cuit) it is possible to generate tests before the circuit is even implemented. This
approach can also be used with a change to the internal architecture and logic de-
sign. This feature could be very useful with biologically inspired hardware sys-
tems.
Embryonics and lmmunotronics 107
i
t e21 1 tq 12
t q40 ~
Valid transilion ~ Q
Fig. 12. Finite state machine representation of system
In contrast to the reliability techniques that have been developed for fault-tolerant
hardware, biology has managed to solve the problem in a remarkably different
way. The fabulously complex defence system in vertebrates has evolved over
hundreds of millions of years to give rise to what we now call the immune system
- a remarkable collection of organs, ducts, and cells comparable in complexity to
the body's nervous system [13]. The immune system is distributed, layered, and
ingenious in its methods of protecting the body from invasion by billions of dif-
ferent bacteria and viruses [14]. If one layer is penetrated, another comes into
play, presenting the invader with progressively more complex and clever barriers.
We concentrate on the acquired component of the immune system here, specifi-
cally the humoral immune response that protects the body from bacterial infec-
tion. Cells of the body and invaders, Of antigens, are distinguished as two different
entities, one should be there, and one should not. The immune system achieves
this through the concept of self/nonself discrimination. Cells of the body define
self, anything else nonself. If it gets this wrong, either way, we are in trouble!
tem and are ideal for a hardware immune system. Many features are already ap-
plied to reliable system design. Embryonics has demonstrated one approach to
distributed fault tolerance by creating cellular electronic systems for example [lOl
If the layers of protection in the human body and existing methods of hardware
protection are compared as in Table 1, we find there is a gap that existing hard-
ware protection systems could potentially benefit from filling. One solution to
completing Table 1 is demonstrated with the development of immunological elec-
tronics, or immunotronics - the creation of an artificial hardware immune system.
Artificial immune systems take their inspiration from the operation of the human
immune system to create novel solutions to problem solving. Although stiH a rela-
tively new area of research, the range and number of applications is already di-
verse [16, 17]. Computer security, virus protection, anomaly detection, process
monitoring, pattern recognition, robot control, and software fauIt tolerance are
some of the applications artificial immune systems are being applied too. One im-
portant feature links alI of these applications - they operate in a software domain.
Our approach demonstrates that artificial immune systems can also exist in the
hardware domain [18].
11 Domain Mapping
12 Choice of Aigorithm
The negative selection algorithm is adopted for the hardware immune system for
two reasons:
• Complex detector set generation benefits a simple operational phase - ideal for
a hardware environment where reduced complexity simplifies the design, re-
duces component count, and promotes distribution throughout the hardware ar-
chitecture.
• Probabilistic detection permits a trade off between storage requirements and the
probability of failing to detect a nonself hardware condition. To cater for
changes in system functionality, the use of a reconfigurable platform such as a
field prograrnmable gate array (FPGA) enable the operation of a system to be
updated or completely changed. The elimination of rigid boundaries between
functional and protection is ideal, a requirement provided by probabilistic de-
tection.
ing re-synthesis using standard development tools to create varying sizes of mem-
ory depending on the desired storage space.
A demonstration system was synthesised for a Xilinx Virtex FPGA, 64 bits
wide, and 128 words deep to create the CAM organisation in Fig. 15. The archi-
tecture of the Virtex FPGA is ideal for constructing 4-bit CAMs using the Look-
up tables (LUTs) [26]. The LUTS are then connected together to create greater
width data. Parallel matching of aII tolerance conditions during operation ensures
single cycle error detection whatever speed the hardware system being protected
is running at. With no speed optimisations tumed on within the Xilinx Foundation
synthesis tools, the XCV300 device that contained the hardware immune system
was estimated to operate at 45 MHz. Considering operational speed for a custom
fabricated device, parallel HCAM searching ensures the system would operate at
the fully required speed of any system it was implemented to protect.
State machine Output~
(Self)
r:~
State
Fig. 14. Structure of the hardware immune system, incorporating the finite state machine to
be protected
112 A. Tyrrell
Masking logic
Data word O
Mask
Data word 1
Data
Found
Fig. 15. Architecture of the partial matching hardware content addressable memory
Partial matching capabilities were further added to the HCAM so that c con-
tiguous bit matching can be implemented rather than only complete string match-
ing. A bit string mask is added that allows each bit of the tolerance conditions to
be selectively inc1uded in a matching operation, or selectively set to a don 't care
condition. With the addition of the masking bits the demonstration system built al-
lows a 32-bit width, 128-words deep partial matching HCAM as shown in Fig. 15.
Matches are selectively made by selecting c contiguous bits to require a match at
any one time.
The synthesised hardware imrnune system was applied to a state machine based
decade counter, the architecture and results of which are shown in [21,27].
What is now required is to incorporate the two bio-inspired architectures into one
highly fault-tolerant system. The use of FPGAs for the implementation of both
ideas makes this easier; however, the limited resources (at least compared to biol-
ogy!) do present some implementation issues. One possible solution is shown in
Fig. 16. This uses two separate networks to transmit different types of data around
the system, in a similar way to the different comrnunication paths present in the
body.
15 Conclusion
The early history of the theory of self-replicating machines is basically the history
of John von Neumann's thinking on the matter [28, 29]. Von Neumann's automa-
ton is a homogeneous two-dimensional array of elements, each element being an
FSM with 29 states. In his historic work, von Neumann showed that a possible
configuration (a set of elements in a given state) of his automaton can implement
a universal constructor able to build onto the array any computing machine de-
Embryonics and Immunotronics 113
Immuno-embryonics
liiiii:iiiiiJI Embryonic Cell
It must be reminded that, in biology, the cel! is the smallest part of the living
being containing the complete blueprint of the being, the genome. On the basis of
this definition, von Neumann's automaton can be considered as a unicellular or-
ganism, since it contains a single copy of the genome, i.e. the description stored
on the tape. Each element of the automaton is thus a part of the cell, i.e. a mole-
cule. Von Neumann's automaton, therefore, is a molecular automaton, and self-
replication is a very complex process due to the interactions of hundreds of thou-
sands of molecules.
Independently of the direct biological inspiration of the embryonics project, the
multicellular structure of complex organisms is of enormous interest for the con-
ception of any large-scale electronic circuit. In fact, biological organisms are
probably the most obvious examples of extremely complex machines capable of a
stunning degree of fault-tolerance. Handling complexity and tolerating faults are
two of the main challenges in the design of the next generations of electronic cir-
cuits.
Embryonic arrays exploit hardware redundancy to achieve fault tolerance.
Spare elements are incorporated at different levels of the embryonics hierarchy,
achieving the resilience of organisms to faults in their constituent molecules and
cells.
114 A. Tyrrell
This work has also demonstrated that taking inspiration from the human im-
mune system, in the form of the negative selection algorithm is suitable for the de-
sign of novel error detection mechanisms for integration into reliable hardware
system. Error detection is performed in real time. Error detection is probabilistic,
permitting a trade-off between storage requirements and the ability to detect an er-
ror within the sequential system. In contrast to existing error detection techniques
that concentrate on single bit errors, and can sometimes fail to detect multiple er-
rors, the hardware immune system is adept to detecting at this task. Generation of
tolerance conditions is still implemented in software; the unique part of this work
is the demonstration of a hardware wrapper that provides fully embedded error de-
tection using principles from immunology. The immune system is also a separate
component, permitting integration in a variety of different systems, either built
with a hardware immune system in mind, or added at a later point. Results have
demonstrated the error detecting abilities of the hardware immune system on a
range of distinctly different benchmark finite state machines.
Work in the field of immunotronics is now progressing in the Bio-Inspired Ar-
chitectures research group at the University of York into the design of microproc-
essor immune systems and distributed hardware immune systems. Immunotronics
is also currently being integrated with other biologicaUy inspired architectures
such as embryonics and evolvable hardware as part of a European Commission's
Future and Emerging Technologies project. Information is available at
http://www.poetictissue.org .
References
L. Wolpert. The Triumph ofthe Embryo. Oxford University Press, New York, 1991.
2 M. Nicolaidis. "Future Trends in Online Testing: a New VLSI Design Paradigm?".
IEEE Design and Test ofComputers, 15(4), 1998, p. 15.
3 D. Mange, M. Tomassini, eds. Bio-inspired Computing Machines: Towards Novel
Computational Architectures. Presses Polytechniques et Universitaires Romandes,
Lausanne, Switzerland, 1998.
4 D. Mange, M. Sipper, A. Stauffer, G. Tempesti. "Towards Robust Integrated Circuits:
The Embryonics Approach". Proceedings ofthe IEEE, voI. 88, no. 4, April 2000, pp.
516-541.
5 C. Ortega, and A.M. Tyrrell "Design of a Basic CeH to Construct Embryonic Arrays",
IEE Proceedings - Computers and Digital Techniques, 145 (3), pp 242-248, May
1998.
Embryonics and Immunotronics 115
1 Background
The dawn of nanoscale science can be traced to a now c1assic talk that Richard
Feynman gave on 29 December 1959 at the annual meeting of the American
Physical Society at the California Institute of Technology'. In this lecture, Feyn-
man suggested that there exists no fundamental reason to prevent the controlled
manipulation of matter at the scale of individual atoms and molecules. Twenty one
years later, Eigler and co-workers [l] constructed the first man-made object atom-
by-atom with the aid of a scanning tunnelling microscope. This was just 7,000
years after Democritus postulated atoms to be the fundamental building blocks of
the visible world. The field derives its name from the SI prefix nano, meaning
, http://nano.xerox.comlnanotechlfeynman.html
118 C.J. McNeil, K.J. Snowdon
~ ~
."
~~
Direct inlerfoce 10 oeil Direct cell 10 inlerface
signalling signalling
Il
SIGNAL PRODUCTION ANO SIGNAL RESPONSE
(fluid, electrical and mechanica modalilies)
nology is not so much about a scale, but about the convergence of physics, chem-
istry and biology.
The properties and functionality of the molecular-scale interface separating the
biological and physical worlds are key to successful integration. Ideally, the prop-
erties and functions of this interface should be indistinguishable from that of
neighbouring cells on the one side, and fully compatible with fluidics, microelec-
tronics and information processing on the other. This means that it should not be
just biocompatible or bioactive, but it should also support controlled material and
information transfer across the physical/biological interface.
The goal of modem medical science is to identify disease at the earliest point in
the disease process, intervene before symptomatic disease becomes apparent, and
monitor the progress of both the disease and the effects of those intervention pro-
cedures. This demands the development of technologies capable of detecting
predisposal to disease conditions and the earliest signatures of emerging disease,
and supporting immediate, specific and highly targeted intervention. Suitable plat-
form technologies must integrate an ability to sense, communicate, respond, and
monitor. Nanotechnology, which is widely predicted willlead to the next techno-
logical revolution, with biomedical applications showing the largest potential for
growth, has the potential to provide enabling components toward fulfilment of
these goals. It thus has the potential to transform health care and medicine.
Nanotechnology increasingly enables controlled component design and fabrica-
tion on atomic and molecular scales. We now have an unprecedented ability to
construct, 'from the bottom up' truly nanoscale biological components (e.g. Fig.
2). Such nanoscale functional components can be self-assembled into larger, mul-
tifunctional systems and complex devices that integrate required functions with
the ability to encapsulate and protect a drug molecule, sense the presence of a de-
fective cell, communicate that presence, and respond. Furthermore, nanotechnol-
ogy enables the precise structural and functional characterisation of such compo-
nents and their interactions with each other and with their environment. In paralIel,
120 C.J. McNeil, K.J. Snowdon
Biological
ceL! ceL! } System
Topograpby }
Cbemistry '---...,..,...-1,.,.- ,..,...--=,.......,..
Protein Engineeriog I } Nanoengineered
1
Topograpby } Bio-interface
Chemistry ..:....:...:........:'-f:........:--'-':........:c....:.,-:........:-=....:.....:..:........:....;:..:..:........:...."+....:....~
Protern Eogineering
PhysicaJ
System
We are now faced with both the opportunity and challenge to integrate these
unprecedented developments in human knowledge across the life sciences, the
physical sciences and engineering, and so enable a step change in our ability to
address clinical priorities such as early, rapid, minimally invasive and highly tar-
geted intervention for cancer, immunological and genetic diseases and HIV. To
capture and exploit, in an effective and coherent manner, synergistic developments
associated with the post-genomic era are key priorities of all three UK Research
Councils (EPSRC, BBSRC and MRC). To reali se the potential offered by these
parallel developments requires not only research, but also the creation of a new
generation of researchers, who can work across traditional disciplines, and who
also know how to work with others at the inteifaces between disciplines. Bio-
medical nanotechnology seeks to shape the next generation of clinical therapeu-
tics, monitoring, diagtlostic and research tools through appropriate and effective
utilisation of nanotechnology, which many predict will be the next technological
revolution. This multidisciplinary approach could revolutionise patient care.
The specific aims of the Network are to seek to combine nanotechnology with
powerful molecular design and protein engineering techniques to develop nanoen-
gineered, robust, communicative biologic al interfaces to the physical and informa-
tion worlds. Since the technology must interact with the biological world on the
molecular and cellular scale, the structure and properties of that interface must be
designed and controlled on that scale. This is the realm of molecular design, pro-
tein engineerit1g, nanotechnology, and the fabrication of functional hybrid devices
and structures. These will be biocompatible, robustly attached to the non-
biological surface, and vossess application-specific components permitting mate-
rial and information transfer across the interface. The resulting platform technolo-
gies will enable us to hamess the full potential of genomic information through
real-time predictive, preventive, point-of-care and personali sed health care provi-
sion, and enable major and affordable advances in orthopaedics and trauma,
pharmaceutical screening devices, environmental and process monitoring, intelli-
gent communications systems, forensics and defence.
It is widely acrepted that the combination of nanotechnology and biomaterials
science will lead to the next technological revolution, with biomedical applica-
tions showing the largest potential for growth. Nanotechnology enables comvo-
nent design and fabrication on atomic and molecular scales and the self-assembly
of such components into larger, multifunctional systems. Furthermore, it enables
the precise structural and functional characterisation of such components and their
interactions with each other and with their environment. This Network will com-
bine macromolecular chemistry, nanotechnology, and powerful methods to re-
engineer proteins and other functional components derived from living systems.
This will result in techniques, devices and systems that can efficiently monitor
molecular interactions within and between cells, providing real-time single mole-
cule sensitivity and personali sed diagnostic tests, more rapid pharmaceutical
screening procedures, and prompt detection of biologic al contamination. The de-
velopment of techniques that would allow us to influence molecular interactions in
real-time would enable us to create e.g. orthopaedic implants able to respond to
and control bone formation. Common to all these examples is the need to create,
and in some cases communicate across, an interface between biologic al and physi-
cal environments.
The recent completion of the Human Genome Project has provided data on the
sequence of the approximately 30,000 human genes. This is a major achievement,
but is in reality merely an entry into the post-genomic age of applied genomics
and proteomics. Understanding the structure and function of these genes and the
proteins that they encode will potentially revolutionise our understanding of hu-
man disease. Pervasive real-time application of that understanding, which this pro-
ject would enable, will revolutionise treatment. The grand challenge identified
within our strategic vis ion appears tantalisingly within grasp following recent ad-
vances. Only with such technology can we interface the 'virtual information' of
the human genome with e.g. the physical reality of prediction, diagnosis and
treatment in medicine.
122 C.J. McNeil, K.J. Snowdon
The initial challenges for Network research activity have been identified as at-
tempting to integrate microsystems technologies, which offer revolutionary possi-
bilities for sensors and sensor arrays for drug screening, with nanoscale assembly
components and in integrating such components into ' smart' therapeutic delivery
vehicles (Fig. 3).
nuclear enhy
protection component
ofDNAfrom
endosomolytic
degradation
component
optimal size
for cellular uptake
condensing targeting
agent with optimal moiety
binding,
biocompatible protection from
opsonization
Fig. 3. Schematic representation of a nano-engineered multicomponent therapeutic delivery
system
It has been recognised that an increasing number of ceH membrane channels are
found to act as targets for medicines and other compounds. The ability to detect
the modulation of ceH membrane channel activity by the binding of therapeutic
agents is considered crucial for rational and efficient drug discovery and design.
Since combinatoriallibraries of potential therapeutic compounds are rapidly grow-
ing, fast and highly sensitive methods for functional drug screening are required.
An attractive possibility is the use of self-assembled tethered membranes contain-
ing specific channel receptors as the sensing element in an otherwise solid-state
biosensing device. Massive arrays of individually addressable microsensors with
integrated fluid hand ling are conceivable. Even very simple sensor designs offer
valuable advances in low-cost sensing for clinical medicine (especiaHy point-of-
care) and the food and hygiene sectors.
Biological sensor systems are predicted to play an important role in preventa-
tive medici ne and early diagnosis. The much-discussed 'artificial nose' containing
a dense array of receptor sites affording unambiguous identification of molecular
species could analyse the breath of patients for known chemical signatures of dis-
eases such as liver cirrhosis and lung cancer [4]. Applications of such devices as
Biomedical Applications of Micro and Nano Technologies 123
chemical sensors in industry, food processing, military combat situations and envi-
ronmental monitoring are promised.
Recent developments [e.g. 5] in scanning probe based microscopies (e.g. AFM,
STM, SECM) and molecular manipulation techniques such as 'optical tweezers'
[6] promise further advances in determining the structure of cell membranes on
the nanometre scale, providing a potential key to understanding processes such as
immune reactions and the development of viral infections. An ability to perform
controlled manipulation procedures on DNA, the creation of artificial enzymes,
proteins and ribozyme catalysts would open a wide range of speculative possibili-
ties. In addition, micro- and nanoscale device fabrication technologies have a po-
tentialIy important role to play in clinic al medicine, psychology and the behav-
ioural sciences in enabling the development of sophisticated, minimalIy invasive,
remotely addressable, yet affordable, sensors for quasi-real-time recording of neu-
rological activity and other biologic al functions.
FinalIy, genetic tests have almost unparalleled scope in medicine and biotech-
nology. With applications in genetic diagnostics in human and animal subjects and
in the detection of pathogens, the demand for genetic information is essentialIy
unlimited and determined only by the cost of information retrieval. The DNA se-
quence for a significant fraction of the 3.3 billion nucleotide base-pairs in the hu-
man genome has already been determined, in principle providing sufficient infor-
mation to measure the altered celIular patterns of expression of these genes in
different physiological conditions and disease states, including cancer. Rapid, in-
expensive methods of detection of these are eagerly sought, to enable clinicians to
identify agreat many pathological conditions with greater speed and certainty.
Dense microarrays of up to 100,000 oligonucleotides are coming into use in ge-
netic research, particularly in the pharmaceutical industry. Much effort is being
devoted to developing compact, inexpensive microarray-based systems which
could ultimately accommodate, on a single substrate, alI processing functions for
DNA amplification, separation, hybridisation and detection of picolitre sample
volumes.
5 Concluding Remarks
References
1. Eigler D.M., Schweizer E.K. (1990) Positioning single atoms with a scanning tunnel-
ling microscope. Nature 344,524-526.
2. Lakey 1.H., Reddy B.V., Murra)-Rust 1. (2001) Macromolecular assemblages. Theory
and simulation. Curr. Opin. Struct. Biol. 11, 139-140.
3. Davis B.G. (2002) Synthesis of glycoproteins. Chem. Rev. 102, 579-602.
4. Thaler E.R., Kennedy D.W., Hanson C.W. (2001) Medical applications of electronic
nose technology: review of current status. Am. J. Rhinol. 15, 291-295.
5. Blackley H.K.L., Davies M.C., Sanders G.H.W., Roberts CJ., Tendler SJ.B., Wilkin-
son M.J., Williams P.M. (2000). In-situ atomic force microscopy study of beta-
amyloid fibrillization. J. MoI. Biol., 298, 833-837.
6. Meiners J.c., Quake S.R. (2000) Femto-Nev/ton force spectroscopy of single ex-
tended DNA molesules. Phys. Rev. Lett. 84,5014-5017.
Macromolecules, Genomes and Ourselves
S. B. Nagl
Department of Biochemistry and Molecular Biology, University College London,
Gower Street, London WClE 6BT, UK
nagl@biochemistry.ucl.ac.uk
J. H. Parish
School of Biochemistry and Molecular Biology, The University of Leeds, Leeds
LS2 9JT, UK
R. C. Paton!
Department of Computer Science, The University of Liverpool, Liverpool L69
3BX, UK
G. J. Wamer
Unilever Research Colworth, Colworth House, Sharnbrook, Bedford MK44
lLQ,UK
1 Preamble
The advent of structural molecular biology and the development of rapid methods
for macromolecular sequencing coincided roughly with the development of digital
computers. Molecular biologists have thus been computer users for several dec-
ades. More recently computational molecular biology has been recognised as a
methodological discipline in its own right and has led to the neologism, 'bioinfor-
matics'. The subject has represented a productive cross-fertilisation of computa-
tional and biological ideas. Arguably it started with a computational solution to
the general problem of aligning strings with the concomitant insertion of gaps by
126 S. B. Nagl et al.
Biological macromolecules are certainly not the largest molecules in the world:
for example a diamond is just one molecule of carbon. However biological mac-
romolecules display a special type of complexity. The structure of a diamond is
based on a simple rule derived from the properties of a saturated tetravalent car-
bon atom. In contrast, aIthough biological macromolecules are based on relatively
simple chemical building blocks such as amino acids and nucleotides, their chemi-
cal properties are dominated by the chiral (asymmetric) nature of the building
blocks. Chemical interactions involving biologic al macromolecules are highly se-
lective and specific. This specificity extends to the interactions that form the rules
that govem the ways in which the macromolecular chains can fold to generate the
conformations (shapes in three dimensions) that are characteristic of these mole-
cules. Although some of the rules are still to be elucidated, one important gener-
alisation is valid: the complexity associated with emergence of nucleic acids and
proteins led to concomitant emergence of a new property - a type of chemically
encoded information. The information includes rules for the folding of a protein
(encoded by the protein sequence itself) and the information for the encoding and
regulated expres sion of genes in DNA (RNA in certain viruses). In this latter case
the proteins involved in the regulation and expression processes can be regarded
as decoding machines and the DNA (or RNA) as an equivalent of a program con-
taining instructions and data. In Sec. 3 we pursue the idea that proteins process in-
formation in the case of enzymic and other functions. Biological macromolecules
are involved in the regulation of cellular activity. Biologically they can be re-
garded as having several roles: structural, regulatory, catalytic and genetic. In this
context we use 'genetic' to refer to the inheritance, transmission and expres sion of
Macromolecules, Genomes and Ourselves 127
genetic information. We note that any one macromolecule can fulfil more than one
of these roles, not simply because of possible ambiguity in the definitions of the
roles (the last three are ali examples of informat ion processing) but a single mac-
romolecule can have different roles: for example, ribosomal RNA is a component
in gene expression, is structural and is an enzyme (catalyst). In this chapter we ex-
clude certain classes of macromolecule, polysaccharides, lipids, lipopolysaccha-
rides, and concentrate chiefly on proteins but in this introduction we put them into
context with DNA and RNA. A biochemist's summary of the roles of these mole-
cules is given in Table 1.
However, when we say that DNA contains what is necessary to specify a living
being, we divest this ceH component of its interrelations with the rest of the net-
work. It is the network of interactions in its entirety that constitutes and specifies
the characteristics of a particular ceH, and not one of its components. That modifi-
cations in those components called genes dramatically affect the structure is very
certain. The error lies in confusi.lg essential participation with unique responsibil-
ity [6]. This has implications for biomedical science and we return to this topic in
the final section.
The sequences of many proteins are known and the majority of these sequences
have been deduced from DNA sequencing studies, notably from the many genome
projects. Several databases are non-redundant, in other words proteins that are
identical or almost identic al in closely related species (e.g. human beings and
chimpanzees) are only represented once. Some proteins are of known structure
and others are of known function (these two sets intersect). Protein sequence data-
bases contain annotations about the source of the protein and such structural and
functional details as are known or have been deduced.
The PIR-International Protein Sequence Database (PSD) [7]is maintained by an
international consortium comprising the Protein Information Resource (PIR) in the
United States, Japan and Germany. As of September 2003 it contained over
142,000 annotated, non-redundant protein sequences. NRL-3D [8] is a supplemen-
tary database to PIR and is produced from sequence and annotation information
extracted from the Brookhaven Protein Databank (PDB) of three-dimensional
structures [9].
SWISS-PROT [10] is a protein sequence database and is the result of coHabora-
tion between the Department of Medical Biochemistry at the University of Geneva
and the European Molecular Biology Laboratory (EMBL). Translated EMBL
(TrEMBL) is a computer annotated supplement to SWISS-PROT in 1996 [l0]
From these several composite databases have been constructed.
vary enormously from protein to protein (or indeed RNA to RNA). In other words
if, in some sense, the evolution of an organism is a summation of the evolution of
its macromolecules, the 'macroscopic' rules (for the cell, the organism) are con-
strained in a different way from the 'microscopic' rules for the molecules.
By using either computational methods for structural three-dimensional alig-
ment, automated analysis of elements of the structure, inspection by human ex-
perts, or a combination of these methods there are three alternative classifications
of proteins of known structure and/or function. SCOP [11] organises proteins in a
hierarchy from class, down through fold and superfamily to family. CATH [12]
employs a hierarchical classification system and a greater degree of automation.
Proteins are grouped on the basis of sequence similarity and a representative of
each group is taken and divided into domains using three automatic domain-
assignment techniques. We note that biochemists use the word 'domain' to mean a
module of a protein structure that they know or believe to be partly independent.
FSSP [13] also relates protein structures on an evolutionary basis, but is a fully
automated procedure and does not assign proteins to classes, folds or families. Ex-
haustive pair-wise structure comparisons are made and the results represented as a
fold tree, which is generated by hierarchical clustering, and as a series of structur-
ally representative sets of folds at varying levels of uniqueness.
A signature is some sort of restricted part of a protein sequence that relates the
protein to a position in a structural/functional classification. A short signature in
this sense is referred to as a 'motif. One approach is to rely only on the sequence
as a source of the signature. In this approach the sequences can be aligned or ana-
lysed in some way to create the signature. Such signatures can be used to construct
'secondary databases'. PROSITE [14] was one of the first secondary databases to
be developed. The rationale behind the content was that the single most conserved
motif observable in a multiple alignment of known homologues could be used to
effectively characterise a protein family, the motifs themselves usually encoding
vital biological functions. The motifs are represented as regular expressions. Thus,
by searching the database, a new sequence could be allocated to a family, or the
domains or functional sites that the sequence contained may be elucidated.
The use of several motifs to characterise a protein family is preferable to using
just one because of the possible benefit to diagnostic performance. For example, a
protein fingerprint [15] offers greater diagnostic reliability over single-motif
methods because if a query sequence fails to match all the motifs in a given fin-
gerprint, the remaining motifs will stiH allow a fairly confident diagnosis. The da-
tabase is called PRINTS-S.
Profiling provides a means of detecting distant sequence relationships in cases
where regular expressions wiH not provide very good discrimination. The com-
plete sequence alignment effective1y becomes the discriminator; the profile is
weighted to indicate where insertions and deletions are allowed, which residues
are allowed at each position and where the most conserved regions are. The pro-
files can be encoded in the form of Hidden Markov Models (HMMs), con si sting
Macromolecules, Genomes and Ourselves 131
of linear chains of match, delete Of insert states which can be used to represent the
sequence conservation within aligned families. An advantage of HMMs is that
they are not limited to searches for motifs. Pfam [16] is a database of protein do-
main families containing curated multiple sequence alignments for each family
and a collection of HMMs for finding these domains in new sequences. In Pfam-A
a collection of hand-edited seed alignments is used to build HMMs using the
HMMER package, to which sequences are automatically aligned to generate final
full alignments. If these do not produce diagnostic HMMs the process is iterated
until a good result is achieved. The seed and full alignments are then coupled with
minimal annotations, database and literature cross-references, and the HMMs
themselves.
An alternative approach is to use as a starting point detailed analysis of those
proteins whose structures are known and thus to generate sparse signatures. Sev-
erai methods [17-21] for the sparse representations of secondary structures are
published. Daniel et al. [22] reali sed that such signatures could be very fuzzy se-
quence-Iength 'motifs' but as these could not be represented as regular expressions
a new alignment algorithm was needed.
INPUT
PATTERN OUTPUTS
substrates
~
..
~
catalyse
-.-...
Pattern Processl
~ Recognition Compute switch
~
~ ... modulate
~ ... integrate
The idea of ceH-as-text has been used to address questions related to reduction be-
tween (say) cytology and molecular biology [25, 26]. Many tools in molecular bi-
ology, both experimental and conceptual devices, have the same effect, namely
they dissect, cut open and reduce the ceH as a whole system into bits and pieces.
An important conceptual anei maybe also perceptual challenge is how to think
about the integrative whole instead of or as a complement to the bags, chips, gels
and spectra of the parts. This section will address these issues with an examination
of the metaphors of cell-as-text and cell-as-ecology.
Proteins exhibit sophisticated information processing capacities. For example,
enzymes can display pattern recognition, memory capacity, context-sensitive ac-
tivity, handling of fuzzy internal and external events, switch-like behaviour, inte-
gration of a number of metabolic pathways and other processes and signal ampli-
fication. It has been argued elsewhere [27] that it is possible to think of enzymes
as playing the central role of verbs in the cellular metabolic and information proc-
essing systern. Like verbs, enzymes can be said to have cases (in the sense elabo-
rated by Fillmore [28]). Within the context of this linguistic metaphor, enzyme
cases would include substrate, product, regulator(s), locations, associations, co-
agent(s) and target site(s). Certain enzymes that exhibit the mood- or voice-like
properties of verbs can be related to their context-sensitivity and also to their in-
ternal configuration and locali sed interactions. The notion of modality in enzyme
action is implicit in a number of recent descriptions including: the fluctuating en-
zyme [29]; the seed-germination model [30] and enzymes as logical agent/verb
[26].
A number of glycolytic enzymes are sensitive to the micro-environments and
ceH types in which they are found. For example hexokinases, which catalyse the
phosphorylation of hexose sugars, come in various isoforms including brain
hexokinase (BHK) and glucokinase (liver). Certain metabolic conditions affect
BHK behaviour with rapid and reversible changes between soluble and particulate
Macromolecules, Genomes and Ourselves 133
fonns in which the latter is more active than the fonner. Another context-sensitive
glycolytic enzyme is phosphofructokinase-l (PFK-l) which catalyses the phos-
phorylation of fructose 6-phosphate (F6P) to fructose 1,6-bisphosphate (Fl,6BP).
In this way PFK-l exhibits sigmoidal kinetics when in free solution but a nonnal
saturation curve when membrane-bound [31]. A special case of a context-sensitive
enzyme that has 'voices' is 6-phospho-fructo-2-kinase/fructose-2,6-
bisphosphatase (6PF2K1F2,6BP). This enzyme catalyses two opposing reactions:
F6P+ ATP F2,6BP+ ADP
F2,6BP F6P+ ATP
As Pilkis el al. [32] note, 6PF2K1F2,6BP has, in addition to a catalytic role, a
key function in intracellular signalling.
Intracellular signalling systems employ inter-communicating networks of
kinases and phosphatases. These enzymes are often multi-functional and highly
integrative. Consider calmodulin-dependent protein kinase II (CaM Kinase II), a
large multimeric and multifunctional enzyme that is derived from four genes. It
acts on upwards of 49 substrates and is very common. CaM Kinase II exhibits a
memory capacity as well as being able to amplify signals. Four functional do-
mains are defined: catalytic, regulatory (inhibitory and CaM binding regions), as-
sociation (with other subunits) and variable (for targeting and localisation). Apart
from the variable domain, the other three domains are highly conserved.
An examination of the eukaryotic transcription factors CBP and p300 provide
further insights into 'glue' relations. CBP/p300 are large multi-functional proteins
that participate in various basic cellular functions, including DNA repair, ceH
growth, differentiation and apoptosis. They act as focal points for multiple pro-
tein-protein interactions and co-activate many other transcription factors including
CREB, nuclear receptors, signal transducer and activator of transcription (STAT)
proteins, p53, and the basal transcription proteins. Using the review by Giles el al.
[33] it is possible to talk about CBP/p300 acting like 'glue' in five ways that relate
to (1) molecule-molecule bindings and interactions, (2) enzymatic processes, (3)
as a physical bridge between various transcription events, (4) acting as histone
acetyltransferases (HATs) -linking transcription to chromatin remodelling - and
(5) mediating negative and positive crosstalk between different signalling pathways.
It is important to note that, from our point of view, this 'glue' is not just that the
molecules have intrinsic adhesive properties, they also provide the ceH with com-
binatorial and cohesive properties at a functionallevel.
Fig. 2 summarises some of the relations between 'gluings' and function for
CBP/p300. Object and process can be subsumed as one general tenn 'glue' with
respect to the multi-functionality of these proteins. What is more, it is possible to
introduce the topological thinking of local ~ semi-Iocal ~ global into the context
in which the 'glue' functions. This relates to the tenns in the boxes on the left of
the diagram. Many verbs can be related to CBP/p300 action, such as bind, interact,
process, bridge, act, link, transcribe, remodel, mediate, talk and signal. The con-
textual and 'gluing' descriptions in Fig. 2 can be used to organise function descrip-
tions in tenns of integration within the hierarchy (context) and interaction across
levels (glue).
134 S. B. Nagl et al.
Crosstalk
Regulation
Bridging
HAT
Fig. 2 is a network that summarises the gluings associated with a family of pro-
teins that participate in several cellular activities. Such a convention could be used
for describing several cases: the complex regulation of nitrogen assimilation in en-
teric bacteria involves glutamine synthase which is not only an enzyme but is in-
volved in the regulation of its own biosynthesis or the pCI protein of the bacterial
virus lambda which operates as an repressor of the expression of certain genes in-
cluding its own and is also an activator of its own expres sion. Multifunctionality
can arise in multidomain proteins (Sec. 2.3) in which the domains have different
functions. An example of medical importance is the thymidylate synthase-
dihydrofolate reductase protein in the malarial parasites. This protein which arose
from a gene fusion associates two different enzyme functions with adjacent steps
in a metabolic pathway required for DNA synthesis. In the next section we address
the problem of more complex systems. Ultimately descriptions such as Fig. 2 may
provide a formal representation. At present the data set is far too incomplete: there
are many examples of multifunctionality yet to be discovered and therefore we use
a more coarse grained model.
