A Visual Process Calculus for Biology
Andrew Phillips
Microsoft Research
7 JJ Thomson Avenue
CB3 0FB Cambridge UK
Abstract. This chapter presents a visual process calculus for designing
and simulating computer models of biological systems. The calculus is
based on a graphical variant of stochastic pi-calculus, extended with mobile
compartments, and the simulation algorithm is based on standard
kinetic theory of physical chemistry. The calculus forms the basis of a
formal visual programming language for biology. The basic primitives of
the calculus are first introduced by a series of examples involving genes
and proteins. More complex features of the calculus are then illustrated
by examples involving gene networks, cell differentiation, and immune
system response. The main benefit of the calculus is its ability to model
large systems incrementally, by directly composing simpler models of
subsystems. The formal nature of the calculus also facilitates mathematical
analysis of models, which in future could help provide insight into
some of the underlying properties of biological systems.
1 Introduction
In many respects, biological systems are like massively parallel, highly complex,
error-prone computer systems. The genetic code of an entire organism is stored in
digital form as DNA, but instead of a binary code there are four letters, G,A,T,C.
A specialised protein travels along the DNA to read the code, like reading a
sequence of instructions in a program. The code itself is divided into genes
with different functions, in the same way that a computer program is divided
into functional modules. The genes produce RNA which is more accessible, like
loading a module into memory. The RNA produces proteins, which fold into
different shapes to perform the essential functions of a living organism. Some of
these proteins can interact with the DNA to switch on different genes at different
times, in the same way that some modules of a computer program can trigger
the execution of other modules. In general, the genetic code can be viewed not
as a single program, but rather as a library of programs that are executed on
demand in response to environmental signals [22].
The Human Genome Project had the ambitious goal of mapping out the
complete genetic code in humans. Scientists had hoped that by listing this code
they would be able to unravel the mysteries of how the human body functions.
Instead, the code raised many more questions than answers. It was then that
the field of Systems Biology rose to the occasion, with the ultimate goal of being
able to understand and predict the behaviour of biological organisms. Systems
Biology can be broadly divided into two complementary approaches. On the one
hand scientists are doing experiments in the lab and studying the results, in
order to infer key properties of biological systems. On the other hand they are
using this knowledge to build detailed models of systems and then testing these
models in the lab. Such models are a powerful tool, allowing scientists to simulate
a range of experiments on a computer before testing the most promising ones,
saving resources and enabling more effective research. They also help to clarify
the mechanisms of how a biological system functions, and are beginning to play
a role in understanding disease.
In recent years we have witnessed a rise in the publication of biological models,
but as these models grow in size they are becoming increasingly difficult to
understand, maintain and extend. Many published models consist of hundreds
of reactions, but as our knowledge of biological systems continues to increase,
models consisting of tens of thousands of reactions will soon be commonplace.
Clearly, writing a model of this size as a single list of reactions is not a scalable
solution, for the same reasons that we would not write a large software
program as a single list of thousands of instructions. Rather, we need a way of
decomposing a large biological model into a collection of smaller, more manageable
building blocks. Since the early days of computer programming, many
high-level languages have successfully been developed to improve program modularity.
But biological programs differ from traditional computer programs in
two fundamental ways: first, they are massively parallel and second, each instruction
has a certain probability of executing, so we are never quite sure what
2
will happen next. These fundamental differences suggest a need for specialised
programming languages for biology.
Over the years, there has been considerable research on programming languages
for complex parallel computer systems. Some of this research is also applicable
to biological systems, which are typically highly complex and massively
parallel. One example of such a language is the pi-calculus [18,33], originally
proposed by Milner, Parrow and Walker [19] to model communicating computer
networks that can dynamically reconfigure themselves over time. A stochastic
extension was later introduced by Priami [28], to model the performance of
computer networks in the presence of communication delays. More recently, the
stochastic pi-calculus has been used to model and simulate a range of biological
systems [17,28], an approach pioneered by Regev, Silverman and Shapiro [30].
In addition, a graphical variant of the calculus has been defined [26], to help
make the calculus more broadly accessible. The stochastic pi-calculus allows the
components of a biological system to be modelled independently, rather than
modelling the individual reactions, enabling large models to be constructed by
direct composition of simple components [1]. The calculus also facilitates mathematical
analysis of systems using a range of established techniques, which in
future could help provide insight into some of the underlying properties of biological
systems. Various stochastic simulators have been developed for the calculus,
such as [28,25], which are used to perform virtual experiments on biological models.
Such in silico experiments can be used to formulate testable hypotheses on
the behaviour of biological systems, as a guide to future experimentation in vivo.
This chapter presents the SPiM calculus, a visual process calculus for designing
and simulating computer models of biological systems. The calculus is
based on a graphical variant of the stochastic pi-calculus, originally presented in
[26], extended with mobile compartments. The basic primitives of the calculus
are first introduced by a series of examples involving genes and proteins. More
complex features of the calculus are then illustrated by examples involving gene
networks, cell differentiation, and immune system response. All of the biological
models in this chapter were simulated using the SPiM simulator [25] and are
available for download online, together with the simulator itself1
.
1.1 SPiM Processes and Chemical Reactions
The Gillespie algorithm [13] was introduced in the 1970s for the exact stochastic
simulation of coupled chemical reactions, and since then a vast number of
chemical and biological models have been simulated using this approach. Even
today, many biological models continue to be published as collections of chemical
reactions, and the format is supported by most modelling standards such as the
Systems Biology Markup Language [15].
Essentially, the SPiM calculus allows a system of chemical reactions to be
rewritten as a collection of modular processes, which interact with each other by
message-passing. Fig. 1 presents a comparison between chemical reactions and
1
http://research.microsoft.com/spim
3
A. Chemical reactions B. SPiM processes
Xp + Y →a X + Yp
X + Yp →d Xp + Y
Xp = a.X
X = d.Xp
Y = a.Yp
Yp = d.Y
Fig. 1. Comparison between chemical reactions (A) and SPiM processes (B) for a
model of protein activation.
SPiM processes for a simple model of protein activation. For both models, the
graphical representation at the top is equivalent to the textual representation
at the bottom. The chemical reaction model of Fig. 1A is constructed by listing
the individual reactions in the system. An active protein Xp can interact with
a protein Y through an a reaction, to produce a protein X and an active protein
Yp. The reverse reaction d can also occur. For the graphical representation,
each shape represents a protein in a particular state and each box represents a
reaction, with inbound edges (arcs) from reactants and outbound edges to products.
In contrast, the SPiM model of Fig. 1B is constructed by describing the
behaviour of the individual components in the system. This is achieved by splitting
the reaction a into two complementary actions, a send a and a receive a, and
similarly for the reaction d. Thus, an active protein Xp can send on a and evolve
to a protein X, which can receive on d and evolve to Xp. Similarly, a protein Y
can receive on a and evolve to an active protein Yp, which can send on d and
evolve to Y . For the graphical representation, each shape represents a protein in
a particular state and each connected graph represents the set of possible states
of a given protein, where a labelled edge represents an action that the protein
can perform in order to change from one state to another. Since Xp can send
on a and Y can receive on a, the two proteins can interact with each other and
evolve to a new state simultaneously. The SPiM model explicitly represents the
fact that the Xp protein evolves to X and the Y protein evolves to Yp after the
interaction takes place. This contrasts with the chemical reaction model, which
does not explicitly state which product comes from which reactant. Although
we can guess by looking at the reactant and product names, we could equally
well interpret the reactions to mean that Xp becomes Yp and Y becomes X.
Another feature of the SPiM model is that the description of the X protein can
4
A. Chemical reactions B. SPiM processes
X1p + Y1 →a X1 + Y1p
X1p + Y2 →a X1 + Y2p
X1p + Y3 →a X1 + Y3p
X2p + Y1 →a X2 + Y1p
X2p + Y2 →a X2 + Y2p
X2p + Y3 →a X2 + Y3p
X3p + Y1 →a X3 + Y1p
X3p + Y2 →a X3 + Y2p
X3p + Y3 →a X3 + Y3p
X1 + Y1p →d X1p + Y1
X1 + Y2p →d X1p + Y2
X1 + Y3p →d X1p + Y3
X2 + Y1p →d X2p + Y1
X2 + Y2p →d X2p + Y2
X2 + Y3p →d X2p + Y3
X3 + Y1p →d X3p + Y1
X3 + Y2p →d X3p + Y2
X3 + Y3p →d X3p + Y3
X1p = a.X1
X1 = d.X1p
X2p = a.X2
X2 = d.X2p
X3p = a.X3
X3 = d.X3p
Y1 = a.Y1p
Y1p = d.Y1
Y2 = a.Y2p
Y2p = d.Y2
Y3 = a.Y3p
Y3p = d.Y3
Fig. 2. Comparison between chemical reactions and SPiM processes for a combinatorial
model of protein activation. (A) For the chemical reaction model there are 18 reactions,
9 forward reactions between Xip and Yj and 9 reverse reactions between Xi and Yjp
with i, j ∈ {1, . . . , 3}. In the general case of i ∈ {1, . . . , N} and j ∈ {1, . . . , M} there
will be N · M forward reactions and N · M reverse reactions. (B) For the SPiM model
there are 12 actions, 6 forward actions for Xip, Yj and 6 reverse actions for Xi, Yjp with
i, j ∈ {1, . . . , 3}. In the general case of i ∈ {1, . . . , N} and j ∈ {1, . . . , M} there will be
N + M forward actions and N + M reverse actions.
be given independently of that of the Y protein, allowing the behaviours of the
two proteins to be modelled separately. A notion of complementary send a and
receive a is needed in order to decompose the system into individual modules
in this way. If we simply wrote Xp = a.X and Y = a.Yp and allowed all the
a actions to interact, there would be nothing to prevent two Xp proteins from
interacting with each other, which should not be possible. The use of complementary
actions in the model is also consistent with the physical characteristics
of the system, since the protein Xp actually sends its phosphate group p to a
receiver protein Y , which then becomes phosphorylated as Yp. Conversely, the
5
phosphorylated protein Yp sends its phosphate group p back to protein X. More
generally, complementary actions can be used to represent a range of biological
interactions, such as two proteins with complementary binding sites, two RNA
strands with complementary sequences, two atoms with complementary electron
configurations, etc.
In addition to improving modularity, the SPiM calculus can also help to
reduce the size of a biological model. Consider the models in Fig. 2, which describe
a collection of proteins that activate each other. For both models, the
graphical representation at the top is equivalent to the textual representation
at the bottom. Each of the proteins X1p, . . . , X3p can activate each of the proteins
Y1, . . . , Y3, and similarly for the proteins Y1p, . . . , Y3p and X1, . . . , X3. For
the chemical reaction model in Fig. 2A, each interaction needs to be described
explicitly, resulting in a combinatorial explosion in the number of reactions. For
the SPiM model in Fig. 2B, instead of explicitly stating which proteins interact
with which other proteins, it is only necessary to describe the channels on which
each protein can send and receive. Implicitly, all the senders on a given channel
can interact with all the receivers on the same channel, resulting in a more
compact model whose size varies linearly with the number of proteins. As the
system grows in size, the advantages of a modular representation become more
noticeable, since the behaviours of individual proteins can be modified without
changing the rest of the model.
2 Basic Examples
This section presents the basic constructs of the SPiM calculus, through a collection
of simple examples. The examples are presented at the level of interacting
genes and proteins, but the constructs of the calculus can equally well be used
at the level of interacting cells or organisms. The simulation results for the examples
were obtained using the SPiM simulator, which executes a SPiM model
according to an exact simulation algorithm [25] based on standard principles of
physical chemistry [13].
2.1 Protein Production
The first example models the production of protein P from a gene G, as shown
in Fig. 3. In the SPiM model of Fig. 3A, the graphical representation at the
top is equivalent to the textual representation at the bottom. The gene G can
do a produce action, after which a new protein P is executed in parallel with
the gene, written (G | P). Graphically, the parallel execution is represented as
a box with outbound edges to G and P. The protein P can then do a degrade
action, after which the empty process is executed, written 0. Graphically, the
empty process is represented as a node with a dotted outline. In order to obtain
the simulation results of Fig. 3B, the produce and degrade actions are associated
with corresponding rates given by 0.1 and 0.001, respectively. Each rate
characterises an exponential distribution, such that the probability of an action
6
A. SPiM model B. Simulation results
G = produce.(G | P)
P = degrade.0
Simulating G with produce = 0.1, degrade = 0.001
Fig. 3. A SPiM model of protein production (A), with associated simulation results
(B).
with rate r occurring within time t is given by F(t) = 1 − e−rt
. The average
duration of the action is given by the mean 1/r of this distribution. Thus, the
produce action will take on average 10 time units, while the degrade action will
take on average 1000 time units. In this example the time units are assumed
to be seconds, but in general they can be any unit that is appropriate for the
system under consideration. The simulation results show the population of protein
P over time. The population of the gene G is always 1 and is not shown.
Initially there is one gene G and no protein P in the system. The number of
proteins increases over time until it reaches an equilibrium of about 100, and
then continues to fluctuate around this equilibrium.
The simulation of the model in Fig. 3 is described in more detail in Fig. 4 and
5. The execution steps in Fig. 4 describe the evolution of the system over time,
where the graphical representation at the top is equivalent to the textual representation
at the bottom. Each frame represents a state of the system, where the
populations of the gene G and protein P are indicated next to the corresponding
nodes in the graph, and the next actions to be executed are highlighted in grey.
The calculations in Fig. 5 show the propensities of the reactions at each step of
the simulation. For the system of Fig. 3 there are two possible reactions, produce
and degrade. The propensity R(produce) of the produce reaction is given by the
number of genes G that can perform this action, written [G], multiplied by the
rate 0.1 of the action. Similarly, the propensity R(degrade) of the degrade reaction
is given by the number of proteins P that can perform this action, written
[P], multiplied by the rate 0.001 of the action. In accordance with [13], the prob-
7
1 2 3 4
G G | P G | 2 · P G | 1 · P
Fig. 4. Execution steps for the protein production model of Fig. 3. (1) Initially there is
one gene G and no protein P. The gene then does a produce action. (2) A new protein
is created in parallel with the gene, which then does a second produce action. (3) A
second protein is created, and one of the proteins then does a degrade action. (4) One
of the proteins is degraded.
