ANALYSIS OF BROWNIAN DYNAMICS SIMULATIONS OF

REVERSIBLE BIMOLECULAR REACTIONS

JANA LIPKOVÁ∗ , KONSTANTINOS C. ZYGALAKIS† , S. JONATHAN CHAPMAN† ,
AND RADEK ERBAN†

Abstract. A class of Brownian dynamics algorithms for stochastic reaction-diffusion models

which include reversible bimolecular reactions is presented and analyzed. The method is a gener-
alization of the λ– model for irreversible bimolecular reactions which was introduced in [13]. The
formulae relating the experimentally measurable quantities (reaction rate constants and diffusion
constants) with the algorithm parameters are derived. The probability of geminate recombination is
also investigated.

Key words. Brownian dynamics, stochastic simulation algorithms, reaction-diffusion problems,

reversible bimolecular reactions

1. Introduction. Brownian dynamics algorithms are used in a number of ap-

plication areas, including modelling of ion channels [8], macromolecules [23], liquid
crystals [30] and biochemical reaction networks [24] to name a few. The main idea is
that some components of the system (e.g. solvent molecules), which are of no special
interest to a modeller, are not explicitly included in the simulation, but contribute to
the dynamics of Brownian particles collectively as a random force. This reduces the
dimensionality of the problem, making Brownian dynamics less computationally in-
tensive than the corresponding molecular dynamics simulations. In a typical scenario,
the position Xi = [Xi (t), Yi (t), Zi (t)] of the Brownian particle evolves according to
the stochastic differential equation


dXi = fi (X1 , X2 , . . . , Xi , . . . ) dt + 2Di dWi ,

where Wi = [Wi,x , Wi,y , Wi,z ] is the standard Brownian motion, Di is the diffusion
constant and fi is the deterministic drift term which depends on the positions of
other Brownian particles. Depending on the particular application area, the drift
term fi can take into account both attractive (e.g. electrical forces between ions of
the opposite charge), repulsive (e.g. steric effects, electrical forces between ions) and
hydrodynamic interactions [15]. In this paper, we focus on algorithms for spatial
simulations of biochemical reaction networks in molecular biology. In this application
area [5, 13, 34], it is often postulated that fi ≡ 0, i.e. the trajectory of each particle
is simply given by


dXi = 2Di dWi . (1.1)

In [13], we used this description of molecular trajectories and analyzed the so called
λ– stochastic simulation algorithm for modelling irreversible bimolecular reactions.
Considering three chemical species A, B and C which are subject to the bimolecular
reaction
1 k
A + B −→ C, (1.2)

∗ Charles University, Faculty of Mathematics and Physics, Sokolovská 83, 186 75 Prague 8, Czech

Republic; e-mail: [Link]@[Link].

† University of Oxford, Mathematical Institute, 24-29 St. Giles’, Oxford, OX1 3LB, United King-

dom; e-mails: zygalakis@[Link]; chapman@[Link]; erban@[Link].

1
2 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

it is postulated that a molecule of A and a molecule of B react with the rate λ

whenever their distance is smaller than the binding (reaction) radius  [9, 32]. This
definition makes use of two parameters λ and  while the irreversible reaction (1.2) is
described in terms of one parameter, the reaction rate k1 . Consequently, there exists
a curve in the λ– parameter space which corresponds to the same rate constant k1 .
In the limit λ → ∞, the model reduces to the classical Smoluchowski description
of diffusion-limited reactions, namely, two molecules always react whenever they are
closer than the reaction radius  [31, 6]. However, having two parameters λ and , we
can choose the reaction radius  close to the molecular radius (which is often larger
than the radius given by the Smoluchowski model [26, 13]) and use k1 to compute
the appropriate value of λ. In this paper, we will study extensions of the λ– model
to the reaction-diffusion systems which include reversible biochemical reactions of the
form
k1
A + B −→
←− C. (1.3)
k2