We have entered the 'post-genome era' and, as techniques for the investigation of
cells at the systems level become more and more refined, are witnessing a revolu-
tionary reorientation in the conceptual foundations of biomedicine. Questions
about the molecular nature of disease can now be asked at the level of differential
activation states of the genome 'transcriptomics' [34], populations of gene prod-
ucts, including splice variants and post-translationally modified versions 'pro-
teomics' [35, 36], or functional interactions in large networks (for example [37]
Macromolecules, Genomes and Ourselves 135
and Fig. 3). In complementary fashion, the structural organisation of cell systems,
from single biomolecules, to molecular complexes, cellular compartments, and in-
dividual cell types, is also being revealed at greater and greater resolution.
Cells are increasingly coming to be seen as systems of interacting structures
with information-processing capabilities. Some of the novel questions rai sed are,
for example: What can global gene expression pattems in model organisms teach
us about the order and logic of the genome? How do genomes and other biologic al
systems evolve? How do large-scale networks of molecular interactions integrate
biologic al signals within cells? How does cellular localisation affect protein func-
tion? It is to be expected that, in the future, the systems properties of biomolecules
will also become a major focus for new research.
"Systems thinking" goes hand in hand with a focus on information processing.
Ideally, explanations are sought as to how the spatial and temporal organisation of
mixed populations of molecules gives rise to the 'smart' properties of biological
systems. Such "smartness" is not restricted to high-level processes like conscious-
ness, but is in evidence at allievels of biological organisation. Currently, much in-
terest is directed towards signal processing by complex interaction networks (Box
1). At the level of single proteins, diverse processes such as signal-mediated acti-
vation of proteins by conformational change, or the parallel-distributed nature of
co-evolutionary mechanisms (Box 2), Can also be approached from this perspec-
tive.
Whilst all biological processes are consistent with the physical and chemical
laws of our universe, and in this sense Can ultimately be 'reduced' to chemistry
and physics, there is a growing awareness that biological phenomena require an
approach that equally addresses the problem of emergence. How do living systems
emerge from the laws of physics and chemistry? In complex systems, emergent
phenomena result from the rule-govemed, non-linear interactions of a large num-
ber of components occurring in a highly context-dependent manner (Box 3). To
come back to the example of consciousness, it arises out of the densely connected
interactions of billions of neurones (and their constituent molecules), and is not a
property of any one brain region, let alone of the neurones themselves. Conscious-
ness is an emergent property of the brain as a whole. Beyond this special case, in-
formation processing can be seen as an emergent property of complex biological
systems in general. One enormous task before us then is the identification of the
rules of the interactions that are embedded in the organisation of living matter.
136 S. B. Nagl eal.
~----~I
Genomics
',,,1,,,,
~----------~
Genome activation patterns:
transcriotomics
Protein population:
Proteomics Molecular
Pathway/network analysis
Organisation:
tissue cel! complex structure
imaging
EM
X-ray, NMR, structural ge-
nomics
integration
Fig. 3. Analysing and modeling complex cellular systems. The global state of complex cell
systems can be investigate at different levels of description, e.g. genome, proteome, cellular
structure, pathway or network. Novel bioinformatics techniques, notably for systems mod-
eling and data mining, are required to integrate the data obtained into models of complex
cellular systems
Macromolecules, Genomes and Ourselves 137
Box 1.
Box 2.
Box 3.
many cellular functions, and therefore are extremely relevant to the study, diagno-
sis and treatment of disease. Consequently, interpretation of the data obtained by
genomic technologies should be sought on this premise. It follows that new com-
putational techniques, able to analyse, model and represent different aspects of
cellular complexity, ought to be a high priority for this new phase of medicine
(Fig. 3).
There has been great excitement in the air, horizons appear to be opening up for
a new way of thinking about biology and medicine, whilst we are also deeply
aware of the huge challenges, both of an empirical and theoretical nature, in bring-
ing this about [38]. A theoretical framework and methodology for the investiga-
tion of complexity and emergence in biology are stilliargely undeveloped. Whilst
the study of complex systems has undergone vigorous expansion over the last dec-
ade, with contributions from a wide range of disciplines, biomedicine has so far
remained almost untouched by these developments. A 'systems vision of life' will
need to apply and extend this knowledge within a biologic al context.
Model choice, has an explicit ethical dimension that scientists have a responsi-
bility to confront. A critic al awareness of which kind of knowledge, and which
kind of goals, are likely to be facilitated by the chosen mode of representation
ought to inform any modelling project.
Models are in a certain sense metaphorical constructions [44]. As such, they carry
with them not only explicit messages but also implicit content. A model is a de-
vice for seeing the world in a particular way. A well-developed scientific model
accumulates a complicated assortment of techniques, interpretations, standards of
proof, and so on; and may well have a cognitive impact far transcending the origi-
nal context in which it was conceived. Much of this remains unwritten, but is un-
derstood by everyone who has been sociali sed within the research tradition associ-
ated with the model.
Metaphors occupy a central place in scientific discourse. It has become increas-
ingly recognised that metaphors, models, and language in general, play a central
role as the means by which 'raw data' are shaped into scientific concepts. This re-
alisation exposes the multiple layers of mediation between 'nature' and the 'sci-
ence of nature': quantitative and qualitative perceptions, descriptions of those per-
ceptions, choices shaped by beliefs and the need to create contextual meanings,
interpretations, and the representations ehosen in alI those mediation steps [45]. A
substantial body of scholarship (for example [46-48]) has described the ways in
which scientists employ metaphors as means to relate the empiric al world to ab-
stract theories. The powerful role of metaphors becomes even clearer when one
realises that they not only perform this explanatory function, but also serve as the
'basic organizing relation of a paradigm' [39]. In other words, metaphors are the
devices which organize the 'entire constellation of beliefs, values, techniques, and
so on shared by the members of a given [seientific] community' [49].
Furthermore, metaphors perform a vital role in creative thought; note, for ex-
ample, the rhetorical question posed by Gould [50] whether 'any brilliant insight
has ever been won by pure deduction, and not by metaphor or analogy'. Another
essential role of metaphors in scientific innovation, especially in biology, should
not be forgotten either. At the present state of scientific communication, meta-
phors are often the only language tools available if one intends to convey pattems
of relationship as the consequential parameters, rather than reductionist, hierarchi-
cal descriptions of cause and effeet [51]. Biology, which aims to explain phenom-
ena of immense complexity, relies strongly on metaphoric explanations. In fact,
when faced with hugely complex phenomena, a metaphorical explanation with a
high degree of fit can be the most 'realistic' description attainable [52]. Judging
from these observations, models/metaphors ean be said to convey facts, help to
eoncretise abstract theories, structure paradigms, as well as being essential devices
for scientific creativity.
Metaphors also convey cultural messages as well as facilitating communication
in science. In fundamental ways, any kind of representation is linguistically con-
structed; we can only know something about the world and about ourselves
Macromolecules, Genomes and Ourselves 143
through language. For example, on the level of primary data representation, design
decisions about genome databases determine what uses can be made of the data -
what can be compared with what. Further interpretation of the data and the dis-
semination of this information also depend crucially on the medium of language.
On a societal scale, 'science' and 'culture ' continuously create and re-create each
other through language. This traffic of ideas, images, metaphors and theories
about "nature" and 'human nature' is bidirectional. It is at this junction point that
wider conceptual frameworks, often only implicit in the linguistic representations,
exert their constraints by enabling certain representations but not others. In this
way, metaphors as 'culturally inherited and linguistically reinforced concepts' [53]
play a tremendously important role in the ongoing transformation of our views of
reality and ourselves.
Fig. 5. A neural network model of the evolution of new functional sites in protein domain
families. The network architecture is that of a c1assic feedforward network whose size can
vary depending on the coevolutionary network to be modelled (closed arrows). Sequence
positions (agents) function as fully connected processing elements (squares) . Each agent is
represented as a binary vector (open arrows). Amino acids are encoded as bitstrings. A het-
ero-associative mapping is performed that maps the input vector matri x (agent states in the
functional site) to the output vector matrix that ranges over a different vector space (states
of coevolving agents). After training, the neural network model encodes the state transition
rules of the coevolutionary network
It is obvious that there are choices to be made. To follow Keller [43], what are
we seeking representations of, and what are we seeking representations for? In a
very fundamental way, the choice of the models we work with, and the choice of
language and metaphor in which we express these ideas, constitute ethical choices
about our attitudes and actions toward other persons and the world. But this wilI
be an on-going reflective process. For what is being proposed here is not simply a
replacement of one set of models/metaphors with another, but recognition that sci-
entists are continuing to construct a discourse with wide ranging implications.
Such awareness demands a stance of self-reflectivity in the inquirer herself or
144 S. B. Nagl etal.
himself. This self-reflectivity aims to make transparent the values and motivations
that inform the research matter, process and goals. As we take up the challenges of
creating new concepts and models for post-genomic biomedicine, we would do
well in probing their implicit metaphoric content. As a community, we might con-
sider it worthwhile to critically explore the possible consequences of adopting cer-
tain conceptual frameworks over others from the widest perspective possible, en-
compassing scientific, medical, social, and cultural dimensions.
Post-genomic biomedicine promises to bring about revolutionary changes in di-
agnosis and intervention, accompanied by a profound transformation in our under-
standing of human bodies, health and disease. There are urgent demands for new
modelling methods and new computational tools to speed up the exploitation of
the huge amount of data becoming available. However, as the directions we em-
bark on now most likely will have significant consequences far into the future, one
may wish to resist these pressures to some extent to allow space for reflection be-
yond the immediate pragmatic challenges. Post-genomic biomedicine may benefit
from a new approach to modelling that includes an ongoing exploration of the
many levels of meanings and prescriptions for action that are inherent in the mod-
els we create.
5 Conclusions
Any scientific or other academic discipline can be viewed from either a historical
or more abstract perspective. In the case of molecular and cell biology the histori-
cal roots are in medicine and the study of human disease, classical biology and
biologic al chemistry. Some of these origins are reflected in the current ways in
which the science is practised today. For example classical biology or natural his-
tory was descriptive, partly anecdotal and led to a preoccupation with the idea of
classification that still dominates aspects of biochemistry (Sec. 2.3). There was an
idea in the relatively early days of molecular biology that, in some sense, the task
of the molecular biologist was to reduce complex biological phenomena to phys-
ico-chemical descriptions. However (Sec. 4.2) currently the complexity of bio-
logical systems is seen differently.
Of several approaches to studying the complexity of natural systems, the semi-
otics of Chandler [54], highlights the new properties that emerge in the progress
from the classes of Subatomic particles, through Atoms, Molecules, Biomacro-
molecules, Cells, Ecoment to Environment. (The words in italics are Chandler's
own names for these classes.) A modem view of biology is that properties includ-
ing reproduction, adaptation, leaming and information processing emerge with in-
creasing complexity. It is beyond dispute that the Environment consists of Sub-
atomic particles but there is no feasible method whereby a detailed study of such
particles could predict the human behaviour that has led to global warming, the
nature of the remarkable population of types of fish in Lake Victoria ....
Computing and biology have a relationship which biologists might regard as an
example of mutualism. Molecular biology would be impossible without computa-
tional analysis of data (Sec. 2) and ideas derived from theoretical considerations
Macromolecules, Genomes and Ourselves 145
provide valuable metaphors for both molecules (Sec. 3) and complex cellular sys-
tems (Sec. 4). An integrative description is at present difficult for reasons stated at
the end section of 3.2 although it is arguable that those practising bioinformatics
might make a major contribution by addressing this point.
Our final point is to comment on the biology-to-computing component of the
mutualistic relationship. We argue that the emergence of properties represented by
the system of Chandler [54] or, altematively, as the process of evolution should
provide novel metaphors in computing. The criterion here is simple but strict: if
the emergence of biological complexity provides useful algorithms, the develop-
ment of such algorithms can become independent of their inspiration. A geneticist
faced with a description of a genetic algorithm (or genetic program) might com-
ment critically on the fact that many facets of actual genetic change are either
omitted or misunderstood. That is irrelevant if the algorithm is efficient and use-
fuI. Rigorous and robust methods inspired by the development of biological com-
plexity can justify themselves and, even if they deviate from the facts of their in-
spiration, should facilitate the analysis of the emergence of properties in more
complex systems.
References
9. Berman, H. M., Westbrook, 1., Feng, Z., Gilliland, G., Bhat, T. N., Weissig, H., Shind-
yalov, 1. N. & Bourne, P. E. (2000). The Protein Data Bank. Nucleic Acids Research
28, 235-242.
10. Bairoch, A. & Apweiler, R (2000). The SWISS-PROT protein sequence database and
its supplement TrEMBL in 2000. Nucleic Acids Research 28, 45-48.
11. Lo Conte, L., Ailey, B., Hubbard, T. J. P., Brenner, S. E., Murzin, A. G. & Chothia, C.
(2000). SCOP: a Structural Classification of Proteins database. Nucleic Acids Re-
search 28,257-259.
12. Orengo, C. A., Michie, A. D., Jones, S., Jones, D. T., Swindells, M. B. & Thomton, J.
M. (1997). CATH - a hierarchic classification of protein domain structures. Structure
5, 1093-1lO8.
13. Holm, L. & Sander, C. (1996). The FSSP database: Fold classification based on struc-
ture structure alignment ofproteins. Nucleic Acids Research 24,206-209.
14. Hofmann, K., Bucher, P., Falquet, L. & Bairoch, A. (1999). The PROSITE database,
its status in 1999. Nucleic Acids Research 27, 215-219.
15. Attwood, T. K., Flower, D. R., Lewis, A. P., Mabey, J. E., Morgan, S. R, Scordis, P.,
Selley,1. N. & Wright, W. (1999). PRlNTS prepares for the new millennium. Nucleic
Acids Research 27,220-225.
16. Bateman, A., Bimey, E., Durbin, R, Eddy, S. R, Howe, K. L. & Sonnhammer, E. L.
L. (2000). The Pfam protein families database. Nucleic Acids Research 28, 263-266.
17. Michie, A. D., Orengo, C. A. & Thomton, J. M. (1996). Analysis of domain structural
class using an automated class assignment protocol. Journal of Molecular Biology
262, 168-185.
18. Russell, R B., Copley, R R & Barton, G. 1. (1996). Protein fold recognition by map-
ping predicted secondary structures. Journal of Molecular Biology 259,349-365.
19. Wamer, G.J, Ison, J.c. & Parish, J.H., (1998). "Protein fold recognition from secon-
dary structure signatures", CCPll Newsletter Issue 6 2.4. Available as
www.hgmp.mrc.ac.uldCCP111newsletter/voI2_4lccpll_article/fuICarticle.html
20. Zhang, C. T. & Zhang, R (1998). A new criterion to classify globular proteins based
on their secondary structure contents. Bioinjormatics 14, 857-865.
21. Zhang, C. T. & Zhang, R (1999). A quadratic discriminant analysis of protein struc-
ture classification based on the helix/strand content. Journal of Theoretical Biology
201,189-199.
22. Daniel, S.C., Parish, J.R., Ison, J.c., Blades, M.J. & Findlay, J.B.C. (1999) Alignment
of a sparse protein signature with protein sequences: application to fold prediction for
three small globulins. FEBS Letters 459, 349-52.
23. Changeux, J. (1965) The control of biochemical reactions. Scientific American 214,
36-45.
24. Paton, RC., Staniford, G. and Kendall, G. (1996) Specifying logical agents in cellular
heirarchies. In Computation in cellular and molecular biological systems (eds.
Cuthbertson, R, Holcombe, M. & Paton, RC.), World Scientific: Singapore, pp. 105-
119.
25. Albrecht-Buehler, G. (1990), In Defense on 'Non moleculuar' Cell Biology, Interna-
tional Review of Cytology, 120, 191-241.
26. Paton, RC. (1997), Glue, Verb and Text Metaphors in Biology, Acta Biotheoretica,
45,1-15.
27. Paton, RC. & Matsuno, K. (1998), Some common yhemes for enzymes and verbs,
Acta Biotheoretica, 46,131-140.
Macromolecules, Genomes and Ourselves 147
28. Fillmore, C.1. (1968), The Case for Case, In Universals in Linguistic Theory (eds.
Bach, E. & Harms, R.T.), HoIt, Rinehart & Winston, pp. 1-88.
29. Welch, G.R. & Kell, D.B. (1986). Not just Catalysts - Molecular Machines in Bioener-
getics. In The Fluctuating Enzyme (ed. Welch, G.R.), John Wiley, pp. 451-492.
30. Conrad, M. (1992), The seed germination model of enzyme catalysis, BioSystems 27,
223-233.
31. Uyeda, K. (1992), "Interactions of Glycolytic Enzymes with Cell Membrane",
Curent. Topics in Cell Regulation 33,31-46.
32. Pilkis, S.1., Claus, T.H., Kurland, 1.J. & Lange, AJ. 1995, "6-Phosphofructo-2-kinasel
Fructose-2,6-Bisphosphatase: A metabolic signalling enzyme", Annu. Rev. Biochem.,
64,799-835.
33. Giles, R.H., Peters, D.1.M. & Breuning, M.H. (1998), Conjunction dysfunction:
CBP/p300 in human disease, Trends in Genetics, 14,178-183.
34. Blohm, D.H. & Guiseppi-Elie, A (2001). New developments in microarray technol-
ogy. Current Opinion in Biotechnology 12, 41-47.
35. Williams, K. L. (1999). Genomes and proteomes: Towards a multidimensional view of
biology. Electrophoresis 20, 678-688.
36. Pandey, A & Mann, M. (2000). Proteomics to study genes and genomes. Nature 405,
837-846.
37. Rain, J.-C. Selig, L., De Reuse, H., Battaglia, V., Reverdy, C., Simon, S., Lenzen, G.,
Petel, F., Wojcik, J., Schachter, V., Chemama, Y., Labigne, A & Legrain, P. (2001).
The protein-protein interaction map of Helicobacter pylori. Nature 409, 211-215.
38. Nagl, S. B. (20ooc). Protein evolution as a parallel-distributed process: A novel ap-
proach to evolutionary modelling and protein design. Complex Systems 12:261-280.
39. Nagl, S. B. (2oo0a). Science and moral agency in a complex world. Paper presented at
the 5th World Congress of Bioethics, Imperial College, London, September 2000.
40. Nagl, S. B. (2000b). Neural network models of protein domain evolution. HYLE-An
International Journal for the Philosophy of Chemistry 6:143-159. An electronic ver-
sion of the paper is available at
41. http://www.uni-karlsruhe.de/-edOllHylelHyle6/nagl.htm.
42. Beauchamp, T. L. and Childress, J. F. (1989). Principles of Biomedical Ethics, Oxford
University Press, Oxford, p. 120.
43. Kemp, M. and Wallace, M. (2000). Spectacular Bodies: The Art and Science of the
Human Body from Leonardo to Now. Hayward Gallery Publishing, p. 24.
44. Keller, H. (1992). Secrets oflife, secrets ofdeath, Routledge, p.5.
45. Holland, J. H. (1998). Emergence, Addison-Wesley, New York. p. 207.
46. Spanier, B. B. (1995) ImlPartial Science, Indiana University Press.
47. Leatherdale, W H. (1974) The Role of Analogy, Model and Metaphor in Science, North
Holland.
48. MacCormac, E R. (1976). Metaphor and Myth in Science and Religion, Duke Univer-
sity Press.
49. Hesse, M. (1996). Models and Analogies in Science, University of Notre Dame Press.
50. Kuhn, T. S. (1970) The Structures of Scientific Revolutions 2nd edition, University of
Chicago Press, Chicago, Ill., p. 175.
51. Gould, S. J. (1996) Why Darwin? The New York Review of Books 4 April, 1996: 10-
14, p. 10.
148 S. B. Nagl et al.
52. Thaler, D. S. (1996). Paradox as path: pattern as map. In The Philosophy and History
of Molecular Biology: New Perspectives, (ed. Sahotra Sarkar), Kluwer, Dordrecht, pp.
233-248.
53. Depew, D. J. & Weber, B. H. (1996). Darwinism Evolving: Systems Dynamics and the
Genealogy of Natural Selection. First paperback edition, MIT Press, Cambridge,
Mass." pp. 21-30.
54. Margulis, L. and Sagan, D. (1995). What is Life?, Weidenfeld & Nicolson Ltd, p. 41.
55. Chandler, J.L.R. (1996) Complexity ill. Emergence. Notation and symbolization.
WESScomm 2,34-37.
56. Nagl, S. B. (2001). Can correlated mutations in protein domain families be used for
protein design? Briejings in Bioinformatics. In press.
57. Cilliers, P. (1998). Complexity & Postmodernism, Routledge, p. 3.
Models of Genetic Regulatory Networks
H. Bolouri, M. Schilstra
Hamid Bolouri
Institute for Systems Biology, Seattle, WA 98103-8904, USA
Division of Biology 156-29, California Institute of Technology, CA91125, USA
H.Bolouri@herts.ac.uk
Abstract. This chapter provides a short review of the modelling of Genetic Regu-
latory Networks (GRNs). GRNs have a basic requirement to model (at least) some
parts of a biological system using some kind of logic al formalism. They represent
the set of alI interactions among genes and their products for determining the tem-
poral and spatial patterns of expres sion of a set of genes. The origin of modelling
the regulation of gene expression goes back to the Nobel-prize winning work of
Lwoff, Jacob and Monod on the mechanisms underlying the behaviour of bacte-
rial viruses that switch between so-called lytic and lysogenic states. Some of the
circuit-based approaches to GRNs such as the work of Kauffman, Thomas, and
Shapiro and Adam are discussed.
A Genetic Regulatory Network (GRN) is simply the set of all int~ractions among
genes and their products determining the temporal and :,patial patterns of expres-
sion of a set of genes. Figure 1 is a cartoon of some of the major interactions that
may be included in a GRN. These include: mRNA transcription, transport out of
(and later into) the cell nucleus, protein synthesis, and protein-protein interactions
within the cytoplasm (including interactions with signaling pathways). Since
genes underlie all aspects of evolution, development and physiology, the topic as
a whole is clearly well beyond the scope of this chapter and its authors! Luckily,
there are numerous very good textbooks that cover specific aspects of GRNs.
The biology of gene structure, regulation and function is beautifully described
in a number of textbooks, e.g. [1,2]. The significance, operation and evolution of
genetic regulatory networks are admirably described in recent books by Eric
Davidson [3] and Sean Carroll et al. [4] For the role of GRNs in development, two
classic textbooks are Scott Gilbert's Developmental Biology [5] and Cells, em-
bryos, and evolution, by John Gerhart and Marc Kirschner [6].
150 H. Bolouri, M. Schilstra
The origins of modeling the regulation of gene expression goes back to the
Nobel-prize winning work of Lwoff, Jacob and Monod on the mechanisms under-
lying the behaviour of temperate bacteriophages, bacterial viruses that switch be-
tween so-called lytic and lysogenic states [7-9]. Mark Ptashne's book [10] gives
an excellent review of the lytic-Iysogenic behavior of bacteriophage lambda, ar-
guably the most studied GRN. Some more recent models of GRNs and the ac-
companying theoretical developments are described in Chaps. 1 to 5 of [11], and
an extensive review of the current literature in this field may be found in [12].
For the rest of this chapter, we will assume the reader is familiar with the basic
biology of gene regulation and concentrate on specific issues of interest to model-
ers. However, to avoid confusion, in the following subsections we first briefly
outline a few basic concepts and facts.
~~
l (,"""
. Gen Cell surface
G,~,m
receptor
populations
( ' '
•.••.••.
. ~ ......ucleu
, ....
........
.'
+J
Ligand
Cell molecule
\VaII Cytoplas
2 What is a Gene?
Over the years, the definition of what constitutes a gene has gradually changed.
Unfortunately, simple definitions like 'one gene, one protein' tend to confuse
rather than help. A commonly accepted definition of a gene is all the DNA se-
quence that is necessary and sufficient to account for the experimentally observed
pattern of production of a specific protein in vivo. Note that this definition implies
there are two aspects to a gene. Firstly, a gene encodes a particular protein. Sec-
ondly, a gene includes any DNA sequence that affects its pattern of expression,
i.e. what cell types it is expressed in, how much, and at what times. Thus, the
Models of Genetic Regulatory Networks 151
DNA encoding a single gene may be divided into two parts: the protein coding
sequence, and the regulatory sequence.
The process by which genes produce proteins involves many steps, and regulatory
interactions can influence the rate and product of each step. The regulation of
transcription, the process by which DNA sequence is 'read' and transcribed into
messenger RNA (mRNA), is probably the most prevalent form of genetic regula-
tion. For the rest of this chapter we will focus on transcriptional control of gene
activity. But mRNA splicing, editing, transport and localization, translation, and
protein modifications alI contribute to the concentration and variety of any protein
produced by a gene and should not be forgotten. Fig. 2 is a cartoon example of
how multiple proteins can interact to regulate gene transcription. Different combi-
nations of proteins can bind to different regulatory sequences in a gene and acti-
vate or repress transcription depending on the history of a cell and extra celIular
signals (see books referred to earlier).
The genomes of more than 45 species have now been sequenced and some 90
more are about to be completed 1. Analysis of these genomes confirms some gen-
eral trends. For example, the genomes of most prokaryotes (single celled organ-
isms without a nucleus, such as bacteria) and cell organelles (such as mitochon-
dria and chloroplasts) are organized in a single, circular DNA. This appears to put
significant constraints on the overall size of the genome. The relatively short gen-
eration time of prokaryotes, and their very long evolutionary history, combine to
make their genomes compact and extremely well optimized. Gene expression in
prokaryotes is usually controlled from a relatively short stretch of DNA just be-
fore the start of the coding sequence, the promoter, to which repressor or activator
proteins can bind. A single promoter may control the transcription of more than
one protein, and sometimes the coding sequences of more than one protein over-
lap.
On the other hand, the genes of eukaryotes (organisms in which alI of the ge-
nornic DNA is contained inside a nucleus) generally have large regulatory se-
quences. The regulatory sequences can be thousands of nucleotides away from
their coding sequences. Transcription factors, the regulatory proteins that bind to
these sequences, control the rate at which transcription is initiated [1]. In simple
single celled eukaryotes, such as yeast, a single transcription factor may some-
I http://www.ncbi.nlm.nih.govIPMGifs/Genomeslbact.html
152 H. Bolouri, M. Schilstra
times be sufficient to initiate transcription, but, in general, the more complex the
organism is, the greater the average number of transcription factors that are in-
volved in the regulation of a single gene.
Transcription
(a)1
11\----
I I I ţ (c)
Regulatory Coding
region region
(d)
5 Modeling GRNs
Some genes (such as those encoding heat shock proteins) are present in large
numbers within the genome. But for the most part, there are just one or two active
copies of each gene per cell. So, even when activators, repressors, or transcription
fac tors are present in large numbers, transcription is fundamentally stochastic. For
some genetic circuits, such as the lambda-phage Iysis-Iysogeny decision circuit
[13] and circuits involving invers ion of DNA segments [14,15], this randomness
appears to be important. It allows cells to exhibit multiple alternative responses to
given conditions. For the most part though, one can average out the randomness
over time. In that case, the rate of change in activity for a gene can be represented
by ordinary differential equations (ODEs). Such time-averaged representations of
gene regulation are simplified by assuming that parts of the system are at steady
state or equilibrium on the time scale of mRNA production [16-18]. This lends it-
self readily to model ling GRNs as continuous sigmoids and autonomous differen-
tiai equations [19]. A further simplification uses 'sum of products' polynomial
functions to describe gene activity ovcr time (see [20]). Further simplifications
Models of Genetic Regulatory Networks 153
use kinetic, qualitative, and Boolean logic models (for a review of the underlying
theoretical issues, see [21-25], for example models see [26-30]).
An example of a logical model of a GRN is shown in Figs. 3 and 4. These
figures are taken from an ongoing collaboration between our group at the Univer-
sity of Hertfordshire and the laboratory of Prof. Eric Davidson at the California
Institute of Technology. The diagrams show a set of interactions among families
of genes inferred from celIular experiments. The model is intended to represent
the necessary and sufficient interactions for endo-mesoderm specification in sea
urchins. A fuller description of the model is beyond the scope of this chapter.
Here, we present the diagrams as examples with which we can give meaning to
some basic concepts. In both figures, gene families are represented by horizontal
lines crossed with a 90° arrow. The coloured lines above the genes represent inter-
actions among gene products (proteins). Where such a line is incident to a gene
symbol, it indicates a regulatory interaction between the protein and the gene. The
manner in which regulatory proteins interact with each other and the regulatory
DNA of the gene they are incident on, is represented below the gene symbols. Ar-
rows indicate positive (excitatory) inputs to the system while bars indicate repres-
sive inputs. The regulatory interactions (in this case, logical) are represented by
the functions indicated in the circles. Fig. 3 shows the 'view from the genome'
[31], i.e. the set of alI interactions possible among the modelled components. In
any one celI, and at any one time, only a subset of these interactions will actually
be taking place. The active interactions are represented by the coloured lines in
Fig. 4, whereas the grey lines represent inactive interactions. Fig. 4 is an example
of what Amone and Davidson [31] call the 'view from the nucleus' .
The history of GRN modelling goes back a long way. For example, in 1971, Stu-
art Kauffman started describing GRNs as Boolean networks [32], and in 1973
Rene Thomas delineated the principles of kinetic logic. Here, we outline only
some of the major trends in modelling over the past 20 years. This is not, by any
means, a comprehensive review!
Prokaryotic systems, and in particular the lysis-lysogeny switch of bacterio-
phage lambda have been studied extensively, and are probably the best place to
start for those new to the field; see for example [16, 17,26,27]. McAdams and
Arkin [33], as well as Gibson & Bruck [34] have shown that stochastic effects
play a fundamental role in the regulation of lambda phage activity (see [35] for an
earlier example of analysis of stochastic effects on gene activity). For a model of
another bacteriophage GRN, see [36]. We previously mentioned the important
work of Thomas and colleagues (e.g. [23]) in formulating a discrete logic formal-
ism ('kinetic logic') for GRNs. This formalism has been used to model a number
of genetic regulatory systems, both prokaryotic and eukaryotic (see [37] and ref-
erences therein).
154 H. Bolouri, M. Schilstra
OT
X
Fig. 4. Sea urchin endo-mesoderrn GRN: view from the nucleus (view in one group of cells
at a particular developmental time)
7 GRN Simulators
Unfortunately, there are currently no simulators with facilities for the range of
GRN representations discussed here . We are currently developing such a resource
that we hope to make public shortly (see http://strc.herts.ac.uklbiolMaria/NetBuilder/-
index.htm).
One of the earliest publicly available simulators with explicit representations
for gene regulation was MetaSim developed by Hans Westerhoff and colleagues
[42]. Under development is Gene-O-Matic, a package, aimed specifically at mod-
elling genetic networks in a multi-cellular context (PI Ute Platzer, German Can-
cer Research Center, Heidelberg, http://mbi.dkfz-heidelberg.de/mbi/research/cellsirn!-
net/index.shtml). A number of knowledge-based simulators that describe GRNs in
terms of predicate logic have also been developed: MolGen [43] , Genesis [44],
GenSim/HepyGene [45]. Of course, one can use general biochemical network
simulators such as
• DBSolve«PI Igor Goryanin, Glaxo-SmithKline, UK
http://websites.ntl.com/-igor.goryanin
• E-Cell (PI Masaru Tomita, Keio Univ., Japan, http://www.e-cell.org/)
• V-Cell (National Resource for Cell Analysis and Modelling,
http://www.nrcam.uchc.edu/index.html )
• Gepasi (PI Pedro Mendes, Virginia Tech http://www.ncgr.org/software/gepasi/)
156 H. Balauri, M. Schilstra
References
Alberts, B., et al., Molecular biology ofthe ceU. 3 ed. 1994, New York: Garland Pub-
Iishing.
2 Lewin, B., Genes VII. 1999, Oxford: Oxford University Press.
3 Davidson, E.H., Genomic regulatory systems. 2001, San Diego Calif.: Academic Press.
4 Carroll, S.B., J.K. Grenier, and S. D. Weatherbee, From DNA to diversity : molecular
genetics and the evolution of animal design. 2001, Cambridge MA: Blackwell Scienee
Ine.
5 Gilbert, S.F., Developmental biology. 2000: Sunderland, Mass: Sinauer Associates Inc.
Models of Genetic Regulatory Networks 157
6 Gerhart, J. and M. Kirschner, Cells, embryos, and evolution. 1997, Malden Mass:
Blackwell Science. 642.
7 Lwoff, A., Lysogeny. Bacteriol. Rev., 1953. 17: p. 269-337.
8 Jacob, F. and J.L. Monod, Genetic regulatory mechanisms in the synthesis of proteins.
J. Mo!. Biol., 1961. 3: p. 318-356.
9 Jacob, F. and 1. Monod, On the regulation of gene activity. Cold Spring Harb. Symp.
Quant.Biol., 1961. 26: p. 193-211,389-401.
IO Ptashne, M., A genetic switch. Phage Lambda and higher organisms. 2 ed. 1992, Ox-
ford: Blackwell Scientific.
11 Bower, J.M. and H. Bolouri, eds. Computational modelling of genetic and biochemical
networks. Computational molecular biology, ed. S. Istrail, P. Pevzner, and M. Water-
man. 2001, MIT Press: Cambridge Mass.
12 Smolen, P., D.A. Baxter, and J.H. Byme, Modelling transcriptional control in gene
networks - Wethods, recent results, andfuture. Bull. Math. Biol., 2000. 62: p. 247-292.
13 Arkin, A., J. Ross, and H.H. McAdams, Stochastic kinetic analysis of developmental
pathway bifurcation in phage lambda-infected Escherichia coli ce lis. Genetics, 1998.
149: p. 1633-1648.
14 Van De Putte, P. and N. Goosen, DNA inversions in phages and bacteria.DNA inver-
sions in phages and bacteria. Trends Genet., 1992.8: p. 457-462.
15 McAdams, H.H. and A. Arkin, It's a noisy business! Genetic regulation at the nanomo-
Iar scale. Trends Genet., 1999(15): p. 65-69.
16 Johnson, A.D., et al., Â repressor and cro-components of an efficient molecular switch.
Nature, 1981. 294: p. 217-223.
17 Ackers, G.K., A.D. Johnson, and M.A. Shea, Quantitative model for gene regulation
by Âphage repressor. Proc. Natl. Acad. Sci. USA, 1982.79: p. 1129-1133.