# R(produce) = [G]·produce R(degrade) = [P]·degrade
1 0.1 = 1·0.1 0 = 0·0.001
2 0.1 = 1·0.1 0.001 = 1·0.001
3 0.1 = 1·0.1 0.002 = 2·0.001
4 0.1 = 1·0.1 0.001 = 1·0.001
... 0.1 = 1·0.1 0.1 = 100·0.001
Fig. 5. Propensity calculations for the execution steps of Fig. 4, where the probability
of a reaction is proportional to its propensity. (1) The system starts with a single gene
G, which can produce a protein P with propensity 0.1. (2) A second protein can be
produced with propensity 0.1. (3) One of the proteins can be degraded with propensity
0.002. Although the propensity of the degrade reaction is much lower than that of the
produce reaction, there is still a small chance that a protein can degrade before another
is produced. (...) Eventually an equilibrium is reached at around 100 proteins, where
the propensity [G] · 0.1 of the produce reaction is equal to the propensity [P] · 0.001 of
the degrade reaction.
ability of executing a given reaction is proportional to its propensity. In order
to calculate the probability of each reaction at a given instant, the simulation
algorithm adds up the propensities of all the possible reactions in the system and
divides the propensity of each reaction by this total. The system reaches an equi-
8
A. SPiM model B. Simulation results
Xp = a.X Y = a.Yp
X = d.Xp Yp = d.Y
Simulating 100·Xp | 100·Y , with ρ(a) = 100, ρ(d) = 10
Fig. 6. A SPiM model of protein interaction (A), with associated simulation results
(B).
librium when the propensity of the produce reaction is equal to the propensity
of the degrade reaction.
2.2 Protein Interaction
The second example models the interaction between proteins, as shown in Fig. 6,
where the SPiM model of Fig. 6A is the same as in Fig. 1A. An active protein
Xp can send on a and evolve to a protein X, which can receive on d and evolve
to Xp. Similarly, a protein Y can receive on a and evolve to an active protein Yp,
which can send on d and evolve to Y . In order to obtain the simulation results of
Fig. 6B, the channels a and d are associated with corresponding rates given by
ρ(a) = 100 and ρ(d) = 10, respectively, where the rate of the channel denotes the
rate of a single interaction on this channel. As in the previous example, each rate
r characterises an exponential distribution with mean 1/r. Thus, an interaction
on channel a will take on average 0.01 time units, while an interaction on channel
d will take on average 0.1 time units. The simulation results show the populations
of proteins Xp and X over time. The populations of Y and Yp are not shown
since they are equal to those of Xp and X, respectively. Initially there are 100·Xp
and 100 · Y proteins. The number of Xp proteins decreases and the number of
X proteins increases over time until the system reaches an equilibrium of about
24·Xp and 76·X proteins, and the populations then continue to fluctuate around
this equilibrium.
The simulation of the model in Fig. 6 is described in more detail in Fig. 7 and
8. The execution steps in Fig. 7 describe the evolution of the system over time,
and the calculations in Fig. 8 show the propensities of each reaction at each step
of the simulation. There are two possible reactions in the system, a and d. The
propensity R(a) of the a reaction is given by the number of possible interactions
9
1 2
100·Xp | 100·Y 99·Xp | 99·Y | X | Yp
3 4
98·Xp | 98·Y | 2·X | 2·Yp 99·Xp | 99·Y | X | Yp
Fig. 7. A sequence of execution steps for the protein interaction model of Fig. 6. (1)
Initially there are 100 · Xp and 100 · Y proteins. An Xp and a Y protein interact on
channel a. (2) The two interacting proteins are converted to X and Yp, respectively.
An additional Xp and Y protein interact on channel a. (3) An X and Yp protein
interact on channel d. (4) The two interacting proteins are converted back to Xp and
Y , respectively.
# R(a) = [Xp]·[Y ]·ρ(a) R(d) = [X]·[Yp]·ρ(d)
1 1000000 = 100·100·100 0 = 0·0·10
2 980100 = 99·99·100 10 = 1·1·10
3 960400 = 98·98·100 40 = 2·2·10
4 980100 = 99·99·100 10 = 1·1·10
... 57600 = 24·24·100 57760 = 76·76·10
Fig. 8. Propensity calculations for the execution steps of Fig. 7, where the probability
of a reaction is proportional to its propensity. (1) The system starts with [Xp] =
[Y ] = 100. An Xp and a Y protein can interact on channel a with propensity 1000000.
(2) An additional Xp and Y protein can interact with propensity 980100. (3) An X
and Yp protein can then interact on channel d with propensity 40. (...) Eventually
an equilibrium is reached at around [Xp] = [Y ] = 24 and [X] = [Yp] = 76, where the
propensity [Xp]·[Y ]·100 of the a reaction is roughly equal to the propensity [X]·[Yp]·10
of the d reaction.
on channel a times the rate of a single interaction. Since each protein Xp can
interact with each protein Y on channel a, the number of possible interactions
on a is given by the population of Xp, written [Xp], times the population of Y ,
10
A. SPiM model B. Simulation results
X = +
b.X Y = +
b.Y
X = −
b.Xp Y = −
b.Y
Simulating 100·X | 100·Y , with ρ(+
b) = 100,
ρ(−
b) = 10
Fig. 9. A SPiM model of protein binding (A), with associated simulation results (B).
written [Y ]. The rate of a single interaction is given by the rate of the channel
a, written ρ(a). Therefore, the propensity R(a) of the a reaction is given by
[Xp] · [Y ] · ρ(a). Similarly, the propensity R(d) of the d reaction is given by
[X] · [Yp] · ρ(d). The system reaches an equilibrium when the propensity of the a
reaction is equal to the propensity of the d reaction.
2.3 Protein Binding
The third example models the binding and unbinding of proteins, as shown in
Fig. 9. In the SPiM model of Fig. 9A, the graphical representation at the top is
equivalent to the textual representation at the bottom. In order to model binding
and unbinding reactions, we associate a given channel b with corresponding
binding and unbinding channels, written +
b and −
b, respectively. A protein X
can send and bind on +
b and evolve to a protein X , which can send and unbind
on −
b and evolve to X. Similarly, a protein Y can receive and bind on +
b and
evolve to a protein Y , which can receive and unbind on −
b and evolve to Y .
The send and receive actions on binding channel +
b represent the fact that the
interacting proteins become bound to each other after the interaction takes place.
The bound proteins can then unbind from each other by interacting on unbinding
channel −
b. In order to obtain the simulation results of Fig. 9B, the channels
+
b and −
b are associated with corresponding rates given by ρ(+
b) = 100 and
ρ(−
b) = 10, respectively. The simulation results show the populations of proteins
X and X over time. The populations of Y and Y are not shown since they are
equal to those of X and X , respectively. Initially there are 100 · X and 100 · Y
proteins. The number of X proteins decreases and the number of X proteins
increases over time until the system reaches an equilibrium of about 3·X proteins
and 97 · X proteins.
11
1 2
100·X | 100·Y 99·X | 99·Y | (−
b) (X | Y )
3 4
98·X | 98·Y | 2·(−
b) (X | Y ) 99·X | 99·Y | (−
b) (X | Y )
Fig. 10. A sequence of execution steps for the protein binding model of Fig. 9. (1)
Initially there are 100 · X and 100 · Y proteins. An X and a Y protein bind to each
other by interacting on channel +
b. (2) The interacting proteins are converted to X
and Y , respectively, and are bound to form a complex (−
b) (X | Y ). An additional
X and Y protein bind to each other by interacting on channel +
b, resulting in 2
complexes 2·(−
b) (X | Y ). (3) An X and a Y protein in one of the complexes unbind
from each other by interacting on the local channel −
b. (4) The two interacting proteins
are converted back to X and Y , respectively.
# R(+
b) = [X]·[Y ]·ρ(+
b) R(−
b) = [(−
b) (X | Y )]·ρ(−
b)
1 1000000 = 100·100·100 0 = 0·10
2 980100 = 99·99·100 10 = 1·10
3 960400 = 98·98·100 20 = 2·10
4 980100 = 99·99·100 10 = 1·10
... 900 = 3·3·100 970 = 97·10
Fig. 11. Propensity calculations for the execution steps of Fig. 10, where the probability
of a reaction is proportional to its propensity. (1) The system starts with [X] = [Y ] =
100. An X and a Y protein can bind by interacting on channel +
b with propensity
1000000. (2) An additional X and Y protein can interact on +
b with propensity 980100.
(3) A pair of bound X and Y proteins can unbind by interacting on −
b with propensity
20. Eventually an equilibrium is reached at around [X] = [Y ] = 3 and [(−
b) (X |
Y )] = 97, where the propensity [X] · [Y ] · 100 of the +
b reaction is roughly equal to
the propensity [(−
b) (X | Y )] · 10 of the −
b reaction.
The simulation of the model in Fig. 9 is described in more detail in Fig. 10
and 11. There are two possible reactions in the system, +
b and −
b. As with
12
the previous example, the propensity R(+
b) of the +
b reaction is given by the
number of possible interactions on binding channel +
b times the rate of a single
interaction, written [X] · [Y ] · ρ(+
b). After the interaction takes place, each X
protein is bound to a single Y protein on a local channel −
b, and the pair of
bound proteins is written (−
b) (X | Y ). The number of possible interactions
on unbinding channel −
b is given by the number of pairs of bound X and
Y proteins, written [(−
b) (X | Y )]. Therefore, the propensity R(−
b) of the −
b
reaction is given by [(−
b) (X | Y )]·ρ(−
b). If we were instead to use a non-binding
interaction on channel b, then every X protein would be able to interact with
every other Y protein, and the equilibrium would be 24 X and Y proteins,
similar to the previous example. But since each X protein can only interact
with a single Y protein, a reaction between X and Y is much less likely to
occur, and the equilibrium is 97 X and Y proteins. This illustrates how the use
of binding interactions can significantly alter the equilibrium of a system.
2.4 Summary
In this section we introduced the basic primitives of the SPiM calculus, using
examples of protein production, protein interaction and protein binding. The
examples were presented in both graphical and textual forms, which could be
used interchangeably.
The first example showed how a gene G could produce a new protein P in
parallel with itself, written G = produce.(G | P), and how a protein P could
degrade to nothing, written P = degrade.0. Both of these definitions gave the
behaviours of the gene G and protein P. In order to run a simulation, initial
populations were specified for processes G and P, and stochastic rates were
associated to the produce and degrade actions, where the average duration of
an action was defined as one over its rate. The propensity of each action was
defined as the number of processes that can perform this action, times the rate
of the action. At each step of the simulation, a given action was chosen with
probability proportional to its propensity. An equilibrium was reached when the
propensities of both actions were roughly the same.
The second example showed how a protein Xp could send on a and evolve
to a protein X, written Xp = a.X, and how a protein Y could receive on a
and evolve to a protein Yp , written Y = a.Yp. Since these two proteins could
send and receive on the same channel, they could interact on this channel and
evolve simultaneously. The reverse interaction could also occur on channel d,
where X = d.Xp and Yp = d.Y . Initial populations were specified for processes
X and Yp, and stochastic rates were associated to the channels a and d, where
the average duration of a channel interaction was defined as one over the channel
rate. The propensity of each channel interaction was defined as the number of
sends on the channel, times the number of receives on the channel, times the
rate of the channel. At each step of the simulation, a given channel interaction
was chosen with probability proportional to its propensity. An equilibrium was
reached when the propensities of both channel interactions were roughly the
same.
13
The third example introduced a notion of binding and unbinding channels,
written +
b and −
b, respectively. The example showed how a protein X could send
and bind on +
b and evolve to a protein X , written X = +
b.X , and how a protein
Y could receive and bind on +
b and evolve to a protein Y , written Y = +
b.Y .
After the binding interaction on channel +
b the two proteins became bound to
each other, written (−
b) (X | Y ), and the bound proteins could then unbind by
interacting on channel −
b, where X = −
b.X and Y = −
b.Y . Initial populations
were specified for processes X and Y , and stochastic rates were associated to
channels +
b and −
b. The key difference with the previous example was that each
X protein became bound to a single Y protein on channel −
b. The propensity
of an unbinding channel interaction was defined as the number of pairs bound
by the channel, times the rate of the channel. The example illustrated how the
use of binding interactions could alter the equilibrium of a system.
The above three examples illustrate some of the basic constructs of the SPiM
calculus. More complex features of the calculus are illustrated in the following
sections, and the full definition of the calculus is given in Appendix 7.
3 Gene Networks
3.1 Repressilator Model
One of the ultimate challenges of Systems Biology is to undertand in detail how
living systems function. Once this has been achieved, in principle we should be
able to engineer living systems to behave in predictable ways. Such engineering
of biological systems could also pave the way for new medical therapies and
treatments. In recent years, scientists have been investigating how to genetically
engineer simple organisms such as bacteria, by inserting combinations of genes
that interact with each other in predefined ways. For example, [11] presents the
results of a well-known experiment, in which E. coli were genetically engineered
to repeatedly glow on and off. A network of three genes was inserted into the
bacteria, consisting of the genes tet, lambda and lac, which mutually repressed
each other. More precisely, the tet gene produced proteins to block the lambda
gene, which produced proteins to block the lac gene, which produced proteins to
block the tet gene, completing the cycle. An additional gfp gene was inserted to
produce a green fluorescent protein (GFP), which made the bacteria glow. The
network was constructed so that the tet gene produced proteins that also blocked
the gfp gene, thereby preventing the production of GFP. When engineered in live
bacteria, the gene network caused the bacteria to repeatedly glow on and off.
The resulting network was called the Repressilator, since the mutual repression
of genes gave rise to oscillations in GFP levels. The gene network was also
reproduced in the descendants of the bacteria, which themselves continued to
glow on and off.
The Repressilator network can be simulated in SPiM by first constructing a
parameterised model of a gene with negative control as shown in Fig. 12, based
on previous work presented in [1] and [2]. The model is similar to the protein
production model of Fig. 3, except that the gene and protein are parameterised
14
A. SPiM model B. Simulation results
G(a, b) = produce.(G(a, b) | P(b))
+ a.B(a, b)
B(a, b) = unblock.G(a, b)
P(b) = degrade.0 + b.P(b)
Simulating G(a, b) | G(b, c) | G(c, a), with
produce = 0.1, degrade = 0.001,
unblock = 0.0001 and ρ(a) = ρ(b) = ρ(c) = 1.
Fig. 12. A SPiM model of a parameterised gene with negative control (A), with associated
simulation results (B).
Gate Auto-inhibitor Bistable Switch Oscillator
G(a, b) G(b, b) G(a, b) | G(b, a) G(a, b) | G(b, c) | G(c, a)
Fig. 13. Modular construction of gene networks using the parameterised gene of
Fig. 12.
by the channels on which they can interact. The gene G is parameterised by
channels a, b, while the protein P is parameterised by channel b. The parameterised
gene G(a, b) can perform one of two actions: either it can produce, after
which a new protein P(b) is executed in parallel with the gene, or it can receive
on a and become blocked to B(a, b). The choice between production and blocking
is written using a (+) symbol. Graphically, this is represented by two outbound
15
edges from node G(a, b), labelled with the produce and a actions, respectively.
If the produce action is chosen then a new protein is produced in parallel with
the gene. If the a action is chosen then the gene becomes blocked to B(a, b) and
is no longer able to produce. The blocked gene can then unblock in order to
become active again. The parameterised protein P(b) can also perform one of
two actions: either it can degrade, after which the empty process is executed, or
it can send on b and continue executing. Therefore, G(a, b) denotes a gene that
is blocked by receiving on a, and that produces proteins which send on b.