This reaction effectively means two reactions, the forward reaction (1.2) which is
modelled with the help of two parameters λ and  (as studied in [13]) and the backward
reaction
2k
C −→ A+B (1.4)
which can be also implemented in terms of two parameters: the rate constant of
the dissociation of the complex C and the unbinding radius σ. Since the reaction
(1.4) is of the first-order, the cleavage of the complex C is a Poisson process with
the rate constant k2 , i.e. the rate constant of the dissociation of C is equal to the
experimentally measurable quantity k2 . The second parameter, the unbinding radius
σ, is the initial separation of the molecules of A and B which are created after a
molecule of C dissociates.
Whenever new molecules of A and B are introduced to the system, we have to
initiate their positions. Since the algorithm considers all molecules as points, it would
make sense to place them at the position where the complex C was just before the
reaction (1.4) occurred, i.e. we would put σ = 0. However, this choice of σ can be
problematic. For example, in the Smoluchowski limit λ → ∞, if two particles start
next to each other, they must immediately react again according to the forward step
(1.2). Andrews and Bray [5] propose a solution to this problem by requiring that
the initial separation of molecules, the unbinding radius σ̄, must be greater than the
binding radius ¯. Here, we generalize the concept of unbinding radius for the λ–
model introduced in [13]. Since λ is in general less than infinity, we can choose the
unbinding radius σ which is less than the binding radius , including the case σ = 0.
This is investigated in detail in Section 3, but we start with the case σ >  in Section
2.
The algorithm for simulating (1.3) has four parameters: the binding radius , the
unbinding radius σ, the reaction rate λ (for the forward step (1.2)) and the rate of
dissociation of C, but we usually only have two experimentally measurable parameters
k1 and k2 . Since k2 is equal to the rate of dissociation of C, the remaining parameters
λ,  and σ will be related to k1 . To simplify the derivation of this relation, we define
the dimensionless parameter α as the ratio of the unbinding and binding radii, i.e.
σ
α= . (1.5)

 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 3

Two cases are considered separately: α > 1 and α ≤ 1, see Figure 1(a). If α > 1,
then the unbinding radius σ is larger than the binding radius . This situation is
investigated in Section 2. In Section 3, we consider the case α ≤ 1. The formula
relating k1 with model parameters λ,  and σ is derived as (2.12) for α > 1 (i.e.
for σ > ) and as (3.4) for α ≤ 1 (i.e. for σ ≤ ). It is given as one equation for
three unknowns λ,  and σ. In particular, there is a relative freedom in choosing the
parameters. For example, considering that  and σ are given, the equations (2.12)
and (3.4) can be used to compute the appropriate value of λ. However, the binding
and unbinding radii are not entirely a choice of a modeller. This is discussed in
Section 4. First, we would like the binding (reaction) radius to be of a size similar to
the molecular radius [13]. Second, we sometimes want to construct algorithms with a
given value of the probability of geminate recombination [5, 2], which is the probability
that a molecule of A and a molecule of B, created from the same molecule of C by
reaction (1.4), react with each other according to (1.2). In Section 4, we discuss how
this extra knowledge can be used to find optimal values of the parameters of the
algorithm. In particular, we find (equation (4.7)) that the geminate recombination
probability is proportional to the inverse of the binding radius  for the parameter
regime relevant to protein-protein interactions.
The analysis in Sections 2, 3 and 4 is done in the limit of (infinitesimally) small
time steps [13]. This provides valuable insights and a lot of interesting asymptotic
behaviour of the algorithm can be investigated. However, if we want to implement the
λ– model on the computer, we have to discretize the stochastic differential equation
(1.1) with a finite time step ∆t which we want to choose as large as possible to
decrease the computational intensity of the algorithm. This is studied in Section 5.
The numerical implementation of the Brownian dynamics algorithm illustrating the
validity of our analysis is presented in Section 6.