18 Keller, A.D., Model genetic circuits encoding autoregulatory transcription factors. 1.
TheoL Bio!., 1995.172: p. 169-185.
19 Mestl, T., E. Plahte, and S.W. Omholt, A mathematicalframeworkfor describing and
analysing gene regulatory networks. J. TheoL Biol., 1995. 176: p. 291-300.
20 Gibson, M. and E. Mjolsness, Modelling the activity of single genes., in Computational
modelling of genetic and biochemical networks., J.M. Bower and H. Bolouri, Editors.
2001, MIT Press: Cambridge, Mass. p. 1-48.
21 Snoussi, E.H., Qualitative dynamics of piecewise-linear dijferential equations: a dis-
crete mapping approach. Dynamics and Stabi'.ity of Systems, 1989. 4(3 & 4): p. 189-
207.
22 Thomas, R. and R. D'Ari, Biologicalfeedback. 1990, Boca Raton Fla. CRC Press.
23 Thomas, R., D. Thieffry, and M. Kaufman, Dynamical behaviour of biological regula-
tory networks. 1. Biological role of feedback loops and practical use of the concept of
the loop-characteristic state. Bull. Math. Biol., 1995. 57(2): p. 247-276.
24 Kauffman, S.A., The origins of order. Self-organizarion and selection in evolution.
1993, New York NY: Oxford University Press.
25 Akutsu, T., S. Miyano, and S. Kuhara, lnferring qualitative relations in genetic net-
works and metabo/ic pathways. Bioinformatics, 2000. 16(8): p. 727-734.
26 Shea, M.A. and G.K. Ackers, The 0R control system of bacteriophage lambda. A
physical-chemical model for gene regulation. J. MoI. Bio!., 1985. 181: p. 211-230.
27 Thieffry, D. and R. Thomas, Dynamical behaviour of biological regulatory networks.
Il. lmmunity control in bacteriophage lambda. Bull. Math. Bio!., 1995. 57(2): p. 277-
297.
158 H. Bolouri, M. Schilstra
28 Thieffry, D., et al., From specific gene regulation to genomic networks: a global
analysis of transcriptional regulation in Escherichia coli. BioEssays, 1998. 20(5): p.
433-440.
29 Yuh, C.-H., H. Bolouri, and E.H. Davidson, Genomic cis-regulatory logic: experimen-
tal and computational analysis of asea urchin gene. Science, 1998. 279: p. 1896-1902.
30 Yuh, C.-H., H. Bolouri, and E.H. Davidson, Cis-regulatory logic in the endo16 gene:
switching from a specification to a differntiation mode of control. Development, 200l.
128: p. 617-629.
31 Amone, M.I. and E.H. Davidson, The hardwiring of development: Organization and
function of genomic regulatory systems. Development, 1997. 124( 10): p. 1851-1864.
32 Kauffman, S.A., Gene regulation networks: a theory for their global structure and be-
havior. Current Topics in Dev. Biol., 1969.6(145).
33 McAdams, H.H. and A. Arkin, Stochastic mechanisms in gene expression. Proc. Natl.
Acad. Sci. USA, 1997.94: p. 814-819.
34 Gibson, M.A. and J. Bruck, Efficient stochastic simulation of chemical systems with
many species and many channels. J. Phys. Chem. A, 2000. 104: p. 1876-1889.
35 Băttner, R., K. Bellmann, and T. Cierzynski, A discrete stochastic simulation model of
the regulation of gene expression with variable control characteristics., in Molecular
genetic information systems: modelling and simulation., K. Bellmann, Editor. 1983,
Academic-Verlag: Berlin.
36 Endy, D., D. Kong, and J.Y. J, lntracellular kinetics of a growing virus: A geneticaliy
structured simulation for bacteriophage 17. Biotechnol. Bioeng., 1997. 55(2): p. 375-
389.
37 Thieffry, D. and R. Thomas. Qualitative analysis of gene networks. in Pacific Sympo-
sium on Biocomputing. 1998. Maui, Hawai, USA: World Scientific.
38 Shimada, T., et al. Knowledge-based simulation of regulatory action in lambda phage.
in 1st IEEE lnt. Symposium on lntelligence in Neural and Biological Systems
(INBS'95). 1995.
39 McAdams, H.H. and L. Shapiro, Circuit simulation of genetic networks. Science, 1995.
269(5224): p. 650-656.
40 Reinitz, J., E. Mjolsness, and D.H. Sharp, Model for cooperative control of positional
information in Drosophila by Bicoid and Maternal Hunchback. J. Exp. Zool., 1995.
271: p. 47-56.
41 Nair, T.M., S.S. Tambe, and B.D. Ku1kami, Analysis of transcription control signals
using artificial neural networks. Computer Applications in the Biosciences, 1995.
11(3): p. 293-300.
42 Stoffers, H.J., et al., METASIM: object-oriented model/ing of celi regulation. Com-
puter Applications in Biosciences., 1992.8(5): p. 442-449.
43 Meyers, S. and P. Friedland, Knowledge-based simulation of genetic regulation in bac-
teriophage lambda. Nucleic Acids Res., 1984. 12( 1): p. 1-9.
44 Friedland, P., et al., GENESIS, a knowledge-based genetic enigeering simulation sys-
tem for representation of genetic data and experiment planning. Nucleic Acids Res.,
1982.10(\): p. 323-340.
45 Karp, P.D., Artificial-intelligence methods for theory representation and hypothesis
formation. Comput. Appl. Biosci., 1991. 7(3): p. 301-308.
46 Koza, J., Genetic programming: on the programming of computers by means of natu-
ral selection (complex adaptive systems). 1992, Cambridge MA: MIT Press. 819.
Models of Genetic Regulatory Networks 159
47 Dellaert, F. and R.D. Beer. Toward an evolvable model of development for autono-
mous agent synthesis. in Artificial Life IV: 4th Int. Workshop on the Synthesis and
Simulation of Living Systems. 1994: MIT Press, Cambridge, Mass.
48 Gallagher, J.c. and RD. Beer. Evolution and analysis of dynamical neural networks
for agents integrating vision, locomotion and short-term memory. in Genetic and Evo-
lutionary Computation Conference (GECCO-99). 1999. Orlando, Fla.
49 Rust, A.G., et al. Developmental Evolution of an Edge Detecting Retina. in 8th Inter-
national Conference on Artificial Neural Networks (ICANN'98). 1998: Springer, Berlin
Heidelberg New York.
50 Rust, A.G. and R Adams. Deve/opmental Evolution of Dendritic Morphology in a
Multi-Compartmental Neuron Model. in 9th International Conference on Artificial
Neural Networks (ICANN'99). 1999: London, IEEE.
51 Rust, A.G., R. Adams, and H. Bolouri. Evolutionary Neural Topiary: Growing and
Sculpting Artificial Neurons to Order. in 7th International Conference on the Simula-
tion and Synthesis of Living Systems (AlifeVIl). 2000.
52 Beerli, R.R, B. Dreier, and C.F. Barbas III, Positive and negative regulation of en-
dogenous genes by designed transcription factor.\·. Proc. Natl. Acad. Sci. USA, 2000.
97(4): p. 1495-1500.
53 Elowitz, M.B. and S. Leibler, A synthetic oscillatory netwurk of transcriptional regula-
tors. Nature, 2000. 403: p. 335 - 338.
A Model of Bacterial Adaptability Based on
Multiple Scales of Interaction : COSMIC
Abstract
Evolution has frequently been seen as a result of the continuous or dis contin-
uous accumulation of small mutations. Over the many years since Darwin,
it has been found that simple point mutations are not the only mechanism
driving genomic change, for example, plasmids, transposons, bacteriophages,
insertion sequences, deletion and duplication, and stress-sensitive mutation
all have a part to play in directing the genetic composition and variation of
organisms towards meeting the moving target that is the environmental ideal
that exists at any one time. These generate the variation necessary to allow
rapid evolutionary response to changing environmental conditions.
Predictive models of E. coli cellular processes already exist, these tools
are excellent models of behaviour. However, they suffer the same drawbacks;
all rely on actual experimental data to be input and more importantly, once
input that data are static. The aim of this study is to answer some of the
questions regarding bacterial evolution and the role played by genetic events
using an evolving multicellular and multispecies model that builds up from
the scale of the genome to include the proteome and the environment in
which these evolving cells compete. All these scales follow an individual based
philosophy, where by genes, gene products and cells are all represented as
individual entities with individual parameters rather than the more typieal
aggregate population levels in a grid. This vast number of parameters and
possibilities adds another meaning to the name of the simulation, COSMIC:
COmputing Systems of Microbial Interaetions and Communications.
162 R.Gregory, RPaton, J.Saundf~rs, and Q.H.\Vu
1 Introduction
Evolution has frequently been seen as a rmmlt of the continuous or discoIltin-
uous accumulation of small mutations. Over the many years since Darwin,
it has been found that simple point mutations arc not the only mcchanism
driving genornic change, for example, plasmids, transposons, bactcriophagcs,
insertion sequences, deletion and duplication, and stress-sensitive mutation
an have a part to play [1,2] in direct ing the genetic composition and variation
of organisms [3] towards rneeting the moving target that is the environmcntal
ideal at any one time. Considering the probability of single point mutations
arising and rcpair mechanisms that may act to counterac:t their ac cumula-
tion, it is unlikely that simple mlltation alone can create rapid diversity. It is
dear that evolutionary change depends more on larger sc:ale changes in ge-
nornic seqllences causcd by sexual and othcr forms of hori7:ontal gcne transfer.
These generate the variation neccssary to allow rapid evolutionary n~sponse
to changing environmental conditions.
Predictive models of E. coli ccllular proccsses already exist; the E-Cdl
project [4] aims to use gene data dircctly in a mathematical model of tran-
scription. The Virtual Cell [5,6] project makes use of user-defined protein re-
actions to simulate compartments at the nudeus and cellular level. Gepasi3 [7]
also models protein reac:tions, but from within au enc:losed box environment.
The BacSim [8] project simulates individual cell growth at the population
sc:ale. Eos [9] is also based at the population sc:ale, but is intendcd as a frame-
work for testing idealised ec:ologies, represented by evolutionary algorithms.
These tools and those that they rely on are excellent models of behaviollr.
However, they sufIer tlw same drawbacks; ali rely on actual experimental
data to be input and more importantly, once input that data is static. The
aim of this study is to answer SOlIle of the questions re gard ing bacterial evo-
lution and the role played by genetic events other than simple point mutation
using an evolving multicellular and multispecie::; modd that builels up from
the scale of the genome. In efIect, it is not bacterial evolution that is being
interrogated, but the co-evolution of bac-teria and any organism that has a
direct efIect on the genetics of those bacteria.
To test these questions, it is necessary to build a model that attempts to
encompass what are considered the important qualities of bacterial evolution
and bacteriallife, but is not overly specificd as to constrain the results. The
model is therefore a careful balance of biological and computational reali-
ties [10] with an emphasis on open-endedness [11]. The biologicalliterature
has many examples of the possible forms of mechanism within the relatively
'simple' example of E. coli, but even this must be carefully constrained. It
is de ar that computer rnodels lack computational power when comparcd to
real world processes.
In focusing attention on aspects of the E. coli system, it is dear that
there are two new insights provideel by the emerging disciplines of genomics
and proteomics. Proteomics is the study of enzyme and protein interactions.
Model of Bacterial Adaptability 163
2 Biology
2.1 DNA, RNA and Proteins
then coiled around itself to a diameter of 11 nm, coiled around it self again to
a diameter of 30 nm, coiled again to a diameter of 300 nm and then wound
in a 700 nm diameter spring like formation.
The E. coli genome is around 4.6 Mbp (million hase pairs) long and
arranged in a circle, composed of many loops each 50 to 100 kbp long joined
by proteins. The DNA is supercoiled, and is further compaeted by protcins
e.g. HU, a dimeric protein that binds DNA non-specifically and wraps the
DNA around itself.
2.2 Transcription
Part of the central dogma proposed by F.H.C. Crick is that RNA is created
from DNA by a transcription phasc where information from DNA is trau-
scribed into RNA. RNA differs from DNA in that it contains uracil in place
of thymine and is frequently single-strandcd, but may loop back on itsclf
caused by sections of intrastrand homology and base pairing. Some reports
(e.g. [22]) have suggested that RN A could be the initial basis for life. The
transcription process is carried out by RNA polym<'fase (an enzyme specific
for the creat ion of RNA species on aDNA template). For the purposes of the
COSMIC simulation, ali references to RNA that follow should be regarded
as mRNA. However it should be appreciated that other types of RNA are
rcquired by living cells as components of thc translation approach.
The first step in the creation of a protcin from the DNA is transcription
by aDNA-dependent RNA polymerasc (a complex protein-based machine).
RNA polymerase catalyses transcription in a process that requires double-
strandcd DNA as well as the nucleotides ATP, GTP, CTP and UTP. Tran-
scription is directional, with mRNA chain growth progressing in the 5' to 3'
direction using aDNA templatc strand with the opposite polarity (3' -t 5')
(the sense strand). The complementary strand (running 5' -t 3') is called
the anti-sense strand. In E. coli the RNA polyrnerase moves at 40 bases per
second at 37 ac, transcribing as it moves.
To start transcription, an RNA polyrnerasc molecule binds to thc doublp
stranded DNA, ideally at a prornoter site. The RNA polyrnerase and ali re-
quired cofactors constitute a transcription complex. The initial base (start)
of the transcribed region after the promoter is known as the + 1 position and
the promoter and any operator sites have negative nucleotide designations,
negative relative to this position. As with replication, the DN A strands must
be unwound to allow transcription both to be initiated and to proceed. Tran-
scription stops at a terminator sequence which contains a self-cornplernentary
region that can form a stem-Ioop or a hairpin structure out of the RNA prod-
uct. This structure ensures that the transcriptioll complex stops and promotes
dissociation of its constitucnt parts.
Model of Bacterial Adaptability 165
There are two dasses of proteins, globular and fibrous. Globular proteins can
be regarded as spherical parti des as they are folded compactly. Most enzymes
are globular. The fibrous proteins have a high axial ratio (length/width)
and are typically used in a structural role. A protein takes a shape dictated
partly by the primary polypeptide sequence and is stabilised by a variety of
forces which hold it together; hydrophilic side chains tend to the outside and
hydrophobic amino acids remain on the inside.
Except for a few catalytic RN A molecules, an enzyme is almost certainly a
protein. An enzyme is a catalyst for reactions that would occur but otherwise
very slowly. COSMIC considers them essential, the large difference in reac-
tion rates makes it not worth considering the case when the reaction occurs
without the enzyme. As a result COSMIC ignores the passive components
and just models the enzymes.
This is also supported by another limitation, a complete system woulel
illvolve the interplay of basie chernical buileling blocks and physical processes,
reqlliring the implementation of alI the chemical and physical processes that
can also possibly occur. Unlike more fixed simulations, COSMIC must be
open cnded and so must sUPPdrt a complete range of possible chemi cal and
physical interactions without favouring one set of fine graineel interactions
over another set. But, as there are more processes than time to implement
them and too many important processes have no predictable function in an
evolving environment, COSMIC is deliberately limiteel to avoiel fine graineel
chemistry and undecidable processes; insteael the focus is on the concepts of
optional transcription and evolutionary mechanisms.
The encoding from D N A to protein function is ill elefineel (hence the
protein foleling problem [12, 23] ). The bases are reaei three at a time as
c:odons anei converteel to 20 different types of amino aci el that are linkeel by
peptiele bonels to make a polypeptiele chain. There are 64 potential codon
combinations, many more combinations than requireel to encoele alI possible
amino aciels, but allowing inclusion of two start coelons, three stop coelons
and the use of varying degrees of redundancy in codon specificity for each
amino acid. Although the base alphabet is 20 letters in size, the polypeptide
can fOfIn words of several hundred amino acids in length. This combineel
with the fact that shape (influenced by the encoeling) elictates the role of
the protein then makes nearly intractable the task of eletermining the role
of a real protein. The shape can take any imaginable three-dimensional form
(obviously the chain cannot go through itself) , using regular patterns that
change over its length.
To further the level of complexity, there is variat ion in gene expression anei
control, as enzymes anei regulatory proteins may be useel to control transcrip-
166 R.Gregory, R.Paton, J.Saunders, and Q.H.Wu
The principal idea of optional transcription cornes from research on the lac
operon found in E. coli. This is only one heavily researched example using
the common idea of optional transcription.
The lac operon is concerned with the use of lactose as a carbon source,
the enzymes that can enable the utilisation of lactose as a carbon and en-
ergy source are only created when lactose is available. The lac operon con-
sists of the structural genes lacZ encoding ,B-galactosidase, lac Y encoding
a galactoside permease and IacA encoding a thiogalactoside transacetylase.
,B-galactosidase is an enzyme that hydrolyses lactose into galactose and glu-
cosp. Galactoside permease aids in lactose transport through the cell wall of
the bacteriurn.
These tluee genes are encoded side hy side in a singlc transcriptional
unit called lacZYA. Relative to this there is an operator site 01 ac between
-5 and +21 bases relative to the transcription start point, just after the
promoter site Plac- If the operator site binds a lac repressor protein then
transcription is strongly repressed. The lac repressor protein it self is encoded
slightly upstream of the lac operon promoter in the lacI gene, which is also
part of the lac operon.
The lacI gene encodes the repressor protein, but the protein itself is active
only as a tetrarner, the gene product only works when in groups of 4. Once this
has taken place, the repressor has a very strong affinity for the lac operator
and also a high affinity with non-operator DNA.
Model of Bacterial Adaptability 167
The tryptophan operon encodes five structural genes that arc required for
tryptophan synthesis. The TINA transcript produced is a single 7-kb long
strand, starting from the operator sit(~ Otrp' As with the expression of the
lac operon, the RNA product is Ilnstable and so reglllation of DNA qllickly
n~gulates the protein encl prodllct, which in this case is tryptophan.
The trpR operon is the sourcc of the trp reprcssor and is locatcd sornc-
wlwre upstream of the trp operon. The operator seqllence is symmetrical and
forms the repressor binding sit<~, whic:h also overlaps with the trp promoter
site hetween bases -21 and +3. The core repressor binding site is a palindrome
18 ilp long. Tlw trp repressor only actively binds the operator site when it
has itself formed a complex with tryptophan. The rcpressor is a dimer and
168 R.Gregory, R.Paton, J.Saunders, and Q.H.Wu
has a structural similarity to CRP protein and the lac repressor, the dimer
needs two tryptophan to be complete. It is the tryptophan that gives the
dimeric structure the correct distance between its two reading heads and its
central core.
The five structural genes encode for enzymes that produce tryptophan.
Tryptophan therefore inhibits its own synthesis by a magnitude of 70 (pre-
sumably under artificial tryptophan simulating conditions), which is much
smaller than that caused by binding of the lac repressor.
So far, the trp operon seems like the lac operon, but for the self-inhibition.
As well as the normal transcription controls there is also an attenuator se-
quence following a le ader sequence using around 162 nt before the first struc-
tural gene trpE. This attenuator is a short area rich in palindromic GC bases
followed by each U bases and it is also rho-independent. If this sequence man-
ages to form a hairpin structure in the transcribing RN A, then it will act as
a terminator and force early termination at around 140 bp long, stopping
before the structural genes have been reached.
The leader itself also has a role to play; divided into four successive se-
quences, 1 and 2, 2 and 3, and 3 and 4 are complementary, and so can bind
to themselves to form a hairpin which stops further RNA transcription. If 2
and 3 bind then this does form a hairpin, but does not stop transcription.
Under normal conditions the binding of 1 to 2 and 3 to 4 is more favourable
than 2 to 3.
Aiso in this leader is an efficient ribosome binding site and successive
codons encoding for tryptophan. Under conditions of low tryptophan avail-
ability the ribosome would pause at this point. Since transcription and trans-
lation are tightly coupled in E. coli, the net effect then is to control trypto-
phan transcription negatively. In reality the hairpin formation between se-
quences 3 and 4 is more likely when tryptophan level is high, the pause
occludes sequence 1 leaving sequence 2 to bind with sequence 3. In the alter-
native case, the pause occurs at the start of sequence 2 and so sequence 2 is
occluded allowing sequences 3 and 4 to form a hair pin.
Given both forms of optional transcription, tryptophan-dependent repres-
sor and tryptophan-dependent attenuation region, the totallevel tryptophan
can be amplified by 700 times, the attenuation sequence giving a lQ-fold
increase and the tryptophan-dependent repressor giving a 70-fold increase.
Generally, attenuation is present in at least six other operons with a role
in amino acid synthesis. An example is the his operon, but in this case the
attenuation mechanism is the only means of control, there is no operator.
cntity with individual parameters. This a110ws for the inclusion of spatial ef-
fects [29] without imposing artificial spatial boundaries of fixed rcsolution.
Given a population of functional polypcptides in the ce11, each has a chance
of reacting with each other and on the genome. The prohahility involved is
based on their type as defined by the genome, their position, their half-life
and their age. Each reaction lasts for a variable time based on the type of
pairing. At this point, it must be pointed out that simplification of the real
biologic al process has lead to COSMIC having no explicit molecule type on
which emwmes react. A single type of generic protein is assumed to exist
and a11 reactions act on that pseudo protein. This protein can be bound by
enzymes, as we11 as the normal process of enzymes acting on each other; the
resulting effect is that the enzymes and the single FAP type are not availablc
to do anything else while bound, a110wing another molecule a chance to bind.
Another simplification is in the area of motility: there is a direct relationship
between motile response and optional transcription. Chemical simulation is
beyond the scope of COSMIC and yet ftage11a motion is primarily created
by proton motive forces over the ce11 wa11 [25], coming from both electri-
cal potential and chemical potential. As a result, this simplification seems
justified.
It is hoped these necessary simplifications have little impact on the cvo-
lutionary results obtained with COSMIC, the primary point of enquiry being
the change in the genome over the course of medium and long term time;
the genome and the concept of the operon [27,30,31] are intended to givc
the COSMIC populat ion their adaptability to respond and adapt; the com-
pleteness of genome operators was therefore seen as important, gene types
such as operators, promoters, repressors, anti-repressors, terminators, RNA
polyrnerase and FAPs (Flage11a Activation Proteins) form the main classes.
These types should give the genorne the ability to express any function re-
quired without the need of realistic metabolites.
4 Model
~ Fl age ll :l
>-< Receplors
:.00.' Gcnomc
0.:.:0 Pathways
000 Proteins
Gp,ne produets ean react with other gene produets and operate direetly
on the genome. Deerypting this relationship requires several steps. Firstly,
because the relationships between gene and protein function are eurrently
intraetable (espeeially for eompletely unknown genes), funetion mapping has
been eonstrained to use only a direct hamming distanee between genes. A
gene (whieh always leads direetly to its gene product) can react with an-
other gene if the hamming distanee is below a speeified threshold. When
specified for all genes, these relations give a picture of whieh gene produet
(enzymes) can reaet. Not ali sequences are genes, control sequences are also
considered in the hamming process. This creates interaction paths between
control sequences and gene products, thus allowing these gene products to
then be called sigma factors, repressors and inducers. A summary of pos-
siblc interactions, sources of transcription products and a map of indirect
or direct attenuation are shown in Fig. 2. It is important to note that the
term 'enzyme' in COSMIC is used loosely to account for a gene product that
influences a chemical reaction.
Given a genome comprising genes, the gene products of each and a set
of relatiollS describing what can happen, the only missing set is of reactions
that are happening. This is based on the relations of possible reactions , each
relation carries with it a set of reaction instances. Each instance carries with
it references to the specific eIl7:yme instances involved as weU as a mutual
position in space and a reaction duration based on the type of relation.
172 R.Gregory, R.Paton, J.Saunders, and Q.H.Wu
•........
The only aspect missing from this system is sense/response networks. For
this COSMIC considers receptors on the cell wall (input) and fiagella motion
(output) to be directly related to the transcription networks. The receptors
represent imitation gene products that exist in a fixed position. These can
bind to normal gene products using a parallel set of relations also based on
hamming distance.
The fiagella response (output) follows in the same way. In both cases, the
effect is not one of conversion, i.e. a receptor places a protein in the cell and
that goes on to be converted to something el se by gene products. Crossing
the protein and chemical scale was never the remit of COSMIC. Instead, the
mode of operation follows that of use, if a gene product is bound at one
location it cannot also be involved in a reaction elsewhere. This is quite a
simplification, but as the role of COSMIC is mainly in evolution it seems
justified.
The largest scale of COSMIC is the environment. To test evolution, a
population of cells must compete side by side for what exists in a common
environment. To this end all cells live in a glucose-rich environment that is
depleted over time by the cells. In common with Genetic Algorithms, better
cells in better regions of the environment will grow faster and so multiply
faster. For this COSMIC uses quantitatively accurate cell growth equations
[8,33,34J.
Model of Bacterial Adaptability 173
The combination of input reward based on cell position and cell position
based on flagelIum out put produces an indirect reward based system that is
the basis on which the simulated E. coli evolve.
Evolution and other causes of cell death are then supplied by the other
organisms and mutations in the farm of both simple point mutation and the
longer gene sequence duplication and gene sequence deletion. This creates
an ecology where it is good to be fit and also good to be genetically unique
[35,36]; an ecology such as this also alIows for a dynamic population size.
5 Implementation
The COSMIC system is clearly very large, the lack of explicit boundaries,
the individual nature of simulation and the need always to model multiple
cells to stimulate evolution alI contribute to this. This requires considerable
computational resources in order to simulate evolution at a reasonable pace.
For this COSMIC currently runs an a GRID enabled cluster of 12 node dual
processor AthlonXP 2000+ machines. This leads to a rapid simulation but is
still slower than real time, on the order of 7:1. In the space of 4 days COSMIC
had evaluated 1,176 bacterial cells, with 170 stillliving at the point the simu-
lation ended. The final environment had turned into a bacterially challenging
patchwork of nutrients, average final genomes were in the range 48 to 822
genes long (168 mean), with 10 to 104,633 enzymes per cell (12,693 mean).
CPU utilisation varied in the range 95-99% creating around 2.5 gigabytes of
data per day for 13 h of simulated bacterial life. This required 11 gigabytes
of total storage space with additional storage for visualisations.
The architecture follows the client server model, the environment being
the server and cells being the clients. This ensures efficient parallelisation as
cell intercommunication is rare and the environment has minimal processing
needs, the overall re suit being a linear growth in computational nodes allow-
ing a linear growth in the product of simulation speed and total simultaneous
number of cells.
6 Results
The COSMIC system clearly has two distant scales, the environment state
aud the state of each cell. COSMIC can generate a vast amount of detailed
raw data on the environment and more importantly, the gene expres sion of
each cell coupled with gene linkage maps. The gene expression data show
that even random loosely connected genomes can achieve a wide range of
transcription rates and connection to the environment.
COSMIC can be considered a batch process, but for the never-ending
naturc of the batch. The architecture ultimately allows a never-ending simu-
lation in which state can be recorded and reloaded while changing the global
174 R.Gregory, R.Paton , J .Saunders, and Q.H.Wu
ceH 177 moving slowly near the bottom right of the picture and the majority
of cells not moving at aH.
The labelling system of cells is based on an ever increasing unique iden-
tifier that a cell obtains upon creation as either a child or a newly initialised
ceH. When a ceH divides the parent retains the original identifier and the child
receives the next unu sed identifier. At that instant both cells are the same
in terms of genetics but differ in the share of enzymes as each enzyme has a
50:50 chance of staying with the parent. It is an implementat ion decision that
parents retain their identifier rather than obtain another identifier; the result
of this decis ion is seen in the following charts which show the figures for cells
42 and 177 over the course of their lives without ever referring to the children
of 177 or the other children of 42 despite them being effectively equivalent
at the instant of division and being subjected to the same level of mutation.
This is for convenience and brevity as these two cells represent a major linage
and contain genetic differences, making them good candidates to single out
and present from aH those available. The convenience comes from avoiding
pasting together results, for instance from ceH 177 then switching to cell 255
and then switching again to cell 276. This is possible but unnecessary when
cells at division are effectively equivalent.
Sturt=881 Otfspring=2S9
End=48599 Max Depth=27
Span=47718
881(86431....
2025(-1.956).... '-"4~->1tJ1I
2981(814881
4469(810481:::: "jo':::,,_>177
5517(83931
5910(81087;":'''' ....
_... ····177->2011
6997(814711....
8675(89641
9639(83071.... ""177->255
9946(814441
11J90(âI4)"'"iir" ....
•• _- ····255->270
111111111111111111111111'1""'"111111111111111111111111111111111111111111
Fig. 5 . Lifetime gene expression of a celllineage 42 over an 82.5 min period . Shades
of grey equate to the log of enzyme concentration. Vertical grey areas fiII unu sed
space later filled by an expanding genome
Fig. 6. Lifetime gene expression of cell lineage 177 over an 97.5 minute period.
Shades of gray equate to the log of enzyme concentrat ion. Vertical gray areas fiII
unused space later filled by an expanding genome.
178 R.Gregory, R.Paton, J.Saunders, and Q.H.Wu
this ceH is also constantly changing, though the speed of change is not so
dramatic. Comparing the start and the end shows how much the new ceH
changed from its parent and it can also be seen how much larger the genome
grew despite growth and reduction being the same probability.
Figure 8 gives some overaH statistics for ceHlineage 42. The top graph shows
ceH volume; starting from some initial value aH ceHs increase to 0.4 fi then
divide, growth being dependent on substrate. Three ceH divisions can clearly
be seen. The next two graphs show the ceH x/y position in the environment
(as shown in environment view figure) , over the range of 0.2 mm . Both
graphs show the ceH is moving at some speed and so must be covering fresh
substrate regularly - an initial requirement of a converged ceH. The third
graph shows total enzyme population over time. As the enzyme population
is divided when the ceH divides, the approximate halving of ceH contents can
be seen coinciding with halving of ceH volume. The penultimate graph shows
receptor activity over time, that is the ability of the optional transcription
network of the ceH to bind with glucose receptors on the ceH waH - it is
regarded as the ceH input. As can be seen, there is some activity, but this
would never be enough to guide the ceH inteHigently around the environment.
The final graph shows fiageHa response, the equivalent of ceH output. This
shows the ability of the network to interact with the fiageHa and so move
the ceH. As can be seen (largely by the x/y graphs) there is sufficient ceH
movement to achieve near maximal growth. The important point here is that
output lacks control, this unconverged ceH keeps the ceH moving regardless
of input.
Figure 9 shows the same overaH parameters as previously but for ceH
177 lineage, the third generation offspring of lineage 42. The most obvious
difference between the two is the number of enzymes, in this case it is now
relatively constant; quickly recovering from ceH divisions. The unusual case
that appears in so many COSMIC ceHs is at time 9.5 x 103 . There is a rise in
Model of Bacterial Adaptability 179
Fig. 7. Partial network interactions over the lifetime of ceH 42. Nodes represent
genes. Inside each node is the particular gene sequence given to that gene, along
with a gene type in shortened form , the long form of which is given in Fig. 2. Edges
show a relationship between the particular genes; this includes reflexive edges which
indicate an inhibited RN A polymerase. Figures on the edges show total usage counts
for binding and unbinding reactions
180 R.Gregory, R.Paton, J.Saunders, and Q.H.Wu
the enzyme population, a fali, a halving caused by cell division, but then the
ceU never reaUy recovers and loses a connection with the environment (both
input and output). COSMIC then eventuaUy kills the ceU as it is considered
no longer viable.
7 Discussion
The COSMIC model is a growing tool with which evolution can be mod-
elled in a greater detail than ever before allowed. The problems this brings
are twofold, the computational effort is significant and stands to limit the
evolution of COSMIC itself. However the biggest hurdle is the shear amount
of data generated. The individual based philosophy is clearly then a double-
edged sword, escaping global averages means always considering individuals.
This leads on to further stages of COSMIC development: tools to analyse the
out put in a meaningful and concise way.
Model of Bacterial Adaptability 181
0.4
0.38
0.36
0.34
0.32
0.3
0.28
0.26
0.24
0.22
0.2
0.1
10 20 30 40 50 00 70 90 100
CcI1 Simlllatioil Time (minu les)
200
~/------------------------ - ----_. 180
160
140
.2 120
100
~
o
80
60
40
20
o
20 30 40 50 60 70 90 100
Cell SimulatiQu Time (min\ltes)
4 500 ~--~---r---r---,----r---~---r---r---,----r---~---r---r.~-,--~r---~---r--~ 4500
4000 4000
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
O ~~ __- L__ ~ __ L-~ __- L__ ~ __ L-~ __- L__ ~ __ L-~ __- L__ ~ __ L-~ ____ o
10 20 30 40 50 60 70 80 90 100
Cell imulatiOIi Tiru@ (tuillutes)
0.09 0.09
0.08 0.0
0.07 0.07
0.06 0.06
0.05 0.05
0.04 0.04
0.03 0.03
0.02 0.02
0.01 0.01
O ~ __ L-~L-~~~
o
10 90 100
Ceu Simulation Time (minutes)
0.25 .----r---......--......-.....'T1. .r r - 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
OL-~~~~~uu~~~u.~~~~~
o
10 20 30 40 00 70 80 90 100
Cetl Sim\llation Timp (minutes)
0.1
0,08
0.06
0.0,1
0.02
O
180 190
Cell Simulation Timc (m inutes)
0.4 ~~---r--~--r-~--~--~--r-~--~--~--r-~--~--~--r-~~~---.--. 0.4
0.35 0,35
r
0.3 0,3
0.2. 0.25
, 0.2 0.2
""
"-
O
0.15
0,1
0,15
0. 1
0.05 0,05
o ~ua~~~~~~~~L-~L-~-L~~~~~~~~~~~~
O
90 100 110 120 130 140 150 160 170 160 190
Cell Simulation Timc (minules)
References
1. Shapiro,.1. A. (1997). "Genome organi7:ation, natural genetic engineering and
adaptive mutation", 71rf'.nd.~ in Genetics, 13,98-104.
2. Shapiro,.1. A. (1999). "Genome System An:hitecture and Natural Genetic En-
gineering in Rvolution", Annals of New Yor'k Academy of Science.~, 870, 23-35.
3. Koch, L. (1993). "Genetic Response of Mic:rohes to Rxtreme Challenges", .Jour-
nal of Theorf'.tico.l Riology, 160, 1-21.
4. Tomita, M., Hashimoto, K., Takaha'ihi, K., Shimi7:u, T., Matsu7:aki, Y.,
Miyoshi, F., Saito, K., Tanida, S., Yugi, K., Venter, .1. c., Hutchison, C. (1999).
