See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261092283

A Modified Stochastic Simulation Algorithm for Time-Dependent Intensity

Rates

Conference Paper · May 2013

DOI: 10.1109/CSCS.2013.101

CITATIONS READS
9 754

2 authors:

Raluca Roxana Purnichescu Purtan Andreea Udrea

Polytechnic University of Bucharest Polytechnic University of Bucharest
31 PUBLICATIONS 284 CITATIONS 61 PUBLICATIONS 382 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Raluca Roxana Purnichescu Purtan on 28 September 2018.

The user has requested enhancement of the downloaded file.

 2013 19th International Conference on Control Systems and Computer Science

A Modified Stochastic Simulation Algorithm for

Time-Dependent Intensity Rates
Raluca Roxana Purnichescu Purtan Andreea Udrea
Faculty of Applied Science, Department of Mathematical Faculty of Automatic Control and Computers, Department
Methods and Models of Automatic Control and Systems Engineering
University “Politehnica” of Bucharest University “Politehnica” of Bucharest
Bucharest, Romania Bucharest, Romania
[email protected] [email protected]

Abstract — There are two main approaches in the modeling well-mixed chemical kinetic systems are based on
mathematical modeling of coupled systems of (bio)chemical the work of Gillespie [5][6], who developed the stochastic
reactions: continuous, represented by differential equations simulation algorithm (SSA) in two variants, the Direct and
whose variables are concentrations or discrete, represented by First Reaction Methods. More recently, Gibson and Bruck [4]
stochastic processes whose variables are numbers of molecules.
The latter approach is used mostly for biochemical systems with
created the Next Reaction Method and Cao et all [3]
a low to moderate number of molecules of certain species and this demonstrated that for certain classes of systems the Direct
kind of systems are typically modeled as continuous time – method can be faster than the Next Reaction method. The
discrete state Markov Process. There are exact stochastic stochastic simulation algorithm produces an exact solution to
algorithms to simulate state trajectories of discrete, stochastic the dynamical chemical system by simulating each molecular
systems and these algorithms are based on methods that are reaction event. For systems with fast reversible or numerous
rigorously equivalent to the Master Equation approach. Two of reactions, the exact simulation algorithm becomes
the most widely used methods for simulating the stochastic computationally costly.
dynamics of a chemical system are the exact stochastic simulation
algorithm (SSA, known also as Gillespie algorithm) and its The basic idea of the SSA is that, at each point in time, a
approximate variant, the tau-leaping algorithm. waiting time to the next reaction and the most likely reaction
This paper describes a modified version of SSA - First to occur must be sampled from a joint probability density
reaction method - by letting the intensity rates of the reactions to function. If the rate constants and/or the number of molecules
be functions of time. The importance of this adaptation is obvious in the system are large, then the time step of the algorithm can
when considering some classes of biological models (for example, be very small and a very large number of steps must be
the one involving circadian rhythm). The underlying assumptions employed for the completion of the simulation. Because of
are that the system is well stirred such that at any moment, each this, Gillespie [7] introduced the Poisson W -leap method (an
reactions occur with equal probability at any position, that each Euler-type method) in which more than one reaction are
allowed to fire in a given time interval W with a frequency
reaction, once occurred, completes instantaneously (there are no
reactions with delay involved) and that the system is non stiff
(there are no different time scales of the reactions involved). extracted from a Poisson distribution. Since then, many
extensions of this idea have been developed, most of them
Keywords — chemical kinetics, stochastic simulation regarding the improvement of the computational efficiency
algorithm, time-dependent intensity rates and finding efficient methods for simulating stiff systems. In
the last 10 years, especially in the biochemical field, the
I. INTRODUCTION variability in external factors (temperature fluctuation, change
in volume, circadian rhythms) has been taken more and more
Stochastic chemical kinetics describes the time evolution into account, and the need for mathematical models which can
of a well-stirred chemically reacting system in a way that include them increased. There have been limited studies
takes into account the fact that molecules dynamics come in concerning this subject [8], [1], [2]. A very interesting
whole numbers and exhibit some degree of randomness in extension of the SSA can be found in the recent paper [9]
their dynamical behavior. In the last 30 years, more and more where the rate constant is replaced with a random variable, i.e.
researchers are using this approach to chemical kinetics in the possess a stationary distribution with a known density.
analysis of cellular systems in biology, where small molecular
populations of only a few reactant species can lead to In this paper we extend the SSA (First Reaction method)
deviations from the predictions of the deterministic differential by letting the reaction intensity be a general function of time.
equations of classical chemical kinetics. In particular, we formulate an optimization problem and
discuss it for a class of non-negative and continuous functions
By representing the chemical reacting system as a jump from the point of view of numerical implementation and time
Markov process, the state of the system is modeled in terms of costs. We also come with new stopping conditions for the
molecule numbers (non-negative integers) and as a chemical proposed algorithm in terms of reaction chain verisimilitude.
reaction fire, discrete state transitions occur, in continuous Until now some conditions for non-negative populations of
time. Many of the existing numerical methods developed for