2. The case α > 1. The λ– model of the forward chemical reaction (1.2) states
that molecules of A and molecules of B diffuse with the diffusion constants DA and
DB , respectively. If the distance of a molecules of A and a molecule of B is less
than , then the molecules react with the rate λ. Considering a frame of reference
situated in the molecule of B, we can equivalently describe this process as the random
walk of a molecule of A which has the diffusion constant DA + DB . This molecule
diffuses to the ball of radius  (centered at origin) which removes molecules of A with
the rate λ [13]. In this frame of reference, the reverse step (1.4) corresponds to the
introduction of new molecules of A at the distance σ from the origin. Let c(r) be
the equilibrium concentration of molecules of A at distance r from the origin. It is a
continuous function with continuous derivative which satisfies the following equation:
 
d2 c 2 dc
(DA + DB ) + − λc = 0, for r ≤ , (2.1)
dr2 r dr
 2 
d c 2 dc
(DA + DB ) + + Q(r − σ) = 0, for r ≥ , (2.2)
dr2 r dr

where Q(r − σ) is a Dirac-like distribution describing the creation of molecules at

r = σ. Let c∞ be the concentration of molecules of A in the bulk, i.e.

lim c(r) = c∞ . (2.3)

r→∞
4 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

To analyze (2.1)–(2.2), we define the following dimensionless quantities


λ k1 r c
β= , κ= , r̂ = , ĉ = , (2.4)
DA + DB  (DA + DB )  c∞

which means that we scale lengths with  and times with 2 (DA +DB )−1 . Substituting
(2.4) into (2.1)–(2.2), we obtain

d2 ĉ 2 dĉ
+ − β 2 ĉ = 0, for r̂ ≤ 1, (2.5)
dr̂2 r̂ dr̂
d2 ĉ 2 dĉ
+ + ω δ(r̂ − α) = 0, for r̂ ≥ 1, (2.6)
dr̂2 r̂ dr̂
where ω is the rate of creation of molecules at r̂ = α. To determine ω, let us note
that the average number of molecules of A produced by the reverse step (1.4) is (at
equilibrium) equal to the average number of molecules of A destroyed by the forward
reaction (1.2), i.e. the equilibrium flux through the sphere of radius 1 is equal to
4πα2 ω. This implies

dĉ 
4πα2 ω = 4π  . (2.7)
dr̂ r̂=1
Following [31], we can relate the flux through the sphere to the macroscopic association
rate and thus the right hand side of (2.7) is equal to the dimensionless rate constant
κ of the forward reaction (1.2). Consequently, we get 4πα2 ω = κ. Substituting
κ/(4πα2 ) for ω, the general solution of (2.5)–(2.6) can be written in the following
form
a1 β r̂ a2 −β r̂
ĉ(r̂) = e + e , for r̂ ≤ 1, (2.8)
r̂ r̂
a4 κ H(r̂ − α)(r̂ − α)
ĉ(r̂) = a3 − − , for r̂ ≥ 1, (2.9)
r̂ 4πr̂α
where H denotes the Heaviside step function and a1 , a2 , a3 , a4 are real constants to be
determined. The boundary condition (2.3) at infinity in the dimensionless variables
read as follows

lim ĉ(r̂) = 1. (2.10)

r̂→∞

Using this condition, the continuity of ĉ at the origin, and the continuity of ĉ and its
derivative at r̂ = 1, we determine the constants a1 , a2 , a3 and a4 in (2.8)–(2.9). We
obtain
4πα + κ sinh βr̂
ĉ(r̂) = , for r̂ ≤ 1,
4πα β cosh β r̂
 
4πα + κ 1 tanh β κ H(r̂ − α)(r̂ − α)
ĉ(r̂) = 1− + − , for r̂ ≥ 1.
4πα r̂ β r̂ 4πr̂α

Substituting ĉ into (2.7) where 4πα2 ω = κ, we obtain

4πα (β − tanh β)
κ= (2.11)
β α − β + tanh β
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 5

(a) (b)
α>1 α≤1 0.5 α=0
α=1
0.4
σ
β
0.3
σ
  0.2

0.1
Section 2 Section 3
0 −4 −3 −2 −1 0
10 10 10 10 10
κ
Fig. 1: (a) Two cases studied in Sections 2 and 3. (b) Dimensionless parameter β
defined by (2.4) as a function of κ for α = 0 (red solid line) and α = 1 (blue dashed
line).