"R-CRLL: Software environment for whole cell simulation'·. Rioinformatics, 15,
72-84.
5. Schaff, .1., Fink, C. c., Slepellenko, R., Carson, .1. H. & Loew, L. M. (1997)
,. A General Comput,at.ional Framework for Modeling Cellular Strudure and
Fllnc:tion", Riophy.~ical .Jollrnal, 73, 1135-1146.
6. Schaff, .1. & T.oew, L. M. (1999). "The Virtual Cell". Pacific Sympo.~itlm on
Riocomptlting, 4, 228-239.
7. Mendes, P. (1997). "GEPASl: a software paekage for modeling the dynamics,
steady states aud control of bioehemical and other systems.", Trend.~ in Rio-
chemical Science, 22, 361-363.
8. Kreft., .1.-U. , Rooth, G., & Winpenny,.1. W. 'T. (1998). "RacSim, a simulator
for inrlivirlllal-baserl morlelling of hac:terial colony growth". Microhiology, 144,
3275-3287.
9. RonRma, R, Shackleton, M. & Shipman, R.. (2000). "Ros - an evolutionary anrl
Reosystem research platform". RT Technology .Jmlrnal 18, 24-31.
10. Way, R C. (2001). "'The role of computation in modeling evollltion'·. RioSy.~
tem.~, 60, 85-94.
11. Kampis, G. (1996). "Self-morlifying systems: a morlel for the C'onstrnc:tive origin
of information'·. RioSy.~tf'.ms, 38, 119-125.
12. Karpllls, K., Sjlander, K., Barret., c., CIine, M., Hallssler, D., Hughey, R.., Holm,
L. & Sander, C. (1997). "Predirting prot.ein Strnrture nsing hirlrlen Markov
models". Prof.f'.in.~: StrlldllTf'., Jiltmdion, and Gf'.nef.ic.~. 134-139.
13. Kent., W ..1. & Haussler, D. (2001). "GigAssemhler: An Algorit.hm for the Initial
Assemhly of t.he Hllman Genome'·. In PTf'..~.~.
14. Freeman, M. (2000). "Feerlback control of intereellular signalling in develop-
ment'·. Nat11rf'., 408, 313-319.
15. Bray, D. (1990). "lnt.ereelllllar signalling a<; a Parallel Distributed Process",
.Jo71rnal of Theoretical Biolog:,! 143, 215-231.
16. Shapiro,.1. A. (1991). "Genomes as smart systems". Gf'.nP.tica, 84,3-4.
17. Bray, D. (1995). "Protein molecules a<; compntational elements in Iiving cells".
Naf.llTf'. 376.
18. R. Paton (1998). "The Ecologies of Hererlitary Information", C:,!hf'.mf'.tic.~ f~
Hllman Knowmg, 5, 31-44.
19. Dllan, Z., Holc:ombe, M. & BeII, A. (2000). ,. A logic: for biological syst.ems".
BioSy.9tf'.m.~ 55, 93-105.
20. Reddy, V. N., Liebman, M. N. & Mavrovollniotis, M. L. (1996). "Ql1alitative
analysis of hiochemical reaction systems". Comp71tf'.r.9 in Bio[ogy and Medicinf'.,
26,9-24.
21. Arkin, A. & ROHS .1. (1994). "Computational Functions in Biological RE'aetion
Networks". Biophy.~ica[ .hl1lmal 67, 560-578.
22. Gilbert., W. (1987). "The Exon Theory of Genes". Sympo.9ia on Qllantitative
Diolog:,!, 52, 901-905.
23. Hallssler, D. (1997). "A Brief Look at Some Machine Leaming Problem in
GE'nE'ties'·. Pro(~eeding of COLT 97.
184 RGregory, RPaton, J.Saunders, and Q.H.Wu
24. Ochman, H., Lawrence, J. G. & Groisman, A. G. (2000). "Lateral gene transfer
and the nature of bacterial innovation". Nature, 405, 299-304.
25. MacNab, R. M. (1996). "Flagella and Motility". Escherichia coli and
Salmonella, cellular and molecular biology. Second Edition, Neidhardt, F.C.
et al. (eds), 1, 123-144.
26. http://www.ecocyc.org
27. Record Jr, M. T., Reznikoff, W. S., Craig, M. L., McQuade, K. L. & Schlax, P.
J. (1996). "Escherichia coli RNA Polymerase (E?70), Promoters, and the Kinet-
ics of the Steps of Transcription Initiation". Escherichia coli and Salmonella:
Cellular and Molecular Biology, Second Edition, Neidhardt, F. C. el al (eds).
2, 792-821.
28. Collado-Vides, J., Gutierrez-Rios, RM. & Bel-Enguix, G. (1998). "Networks of
transcriptional regulation encoded in a grammatical model". BioSystems 47,
103-118.
29. RayChaudhuri, D., Gordon, G.S. & Wright, A. (2001). "Protein acrobatics and
bacterial cell polarity". PNAS, 98, 1322-1334.
30. Turner, P. C., McLennan, A. G., Bates, A. D. & 'Vhite, M R. H. (1997). " Instant
Notes in Molecular Biology". BIOS Scientijic Publishers., Abingdon.
31. Wagner, R. (2000). "Transcription Regulation in Prokaryotes". Oxford Univer·-
sity Press.
32. Gregory, R, Paton, R.C., Saunders, J.R & WU, Q.H. (2002). "COSMIC: Com-
put ing Systems of Microbial Interactions and Communications 1 - Towards a
Modelling Framework" , In preparation for submission to Biosystems.
33. NeijsseI, O. M., Teixeira De Mattos, M. J. & Tempest, D. W. (1996). "Growth
Yeild and Energy Distribution". Escherichia coli and Salmonella: Cellular and
Molecular Biology, Second Edition, Neidhardt, F.C. et al. (eds), 2, 1683-1692.
34. Bremer, H. & Dennis, P. P. (1996). "Modulation of Chemical Composition
and Other Parameters of the Cell by Growth Rate". Escherichia coli and
Salmonella: Cellular and Molecular Biology, Neidhardt, F.C. et al (eds) 2,
2nd Ed. ASM Press, Washington DC. 1553-1569.
35. Krebs, J. R & Davies, N. B. (1987). "An introdllction to Behavioural Ecology".
Blackwell Scientijic Publications, Oxford.
36. Devine, P., Paton, R. & Amos, M. (1997) "Adaptation of Evolutionary Agents
in Computational Ecologies". Biocomputing and emergent computation: Pro-
ceedings of BCEC97. World Scientijic, Lundh, D., OIsson, B. & Narayanan, A.
(eds.), World Scientific, Singapore. ISBN 981-02-3262-4, 66-75.
Stochastic Computations in N eurons and
Neural Networks
Jianfeng Feng
Abstract
1 Introduction
Neurons fire randomly, at least in the cortex. As a consequence, it is a long
standing question in neuroscience: what are the advantages of random firing
in neuronal computation? Or what are the functional roles of random firing?
There is a corresponding question in computer science: can we make use
of the randomness, mimicking the random firings in neurons, to solve some
practical problems? There are many publications in the literature to explore
these issues [1]. Here we review some of our recent results and, hopefully,
they will shed some new lights onto these fundamental problems.
It has recently been argued that statistics might be the foundation of neu-
roscience [2]. In statistics, at least in cert ain periods, it was widely accepted
that the statistical decision theory is the basis of the whole of statistics.
Hence applying the decision theory in statistics to, or developing a decis ion
theory for, neuroscience is of vital importance. The actual neuron mechanisms
underpinning the discrimination activity remain one of the most significant
and puzzling problems in neuroscience, despite there having been mounting
experimental and theoretical results devoted to the topic (for example see
recent reviews [3,4]). In a series of experiments, Newsome and his colleagues
have compared single neuron activity with psychophysical experimental data.
186 Jianfeng Feng
They found , surprisingly, that the information extracted from single neuron
activity in middle temporal (MT) is almost enough to account for psychophys-
ical experimental data. Hence an observation of the firing rates of a single
neuron, at least in MT, contains enough information to further guide mo-
tor activity. Imagining the enormous number of neurons in the cortex, their
findings are striking and open up many interesting issues for further theo-
retical and experimental study. Interestingly, similar findings are reported in
somatosensory pathways [4] as well. In line with these experimental results ,
we concentrate on the relationship of the input and output firing rates of
a single neuron. The first issue we are going to address is quite straightfor-
ward (a discrimination task, which is the basis of decision theory, see Fig.
1). Suppose that a neuron receives two set of signals (coded by firing rates)
Mixed
....
Fig. 1. For two mixed signals (left), after neuronal transformation, will they become
more mixed or more separated?
~ What are the implications of the informax prin cip le with a network of the
IF neurones? In other words, what is the outcome of the informax principle
learning? After learning, do the weights tend to self-organize themselves,
represent inputs or accomplish something else?
~ What is the computational capacity of an IF neuron? Can the IF neuron
be applied to solving practic al problems? We have seen applications of the
IF model, or more general spiking neuron; to solving engineering problems,
where the information contained in spike intervals is exploited. We test the
computational capacity of the IF neuron in blind separations, using a rate
coding assumption.
~ What is the optimal value of the ratio between inhibitory and excitatory
inputs? We alI know there are inhibitory inputs in the neural system, but
the functional role of them remains elusive.
Under the informax principle, the learning rule developed for the IF model
is complicated and so a complete theoretical treatment is impossible. For
some ideal cases, we could understand its underpinning mechanisms [20]. For
general cases, we have to resort to numeric al simulations and alI the questions
raised before are answered.
The first neuron model we use here is the classical lF model [15,16,25).
When the membrane potential Vi is below the threshold Vihre, it is given by
with Ei(t),Ii(t) as Poisson processes with rates Ai,E and Ai,! respectively,
a > D, b > O are magnitude of each EPSP and lPSP, p and q are the total
number of active excitatory and inhibitory synapses. Once Vi crosses vthre
from below a spike is generated and vt is reset to Vrest , the resting potential.
This model is termed the lF model. The interspike interval of efferent spikes
is
T = inf {t : vt ~ l'thre}
More specificaUy, synaptic inputs take the foUowing form (p = q)
P P
Isyn(t) = a LEi(t) - b L1j(t)
i=l j=<
Pc P Pc P
where Ei(t), i = 1,··· ,Pc are correlated Poisson processes with an identical
rate Aj,j = 1,2, Ei(t) are Poisson processes with a firing rate ~i-Pc in-
dependently and identically distributed random variables from [O, 100], i =
Pc + 1,'" ,P, Ii(t),i = 1,···,p have the same property as Ei(t), but with
a firing rate of r Aj, j = 1,2 or r~i-pc for r E [O, 1] representing the ratio
between inhibitory and excitatory inputs.
From now on, we further use diffusion approximations to approximate
synaptic inputs [25] and without loss of generality we assume that a = b and
Vrest = O.
P-Pc P-l1c
Isyn(t) = apcAjt +a L ~it - bpcrAjt - b L r~it
i-1 i-1
where B t is the standard Brownian mot ion and j = 1,2. We first consider
the case that a neuron receives independent inputs. As we might expect, the
output from a single neuron does not contain enough information for the
discrimination task (results not shown, see next section), with the ratio of
inhibitory to excitatory inputs spanned from nil to one (exactly balanced
inhibitory and excitatory input). We then turn to the situation that a small
amount of correlations are added to the synaptic inputs which code coherently
moving dots. For the simplicity of notation we assume that the correlation
coefficient between ith excitatory (inhibitory) synapse and jth excitatory
(inhibitory) synapse is C > O. The correlation considered here reflects the
correlation of activity of different synapses, as discussed and explored in [5].
It is not the correlation of single incoming EPSP or IPSP which could be
expressed as Cij(t - ti) for the EPSP (IPSP) at time t of the ith synapse and
time ti of the jth synapse. We refer the re ader to [5] for a detailed discussion
on the meaning of the correlation considered here.
In summary, suppose that a neuron receives P synaptic inputs. The goal
of the postsynaptic neuron is to discriminate between two types of inputs:
1. Pc excitatory Poisson inputs fire at a rate Al and Pc inhibitory Poisson
inputs fire at a rate r Al with r E [O, 1].
2. Pc excitatory Poisson inputs fire at a rate A2 (A2 i- Al) and Pc inhibitory
Poisson inputs fire at a rate r A2 with r E [O, 1].
In both cases, the neuron receives 'noise' Poisson inputs consisting of p - Pc
excitatory inputs and the same number of inhibitory inputs. We as sume that
'noise' excitatory inputs are uniformly distributed between O and 100 Hz,
and 'noise' inhibitory inputs are between O and lOOr Hz. Without loss of
generality, we always assume that A2 > Al,
Stochastic Computations in Neurons and Neural Networks 191
,
I \
,,
I
,
,,
,, ,
, ,
Ul I p (out}(Â) ,
E '2 ,
~ I \
O>
o
Cii
I ,,
3 Theoretical Results
and
and denotc
(2)
If it is clear from the context about the dependence of a( A1 , A2 ' e, r) on e, 1" , we
somctimes simply write 0(A[,A2,c, r) as 0(Al , A2)' Hence for fixed (A1 , A2),
O(A[, A2) gives us the critical value of Pc: when Pc > O:(Al' A2) the input
patterns arc pcrfectly separable in the sense that the the olltPllt firing rate
histograms are not mixcd with TPM=O; when Pc < ct(A[, A2) the inpllt pat-
terns might Bot bc separable with TPM> O. Note that we consider the worst
case here and in practical applications , the critica) vaille of Pc at which the
inpllt pattcrns are perfcctly separable, as fOllnd in the previolls section, is in
192 Jianfeng Feng
generallower than 0:( Al, A2' e, r). From now on, all figures are generated using
the same parameters as in the previous section, if not specified otherwise.
Here is the basic idea of our approach. As pointed out before, it is not easy
to directly calculate the distribution of (T). Nevertheless, the discrimination
task is only involved in the left most point of P2(A), i.e. R min(A2), and the
right most point of Pl (A), i.e. Rmax(Al), provided that both P2 and Pl are
positive only in a finite region. This is exactly the case for the model we
considered here sin ce neurons fire within a finite region.
(3)
15
• Pc = 60
• • o •
• • O
+ Pc = 55
oPe = 50
10
• • - 10
;e
o
5
• • O
O
i3 - 20 • •
O
O
-3
• • • p,= 40
o
O o + P, = 35
- 10 O -4 o o Pc = 30
O o
- 15
o 0.2 0.4 0.6 0.8 -5 o 0.2 0.4 0.6 0 .8
Ratior Ra~o,
25 10
20 Pe = 60
o •
+ Pc = 55
•
15 o Pc = 50
•
10
•
- 10
• O
;e
o 5
• • ;e
o -20
O
O
O
O
O • Pc = 40
O + Pc = 35
-40 O
- 10 O
O o Pc : 30
O O
-150
0.2 0.4 0.6 0.8 o 0.2 0.4 0.6 0.8
Aalior Ratior
20 10
•
• Pc = 60 5
15
+ Pc = 55 o
o Pc = 50
10 -5
•
• •
- 10
• •
• •
'"i5 13 -15
• •
O
O - 20
-25
• O
O
O . Pc = 40
O O + Pc::;; 35
O O
-1 O o Pc ~ 30
-35 O
O
- 150 -40
0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
Ratio r Ratior
Fig. 3. Diff=R l11 in('\2) - Rmax (.\I) for a = 0.5 (upper panel) , a = 1 (middle panel)
and a = 2 (bottom panel) with .\1 = 25 Hz, .\2 = 75 H:>: and c = 0.1. It is easy to
read out a(.\I, '\2, c, r)
194 Jianfeng Feng
(5)
In Fig. 4 some numeric al results of a(Aj, A2) are shown. It is easily seen
that when e = O, a(A[, A2) is independent of r.
\Ve want to point out another amazing fact from Theorem 2. a(A[, A2, O, r)
and a(Aj, A2, e, 1) with e> O are both independent of a, 1Ithr·e, L. When r =
1, c = 0.1, A[ = 25 HII, A2 = 75 Hz and Amax = 100 Hz and p = 100, we have
0(25,75,0.1,1) = 32.5133,0(25,75,0,1) = 66.6667 (see Figs. 3 and 4). Hence
we conchIde 32.5133 < a(25, 75, 0.1, r) < 66.6667 for r E (0,1).
Finally we are in the position to answer one of the questions rai sed
in the introduction: a large coefficient of variation (CV) implies a srnall
rY (Al, A2 , e, r). Note that the CV of interspike intervals is the the variance
when we calculate the mean interspike intervals. In other words, for each
fixed realization of ~i, i = 1,···,p - Pc, it is the variation of T. When we
calculate a(Aj, A2' e, r), the variance of firing rates histogram is mainly in-
troduced via the masking 'noise'. In other words it is the variatioll of (TI.
Therefore these are different sources of noisc. By in cre as ing the number of
interspike intervals, we can reduce the variance of the first kind. Note that
in the previous section we deliberately employed a small number of spikes
(100), which might be close to the biological reality, to estirnate (TI. Tl]{'
Stochastic Computations in Neurons and Neural Networks 195
In tlw previous subsections, we only consider the out put firing rate his-
tograms. It is certainly interesting to compare the input histograms with
out put histograms. As before, let A be the set of illput frequency of the
model. For a fixed (Al E A, A2 E A) with Al < A2 we have two corresponding
histograrns pi (A) and P~ (A) of input firing rates, i.e. pi(A) is the histograrn
of PcAl + Lf~ic ~i and p2(A) is the histogram OfPc A2 + Lf~i' ~i' Define
70 100
65
+
+
9
*******
N 55 +
+
!.
(il +
+
ţ..
iii
(\1
50
+ ~ 85 • c= O +
lI'l
1l'
• c =o
(\J + C =0.05
45 + 'B' 80 O C =0.1
+ c = 0.05
40 o c = 0 .1
75
35
300 700
0.2 0.4 0 .6 0 .8 0.2 0.4 0.6 0.8
Ratio r Ratio r
Fig. 4. 0'(>'1, >'2) (Hz) versus ratio r with >'1 = 25 Hz and >'2=75 Hz and >'1 = 25
Hz and >'2=30 Hz. It is easily seen that when c = O, 0'(>'1 , >'2) is flat, otherwise it
is a decreasing function of r
4O r----r----r----r----~--~--~
20
O .,,-
-20 .. '
.. ' ".-'
N
;
==
'il
"S
%-80 - r= 1
O . r =O
-100
- 120 c=O.1
- 140 .
-1~00~--80
.....---..:'60::---
_4~0----!:
20::---~--:'·
5000
Input ditf (Hz)
Fig. 5. Rmin(>'2) - Rmax (>'I) vs. R~'in(>'2) - R:nax (>'I) which is a function of pc ,
for c = O (right) and c = 0.1 (left)
Stochastic Computations in Neurons and Neural Networks 197
4 Informax Principle
'li; Q( (88x Y) -1 ~
8w
(88xY) (6)
For the purpose of the following sections, we redefine the IF model. Sup-
pose that a neuron receives :CPSPs (excitatory postsynaptic potentials) at
p synapses and IPSPs (inhibitory postsynaptic potentials) at p inhibitory
synapses. When the membrane potential ~(i) of ith neuron is between the
rest ing potential V rest and the threshold Vlhre, it is given by
dv:(i)
t = -L(v:(i)
t - v:rest )dt + dI(i)
syn (t) (7)
where L is the decay rate and synaptic inputs
j==l j==l
with Ei(t), Ii(t) as Poisson processes with rate Af and A{ respectively, w~ >
O, w{j > O being magnitudc of cach EPSP and IPSP. As before, the IF model
ean be approximated by
dvt(i) -
- - L( V t(i) - V.rest )dt + d~(i) ()
Zsyn t
where
p p
(9)
(10)
We only consider the case of rate coding since then a rigorous input-
out put relationship of firing rates is known for the IF model. By rate coding,
we mean that the information is carried by the firing rate of a neuron. It is
weU known in the literature that the input-output relationship of a neuron
takes a sigmoidal form and this is the basis of neural computations developed
in the past decades. The input-output relationship of an IF model (see Fig.
7) takes a sigmoidal function as well (not surprising at aU), but it depends
not only on the mean of inputs, but also on the variance of inputs. The
latter feature enables us to derive novel learning rules which, to the best of
our knowledge, have not been reported in the literature and exhibit some
intriguing phenomena. The importance that a neuron might use higher order
statistics to compute has been recognized early in the literature (see for
example [2]).
5 Learning Rule
Remember that the mean interspike interval of the IF model with Poisson
inputs is given by
g(x)dx (11)
where
Stochastic Complltations in Neurons and Neural Networks 199
Output
Inp ut
<
I I I I!
I I I I
IIII
I
111 1 I I I I
r _ 1
........... ..........
...... ..
......
°o~~~--~--~--~--~~
~--~--~--~---:~~
I npu t (HZ)
Fig. 7. Out put (Hz) versus input (kHz, AL = A2 = A3) of the IF model with
n = 3 , W·i.i = 0.5,i , j = 1, ·· ·,3,AI = A2 = A3, Vi"rc = 20 mV,Vr e st = O mV a.ud
L = 1/20
âlJI
lJI
âWi j
Wij cx (12)
where
and
200 Jianfeng Feng
where
Therefore
11 -(
J - -1
)p[rrP ((Ti(r)) 1+ T
i=1 ref )2
1. [f-(
~ -1
)i+iB(Ti(r))IA
BAj ij
1]
which yields
(13)
!
Defining
Ei ~ (ţ w;;Â;)L(1+ r)
ei = (L WijAj)(l - r)
j=1
letting Vrest = 0,
Stochastic Computations in Neurons and Neural Networks 201
and
2(1 - r)E; + 3w;jAjL(1 -r 2 ) + 2WijL(1 + r)(xL - e;)
2E3
3WijAjL(1 + r)[2Wij(1'- r)E; + w;jL(1 + r)(xL - e;)]
if k =j
2E?
WikAjL(1 - r2)(4wij - Wik)
2E3
3WijAjL(I + r)[2wik(1 - r)E; + w;k L (1 + r)(xL - ei)] if k =f. j
2E?
we arrive at
For the general case the learning rule presented here is too complex to
be explored theoretically, nevertheless we can simulate it numerically, as pre-
sented in the next section.
alJI
aWij
(17)
IJI
where 'Ii is the efferent firing rate (in unit of l/ms) ofthe ith neuron. There-
fore if we fix (clamp) the efferent firing rate, we could have a version of
supervised learning.
We simulate the learning rule with the following parameters: p = 6, 'Ii =
50 HZ,Ai = 500 Hz for i = 1,2,3, Ai = 2,000 Hz for i = 4,5,6,
r E [0,1], vthre = 20, Vrest = O and L = 1/20, using Matlab. Thc total
cxcitatory input is of 7,500 Hz. The input is equivalent to cach rwuron re-
ceives 100 active synaptic inputs, each with 75 Hz, which are physiologically
plausible parameters [17]. Note that for the purpose of visualizing results,
here we scale down p and scale up A correspondingly. The initial weights are
random variables from U(O, 9) and step size oflearning is 0.05. After 800 time
steps of learning, the synaptic weights are stable, for aH cases we considered
here.
We carry out simulations for r = 0,0.5 and 1. It is shown (see [24]) that
weights become dilution, after learning. For example when r = O we have
(W1l (O), W12 (O), W13 (0), W14 (0), W!.5 (O), W16 (O)) = (7.8,2.9,2.25.5,9.1,5.2)
and
(WIl (800), W12 (800), W13 (800), W14(800), W15 (800), W16 (800)) =
(9.5, O, O, O, 5.3,1.4)
namely the connections
2W .---~---------~--~----~
"
12
,o
î
oS 8
r= 1
!!i
~
6
~
"
w
eel Cell
2. 5, ,..---~----~----_----~--__,
r. l c
§ :~
~
~ 1.5 ~
1> "O
r . 0.5 C
<'
.! ' ~
0.6 r. ,
.11
"u'8 0.5
~ O.•
O.S ( =0
(=0 0.3
02
O,
Cell
O., , 3 Celi •
Fig. 8. Mean and CV of output ISIs. Leit are initial values and right are values
after learning
10
Qutput
B
4 Inpl".Il
Tlme
Fig. 9. Input signals (lowe1' trace of each figure) and 01ltput signals (uppe1' trace)
after blind separations with l' = 0.5
a(t) = 2 sin(400t)cos(30t)
{ b(t) = 2 sign{sin[500t+9cos(40t)]} (18)
c(t) = 2 (rand - 0.5)
for t 2: O and rand is the uniform random numbers from [O, 1]. The input
signals are mixed with the matrix
Stochastic Computations in Neurons and Neural Networks 205
and a constant vector (2,2,2) is added (to ensure that input signals are
positive), i.e. the received signals are
3.00002.45000.1000)
( 1.0000 5.0000 0.3000
0.4450 3.0000 9.0000
The weights are updated with a rate of l/(t + 100)1/2, an initial state of
randomly generated number from U(O, 10). After 800 times of iterations we
obtain
0.36790.00525.6465)
w(800) = ( 0.1533 1.26190.1755
6.3680 4.4491 0.0023
which gives us the results as shown in Fig. 9. The out put signals are obtained
via
7 Discussion
Time
Fig. 10. Input signals (lower trace of each figure) aud output signals (upper trace)
after blind separations with r = O
28
"00
It Of"ofion 1in"los
Fig. 11. Efferent interspike intervals versus tirne with r = O, see Fig. 10
Fig. 12. Upper panel is the original faces , each with 287 x 387 pixels. Lower panel
is obtained after 100,000 iterations, r = 0.5
alJl
âWij
(19)
IJI
208 Jianfeng Feng
if we simply use "fi and Ai as instantaneous firing rate, i.e the inverse of
interspike interval, a learning rule based upon the time is obtained [28 32].
We will explore it in further publications.
The approach presented here is quite general. In principle, once we know
the relationship of the input-output of a neuron, we can obtain the learning
rule of the neuron. As we have mentioned in Sec. 1, we have accumulated
many results on the in put-out put relationship of a single neuron, for example
the IF-FHN model [33]. We expect our approach can also shed more light
onto the cod ing problem [29].
It is also very interesting to note that for all cases we considered the best
separation is achieved for a(t), which is supposed to be a more natural signal
than the other two. Whether it is an intrinsic property of the IF model is
not clear at the moment. There are many issues to be further explored in the
future. For neuron decision theory:
- We have only attempted to accomplish the discriminating task and have
not included time constrains. Definitely it is of vital importance for a neu-
ronal system 1,0 tel! one signal from the other within as short a time window
as possible.
- We have tested our model with static inputs. It is an interesting question to
generalize our results here to time-varying inputs as reported in [4]. Such
a study might be helpful to clarify the ongoing debate on the advantages
of 'dynamical stimuli' over the 'static stimuli'.
- The input signal used here is very naive. To transform the image of moving
dots to input signals specified in the present paper requires a neural net-
work to preprocess the image. Hence to devise a network model (spiking
neural networks or Reichardt detector [34]) to reproduce our results is one
of our ongoing research topics. We expect that such a study could provide
us with a template to compare models with psychophysical experiments.
- A neuronal system without learning is a 'dead' system. In actual situa-
tions, we all know that learning is prevailing in neuronal systems. Hence a
reasonable learning rule should improve the neuronal capability of discrim-
ination of different input signals. There are severallearning rules reported
in the literature. To assess the impact of them on discrimination tasks is
an intriguing issue.
Discriminating between different input signals is probably a more fun-
damental constraint on the neural system than others such as maximizing
input-output informat ion or redundancy reductions, a view recently echoed
in [2]. To understand it will reveal principles employed by neuronal systems
which remain mysterious to us. The issue discussed here is currently a hot
topic in neuroscience (for example see [35]). Our approach provides us with
a solid theoretical foundation for further study and we expect that our ap-
proach will also open up many interesting questions to be furt heI' investigated
in the future.
Stochastic Computations in Neurons and Neural Networks 209
References
1. Albright T.D., Jessell T.M., Kandel E.R., and Posner M.I. (2000), Neural sci-
ence: A century of progress and the mysteries that remain. Cell100, sl-s55.
2. Barlow H. (1986) Perception: What quantitative laws govern the acquisition of
knowledge from the senses? in Functions of the Brain ed. C. Coen, Clarendon
Press: Oxford.
3. Parker A.J., and Newsome W.T. (1998) Sense and the single neuron: probing
the physiology of perception. Annu. Rev. Neurosci. 21 227-277.
4. Ricciardi, L.M., and Sato, S. (1990), Diffusion process and first-passage-times
problems. Lectures in Applied Mathematics and lnformatics ed. Ricciardi, L.M.,
Manchester: Manchester University Press.
5. Feng J. (2001) Is the integrate-and-fire model good enough? -a review. Neural
Networks 14 955-975.
6. van Vreeswijk C., Abbott L.F., and Ermentrout G.B. (1994), Jour. Computat.
Neurosci. 1 313-321.
7. Feng J., and Liu F. (2003) Discriminating between input signals via single
neuron activity (submitted).
8. Hopfield, J.J., and Brody, C.D., (2000), What is a moment? "Cortical" sensory
integration over a brief interval, PNAS97: 13919-13924
9. Hopfield, J.J., and Brody, C.D., (2001), What is a moment? Transient syn-
chrony as a collective mechanism for spatiotemporal integration, PNAS98:
1282-1287.
10. Feng J., and Brown D. (2000b), Impact of correlated inputs on the out put of
the integrate-and-fire model. Neural Computation 12, 711-732.
11. Salinas E., and Sejnowski T. (2000) Impact of correlated synaptic input on
ouput firing rate and variability in simploe neuron mdoels. Journal of Neuro-
science, 20, 6196-6209.
12. Salinas E., and Sejnowski T. (2001) Correlated neuronal activity and the flow
of neural information. Nature Reviews Neuroscience 2, 539-550.
13. Grossberg, S, Maass W., and Markram H. (2001) Neural Networks (Special
Issue), voI. 14.
210 Jianfeng Feng
N.Monk
1 Introduction
in pattern formation [1]. In many cases the nature of the signalling interaction is
such that a single signaHing ceH influences the behaviour only of its immediate
neighbours (for example, the Boss-Sevenless interaction in the developing fly
eye[2]). However, in some cases juxtacrine signalling directs pattern formation in
large populations of cells. This occurs when every cell can act both as a source
and as a recipient of signalling. Patterning depends crucially on the ability of each
cell to regulate the strength with which it signals with its neighbours in response to
the level of signalling it receives. This type of signalling can be described as activ-
ity-regulated juxtacrine signalling (ARJS) [3].
2 Biological Setting
Active Delta
Inactive Delta
Fig. 1. A simplified scheme iIIustrating the main features of Delta-Notch signalling. Delta-
mediated activation of Notch leads both to down-regulation of Delta activity and to signal
transduction
Spatial Patteming in Explicitly Cellular Environments 213
where (ţ), denotes the average value of the variable q in the cells immediately
neighbouring the cell labelled by i. ka and kd are the association and dissociation
rates of the ligand-receptor complex, respectively, and ţJţdenotes the linear deg-
214 N. Monk
radation rate of the variable ~. Pa and P, are functions encoding the dependence of
the rate of production of A and R on C. The first two terms in each equation de-
scribe ligand-receptor binding, the third term describes linear degradation, and the
fourth term (where present) represents regulated production. This formalism, the
explicit binding model (EBM), has been used in the work of Owen et al. [13-15]
and Wearing et al. [16].
Ligand-Receptor binding
A+R c kQ ·e
(b) /yj
Fig. 2. a The general scheme for a two-component juxtacrine signalling system in a line of
cells. Ai' Ri and Ci denote the levels of ligand, free receptor and Iigand-receptor complex in
cell i, respectively . Blue arrows represent ligand-receptor binding interactions (juxtacrine
signalling); red arrows represent intracellular regulation. b Simple monomeric reversible
binding reaction between Iigand (A) and receptor (R) to generate a bound ligand-receptor
complex (C)
which the model equations are continuous in both space and time. This approach
will not be discussed here; an outline can be found in Monk et al. [3]. Conversely,
Luthi et al. have studied models of juxtacrine signalling that are discrete in both
space and time [19]. Such systems are essentially cellular automata with continu-
ous state variables, and will also not be discussed here. A good review of the use
of cellular automata in biological modelling that is of Ermentrout and Edelstein-
Keshet [20].
Spatially discrete models have sometimes been used to study pattern formation
in lattices of cells coupled by cell-to-cell diffusion (for example, [21]). It is impor-
tant to note that the models discussed here are fundamentally different from such
'discrete diffusion' models, since the interaction between neighbouring cells de-
pends only on absolute levels of ligand activity, and not on the difference in
ligand activity between cells. This distinction leads to significant differences in the
types of behaviour exhibited by the two classes of models.
4 Pattern Formation
The model equations presented above exhibit a range of steady state and travelling
patterns. Patterning in the level of ligand-receptor complex is of particular inter-
est, since this level can have a direct influence on cell properties, such as state of
differentiation, proliferation rate, etc. Techniques for the analysis of these patterns
can be found elsewhere [13-18], and will not be covered in any detail here.
Rather, the general classes of pattern that have been found in analytical and nu-
merical studies will be outlined. Quite generally, the types of patterns observed
depend on the form of the production terms Pa and P" or of the feedback function
g. The following cases have been studied:
1. IBM with g monotonic decreasing [17]
2. IBM with g monotonic increasing [18]
3. EBM with Pa and P, monotonic increasing [13-16]
The case of IBM with g monotonic decreasing has been proposed as a simple
model that captures the essence of Delta-Notch signalling [17]. As discussed
above, the operation of DeIta-Notch signalling in an initially equivalent popula-
tion of cells most commonly leads to the development of a spacing pattern, in
which scattered cells with low Notch activity are surrounded by cells in which
Notch activity is high. Equations (4) and (5) have been studied in one- and two-
dimensional arrays of cells. Two classes of steady state solutions are possible: (a)
a homogeneous state; (b) a spacing pattern in which cells have either high or low
levels of receptor activity, cells with low receptor activity (primary fate cells) be-
ing separated by a characteristic distance with intervening cells having high recep-
tor acti vi ty .
216 N. Monk
Which pattern develops over time depends both on the initial conditions and on
the form of the feedback function g; if the steepness of g is below a critic al value,
then the homogeneous state is stable to perturbations and will develop from any
random initial conditions (with slight discrepancies at the boundaries in general).