978-0-7695-4980-4/13 $26.00 © 2013 IEEE 365

DOI 10.1109/CSCS.2013.101
molecules were discussed and implemented for the property (P2) exist, and then determining the forms of those
approximate binomial W -leap method [10]. functions. This effort typically focuses on identifying for each
RP a reaction probability rate constant cP 2. Then, using the
addition law of probability theory, the probability (P2) can be
II. THEORETICAL AND COMPUTATIONAL computed as the sum of cP dt over all distinct combinations of
FRAMEWORK
RP reactant molecules in the current state
The state of a chemical system in the stochastic framework x X (t ) X 1 (t ), X 2 (t ),..., X N (t ) .
is defined by the number of each species involved and changes
discretely with every reaction that is executed. This
If we denote the number of all distinct combinations of RP
simplifying approach is made under the assumption that
nonreactive collisions occur far more often than reactive reactant molecules in the system at any time t , when there are
collisions and, hence, the fast dynamics of motion can be exactly Xi of the S i molecules (i 1, 2, ..., N )
neglected.
by hP ( X 1 (t ), X 2 (t ),..., X N (t )) , we can write the propensity
A typical process can be defined as a well-stirred system
of molecules of N chemical species {S1 , S 2 , ..., S N } , which functions for each reaction channels RP as
interact through M elemental1 reaction channels aP ( x ) dt cP hP ( x ) dt .
{R1 , R2 , ..., RM } . Our goal is to estimate the state vector
The vectors hP ( x ) are products of combinatorial
X (t ) X 1 (t ), X 2 (t ),..., X N (t ) , given that the system starts
functions, corresponding to each particular chemical reaction
in state X (t0 ) x0 at some initial time t0 . Assuming that the structure. Each vector is easily obtained from the reaction
system obeys the mass-action law, each reaction channel RP channel by counting the numbers of each molecule species
that are consumed and produced in the reaction.
can be characterized mathematically by two quantities [5]:
The form of cP is highly reaction specific –
(P1) The state-change vector ("jump" vector)
vP (vP 1 , vP 2 , ..., vP N ) , where vP i is the experimentally deduced and usually depends on the system
volume and temperature.
stoichiometric vector for the specific reaction RP ,
If cP are constants, the propensity rates are time
defined to be the change in the {Si } vector of independent the fluctuations in the system are due to intrinsic
molecular populations induced by a single noise.
occurrence of the particular RP reaction event; thus,
If cP is a time-dependent function, it corresponds to time
an RP reaction induces the state change x o x  vP . dependent propensity rates and the fluctuations in the system
(P2) The intensity (propensity) rate aP defined arises from variability in external factors (extrinsic noise).

by: aP ( x ) dt the probability, given X (t ) x , that one

Example: Consider a system with 4 molecular populations
particular RP reaction will occur inside : in the
(N 4) which interact via 3 reaction channels:
next infinitesimal time interval [t , t  dt ) . The
propensity rates are transition probabilities per unit
S1  S 2 o 2 S1  S 3
time for all M reactions.
Definition (P2) can be viewed as the fundamental premise S 2  2 S3 o S4
of stochastic chemical kinetics, because everything else in the
theory follows from it ([5], [6]). This requires finding the S4 o S2
specific conditions under which functions aP ( x ) having the
We denote by x ( x1 , x2 , x3 , x4 ) the state of the system at
1 a time moment t .
 An elemental reaction is one that occurs essentially
instantaneously. In practice, there are only two types of
elemental reaction: unimolecular, in which a single molecule 2
changes form; and bimolecular, in which two molecules  cP dt probability that a particular combination (randomly
collide and a chemical change occurs as a result. All other chosen) of RP reactant molecules will react accordingly in the
types of reaction (trimolecular, reversible, etc.) are made up of
a series of two or more elemental reactions. next infinitesimal time interval [t , t  dt ) 