The use of complementary actions in the model is consistent with the physical
characteristics of the system, since G(a, b) denotes a gene with promoter region
a, which produces proteins P(b) with binding site b. When the promoter region is
free, the gene is able to repeatedly produce proteins. When the promoter region a
is blocked by a protein with a complementary binding site a, the gene is no longer
able to produce. The parameterised gene can be used to construct networks of
arbitrary complexity, as illustrated in Fig. 13. An auto-inhibitory gene can be
defined as G(b, b), i.e. a gene with promoter region b that produces proteins with
binding site b. Thus, the proteins produced by the gene will in turn cause the
gene to block. In this setting, the parameters (a, b) of the gene are replaced with
actual channels (b, b), resulting in a gene that blocks itself. A bistable network
can be defined as G(a, b) | G(b, a), i.e. two genes that block each other. The first
gene uses channels (a, b) as usual, while the second gene uses channels (b, a).
Similarly, an oscillator network can be defined as G(a, b) | G(b, c) | G(c, a).
The simulation results for the oscillator network of genes are shown in Fig. 12B,
where the produce, degrade and unblock actions are associated with corresponding
rates given by 0.1, 0.001 and 0.0001, respectively, and the channels a, b, c are
associated with rate 1. The results show the populations of proteins P(a), P(b),
P(c) over time. Initially, about 100 P(a) proteins are produced while the other
two proteins are absent. Subsequently, about 100 P(b) proteins are produced,
which repress the production of P(a). Eventually, about 100 P(c) proteins are
produced, which repress the production of P(b), completing the cycle. The alternate
production of proteins continues in this order until the end of the simulation.
From the simulation results it is not immediately clear how the oscillations are
initialised, or why they proceed in a specific order. To understand the mechanism
by which the gene network is able to produce these oscillations, it is necessary
to closely examine the interactions between genes and proteins over time.
3.2 Model Execution
The simulation of the model in Fig. 12 is described in more detail in Fig. 14
and 15. Fig. 14 shows the first few execution steps for the simulation results
of Fig. 12, where the definitions of G(a, b), G(b, c) and G(c, a) are expanded so
that the behaviour of each gene is represented as a separate graph. The graph
for each gene is simply a copy of the model in Fig. 12, but with parameters
(a, b) being replaced by (a, b), (b, c) and (c, a), respectively. Fig. 15 shows the
propensities of the different reactions for each of the execution steps of Fig. 14.
16
1
2
3
4
Fig. 14. A sequence of execution steps for the repressilator model of Fig. 12. (1) Initially
there is one copy of each gene. The gene G(a, b) produces a protein P(b). (2) Protein
P(b) sends on b and causes G(b, c) to block. (3) Gene G(c, a) produces a protein P(a).
(4) Protein P(a) sends on a and causes G(a, b) to block.
17
# R(produce) R(unblock) R(degrade) R(a) R(b) R(c)
= [G(a, b)]·produce
+[G(b, c)]·produce
+[G(c, a)]·produce
= [B(a, b)]·unblock
+[B(b, c)]·unblock
+[B(c, a)]·unblock
= [P(b)]·degrade
+[P(c)]·degrade
+[P(a)]·degrade
1 0.1 + 0.1 + 0.1 0 0 0 0 0
2 0.1 + 0.1 + 0.1 0 0.001 + 0 + 0 0 1·1·1 0
3 0.1 + 0 + 0.1 0 + 0.0001 + 0 0.001 + 0 + 0 0 0 0
4 0.1 + 0 + 0.1 0 + 0.0001 + 0 0.001 + 0 + 0.001 1·1·1 0 0
5 0 + 0 + 0.1 0.0001 + 0.0001 + 0 0.001 + 0 + 0.001 0 0 0
... 0 + 0 + 0.1 0.0001 + 0.0001 + 0 0 + 0 + 100·0.001 0 0 0
Fig. 15. Propensity calculations for the execution steps of Fig. 14, where the propensity
R(a) of the a reaction is defined as [G(a, b)]·[P(a)]·ρ(a). The propensities R(b) and R(c)
are defined in a similar fashion. (1) The simulation starts with one copy of each gene,
which can each produce a protein with propensity 0.1, such that the total propensity of
the produce reaction is 0.3. (2) A protein P(b) is produced, which can block gene G(b, c)
with propensity 1. (3) The gene G(b, c) becomes blocked, preventing the production
of protein P(c). But genes G(c, a) and G(a, b) can still each produce with propensity
0.1 (4) A protein P(a) is produced, which can block gene G(a, b) with propensity 1.
(5) Now both G(a, b) and G(b, c) are blocked, with a very low propensity to unblock,
leaving G(c, a) to produce freely. (...) Eventually an equilibrium is reached between
production and degradation of protein P(a), where R(produce) = R(degrade) = 0.1.
This represents the first protein cycle of the simulation.
The propensities demonstrate how the first protein cycle is produced by a sequence
of high-probability reactions. There are three possible delay reactions
in the system: produce, degrade and unblock. The propensity of the produce
reaction is given by the sum of the production propensities of each of the genes.
The production propensity of gene G(a, b) is given by the population of the gene
times the rate of production, written [G(a, b)]·produce, and similarly for genes
G(b, c) and G(c, a). Thus, the propensity of the produce reaction is given by
[G(a, b)]·produce + [G(b, c)]·produce + [G(c, a)]·produce. The choice of a particular
gene to produce happens in two stages: first, the produce reaction is chosen
and second, one of the active genes is chosen to produce with equal probability.
The propensities of the unblock and degrade reactions are defined in a similar
fashion. There are three possible interactions in the system, on channels a,b
and c. The propensity of an interaction on channel a is given by the number of
proteins P(a) times the number of genes G(a, b), times the rate of the channel,
written [P(a)]·[G(a, b)]·ρ(a). This essentially determines the propensity of gene
G(a, b) to be blocked by protein P(a). The greater the population of repressor
18
Fig. 16. A sequence chart of interactions between genes and proteins for the execution
steps of Fig. 14. (1) Initially there is one copy of each of the genes G(a, b), G(b, c),
G(c, a). Gene G(a, b) produces a protein P(b) in parallel with itself. After the production,
the populations of G(a, b) and P(b) are both 1. (2) Protein P(b) sends on
channel b and interacts with gene G(b, c), which becomes blocked to B(b, c). After the
interaction, the populations of P(b), G(b, c) and B(b, c) are 1, 0 and 1, respectively.
(3) Gene G(c, a) produces a protein P(a) in parallel with itself. (4) Protein P(a) sends
on channel a and interacts with G(a, b), which becomes blocked to B(a, b). (5) Subsequently,
G(c, a) produces multiple proteins P(a), some of which are degraded. In
addition, B(b, c) unblocks to G(b, c), which is immediately blocked again by P(b).
proteins, the greater the blocking propensity. The propensities of the interactions
on channels b and c are defined in a similar fashion.
When examining the simulation results of Fig. 12 or the cartoon strip of
Fig. 14, it is difficult to understand the causal chain of events leading up to the
first protein cycle. In order to clarify the interactions that take place, together
with the consequences of these interactions, Fig. 16 represents the execution
steps of Fig. 14 as a sequence chart, based on an approach presented in [3].
In general, a chart represents a sequence of reactions from a single simulation
run. Each vertical line in the chart represents a separate species, where time
proceeds downwards. The vertical line is labelled with the current population
of the species, which can evolve as the simulation progresses. Each horizontal
line from one species to another represents a parallel composition of species.
The line connects the populations of two or more species, and represents the
fact that these populations are all executing simultaneously. Arrows are used
to denote reactions that take place as the simulation progresses. Vertical arrows
19
Fig. 17. A 3D representation of the gene with negative control from Fig. 12.
represent delay reactions, while horizontal arrows represent interactions between
two species, where the arrow is labelled with the channel on which the interaction
takes place. The chart shows how the different genes and proteins interact with
each other to initiate the first protein cycle.
3.3 3D Visualisation
This section describes how the repressilator model of Fig. 12 can be simulated
in three dimensions, in order to visualise the interactions between genes and
proteins over time. The three-dimensional models are displayed using Network
3D2
. Fig. 17 shows a 3D representation of a gene with negative control, based
on the model in Fig. 12. The main difference with the 2D representation is that
each action is represented as a node instead of as an edge. As a result, there are
now two types of nodes in the model: action nodes and species nodes. Action
nodes are shown in green and are of fixed size, while species nodes are of variable
colour and size, where the volume of the node is proportional to the population
of the species. The node labels are hidden by default, but can be revealed by
clicking on individual nodes. There are also two types of edges: transition edges
and interaction edges. Transition edges denote a transition from one species
to another, and go from a species node to an action node, and then from an
action node to zero or more species nodes. For example, the a action has one
inbound edge from G(a, b) and one outbound edge to B(a, b), since the gene can
transition from an active state to a blocked state by receiving on a. The produce
action has one inbound edge from G(a, b) and two outbound edges to G(a, b)
and P(b), since the gene can produce a new protein in parallel with itself. The
degrade action has only one inbound edge from P(b) and no outbound edges,
since the protein disappears after the degrade action is executed. Interaction
edges denote an interaction between two species, and go from an action node
2
Network 3D is developed by Rich Williams of Microsoft Research.
20
Fig. 18. A 3D representation of the repressilator model of Fig. 12, where the behaviour
of each gene is represented as a separate subgraph.
that is sending on a given channel to a corresponding action node that is receiving
on the same channel. There are no interaction edges in Fig. 17, since there are
no corresponding send and receive actions on the same channel.
Fig. 18 shows a 3D representation of the repressilator model of Fig. 12, where
the definitions of G(a, b), G(b, c) and G(c, a) are expanded so that the behaviour
of each gene is represented as a separate subgraph, similar to Fig. 14. There
are three subgraphs in the figure, one for each gene, represented in orange,
blue and red, respectively. The species nodes at the bottom of each subgraph
represent active and blocked genes, which have a population of 0 or 1, while the
species nodes at the top represent proteins, which have variable populations. The
subgraphs are connected to each other via interaction edges, which go from the
send action of a protein in one subgraph to the receive action of a corresponding
gene in another subgraph. For example, there is an edge from the send action c
of the blue protein to the receive action c of the red gene, representing the fact
that the blue protein P(c) can switch off the red gene G(c, a).
Fig. 19 shows a sequence of 3D pictures from a simulation of the repressilator
model of Fig. 12. The population of a species is visualised by dynamically altering
the size of the corresponding species node. When the population is empty, the
species node is obscured. The interactions between species are visualised by
drawing an edge from a send action to a corresponding receive action on the
same channel. When there are no senders or receivers the interaction edge is
obscured. The 3D layout allows complex interactions to be visualised, while
avoiding edge crossings that would have been unavoidable in 2D. The software
can be used to rotate or zoom into the model, in order to focus on a subset of
interactions and species. The 3D simulation is complementary to the 2D plots
of species populations over time, since it allows the interactions between species
to be visualised as the simulation progresses.
21
1 2
3 4
5 6
Fig. 19. A sequence of 3D pictures from a simulation of the repressilator model of
Fig. 12. (1) Initially there is a large population of red proteins, which switch off the
orange gene. (2) The blue gene then switches on and starts producing proteins, which
switch off the red gene. (3) Since no more red proteins are produced, the population
of red proteins slowly decreases over time, until all of the red proteins are degraded.
(4) The orange gene then switches on and starts producing proteins, which switch off
the blue gene. (5) Since no more blue proteins are produced, the population of blue
proteins slowly decreases over time, until all of the blue proteins are degraded. (6) The
red gene then switches on and starts producing proteins, which switch off the orange
gene, completing the cycle.
3.4 Model Refinement
Once we understand how the repressilator model functions, we can further refine
the model in order to improve the regularity of oscillations, as discussed in [2].
22
A B
degrade = 0.001,unblock = 0.0001, r = 1 degrade = 0.0001,unblock = 0.00001, r = 1
C D
degrade = 0.0001,unblock = 0.0001, r = 1 degrade = 0.0001,unblock = 0.0001, r = 10
Fig. 20. Parameter variation for the repressilator model of Fig. 12, with produce = 0.1
and ρ(a) = ρ(b) = ρ(c) = r.
We can start by adjusting the rates of the model, as shown in Fig. 20. We
observe that each protein cycle is characterised by a dominant protein whose
population stabilises at an equilibrium between production and degradation,
given by produce
degrade . In Fig. 20A the average population for a given protein cycle
is about 100, but this fluctuates significantly due to stochastic noise. We can
limit the relative size of the fluctuations by decreasing the degradation rate to
0.0001, resulting in a dominant population of about 1000, as shown in Fig. 20B.
We observe that when one protein is dominant the other two proteins are absent
and their corresponding genes are blocked, where one of the blocked genes is
actively repressed. For example, when the blue protein is dominant both the
red and orange proteins are absent, where the red gene is actively repressed and
the orange gene is waiting to unblock. This is illustrated in step 3 of Fig. 19.
The duration of protein cycles is highly irregular, since it depends mainly on
the rate of unblocking of the unrepressed gene, which is relatively slow and is
characterised by an exponential distribution. We can reduce this variability by
increasing the rate of unblocking to unblock = 0.0001, as shown in Fig. 20C. In
this setting, the gene unblocks more rapidly, but is immediately blocked again
by any remaining repressors. The rapid unblocking means that once the last
repressor has been degraded, the gene can become active soon after. Since the
repressor degradation curve is fairly regular, we observe an increased regularity
in the oscillations. Unfortunately, since the repressed gene is repeatedly trying
to unblock, this also increases the likelihood that it will produce a protein before
23
G(lac, tet) | G(tet, lambda) | G(lambda, lac) | G(tet, gfp)
Fig. 21. Simulation results for the repressilator model of Fig. 12, with an additional
GFP reporter, where produce = 0.1, degrade = unblock = 0.0001 and r = 10.
all of its repressors are degraded, which can result in large fluctuations in protein
levels, as observed in Fig. 20C. We can compensate for this by increasing the
rate of gene repression to r = 10.0, as shown in Fig. 20D. In this setting, even
if there are only a few repressor proteins remaining, they will still have a high
probability of blocking the corresponding gene, meaning that a gene will only
become active once all of its repressors are degraded. This property, together
with the regular degradation curve of repressors, gives rise to protein cycles of
regular duration and amplitude.
If we were to implement a new version of the repressilator inside a bacterium,
we might try to use genes and proteins whose reaction rates correspond roughly
to those of Fig. 20D. The simulation results for such a repressilator are shown in
Fig. 21, with genes for tet, lambda and lac, together with a reporter gene for gfp.
Note that GFP is produced in absence of the tet protein, where the intermittent
expression of GFP corresponds to the bacterium repeatedly glowing on and off.
In addition to adjusting the rates of the model in Fig. 12A, we can also
refine the model to include transcription and translation steps explicitly, together
with repressor binding and unbinding, as shown in Fig. 22A. The gene G(a, b)
can transcribe, after which a new mRNA strand M(b) is produced in parallel.