When the steepness of g is above this critical value, then a spacing pattern devel-
ops. The precise form of the spacing pattern depends on the initial and boundary
conditions, and on the form of g. Typical steady state patterns in one- and two-
dimensional arrays of cells are shown in Figs. 3 and 4.
10 15 20 25 30 35 40 45 50
0011 number
(b) 40
00
300
Q,) 06
.5;
.., 200
100
5 10 15 20 2S 30 35 40 45 50
cellnum ber
Fig. 3. Time-course (in arbitrary units of time) of development of steady state pattems of a)
Delta and b) Notch activities in a line of cells. Cells initially have random low Ievels of
Delta and Notch activity; boundary cells receive Delta signalling only from their
neighbours within the line (zero boundary conditions). Note that the system approaches the
(unstable) homogeneous steady state before levels of Delta and Notch activity diverge
(around time 80)
In one-dimensional arrays in which alI cells initially have roughly the same lev-
els of Delta and Notch activity, the typical pattern that results has a periodicity of
2. That is, cells with high Notch activity are flanked by two cells with low Notch
activity, and vice versa. It is typical for 'defects' to arise in this pattern, where two
neighbouring cells both have high levels of Notch activity. However, pattems in
which two neighbouring cells both have low levels of Notch activity have never
been observed. Alternative patterns can be forced by biasing the initial levels of
Delta and Notch activity. For example, if the initial conditions are biased such that
cells with lower Notch activity (andlor higher Delta activity) are separated by two
cells with higher Notch activity (andlor lower Delta activity), then this prepattern
will be reinforced resulting in a stable period-3 pattern. An example is shown in
Fig. 5a. However, if one attempts to generate the con verse period-3 pattern, in
Spatial Patteming in Explicitly Cellular Environments 217
which high Notch activity cells are separated by two low Notch activity cells, the
initial conditions are overcome (even if they are strongly biased), and a typical pe-
riod-2 pattern with defects results (see Fig.5b.)
Fig. 4. Typical steady state pattern of activity of the Notch signalling pathway in an array
of hexagonal cells. The cells surrounding the array have zero Delta activity (zero boundary
conditions). White represents high Notch activity while black represents low Notch activity.
Note the regions of close hexagonal packing separated by channels of high Notch-activity
cells
(a)
aoo
oa
600
0.6
~ 400
0.4
200 02
10 15 20 25 30 s, 40 4, ,0
ce ll nuntle r
[b)
800
08
600
06
g 400
200 . ~
1- - 1· ..
,
10 15 20 25 so s, 40 " 50
O!! II nul'1"hl!r
Fig. 5. Time-eourse of Noteh aetivity in a line of eells with biased initial eonditions. a) A
foreed period-3 pattern resulting from initial eonditions sueh that eells with a bias towards
high Delta aetivity were separated by two eells biased towards low Delta aetivity. Note that
the left-hand boundary eondition forees a small region of period-2 pattern that overeomes
the bias in the initial eonditions. b) The effeet of initial eonditions biased towards a period-
3 pattern in whieh two adjaeent eells have high levels of Delta aetivity. This period-3 pat-
tern is unstable, and the 'defaul!' period-2 pattern with oeeasional defeets soon develops
IBM with g monotonie inereasing can be considered as a model for the propaga-
tion of asigna! through a population of cells under the influence of a spatially 10-
calised signa! source [18] . Due to the absence of any feedback to balance the posi-
tive feedback inherent in this model, it does not appear to be capable of generating
pattems de nava. As with ali the models discussed here, there exist one or more
homogeneous steady states of the model, in which ali cells ha ve equallevels of re-
ceptor activation (with slight deviations at boundaries in practice). If a boundary
condition is imposed that varies significantly from the homogeneous steady state,
then this boundary can have an effect on the level of receptor activation that
spreads into the cell population. This effect can take the form of either a stable
gradient or a propagating front [18] .
In the case that only one stable homogeneous steady state exists, which is just
the case that the functionj(g(C)/,u) - ,u,C has on!y one zero, then a fixed deviation
from this steady state results in the formation of a stable gradient of receptor acti-
vation. A typical example of such a gradient is shown in Fig. 6. In this example,
the on!y homogeneous steady state is the zero state and a fixed non-zero boundary
condition generates a gradient over a range of 30 eells or so.
Spatial Patteming in Explicitly Cellular Environments 219
005
004
003
0C12
001
6 10 16 20 26 30
C&l1 number
Fig. 6. A typical steady state gradient of receptor activity resulting from unilateral constant
stimulation of line of cells
While the gradients formed in this c1ass of model have an appearance similar to
those formed by the diffusion of a substance from a locali sed source, they are
quite different in some respects. Perhaps most significantly, the range over which
a gradient can form is restricted. This is in contrast to gradients formed by diffu-
sion, whose range can be extended both by increasing the strength of the signal
source and by increasing the diffusion wavelength (given by the balance of the
diffusion coefficient and the decay rate of the diffusing substance). Range restric-
tion stems from two features of the juxtacrine model. Firstly, if the strength of the
signal source is increased, the receptors on the cells neighbouring the source even-
tually become fully saturated; further increases in signal strength will then have no
effect on the form of the gradient. Secondly, if the steepness of f and/or g is in-
creased (roughly equivalent to increasing the diffusion wavelength), then a point is
reached where additional homogeneous steady states come into existence. In this
case, travelling fronts rather than gradients are generated. Thus, while these two
strategies can be effective in increasing the range of gradients in the juxtacrine
model, there are strict upper limits to their use, beyond which they either have no
effect Of lead to the generation of distinct modes of response.
In the case when there exist two or more stable homogeneous steady states, a
locali sed perturbation to one of these states can result in the establishment of a
travelling front propagating away from the site of the disturbance. Typically, two
such stable steady states are separated by an unstable homogeneous steady state
(see Fig. 7).
220 N. Monk
Fig. 7. A plot of the functionj(g(C)/,u) - ,u,C showing the existence of three homogeneous
steady states, Co' C" C,. Coand C, are stable, C, is unstable. In this example, ,u" = ,u, = l.t(~)
= ;/(0.2 + ;) and g(~) = ~
Let the levels of receptor activation in the two stable states be CII and C,. and
that in the intervening unstable state be C" with C2 > C, > Ca' Then if the cells in
an array are at Cu' a fixed local perturbation of level C p will generate a travelling
front if Cp > C,. Behind this front, aII cells will be at C,. An example is shown in
Fig. 8. Conversely, if the cells are initially at Cl' then a perturbation CI' < C, will
generate a travelling front behind which cells are at C()' Once established, these
fronts move across an array of cells at a constant speed that can be determined us-
ing linear analysis about the homogeneous states.
The richest model in terms of de novo pattern formation is EBM with Pa and P,
monotonically increasing [13-16]. Like IBM with g monotonically increasing,
discussed in the previous section, this model can generate both spatial gradients
and travelling fronts in response to a locali sed perturbation to a stable homo gene-
ous steady state. However, in this case the existence of feedback on the level of
free receptor allows gIadients of unrestricted range to be formed (although the
time taken to form the gradient increases with the range). Details are contained in
Owen el al. [13, 14].
Spatial Patteming in Explicitly Cellular Environments 221
1800 0.9
1600 0.8
1400 0.7
1200 0.6
(])
E 1000 0.5
:;J
800 0.4
600 0.3
400 0.2
200 0.1
O
10 20 30 40 50 60 70 80 90
cell number
Fig. 8. A typical travelling front of receptor activity resulting from unilateral constant
stimulation of line of cells.
For certain forms of the feedback functions Paand P" the homogeneous steady
states of this model are unstable to spatially inhomogeneous perturbations, leading
to the formation of spatial patterns [15, 16]. Two types of perturbation to an array
of celIs at the homogeneous steady state have been studied. When a small pertur-
bation is applied to cells along a line in a two-dimensional array, a pattern propa-
gates out from this line. The resulting pattern consists of either continuous or bro-
ken Iines of cells with elevated levels of receptor activation, running parallel to the
line of initial perturbation (see Fig. 9a, b). When small random perturbations are
applied throughout the array, roughly periodic patterns, with a characteristic wave-
length, result (see Fig. 9c). These patterns are formed typically from locali sed
small collections of cells that have a higher level of receptor activation than the in-
tervening cells. Sensitivity analysis shows that both types of pattern are quite ro-
bust to changes in initial conditions and parameter values [15].
This mechanism of pattern formation, which has been termed lateral induction
by Owen et al., is a new and potentially important mechanism for spontaneous
pattern formation . Patterning depends on the establishment of a positive feedback
that enhances receptor activity, and on the spatiallocalisation of this enhancement.
This is achieved in the model by having strong feedback on free receptor (P,
steep) together with weak feedback on ligand (Pa shallow). A cell that develops a
higher level of free receptor than its neighbours (due to random tluctuation) wiU
consequently develop elevated levels of bound receptor by binding of ligand from
neighbouring cells. In turn, this will increase the level of free receptor further,
driving a positive feedback. Neighbouring cells cannot propagate due to the fact
that a high proportion of ligand on these cells is bound to the excess of receptors
on the initially perturbed cell. It is this fact that, in the case of weak ligand feed-
back, ensures that cells developing high levels of receptor activation remain spa-
tially isolated.
222 N. Monk
(a)
3000 6000
(b)
2000 6000
(e)
3000 6000
5 Further Developments
References
1. Fagotto, F. and Gumbiner, B. (1996). CeH contact-dependent signaling. Dev. Bio/. 180,
445-454.
2. Krămer, H., Cagan, R. L. and Zipursky, S. L. (1991). Interaction of bride of sevenless
membrane-bound ligand and the sevenless tyrosine-kinase receptor. Nature 352, 207-
212.
3. Monk, N. A. M., Sherratt, J. A. and Owen, M. R. (2000). Spatiotemporal patteming in
mode1s of juxtacrine interceUular signalling with feedback, in Mathematical Models
for Biological Pattern Formation. (Ed. P. K. Maini and H. G. Othmer), pp. 165-192,
Springer, Berlin Heide1berg New York.
4. Weinmaster, G. (1998). Notch signaling: direct or what? Curr. Opin. Genetics Dev. 8,
436-442.
5. Heitz1er, P. and Simpson, P. (1991). The choice of ceU fate in the epidermis of Droso-
phila. Cell 64, 1083-1092.
224 N. Monk
25. Panin, V. M., Papayannopoulos, V., Wilson, R. and Irvine, K. D. (1997). Fringe modu-
lates Notch-ligand interactions. Nature 387, 908-912.
26. Christensen, S., Kodoyianni, V., Bosenberg, M., Friedman, L. and Kimble, 1. (1996).
lag-l, a gene required for lin-12 and glp-l signaling in Caenorhabditis elegans, is ho-
mologous to human CBFl and Drosophila Su(H). Development 122,1373-1383.
27. de Celis, J. F., Bray, S. and 8arcia-Bellido, A. (1997). Notch signalling regulates
veinlet expres sion and establishes boundaries between veins and interveins in the Dro-
sophila wing. Development 124,1919-1928.
28. Wilkinson, H. A., Fitzgerald, K. and Greenwald, 1. (1994). Reciprocal changes in ex-
pression of the receptor lin-12 and its ligand lag-2 prior to commitment in a C. elegans
cell fate decision. Ce1l79, 1187-1198.
29. Eaton, S. (1997). Planar polarization of Drosophila and vertebrate epithelia. Curr.
Opin. Cell Biol. 9, 860-866.
30. Usui, T., Shima, Y., Shimada, Y., Hirano, S., Burgess, R. W., Schwarz, T. L., Ta-
keichi, M. and Uemura, T. (1999). Flamingo, a seven-pass transmembrane cadherin,
regulates planar cell polarity under the control of frizzled. Ce1l98, 585-595.
31. Strutt, D. 1. (2001). Asymmetric localization of frizzled and the establishment of cell
polarity in the Drosophila. MoI. Cell. 7, 367-375.
32. Goodyear, R. and Richardson, G. (1997). Pattern formation in the basilar papilla: evi-
dence for cell rearrangement. J. Neurosci. 17,6289-6301.
33. Salazar-Ciudad, 1. , Garcia-Fernaodez, J. and Sole, R. (2000). Gene networks capable
of pattern formation: from induction to reaction-diffusion. J. Theor. Biol. 205, 587-
603.
34. Page, K. M., Maini, P. K., Monk, N. A. M. and Stern, C. D. (2001). A model of primi-
tive streak initiation in the chick embryo. J. Theor. Biol. 208, 419-438.
Modelling the GH Release System
1 Introduction
This chapter describes a model of the hypothalamic and pituitary components in-
volved in controlling growth hormone release. The model has been developed by
gathering and attempting to formalise the experimental data on the system but has
been kept as simple as possible, focusing on the functional rather than mechanical
properties of its components. In this way it has shown that a relatively simple
model can be capable of producing complex behaviour and accurately reproducing
the behaviour and output of a real brain system.
2 Research Background
Much of the information communicated between cells and systems in the brain is
represented not so much by the amplitude of activity but by the pattern of activity
over some time period. This has perhaps always been a more obviously useful and
robust alternative but it also requires that we attribute memory to very low-Ievel
components of the brain.
One part of the brain that exhibits very definite temporally patterned activity is
the neuroendocrine system. This system also has the advantage of producing an
easily accessible and relatively simple output in the form of hormone release into
the bloodstream. We know that for several hormones, inc1uding growth hormone,
a pulsatile pattern of release is optimal for their actions within the body. In male
rats, as in humans and many other animals, Growth Hormone (GH) is released in
large pulses every 3h. A pattern of large pulses rather than continuous release al-
lows a maximal response to the hormone without desensitising the target receptors
and it is also believed that the pattern of mixed high and low activity may be used
to instruct different systems with some responding to high and some low activity.
These patterns are triggered by neurons but they operate on a timescale of hours,
far longer than the milliseconds over which action potentials fire. We need to un-
derstand how processes which operate at very different speeds can integrate with
each other and what the dynamics of such connections might be.
228 D. J. MacGregor. G. Leng. D. Brown
The neuroendocrine system forms the brain 's interface to much of the body's
int~rnal systems and its function is essential to many fundamental processes, par-
ticularly development and reproduction. Better understanding of these systems
will make us much more able to diagnose and treat patients with hormonal disor-
ders. Many children suffer from deficient GH release and in rarer cases over-
release of GH. We already have artificial peptides that can stimulate and control
GH release but gre ater understanding will allow us to administer them more effec-
tively and also to more finely diagnose the problems in individual cases.
3 GH Research
male female
- U
0.8
0.4
el
E
UJ
Z
O
0.8
c
- ~
O
~
o:: 0.4
O
I
I O .,.. AA. ft ,.,Ăe ~
1-
3: 0.8
O
o::
~ 0.4
O
O 2 4 6 8
~.4
O 2 4 6
, ~
8
TIME (h)
Fig. 1 : Redrawn from [28] GH release in conscious male and female rats.
Modelling the GH Release System 229
GH is released from the anterior pituitary under the control of two hypotha-
lamic peptides, GH releasing hormone (GHRH) and somatostatin. GHRH stimu-
lates GH release and also synthesis. Somatostatin inhibits release by blocking the
pathway by which GHRH acts. In this way, GHRH is responsible for individual
bursts of GH release but somatostatin seems to be responsible for the overall pat-
tern of release by exerting permissive control over the somatotroph release re-
sponse, and probably also through hypothalamic mechanisms.
These peptides are synthesised by two groups of hypothalamic cells, the GHRH
neurons in the arcuate nucleus and the somatostatin neurons in the periventricular
nucleus. These neurons have axons that project to the median eminence where
they release the peptides into the portal blood supply to be carried to the pituitary.
It was these projections that helped to identify the two groups of neurons. Unfor-
tunately the portal supply' s very small volume makes its content very difficult to
measure with any decent temporal resolution. In addition to this, GHRH neuron
activity is knocked out by anaesthesia making it impossible to record endogenous
behaviour. Recording from somatostatin cells is particularly difficult because they
are spread in a very thin layer. These problems mean that there is no direct way to
discover the real patterns of GHRH and somatostatin release in the working sys-
tem and so more indirect methods have to be employed.
The experiments that have been used to investigate the system can be divided into
three groups, behavioural, electrophysiological and anatomical. Behavioural ex-
periments look at how the system as a whole works, and so the final output, GH
release into the bloodstream, is measured. These are invivo experiments, usually in
conscious animals, which consist of adding or knocking out substances or path-
ways to test what effect they have on GH release.
In electrophysiological experiments we go directly to the neurons thought to be
involved in the system to measure their activity. Although the ability to do this is
limited in the GH system, we do still have the options of recording and stimulating
the GHRH neurons and stimulating the somatostatin neurons, since stimulation
requires less accurate electrode placement. We can still test the GHRH neurons'
response to artificial stimulation in order to determine their properties even though
we can't record normal activity.
Anatomical experiments investigate the system at a lower level, determinmg
the properties of individual components, such as groups of cells, and the connec-
tions between them. This will include a range of molecular and cellular techniques
used to find out what types of receptors each group of cells has, what substances
act on them and what substances they release. This is probably the most difficult
level at which to draw any definite conclusions and individual experimental re-
sults will really only give clues. A single part of the brain can hold groups of cells
performing many different tasks and it can be difficult to be sure that results relate
to the system under investigation. Many substances will appear to act on the cells
being examined, but only some of these will pIay a real role in the action of our
system. The best evidence comes when severai experiments point to the same
230 D. 1. MacGregor, G. Leng, D. Brown
conclusion from different directions, i.e. we find the receptors for a substance, the
substance has a functionally useful effect on the cells, and another group of cells
in our system releases this substance.
The central adrenergic system also appears to play a role in GH release, possi-
bly acting as direct input to GHRH neurons. The centrally acting adrenergic recep-
tor agonist clonidine increases GH release [13] and receptor antagonists [14] and
adrenaline synthesis inhibitors [15] block GH release. The action of clonidine is
blocked by GHRH antisera, but not somatostatin antisera, pointing at the GHRH
site of action.
We can't measure endogenous GHRH neuron activity but what has been done is
to stimulate the GHRH neurons with varying pattems of signal and measure the
GH release in response. By comparing this with normal GH release we can at least
narrow down the range of feasible pattems of activity that might exist in the natu-
ral system, although the detail of any conclusions is limited by the temporal reso-
lution at which we can measure GH release, on a scale of minutes compared to the
much finer timescale on which electrophysiological experiments can work.
Stimulating the GHRH neurons at varied pulse frequencies shows no difference
in GH release levels but increasing the number of pulses produces a large non-
linear increase in GH release [16]. Giving larger doses of GHRH directly to the pi-
tuitary does increase GH release but not in the same non-linear fashion suggesting
that the non-linear relation is at the le vei of the hypothalamus or may also involve
somatostatin.
The other type of experiment which has been conducted is stimulating the
periventricular nucleus and testing the response in the arcuate neurons [17]. Prob-
able GHRH neurons are identified among the arcuate cells by testing for a re-
sponse to stimulation at the axonal terminals in the median eminence. Most of the
GHRH neurons are inhibited during periventricular stimulation and also show a
rebound hyperactivation after the periventricular stimulation.
sults tell us that GH pulse amplitude is related to the size of GHRH pulse and also
that each GH pulse requires a corresponding pulse of GHRH release.
If somatostatin alone is knocked out, the low basal GH secretion between bursts
in males is increased [20], suggesting that somatostatin is responsible for the re-
duced GH release, rather than lower GHRH activity. This also means that there is
either a certain level of GH release from the pituitary independent of GHRH, but
modulated by somatostatin or that there is stiH GHRH activity between GH bursts
which is being controlled by somatostatin either at the hypothalamic level or at the
pituitary.
Giving GHRH injections to male rats without knocking out endogenous soma-
tostatin produces very variable GH pulses. Female rats, however, produce more
regular GH pulses in response ta GHRH. These are smaller than the largest GH
pulses observed in males but are also larger than the smallest GH responses in
males. If a female rat is given three-hourly injections of GHRH over several days
it wiH produce a pattern of GH release very similar ta the natural pattern observed
in males [21]. The animals become entrained into the artificially stimulated pattern
eventually showing no endogenous release between induced GH pulses. This
would require a change in the patterns of hypothalamic GHRH and somatostatin
release and sa there must be some hypothalamic action either by GHRH directly
or through some other feedback mechanism.
If a male rat is given repeated injections of GHRH it will continue ta only pro-
duce a large GH response every 3 h., even with much more frequent GHRH injec-
tions [22]. This refractory period is not likely to be due to a depletion of GH at the
pituitary since it takes 24 h. of continuous GHRH before significantly depleting
pituitary GH content, and female rats continue to respond to even frequent GHRH
injections. It is more likely that the refractory period is due to a cyclic pattern of
somatostatin release.
When male rats are infused with a sufficient doses of somatostatin, GH release
is abolished [23]. At a lower does the endogenous pattern of release remains, but
with smaller amplitude GH pulses. When the somatostatin infusion is stopped
there is a large rebound pulse of GH release, which increases in size with the
somatostatin dosage. A small rebound GH release in response ta somatostatin
withdrawal can be observed with pituitary preparations invitro, but this is of much
smaller magnitude than the invivo rebound. The rebound effect is reduced by the
addition of GHRH antiserum and also by urethane anaesthesia [24], which knocks
out hypothalamic GHRH activity, suggesting an active hypothalamic component
rather than just the removal of the somatostatin inhibitory influence at the pitui-
tary.
If female rats are given a three-hourly pattern of somatostatin infusions, on for
150 min and off for 30 min they produce a pattern of GH release very similar ta
the male pattern. Using a more sinusoidal pattern of somatostatin does produce
three-hourly bursts of GH release but more extended and of lower amplitude than
those observed in the male. These results alI support the idea of high levels of
somatostatin between GH bursts, the inhibitory effects of which start and stop
fairly abruptly, either because of the dynamics of somatostatin's effect within the
hypothalamus, or directly because of its release pattern.
Modelling the GH Re1ease System 233
BehaviounJ experiments have also demonstrated that GH can have a major ef-
fect on its own release. Giving an infusion of GH to male or female rats sup-
presses the endogenous pattern of release [25]. However, unlike GHRH and soma-
tostatin which act fairly immediately, the effect is over quite a long timescale,
taking in the range of 1 h. to develop, and 1 to 2 h. for normal release to recover
after the infusion.
What is uncertain is whether the suppression works by increasing somatostatin
release or by inhibiting GHRH release. In female rats, when GHRH injections'are
given during a GH infusion there is still a large GH response to each injection, al-
though occasionally with some reduction in amplitude compared to injections be-
fore the GH infusion [26]. If somatostatin is infused then there is no response to
the GHRH injections. However, in male rats, which are always only intermittent1y
responsive to GHRH injections, only one GHRH injection produces a GH pulse
during a 6 h. infusion of GH. The response they do give, compared to responses
before and after the infusion, seems to follow the natural three-hourly pattern of
GHRH responsiveness, but with an extended refractory period. This suggests more
than just suppression of GHRH by the GH infusion, perhaps prolonged soma-
tostatin release. Smaller pulses of GHRH during GH infusion in females still pro-
duce large GH pulses, evidence against somatostatin stimulation. The response to
GH seems to vary across gender, which given the general variation in GH release,
makes GH a good candidate for an endogenous feedback mechanism. There are
other substances such as IGF (insulin-like growth factor) which are also thought to
mediate GH feedback but these act too slowly to directly control the normal re-
lease pattern. If GH is given in more natural injections, rather than infusion, then
after one or two three-hourly injections, the rats become entrained in a similar
manner to repeated GHRH injections, producing pulses of GH in synchronisation
with the injections [27]. More frequent 1.5-hourly injections cause the period be-
tween GH pulses to increase, instead of synchronising to the more frequent pat-
tern. This suggests that it may be GH itself that triggers the 3 h. refractory period.
The first stage in building a system model is to define the desired output. The
model GH system needs to be able to reproduce the pattern of GH release in the
rat. The male pattern is the best one to use initially, sin.::e this is a better defined
pattern, but we would also hope that it can easily be adapted to reproduce the fe-
male pattern. The GH system in the female is likely to be similar with only a few
variations which produce the different behaviour. GH release in the male rat oc-
curs in three-hourly bursts of severallarge pulses of GH over roughly 1 h. (the pe-
riod of apparent pulse activity varies in different results). Between the bursts there
is very low basal release of GH. In modelling a whole system at this level it is
more important to be able to reproduce the characteristics of the real data rather
than exact quantitative details.
The next stage is to lay out the system' s components and what we know of how
they behave and interact with each other.
234 D. J. MacGregor, G. Leng, D. Brown
hypothalamus
+
Portal blood
supply
This ba sic model is used to try and think of a working system. The pattern of
output repeats in a cycle and so in the absence of any outside generator producing
this pattern the system needs to have its own cycle. GHRH triggers GH release,
which feeds back to increase somatostatin release, which would inhibit GHRH and
GH release. EventualIy the somatostatin stimulated by GH would falI away and
GHRH would once again trigger GH and repeat the cycle. However, what this is
missing is an input for the GHRH pulses, and so we go back to the experimental
data. Here we find two possibilities, an adrenergic input and the rebound GHRH
Modelling the GH Release System 235
4.1 Simplifications
This is a model of the system at a functional rather than mechanicallevel, and ma-
jor simplifications have been made in representing the system's components.
GHRH and somatostatin are both released and controUed by groups of thousands
of neurons but each is represented by a single variable which measures the level of
the peptide in the system. We are assuming that each group of neurons can be
treated as a single unit. This works in this system for several reasons, the main one
being the timescale of the model. The model works on a scale of minutes, whereas
neurons work on milliseconds and so activity is averaged out. We know that the
neurons do synchronise because this is necessary in any system to produce pulsa-
tile output. Without synchronisation changes in overaU activity will be much less
sudden. The other major reason is the way in which electrical activity is trans-
duced into hormone release. Although individual action potentials trigger individ-
ual releases of peptide, the release from individual neurons aU diffuses into a
common transport channel to the pituitary, and so the GH releasing cells experi-
ence overall GHRH activity rather than the actions of single cells.
The state of the system is stored as a set of variables that represent measures in the
system such as hormone blood level, electrical activity or number of free recep-
tors. The functions of each component may be further broken down into more
variables but this will only be done with the aim of reproducing the component's
behaviour rather than just its mechanism. The basic GH model based on the dia-
gram above has five variables, representing the blood concentrations of GHRH,
GH and somatostatin, the level of releasable somatostatin and the number of free
GHRH receptors at the pituitary. This last variable comes from previous work,
which developed a model of the pituitary component of the system, relating GH
release to the levels of GHRH and somatostatin.
Each variable has a corresponding differential equation which models its be-
haviour and these equations will contain parameters which relate to some measure
such as a threshold or synaptic strength or may belong to a more obscure mathe-
236 D. 1. MacGregor, G. Leng, D. Brown
matical construction. The model also contains other equations, which calculate
values such as level of receptor activation for use in the differential equations.
Hill equations are used to model the actions of the peptides on the groups of
neurons. In general, these equations model the effects of ligand-receptor binding
and allow a variable threshold for activation and variable steepness using the Hill
coefficient which measures the degree of cooperativity of the binding. If binding
is cooperative, then the binding rate is affected by the current level of bound com-
plex, so if the effect is positive, the binding rate will increase as the level of bound
complex increases. If the effect is negative then the binding rate is reduced. Essen-
tially, this controls the steepness with which the level of binding increases. These
equations give a biologically realistic, but still simple, way of getting a measure of
activation where there is a substance binding to receptors. The value ranges from O
to 1, with a higher value indicating a higher level of binding, or activation.
For this model a type of variable has been developed which represents a measure
of the ability to release a substance, or 'releasability'. This is deliberately vague
because it is intended to represent behaviour rather than a particular mechanism.
The real biological substrate could be something like vesicle mobilisation or the
number of activatable receptors. The idea is that they allow a substance to charge
up the ability to release something without directly triggering its release. They
form a kind of memory that allows a substance to trigger an effect which takes
place over a longer timescale or at a different time point to when the original trig-
ger substance was experienced. This was originally developed in order to model
rebound effects. The storage variable could be charged up during inhibition while
its release was blocked and then allowed to drain quickly after inhibition to pro-
duce a large rebound spike. It has also been used to model the effect of GH on
somatostatin, by getting GH to charge up a somatostatin release variable so that
the relatively short period of GH pulses can trigger a much longer period of soma-
tostatin release.
Even fairly complete models of self-contained systems usually need some sort of
external input to control them. When the system of equations is run as a computer
program the variables are progressively calculated at discrete time points. Inputs
are usually given to the system by perturbing the variables, changing them to a
specified value at a specific time point. A whole series of inputs can be defined in
order to form a pattern, such as a series of spikes and this whole pattern of inputs
is known as the protocol. Each variable can have its own protocol but usually only
a few of them will be controlled in this way.
Modelling the GH Release System 237
6 The Model
The concentrations of GHRH (r) and somatostatin (s) are modelled by equations,
which represent release rate and the rate of clearance in the bloodstream:
dr
-=1, -k6 r (1)
dt
ds
-=1 -k s (2)
dt s 7
1, and I, are the release rates. The values k6 and k7 give the clearance rates, mod-
elled as proportional to the current concentration.
The effect of somatostatin on the processes of receptor recycling and GH re-
lease is modelled by calculating a level of somatostatin activation <I>(s), which is a
non-linear function of somatostatin concentration :
1
<D(s)=----- (3)
1 + e -(loglO S-So) / "'o
This equation produces a form of sigmoid. The level of activation varies be-
tween O and 1. After s reaches a certain value, controlled by s", the activation
value will increase quickly and then level off at a plateau. In many biological sys-
tems the effect of a substance varies as its log rather than linearly, i.e. at low con-
centrations a small increase has a large effect, but the same increase at a larger
238 D. J. MacGregor, G. Leng, D. Brown
concentration has less effect, and so the base 10 log of sis used instead of just s it-
self.
The number of free GHRH receptors, f, is modelled by :
dh
-=[k4 +k,(l-<l>(s))][r+c]f (5)
dt .
This models GH release as the product of the rate of receptor binding, (r + c)f
and asum that allows release at rate k4 independently of somatostatin and extra re-
lease at rate ks which is modified by somatostatin activation. Normally k, would be
much larger than k4 making the somatostatin blockable component much more
significant.
The full model adds to the pituitary model the hypothalamic connections and
components which control the GHRH and somatostatin neurons. One new storage
variable, S" is used to give a measure of the ability of the periventricular soma-
tostatin neurons to release somatostatin. The model adds two new connections, an
inhibitory link from somatostatin to the GHRH neurons and a feedback link from
GH to somatostatin that increases the level of releasable somatostatin. It also adds
a delay to the GH level representing the time it takes from pituitary release to have
an effect on the somatostatin neurons. Without this delay GH will effectively auto-
inhibit by causing the somatostatin level to immediately rise, making the GH pulse
much smaller and shorter. The new links are modelled by two Hill equations, ra"
and Saci' which give the level of activation at the GHRH and somatostatin neurons
respectively:
(6)
(7)
ModelJing the GH Release System 239
The parameters thI and th2 give the thresholds for activation (really the level
for half maximal receptor binding) and nI and n2 are the Hill coefficients giving
the steepness with which the level of activation increases. The r equation is sub-
acl
tracted from 1 because it is being used for inhibition. The new equation for GHRH
release adds the connection from somatostatin by allowing r to modify the input
acl
activi ty 1,.
dr
dt = Irract - k 6r (8)
The level of somatostatin activation by GH, h'lCl is used in the new s, equation :
(9)
(10)
Combined with the pituitary model these equations implement the basic GH
system model described in Fig. 3.
240 D. J. MacGregor, G. Leng, D. Brown
Random
adrenergic constant
~ Electrical inputs
~
B~
, ~----------J>_ G ..-- +
G
+
Fig. 3. The GH system model. The model uses a previous model of the pituitary GH release
response to GHRH and somatostatin, from earlier work by Elinor Stephens. It includes, in
addition to the basic components, a random adrenergic input to the GHRH neurons, con-
stant electrica! input to the somatostatin neurons, a variable representing the releasable
store of somatostatin, s" an inhibitory connection from the GHRH to somatostatin neurons
and a delay value for GH feedback. This figure also shows the experimental inhibitory link
from GHRH to somatostatin
The set of equations has been implemented as a computer program that runs the
equations over a fixed number of time steps, recording the value of each variable
at each step. Each variable can then be displayed as a graph which shows how it
changes with time and the graphs can be viewed together to examine how the be-
haviour of the variables relate to each other. The model' s parameters and the input
protocols can be altered and the model immediately run again to show what effect
the change has. The equations themselves can be altered by editing the program
code and then restarting the model program.
Modelling the GH Release System 241
The parameters from the pituitary model were fixed by previous work which fitted
the pituitary model to real invitro experimental data. The pituitary model is able to
match the dynamics of its individual components in real data and also the behav-
iour of the pituitary GH release system as a whole, specifically desensitisation to
GHRH, and the smaller pituitary version of rebound GH release following soma-
tostatin withdrawal. The pituitary model and its parameter values are assumed to
be a sound basis on which to build and test the hypothalamic components of the
model.
The experimental data on the hypothalarnic systems only indicate the existence
and overall effect of individual substances or connections and so there are no data
from which to direct1y derive values for the hypothalamic parameters. The values
have had to be determined by altering them so that the model produces the desired
pattern of GH output with plausible patterns of GHRH and somatostatin. It has
been attempted to retain biological plausibility with alI the parameter values. The
Hill equations can be made to behave like switches if the Hill coefficients are very
large but the dynarnics of individual receptor-triggered processes are seldom this
sharp and so the coefficients have been kept at small values. The thresholds of
these equations have been set so that an appropriate level of the triggering sub-
stance (somatostatin or GH, in the current model) is able to get the desired re-
sponse, i.e. blocking GHRH release or sufficient1y charging releasable soma-
tostatin. It is also desirable to put parameter values in a range where small changes
do not have a major effect on behaviour. Biological systems need to be robust and
the model should be the same. This characteristic is often a good indicator that a
model is a good lepresentation of the real biological system.
Many of the parameters will depend on each other, hence there are likely to be
many sets of parameters that will produce the same behaviour. This goes some
way to limit the range of behaviour but it will still be vast, increasing exponen-
tially with the number of parameters. Testing the model' s range of behaviour, be-
yond the target pattern of GH output, through varying the parameter values is an
essential part of the experimentation in order to understand the significance of the
model's components and their parameter values. What is desirable is some auto-
mated system to search the whole space of alI possible parameter values and dis-
cover the model's full range of behaviour. To do this we need to define precisely
how to assess the model's performance so that it can be implemented as part of the
model program. The important characteristics of the GH output pattern need to be
defined.