which is the desired relation between the measurable quantities and the model pa-
rameters. Using (1.5) and (2.4), the condition (2.11) can be equivalently expressed
in terms of the measurable rate constant k1 and diffusion constants DA , DB , and the
model parameters (binding radius , unbinding radius σ and the rate λ) as follows
    
λ λ
4πσ(DA + DB )  DA +D B
− tanh  DA +DB
k1 =      . (2.12)
λ λ λ
σ DA +DB −  DA +DB + tanh  DA +D B

Remark: If we take the limit of α → ∞ in (2.11), we obtain

lim κ = 4π(1 − β −1 tanh β). (2.13)
α→∞

This is exactly the same expression as in [13] for the original λ– model, which de-
scribes only the bimolecular reaction (1.2). However, this should not be a surprise,
since by taking the limit α → ∞, we effectively remove the reverse reaction (1.4) from
the system. Passing to the limit β → ∞ in (2.13), we obtain the relation
k1 = 4πρs (DA + DB ) (2.14)

where ρs is the radius in the Smoluchowski model of diffusion-limited reactions [13, 31].
3. The case α ≤ 1. If σ ≤ , then the equilibrium equations (2.5)–(2.6) together
with the boundary condition (2.10) at infinity have to be replaced by one equation

d2 ĉ 2 dĉ κ δ(r̂ − α)
2
+ − β 2 ĉ + = 0, for r̂ ≤ 1, (3.1)
dr̂ r̂ dr̂ 4πα2
with the boundary condition ĉ(1) = 1. This takes into account the fact that there
is no diffusive flux for r̂ > 1, i.e. ĉ(r̂) = 1 for r̂ > 1. The general solution of the
second-order ordinary differential equation (3.1) is given by
a1 β r̂ a2 −β r̂ κ H(r̂ − α) sinh(βr̂ − βα)
ĉ(r̂) = e + e − ,
r̂ r̂ 4παβ r̂
6 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

where a1 and a2 are real constants which are determined by the boundary condition
ĉ(1) = 1 and the continuity of ĉ at the origin. We obtain

4παβ + κ sinh(β − βα) sinh βr̂ κ H(r̂ − α) sinh(βr̂ − βα)

ĉ(r̂) = − . (3.2)
4πα β sinh β r̂ 4παβ r̂
Since there is no diffusive flux at r̂ = 1 at equilibrium, we have
dĉ
(1) = 0.
dr̂
Evaluating this condition for (3.2), we get

4πα (β − tanh β)
κ= (3.3)
cosh(β − βα) tanh β − sinh(β − βα)

which can be expressed in terms of the experimentally measurable quantities k1 , DA

and DB , and the model parameters , σ and λ as
    
λ λ
4πσ(DA + DB )  DA +D B
− tanh  DA +DB
k1 =          . (3.4)
λ λ λ
cosh ( − σ) DA +DB tanh  DA +DB − sinh ( − σ) DA +D B

3.1. Asymptotic behaviour. Let us consider that the binding radius  is fixed.
Since k1 , DA and DB are typically given by experiments, the dimensionless parameter
κ is a fixed nonnegative constant. Taking the limit α → 0 in (3.3), we obtain

lim κ = 4π(cosh β − β −1 sinh β). (3.5)