Alternatively, the gene can receive and bind on +
a and become blocked by a
repressor, after which it can receive and unbind on −
a and release the repressor,
becoming active once more. The mRNA strand M(b) can translate and produce
a new protein P(b) in parallel. Alternatively, the strand can degrade. The protein
P(b) can send and bind on +
b and become bound to a gene, after which it
can send and unbind on −
b. Alternatively, the protein can degrade. Simulation
results for the detailed model are shown in Fig. 22B. Note that the protein levels
fluctuate much more widely than in Fig. 12B, since they amplify any small
fluctuations that occur in the levels of mRNA. Although the detailed model
is much more complex than the simplified model of Fig. 12, similar analysis
can still be applied in order to understand how the model functions. However,
further refinements to the model are required in order to reduce the fluctuations
in mRNA levels. One way of reducing fluctuations is to introduce some form
24
A. SPiM model B. Simulation results
G(a, b) = transcribe.(G(a, b)|M(b))
+ +
a.B(a, b)
B(a, b) = −
a.G(a, b)
M(b) = translate.(M(b) | P(b))
+ degrade .0
P(b) = +
b.P (b) + degrade.0
P (b) = −
b.P(b)
Simulating G(a, b) | G(b, c) | G(c, a), with
transcribe = translate = 0.1,
degrade = degrade = 0.001, ρ(+
a) = ρ(+
b) =
ρ(+
c) = 1 and ρ(−
a) = ρ(−
b) = ρ(−
c) = 0.5.
Fig. 22. Refining the gene with negative control of Fig. 12 to include transcription,
translation and repressor binding.
of cooperativity in the model, for example by allowing proteins to form dimers
or even tetramers before repressing a gene. We omit the extensions here, but
additional details can be found in [2]. The main point to note is that we can
continually refine our parameterised model of a single gene without changing the
structure of the overall gene network, which remains fixed at G(a, b) | G(b, c) |
G(c, a). This illustrates the modularity of our approach, where module definitions
can be continually refined without changing the main program.
25
A. Gate with Inducer B. D038 Logic Gate
Gi(a, b, i)
Gi(tet, tet, aTc) | Gi(tet, lac, IPTG)
| G(lac, lambda) | G(lambda, gfp)
Fig. 23. Modular construction of gene networks from two parameterised models of a
gene. (A) A parameterised gene whose proteins are inhibited by an inducer i. (B) A
gene network that behaves like a boolean logic gate. The proteins aTc and IPTG act
as boolean inputs, while GFP acts as a boolean output. Depending on the presence or
absence of inhibitors, the GFP levels are either high or low. When implemented in live
bacteria, the presence or absence of inducers determines whether or not the bacteria
will glow.
3.5 Genetic Logic Gates
A substantial extension to the original repressilator experiment is described in
[14], in which a collection of logic gates were implemented in live bacteria. Rather
than producing oscillations in GFP levels, the experiment aimed to construct
gene networks that were able to function as boolean logic gates. The extension
was achieved by introducing inducers aTc and IPTG to inhibit the tet and lac
proteins, respectively. The presence or absence of the inducers represented the
logical inputs to the network, while the presence or absence of GFP represented
its logical output. In total, 125 different networks of 3 promoter-gene units were
constructed, using a combination of 3 different genes and 5 different promoters.
For each of these networks, the level of GFP was measured in presence or absence
of aTc, IPTG, or both.
We can model these networks in the SPiM calculus by defining a parameterised
gene in which the produced proteins are switched off by an inducer, as
shown in Fig. 23A. The process Gi(a, b, i) represents a gene that is blocked by
receiving on a, and that produces proteins which send on b, where the produced
proteins are inhibited by receiving on i. We can use the gene Gi(a, b, i) together
with the gene G(a, b) of Fig. 12 to construct the networks described in [14]. One
of these networks is illustrated in Fig. 23B. The first gene in the network is inhibited
by tet and produces tet, the second gene is inhibited by tet and produces
lac, while the third gene is inhibited by lac and produces lambda. Finally, the
network is fixed so that the lambda gene inhibits gfp, meaning that the bacteria
glows in absence of lambda. The two inputs to the system are aTc and IPTG,
where aTc inhibits the tet proteins and IPTG inhibits the lac proteins.
Fig. 24A shows the definition of the parameterised gene with inducers. The
definition is similar to that of Fig. 12, except that the gene is parameterised by
26
A. SPiM model B. Simulation results
D038 D038|I(aT c)
D038|I(IP T G) D038|I(aT c)|I(IP T G)
Gi(a,b,i) = produce.(Gi(a,b,i)|Pi(b,i))
+ a.Bi(a, b, i)
Bi(a,b,i) = unblock.Gi(a, b, i)
Pi(b, i) = degrade.0
+ b.Pi(b, i) + i.PI
I(i) = i.I(i)
Simulating the D038 network of Fig. 23, with
produce = 0.1, degrade = 0.001, unblock =
0.25 and all channel rates set to 1. D038 is defined
as Gi(tet, tet, aTc) | Gi(tet, lac, IPTG) |
G(lac, lambda) | G(lambda, gfp).
Fig. 24. A SPiM model of a parameterised gene with negative control and protein
inhibition (A), with associated simulation results (B). The model is similar to that
of Fig. 12, except that the gene is parameterised by an inducer i, which inhibits the
produced proteins.
an inducer i, which inhibits the produced proteins. The protein Pi(b, i) can block
a gene by sending on b, or it can be inhibited by receiving on i. Alternatively,
the protein can degrade. The behaviour of an inducer I(i) is straightforward,
in that it can inhibit a protein by sending on i. Fig. 24B shows the simulation
results for the gene network of Fig. 23B, where the green line represents the
levels of GFP. Each plot represents a simulation of the network with different
combinations of inputs. We observe that GFP is high in presence of I(aTc) and
in absence of I (IPTG), which gives a logical characterisation of the network. The
outcome of the simulation is not immediately obvious by looking at the network
construction of Fig. 23, not least because of the presence of a head feedback loop,
where one of the genes is repressed by the protein it produces. A more detailed
explanation of the simulation results is given in [1]. The main point to note here
27
is that we can model a wide range of genetic logic gates using a small number
of simple genetic building blocks.
4 C. elegans Development
The biological examples presented so far have all been modelled at the level of
interacting genes and proteins. In this section we show how the SPiM calculus
can also be used to model systems at the level of interacting cells. We consider
an example from developmental biology, relating to the development of the egglaying
system of the C. elegans nematode. C. elegans has been used to study
a number of fundamental cellular processes, and many breakthroughs relating
to human physiology were first discovered in C. elegans. Examples include genetic
regulation of organ development and programmed cell death, together with
RNA interference - gene silencing by double-stranded RNA, both of which were
awarded the Nobel prize in Physiology or Medicine in 2002 and 2006, respectively.
More recently, the Nobel Prize in Chemistry was awarded for the discovery
and development of the green fluorescent protein, GFP, for which a significant
part of the research was carried out in C. elegans.
In this section we consider the phase of development in C. elegans when six
Vulval Precursor Cells (VPCs) decide to adopt a primary, secondary or tertiary
fate depending on their positions with respect to an Anchor Cell (AC). The
decision of a precursor cell to adopt a particular fate is an important part of the
development of the organism, which will ultimately determine whether or not
the adult worm can successfully lay eggs. More details about the system can be
found in [36,37,12] and a recent review is presented in [35]. Various models of
VPC differentiation have been published recently, such as [9,12,21] to cite a few,
in which the differentiation process has been highly simplified. Here we present
an even simpler model, which illustrate some of the main factors involved during
development. While many biological details are missing, our approach allows
further model refinements to be made so that relevant details can be included
incrementally.
Fig. 25 illustrates some of the main events that take place during normal
differentiation of VPC cells in C. elegans. Once the cells have adopted their
respective fates, they undergo further stages of proliferation and differentiation
to develop into the egg-laying system of the adult worm.
4.1 Individual Cell Models
We start by constructing a simplified model of an anchor cell, together with a
generic model of an undifferentiated precursor cell that can adopt one of three
fates: primary, secondary or tertiary. The cell models are given in Fig. 26, where
the graphical representation at the top is equivalent to the textual representation
at the bottom. The anchor cell AC can repeatedly send on ac, which represents
a persistent inductive signal. A precursor cell can have four possible states, representing
the undifferentiated cell V , the primary cell fate V 1 and the secondary
28
lateral signal
AC
P3.p P8.pP7.pP6.pP5.pP4.p
2°
1°
3°
P3.p P8.pP7.pP5.pP4.p
1°P3.p P8.pP4.p 2°
2° 1° 2°3°3°
inductive signal
1.
2.
3.
4.
Fig. 25. Some of the main events that take place during normal differentiation of VPC
cells in C. elegans. (1) Initially the six cells P3.p - P8.p are undifferentiated. The anchor
cell sends an inductive signal, which is strongest for the cell P6.p directly in front of it.
(2) The strong inductive signal causes cell P6.p to adopt a primary fate, after which
it sends a lateral signal to its immediate neighbours. (3) The lateral signal causes cells
P5.p and P7.p to adopt a secondary fate. (4) Eventually, the remaining cells adopt a
tertiary fate.
AC = ac.AC
V (x, l, r, k) = ac · k.V 1(x, l, r) + l.V 2(x) + r.V 2(x) + w.V 3(x)
V 1(x, l, r) = l.V 1(x, l, r) + r.V 1(x, l, r)
Fig. 26. A SPiM model of an Anchor Cell and a Vulval Precursor Cell, which are
instrumental in the development of the egg-laying system of the C. elegans nematode.
The undifferentiated precursor cell V can adopt one of three fates: a primary fate V 1,
a secondary fate V 2 or a tertiary fate V 3.
29
V
AC
VVVVV
s4 s8s7s6s5
ac·low ac·low ac·med ac·high ac·med ac·low
83 4 5 6 7
AC | V (3, s3, s4, low) | V (4, s4, s5, low) | V (5, s5, s6, med)
| V (6, s6, s7, high) | V (7, s7, s8, med) | V (8, s8, s9, low)
Fig. 27. System model of cell-cell interactions during VPC differentiation in C. elegans.
The model is obtained by placing an anchor cell in parallel with six different VPC cells,
using the generic model of a VPC cell described in Fig. 26.
and tertiary cell fates V 2 and V 3, respectively. The undifferentiated cell V is parameterised
by its position x inside the worm, which is represented by a number
between 3 and 8. Cells with successive numbers are considered to be adjacent to
each other, and the cell at position 6 is considered to be directly opposite the
anchor cell. The undifferentiated cell is also parameterised by a rate k, which
represents the strength of its interaction with the anchor cell. Cells with a low
k are considered far away and only interact weakly with the anchor cell, while
cells with a high k are considered nearby and interact strongly with the anchor
cell. For simplicity we assume only three levels for k: low, medium and high.
Each undifferentiated cell is also characterised by its left and right neighbours,
given by the channels l and r, which correspond to the lateral signals that a
cell can use to communicate with its neighbours. The intuition is that each cell
will interact with its left neighbour on channel l and with its right neighbour on
channel r. We assume that once the precursor cells have adopted a particular
fate, they no longer interact with the anchor cell. Initially, a precursor cell starts
out undifferentiated as V . One of three things can then happen:
1. It can receive a signal on ac at rate k from the anchor cell, and adopt a
primary fate V 1. The primary cell will then send a persistent signal to its
left and right neighbours on l and r, respectively.
2. It can receive a signal from its left or right neighbour on l or r, respectively,
and adopt a secondary fate V 2.
3. In absence of stimulus it can eventually adopt a tertiary fate V 3, at rate w.
4.2 System Model
A model of the full developmental system is shown in Fig. 27, which consists
of the anchor cell and the six VPC cells, where the graphical representation
30
at the top is equivalent to the textual representation at the bottom. Each of
the VPC cells is an instance of the generic model of a single cell described in
Fig. 26. The generic model is instantiated with different parameters depending
on the position of the cell, the strength of its interaction with the anchor cell,
and the neighbouring cells with which it can communicate. Cells are assumed to
communicate with their left and right neighbours on dedicated channels, where
the left channel of a given cell is equal to the right channel of its left neighbour.
For example, the cell V (6, s6, s7, high) at position 6 can communicate with its
left neighbour on channel s6 and with its right neighbour on channel s7. It can
also communicate with the anchor cell at a high rate, since it is assumed to be
in close proximity to the anchor cell. Note that channels s3 and s9 in Fig. 27
are unused, since the cells to the left of position 3 and to the right of position 8
are not represented. The model can be further extended to represent these cells
if necessary.
4.3 Simulation Results
In order to simulate the process of VPC differentiation, we assign approximate
rates to the generic VPC model of Fig. 26 and to the overall system model of
Fig. 27. The rates of low, medium and high inductive signalling are given by 0.0,
1.0 and 10.0, respectively, while the rate ls of lateral signalling is given by 1.0. We
assume that the rate of lateral signalling is the same for all precursor cells, with
the rates of channels s4, . . . , s8 being equal to ls. The rate w for a cell to adopt
a tertiary fate in the absence of external signals is given by 0.0001. Although
these rates are entirely arbitrary, they give an initial approximation for the
relative speeds of the various interactions that take place during development,
and can be adjusted accordingly as new insight is gained from simulation and
experimentation.
The simulation results of Fig. 28 show the total population of undifferentiated
cells V and of differentiated cells V 1, V 2, V 3 over time. Initially there are
6 undifferentiated cells, with one of the cells adopting a primary fate almost
immediately. Two of the cells then adopt secondary fates in quick succession,
and the remaining cells eventually adopt a tertiary fate. In order to determine
which cells adopt which fates, we plot the number of cells V, V 1, V 2, V 3 at each
of the 6 positions as shown in Fig. 29, where each circle represents a precursor
cell. The horizontal coordinate denotes the cell position, which ranges from 3 to
8. The colour denotes the cell fate, where an undifferentiated cell V is shown in
grey, and a cell that adopts a primary, secondary or tertiary fate is shown in red,
blue or yellow, respectively. The plots are generated automatically during the
simulation, by using the parameter x in processes V, V 1, V 2, V 3 as the horizontal
coordinate for each process. As a given cell adopts a particular fate, the process
V (x, l, r, k) of each cell transitions to either V 1(x, l, r), V 2(x) or V 3(x), and
the processes at each position x are plotted accordingly. In this way, the SPiM
simulation generates a movie of cell differentiation over time. We can take snapshots
of the simulation to separate out the individual steps, as shown in Fig. 29.
For example, the first frame plots processes V (3, s3, s4, low) | V (4, s4, s5, low) |
31
Fig. 28. Simulation results for the VPC model of Fig. 27. The results show the populations
of undifferentiated, primary, secondary and tertiary cells over time.
1 2 3
4 5 6
Fig. 29. Coordinate geometry plots for the simulation results of Fig. 28. The results
show the change in cell fate over time for cells at positions 3 to 8, where the horizontal
coordinate denotes the cell position. Initially all the cells start out as undifferentiated
(not shown). The cell at position 6 then adopts a primary fate, after which the cells at
positions 7 and 5 adopt a secondary fate. One by one the remaining cells then adopt a
tertiary fate.