The goal of the model is to be able to reproduce the pattern of GH release in the
rat, Modelling is mainly concerned with reproducing the qualitative characteristics
of the data but it should at least show the same relative differences in levels of
hormone. A set of characteristics must be defined that is sufficient to differentiate
the patterns that meet the requirements from those that do not. Since this is nu-
242 D. 1. MacGregor, G. Leng, D. Brown
meric data, these will at least need to start out as measurements, so we list what
measurements might determine a male rat type pattern :
• Pulse amplitude
• Pulse duration
• Burst duration
• Number of pulses per burst
• Inter-burst GH level
• Inter-burst duration
The first stage is probably deciding whether regular bursts exist at alI. If this is
established then the individual bursts can be identified and then the measurements
can be made. A male rat type pattern requires approximately three-hourly bursts
with several high amplitude pulses and very low levels of release between bursts.
The female pattern may not have regular bursts but should at least have fairly fre-
quent pulses of at least a third of the amplitude seen in the males. These judge-
ments are easy to make with the eye but are difficult to sufficient1y formali se for
the sake of conventional computer implementation.
The basic version of the GH system model uses the equations described above
with regular patterns of input to the GHRH and somatostatin neurons to success-
fully produce a pattern of GH release similar to the male rat. The somatostatin
neurons are given constant fixed stimulation so they release somatostatin when-
ever s, is charged, i.e. somatostatin ready for release after being charged by GH.
The GHRH neurons get a regular pattern of short pulses, 1 or 2 min long and usu-
ally spaced 15 min apart. These pulses reproduce the multiple peaks observed in
GH release but they do not directly control the timing of the GH bursts. Fig. 4 il-
lustrates the model output.
The model begins running with somatostatin already charged to a low level and
so the frrst GHRH pulses are inhibited, but when the somatostatin level falls as s,
discharges the Somatostatin-GHRH inhibition is removed and pulses of GHRH
are released triggering several pulses of GH release from the pituitary. After the
30 min delay the GH pulses cause s, to recharge and the somatostatin level in-
creases again, ending the GH burst.
In Fig. 5 the model produces a similar pattern of GH release to the male rat. It
shows the same multiple pulse bursts of GH release following a similar timescale.
In Fig. 5 there is approximately 2.5 h. between bursts, but by increasing the delay
to 40 mins the model can be made to match the three-hourly pattern. How closely
we can analyse the shape is limited by the resolution of real data. Most results
measure the GH level at intervals of at least 5 mins, and so although we do know
that there are multiple peaks, we do not know how sharp these really are or how
Modelling the GH ReIease System 243
suddenly the GH burst starts and finishes. The one difference with the model that
is apparent is that in real data the first pulse in the burst is often the largest
whereas in the model the remaining somatostatin when the first GHRH pulses fire,
causes the first pulses to be smaller. This is less apparent when the model uses
random intervals for the GHRH input pulses and the random pulses make a better
match to real data. The effect is mainly controlled by the Hill equation which
models the somatostatin-GHRH link, raCl. If the binding coefficient, nI is made
larger then the inhibition is much sharper, acting like a switch and eliminating the
partially inhibited GHRH pulses.
R.n ..
,O),
.r,
'",o),
:~1 )/. III
'IlO
'" ",
orn
HO
"ca
• ., ", 'OI ",
."',rn .
".---------~--------
o !)l. II)) ,~ ,:xl
". ,
1(°1
~: i ~--- -r'''': _
o lCO l!iO ZXI
4(Oi
".
+
•
~lO
'DO
Fig. 4. The model program. This shows the software developed to run the model, running
on a PC in Windows. The main window displays graphs of the variables as the model's
outpUL The boxes to the right allow aII of the model's parameters and input protocols to be
modified. The boxes to the Iert of the graphs allow the scale to be altered, allowing to the
viewer to zoom in or out on the data. The info box displays the values pointed at by the
mouse
244 D. J. MacGregor, G. Leng, D. Brown
! \1 111111111111111111111111111111111 i ~I IIIIIIIIIIII~IIII i 1
HI =r:rvJ HI=- ~ ~ d
Il ~ Il ~ ~
oo
,J.
100 200
,J.
Thle (mlnul")
300 400
I
500 OO
J
100 200 300
Tlmo(mlnutos)
400
Fig. S. Results using the current basic version of the model, regular (Ieft) and random poi-
500
I
son (right) pulses of input to the GHRH neurons, GH feedback direct to s" and no link
from GHRH to somatostatin. GHRH pulse interval 15 min, GH feedback delay 30 min.
This gives a regular pattern of GH release bursts, with spacing similar to real release pat-
terns in the male rat. With the random input pulses the bursts vary in shape and pulse am-
plitude but the three-hourly pattern remains
There is indirect experimental evidence for an inhibitory link from GHRH neurons
to somatostatin neurons. It was not obvious without testing what effect this would
have on the model.
A new Hill equation, saCI' was used to model the action of GHRH on soma-
tostatin :
n3
sacI
=1- 3
r 3 (11)
r n + th3 n
This is similar to raCl' with its own parameters for threshold and binding coeffi-
cient, th3 and n3. This is added to the somatostatin equation by allowing saCI to
modify the release component :
(12)
Modelling the GH Release System 245
The effect of this new component has been to make GHRH bistable, producing ei-
ther large or very small pulses rather than a range controlled by the somatostatin
inhibition. This is similar to making the somatostatin inhibition sharper, but makes
the partiaIly inhibited GHRH pulses larger instead of removing them. Small
GHRH pulses are too small to inhibit somatostatin but larger pulses that are par-
tially inhibited are able to inhibit somatostatin, thus removing their own inhibition
and becoming very large pulses. This in turn increases the level of GH release and
so increases the GH effect on somatostatin prolonging the period between bursts
(Fig. 6). This works particularly well with random GHRH input pulses, countering
the variation in GHRH pulse size and making the GH bursts much more regular in
shape and spacing.
i 1 .111 II II 1 i 1I I I I III
i~l~J\~B I~l ~ ~0J
il l ~ Â i~[ . l ~ ~I
oo 100 200 300
T1mo (nWMrt..)
_
I
500 oo 100 200 300
T1mo (_ )
_ 500
Figure 6. As Fig. 5 but with the addition of an inhibitory link from the GHRH to soma-
tostatin neurons, suggested by experimental results. This creates a bistability in GHRH ac-
tivity, which produces much larger GHRH pulses and in turn larger pulses of GH release. It
slightly increases the time between GH bursts
tween pulses within the GH burst. The model has been used to investigate this by
adding a new GH activation equation, hacr2 and allowing this to modify the stimula-
tory component of the somatostatin equation.
h = hdelay2
n4
(13)
act2 hdelay2
n4
+ th4
n4
(14)
By using the product to modify the re1ease component this gives GH fuU con-
trol over somatostatin release, since 1, is just a constant. To combine the two rather
than replace the sum would be used instead. Initial testing with the model bears
out the predictions. The long bursts of GH release are replaced by frequent, short,
low amplitude pulses. GH inhibits itself and also Ioses the Iong periods of Iow ac-
tivity since there is no stimulus to maintain somatostatin release between GH
puIses (figure 7). However, there is an interesting effect if th4 is made very small
so that the Iow basallevels of GH release can stimulate somatostatin release. This
retums burst release and also extends the periods of low GH release by making the
rate at which somatostatin release decays more linear. The decay rate will nor-
maUy be non-linear because re1ease is proportional to the amount of charge left on
S, but at the same time GH release is gradually decreasing with the faU in soma-
tostatin. A1lowing GH to stimulate means that, while the pool of reIeasabIe soma-
tostatin is decreasing, the stimulation is increasing and the two effects combine to
create a longer more linear dec1ine in somatostatin release. It is unlike1y that such
low levels of GH have an effect in the real system but this is a good illustration of
how a model can find unexpected behaviour.
8 Conclusions
Modelling at the system level provides a powerful tool for making use of experi-
mental data. It forces the conc1usions from real data to be formali sed and provides
a structure for testing these conc1usions. It illustrates how important it is to assess
components as part of a whole system rather than in isolation and it provides a
method for doing this. When a model system is built it can then be used to predict
the results of experiments so that we can better direct investigations in the real
system or simulate experiments which cannot be done in animal models. Model-
ling is the practical form of the more functional rather than implementation based
language which is required to examine the brain at a higher level, looking at whole
systems rather than just biological components. It carries the risk of oversimplify-
ing, but this should not be a problem so long as the model' s design has a proper
basis in experimental results. BioIogy is elegant and much of the apparent hope-
Modelling the GH Release System 247
less complexity disappears when pieces are successfuIly distilled down to their
functional descriptions.
~ ~ f ~1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 i1 1
II II 11111111 II IIII I
"
~l+-LI-"-'l . L. !JI I~ I li. .,
J:
a:
J:
It oo ,'00" , "'200"
line (mlooos)
" 300 ' 400
I
500
~~I ~
oo '00 200
A
:)00
line (minute.)
400
I
500
Fig. 7 : Direct GH feedback to somatostatin. On the left, for comparison the same random
input pattern as Fig.5 is used. The tonic somatostatin electric al stimulation is replaced by
direct stimulation from GH. Making somatostatin release entirely dependent on GH feed-
back breaks the ability to maintain a prolonged somatostatin release and the immediate in-
crease in somatostatin release with each GH pulse causes the pulses to be shortened and
have smaller amplitude. Together these cause the loss of pulsatile release. However, on the
right, if th4 is made small enough for basal GH to stimulate somatostatin then pulsatile re-
lease returns
The model of the GH systern shows that only a few components, using very
simple dynamics, are sufficient to produce the pulsatile pattern of release. It is also
able to suggest what effects other components or connections might have on the
release pattern. It indicates that we need more experimental data in particular on
how GH affects somatostatin release, since this connections appear to play the
most important role in mediating the long delay between bursts of GH release. The
model provides better possible explanations for experimental results, such as the
non-linear increase in GH release with increasing stimulation of the GHRH neu-
rons. With the inhibitory connection from GHRH to somatostatin, stimulation of
GHRH increases GHRH release and also decreases somatostatin release, causing
the non-linear increase in GH release. This suggests an experiment repeating the
GHRH stimulation, with the addition of somatostatin antiserum to test whether the
non-linear increase in GH remains.
248 D. J. MacGregor, G. Leng, D. Brown
References
1. Sawchenko, P. E., Swanson, L. W., Rivier, 1., Vale, W. W. (1985). "The distribution
of GH releasing factor (GRF) immunoreactivity in the CNS of the rat: An immunohis-
tochemical study using antisera directed against rat hypothalamic GRF." Journal of
Comp. Neurology 237: 100-115.
2. Wiegand, S. J. and Price, J. L. (1980). "The ceUs of origin of the afferent fibres to the
median eminence of the rat." Journal ofComp. Neurology 192:1-19.
3. Dierickx, K. and Vandesande, F. (1979). "Immunocytochemicallocalization of soma-
tostatin containing neurons in the rat hypothalamus." Cell Tissue Research 201: 349-
359.
4. Aguila, M. C. and McCann, S. M. (1993). "GH increases somatostatin rele ase and
messenger ribonucleic acid levels in the rat hypothalamus." Brain Research 623: 89-
94.
5. Kamegai, J., Minami, S., Sugihara, H., Higuchi, H. and Wakabayashi, 1. (1994) "GH
induces expression of the c-fos gene on hypothalamic neurons in hypophysectomized
rats." Endocrinology 135: 2765-2771.
6. Epelbaum, J., Moyse, E., Tannenbaum, G. S., Kordon, C. and Beaudet, A. (1989).
"Combined autoradiographic and immunohistochemical evidence for an association of
somatostatin binding sites with GH-realising factor-containing nerve ceU bodies in the
rat arcuate nuc1eus." Journal of Endocrinology 1: 109-115.
7. Meister, B., Hokfelt, T., Vale, W. W. and Goldstein, M. (1985). "GH- releasing factor
(GRF) and dopamine coexist in hypothalamic arcuate neurons." Acta Physiologica
Scandinavica 124: 133-136.
8. Meister, B. and Hokfelt, T. (1988). "Peptide- and transmitter-containing neurons in the
mediobasal hypothalamus and their relation to the GABAergic systems: possible roles
in control of prolactin and GH secretion." Synapse 1: 585-605.
9. Meister, B., Hokfelt, T., Vale, W. W., Sawchenko, P. E., Swanson, L. W. and Gold-
stein, M. (1986). "Coexistence of tyrosine hydroxylase and GH- releasing factor in a
sub-population of tuberoinfundibular neurons of the rat." Neuroendocrinology 42:
237-247.
10. Wakabayashi, 1., Miyazawa, Y., Kanda, M., Miki, N, Demura, R., Demura, H. and
Shizume, K. (1977). "Stimulation of immunoreactive somatostatin rele ase from hypo-
thalamic synaptosomes by high K+ and dopamine." Endocrinology IJm. 24:601-605.
Il. Kitajima, N., Chihara, K., Abe, H., Okimura, Y., Fujii, Y., Sato, M., Shakutsui, S.,
Wantanabe, M. and Fugita, T. (1989). "Effects of dopamine on immunoreactive GH-
releasing factor and somatostatin secretion from rat hypothalamic slices in vitro." En-
docrinology 124: 69-76.
12. Locatelli, V., TorseUo, A., Redaelli, M., Ghigo, E., Massara, F. and Muller, E. E.
(1986). "Cholinergic agonist and antagonist drugs modulate the GH response to GH-
releasing hormone in the rat: evidence for mediation by somatostatin." Journal of En-
docrinology 111: 271-278.
13. Day, T. A. and Willoughby, J. O. (1980). "Noradrenergic afferents to the median emi-
nence: inhibitory role in rhythmic GH secretion." Brain Research 202: 335-345.
14. Amold M. A. and Femstrom, J. D. (1980). "Administration of antisomatostatin serum
to rats reverses the inhibition of pulsatile GH secretion produced by injection of me-
tergoline but not yohimbine." Neuroendocrinology 31: 194-199.
Modelling the GH Release System 249
15. Terry, L c., Crowley, W.R. and Johnson, M.D. (1982). "Regulation of episodic GH
secretion by the central epinephrine system." Journal of Clinical Invest. 69: 104-112.
16. Dickson, S. L., Leng, G. and Robinson, 1. (1993). "Growth-Hormone Release Evoked
By Electrical-Stimulation of the Arcuate Nucleus in Anaesthetized Male-Rats." Brain
Research 623(1): 95-100.
17. Dickson, S. L., Leng, G. and Robinson, 1. (1994). "Electrical-Stimulation of the Rat
Periventricular Nucleus Influences the Activity of Hypothalamic Arcuate Neurons."
Joumal of Neuroendocrinology 6(4): 359-367.
18. Wehrenberg, W. B., Brazeau, P., Luben, R., Bohlen, P. and Guillemin, R. (1982). "In-
hibition of the Pulsatile Secretion of Growth-Hormone By Monoclonal-Antibodies to
the Hypothalamic Growth-Hormone Releasing- Factor (Grf)." Endocrinology 111(6):
2147-2148.
19. Wehrenberg, W. B., Brazeau, P., Luben, R., Ling, N. and Guillemin, R. (1983). "A
Non-Invasive Functional Lesion of the Hypothalamo-Pituitary Axis For the Study of
GH-Releasing Factor." Neuroendocrinology 36(6): 489-491.
20. Ferland, L., Labrie, F., Jobin, M., Arimura, A. and Schally, A. V. (1976). "Physiologi-
cal role of somatostatin in the control of GH and thyrotropin secretion." Biochemical
& Biophysical Research Communications 68(1): 149-56.
21. Clark, R. G. and Robinson, 1. C. A. F. (1985a). "Effects of a Fragment of Human GH-
Releasing Factor in Normal and Little Mice." Journal of Endocrinology 106(1): 1-5.
22. Clark, R. G. and Robinson, 1. C. A. F. (l985b). "Growth-Hormone Responses to Mul-
tiple Injections of a Fragment of Human GH-Releasing Factor in Conscious Male and
Female Rats." Joumal ofEndocrinology 106(3): 281-289.
23. Clark, R. G. and Robinson, 1. C. A. F. (1987). "Growth-Hormone (GH) and Body-
Growth Responses to Intermittent Somatostatin (SS) Infusions in the Rat." Joumal of
Physiology-London 382: 33.
24. Clark, R. G., Carlsson, L. M. S., Rafferty, B. and Robinson, 1. C. A. F. (l988a). "The
Rebound Release of Growth-Hormone (GH) Following Somatostatin Infusion in Rats
Involves Hypothalamic GH-Releasing Factor Release." Joumal of Endocrinology
119(3): 397-404.
25. Clark, R. G., Carlsson, L. M. S. and Robinson, 1. C. A. F. (l988b). "Growth-Hormone
(GH) Secretion in the Conscious Rat - Negative Feedback of GH On Its Own Re-
lease." Joumal ofEndocrinology 119(2): 201-209.
26. Carlsson, L. M. S., Clark, R. G. and Robinson, 1. C. A. F. (1990). "Sex Difference in
Growth-Hormone Feedback in the Rat." Joumal of Endocrinology 126(1): 27-35.
27. Carlsson, L. M. S. and Jansson, 1. O. (1990). "Endogenous Growth-Hormone (GH)
Secretion in Male-Rats Is Synchronized to Pulsatile GH Infusions Given At 3-Hour In-
tervals." Endocrinology 126(1): 6-10.
28. Clark, R. G. and Robinson, 1. C. A. F. (1988). "Paradoxical growth-promoting effects
induced by pattemed infusions of somatostatin in female rats." Endocrinology 122:
2675-2682.
Hierarchies of Machines
M. Holcombe
Abstract. Computational models have been of interest in biology for many years
and have represented a particular approach to trying to understand biologic al
processes and phenomena from a systems point of view. One of the most natural
and accessible computational models is the state machi ne. These come in a variety
of types and possess a variety of properties. This Chapter discusses some useful
ones and looks at how machines involving simpler machines can be used to build
plausible models of dynamic, reactive and developing biologic al systems which
exhibit hierarchical structures and behaviours.
Much of the early work using state machines and related models for modelling
biologic al processes and systems was rather abstract and high level and probably
seemed, to many, to be of more philosophical than practical value. There have,
however, been some advances in the development of more realistic models and
the current state of computer science research provides us with new opportunities
both through the emergence of models that can model seriously complex systems
but also the support that modem software can give to the modelling process. This
chapter describes a few of the early simple models and then goes on to look at
some new ideas in the area. Some general principles relating to how new and
emerging computational techniques can help us to represent and understand ex-
tremely complex models conclude the paper.
Computational models are models of systems inspired by the model of an in-
formation processing system, in its most common manifestation a digital com-
puter but it is not as restrictive as that in practice.
The principle philosophy for this work is the representation of some aspects of
cellular metabolism as computation of suitably defined data. The main benefit is
the opportunity to make use of some of the theoretical models of general, concur-
rent computing systems in the hope that :
• the computational models may enlighten the biochemical theory
• the analysis of successful parallel biochemical systems will enlighten the the-
ory of parallel computing.
252 M. Holcombe
extern al
...
..
inputs
... internal
states
system
outputs
The most basic discrete model: finite state machine (see Fig. 2).
• a set of internal states
• a set of external inputs - events
• a set of system outputs - actions (this is an optional feature, in some machines
there are no explicit outputs)
• a transition structure to link it alI together
inputs are: a, b, c
outputs are: x, y
How does this model work? RecaB that the system is always in one of the in-
ternal states and state change is caused by the receipt of inputs.
• If in statel and input (event) a occurs then it changes to state2 and outputs x.
• If in state 1 and input b occurs then it changes to state3 and output y occurs.
Hierarchies of Machines 253
The system then waits for the next input to occur before the next state change.
This continues as long as the system continues to function. These models have
been used to analyse metabolic pathway mociels - [1] and we can iIIustrate this for
the simple Krebs (tricarboxylic acid) cycIe, Fig. 3.
If we regard the inputs as being CI, C2 and C3 which are specific co-enzymes
that drive the cycIe and the intermediate substrates as being the states of the sys-
tem then it behaves Iike a state machine.
Enzymes that are required for each reaction are assumed to be present in suffi-
cient concentration.
This is a simple model of organisation which ignores many factors, such as re-
action kinetics, enzyme production, etc. Several quite complex metabolic path-
ways have been modelled in this way.
The algebraic theory of these machines can be used to construct decomposi-
tions into simple components (generated by finite simple groups - a theory that is
now well understood in mathematics). What the theory is saying is that any sys-
254 M. Holcombe
tem of this type can be broken up naturally into a collection of subsystems, manu-
factured from simple mathematical objects, which are joined together as systems
in two main ways - parallel and serial connections. For a full description of the
decomposition theory and its application to the Krebs cycle see Holcombe [2]
which is an extension of the holonomy decomposition theory of EiIenberg, [3]. In
the case of the Krebs cycle the algebraic structure of the machine can be decom-
posed to form an equivalent machine which is buiIt up from cyclic groups of or-
der 2 and 3 connected together as wreath products which are also combined with
some simple aperiodic semigroups of order 2 and 3.
What does this decomposition mean, biochemically? It does not seem immedi-
ately clear and for this reason research in this area seems to have faded away.
However, the impact of a greater understanding of the genomic basis for the pro-
duction of the enzymes to drive the system might throw new light on the problem.
Many systems can be modelled in this way but:
• It is unsuitable for complex systems because of state space explosion.
• The functions represented by these machines are too simple for many situa-
tions.
• Continuous behaviour is not modelled.
• Concurrent systems are not modelled well without large problems.
• Communication is hard to model this way.
Cellular automata, however, are built from these machines and have proved
useful in some cases, for example:
• Models of simple development
• Models of simple ecologies.
The idea of a computational model stems from the work of Turing, and others,
who, before digital computers were ever built wondered exactly what could be
computed by a hypothetical machine, and this led to the identification of a class of
generali sed state machines, Turing machines, which seemed to capture the notion
of computation. Turing machines are directly related to other approaches to the
definition of computation, recursive functions, Markov algorithms etc. and this
provided more evidence that the notion of computation had been defined in a
plausible manner. Recent work on the theory of quantum computers has taken this
work into a new direction.
The idea that a cell, a system of cells or a complete organism can be viewed as
carrying out some form of computation has been a significant factor in the appli-
cation of computational models to the modelling and simulation of biology. We
have seen that a particularly simple computational model, the finite state machine,
can be used to model aspects of a metabolic pathway and this also provides some
interesting connections between biology and computing.
Hierarchies of Machines 255
The machines that we have considered are very simple and thus very limited in
their capabilities and do not represent a realistic basis for modelling biologic al
systems an processes. The Turing machines, on the other hand, are too abstract
and unwieldy to provide a suitable answer. We investigate some new model types
and see how they fare.
2.1 X-machines
outputs
internal
state
.. inputs
((jjITiTIil=~~~",--:- proces.ing
ite
(eg.mitocboodria)
nucleus_4+--~
proce-ssing si e
The analogy with the VLSI model is strengthened when we consider some of
the sorts of processing that might be involved, for example we have simple meta-
bolic processes that might correspond to the behaviour of a register, that is infor-
mation (molecules) might be stored temporarily at a location prior to being further
transformed by another processing site, we might have sites performing fairly
simple processing akin to that of an adder which operates when its input places
contain appropriate values and then there are structures like multipliers that oper-
ate continuously. The overall model of the system could then be determined with
reference to the data types involving the various active sites and the functions op-
erating on the system described at this level.
The fundamental data type will be of the form:
x= INmsl x INms4 x [MSI) x [MS2) x [MS3) x[ MS4) x [N) x [PI) x [P2) x [P3) x
OUTms2 x OUTms3
where each element defines the type of values that the site can deal with and
would be defined in a formal manner following our previous work ([1,5]), INmsl
and INms4 are the input datatypes at msi and ms4 respectively and OUTms2 and
OUTms3 are the output data types at ms2 and ms3 respectively.
The next stage is to identify when the various sites are active and from this we
construct a top level model of the system using the specific processing functions
that determine what happens to the data.
In Fig. 6. we note the existence of functions indicated by the arrows, thus we
can list the functions and the appropriate data types. As a convention let us indi-
cate input datatypes and output datatypes to sites as follows :
For site K we define IN(K) and OUT(K) to be the appropriate types and
COMP(K) : IN(K) -+-> OUT(K) to be the specific processing function (which
Hierarchies of Machines 257
Fig.6. A possible state space describing the functional dynarnics of the ceH
Retuming to our hypothetical ceH, we note that in some cases there are rela-
tionships between types, for example,
The environmental inputs and outputs are: INmsl, INms4, OUTms2, OUTms3.
where:
To model systems that operate concurrently and which communicate with each
other in a more efficient way we introduce a new approach. Consider a number of
separate stream X-machines, Fig. 8, which have the following properties:
1. There are certain communication channels between some of the machi nes.
2. Some of the states in these machi nes are solely used for communicating mes-
sages to other machines.
3. The other states are ordinary processing states.
Each machine works separately and concurrently in an asynchronous manner
until it reaches a communication state.
It then waits to either:
• send a message to another machine.
• or receive a message from another machine which may not be ready to send it.
The receiving machine must then wait for the message.
communicating state
..-.......-... _... _-_.-~-- - -_.-
channelA
machine 2
chaonelB
AII of the machines discussed previously are discrete, so only instantaneous proc-
essing can be modelled and only finite discrete data are processed.
Continuous functions and real valued data cannot be incorporated into tradi-
tional finite state machine models.
The hybrid X-machine (or just hybrid machine) [7], overcomes this.
A hybrid machine has states and transitions as usual and responds to discrete
events and performs discrete actions which are observable. The internal memory
consists of:
• a set of discrete variable
• a set of continuous variables.
When it is in a given state there are sets of equations that apply to the system's
continuous variables and aII the while it is in that state, with time progressing,
these variables change according to these equations.
When either an appropriate external event occurs or a leaving condition is met
(e.g. a set point) the system moves to its next state where a different set of equa-
tions take over, see Fig. 10.
Hierarchies of Machines 261
event/action
eventlaction
Agents are autonomous systems that interact with their environment - and other
agents, using a set of rules to describe their behaviour.
They can be simple reactive systems or they can be very sophisticated with
complex strategies and learning capabilities. The important aspect of them is that
they are a low level phenomenon whose behaviour in their environment emerges
according to the rules. In a different environment the behaviour may change.
Agents can cooperate or compete with one another.
Many biologic al processes seem to behave like agents, an interesting example
being the system of ants and their behaviour as they form trails between the nest
and a food source, Fig. 12.
pheromone
trail ---Ij.
obstacle
ants oest
Fig. 12. A diagram of an insect trail
Suppose that these ants are simple agents that obey rules relating to the concen-
tration of pheromone and communication events with other ants.
Rules in order of priority might be.
1. If detect obstacle then change direction
2. If detect pheromone then follow trail
3. If meet ant with food then continue
4. If find food pick up, turn round and lay trail
5. If meet ant with no food then turn round
6. If true then move randomly
etc.
This agent can be described in full detail using an X-machine model, where
each function is defined by the rules and the associated inputs and memory. Mem-
Hierarchies of Machines 263
ory will be the position; whether with or without food; the direction the agent is
moving and other relevant parameters. See Fig. 13 for a state diagram of such an
agent. The details of the function definitions are left out in this review due to
space limitations.
move
change
direction
move
This will be a hybrid machine. Whilst in the following traiI state the motion
could be determined by standard equations of motion. A refinement is to intro-
duce a probabilistic dimension to the possible state changes.
Other examples of agents can be found in the immune system, e.g. T cells, B
cells etc., different molecular species could also be interpreted in this way.
The next issue is how to model communities of agents. Suppose that we had a
collection of agents, in this case represented as individual hybrid X-machines. We
now have to try to identify the communication channels and how these might
work. Suppose that there are N agents and each is potentially able to communicate
with any other. We thus have an N x N communication matrix.
Suppose that an agent can only communicate with other agents within a spe-
cific distance of it. There is a global memory that maintains the current position of
each agent (an N-vector of coordinates). At any communication state an agent can
interrogate this memory to ascertain which other agents are within communication
distance. A number of strategies can be used to determine which and how many
attempts at a communication can be made. The agent then puts data into the ap-
propriate slots of the communication matrix and continues processing, moving or
whatever.
264 M. Holcombe
Some of these data may time out if the agent it is intended for fails to retrieve
it. Similarly, when an agent is within reach of another and wishes to do so it can
retrieve data from the specific slot. Agents can die in which case both the row and
column in the matrix become empty (void). Agents can be created, in which case
the matrix is expanded to an extra row and column and the memory extended as
appropriate.
Some simple ground rules must cover situations where two agents try to oc-
cupy the same co- ordinates. We should be able to simulate communities of sim-
ple agents in this way - when N becomes very large this may be a problem. Hope-
fully emergent properties can be identified, analytically or numerically. This
needs further investigation.
4 Hierarchies of Machines
Suppose that we are looking at some specific metabolic processing within a cellu-
Iar component. This might be modelled using a simple state machine or an X-
machine. It requires the existence of a number of conditions to be realistic, for ex-
ample, certain enzymes will have to be present in sufficient concentrations at the
location or region of the processing. These are in turn controlled by the genes that
are currently expressed and this is dependent on many other factors including
metabolic activity in other parts of the system, intracellular messages and envi-
ronmental influences.
We could thus consider the following hierarchical sort of model for a simple
cell, Fig. 14:
Hierarchies of Machines 265
external signal
and events
I H~Hlll communication between levels
genetic
regulatio
ba sic metaboli
level
systems. One important factor is what the overall system behaves like. Because it
has been proved that the X-machine model is fuHy general and that systems of
communicating X-machines are still X-machi nes it is plausible for us to treat the
model of the complete cell as an X-machine, albe it one composed of many levels
of intercommunicating systems of communicating X-machines.
One issue that has not yet been addressed is the fundamental fact that ceHs and
organisms do not remain structurally static but develop, divide, differentiate and
ultimately die. We will try to look at this through the issue of modelling tissue.
If we have a model of a cell we need to see how it might fit within a model of the
collection of cells that form a coherent supersystem such as a tissue system.
The first point is that the cell model must reflect the fact that some cells change
their behaviour during their life. This can be modelled by the switching on of
parts of the hierarchy as a result of different events and internal factors. This can
be modelled by using the communication matri x and the memory to control what
parts of the system are active and which parts are inactive or faulty. In the latter
case we need to be able to intervene in the model and to 'damage' some transi-
tion.
For tissue we have another issue. We will start with the model of an individual
ceH and consider how tissues might be modelled. The obvious point is to use the
communicating X-machine approach but there is a problem in how we model the
growth of the tissue from a starting point of a single cell or group of cells. Mitosis
introduces a fundamentally new concept, that of cell division and the treatment of
this in the model needs to be addressed.
Again, the approach is to structure the memory. Here we have a memory that is
based round a single membrane system, together with, possibly other pieces of in-
formation, time, size and position being some possibilities. It each are associated
with a specific ceH then we can describe this as a vector of parameters, one of
which is a membrane system which itself has significant internal structure and
subdivisions. We call this the basic cel! memory structure.
The development of two cells from an existing cel! wil! take place in a particu-
lar state, a mitosis state, during this phase the structure of the memory will
change. The difference between the initial memory value on the entry to the mito-
sis state and the memory value on exit from this state will differ as fol!ows, Fig.
15:
A natural way of dealing with this is to construct a binary tree of the basic cell
memory structure, which is, in effect, the membrane structures and the other pa-
rameters associated with each cell, to represent the process of cel! division which
maintains the information about the individual cell
The state change describing the mitosis in Fig. 15 would ne represented in the
form of the transformation: C CI. C, where CI' C, represents the two new daugh-
Hierarchies of Machines 267
ter cells and the variable values involved in the original ceH C are distributed
amongst the two new ones. More work is needed to produce an effective model of
mitosis.
®O ~O
o
original ceH daughter ceH daughter ceH
Fig. 15. Mitosis
References
1. K. Krohn, R. Langer & J. Rhodes, Algebraic principles for the analysis of a biochemi-
cal system, J. Comp, and System Sci. 1,119-136,1967.
2. M. Holcombe, Algebraic Automata Theory, Cambridge University Press, Cambridge,
1982.
3. S. Eilenberg, Automata, languages and machines, volume, B, Academic Press, New
York, USA. 1976.
4. M. Holcombe & F. Ipate, Correct systems - building a business process solution,
Springer, Berlin Heidelberg New York, 1998.
5. N.J. Talbot, Trends in Microbiology, 9, 1995.
6. T. Balanescu, M. Holcombe, A. J. Cowling, H. Gheorgescu, M. Gheorghe, C. Vertan,
"Communicating Stream X-machines Systems are no more than X-machines". Journal
of Universal Computer Science, Volume 5, no. 9, 494-507,1999.
7. Z. Duan, M. Holcombe, A. BeII, A logic for biologic al systems, BioSystems, 55, 93-
105,2000.
8. E. Clarke, E. Emerson & A. Sistla, Automatic verification of finite state concurrent
systems using temporal logic specifications, ACM Trans. Prog., Lang., & Systems, 8,
244-263, 1986.
Models of Recombination in Ciliates
P. Sant
Department of Computer Science, King' s College, London WC2R 2LS, UK
sant@dcs.kcl.ac.uk
M.Amos
School of Biological Sciences and Engineering and Computer Science, University
of Exeter, EX4 4JH, UK
Abstract. Ciliate is a term applied to any member of a group of around 10,000 dif-
ferent types of single-celled organisms that are characterised by two unique fea-
tures: the possession of hair-like ci/ia for movement, and the presence of two nu-
clei instead of the usual one. One nuc1eus (the micronucleus) is used for sexual
exchange of DNA, and the other (the macronucleus) is responsible for cell growth
and proliferation. Crucially, the micronuc1eus contains an 'encoded' description of
the working macronuc1eus, which is decoded during development. This encoding
"scrambles" functional gene elements by both the permutation of coding sequences
and the inc1usion of non-coding sequences. A picture of the ciliate Oxytricha nova
is shown in Fig. 1. During development, ciliates reorganise the material in the mi-
cronuc1eus by removing non-coding sequences and placing coding sequences in
the correct order. This 'unscrambling' may be interpreted as a computational proc-
ess during which up to 95% of the original sequence is discarded. The exact
mechanism by which genes are unscrambled is not yet fully understood. We frrst
describe experimental observations that have at least suggested possible mecha-
nisms. We then describe two different models of the process. We conc1ude with a
discussion of the computational and biological implications of this work.