α→0

Since the left-hand side is a nonnegative constant and the right-hand side an increasing
function of β, we can solve (3.5) for β. We denote the unique solution of (3.5) as βc .
We have βc > 0 because the right hand side of equation (3.5) approaches zero in the
limit β → 0.
Considering typical values of the diffusion and reaction rate constants for proteins
[7], namely DA = DB = 10−6 cm2 s−1 , k1 = 106 M−1 and  = 1 nm [1], we find
that κ  8.33 × 10−3 and βc  4 × 10−2 , i.e. both κ and βc are small parameters.
Considering small β and α of order 1, the leading order term in the expansion of (3.3)
is 4πβ 2 /3 which is independent of α. Consequently, we observe that (3.5) is actually
a good approximation of (3.3) even for α of order 1. This point is illustrated in Figure
1(b), where we plot β defined in (2.4) as a function of κ for α = 0 and α = 1. As we
can see, it is only when κ becomes of order 1 that the rate β calculated with (3.3)
slightly differs from the one calculated using (3.5). This implies that there exist a
realistic parameter regime for κ for which the parameter α is not influencing the value
of the removal rate β, and α can thus be set to 0 or 1. Moreover, this implies for this
particular parameter range of κ, we can completely drop the concept of the unbinding
radius σ.
4. Geminate recombination. In Sections 2 and 3, we derived formulae (2.12)
and (3.4) relating the algorithm parameters with the experimentally measurable quan-
tities. In both cases α > 1 and α ≤ 1, we have one equation for three unknowns ,
σ and λ. The binding radius  describes the range of interaction between molecules.
Postulating that  is comparable to the experimentally measurable molecular radius,
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 7

we are left with two unknowns σ and λ related by one condition (2.12) (resp. (3.4)).
Using the dimensionless parameters (2.4), we can also formulate it as one equation
(2.11) (resp. (3.3)) for two unknowns α and β. In particular, different choices of
these parameters lead to the same reaction rates. If we want to uniquely specify α
and β, we will need an extra equation. In this section, we show that different pairs of
α and β (which lead to the same reaction rates) correspond to different probability
of geminate recombination (which is properly defined in the next paragraph). This
observation can be used to find the missing relation between α and β.
When a molecule of C dissociates, one molecule of A and one molecule of B are
introduced to the system. They can have two possible fates. Either, they react again
to form the same complex C, or they diffuse away from each other. The first case
is called geminate recombination [2, 5]. We denote by φ the probability of geminate
recombination, i.e. the probability that the newly born pair of A and B reacts again.
To derive a formula relating φ, α and β, we denote by p(r̂) the probability that a
molecule of A, which is introduced in distance r̂ from a molecule of B, will react
with B before escaping to infinity. The probability p(r̂) is a continuous function with
continuous derivative. Conditioning on the first move [6, Chapter 3], it satisfies the
equations

d2 p 2 dp
+ = β 2 (p − 1), for r̂ ≤ 1, (4.1)
dr̂2 r̂ dr̂
d2 p 2 dp
+ = 0, for r̂ ≥ 1, (4.2)
dr̂2 r̂ dr̂
and the boundary condition

lim p(r̂) = 0. (4.3)

r̂→∞

Solving (4.1)–(4.3), we get

sinh(r̂β)
p(r̂) = 1 − , for r̂ ≤ 1,
r̂ β cosh β
β − tanh β
p(r̂) = , for r̂ ≥ 1.
r̂ β
Whenever the reverse reaction (1.4) takes place, the initial separation of molecules of
A and B is equal to α (in dimensionless variables). Consequently, the probability φ
of geminate recombination is given as φ = p(α), i.e.

sinh(αβ)
φ=1− , for α ≤ 1, (4.4)
αβ cosh β
β − tanh β
φ= , for α ≥ 1. (4.5)
αβ
If a modeller wants to design an algorithm with a given value of the probability φ
of geminate recombination, then equations (4.4), (4.5) will give the second condition
relating the parameters α and β. The first one is (2.11) (resp. (3.3)).
4.1. Asymptotic behaviour. As we observed in Section 3.1, realistic parame-
ters for protein-protein interactions lead to a small value of the dimensionless param-
eter β. In particular, the second condition relating α and β is not needed because
8 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

(a) (b)
−4
x 10 −2
12 10
(4.7)
(4.6)
−3
10
10

φ φ 10
−4

8
−5
10

6
−6
10

4 −7
10

−8
2 10 −1 0 1 2 3 4
0 0.5 1 1.5 2 10 10 10 10 10 10

α  [nm]
Fig. 2: (a) Geminate recombination probability φ as function of α. We use DA =
DB = 10−6 cm2 s−1 ,  = 1 nm and k1 = 106 M−1 . (b) Comparison of the geminate
recombination probability φ calculated by (4.6) and (4.7). We use DA = DB = 10−6
cm2 s−1 and k1 = 106 M−1 .