V (5, s5, s6, med) | V 1(6, s6, s7) | V (7, s7, s8, med) | V (8, s8, s9, low), with process
V 1(6, s6, s7) at position 6 representing a primary fate and the remaining
processes representing undetermined fates. As the simulation progresses, the processes
interact with each other according to the behaviour outlined in Fig. 26,
with the processes V (7, s7, s8, med) and V (5, s5, s6, med) transitioning to V 2(7)
and V 2(5), respectively, followed by the remaining processes transitioning to
V 3(4), V 3(8) and V 3(3).
We can interpret the simulation results based on the model construction and
the chosen parameters. The precursor cell at position 6 adopts a primary fate
by interacting with the anchor cell at rate high = 10. The average duration of
this interaction is 0.1, which explains the rapid transition time observed in the
simulation. Next, the primary cell at position 6 can signal to its left and right
neighbours to adopt a secondary fate at rate ls = 1.0. The average duration of
these transitions is 1, which is 10 times slower than the initial inductive signal
to the cell at position 6. In absence of external signals, the remaining cells adopt
a tertiary fate at rate w = 0.0001, which explains the length of time taken for
these cells to differentiate.
32
Although the rates are chosen arbitrarily, we observe that if we increase the
rate w then the cells are able to adopt a tertiary fate more quickly. Since the
simulation is stochastic, this means that the cells at positions 5 to 7 will have a
greater chance of adopting a tertiary fate, even in the presence of inductive and
lateral signals, which could potentially result in abnormal development. Note
also that the cells at positions 5 and 7 receive an inductive signal at rate 1,
which is 10 times weaker than the inductive signal to the cell at position 6.
Although the inductive signal is much weaker, there is still a one in ten chance
for a cell at position 5 or 7 to adopt a primary fate, resulting in abnormal
development. Indeed, if we simulate the model several times we observe that cells
5 or 7 adopt the wrong fate roughly 10 percent of the time. Clearly the VPC
model is not satisfactory, since in reality the process of VPC differentiation is
extremely robust. Simply changing the ratio of the high to med inductive signals
to something like 1000 : 1 in order to improve robustness is not acceptable either,
since in reality the strength of the inductive signal does not differ by such a
large amount between neighbouring cells. Thus, although the high-level model
we have presented can mechanistically reproduce some of the stages of VPC
differentiation, it cannot reproduce the inherent robustness of the process, and
more refined models are needed.
4.4 Refined Model
To illustrate our modelling approach, we describe a refined version of the VPC
model of Fig. 26, based on the early results of [37]. The refined model is presented
in Fig. 30. One of the simplifying assumptions of the initial model was to
represent a change in cell fate by the activation of a single process inside the cell,
representative for example of the activation of a single gene. In the refined model
we now represent a change in cell fate by the accumulation of a particular protein,
where V 1, V 2, V 3 represent proteins with primary, secondary and tertiary
functions, respectively. As with the simplified model, an undifferentiated cell V
is parameterised by its position x, the strength k of its interaction with the anchor
cell, and the channels l and r on which it can communicate with its left and
right neighbours. As a result, the overall model of the VPC system remains the
same as in Fig. 27, while the behaviour of an undifferentiated cell V (x, l, r, k)
is substantially refined. Instead of consisting of a single process that changes
state over time, the cell V now consists of multiple processes running in parallel.
The processes gV 1, gV 2, gV 3 represent the genes for proteins V 1, V 2, V 3, while
the processes gV ul and gLin12 represent the genes for proteins V ul and Lin12.
The genes are executed in parallel with each other inside the cell, alongside the
Muv proteins. The gene gV ul produces proteins in the presence of a persistent
inductive signal on channel ac, where the rate of production is determined by the
parameter k. The V ul proteins then initiate a persistent communication with
neighbouring cells by sending on l and r. The V ul proteins can also induce the
production of proteins V 1 by sending on vul. The proteins V 1 are specific to
cells that adopt a primary fate and can persistently send on v1, which causes the
Lin12 receptors to degrade. This prevents the cell from responding to its own
33
V (x, l, r, k) = (vul, v1, v3, v3, lin12, muv)
( gV ul(x, l, r, vul, muv, k) | gLin12(x, lin12, v1, v3, l, r)
| gV 1(x, v1, v2, vul) | gV 2(x, v2, lin12) | gV 3(x, v3, vul) | Muv(x, muv))
gV ul(x, l, r, vul, muv, k) = ac · k.(gV ul(x, l, r, vul, muv, k) | V ul(x, l, r, vul, muv))
gLin12(x, lin12, v1, v3, l, r) = l.(gLin12(x, lin12, v1, v3, l, r) | Lin12(x, lin12, v1, v3)
+ r.(gLin12(x, lin12, v1, v3, l, r) | Lin12(x, lin12, v1, v3)
gV 1(x, v1, v2, vul) = vul.(gV 1(x, v1, v2, vul) | V 1(x, v1, v2))
gV 2(x, v2, lin12) = lin12.(gV 2(x, v2, lin12) | V 2(x, v2))
gV 3(x, v3, vul) = w.(gV 3(x, v3, vul) | V 3(x, v3, vul))
V ul(x, l, r, vul, muv) = l.V ul(x, l, r, vul, muv) + r.V ul(x, l, r, vul, muv) + vul + muv
Lin12(x, lin12, v1, v3) = lin12.Lin12(x, lin12, v1, v3) + v3 + v1
Muv(x, muv) = muv.Muv(x, muv)
V 1(x, v1, v2) = v1.V 1(x, v1, v2) + v2 + dV 1
V 2(x, v2) = v2.V 2(x, v2) + dV 2
V 3(x, v3, vul) = v3.V 3(x, v3, vul) + vul + dV 3
Fig. 30. A refined SPiM model of the Vulval Precursor Cell of Fig. 26. The proteins
V 1, V 2, V 3 represent primary, secondary and tertiary functions which correspond to
particular cell fates.
lateral signals on l and r, thereby preventing it from adopting a secondary fate.
34
Fig. 31. Simulation results for the VPC model of Fig. 30, with low = 0.01, med = 2,
high = 10, v1 = 10, v2 = 0.1, v3 = 1, dV 1 = 0.1, dV 2 = dV 3 = 0.05 and all remaining
rates set to 1. The results indicate the populations of processes V 1, V 2 and V 3, which
represent primary, secondary and tertiary fates, respectively, for each of the 6 cells.
Both plots represent different views of the same simulation. The left plot represents a
3D view of the 18 different protein concentrations over time, 3 for each cell. The right
plot gives another view of the same results, which displays the maximum populations
for each of the proteins. Each cell has a dominant population of a particular protein,
which corresponds to a particular cell fate.
Conversely, the gene gLin12 produces active receptor proteins when it receives
a signal on l or r from its left or right neighbours. The Lin12 receptors send
on lin12 and cause the production of proteins V 2, which are specific to cells
that adopt a secondary fate. The proteins V 2 also cause the degradation of proteins
V 1 by sending on v2, thereby preventing the cell from adopting a primary
fate. Finally, the gene gV 3 is constitutively expressed at rate w and produces
proteins V 3 that are specific to cells with a tertiary fate. These proteins are
in turn inhibited by receiving on vul. Thus, there is a complicated network of
interactions to ensure that the cells reliably adopt specific fates depending on
the signals they receive. The interactions inside a particular cell take place on
channels vul, v1, v2, v3, lin12 and muv. In order to ensure that the signals inside
a given cell do not interfere with those inside neighbouring cells, the channels
are restricted to each cell, represented by placing them in brackets inside the
cell definition. It is this notion of restricted channels that enables boundaries
between cells to be established in the model.
Initial simulation results for the VPC model of Fig. 30 are presented in
Fig. 31. The results show the populations of processes V 1, V 2 and V 3 for each
of the 6 cells. Since these processes are parameterised by their location x, there
are 18 different processes altogether. The first three lanes represent populations
of V 1, V 2, V 3 for the cell at position 3, the next three lanes represent populations
of proteins for the cell at position 4, etc. We observe that the cell at position 6
adopts a primary fate, since it has a predominance of proteins V 1, while the cells
at positions 5 and 7 adopt a secondary fate, with a predominance of V 2 proteins.
35
Finally, the cells at positions 3,4 and 8 adopt a tertiary fate, with a predominance
of V 3 proteins. Rather than making a decision after receiving a single signal from
a neighbouring cell, repeated signalling is required for a cell to adopt a particular
fate. This improves the robustness of the system, since the decision to adopt a
particular fate is now made over a period of time. An important mechanism
is the amplification of a relatively small difference in inductive signal between
cells, so that each cell can reliably adopt the correct fate. Some of these issues are
discussed in more detail in [9]. The main point we wish to make here is that we
can start with a highly simplified model of a cell, and instantiate multiple copies
of this model with different parameters to represent a system of communicating
cells. We can then refine the model of a cell without changing the overall system
definition. Additional details can be included in subsequent refinements, based
on published results in [35,9,12,21]. This highlights the main motivation for our
approach, which is to construct complex systems from simpler building blocks,
and to be able to refine these building blocks without having to reconstruct the
entire system.
5 SPiM with Compartments
The previous section described how the SPiM calculus can be used to model
interactions between cells with fixed locations. In this section we describe how
the calculus can be extended with a notion of mobile compartments, in order to
model interactions between cells that can move relative to each other. The ambient
calculus [5] by Cardelli and Gordon originally introduced a notion of mobile
compartments to model computation carried out in mobile devices, together with
mobile code that can move between devices. Since the original calculus, a range
of variants have been introduced to model various aspects of security and mobility
of distributed computer systems. More recently, the bioambient calculus
[31] was introduced by Regev and colleagues to model molecular localisation and
compartmentalisation in biological systems. In this section we present an extension
of the SPiM calculus with compartments, which is essentially a graphical
variant of the bioambient calculus.
5.1 Basic Primitives
We introduce the basic primitives of the SPiM calculus with compartments by
means of a simple example. The example models some of the interactions that
can take place between cells of the human immune system, and represents a tiny
fraction of the processes taking place during an immune response.
Dendritic cells are part of the adaptive immune system, one of the main defence
mechanisms of the human body against viral infections. Dendritic cells act
as sentinels, patrolling areas of the body that can be exposed to infection, such
as the lungs and the tissue underneath the skin or surrounding the intestines. If a
dendritic cell becomes infected by a virus such as measles, the virus hijacks some
of the internal machinery of the cell to produce more copies of itself, consuming
36
some of the resources of the cell in the process. When infected with a virus,
a dendritic cell can detect the infection by means of internal receptors, which
cause the cell to become activated. Once activated, the dendritic call can travel
to a nearby lymph node, where it presents fragments of some of the proteins
being made by the virus. The fragments are presented by means of MHC class
I molecules, which capture the fragments from inside the cell and present them
at the cell surface. The MHC class I molecules provide an external signal that
the cell is infected, by presenting specific parts of the viral proteins as evidence
of the infection. Once inside a lymph node, the dendritic cell with MHC class I
molecules on its surface will encounter a large number of naive CD8 T cells, some
of which will recognise the viral protein fragments being presented and become
activated as a result. Only a very small proportion of the total population of CD8
T cells will be able to recognise a particular protein fragment from a particular
virus. This is because the T cell repertoire needs to be broad enough to recognise
potentially billions of viral proteins. This broad repertoire is achieved by a
process of genetic recombination, involving a random mix-and-match of different
genetic sequences during T cell development. Thus, each T cell develops receptors
that can only recognise specific protein fragments. When the receptors of a
given CD8 T cell recognise a particular protein fragment being presented on the
surface of an infected dendritic cell, the T cell can become activated, provided
it also receives the necessary co-stimulatory signals. The activated CD8 T cell
then proliferates to make more copies of itself, creating an army of such cells to
effectively handle the viral infection. The activated T cells leave the lymph node,
travelling through the blood stream until they reach the site of infection. During
their journey, they are guided by specific chemokine signalling molecules, while
specific addressin and selectin molecules allow them to identify and enter the site
of infection. On arrival, the activated CD8 T cells are able to use their receptors
to target the infected cells. This is because the infected cells are also presenting
protein fragments from the virus with which they are infected, by means of MHC
class I molecules. The T cell army can recognise the specific protein fragments
being presented by the infected cells, and can target those cells for destruction
in order to help contain the viral infection.
Here we have described only a tiny fraction of the events taking place during
a CD8 T cell response. More details are available in standard text books such as
[34,20], some of which are searchable online3
. A recent review of the presentation
of protein fragments by MHC class I molecules can also by found in [38].
Fig. 32 presents a SPiM model of some of the basic processes that can take
place during an immune response to a viral infection. The graphical representation
at the top is equivalent to the textual representation at the bottom. The
example illustrates the main primitives of the SPiM calculus with compartments.
We model the site of infection as a particular compartment, which represents
a region of tissue cells belonging to an infected organ (such as the liver or a
lung). The compartment is drawn as a box around a collection of processes, all
of which are executing in parallel inside the compartment. We also model a cell
3
http://www.ncbi.nlm.nih.gov/
37
1 1
1 1 1
outTissue tissue Lymph_Node in lymph
Dendritic_Cell CD8Naive
Activated
s2s activate
Presenting s2s
out
Migrating
in lymph
tissue
activate
Infected
1
Tissue = out tissue.Tissue Lymph Node = in lymph.Lymph Node
Infected = out tissue.Migrating Naive = s2s activate.Activated
Migrating = in lymph.Presenting
Presenting = s2s activate.Presenting
Tissue | Dendritic Cell | Infected | Lymph Node | CD8 | Naive
Fig. 32. A SPiM model of some of the basic processes that can take place during an
immune response to a viral infection.
as a compartment, with its own internal processes. In this case, the compartment
is a means of separating the computation taking place inside the cell boundary
from the computation taking place outside the cell. When a dendritic cell becomes
infected, it can become activated and migrate from the site of infection
towards a secondary lymphoid organ, such as a nearby lymph node. In the SPiM
calculus, the movement is represented by the out tissue action. The out keyword
denotes the direction of movement, while the send on tissue denotes the
channel over which the movement takes place. In parallel, the tissue contains a
persistent out tissue action, which allows cells to leave the tissue. Thus, the send
out tissue is executed inside the compartment that is leaving, while the receive
out tissue is executed inside the compartment that is allowing the leave to take
place. The sender and receiver can then interact on channel tissue, after which
the sender compartment moves outside the receiver compartment. Note that the
entire compartment moves together with all of its contents after the interaction
takes place. The dendritic cell is then in a Migrating state and can enter a lymph
node by the in lymph action. The in keyword denotes the direction of movement,
while the send on lymph denotes the channel over which the movement
38
takes place. In parallel, the lymph node contains a persistent in lymph action,
which allows cells to enter the lymph node. The send in lymph is executed by
the compartment that is entering, while the receive in lymph is executed by the
compartment that is allowing the enter to take place. Once the dendritic cell has
entered the lymph node, it can present the MHC class I molecules on its surface
to nearby T cells. This is represented by a persistent send on s2s activate. The
s2s keyword stands for sibling to sibling, and denotes a communication from
one sibling compartment to another.