The macronuc1eus consists of millions of short DNA molecules that are 'snipped
out' of or excised from the micronuc1eus. Each macronuc1ear molecule corre-
sponds to an individual gene, varying in size between 400 b.p. (base pairs) and
15,OOOb.p. (the average size is 2,000b.p.) The macronuc1ear DNA forms a very
small proportion of the micronuc1eus, as up to 95% of micronuc1ear DNA forms
intergenic 'spacers', and is eliminated during genetic excision (that is, only 5% of
the micronuc1eus is coding DNA).
270 P. Sant, M. Amos
During macronuclear development, individual genes are excised from the mi-
cronucleus, and are, after the completion of this process, present as individual
short molecules in the macronucleus. The formation of the macronucleus triggers
the process of transcription, whereby genes are read and transcribed into RNA, the
'blueprint' for proteins.
MDSs i 2 3
~I____~~~__~rl~ ______~ (a)
IESi IES2
IESI
2-7 b.p.
Fig. 2. a Schematic representation of interruption of MDSs by IESs. b Repeat sequences
f1anking an IES
We now present a review of two models that attempt to shed light on the process
of macronuclear gene assembly. In [3] Landweber and Karl propose an initial
model that was subsequent1y enhanced in [4], where a formal model for gene rear-
rangement was presented.
Landweber and Karl propose two main operations that model the process of in-
tra- and intermolecular recombination. These can be used to unscramble a micro-
nuclear gene to form a functional copy of the gene in the macronucleus. Both of
these operations are based on the concept of repeat sequences 'guiding' the un-
scrambling process.
pointers are represented by arrows, -+ being the outgoing pointer and +- being the
incoming pointer.
The frrst operation is the simplest, and is referred to as loop, direct-repeat exci-
sion. This operation deals with the situation depicted in Fig. 3, where two MDSs in
the correct (i.e., unscrambled) order are separated by an IES.
The operation proceeds as follows. The strand is folded into a loop with the
pointers aligned, and then staggered cuts are made (Fig. 3b and c). The pointers
connecting the MDSs then join them together, while the IES self-anneals to yield
two molecules, one linear and the other circular (Fig. 3d).
The second operation is known as hairpin, inverted repeat excision, and is
used in the situation containing inverted sequences (sequence y in Fig. 4a). The
molecule folds into a hairpin structure (Fig. 4b), cuts are made (Fig. 4c) and the
inverted sequence is re-inserted (Figure 4d), yielding a single molecule.
The third and final operation in the Rozenberg et al. model is double-loop,
alternating direct repeat excision/reinsertion. This operation is applicable in
situations where molecules have an alternating direct repeat pointer pattern, as
depicted in Fig. 5.
y Fi z (a)
x Fi z
y y y
(b) (e) (d)
Fig. 3. Excision
274 P. Sant, M. Amos
X
~ Y ~ z (a)
X
~
Y Y
z
~
(b) (e)
X
(d)
z
Fig. 4. Inversion
y z g u w
Fig.5. Alternating direct repeat pattern
The molecule folds into a double loop, with one loop being aligned by one
pointer pair, and the other loop being aligned with the second pointer pair (Fig.
6a). Cuts are then made «Fig. 6b), and the sections representing u and y exchange
positions by a process of reinsertion (Fig. 6c).
The three operations presented are intramolecular (as opposed to the intermo-
lecular operations of Landweber and Karl) in that a molecule 'reacts with' (i.e.,
folds on) itself. The process by which gene assembly takes place using these op-
erations and the computational properties of the system are discussed in detail in
[6].
Although the model presented may appear rather abstract, it has been success-
fully applied to the assembly of real genes, inc1uding the Actin 1 gene of Urostyla
grandis and Engelmanniella mobilis, the gene encoding a telomere binding pro-
tein in several stichotrich species and assembly of the gene encoding DNA poly-
merase a in Sterkiella nova. Descriptions of these applications are presented in [7].
Models of Recombination in Ciliates 275
X -X~
\
Y y
u) u
)
~//
w
~ (a)
w
~ (b)
g
=:~\
X
r: y ~/
) u)
w
(c)
Fig. 6. Excisionlinversion
3 Discussion
In this paper we have described two models for the assembly of genes in ciliates.
Although the fundamental molecular mechanisms underlying the operations within
these models are still not well understood, they do suggest possible areas of ex-
perimental enquiry. Looking further ahead, it may well be that in the future these
mechanisms may even be exploited by using ciliates as prototype cellular com-
puters.
276 P. Sant, M. Amos
References
M. J. Fisher
School of Biological Sciences, University of Liverpool, L69 3BX, UK
fishennj@liv.ac.uk
G. Malcolm, R. C. Paton!
Department of Computer Science, University of Liverpool, Chadwick Building,
Peach Street, Liverpool L69 7ZF, UK
Knowledge about the subtle intricacies of molecular processes in cells and their
sub- compartments is increasing rapidly. Cells are highly structured, hierarchically
organised open systems. Contemporary models must take account of spatial het-
erogeneity, multi-functionaIity and individuality. For example, an enzyme can be
described as a 'smart thennodynarnic machine' which satisfies a 'gluing' (functo-
rial) role in the infonnation economy of the cell [1]. We exploit these views by
drawing comparisons between enzymes and verbs (see later and also elsewhere in
this volume). This text/dialogical metaphor also helps refine our view of proteins
as context-sensitive infonnation processing agents [2]. Many proteins, such as en-
zymes and transcription factors display a number of 'cognitive' capacities inc1ud-
ing: pattern recognition, handIing fuzzy data, memory capacity, multifunctionality,
signal amplification, integration and crosstalk, and context-sensitivity (see e.g., [3,
4].
In that they are able to interact with other molecules in subtle and varied ways,
we may say that many proteins display social abilities. This is c1early demon-
strated by Welch (e.g., [5, 6]). The social dimension to enzyme agency also pre-
278 M. J Fisher et al.
supposes proteins have an underlying ecology in that they interact with other
molecules including substrates, products, regulators, cytoskeleton, membranes,
water as well as local electric fields. The metaphor also facilitates a richer under-
standing of the information processing capacities of cells (e.g., [7]). This can be
characterised as an ability to act in a flexible, unscripted manner - another feature
of adaptive agents.
There are several ways of implementing computational agents in computer
software. We have investigated a number of approaches to this all from within an
individual based modelling (IBM) perspective (see [8, 9]). This type of approach
lends itself to fine grain simulation although there are many difficulties with the
variety, reliability and availability of experimental data. Some examples of this
work includes:
• Proteins as classifier systems - in particular, we examined a fuzzy genetics
based leaming classifier approach to 'designing' calmodulin-dependent protein
kinase [3].
• Proteins as logical agents - we demonstrated how 'simpler' enzymes such as
cyclin dependent kinase (Cdk) could be specified for implementation in logic
programs ([3]; see also [10]). Investigations of fuzzy logic models of aspects of
hepatic glucose metabolism were also undertaken [11].
• Proteins in SWARM-based systems - this was concemed with certain aspects
of the E. coli signal transduction system and involved a spatially explicit IBM
approach [12].
• Proteins and algebras - this work comes from programming semantics. Not
only do we argue that it is useful to view the signaling ecology as a vast parallel
distributed processing network of agents operating in heterogeneous microenvi-
ronments, we seek to develop mathematical and semantic methodologies that
might help clarify this analogy between biological and computational systems.
Some examples are discussed in a later section.
A related subject to proteins as algebras concems work on proteins as verbs.
This work grew from a simple analogy, that cells in some ways share systemic
properties that are like 'texts' [7]. Within this framework, enzymes (and particu-
larly signalling enzymes) play an integrative or 'gluing' or functorial role in the
cell [2]. Within this analogical framework it is possible to view multifunctional
processing capabilities of many proteins in terms of process or verb. For example,
CaM Kinase II is a large multimeric enzyme that acts on upwards of 50 substrates
and has four functional domains:
• Catalytic
• Regulatory (it has both inhibitory and CaM binding regions)
• Variable (for targetting and localisation)
• Association (with other subunits)
These functions can be expres sed in terms of verbs (e.g., targetting, catalysing,
regulating etc), or as nouns. Fig. 1 summarises the interactions of the four CaM
Kinase processes in a diagram in which processes are modelled as nodes (vertices)
and interactions as arcs (edges). Using the language of Category Theory it is pos-
Developing Algebraic Models of Protein Signalling Agents 279
The patterns of activity of transcription fac tors provide further insights into
colimit or 'gluing' relations. For example, CBP/p300 are large multifunctional pro-
teins that participate in various basic cellular functions, including DNA repair, cell
growth, differentiation and apoptosis. They are important networking proteins in
that they act as foc al points for multiple protein-protein interactions and co-
activate many other transcription fac tors including CREB, nuclear receptors, sig-
nal transducer and activator of transcription (STAT) proteins, p53, and the bas al
transcription proteins. It is possible to talk about CBP/p300 acting like 'glue' in a
number of ways that relate to molecule-molecule bindings and interactions, en-
zymatic processes, as a physical bridge between various transcription factors, act-
ing as histone acetyltransferases (HATs) - linking transcription to chromatin re-
modelling, and mediating negative and positive crosstalk between different
signalling pathways.
This notion of 'gluing' is a very important concept for appreciating multifunc-
tional activities at a number of levels of intracellular scale. This 'glue' is not just
that the molecules have intrinsic adhesive properties, they also provide the cell
280 M. J Fisher et al.
Exterior
Membrane
Interior
8 ~
+ ACTIVATION
B ..
PHOSPHORYLATION
@ ~ ! ACTIVATION
PHOSPHORYLATI~
®@8
The classical secondary messenger signalling system
Fig. 2. The c1assical secondary messenger signalling system
depends upon its association with structural elements of the ceH via anchor pro-
teins (AKAPs - A kinase anchor proteins; [17]). Specific isoforms of CaMK also
possess 'positional' information; in that 'nuclear-specific localisation' sequences
target this enzyme to the ceH nucleus and, consequently, a role in the phosphoryla-
tion ofproteins involved in the control of gene expression [18].
Perhaps the most sophisticated example of spatial organisation of signalling
pathway components concems the mitogen-activated protein kinase cascades
(MAPK cascades). The spatial organisation of these protein kinase cascades leads
to distinct cellular responses to a diverse group of environmental [19]. MAPK
cascades are organised into discrete parallel signalling complexes by 'scaffold
proteins' - the signalling cascade components are brought into close physical con-
tact allowing rapid and direct transfer of signalling information (see Fig. 3). An in-
triguing feature of these signalling pathways is that, despite sharing common
components, they are normally extremely weH insulated from each other and show
Httle if any 'cross-talk or cross-activation' [20].
Mating Pheromone
!
·". . 1-·. .· Solute change.
~
Cell membrane
+
+ Q +
+
e
+
8 +
+ ?
+ +
'osmoadaptation' gene.
'mating' gene.
'filamentation' gene.
Note that both proteins coexist in both the left- and right-hand sides of this rule;
alI that changes in this transformation is the phosphory!ation of the proteins. As a
concrete example of applying this rule, given the protein soup term
8te11(n) ~ 8te11(n-1)
Again, we assume that n is greater than O. We as sume we have similar phos-
phatase rules for each of the proteins we are interested in.
The various interactions relevant to the yeast mating pathway of Fig. 3 are
given by the three rewrite rules below. The first two represent the cascade of acti-
vation facilitated by the Ste5 'scaffold' protein, where SteI! activates Ste7, which
in turn activates Fus3. First, the activation of Ste7 by Stel!, in the presence of an
Ste5 protein:
We have broken down the effects of the scaffold protein into two stages (the
first two rules above) for the purpose of illustration only. There is some evidence
that the effect of the Ste5 scaffold protein is processive: that is, both phosphoryla-
tion reactions occur at the same time [21].
As an example, once the mating pathway is activated, we may see the following
happen:
Ste5(1) E9 Ste11(1) E9 Ste7(O) EE> Ste11(1) EE> Fus3(O) EE> Ste12(O) EE> Ste12(O)
Ste5(1) EE> Ste11 (1 ) EE> Ste7(1) EE> Ste11 (1 ) EE> Fus3(O) EE> Ste12(O) EE> Ste12(O)
Ste5(1) EE> Ste11 (1) EE> Ste7(1) EE> Ste11 (1) EE> Fus3(1) E9 Ste12(O) EE> Ste12(O)
Ste5(1) EE> Ste11 (1) EE> Ste7(1) EE> Ste11 (1) EE> Fus3(1) EE> Ste12(1) EE> Ste12(O)
Ste5(1) EE> Ste11(1) EE> Ste7(1) EE> Ste11(1) EB Fus3(1) EB Ste12(1) EB Ste12(1)
Ste5(1) EB Ste11 (1) EE> Ste7(1) EB Ste11 (O) EB Fus3(1) EB Ste12(1) EB Ste12(1)
At each step we highlight in bold-face the proteins that match the left-hand side
of a rewrite rule. The end re suIt is that both the SteI2 proteins are phosphorylated,
while in the last rewrite one of the SteII proteins loses a phosphate group through
phosphatase action.
This algebraic model has the virtues that it faithfully represents the signalling
pathway cascade of the yeast mating response, and that it does so computationally:
the rewrite rules described above constitute a program in an algebraic program-
ming or specification language such as OBJ [22] or Maude [23], from which we
borro\V the term 'associative-commutative soup'). This computational element al-
lows us to construct simulations of protein interactions as OBJ or Maude pro-
grams. A disadvantage of our model, however, is its simplicity: it oversimplifies
the protein interactions involved in the signalling pathways in several ways. One
of these ways involves the effects of the scaffold proteins. According to the re-
write rules given above, the interactions between SteII, Ste7 and Fus3 take place
in any space which simply contains an Ste5 protein; there is no notion of the pro-
teins binding in any way to the 'scaffold'.
We could characterise the interactions between proteins and scaffold proteins
by introducing terms such as:
which is intended to represent an Ste5 protein which has bound an Stell and an
Ste7 protein. The rewrite rule capturing the phosphoryIation of Ste7 by Ste 11 then
becomes:
References
Paton, RC. & Matsuno, K. (1998), Some Common Themes for Enzymes and Verbs,
Acta Biotheoretica, 46, 131-140.
2 Paton, Re. (1997), Glue, Verb and Text Metaphors in Biology, Acta Biotheoretica,
45,1-15.
3 Paton, RC., Staniford, G. & Kendall, G. (1996), Specifying Logica! Agents in Cellu-
Iar Hierarchies, in Cuthbertson, R., Holcombe, M. & Paton, R. (eds), Computation in
Cellular and Molecular Biological Systems, World Scientific: Singapore, 105- 119.
4 Fisher, M.J., Paton, Re. & Matsuno, K. (1999), lntracellular Signalling Proteins as
'Smart' Agents in Parallel Distributed Processes, BioSystems, 50, 159-171.
5 Welch, G. R. (1987), "The Living Cell as an Ecosystem: hierarchical analogy and
symmetry", Trends Ecol. Evol., 2, 305-309.
Developing Aigebraic Models of Protein Signalling Agents 287
6 Welch, G. R. & Keleti, T. (1987), "Is cell metabolism controlled by a molecular de-
mocracy or a supramolecular socialism?", TIBS, 12, 216-217.
7 Paton, R.e. (1993), Some Computational Models at the Cellular Level, BioSystems,
29,63-75.
8 Devine, P. & Paton, R.C. (1997a), Biologically-inspired Computational Ecologies: a
Case Study, AISB Evolutionary Computation Workshop, appears in Corne D &
Shapiro, J. (eds) LNCS Springer, Berlin Heidelberg New York.
9 Devine, P. & Paton, R.e. (1997b), Individual Based Modelling in an Explicitly Spa-
tio-temporal Ecosystem, Proceedings of IMACS 97 World Congress, Berlin.
10 Staniford, G. & Paton, R.e. (1994) Simulating Animal Societies using Distributed
Agents, Proceedings of ECAI Workshop on Distributed A.I., also appears in
Wooldridge, M.J. & Jennings, N.R. (eds) Intelligent Agents, Lecture Notes in AI 890,
Springer: Berlin, 145-159.
II Butler, M.H. (1999), Information Processing in Liver Glucose Metabolism, Unpub-
lished PhD Thesis, The University of Liverpool.
12 Clark, L. (2000), A Distributed Information Processing Model of Bacterial Chemo-
taxis, Unpublished PhD Thesis, The University of Liverpool.
13 Ehresmann, A.C. & Vanbremeersch, J-P. (1987), Hierarchical evolutive systems ... ,
Bul. Math. Bio. 49 [1]: 13-50.
14 Takio, K., Smith, S. B., Krebs, E. G., Walsh, K. A. & Titani, K. (1984) Amino acid
sequence of the regulatory subunit of bovine type II adenosine cyclic 3' ,5' -phosphate
dependent protein kinase. Biochem. 23,4200-4206.
15 Schulman, H. & Lou, L. L. (1989) Multifunctional Ca2+/calmodulin-dependent pro-
tein kinase: domain structure and regulation. Trends Biochem. Sci. 14,62-66.
16 Hubbard, M. & Cohen, P. (1993) On target with a mechanism for reversible phos-
phorylation. Trends Biochem. Sci. 18, 172-177.
17 Rubin, C. S. (1994) A kinase anchor proteins and the intracellular targeting of signals
carried by camp. Biochim. Biophys. Acta 1224,467-479.
18 Heist, E. K. & Schulman, H. (1998) The role of Ca'+/calmodulin-dependent protein
kinases within the nucleus. Cell Calcium 23, 103-114.
19 Sprague, G. F. (1998) Control of MAP Kinase signalling specificity or how not to go
HOG wild. Genes & Dev. 12,2817-2820.
20 Fisher, M., Malcolm,G & Paton, R.C. (2000), Spatio-Iogical Processes in Intracellular
Signalling, BioSystems, 55, 83-92.
21 Levchenko, A., J. B. and P. W. Sternberg. Scaffold proteins may biphasically affect
the levels of mitogen-activated protein kinase signaling and reduce its threshold prop-
erties. PNAS voI. 97, no. 11, pp. 5818-5823, May 2000.
22 Goguen,1. A., T. Winkler, J. Meseguer, K. Futatsugi and 1.-P. Jouanaud. Introducing
OBJ3. In Joseph A. Goguen and Grant Malcolm, eds., Software Engineering with
OBJ: Algebraic Specijication in Action. Kluwer, Dordrecht, 2000.
23 Clavel, M., S. Eker, P. Lincoln and J. Meseguer. Principles of Maude. Electronic
Notes in Theoretical Computer Science, voI. 4, pp. 65-89. Eisevier Amsterdam, 1996.
24 Elion, E. E., Pheromone response, mating and ceH biology. Current Opinion in Mi-
crobiology, voI. 3, pp. 573-581. Elsevier Science, Amsterdam, 2000.
25 Courcelle, B., Graph rewriting: an algebraic and logic approach. In Jan van Leeuwen,
ed., Formal Models and Semantics, voI. B of Handbook of Theoretical Computer Sci-
ence, Eisevier Amsterdam, 1994.
26 Goguen, J., Objects. International Journal of General Systems, voI. 1. pp. 237-243,
1975.
288 M. J Fisher et al.
Abstract
The aim is to explain and explore some of the current ideas from category
theory that enable various mathematical descriptions of hierarchical struc-
tures. We review some aspects of the history and motivations behind the
development of category theory and how it has impacted on developments
in theoretical biology and theoretical computer science. This leads on to a
discussion of hierarchical systems and a discussion of some simple examples.
The important idea of colimit is then introduced. Towards the end of the
chapter a number of open questions and problems are discussed.
1 Introduction
This chapter seeks to explore some of the current ideas that provide abstract
models for hierarchical systems in general and ceH systems in particular. The
models are based on category theory and as that theory is to a large extent
relatively unknown to workers in (mathematical and computational) biology,
the chapter will introduce some of the elementary language and concepts
of that theory. We will attempt to do this through fairly 'common or gar-
den' mathematical situations which themselves have aspects of hierarchical
structures about them. Our aim is to present enough of the ideas to make
it possible for the enterprising reader to start delving further into the way
in which others (Rosen [1], Ehresmann and Vanbremersch, see for instance,
[2)) have seen category theory as a potentiaHy usefullanguage and toolkit of
concepts for use in this area.
290 R.Brown, R.Paton, and T.Portcr
1'he first paper in category theory was by Eilenberg and Mac Lane in 1945
[3]. It airned to describe (i) interaction and comparison within a giWIl con-
text (topological spaccs, groups, other algebraic structures, etc.) and (ii) in-
teractions between different contexts, for instance within the an~a of pun'
mathematics known as algebraic topology, problf'rns in the theory of spaces
arc attackcd by assigning various types of algebraic gadgetry to spaces, thllS
translating the topological problern to a, hopdully more tractable, algebraic
one. A category consists of objects anei 'rnorphisms' between thern mor-
phisrns can be thought of as 'structure preserving mappings between objects'.
Even as early as the 1950s anel 1960s, category theory hael bccome a highly
successful language providing geJl(~ral tools for revealing common structure
between different contexts. New concepts (limits, colimits, products, coprod-
uds, etc.) were defined abstractly and these elefinitions highlighteel the prop-
erties that had beeI! there in the examples but had often been hidden by
context specific details. Already in la58 ROSCIl had tried USillg elementary
categorical language to help with the modelling of biological systems. In t.he
1960s, Lawvere's thesis and relatecl work by llenaboll and others at abOlit the
salIle time, sheclnew light on what it Illeant for el structure to be 'algebraie'.
This linked up with thc semantics of formal languagcs and thus with logic.
Lambek anei Lawvere (1968-1970s) showed that there was an int(~rpretation
of the typed ).-caIculus, which was a rich model for some aspects of the logic al
theory of sets, within category theory. Here formulae are interpreted as the
objects, anei proofs as the morphisms. (For ilS OIle important point is that
the morphisms are no longer structure preserving 'mappillgs', they are just
'rnorphisIlls'!) 1'his gave:
(a) Models for set tlwory, but sonwthing much richer is true, the 'sets' are
'variable sets'; anei they can be applied in many more c:ontexts and have their
Categorical Language and Hierarchical Models for Cel! Systems 291
{1,3,5,9,15,45}
Levcl
45 3
15
/~ 9 2
;)
/~/ 3 1
~/
1 O
This is an example of a partially ordered set (poset). The above diagrarn is
called the Hassc diagmm of the poset. ~ote that only the essential 'gener-
ating' arrows are shown here. A poset is not simply a bag containing parts
there is also an ordering among the parts. If we include the 'composite'
arrows, we get a more complicated diagram.
Categorical Language and Hierarchical Modcls for Cel! Systems 293
Level
9 2
1 o
togethcr with a loop at each vertex i sin ce i exactly divides itself! Each level
of the 'hierarchy' measures in some sense the 'complexity' of the objects at
that level.
In going from simple diagrarn of the partially ordered set (poset) to that
including the composite arrows we are starting the transition from posets to
categories. As we do so, further structure is added to the system. What is a
category in this sense?
A category COIlsists of:
- A coUection of objects C = {i, j, k, ... , etc.}.
Collections of ar·row8 or links or morphisms (alternative terms depending
on taste, context etc.).
Each arrow has a 80urce ami target and C(i,j) wiU denote the set of arrows
from source -i to target j, if f E C(i,j) we might also write f : i ---+ j Of
. f .
~---+J.
. ---+
f ·J, 9, k
J. --,
l
1. Do Objects - divisors of 45
2. D 1 Objects - divisors of 45
Dl (i, j) = set of paths from i to j in the diagram of Div( 45).
Thus, for instance,
1 -t 5 -t 15 -t 45 }
Dl(1,45) = { 1-t 3 -t 15 -t 45
1 -t 3 -t 9 -t 45
Note that aU paths (following the arrows) are used and are considered to be
distinct. This category Dl is caUed the fr'ee category an the Hasse diagrarn
of Div(45). That diagram is a directed graph and given any directed graph,
r, define
C:= FCat(r)
by
Objects (C) = V(r), the set of nodes or vertices of r;
C(u, v) = the set of directed paths in r from vertex u ta vertex v.
Composition : concatenation of paths.
Identity : 'empty path' at a vertex.
As the notation suggests, this category is called the free categor'y an r.
3. Our next example D 2 will be more complicated than a category.
In a category, C, each C(i,j) is a set; in an enriched category, it will have
more structure - in our example, each D 2 (i, j) will itself be a category
(in fact a poset):
Categorical Language and Hierarchic:al Models for Cel! Systems 295
As before, the objects of D'2 will be the divisors of 45 and D 2 (i, j) will be
empty if i does not divide ji however if i does divide j then D 2 (i,j) will
be the poset Div(j li) of divisors of the quotient of j by i. For example
D 2 (3,45) = Div(15) and so its Hasse diagram looks like
given by the product . of numbers - this mapping preserves the order, and
so is called an 'enrichcd composition'. Thus D 2 is an example of an 'order
enriched' category. (Uses of order cnriched categories relevant to the sub-
ject matter of this chapter include the book by Arbib and Manes already
mentioned [5] and Goguen's book [8] in which a special form there calleei
a ~-catcgory is used to provide structure for sign systems within algcbraic
semiotics. )
4. Some important further examples are called 'Big' categories because there
are so many objects (technically, the objeets do HOt form a set).
The generic form of thcse is:
All rnathematical objects with some specified structurc
All rnorphisms, i.e. mappings that preserve that structure
Here are some cxarnples:
Name Objeets Morphismsl arrows
COP(i,j) == C(j,i)
with composition suitably adjusted. For posets, this corresponds to putting
the reverse order on the poset and so to turning the Hasse diagram upside
down.
Because of this any construction applied to categories comes in two
flavours, since the construction can also be applied t.o the opposite cat-
egory; these two are usually distinguished by giving one of them the prefix
'co'. Thus we will have below limits and colimits, and for different appli-
cations we favour one rather than the otheL We will be chiefly concerned
with colimits which are useful for describing how a structure is made from
components by putting them together: the device for describing 'put.ting
t.hem together' is called a 'cocone'.
COlilllits
As an illustrative example of this we give the least common multiple
lcm(a, b) of two numbers a, b. Thus lcm(a, b) is the least of the common
multiples of a, b and so lcm(9, 15) == 45, lcm(3, 5) == 15, and sa ono
We do need a bit more precision, of course. In the natural numbers, 1,2,
... , we say c is a common multiple of a and b if c is a multiple of a (sa
there is some d with c == da), and c is a multiple of b (so there is some e
with c == eb). We note that in the diagram for D'iv(n) for any such c, a, b, we
would have
a b
This diagram is a cocone (by convention, not a cone) on the pair a, b. It
says c is a common multiple of a and b.
Now bring 'least' into play:
r
lcm(a, b)
a
/~ b
Categorical Language and Hierarchical Models for CeH Systems 297
This type of situation can be abstracted and generalised to give the notion
of 'colimit'. The 'input data' for a colimit is a diagram D, that is a collection
of some objects in a category C and some morphisms between them. So a
diagram D could look like this:
D= ;> •
/ / .
/
~
Now \Ve need the notion of a cocone with base D and vertex an object C.
This wiU look like:
We also have to make the condition that each of the the triangular faces of
this cocone is commutative, where a triangle of morphisms
/'
/'
/'
/'
.-\.."",.... . . ...:
/'
/'
/'
~ \>,:.. :~::
1 .....
Again, all triangular faces of the combined picture are commutative. Now
stripping away the 'old' cocone gives the factorisation of the cocone via the
colimit:
/'
/'
/'
/'
/'
/'
/'
/'
/'
!i.':"''''';::>:.::: ...
~ .... . ...
.~ .. .
Intuitions:
The object colim(D) is 'put together' from the constituent diagram D by
means of the colimit cocone. From beyond (or above in our diagrams) D, an
object C 'sees' the diagram D 'mediated' through its colimit, i.e. if C tries to
interact with the whole of D, it has to do so via colim(D). The colimit cocone
is a kind of program: given any cocone on D with vertex C, the output will
be a morphism colim(D)-+ C.
Example
The lcm is the colimit of the diagram
Categorical Language and Hierarchical Models for Cell Systems 299
a~ /1
gcd(a, b)
The gcd, from a lower level of the hierarchy, 'measures' the interaction of a
and b.
Some people have viewed models of biological organs as colimits of the
diagrams of interacting cel!s within them.
WARNING. Often colimits do not exist in a category C for some diagrams.
However, one can add colimits in a complet ion process, i.e. freely for a class
of diagrams, and then compare these 'virtual colimits' with any that happen
to exist. An example of this process would seem to be the introduction in
neural net theory in the 1980s of the notion of 'virtual neurons' or 'assem-
blages' where an interacting subnet of 'neurons' exhibited behaviour as if it
was a (more powerful) single neuron. Perhaps the superstates considered by
Bel! and Holcombe [9] are similarly 'formal' colimits. Instances of biologic al
situations that lead to diagrams of this sort and hence to colimits occur in
the work of Ehresmann and Vanbremersch, mentioned below in more detail,
in Dioguardi [10] where they are used to model 'the hepatone', which is a
model of the interaction of the major cel! types in the liver, and, generical!y,
in the discussion of 'glue' in the study of integrative biology by the second
author ([11, 12]). Dioguardi's formulation of the 'hepatone' and subsequent
extension to incorporate also the 'hepatonexon' is an important illustration of
the use of mathematical thinking to clarify the idea of biological functioning
units. It is especial!y interesting as a number of models have been proposed
for units of hepatic function.
It is important to note that a colimit with its defining cocone has more
structure than merely the sum of its individual parts, sin ce it depends on
the arrows of the diagram D as well as the objects. Thus the specification
for a colimit object of the cocone which defines it can be thought of as a
'subdivision' of the colimit object. It would be interesting if the folding of a
one-dimensional protein sequence to a three-dimensional functioning struc-
ture could be seen as a colimit operation.
Much of this has been leading up to an introduction to the notion of a
(categorical model for a) hierarchical system as formulated by Ehresmann
and Vanbremersch [2].
They consider a basic category 1HI with specified objects and arrows. A
partition of Obj (IHI), the set of objects of 1HI into p + 1 classes (levels) labelled
300 R.Brown, R.Paton, and T.Porter
0,1, ... ,p such that each object AI at level n + 1 (with n < p) is thc coli mit
in JH[ of a diagram A of linked objects at level n.
We refer the re ader to the papers of Ehresmann and Vanbremersch aud
their web page
http://perso.wanadoo.fr/vbm-ehr/
and also to related pages such as that of Amiguet
http://iiun.unine.ch/people/mamiguet/index.html
Queries
(i) Why here is A only made up of objects anei links at level n? In 'ce11
systems', does one need shared objects of lowcr levels within the diagram?
(ii) How can one handle mathematica11y, and then computationally, thc
properties of AI, given knowledge of A?
Parting Thoughts
(a) To model manufacturing control systems, models such as Petri ncts,
timcd event graphs, etc. exist in numerous ftavours, stochastic, !1J,zzy, etc.
These seem 'enriched versions'. Is there a way of handling hicrarchical systcms
in which these 'enrichments' play a significant role. Some small progress has
been made in this direction - but so far it is inconclusive. For a model of
computation using enriched categories see [13].
(b) E- V hiemrchical systems try to model ce11 systems and do consider
weighted arrows. Would a variant of their theory. but using (poset or lattice)
enriched categories enable an amalgam of thcir rich conceptual basis with the
rich computational machinery already dcvelopcd from (a) above?
(c) Are there 'formallanguage' aspects of the hierarchical systems, capable
of providing models for cell systems?
Therc is a good practical interpretat ion of linear logic for manufacturing
systems [14]. There are many levels of manufacturing process in cellular sys-
tems - from the synthetic processes producing intracelllllar post-translational
prodllcts tn material targeted for export (notably protein products). One also
needs to model the vertical as well as the horizontal informat ion proccssing
involved in such multiple levels of ce11ular interaction.
(d) What might be a feasible successful biological model in the above
context? Reca11 that fractals have becn considered successful becausc they
showed that complex variation could result from a very simple model. How-
ever, many fractals are very simple, since they are defined by iterated function
systems, based on iterates of a single map. Examplcs need to be developed
of the next level of complexity, where also some actual computation, rather
than experimentation, is feasible because of the algebraic c:onditions imposed
by the struc:ture of the system. Thus a researc:h programme would be to com-
bine the algebra of rewriting [15], whic:h considers the c:onsequences of rules,
Categorical Language and Hierarchieal Models for CeH Systems 301
with some continuous variat ion as in fractals, to see how a range of 'colimit
structures' can develop. A generalisation of rewriting to categories and to
actions of catcgories is given in [16].
(e) We should also note the work of Dampney and Johnson on informat ion
systems [17], which showed that simple commutative diagram confliderations
could have useful consequences in simplifying a complex system (and so sav-
ing money). Since efficiency is of importance for biological systems, we would
hope to find examples of analogous considerations.
f) Another potential area for development is that of 'higher dimensional
algebra', see the Introduction given in [18]. This shows that one of the con-
tributions of category theory is not only to give a useful general language
for describing structures but also, in a self-reference mode, that in order to
describe the array of structures which have arisen in mathematics new math-
ematical structures have been found needed, and that these structures have
proved of independent interest. Not only that, a crucial step in developing
category theory is to have an algebraic operation, composition of arrows,
which is defined under a geometric condition, that the source of one arrow
is the target of the other. So one is led to envisage more general kinds of
compositions. An overall slogan in one aspect of the applications of these
more general structures was:
Algebraic inverses to subdivision.
That is, we may know how to cut things up, sub divide them, but do we have
an adequate algebra which encodes the structure and rules which govern the
behaviour of the re suIt of putting them together again? It was found, as is
described with references to the literature in [18], that there are forms of
what are called multiple categories which do have convenient properties in
some situations in this regard. These ideas have led to new mathematics
which has enabled new descriptions and new computations not available by
other means. The enriched categories to which we referred earlier can also be
regarded as forms of multiple categories.
The situation is even more elegant in that we generally think that com-
position is described mathematically by forms of algebra. There is a growing
body of mathematics called 'co-algebra' (see for example [19]) which seems
to give a possible language for subdivision. The combination of these two
strands of composition and subdivision could well be important for broader
applications in the future.