different values of α lead to the same results. Considering the same parameters as
in Section 3.1, we plot the geminate recombination probability φ as a function of the
dimensionless ratio α in Figure 2(a). To compute this plot, we use (2.12) or (3.4) to
calculate β for a given value of α. Then we calculate φ using (4.4),(4.5). In Figure
2(a), we observe that the probability φ of geminate recombination is close to zero for
all values of α. If α = 0, then (4.4) implies
1
φ=1− , (4.6)
cosh βc

where βc satisfies (3.5). Since βc  1, equations (4.6) and (3.5) give

1 2 4πβc2
φ≈ β , and κ≈ .
2 c 3
Combining these two equations we obtain φ = 3κ/(8π). Substituting (2.4) for κ, we
get
3ρs
φ= (4.7)
2
where ρs is the reaction radius corresponding to the Smoluchowski model given by
(2.14). In Figure 2(b), we plot the geminate recombination probability φ as a function
of  for α = 0. We use the same values of DA , DB and k1 as in Figure 2(a) and we
vary  from 1Å (0.1 nm) to thousands of nanometers. We observe that the formula
(4.6) (together with (3.5)) gives the same geminate recombination probability φ as
the approximation (4.7). Finally, let us note that by taking the limit β → ∞ in (4.5),
we obtain φ = α−1 = /σ, which is the expression for the geminate recombination
probability used in [5].
5. Stochastic simulation algorithm for large time steps. To implement λ-
 model on a computer, we have to discretize (1.1) using a finite time step ∆t. Using
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 9

the Euler-Maruyama method [27, 14], the position [Xi (t + ∆t), Yi (t + ∆t), Zi (t + ∆t)]
of the i-th molecule at time t + ∆t is computed from its position [Xi (t), Yi (t), Zi (t)]
at time t by


Xi (t + ∆t) = Xi (t) + 2Di ∆t ξx ,


Yi (t + ∆t) = Yi (t) + 2Di ∆t ξy , (5.1)


Zi (t + ∆t) = Zi (t) + 2Di ∆t ξz ,

where ξx , ξy , ξz are random numbers which are sampled from the normal distribution
with zero mean and unit variance. If ∆t is “very small”, then the computer implemen-
tation of the reversible reaction (1.3) is straightforward. We use (5.1) to update the
position of every molecule using Di = DA for molecules of A, Di = DB for molecules
of B and Di = DC for molecules of C. Whenever the distance of a molecule of A from
a molecule of B is less than the reaction radius , the molecules react according to
the forward reaction (1.2) with probability Pλ = λ ∆t. The probability of the reverse
reaction (1.4) during one time step is equal to k2 ∆t. If the complex C dissociates,
then we introduce one molecule of A and one molecule of B in a distance σ apart.
This computer implementation of the reversible reaction (1.3) will only work if
the time step ∆t is chosen so small that Pλ = λ ∆t  1, k2 ∆t  1 and γ  1, where
γ is given by


2(DA + DB )∆t
γ= , (5.2)