366
 TABLE I System characteristics wP ( X t , t X s , s ) M

wt
¦ [ P( X t
 vP , t X s , s ) u cP hP ( X t  vP ) 
(4)
Type Reaction description hP ( x ) and vP ( x ) P 1

 P ( X t , t X s , s ) u cP hP ( X t )]
h1 ( x ) x1 x2
R1 S1  S 2 o 2 S1  S 3
v1 (1, 1,1, 0) or, after replacing aP ( x ) with cP hP ( x ) :
1 wP ( X t , t X s , s ) M

h2 ( x ) x2 x3 ( x3  1)
wt
¦ [ P( X t
 vP , t X s , s ) u a P ( X t  vP ) 
(5)
R2 S 2  2 S3 o S4 2 P 1

v2 (0, 1, 2,1)  P ( X t , t X s , s ) u aP ( X t )]

h3 ( x ) x4 In principle, the CME completely determines the

R3 S4 o S2
v3 (0,1, 0, 1) function P( X , t X , s ) , subject to some initial conditions. The
t s

CME equation is an exact consequence from the reaction

channels characterized by intensity rates and state-change
After any of the reactions fires, the state of the system will vectors. If the equation can be solved (i.e. for a system with
change accordingly: x o x  vP . very few states), then the process X (t ) is fully tractable.
However, an exact solution of CME can rarely be obtained
A. The Chemical Master Equation and the difficulty comes from the high dimension, which
equals the total number of possible states of the system under
It has been proven ([5], [6]) that the behavior of the study.
N species population vector X (t ) is a jump type Markov
process on the non-negative N -dimensional integer lattice, B. The Stochastic Simulation Algorithm
i.e. the transition probability depends only on the current state The intractability of the CME lies also in the fact that the
and not on the previous states. The probability that a certain deterministic approach tries to solve simultaneously for the
reaction RP will take place in the next infinitesimal time probability of all possible trajectories. An alternate approach
interval [t  dt ) is given by: for describing the behavior of the system is to generate a
numerical realization of X (t ) , i.e., a single trajectory of
P ^ N (t  dt )  N (t ) 1` a P ( X (t )) dt  o ( dt ) 3 (1) X (t ) versus t. This is not the same as solving the CME
numerically [5]. The SSA proposed by Gillespie in [5] with
Consequently, the probability that no reaction will occur in both of its variants (The Direct Method and The First Reaction
[t  dt ) is: Method) is an exact algorithm because the generated
M
trajectories are chosen according to the correct probability
P ^ N (t  dt )  N (t ) 0` 1  ¦ a ( X (t ))dt  o(dt )
P
(2) distributions, i.e. the probability of generating a given
trajectory with the simulation algorithm is exactly the
P 1
probability that would come out of the solution of the CME.
and the probability that more than one reaction occur in Due to the fact that SSA is generating trajectories with the
[t  dt ) is: correct probability, any parameters of interest can be estimated
by generating many trajectories.
P ^ N (t  dt )  N (t ) ! 1` o (t ) (3)
If we consider a system of N molecular species (states)
The corresponding probabilities for these three events that interacting via M chemical reactions (transitions), at each
can take place in the interval [t  dt ) are rigorously deduced time instant the system is in exactly one state and there are at
most M possible transitions from that state to another. The
by D.T.Gillespie in [5], and we denoted by N (t ) the number main two questions are „which reaction occurs next ?” and
of events (reactions RP ) that occurred by the time t . „when does it occur ?”. In answer, Gillespie proposed two
exact simulation methods.
The Chemical Master Equation (CME) is just another
Further, we will focus on the First Reaction Method,
name for the Kolmogorov forward equation of Markov
Processes. It contains one probability variable for each assuming, for now, that all reaction probability rates c P are
possible state of the system and represents a set of linear time independent.
differential equations with constant coefficients for the
conditional probability: Taking into account that relations (1), (2), (3) hold for each
reaction channel, the reactions are independent Poisson
P( X t , t X s , s) P{X (t ) X t , given that X (s) X s , for s d t} processes with intensity (propensity) rate aP ( x ) . Therefore the
distribution of the first event time (firing of the each reaction)
is exponentially distributed with a parameter of aP ( x ) .
3
 o ( dt ) denotes terms that are negligible for small dt Considering the whole system, the reaction that will fire first