Another theme related to 'algebraic inverses to subdivision' is 'non-
commutative methods for local-to-global problems'. See [18] for an example
of how a two-dimensional structure proposed in 1932 for geometric purposes,
and in which the operations were always defined, was found to reduce to
302 R.Brown, R.Paton, and T.Porter
4 Conclusion
We hope that pointing out the existence of this categorical rnathernatics wiU
help the formulat ion of applications and also suggest ways to new forrns
of mathernatics required for the biological applications. Category theoretic
applications to biological systems such as those of Rosen, Ehresrnann and
Vanbremersch, and the chapters in this volume by .vlalcolm and Fisher, by
Paton, by Wolkenhauer, help to strengthen the importance of relational as
well as hierarchical thinking in biology.
References
O. Wolkenhauer
Systems Biology & Bioinformatics Group, Department of Computer Science,
University of Rostock, www.sbi.uni-rostock.de
o. wolkenhauer@umist.ac.uk
W. Kolch
Institute of Biomedieal and Life Science, University of Glasgow,
Cancer Research UK, Beatson Laboratories, Garscube Estate,
Switchback Road, Glasgow G611BD, UK
K.-H. Cho
School of Electrical Engineering, University of Ulsan, Ulsan, 680-749,
South Korea
Gene expres sion is the process by which information stored in the DNA is trans-
formed via RNA into proteins. While the availability of genome sequences is
without doubt a revolutionary development in the life sciences, providing a basis
for technologies such as microarrays, the principal aim of the post-genome era is
to understand the organisation (structure) and dynamies (behaviour) of genetic
pathways. The area of genomics reflects this shift of focus from molecular charac-
terisation of components to an understanding of the functional activity of genes,
proteins and metabolites. This shift of focus in genomics requires a change in the
way we formally investigate cellular processes: Here we suggest a dynamic sys-
tems approach to gene expres sion and regulation, an approach we refer to as sys-
tems biology or genomie eybernetics.
Later we are going to provide an example for intracellular dynamics by means
of a mathematical model for a signalling pathway. However, looking at cells inter-
acting in the morphological development of an organism provides another exam-
ple for the importance of a dynamic-systems perspective of gene expression and
regulation. For differentiation of cells in development we find that the relation be-
tween the genome of a cell and the reactions which occur in the cells we require a
conceptual framework for both spatial and temporal aspects in order to capture the
relationship between an intern al programme and dynamic interactions between the
ceH and its environment. The environment may be other cells, physical constraints
or external signals to which the cellular system can respond. While we suppose
that the cells in a developing organism can possess the same genome, they never-
theless can develop and respond completely differently from one another. To an-
swer why and how this can happen one ought to study gene expression as a tem-
poral process. The principle chaHenge for systems biology is then to answer the
following questions [adopted from 1]:
1. How do cells act and interact within the context of the organism to generate
coherent wholes?
2. How do genes act and interact within the context of the cel! as to bring about
structure andfunction?
Asking how genetic pathways are dynamieally regulated and spatially organ-
ised, we distinguish between the aetion and interaetion of genes and cells respec-
tively (intra- and intercellular dynamics). For example, considering morphological
development, to what extent do genes control the process or do genes only partici-
pate in a reactive fashion? Many decisions in development are induction events
mediated by the contact with the surroundings. The multicellular context therefore
determines what happens to the individual cell. For example, cancer cells have lost
this ability to respond and therefore disregard tissue organisation and grow unre-
strictedly and invasively. It seems that cells and eventually organs have an inher-
ent developmental programme which they execute unless instructed otherwise.
Since the 1960s it is known that the most basic cellular processes are dynamic,
feedback controlled and that cells display anticipatory behaviour. In the 1960s,
investigating regulatory proteins and the interactions of allosteric enzymes, Fran-
cois Jacob and Jaques Monod introduced the distinction between 'structural genes'
Mathematical Systems Biology: 307
(coding for proteins) and 'regulatory genes', which control the rate at which struc-
tural genes are transcribed. This control of the rate of protein synthesis gave the
first indication of such processes being most appropriately viewed as dynamic sys-
tems. With the lack of experimental time-course data, mathematical models of
gene regulatory networks have so far focused on ordinary or stochastic differential
equations and automata [2, 3]. For such models to be specific they only consider a
small number of genes and for simulations of many genes interacting, the relation
to experimental data is lost. The problem, also known as Zadeh's uncertainty prin-
ciple is further discussed below. It is clearly important to explore the principal
limits of how we can balance the composition of components on a large scale, pre-
serving the integrity of the whole system, with the individuality of its components,
and without losing too much accuracy on the small scale. Since the two organisa-
tional levels (gene versus genome or ceH versus tissue/colony) are very different
with regard to how we can observe and represent them, different areas of research
have evolved around these organisational and descriptional levels. For example,
while differential equations have been used to develop accurate or predictive
models of individual genes in a particular organism and context [2], Boolean net-
works modeling hundreds and thousands of interacting genes have been successful
in capturing evolutionary aspects at the genome level [4].
The challenge is to develop a conceptual framework, which integrates these
models through abstraction (i.e., generalisation). For even the simplest of biologi-
cal systems we find that a whole range of techniques, ranging from time series
analysis (regression models), dynamic systems theory (rate equations, behavioural
models), automata theory (finite state machines) and various others are likely to be
considered. The validation and evaluation of any mathematical model with ex-
perimental data will further require pattern recognition techniques such as multi-
variate clustering and component analysis. There is therefore agreat need for inte-
gration of mathematical models and to formalise the modeling process itself.
Possible approaches which may be able to integrate or unify these distinct meth-
odologies are briefly discussed in the following section.
regulated through further maps from the previously introduced co-domain and the
set of mappings between the two spaces. While Rosen captured this transforma-
tion-regulation process using category theory, it is possible to derive conventional
models such as automata, state-space representations and regression models from
them [11]. In [12] we discussed how automata and state-space models can be con-
sidered as special cases (or 'realisations') of (T,R)-systems. The shift of focus
from molecular characterisation to understanding the dynamics of pathways in ge-
nomics is reflected in the change of the definition of the objects in the domain and
co-domain to become sequences of data obtained from time course experiments.
Further below we return to the discussion about how the change of thinking in ge-
nomics should be reflected in mathematical modelling of biological systems.
Constraints on the nature of mappings and therefore the c1ass or categories of
functions and its structure arise 'naturally' from biological considerations. For in-
stance, gene products usually have more than one biological function which fre-
quently depends on the state of the cell (metabolic, other signaling, etc.). To give
one extreme example, beta-catenin is a structural protein of cell-cell adhesions at
the cell membrane, where it helps in gluing cells together. However, it also can
work as a transcription factor in the nuc1eus as the endpoint of the so-called wnt
pathway, which is an extremely important developmental pathway. Any devia-
tions from expected behaviour have catastrophic consequences in the development
of the organism. Thus, a mapping or the class of mappings must be able to ac-
commodate dynamic changes. Sometimes two different genes may lead to the
same biologic al function. Gene knock-out studies show that the function of a de-
leted gene can sometimes be replaced by another gene or genes. For instance,
there are several Ras genes, three of which have been knocked out in mice: Har-
vey-Ras, Kirsten-Ras and N-Ras. H-Ras and N-Ras knock-out are almost normal,
but the K-Ras knock-out is lethal. The work of Casti [11, 13], which extends
Rosen's work on (M,R)-systems and considers regulation in dynamic metabolic
systems, could provide an interesting starting point to investigate this problem.
Conventional systems theory considers inputs (independent variables) trans-
formed into outputs (dependent variables). The inputloutput point of view, al-
though suitable for the engineering and physical sciences, is unsuitable for cellular
systems or gene networks as these systems do not have an obvious signal flow of
direction. In contrast, in the 'behavioural approach' [14] systems are viewed as de-
fined by any relation among dynamic variables and a mathematical model is de-
fined as a subset of a universum of possibilities. Before we accept a mathematical
model as an encoding of the natural system, all outcomes in the uni verse are pos-
sible. The modelling process then defines a sub set of time-trajectories, taking on
values in a suitable signal space, and thereby defines a dynamic system by its be-
haviour rather than its inputs and outputs. While the definition of causal entail-
ment via designated 'inputs' and 'outputs' remains the primary objective for the
biological scientist, its definition follows that of a dynamic system in terms of
time trajectories. Willems' behavioural framework fits therefore very well the
situation in which we obtain experimental data. For example, microarrays provide
us with large sets of short time series for which dependencies have to be identified
from the data rather than being defined a priori.
310 O. Wolkenhauer et al.
function', one could altematively consider cells as related but essentially inde-
pendent components with an intemally defined programme for development, in-
cluding mechanisms in response to environmental changes or inputs. The com-
parison and combination of both modelling paradigms could lead to a number of
interesting questions related to how the scientist interprets causal entailment in
biological systems.
In general, causation is a principle of explanation of change in the realm of
matter. In dynamic systems theory causation is defined as a (mathematical) rela-
tionship, not between material objects, but between changes of states within and
between components. In biology causation cannot be formally proven and a 'his-
torical approach' is the basis for reasoning, i.e., if correlations are observed con-
sistently and repeatedly over an extended period of time, under different condi-
tions and by different researchers, the relationship under consideration is
considered 'causal'. This approach is surprisingly robust, although exceptions
have been found to almost any single dogma in biology. For instance, some vi-
ruses contain RNA genomes which they copy into DNA for replication and then
have the host ceH transcribe it back into RNA.
certain variables according to specific input pattems to the system) but further
suggests that the systems biologist should be part of the experimental design proc-
ess rather than being 'delivered' a data set for analysis.
linearisation reduction
physico-chemical
simulation
principles
measurement parameter
and observation estimation
pre-processin realisation
Fig. 1. Mathematical modelling of biological systems can follow two routes - 'modelling',
guided by experimental data, and 'identification' from experimental data. In both cases, we
rely on numerous assumptions and simplifications [12]
Dnce the experimental design is completed and data are being generated, the
question of which kind of mathematical model and which structure it should have
arises. In the theory of dynarnic systems we generally have to make a decision
whether to regard the process as a deterministic non-linear system but with a neg-
ligible stochastic component or to as sume that the non-linearity to be only a small
perturbation of an essentially linear stochastic process. Genuine non-linear sto-
chastic processes have not yet been shown to be applicable for practic al time-
series analysis. Although natural phenomena are never truly linear, for a very large
number ofthem linear (stochastic) modelling is often the only feasible option. The
dilemma with, for example, microarray time course experiments is that hundreds
of variable are sampled at only a few sample points with replicates considered a
luxury. This naturally gives rise to questions regarding the limitations of stochastic
linear modelling in the context of such data.
An interesting question in the context of the semantics of mathematical models
is the role of 'noise' or random fluctuations in general. In biology, the role of ran-
dom variation is often illustrated with examples related to evolution and intrace1-
lular fluctuations of regulatory molecules. For the latter the question is usually an-
swered by the number of molecules involved, fewer molecules usually suggesting
a stochastic model while large numbers of molecules often permit a deterministic
model. While in the former case variation is an intrinsic aspect of the natural sys-
tem under consideration, a noise term in a description or formal representation is
often used to 'cover up' variations that cannot be explained with the given model
and hence relates to a limitation in the observation and explanation of the phe-
nomena. The question then is to whether a mathematical model is considered to
Mathematical Systems Biology 315
which precision and significance (or relevance) become almost exclusive charac-
teristics. Overly arnbitious attempts to build predictive models of cells or subcel-
luar processes are likely to experience the fate of historians and weather forecast-
ers - prediction is difficult, especially if it concems the future ... , and these
difficulties are independent of the time, arnount of data available or technological
resources (e.g. computing power) thrown at the problem.
The problem is that perturbations to cells have multi-gene / multi-transcript /
multi-protein responses, 'closing' the system, i.e., restricting the model to a small
set of variables, assuming constancy of some variables, inevitably leads to an of-
ten unacceptable level of uncertainty in the inference. In other words, the prob-
lems of applying systems theory in biology can be summarised by
(a) The difficulty of building precise and yet general models
(b) The 'openness' ofbiological systems, the fact that these systems are hierar-
chical and highly interconnected
NATURAL SYSTEM
Represents
Measurement ----+. Mathematical Modelling Simple Systems
(axioms, equalions, diagrams) Quantilatively
1 Precise Inference
----+.
j
Inferential Entailment but
Inaccurate Conclusions
Describes
Observation ---+. Empirical Analysis ----+. Complex Systems
(naturallanguage, diagrams, pictures) Qualitatively
1
Causal Enlailment
Accurale Conclusions
bul
Imprecise Reasoning
Mitogens
GrOHth factors
T=
dx (t)
-kl 'X I(t)x 2(t)+k2 ·x (t) 3
dx7 (t)
~=-kl'XI(t)X2(t)+k2 ·x/t)+k3 ·x3 (t)
dx/t) _
-----;Jţ-kl ·X I (t)X2 (t)-k2 ·x/t)-k3 'X/t)
dx4 (t) _
-----;Jt-k3 ·x3 (t)-k4 'X4 (t)
,
rameter values have to be identified.
complex (ES)
Fig. 4. The pathway model is constructed from basic reaction modules like this enzyme ki-
netic reaction for which a set of four differential equations is required
320 O. Wolkenhauer et al.
Raf-l* RKIP
9~
~
ii;
0.5 .. 0.51----'...--__7Y
1>
~
i 0.5t:===~:=:=:~0K'~~t=:3 a
15
~ 0.4
.. 0.5
15
~ I I
I
I
_ 1 ___ J. ___ L ___ 1__ _
.: 0 .4 -
~O.3 I I I
~
0.3 '-----'-----'-----'-----'-----'
o 6 10
t time t: time
x 10-3parameler esUmation for k2~ k2~O.OO72 x 10~ parametef eSlimatlon for k4: k4=O.OO24S
8~~~~~---------, 3 r~
""'--C~
.k-U~
.1-
00""'--------'
~
1251-=:..!:~~...J-':~~.r""'-=9
'5 2
~
I I I
___ .i ___ L.. ___ 1__ _ ~ _ __ J. ___ L ___ 1__ _
=- 1.5
~
1'--_-'--_-'-_--'-_--'-_--'
o 10
I: llme t Ume
Fig. 6. IIIustration for parameter estimation from time-series data: the upper left shows
Raf-I */RKIP complex association parameter kl, the upper right shows Raf-
I */RKIPIERK-PP association parameter k3, the lower left shows Raf- I * and RKIP disso-
ciation parameter k2, and the lower right shows ERK-PP and Raf-I */RKIP complex disso-
ciation parameter k4
Suppress,on oI Ra~ 1 ',nase aCIMty by RKlP' Ra" Supp'ession of Raţ 1 kmase aclMly by RKlP ERK-P
-003
~002
Q.
~001
w
° .. '
50
o
°
reaclion tlme [seci concenlral,on 1!1M1 reacllon hme [sec) ooncentrallon 1!1M1
Supp,.ssion of R.~l kin ... aClml y by RKlP, RKlP Supp,ession 0/ Ra/· I kina.e aelMly by RKlP: RKlP·P
0.02
002
~
:tom
Si!
a::
o
50
o
°
reaclion I,me Isecl concenlralion 1!1M1 reOlr.tlOn lim81sac1 concen'r.,ion 1!1M1
Fig. 7. The simulation results according to the variation of the concentration of RKlP: The
upper left shows the change of concentration of Raf-l *, the upper right shows ERK, the
lower left shows RKlP, and the lower right shows RKlP-P
Supp,ession ofR,!- 1 kinase aC1M11 by RKlP. MEK·PP Suppt.ssion of Raf.1 kin... aelNily by RKlP' RKlP-P·RP
~2.5 ~Olli
Q. 0.04
8: 2 a;
::l:
w
~ 0.02
::< 15
50
Si!
a; °
50
reaction lim e [sec) o concenlral,on I!1MI reaction time lsecl o concenlral,on 1!1M1
Suppress,on of R,f.1 kinase oclM1y by RKlP. ERK·P-MEK·PP Suppr••sion of R,ţ 1 k,nas. IclMly by RKlP: RP
3
2.98
~2.96
Q:"2.94
a:: 2.92
50
reactio" limelsecJ
° concenlralion I!1MI ° reaction hma Isec) o concenl,al;on I!1MI
Fig. 8. The simulation results according to the variation of the concentration of ERK-PP
(continued): the upper Iert shows the change of concentration of MEK-PP, the upper right
shows RKlP-P-RP, the lower Iert shows ERK-P-MEK-PP, and the lower right shows RP
The kind of models and the modelling approach which we introduced here have
already proven to be successful (i.e. useful to the biologist) despite the many as-
sumptions, simplifications and subsequent limitations of the model. A challenge
for systems biology remains: how can we scale these models up to describe not
Mathematical Systems Biology 323
only more complex pathways but also to integrate information and capture dy-
namic regulation at the transcriptome, proteomic and metabolomic level.
Especially MAP kinase pathways have been investigated by various groups us-
ing a variety of mathematical techniques [17, 18] and the co-existence or generali-
sation of different methodologies raises questions about the biologic al systems
considered, for mathematical models to have explanatory power, hence being use-
fuI to biologists, the semantics or interpretation of the models matters. Do we as-
sume that a cell is essentially a computer or machine - executing logical pro-
grammes, is it a biochemical soup, an essentially random process or are
independent agents interacting according to a set of pre-defined rules? Is noise an
inherent part of the biological process or do we introduce it as a means to repre-
sent unknown quantities and variations?
Real-world problems and challenges to apply and develop research in the area
of systems biology are abounding. For example consider the development of
mathematical models used to analyze and simulate problems in development such
as what is sometimes called asymmetrical division. This describes the phenome-
non that when a stern cell divides, one daughter cell differentiates whereas the
other remains a stern cell. Otherwise our stern cells would get depleted. This phe-
nomenon happens although the dividing stern cell has the same pattern of gene
expres sion and is exposed to the exact same environmental cues. Another applica-
tion would be the mathematical modeling of differential gene expression and regu-
lation of transcription during a bacterial or viral infection. The theoretical work
could be guided by the analysis of DNA microarray data, which are available for a
number of organisms.
The discussion above outlined a dynamic systems framework for the study of gene
expres sion and regulation. We are interested in the interface between internal cel-
lular dynamics and the external environment in a multi-cellular system. A defini-
tion of complexity in the context of modelling gene expression and regulation is
given and the background and perspective taken are described in detail. While the
motivation is to investigate some fundamental questions of morphological devel-
opment, differentiation and responses to environmental stress, the proposal is to
focus these questions on a limited set of problems, methodologies and experimen-
tal techniques. The use of a model is to see the general in the particular, i.e., the
purpose of a mathematical model of cellular processes is not to obtain a perfect fit
to experimental data, but to refine the biologic al question and experiment under
consideration.
The central dogma of systems biology is the fact that the cell and its inter- and
intracellular processes describe dynamic systems. An understanding of regulatory
systems therefore requires more than merely collecting large amounts of data by
gene expression assays. If we are to go beyond association to an understanding of
causal entailment, we need to go beyond the data mining approach. The systems
approach is characterised by systematic manipulation of the system behaviour.
324 O. Wolkenhauer et al.
References
R. Sole and B. Goodwin (2000): Signs of Life: How Complexity Pervades Bi-
ology? Basic Books, New York.
2 J.J. Tyson and M.C. Mackey (2001): Molecular, Metabolic and Genetic Con-
trol. Chaos, VoI. 11, No 1, March 2001 (Special Issue).
3 J. Hasty, D. McMillen, F. Isaacs and J.J. Collins (2001): Computational Stud-
ies of Gene Regulatory Networks: In Numero Molecular Biology. Nature Re-
views Genetics, VoI. 2, No 4, 268-279, April2001.
4 S.A. Kauffman (1995): At Home in the Universe: The Search for Laws of
Self-Organisation and Complexity. Oxford University Press, New York,
1995.
Mathematical Systems Biology 325
C. G. Johnson
Abstract. This chapter is concerned with how computational ideas can be used as
the basis for understanding biological systems, not by simulating such systems, but
by taking a computational stance towards the way such systems work. A number of
issues are addressed. Firstly the question of what kinds of computer science are
needed to help understand computational processes which happen outside of con-
ventional computing machines. The second issue addressed places computational
constraints on how the world can act into Dennett's framework of grades of possi-
bility. The final main section considers the issue of changes in the world, and when
it is meaningful to regard such changes as carrying out computations.
1 Introduction
'Who could believe an ant in theory? A giraffe in blueprint? Ten thousand doctors of what' s
possible could reason half the jungle out of being.' John Ciardi [1]
It is a tired cliche of popular psychology that we only use 10% of our brains. It
is unlikely that this is true, but it is interesting to consider how a statement like this
might be interpreted. Is this a physical statement? Could we cut out the 90% that
isn't being used, throw it away, and stiH function normally? Is this a biological
question, meaning that we are only using 10% of the available neuronal pathways
or only activating 10% of the signals that we could? This is perhaps closer, but stiH
not ideal. Does it mean that we could store 10 times as much 'stuff' if we were
working at fuU capacity? Think 10 times as quickly? These are stiH fairly ill-
defined questions, but they conform to the intuitions which people have about the
brain, aud they are at heart computational questions. They are questions about the
capacity of an entity for the storage of information and its ability to process that in-
formation.
The point of this story is to illustrate that we already think about biologic al
processes in terms of computational ideas, even if in an informal way. It is not sur-
prising to find that we think about the brain in this way, given both the popular
view of brains as being essentially computers and the view dating back to the early
days of computing of computers as 'electronic brains'. However, it is only a short
step from this to start thinking about other parts of the body (whether at the tissue
level or the cellular level) in computational terms.
Many cellular systems have an information processing aspect. The immune sys-
tem is weU studied from this perspective [2, 3], and there is potential to view the
signal transduction system in this way. What kinds of computer science are needed
to help understand these kinds of system?
One example of a piece of computer science theory that could provide a tool for
the understanding of natural computation is a theory of complexity. If we consider
computation to be something that is grounded in the world, then how does that in-
f1uence our view of computational complexity? What kinds of complexity in nature
arise out of the presence of computations in natural systems? Clearly we can al-
ways define certain formal system& as being what we mean by the term computa-
tion, and then derive/define certain measures of complexity with respect to this
definition. However, if we want to apply computational ideas in the context of ana-
lysing transformations of the world then we might want to not ground these in par-
ticular axiomatisations of computation, as we might not be able to show that the
physical system conforms to that axiomatisation.
What Kinds of Natural Processes Can Be Regarded as Computations? 329
We can generalize this idea to ali natural systems as follows. Given any system
in the world and some idea of what we mean by a computer either we can simulate
it on the computer or not. Thcre are two variants of this. In the first we consider
what can be simulated at aII. For example we cannot accurately simulate most non-
330 C. G. Johnson
discrete systems using a computer with a finite memory (we can clearly simulate
some such systems, such as relationships between two intervals on the real line
which can be described by a finitely describable function). This is regardless of the
amount of time we take; even given any finite number of timesteps we cannot even
represent the initial state of the system exact1y. Other systems admit an inefficient
simulation. For example a problem like factoring composite integers is hard (in the
technical sense) on conventional computers, yet proposed quantum computers [9]
provide a technology on which polynomial-time algorithms for the same problem
can be executed.
The consequence of this is that if we cannot simulate the system (efficient1y, or
at all) on the computer then theoretically there is a property of the world which can
be used as a substrate for computation. Clearly whether the particular property ad-
mits its use for computing artificially created problems will vary from case to case.
In particular a significant feature is whether the system admits control over its in-
puts; many computations are happening in the natural world which cannot take in-
puts other than the ones which they receive as part of a larger system. Therefore we
cannot say that merely observing the act of computation in a natural system pro-
vides the core of a practical computation system.
3 Grades of Possibility
There seem to be at least four different kinds or grades of possibility: logical, physical,
biological, and historical, nested in that order. Daniel Dennett [10]'
Does computation have a role to play in explaining what is possible in the
world? It has been suggested by Dennett [10] that there is a hierarchy of 'grades of
possibility' of things in the world. He suggests the fOllowing as such a hierarchy
(with the possibility of other grades being added):
• Logical
• Physical
• Biological
• Historical
In order for something (a biological something, that is) to actually exist in the
world, it has to be possible at all levels. However, given a putative object which
does nof exist in the world, that non-existence can be explained at one of the levels.
Some things are not logically possible, for example an object which simultane-
ously exists and doesn't. In order to explain the impossibility of such an object a
logical explanation suffices; it is not necessary to go further in the hierarchy to ex-
plain the impossibility of such an object. Just because a putative object contains
characteristics which associate it with a point on the hierarchy doesn't necessarily
place it there; lOm high ants would be biological objects, but it is not necessary to
go as far as biological in the hierarchy to explain their absence in the world; phys-
ics will do that for us. Therefore for every putative object each of those stages can
be examined and whether it is possible at that level or not determined. Broadly any
object placed at one level must also be possible in the previous level, though there
What Kinds of Natural Processes Can Be Regarded as Computations? 331
are complexities, in particular the 'historical' level can contain objects which are
physically possible but which have no biologic al aspect, so it is impossible to place
them meaningfully in or out of the biological category. There are a number of ways
to deal with this consistently, e.g. allowing non-biologic al objects through the bio-
logical layer without being classified or branching the hierarchy into 'biologica\'
and 'non-biological' branches. The final stage in the hierarchy is concerned with
what has actually happened; thus it is necessary to fix a particular point in time be-
fore it is possible to make statements about that final point in the hierarchy.
It is an interesting thought-experiment to take each of the grades in the hierarchy
and think of some putative object which is impossible because of precisely that
reason, i.e. the reason 'earlier' in the hierarchy admits it, whilst the current reason
is sufficient to dismiss it without needing to go further into the Iist.
[11, 12]. Exploration of this idea would take us too far from our main discussion
here.
It may be that 'computational' can be meaningfully placed at a number of points
in the hierarchy, and that these different placements give a taxonomy of different
kinds of computational phenomena.
4.1 Observability
Is the notion of a change in the world being observed essential to the idea of in-
cluding it in the set of computations? This idea can be unpacked in two directions.
Firstly it doesn't seem essential that the computation itself be observable, only
that the inputs and outputs be. In normal computing, we are happy with the idea
that a user of a system (whether that user is a human or another computer system
making an automated enquiry) interacts with the system by specifying input and in
turn receives output; they do not need to see the states which the machine takes on
What Kinds of Natural Processes Can Be Regarded as Computations? 333
in between. Indeed it seems natural to extend the idea of computation to those sys-
tems where the changing state cannot be observed without disturbing the process,
as in quantum computing.
Secondly we can concentrate on the question of what is doing the observing. It
does not seem necessary to restrict the observer to being a conscious entity; it
would seem reasonable to suggest that in a multi-component system, one compo-
nent can carry out a computation and pass its output onto another. It may be the
case that a system can be self-observing.
The aim of considering observability is to attempt to exclude those parts of the
world which are changing but not affecting other parts; however, this does not
seem to be a significant part of 'computing'. Whilst transformations happen in the
world without being observed (in the broad sense of passing their data onto another
system), it does not seem that we should exclude these from what we regard as
computations, or that this is a significant distinction (it is akin to 'when a tree falls
in the woods, does it make a sound?' - entire1y dependent on definition).
Another factor which we may want to take into account in developing a distinction
between computation and non-computation is the flexibility that an external system
has to change the input. An important characteristic of computing is that computers
act on different data; they don't just do the same action aH the time. Still important,
though perhaps less core to the idea of computing, is the idea of programmability.
The ability to influence the system by adding new information would seem to be a
core idea in ascribing an idea of 'computing' to an action in the world. Again this
is well illustrated by the protein folding problem; one of the reasons that we can
easily apply computational reasoning to understanding that problem is that we can
put information into the system as symbol strings, and the range of inputs is vast.
4.6 Summary
Clearly we could consider other properties. However, it seems that we are begin-
ning to tease out what we might mean by a 'computation', and what transforma-
tions we might ascribe to the category 'not a computation'. Clearly this is a topic
around which much future discussion could revolve.
In particular it is interesting to speculate whether it is only a historical happen-
stance that our first encounter with the concept of computation was through the
synthetic creation of computing devices? Could we instead have come across some
of the important ideas in computing by an analytic study of biologic al systems? If
so, which concepts would have most easily been discovered through such an ana-
lytic study? What systems, and what commonalities between systems, might have
suggested these concepts? Are there other concepts, now regarded principally as
analytic scientific concepts, which were originally discovered as synthetic engi-
neering ideas? If so, how did the transition occur from the concept being seen as
purely synthetic to it being seen as a scientific concept which could be applied in
an analytic fashion to understand objects in the natural world? If we believe that
this is important, how do we go about encouraging such an enrichment of ideas in
the context of concepts from computing?
How can we revisit computational ideas without being overly distracted by the
kind of computation that we see on computational devices? What would computer
scientists be studying if the computer had not been invented?
References
1. J. Ciardi. The Collected Poems of John Ciardi. University of Arkansas Press, 1997. ed-
ited by Edward M. Cifelli.
2. S. Forrest and S.A. Hofmeyr. Immunology as information processing. In L.A. Segel
and 1. Cohen, editors, Design Principles for the Immune System and Other Distributed
Autonomous Systems. Oxford University Press, Oxford, 2000.
3. A. S. Perelson and G. Weisbuch. Immunology for physicists. Reviews of Modem Phys-
ics,69(4): 1219-1267,1997.
4. A. Fraenkel. Complexity of protein folding. Bulletin of Mathematical Biology, 55(6):
1199-1210,1993.
5. A. Berger and T. Leighton. Protein folding in the hydrophilic-hydrophobic (HP) model
is NP-complete. Joumal ofComputational Biology, 5(1): 27-40,1998.
6. P. Crescenzi, D. Goldman, C. Papadimitriou, A. Piccolboni, and M. Yannakakis. On
the complexity of protein folding. Journal of Computational Biology, 5: 423-465,
1998.
7. C. Levinthal. How to fold graciously. In J. T. P. DeBrunner and E. Munck, editors,
Mossbauer Spectroscopy in Biological Systems: Proceedings of a meeting held at AI-
lerton House, Monticello, Illinois, pages 22-24. University of Illinois Press, Chicago,
III., 1969.
8. R. P. Feynman. Simulating physics with computers. International Journal of Theoreti-
cal Physics, 21: 467-488, 1982.
336 C. G. Johnson
Amos,M.
School of Biological Science and Engineering and Computer Science,
University of Exeter, Exeter EX4 4JH, United Kingdom
Bolouri, H.
Institute for Systems Biology, Seattle, W A 98103-8904, and Division of
Biology, 156-29 California Institute of Technology, CA 91125 USA
Brown, D.
School of Biomedical and Clinica! Laboratory Sciences,
University of Edinburgh, Hugh Robson Building, George Square,
Edinburgh EH8 9XD, United Kingdom
Brown, R
Mathematics Division, School of Informatics, University of Wales,
Bangor, Gwynedd LL57 1UT, United Kingdom
de Castro, L. N.
School of Electrical Engineering and Computing, State University of
Campinas, 13081-970 Campinas, Sâo Paulo, BraziI
Cho, K.-H.
School of Electrical Engineering, University of Vlsan,
Vlsan, 680-749 South Korea
Feng, J.
COGS, Sussex University, Brighton BNI 9QH, and Newton Institute,
Cambridge University, Cambridge CB3 OEH, United Kingdom
Hart, E.
School of Computing, Napier University
Edinburgh, Scotland, United Kingdom
Holcombe, M.
Department of Computer Science, University of Sheffield,
Regent Court, Portobello Street, Sheffield SI 4DP, United Kingdom
Kolch, W.
Institute of Biomedical and Life Sciences, University of Glasgow
CRC Beatson Laboratories, Garscube Estate,
Switchback Road, Glasgow G61 lBD, United Kingdom
Leng,G.
School of Biomedica1 and Clinical Laboratory Sciences,
University of Edinburgh, Hugh Robson Building, George Square,
Edinburgh EH8 9XD, United Kingdom
MacGregor, D. J.
School of Biomedica1 and Clinical Laboratory Sciences,
University of Edinburgh, Hugh Robson Building, George Square,
Edinburgh EH8 9XD, United Kingdom
Monk,N.
Centre for Bioinformatics and Computational Biology,
Division of Genomic Medicine, University of Sheffield,
Royal Hallamshire Hospital, Sheffield S 10 2JG, United Kingdom
Nagl, S. B.
Department of Biochemistry and Molecular Biology,
University College London,
Gower Street, London WClE 6BT, United Kingdom
Parish,1.H.
School of Biochemistry and Molecular Biology, University of Leeds,
Leeds, LS2 9JT, United Kingdom
List of Contributors 339
Porter, T.
Mathematics Division, School of lnformatics, University of Wales,
Bangor, Gwynedd Ll57 1UT, United Kingdom
Sant, P.
Department of Computer Science, King's College
London WC3R 2LS, United Kingdom
Schilstra, M.
Science and Technology Research Centre, University of Hertfordshire,
College Lane, Hatfield, HertfordshireALlO 9AB, United Kingdom
Tateson. R.
Future Technologies Group, lntelligent Systems Lab,
BT Exact Technologies, PPlIl2 Orion Building, Adastral Park,
Martlesham, Ipswich IP 3RE, United Kingdom
Timmis, J.
Computing Laboratory, University of Kent
Canterbury, Kent, CT2 7NF, United Kingdom
Tyrrell, A.
Bio-inspired Architectures Laboratory, Department of Electronics,
University of York
York YOlO 5DD, United Kingdom
Warner, G. 1.
Unilever Research Colworth
Colworth House, Sharnbrook, Bedford MK44 ILQ, United Kingdom
WU, Q. H.
Department of Electrical Engineering, University of Liverpool
Liverpool L69 3BX, United Kingdom
Wolkenhauer, O.
Department of Biomolecular Sciences
and Department of Electrical Engineering and Electronics,
Control System Centre, UMIST, Manchester M60 IQD, United Kingdom
Index
CATH 130
fault diagnosis 78
ceH 12, 98, 264, 306
fault tolerance 11, 107
ceHlineage 175
Feng 185
cellsim 20
Feynman 117,329
Cho 305
field programmable gate array 97,97
ciliate 269
finite state machi ne 106, 252
cIassification 126
Fisher 277
clonal selection 62
fitness 39
coli mit 279, 296
function 121, 127, 134, 277
communication network 10
communicating x-machi ne 259 'glue' 134,280
complex systems 135, 138,328 GH releasing hormone 229, 234
GH model, 237, 240
342 Index