i.e. γ is the ratio of the average step size in one coordinate during one time step
over the reaction radius . In this section, we show how the restrictions on the time
step ∆t can be removed. First of all, the probability that the complex C dissociates
during the time interval (t, t+∆t) is equal to 1−exp(−k2 ∆t), i.e. the reverse reaction
(1.4) is easy to implement for arbitrary time step ∆t. We simply use 1 − exp(−k2 ∆t)
instead of k2 ∆t as the probability of dissociation of C during one time step. To relax
the restrictions γ  1 and Pλ = λ ∆t  1, we slightly reformulate the algorithm [13].
As before, it will make use of three parameters: the reaction radius , the unbinding
radius σ and the reaction probability Pλ of the forward reaction (1.2). We postulate
that a molecule of A and a molecule of B (which are closer than the reaction radius )
react with probability Pλ ∈ (0, 1] during the next time step. Therefore, the computer
implementation of the reversible reaction (1.3) will make use of the following three
steps:
[i] If the distance of a molecule of A from a molecule of B (at time t) is less than
the reaction radius , then generate a random number r1 uniformly distributed
in (0,1). If r1 < Pλ , then the forward reaction (1.2) occurs, i.e. the molecules of
A and B are removed from the system and a new molecule of C is created.
[ii] For each molecule of C, generate a random number r2 uniformly distributed in
(0,1). If r2 < 1 − exp(−k2 ∆t), then the reverse reaction (1.4) takes place, i.e.
the complex C dissociates, and one molecule of A and one molecule of B are
introduced a distance σ apart.
[iii] Use (5.1) to update the position of every molecule.
The steps [i]–[iii] are repeated during every time step. In order to use this algorithm,
we need to find equations relating parameters , σ and Pλ with the experimentally
measurable quantities. If ∆t is small, then one condition is given as (2.12) (resp.
10 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

(3.4)) where Pλ = λ ∆t  1. However, if Pλ is close to 1, we have to modify the

derivation of these conditions, replacing partial differential equations (2.1)–(2.2) by
suitable integral equations [13, 12].
First of all, the conditions depend on the ordering of steps [i]–[iii], i.e. on the
ordering of subroutines of the algorithm. Consider the case Pλ = 1 and α = σ/ < 1.
If we ordered the subroutines as [ii], [i] and [iii], then each dissociation of a complex C
in step [ii] would introduce two new molecules of A and B which are a distance σ apart.
Since [ii] would be immediately followed by [i], the new molecules would have to react
again because Pλ = 1 and their separation is less than . In particular, there would
be no chance to correctly implement this model for Pλ = 1 and α = σ/ < 1. On the
other hand, if we order the subroutines as [i], [ii] and [iii], then the dissociation of C
is followed by diffusion of molecules, i.e. the new molecules of A and B can diffuse
away of each other. In the rest of this paper, we assume that the subroutines are
ordered as [i], [ii] and [iii] during each time step. Finally, we also have to specify the
position of a new molecule of C which is created by the forward reaction in the step
[i]. A natural initial condition for C is the centre of mass of reactants A and B. In
our illustrative simulations, we will consider that reactants A and B have the same
mass which means that the molecule of C is placed halfway between the reactants.
In a similar way, we initialize positions of molecules A and B created by the reverse
reaction in the step [ii]. We will place them a distance σ apart so that their centre of
mass is equal to the position of complex C before it dissociates. The initial direction
of the line connecting molecules A and B is chosen randomly.
As in the case of (2.1)–(2.2), we consider a frame of reference situated in the
molecule of B, i.e. molecules of A diffuse with the diffusion constant DA + DB and
are removed in the ball around origin with probability Pλ during each time step. We
use the dimensionless parameters given by (1.5), (2.4) and (5.2). Let ck (r̂) be the
concentration of molecules of A at the distance r̂ from the origin. Each step of the
algorithm changes the concentration which can be schematically described as follows:

[i] [i] [ii] [ii] [iii]

ck (r̂) −→ ck (r̂) −→ ck (r̂) −→ ck+1 (r̂),

[i] [ii]
where ck (r̂) (resp. ck (r̂)) is a concentration at the distance r̂ from the origin after
step [i] (resp. [ii]). Using the definition of steps [i]–[iii], we find
[i]
ck (r̂) = (1 − Pλ )χ[0,1] (r̂)ck (r̂) + χ(1,∞) (r̂)ck (r̂), (5.3)
[ii] [i]
ck (r̂) = ck (r̂) + ω δ(r̂ − α), (5.4)
∞
[ii]
ck+1 (r̂) = K(r̂, r̂ , γ)ck (r̂ ) dr̂ , (5.5)
0

where ω is a constant describing the production of molecules of A in one time step

and K(z, z  , γ) is a Green’s function for the diffusion equation given by
 
z (z − z  )2 (z + z  )2
K(z, z , γ) = √ exp − − exp − .
zγ 2π 2γ 2 2γ 2

Substituting (5.3) and (5.4) in (5.5), we obtain

1 ∞
ck+1 (r̂) = (1 − Pλ ) K(r̂, r̂ , γ)ck (r̂ ) dr̂ + K(r̂, r̂ , γ)ck (r̂ ) dr̂ + ω K(r̂, α, γ).
0 1
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 11