In general, the lymph nodes are meeting places where dendritic cells can
present protein fragments of the viruses with which they are infected, and can
recruit T cells to help fight the infection. Here we represent a CD8 T cell as a
compartment, containing the CD8 process in parallel with the Naive process.
The CD8 process can represent a range of cellular processes which are necessary
for the survival and normal functioning of the CD8 T cell. In our model, the
Naive process represents the current state of the CD8 T cell. The cell can
transition from Naive to Activated by receiving on s2s activate from a nearby
cell. Here we represent the activation step as a single communication from the
dendritic cell to the naive CD8 T cell on the activate channel. In reality, of
course, the mechanism is much more complicated. In particular, the activation
will only be successfully if the MHC class I molecules and their contents on the
surface of the dendritic cell match the receptors on the surface of the T cell.
We can represent this match in an abstract way, by requiring the sender and
receiver cells to communicate on the same activate channel. Once the T cell
becomes activated, it can then leave the lymph node and circulate in the blood
until it reaches the infected tissue, though these later events are not represented
in Fig. 32.
Fig. 33 illustrates the sequence of events that take place when executing the
model of Fig. 32. For each frame in the figure, the graphical representation at
the top is equivalent to the textual representation at the bottom. The number
of copies of a given process is indicated next to the corresponding node in the
figure. By default we assume that there is a single copy of each compartment.
The example illustrates the in and out movement primitives, which allow compartments
to move in and out of each other, respectively. When a compartment
moves it travels with its entire contents, and when it reaches its new location it
can interact with additional compartments that may not have been present in
its previous location. For example, in the case of the dendritic cell moving inside
a lymph node, on arrival it can interact with Naive CD8 T cells that were not
accessible in its previous location. The interaction between two compartments
is achieved using the s2s communication primitive, which allows two sibling
compartments to communicate with each other. There are also primitives for
allowing communication from a parent to a child (p2c ) and from a child to
its parent (c2p ), respectively, together with a merge primitive for merging two
compartments. These primitives, together with the full definition of the SPiM
calculus with compartments, are given in Appendix 7.5.
39
1 2
outTissue tissue Lymph_Node in lymph
Dendritic_Cell CD8Naive
Activated
s2s activateout
in lymph
tissue
Infected
1 1
1 1 1 1
Presenting s2s activate
Migrating
outTissue tissue Lymph_Node in lymph
Dendritic_Cell CD8Naive
Activated
s2s activateout
in lymph
tissue
Infected
1 1
1
1 1 1
Presenting s2s activate
Migrating
Tissue | Dendritic Cell | Infected |
Lymph Node | CD8 | Naive
Tissue | Dendritic Cell | Migrating |
Lymph Node | CD8 | Naive
3 4
outTissue tissue
Lymph_Node in lymph
Dendritic_Cell CD8Naive
Activated
s2s activateout
in lymph
tissue
Infected
1
1
1
1 1 1
Presenting s2s activate
Migrating
outTissue tissue
Lymph_Node in lymph
Dendritic_Cell CD8Naive
Activated
s2s activateout
in lymph
tissue
Infected
1
1
1
1
1
1
Presenting s2s activate
Migrating
Tissue |
Lymph Node | CD8 | Naive |
Dendritic Cell | Presenting
Tissue |
Lymph Node | CD8 | Activated |
Dendritic Cell | Presenting
Fig. 33. The sequence of events that take place when executing the model of Fig. 32.
(1) The infected dendritic cell can leave the tissue by sending on out tissue. (2) The
migrating dendritic cell has left the tissue, and can now enter the lymph node by
sending on in lymph. (3) The presenting dendritic cell can activate a nearby Naive
CD8 T cell by sending on s2s activate. (4) The CD8 T cell is now activated.
40
5.2 Simple Model of Viral Infection and T Cell Response
Fig. 34 extends the simple model of Fig. 32 to include activation and proliferation
of T cells. The model of a dendritic cell is extended to include an initial
Clean process, which can receive on s2s infect from a nearby cell and become
Infected. We also allow the Presenting process to die, which represents the
fact that the dendritic cell will not continue presenting indefinitely. We start
with a nominal population of 10 dendritic cells, where the initial population is
indicated next to the compartment boundary. We provide a basic model for the
other cells in the tissue, which start out in an uninfected state and can receive
on s2s infected, becoming infected as a result. Once infected, the cell can then
infect other neighbouring cells by repeatedly sending on s2s infect. In this way,
a virus inside a single cell can infect a potentially infinite number of neighbouring
cells. The infected cell can also be killed by receiving on s2s kill. We start with
a nominal population of 1000 uninfected cells, together with a single infected
cell. We also extend the model of the tissue by allowing multiple cells to enter
or leave, represented by persistent receive actions on in tissue and out tissue,
respectively. Similarly, we allow multiple cells to enter or leave the lymph node,
represented by persistent receive actions on in lymph and out lymph, respectively.
Finally, we extend the model of a CD8 T cell to represent the processes
that take place after activation, namely T cell proliferation and migration to
the site of infection, followed by killing of the infected cells. Once activated, the
CD8 T cell can either proliferate or die, which ensures that it does not continue
to proliferate indefinitely. After the proliferate action is executed, a new cell is
created in parallel with the Parent process. The creation of a new cell is represented
by a triangle node, where the successor nodes to the triangle represent
the contents of the newly created cell. In this example, the new cell executes the
Child process, which sends on out divide. This is a simple way of representing
cell division, where the newly created cell leaves it parent and then executes
alongside the parent. The Parent process allows the newly created cell to leave
by receiving on out divide. Once the child cell has left its parent, it executes the
Activated and CD8 processes, representing a newly created CD8 T cell in an
activated state. Meanwhile, the parent cell can leave the lymph node by sending
on out lymph and then enter the tissue by sending on in tissue. On arrival, it
can kill neighbouring infected cells by repeatedly sending on s2s kill, until eventually
it dies. Note that the newly created T cell can perform the same actions
as the activated parent. In this way, a single activated T cell can produce a large
population of cells, in order to effectively fight a viral infection.
Of course, this model is still an extreme simplification of the events that
actually take place during T cell activation. In particular, the activated CD8 T
cell will only be able to kill infected cells that it recognises as being infected.
As with T cell activation, this will depend on whether or not the T cell receptors
recognise the viral protein fragments being presented on the surface of the
infected cell by MHC class I molecules. We can model this complementarity between
the presenter and the receiver by requiring the T cells and the infected
cells to interact on the same kill channel. Thus, different signalling and receptor
41
die
proliferate
outTissue tissue Lymph_Node out lymph
Dendritic_Cell CD8Naive
Activated
s2s activate
Presenting s2s
out
Migrating
in lymph
tissue
activate
Infected
in tissue in lymph
Infected_Cell s2s
Cell
infect
s2s infect
s2s kill
Clean
s2s infect
Leaving
out divide
out divide
Killer s2s
out
Entering
in
lymph
tissue
kill
die
Parent
Child
1 1
1 11 1
1
1000
10
die
Infected_Cell s2s infect
s2s kill 1
T issue = in tissue.T issue Lymph Node = in lymph.Lymph Node
+ out tissue.T issue + out lymph.Lymph Node
Clean = s2s infect.Infected Naive = s2s activate.Activated
Infected = out tissue.Migrating Activated = proliferate.(P arent | Child )
Migrating = in lymph.P resenting + die
P resenting = s2s activate.P resenting Child = out divide.(Activated | CD8)
+ die P arent = out divide.Leaving
Cell = s2s infect.Infected Cell Leaving = out lymph.Entering
Infected Cell = s2s infect.Infected Cell Entering = in tissue.Killer
+ s2s kill Killer = s2s kill.Killer + die
T issue | 10 · Dendritic Cell | Clean |
1000 · Cell | Infected Cell
| Lymph Node | CD8 | Naive
Fig. 34. A simple model of viral infection and immune response.
42
Fig. 35. Simulation results for the viral infection model of Fig. 34. The results show the
total populations of uninfected and infected tissue cells, together with the populations
of presenting dendritic cells and killer T cells over time. We assume that proliferate =
0.5, divide = 0.01, activate = 0.2, killer = 0.005, infect = 0.00001 and the remaining
rates are equal to 1. We start with a single infected tissue cell among a population of
1000 uninfected cells and 10 dendritic cells. When a dendritic cell becomes infected,
it moves to a lymph node where it activates a corresponding CD8 T cell. The T cell
proliferates and the newly produced T cells move to the site of infection, where they
kill the infected cells to contain the spread of the virus. After the infection is cleared,
the T cells eventually die off.
molecules can be represented by sending and receiving on different channels. In
this simplified model we have also represented proliferation by only allowing a
given cell to divide once before leaving the lymph node, when in reality the same
parent cell can divide multiple times. More generally, the lymph node also has a
much more complicated structure, and the behaviours of the individual cells will
probably require millions of lines of code in order to be accurately represented.
Nevertheless, we can continually refine our models to include additional detail
as needed, and we can use our highly simplified models to study the dynamic
interactions between cells at an abstract level.
Fig. 35 shows a simulation of the model of Fig. 34. We start with 1000 tissue
cells and a single infected tissue cell. In addition, 10 dendritic cells are assumed
to be patrolling this particular area of tissue. When one of the dendritic cells
becomes infected, it moves to a nearby lymph node, where it chances upon a
Naive CD8 T cell that recognises the viral protein fragments being presented
43
on its surface and becomes activated as a result. The CD8 T cell proliferates
and moves to the site of infection. Meanwhile, the infection is spreading between
neighbouring cells. We can see this by the decrease in population of uninfected
cells, shown in green, and the increase in population of infected cells, shown in
red. As the killer T cells arrive, they destroy the infected cells as a means of
preventing the spread of the virus. We observe a sudden drop in the population
of infected cells, which coincides with a rise in the population of killer T cells.
Once the infection has been cleared, there are no more infected cells to stimulate
an immune response, and the killer T cells die off. We can try different parameter
values for the simulation to observe under what conditions the infection is
contained and how many of the cells become infected. We can also observe the
trade-offs between rapid proliferation with a large population of killer T cells,
and slower proliferation with less T cells. In reality, the immune system is often
trying to strike a balance between responding too strongly and using unnecessary
resources, or not responding strongly enough and being unable to clear the
infection. Thus, the T cells need to proliferate at just the right rate depending
on the level of presentation by the dendritic cells, and to quickly die off once the
infection has been contained.
The are a number of simple ways in which the model can be extended. For
instance, rather than having a fixed population of tissue cells we can include a
steady-state population, where cells are continually dying and proliferating. We
can also include many other types of cells such as helper T cells, which send
activation signals to regulate the overall immune response. Furthermore, when
a dendritic cell becomes infected by a virus such as measles, the dendritic cells
can also cause the virus to spread to other cells in the lymph node, leading to a
sophisticated battle between immune system and virus.
6 Conclusion
In this chapter we have presented a visual process calculus for designing and
simulating computer models of biological systems. We have started with simple
models of interacting genes and proteins, followed by more complex models of
gene networks. We have also presented simple models of developmental processes
by constructing a generic model of a particular cell type, and then instantiating
this model to represent a number of identical cells in different locations with
different environmental signals. Finally, we have used the calculus to model cells
that can move, proliferate and interact with each other in different physical and
logical locations. The calculus we have presented builds on previous work of
[29,28,32,31] among many others, but there is still scope for many extensions,
for example to handle spatiotemporal dynamics or more realistic biological interactions.
One example is the use of primitives for biological membranes, as
described in [4], which more accurately represent membrane formation and interaction.
More generally, there are a wide variety of process calculi being developed
for biology, such as [27,10,8,7,23] to name but a few. In future, these and other
process calculi could help to form the theoretical foundations for novel program-
44
ming languages for biology. Such languages will enable us to take modelling,
simulation and analysis of biological systems to a stage where we can handle
models consisting of millions of lines of code in a scalable, modular fashion, just
as we have achieved for more traditional software systems.
Acknowledgements Thanks to Hillel Kugler for discussions and helpful comments
on C. elegans modelling and to Rosie Bloxsom for contributing a stochastic VPC
model based on [37]. Thanks also to Alistair Bailey, Tim Elliott and Sandra
Phillips for helpful feedback.
45
π ::= r Delay at rate r
| x( ˜m)r
Receive names ˜m on channel x at rate r
| x ˜n r
Send names ˜n on channel x at rate r
| x( ˜m)r
Send restricted names ˜m on channel x at rate r
M ::= π1.P1 + . . . + πN .PN Choice between actions, N ≥ 0
P ::= 0 Null
| X(˜n) Species X with parameters ˜n
| P1 | . . . | PN Parallel composition of processes, N ≥ 2
| (x1) . . . (xN ) P Restricted channels x1, . . . , xN in P
D ::= P Definition of a process
| M Definition of a choice
E ::= X1( ˜m1)=D1, . . . , XN ( ˜mN )=DN Definition of species Xi with parameters ˜mi
S ::= E P System with environment E and process P
Definition 1. Syntax of SPiM.
7 The SPiM calculus
This appendix presents the technical definition of the SPiM calculus, a variant
of stochastic pi-calculus for biological modelling.
7.1 Syntax
The syntax of the calculus is presented in Definition 1, where r denotes a rate
belonging to the set of real numbers, x denotes a channel, X denotes a species,
˜n denotes a list of names n1, . . . , nN , and ˜m denotes a list of names that are
pairwise distinct. Each channel x is also associated with a corresponding rate
given by ρ(x). Rates are used to characterise the speed of a reaction according
to an exponential distribution, such that the probability of a reaction with rate
r occurring within time t is given by F(t) = 1 − e−r·t
. The average duration of
the reaction is given by the mean 1/r of this distribution.
An action π can be a delay at rate r, a receive x( ˜m)r
of names ˜m on channel
x at rate r, a send x ˜n r
of names ˜n on channel x at rate r, or a send x( ˜m)r
of
restricted names ˜m on channel x at rate r. A choice M denotes a competition
between zero or more actions of the form π.P, where process P is executed after
performing action π. A process P can be a empty 0, a species X(˜n) with name X
and parameters ˜n, a parallel composition of two or more processes P1 | . . . | PN ,
or a process (x1) . . . (xN ) P with restricted channels x1, . . . , xN . The restricted
46
channels are used to represent the formation of a complex. For example, the
process (x) (X1(x) | X2(x)) represents a complex of two species, X1 and X2,
bound to each other on channel x. The restriction denotes a local channel x on
which the two species can interact in order to split the complex.
An environment E consists of a set of species definitions of the form X( ˜m)=
D, where X is the name of the species, ˜m are its parameters and D is its
corresponding definition, which can be a process P or a choice M. It is assumed
that fn(D) ⊆ ˜m, where fn(D) denotes the set of free names of D, given that
(x) P binds the name x in P, and x( ˜m)r
.P and x( ˜m)r
.P bind the set of names
˜m in P. Definitions D are also assumed to be equal up to renaming of bound
names. In addition, for each definition X( ˜m) = D, any calls to X inside D can
only occur after an action π. This prevents potentially infinite processes such as
X() = (X() | X()). Finally, a system S consists of a global constant environment
E, together with a process P.