367
(and will be the only reaction firing in the time interval As in the homogeneous case, we are interested to find the
[t , t  dt ) , according to (1), (2) and (3)) is, naturally, the one time of the first event for each reaction. We shall generate a
with the smallest time of first event. sequence of M random variables ( r1 , r2 , ..., rP ,..., rM ) , uniform
In order to find the time of the first event for each reaction, distributed in (0,1) and the time of the first firing for the P -
is a common practice to generate a sequence of M random th reaction will be the positive solution W P (with a minimal
variables ( r1 , r2 , ..., rP ,..., rM ) , uniform distributed in (0,1) and
value, if there are multiple solutions) of the integral equation:
apply the Probability Integral Transform to obtain a
corresponding sequence of M random variables exponentially t W P §1·
distributed : aP ( x ) ˜ e
 aP ( x ) ˜W P
. The time of the first firing for the ³ aP ( X (t ), u ) du ln ¨ ¸ (6)
© rP ¹
t

1 §1· and the reaction that will fire is the reaction corresponding to
P -th reaction will be W P ln ¨ ¸ and the reaction
aP ( x ) © rP ¹ tmin min ^W P ` . Finally, updating the state vector with the
P 1... M

that will fire is the reaction corresponding to tmin min ^W P ` . corresponding state change vector and the incrementing the
P 1... M
time by tmin , the algorithm is running until a stop condition is
This is the main core of the First Reaction Method. enabled.
Updating the state vector with the corresponding state change
The classical stopping condition is: the number of desired
vector and the incrementing the time by t min , the algorithm is
steps or the sum of the populations dimensions.
running until a stop condition is enabled.
In addition we consider that the following criterion must
The First Reaction Method is the algorithm implementing be incorporated in the stop condition: strict negative values for
the above strategies. In [5], Gillespie proved that is an exact each one of the populations.
simulation algorithm.
Note that X (t ) is constant in the above integrals (i.e. the
III. MODIFIED FIRST REACTION METHOD ALGORITHM hP vectors are constants) because no reaction take place in the
Consider a system in which the probability rates cP are time interval >t , t  W P
. In case of a time independent
time dependent, i.e. cP (t ) . For example, any modification in intensity (propensity) rate aP ( x ) , equation (6) becomes
the temperature or volume of the reacting system could lead to
1 §1·
changes in the specific probabilities rates cP (t ) . In WP ln ¨
¸ (the same as for the homogeneous case).
aP ( x ) © rP ¹
biochemistry is a well-known fact that external factors as
circadian rhythms can play a critical role in absorption and Another important aspect of the above approach is the
metabolize of different substances. Some of these external
influences have been monitored by repeatedly and carefully specific class of cP (t ) functions. In our future work we will
conducted experiments, so there is a large amount of data we study in depth this aspect.
can rely on for constructing some pertinent classes of specific
time dependent functions. Modified Algorithm for First Reaction Method
1) Initialization:
Using time dependent probability rates cP (t ) (at least for - Set the initial numbers of molecules for each species (in the
some of the reactions of the system), we have to change state vector X (t ) ) ;
accordingly the intensity rates: aP ( X (t ), t ) dt cP (t ) hP ( X (t )) dt .
- Set the specific probability rate constants cP (t ) for all
Every reaction is a non-homogeneous Poisson process. The
elemental reactions;
number of events for each reaction RP in any interval t1 , t 2 @
- Set the functions hP for all reactions;
has a Poisson distribution with parameter ³ aP ( X (t ), t ) dt .
t2

t 1 - Set the state-change vectors vP for all reactions;

In other words: - Set t 0.

ª ³ t aP ( X (t ), t ) dt º ˜ e ³
t2
k  aP ( X ( t ), t ) dt
2 t1
2) Calculate the propensity functions aP ( X (t ), t ) for each
P ^ X (t1 )  X (t 2 ) k `
¬ t 1 ¼
reaction .
k!
3) Generate the sequence of random numbers
It can be shown by straightforward calculations that
relations (1), (2) and (3) hold for time dependent intensity ^r1 , r2 ,..., rP ,...rM ` from a unit-interval uniform distribution.
rates.