We define a number of syntactic conventions: we assume that the parallel
composition operator (|) has the lowest precedence among the operators of the
calculus; we abbreviate an action π.0 followed by the empty process to π, and a
sequence of restricted channels (x1) . . . (xN ) P to (x1, . . . , xN ) P; we abbreviate
a parallel composition of N identical processes P | . . . | P to N ·P. We also
allow some or all of the parameters of a species to be hidden in cases where
these parameters are unchanged throughout the system. We can recover the
parameters by observing the free names of the species definition. In order to
model the binding and unbinding of species we assume that each channel x is
associated with a unique unbinding channel −
x and vice-versa, where x and −
x
represent two distinct names. We write +
x as an abbreviation for x(−
x) and +
x
as an abbreviation for x(−
x). We illustrate these abbreviations with the following
three identical systems, which model the binding and unbinding of proteins X
and Y , as described in Fig. 9:
X = +
b.X , Y = +
b.Y ,
X = −
b.X, Y = −
b.Y
X | Y (1)
X = b(−
b).X , Y = b(−
b).Y ,
X = −
b.X, Y = −
b.Y
X | Y (2)
X(b) = b(−
b).X (b, −
b), Y (b) = b(−
b).Y (b, −
b),
X (b, −
b) = −
b.X(b), Y (b, −
b) = −
b.Y (b)
X(b) | Y (b) (3)
The first system models the binding and unbinding of proteins X and Y on
channel b, using the syntactic abbreviations for binding. The second system expands
the binding abbreviations, while the third system explicitly shows all of
the parameters of the different proteins. The abbreviations allow a more compact
representation of systems, without loss of information. Note that here we
47
X1(˜n1) = x ˜n r1
.P1 + M1
X2(˜n2) = x( ˜m)r2
.P2 + M2
X1(˜n1)|X2(˜n2)
ρ(x)·r1·r2
−→ P1 |P2{˜n/ ˜m}
(4)
X1(˜n1) = x( ˜m)r1
.P1 + M1
X2(˜n2) = x( ˜m)r2
.P2 + M2
X1(˜n1)|X2(˜n2)
ρ(x)·r1·r2
−→ ( ˜m)(P1 |P2)
(5)
X(˜n) = r.P + M
X(˜n)
r
−→ P
(6)
P
r
−→ P
(x) P
r
−→ (x) P
(7)
P
r
−→ P
P | Q
r
−→ P | Q
(8)
Q ≡ P
r
−→ P ≡ Q
Q
r
−→ Q
(9)
Definition 2. Reaction rules in SPiM. The condition X(˜n) = D is true if there is a
definition in the system environment such that X( ˜m) = D and D = D{˜n/ ˜m}.
P | 0 ≡ P (10)
P1 | P2 ≡ P2 | P1 (11)
P1 |(P2 | P3) ≡ (P1 | P2)|P3 (12)
π1.P1 + π2.P2 ≡ π2.P2 + π1.P1 (13)
(x) 0 ≡ 0 (14)
(x) (y) P ≡ (y) (x) P (15)
(x) (P1|P2) ≡ P1|(x)P2 if x /∈fn(P1)(16)
X(˜n) ≡ P{˜n/ ˜m} if X( ˜m)=P (17)
Definition 3. Structural congruence axioms in SPiM. Structural congruence is reflexive,
symmetric and transitive, and holds in any context inside a process or a choice.
have defined binding and unbinding primitives purely by means of syntactic abbreviations.
A more detailed treatment of such primitives is given in [6], which
defines a novel calculus for binding and unbinding.
7.2 Reaction
The reaction rules of the calculus are presented in Definition 2. The rules are
of the form condition
reaction , where the given condition must be satisfied in order for
the given reaction to be possible. The notation P
r
−→ P states that the process
P can evolve to P by performing a reaction at rate r. If a species X(˜n) can
do a delay r.P in competition with actions M, then the species can evolve to a
process P at rate r (6). If a species X1(˜n1) can do a send x ˜n r1
.P1 in competition
with actions M1, and in parallel a species X2(˜n2) can do a receive x( ˜m)r2
.P2 in
competition with actions M2, then the two species can interact at the rate of the
channel x times the rates r1 and r2 of the actions. After the interaction takes
place, the processes P1 and P2 are executed in parallel, where the parameters
˜m are replaced with ˜n in process P2, written P2 {˜n/ ˜m} (4). If the species X1(˜n1)
can do a restricted send x( ˜m)r1
.P1 in competition with actions M1 then the two
species become bound on channels ˜m after the interaction takes place, and the
48
result of the interaction is the complex ( ˜m) (P1 | P2) (5). All of these reactions
can also take place inside a restriction (7), inside a parallel composition (8)
and up to re-ordering of processes (9). The re-ordering is defined by structural
rules, according to Definition 3. Essentially, the rules allow the re-ordering of
parallel processes (11), actions (13) and restricted channels (15), and the removal
of empty parallel processes (10) and unused restricted channels (14). Parallel
composition is associative (12), and the scope of a restricted channel can be
extended over parallel processes that do not use the channel (16). Finally, a
species X(˜n) defined in the environment as X( ˜m) = P can be replaced with
its process definition, where the parameters ˜m are replaced with ˜n in process P
(17).
7.3 Reaction Probability
According to standard principles of chemical kinetics [13], the probability of a
reaction should be proportional to its rate. In order to compute this probability,
we therefore need a way of computing the rates of each of the individual reactions
in the system. We do this using a notion of standard form, presented in Definition
4, together with a set of indexed reaction rules, presented in Definition 5.
The approach is inspired by previous work presented in [16]. A given process can
be converted to standard form by application of the structural rules in Definition
3. In particular, all of the restricted channels can be moved to the top-level
by application of rule (16), and each of the species defined as a process can be
expanded by application of rule (17), until only species defined as a choice of
actions remain. Once a process is in standard form, we can identify each species
Xi(ni) by its position i in the parallel composition X1(n1) | . . . | XN (nN ), and
we can identify each action πj.Pj of a given species by its position j inside the
choice π1.P1 + . . . + πM .PM . Thus, each action can be identified by a pair (i, j).
These action identifiers can then be used to identify all of the reactions in a
system. A reaction identifier w can be a pair of indices (i, j) denoting a delay
action j inside choice i, or a tuple of four indices (i1, j1, i2, j2) denoting a send
action j1 inside choice i1 interacting with a receive action j2 inside choice i2.
The indexed reaction rules of Definition 5 are analogous to those of Definition 2,
except that they assign an identifier w to each reaction, based on the position
of each action in the system. The notation P
r,w
−→ P states that the process P
can evolve to P by performing a reaction with identifier w at rate r.
We can use the indexed reaction rules to compute the probability of individual
reactions, as shown in Definition 6. The propensity ρ(P) of a process P
is defined as the sum of the rates of all the reactions of the process (24). The
probability Pr(P
r,w
−→ P ) that the process P can perform a particular reaction
w with rate r and evolve to P is given by the rate of the reaction divided by
the propensity of the process (25). Similarly, the probability Pr(P −→ P ) that
the process P can reduce to P is given by the sum of the probabilities of all the
reactions for which P can reduce to P (26). These reaction probabilities form
the basis of the SPiM stochastic simulation algorithm. More details about the
algorithm and its corresponding implementation are presented in [25].
49
(x1) . . . (xM ) (X1(˜n1) | . . . | XN (˜nN )) (18)
Definition 4. A process P is in standard form if it is in the form given by (18), such
that all restricted channels are at the top-level, and each species Xi(ni) is defined as a
choice of actions Mi in the environment.
Xi1 (˜ni1 ) = j∈J1
πj.Pj πj1 = x ˜n r1
j1 ∈ J1
Xi2 (˜ni2 ) = j∈J2
πj.Pj πj2 = x( ˜m)r2
j2 ∈ J2
i1 = i2 i1, i2 ∈ I
i∈I Xi(˜ni)
(ρ(x)·r1·r2),(i1,j1,i2,j2)
−→ Pj1 | Pj2 {˜n/ ˜m} | i∈I\i1,i2
Xi(˜ni)
(19)
Xi1 (˜ni1 ) = j∈J1
πj.Pj πj1 = x( ˜m)r1
j1 ∈ J1
Xi2 (˜ni2 ) = j∈J2
πj.Pj πj2 = x( ˜m)r2
j2 ∈ J2
i1 = i2 i1, i2 ∈ I
i∈I Xi(˜ni)
(ρ(x)·r1·r2),(i1,j1,i2,j2)
−→ ( ˜m) (Pj1 | Pj2 ) | i∈I\i1,i2
Xi(˜ni)
(20)
Xi1 (˜ni1 ) = j∈J πj.Pj πj1 = r i1 ∈ I, j1 ∈ J
i∈I Xi(˜ni)
r,(i1,j1)
−→ Pj1 | i∈I\i1
Xi(˜ni)
(21)
P
r,w
−→ P
(x) P
r,w
−→ (x) P
(22)
P
r,w
−→ P ≡ Q
P
r,w
−→ Q
(23)
Definition 5. Indexed reaction rules in SPiM. Processes are assumed to be in standard
form, in accordance with Definition 4. A choice between zero or more actions
π1.P1 + . . . + πN .PN is abbreviated to a sum i∈{1...N} πi.Pi and a parallel composition
of zero or more processes P1 | . . . | PN is abbreviated to a product i∈{1...N} Pi.
ρ(P)
P
r,w
−→
r (24)
Pr(P
r,w
−→ P ) if P
r,w
−→ P then
r
ρ(P)
else 0 (25)
Pr(P −→ P )
P
r,w
−→P
r
ρ(P)
(26)
Definition 6. Reaction Probability in SPiM. (24) defines the propensity of process
P, where P
r,w
−→ means that P can perform a reaction w at rate r. (25) defines the
probability that process P can perform a particular reaction w at rate r and evolve to
P . (26) defines the probability that process P can evolve to P .
50
Definition Choice Parallel Species Restriction Null
D π1.P1 + . . . + πN .PN P1 | . . . | PN X(˜n), X( ˜m)=D ( ˜m) P 0
D P1 PN
p1 pN
... P1 PN...
~ ~m:=n
X P
~(m)
∅
Fig. 36. Graphical representation of a static environment E in SPiM, where the graphical
representation at the bottom is equivalent to the textual representation at the top.
Process Species Restriction Null
P X(˜n), X( ˜m)=D ( ˜m) (X1(˜n1) | . . . | XN (˜nN )) 0
P
1
X
~ ~m:=n
XNX1
1
~ ~m1:=n1
~ ~mN:=nN
~(m)
∅
Fig. 37. Graphical representation of a dynamic process P in SPiM, where the graphical
representation at the bottom is equivalent to the textual representation at the top.
7.4 Graphical Representation
The syntax of the SPiM calculus supports a concise graphical representation,
originally presented in [26]. A system E P is displayed in two parts, a static
environment E which remains constant over time, and a dynamic process P
which is updated after each reaction.
The static environment E is displayed as a graph of nodes and edges, as
shown in Fig. 36. Each choice M or process P in the environment is displayed as
a node in the graph, and each definition X( ˜m)=D assigns an identifier X and
a label X( ˜m) to a given node D. A choice π1.P1 +. . .+πN .PN is displayed as an
elliptical node with edges to nodes P1, . . . , PN . Each edge to a node Pi is labelled
with an action πi and denotes an alternative execution path in the system. A
parallel composition P1 | . . . | PN is displayed as a rectangular node with edges
to nodes P1, . . . , PN . Each edge to a node Pi denotes a concurrent execution
path in the system. A species X(˜n) with definition X( ˜m)=D is displayed as a
51
X(n) = r.P1 + . . . + πN .PN X(n)
r
−→ P1
r
X(n)
P1 PN
r pN
...
1
X(n)
P1 PN
r pN
...
1
X(n) = x.P1 + . . . + πN .PN ,
Y (m) = x.Q1 + . . . + πM .QM
X(n) | Y (m)
ρ(x)
−→ P1 | Q1
r(x)
Y(m)
Q1 QM
x pM
...
1
X(n)
PN P1
pN
...
_
x
Y(m)
Q1 QM
x pM
...
1
X(n)
PN P1
pN
...
_
x
1
1
X(n) = x(u).P1 + . . . + πN .PN ,
Y (m) = x(u).Q1 + . . . + πM .QM
X(n) | Y (m)
ρ(x)
−→ (u) (P1 | Q1)
r(x)
Y(m)
Q1 QM
x(u) pM
...
1
X(n)
PN P1
pN
...
_
x(u)
Y(m)
Q1 QM
x(u) pM
...
1
X(n)
PN P1
pN
...
_
x(u)
1
1 u
Fig. 38. Example graphical reactions in SPiM, where the graphical representation at
the bottom is equivalent to the textual representation at the top. The examples illustrate
the three main reaction rules (6), (4) and (5) of Definition 2, respectively. A given
process is executed graphically by first applying one of the calculus reaction rules and
then displaying the resulting process.
rectangular node with an edge to the node identified by X. If ˜m = ˜n then the tip
of the edge is labelled with the substitution ˜m := ˜n. This represents the passing
of parameters from one node to another. Note that we only need to display the
mappings in the substitution that are different from the identity, since we only
need to indicate those parameters that are modified by the substitution. We
also abbreviate the notation so that edges from a choice or parallel composition
node to a species X(˜n) are connected directly to node X, without displaying the
additional rectangle. For example, an action π.X(˜n) with X(˜n) = D is displayed
as
π
−→X instead of
π
−→ −→X. A restriction ( ˜m) P is displayed by placing the
restricted channels ( ˜m) next to node P, and the null process 0 is not displayed.
The dynamic process P is displayed by attaching tokens to nodes in the graph
to represent currently executing species, and by connecting restricted names to
52
tokens to represent the formation of complexes, as shown in Fig. 37. A species
X(˜n) with definition X( ˜m) = D is displayed by attaching a substitution token
˜m := ˜n to the node identified by X. A parallel compositions of species X1(˜n1) |
. . . | XN (˜nN ) that share restricted names ( ˜m) is displayed by drawing a dotted
edge from names ˜m to the tokens of each of the species. This represents the
formation of a complex of species. The null process 0 is not displayed.
We introduce a number of convenient abbreviations for the graphical representation
of dynamic processes. If N identical tokens are attached to the same
node, they can be replaced with a single token preceded by the number N. Furthermore,
if ˜m = ˜n the token can be displayed as just the number N. If there
are N copies of a restriction ( ˜m) P they can be displayed by placing the number
N next to the restricted names ˜m. The scope of a restricted name can be further
clarified by only drawing a dotted edge to species that use this name. For example,
a restriction (x1, x2) (X1(x1) | X2(x1, x2) | X3(x2)) can be displayed as:
X1(1) · · · x1 · · · (1)X2(1) · · · x2 · · · (1)X3, assuming that the substitution tokens are
all empty and displayed as (1). The graphical representation clearly shows that
X1 and X3 do not share any restricted names, and are therefore not connected
to each other directly. Finally, in some cases a graph can appear crowded when
multiple tokens with different substitutions are attached to the same node. In
order to improve readability in these cases, it is convenient to display a separate
copy of the entire graph for each of the different tokens.
Once we have formally defined a graphical syntax in this way, we can use it
to display the execution of a model. Essentially, the execution of a system can be
visualised by applying the reaction rules to a given process in SPiM, and then
displaying the resulting process after each reaction. Thus, we do not need to
formally define the graphical reactions rules. We illustrate a number of example
reactions by displaying the process before and after the reaction takes place, as
shown in Fig. 38.
7.5 SPiM with Compartments
The SPiM calculus can be extended with mobile compartments, resulting in
a calculus similar to the bioambient calculus originally presented in [31]. The
syntax of SPiM with compartments is a direct extension of Definition 1. The
syntax of processes is extended with the notion of an ambient P consisting of
a process P inside a compartment. The syntax of actions is extended so that
an action π can also be a receive γx( ˜m) of names ˜m on channel x from an
ambient, a send γx ˜n of names ˜n on channel x to an ambient, or a send γx( ˜m)
of restricted names ˜m on channel x to an ambient, where γ denotes the direction
of communication. This can be from a child ambient to its parent (c2p ), from a
parent ambient to a child (p2c ), or from one sibling ambient to another (s2s ).
In addition, an action π can be a move µx on channel x or an accept µx on
channel x, where µ denotes the direction of movement. This can be an ambient
entering one of its siblings (in ), a child ambient leaving its parent (out ) or a
merge of two sibling ambients (merge ).
53
P, Q ::= . . .
| P Ambient containing process P
π ::= . . .
| γx( ˜m) Ambient receive
| γx ˜n Ambient send
| γx(˜n) Ambient restricted send
| µx Ambient move
| µx Ambient accept
γ ::= s2s Sibling
| c2p Parent
| p2c Child
µ ::= in Enter
| out Leave
| merge Merge
Definition 7. Syntax of SPiM with compartments, which extends the syntax of Definition
1, where γ represents a direction of communication and µ represents a direction
of movement.
The reaction rules for SPiM with compartments are presented in Definition 8,
and extend the reaction rules of Definition 2. The rules (27) - (29) allow two
ambients to communicate with each other. If a species X1(˜n1) can do a send
γx ˜n r1
.P1 in competition with actions M1, and in parallel a species X2(˜n2) can
do a receive γx( ˜m)r2
.P2 in competition with actions M2, then the two species
can interact at the rate of the channel x times the rates r1 and r2 of the actions.
After the interaction takes place, process P1 continues executing in the sending
ambient, and process P2 continues executing in the receiving ambient, where the
parameters ˜m are replaced with the names ˜n in process P2, written P2 {˜n/ ˜m}.
The communication can be from a child to a parent (27), from a parent to a
child (28) or from one sibling to another (29).
The rules (30) - (32) allow two ambients to move in or out of each other, or to
merge with each other. If a species X1(˜n1) can do a move µxr1
.P1 in competition
with actions M1, and in parallel a species X2(˜n2) can do an accept µxr2
.P2 in
competition with actions M2, then the two species can interact at the rate of
the channel x times the rates r1 and r2 of the actions. After the interaction
takes place, process P1 continues executing in the moving ambient, and process
P2 continues executing in the accepting ambient. The move can be one sibling
entering another (30), a child leaving its parent (31) or two siblings merging (32).
In the case of the merge, the contents of both ambients are merged in a single
ambient. In addition, all of these reactions can take place inside an ambient (33).
The structural rules for SPiM with compartments are extended with the
rules in Definition 9, which allow empty ambients to be deleted (34) and the
scope of restricted names to be extended outside an ambient (35). The reaction
probabilities for SPiM with compartments can be derived in a similar way to
Definition 5 and Definition 6, but we omit the details here. More information
54
X1(˜n1) = c2p x ˜n r1
.P1 + M1 X2(˜n2) = c2p x( ˜m)r2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2{˜n/ ˜m}
(27)
X1(˜n1) = p2c x ˜n r1
.P1 + M1 X2(˜n2) = p2c x( ˜m)r2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2{˜n/ ˜m}
(28)
X1(˜n1) = s2s x ˜n r1
.P1 + M1 X2(˜n2) = s2s x( ˜m)r2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2{˜n/ ˜m}
(29)
X1(˜n1) = in xr1
.P1 + M1 X2(˜n2) = in xr2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2
(30)
X1(˜n1) = out xr1
.P1 + M1 X2(˜n2) = out xr2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2
(31)
X1(˜n1) = merge xr1
.P1 + M1 X2(˜n2) = merge xr2
.P2 + M2
Q1 | X1(˜n1) | Q2 | X2(˜n2)
ρ(x)·r1·r2
−→ Q1 | P1 | Q2 | P2
(32)
P
r
−→ P
P
r
−→ P
(33)
Definition 8. Reaction rules for SPiM with compartments, which extend the reaction
rules of Definition 2. The condition X(˜n) = D is true if there is a definition in
the system environment such that X( ˜m) = D and D = D{˜n/ ˜m}. For each of the
communication rules (27), (28) and (29) there is also a corresponding rule for restricted
communication, along the lines of (4) (not shown).
0 ≡ 0 (34)
(x) P ≡ (x) P (35)
Definition 9. Structural congruence axioms for SPiM with compartments, which extend
the axioms of Definition 3.
about a stochastic simulation algorithm for bioambients and its corresponding
implementation are presented in [24].
55
Definition Ambient
D . . . P
D . . . P
Process Ambient
P . . . P
P
1
. . .
P
1
Fig. 39. Graphical representation for SPiM with compartments. The static representation
of an ambient extends the display of static environments E in Fig. 36, while the
dynamic representation extends the display of dynamic processes P in Fig. 37.
The SPiM calculus with compartments also supports a concise graphical
representation, as shown in Fig. 39. In a static environment, an ambient P is
displayed as a triangular node with an edge to process P. In a dynamic process,
the ambient is displayed as a box around the process P, where each box contains
its own copy of the graph representing P. If N identical ambients are executing
in parallel, only one copy of the ambient is represented, with the number N next
to the surrounding box. Once again, the graphical reaction rules do not need
to be defined explicitly. Instead, the execution of a system can be visualised by
displaying the system before and after each reaction. We illustrate a number of
example reactions, as shown in Fig. 40.
56
X(n) = s2s x.P1 + . . . + πN .PN ,
Y (m) = s2s x.Q1 + . . . + πM .QM
P | X(n) | Q | Y (m)
ρ(x)
−→ P | P1 | Q | Q1
s2s x
_
s2s xs2s x
_
s2s x r(x)
Y(m)
Q1 QM
pM
...
1
X(n)
PN P1
pN
...
Y(m)
Q1 QM
pM
...
1
X(n)
PN P1
p1
...
1
1
P Q P Q
1 1 1 1
X(n) = in x.P1 + . . . + πN .PN ,
Y (m) = in x.Q1 + . . . + πM .QM
P | X(n) | Q | Y (m)
ρ(x)
−→ P | P1 | Q | Q1
in x
_
in x
_
in x r(x)
Y(m)
Q1 QM
in x pM
...
1
X(n)
PN P1
pN
...
Y(m)
Q1 QM
pM
...
1
X(n)
PN P1
p1
...
1
1
P Q P Q
1 1 1 1
X(n) = out x.P1 + . . . + πN .PN ,
Y (m) = out x.Q1 + . . . + πM .QM
P | X(n) | Q | Y (m)
ρ(x)
−→ P | P1 | Q | Q1
out x
_
out x
_
out x r(x)
Y(m)
Q1 QM
out x pM
...
1
X(n)
PN P1
pN
...
Y(m)
Q1 QM
pM
...
1
X(n)
PN P1
p1
...
1
1
P Q P Q
1 1 1 1
Fig. 40. Example graphical reactions in SPiM with compartments, where the graphical
representation at the bottom is equivalent to the textual representation at the top. The
examples illustrate the reaction rules (29), (30) and (31) of Definition 8, respectively.
References
1. Ralf Blossey, Luca Cardelli, and Andrew Phillips. A compositional approach to
the stochastic dynamics of gene networks. Transactions in Computational Systems
Biology, 3939:99–122, January 2006.
2. Ralf Blossey, Luca Cardelli, and Andrew Phillips. Compositionality, stochasticity
and cooperativity in dynamic models of gene regulation. HFSP Journal, 2(1):17–
57
28, February 2008.
3. Johannes Borgstroem, Andrew Gordon, and Andrew Phillips. A chart semantics
for the pi-calculus. Electronic Notes in Theoretical Computer Science, 194(2):3–29,
January 2008.
4. Luca Cardelli. Brane calculi. In CMSB’04, pages 257–278, 2004.
5. Luca Cardelli and Andrew D. Gordon. Mobile ambients. Theoretical Computer
Science, 240(1):177–213, 2000. An extended abstract appeared in FoSSaCS ’98:
140–155.
6. Luca Cardelli and Gianluigi Zavattaro. On the computational power of biochemistry.
In Algebraic Biology, volume 5147 of LNCS, pages 65–80. Springer, July
2008.
7. Nathalie Chabrier-Rivier, Fran¸cois Fages, and Sylvain Soliman. The biochemical
abstract machine BIOCHAM. In V. Danos and V. Sch¨achter, editors, Proc. CMSB,
volume 3082 of LNCS, pages 172–191. Springer, 2004.
8. Federica Ciocchetta and Jane Hillston. Bio-pepa: An extension of the process
algebra pepa for biochemical networks. Electron. Notes Theor. Comput. Sci.,
194(3):103–117, 2008.
9. Paul W. Sternberg Claudiu A. Giurumescu and Anand R. Asthagiri. Intercellular
coupling amplifies fate segregation during caenorhabditis elegans vulval development.
PNAS, 103(5):1331–1336, January 2006.
10. Vincent Danos and Cosimo Laneve. Formal molecular biology. Theor. Comput.
Sci., 325(1):69–110, 2004.
11. Michael B. Elowitz and Stanislas Leibler. A synthetic oscillatory network of transcriptional
regulators. Nature, 403(6767):335–338, January 2000.
12. Jasmin Fisher, Nir Piterman, Alex Hajnal, and Thomas A. Henzinger. Predictive
modeling of signaling crosstalk during c. elegans vulval development. PLoS
Computational Biology, 3(5):0862–0873, May 2007.
13. Daniel T. Gillespie. Exact stochastic simulation of coupled chemical reactions. J.
Phys. Chem., 81(25):2340–2361, 1977.
14. C˘alin C. Guet, Michael B. Elowitz, Weihong Wang, and Stanislas Leibler. Combinatorial
synthesis of genetic networks. Science, 296:1466–1470, May 2002.
15. Michael Hucka et al. The systems biology markup language (sbml): a medium
for representation and exchange of biochemical network models. Bioinformatics,
19(4):524–531, 2003.
16. C´eline Kuttler, C´edric Lhoussaine, and Joachim Niehren. A stochastic pi calculus
for concurrent objects. In Algebraic Biology, volume 4545 of LNCS, pages 232–246.
Springer, August 2007.
17. Paola Lecca and Corrado Priami. Cell cycle control in eukaryotes: a biospi model.
In BioConcur’03. ENTCS, 2003.
18. Robin Milner. Communicating and Mobile Systems: the π-Calculus. Cambridge
University Press, 1999.
19. Robin Milner, Joachim Parrow, and David Walker. A calculus of mobile processes,
part I/II. Journal of Information and Computation, 100:1–77, September 1992.
20. Kenneth M. Murphy, Paul Travers, and Mark Walport. Janeway’s Immunobiology.
Garland Science, seventh edition, 2008.
21. Rami Marelly Lara Appleby Jasmin Fisher Amir Pnueli David Harel Michael J.
Stern E. Jane Albert Hubbard Na’aman Kam, Hillel Kugler. A scenario-based
approach to modeling development: A prototype model of c. elegans vulval fate
specification. Developmental Biology, Volume 323(1):1–5, November 2008.
22. Denis Noble. The Music of Life: Biology Beyond the Genome. Oxford University
Press, 2006.
58
23. Michael Pedersen and Gordon Plotkin. A language for biochemical systems. In
M. Heiner and A. M. Uhrmacher, editors, Proc. CMSB, LNCS. Springer, 2008.
24. Andrew Phillips. An abstract machine for the stochastic bioambient calculus.
Electronic Notes in Theoretical Computer Science, 227:143–159, January 2009.
25. Andrew Phillips and Luca Cardelli. Efficient, correct simulation of biological processes
in the stochastic pi-calculus. In Computational Methods in Systems Biology,
volume 4695 of LNCS, pages 184–199. Springer, September 2007.
26. Andrew Phillips, Luca Cardelli, and Giuseppe Castagna. A graphical representation
for biological processes in the stochastic pi-calculus. Transactions in Computational
Systems Biology, 4230:123–152, November 2006.
27. C. Priami and P. Quaglia. Beta binders for biological interactions. In V. Danos and
V. Sch¨achter, editors, Proc. CMSB, volume 3082 of LNBI, pages 20–33. Springer,
2005.
28. C. Priami, A. Regev, E. Shapiro, and W. Silverman. Application of a stochastic
name-passing calculus to representation and simulation of molecular processes.
Information Processing Letters, 80:25–31, 2001.
29. Corrado Priami. Stochastic π-calculus. The Computer Journal, 38(6):578–589,
1995.
30. A. Regev, W. Silverman, and E. Shapiro. Representation and simulation of biochemical
processes using the pi- calculus process algebra. In Pacific Symposium
on Biocomputing, volume 6, pages 459–470, 2001.
31. Aviv Regev, Ekaterina M. Panina, William Silverman, Luca Cardelli, and Ehud Y.
Shapiro. Bioambients: an abstraction for biological compartments. Theor. Comput.
Sci., 325(1):141–167, 2004.
32. Aviv Regev and Ehud Shapiro. The π-calculus as an abstraction for biomolecular
systems. Modelling in Molecular Biology, pages 219–266, 2004.
33. Davide Sangiorgi and David Walker. The π-calculus: a Theory of Mobile Processes.
Cambridge University Press, 2001.
34. Lauren M. Sompayrac. How the Immune System Works. Wiley-Blackwell, third
edition, 2008.
35. Paul W. Sternberg. Vulval development. WormBook, June 2005.
36. Paul W. Sternberg and H. Robert Horvitz. Pattern formation during vulval development
in c. elegans. Cell, 44:761–772, March 1986.
37. Paul W. Sternberg and H. Robert Horvitz. The combined action of two intercellular
signaling pathways specifies three cell fates during vulval induction in c. elegans.
Cell, 58:679–693, August 1989.
38. Jonathan W. Yewdell, Eric Reits, and Jacques Neefjes. Making sense of mass
destruction: quantitating mhc class i antigen presentation. Nature Reviews Immunology,
3:952–961, 2003.
59