368
 4) For every reaction, find the positive solution W P (with a In spite of the fact the time increases for the simulation
minimal value, if there are multiple solutions) of the integral duration, the proposed algorithm solves the general problem of
time varying intensity rates in a general manner. The
t W P §1· numerical implementation can be adapted for each class of
equation ³ aP ( X (t ), u ) du ln ¨ ¸ ; functions characterizing the intensity rates and optimal
© rP ¹
t
solution that decrease the simulation duration can be found.
5) Find tmin min ^W P ` and store the index of the reaction IV. CONCLUSIONS AND FURTHER WORK
P 1... M

with W P t min ; In this paper we propose a modified version of SSA - First

reaction method considering the intensity rates of the reactions
6) Update the state vector to reflect that reaction RP fired: to be functions of time. This modification gives an added value
to the original algorithm in the context of many bio-chemical
X (t ) X (t )  vP ; Set t m t  tmin ; processes with variable comportments. The mathematical
aspects and a series of numerical implementations problems
7) Go to step 2 or stop. and solutions are discussed. An example of class of functions
for which the algorithm can be applied without the otherwise
As it was already pointed out, SSA is time consuming inherent drawbacks in terms of is given.
especially if the number of reactions and populations is large.
If we consider the numerical implementation of step (4) we are Improvements and further work:
in fact confronted with an optimization problem: searching for
- simplifying the expression of the equality in the 4th step of
the minimal solution of the equation over the time interval
the algorithm in order to permit a more efficient numerical
(t , t max ) ,where t is the present time and t max is the simulation implementation
duration in order to determine the value for W P . - defining and analyzing other classes of functions inspired
directly from bio-chemistry.
There are some general methods used for determining the
most suitable value for t max function of the number of ACKNOWLEDGMENT
populations and reactions involved [5]. For the first author, this research is supported in part by
POSDRU/107/1.5/S/76813.
Numerically computing the integral at step (4) over the
time interval (t , t max ) and looking for the minimal solution of REFERENCES
[1] D.F. Anderson, “A modified next reaction method for simulating
the equation it would certainly raise a computation time chemical systems with time dependent propensities and delays”, Journal of
problem. Chemical Physics, vol.127, 2007, 214107
We consider that, in spite of this aspect, there are certain [2] A.P. Arkin, J. Ross and H.H. McAdams, “Stochastic kinetic analysis
classes of functions that are extremely common in bio- of developmental pathway bifurcation in phage lambda-infected Escherichia
coli cells”, Genetics, vol.149, no.4,1998, pp.1633-1648
chemistry for which the numerical algorithm can be
successfully implemented in terms of time costs. [3] Y. Cao, H. Li and L. Petzold, “Efficient Formulation of the Stochastic
Simulation Algorithm for Chemically Reacting Systems”, Journal of
One of these classes is represented by periodic functions Chemical Physics, vol.121, no.9, 2004, pp.4059-4067
(modeling, for example, the circadian rhythms, that have [4] M.A. Gibson and J. Bruck, “Efficient Exact Stochastic Simulation of
sinusoidal properties). For this specific class of functions the Chemical Systems with Many Species and Many Channels”, Journal of
search for a minimal solution takes place only on the interval Physical Chemistry, no.104, 2000, pp.1876-1889

§ t , § ª t º  1 · ˜ T · where T denotes the period of the [5] D.T. Gillespie, “Exact stochastic simulation of coupled chemical
¨ ¨« » ¸ ¸ reactions”, Journal of Physical Chemistry, no.81, (1977)pp.2340-2361
© © ¬T ¼ ¹ ¹
[6] D.T. Gillespie, “A general method for numerically simulating the
considered function. stochastic time evolution of coupled chemical reactions”, J. Comput. Phys.,
no.22, 1976, pp.403-434
Also, keeping in mind that: aP ( X (t ), t ) dt cP (t ) hP ( X (t )) dt
[7] D.T. Gillespie, “Approximate accelerated stochastic simulation of
chemically reacting systems”, Journal of Chemical Physics, vol.115, no.4,
where, in fact, hP ( X (t )) has a constant value during the 2001, pp.1716-1733
numerical implementation of this step, one only has to [8] P. Lecca, “A time-dependent extension of Gillespie algorithm for
compute the intersection between the sinusoid and a straight biochemical stochastic S -calculus”, 2006, Proceedings of the 2006 ACM
line, parallel to the time axes denoting the constant: symposium on Applied Computing 23-27 April 2006, Dijon, pp.137-144
[9] C.Y Lee, “Stochastic simulation of biochemical systems with
1 §1 · randomly fluctuating rate constants”, 2012, arXiv:1202.1266[q-bio.QM]
ln ¨ ¸.
hP ( X (t )) © rP [10] T. Tian, K. Burrage, “Binomial leap methods for simulating
¹ stochastic chemical kinetics”, Journal of Chemical Physics, vol.121, no.21,
2004, pp.10356-10364
Other classes of functions can be found for which the
algorithm does not dramatically increase in terms of duration.

369

View publication stats