We are interested to find the fixed point g(r̂) of this iterative scheme [13]. At steady
state, the mass lost in (5.3) is equal to the mass added in (5.4), i.e. 4πα2 ω =
1
Pλ 0 g(z)4πz 2 dz. Consequently, g(r̂) satisfies the following equation
1 ∞
g(r̂) = (1 − Pλ ) K(r̂, r̂ , γ)g(r̂ ) dr̂ + K(r̂, r̂ , γ)g(r̂ ) dr̂
0 1
1
Pλ K(r̂, α, γ)
+ g(z)z 2 dz. (5.6)
α2 0

Then the rate of removing of particles during one time step is

1
κ = Pλ 4πz 2 g(z)dz. (5.7)
0

where κ is the dimensionless reaction rate given by

k1 ∆t
κ= . (5.8)
3

It is worth noting that κ is defined with the help of the time step ∆t and it is
therefore different from κ defined by (2.4). In Figure 3, we plot κ as a function of γ,
for different values of probability Pλ and ratio α. Figure 3(a) is calculated for Pλ = 1,
which corresponds to the Andrews and Bray model [5]. Panels (b), (c) and (d) in
Figure 3 correspond to Pλ = 0.75, Pλ = 0.5 and Pλ = 0.25, respectively. In each
panel, the κ-γ curves are plotted for the values of ratio α equal to 0, 0.5, 0.7, 0.8,
0.9, 1, 1.6, 2.5, 4, 6.3 and 10, starting always from the top in each panel. To solve
equation (5.6) numerically, we first truncate the integrals to the finite domain [13].
This means that we choose a large real number S and we approximate g(r̂) ≡ 1 for
r̂ > S. Then the equation (5.6) reads as follows
1 S
g(r̂) = (1 − Pλ ) K(r̂, r̂ , γ)g(r̂ ) dr̂ + K(r̂, r̂ , γ)g(r̂ ) dr̂
0 1
1 ∞
Pλ K(r̂, α, γ)
+ g(z)z 2 dz + K(r̂, r̂ , γ) dr̂
α2 0 S

Then we use the Simpson rule to discretize this integral equation and we obtain a
system of linear equation which is solved numerically.
5.1. Probability of geminate recombination. In Figure 3, we observe that
there exist various combinations of the parameters γ, Pλ and α for which we obtain
the same value of the dimensionless reaction rate κ. Even if we fix γ (which is, roughly
speaking, equivalent to choosing the time step ∆t), there are still different choices of
pairs Pλ and α which lead to the same reaction rate. For example, κ = 1 and γ = 0.5
can be achieved both for Pλ = 0.5, α = 4.4258 and Pλ = 0.25, α = 0.8887. As we
observed in Section 4, one possible way to distinguish different sets of parameters is
by studying the geminate recombination probability.
To compute φ, we study the following auxiliary problem. We consider a molecule
escaping by discretized diffusion from the disk of radius 1 centered at the origin. This
disk can capture the particle with probability Pλ at each time step. Let p(r̂) be the
probability that a molecule starting at r̂ is captured before it escapes to infinity. It
12 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

(a) (b)
6 6
α=0 α=0
α=1 α=1
5 5

4 4
κ κ
3 3

2 2

1 1

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
γ γ
(c) (d)
6 6
α=0 α=0
α=1 α=1
5 5

4 4
κ κ
3 3

2 2

1 1

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
γ γ
Fig. 3: Relation of κ and γ for different values of α and Pλ : (a) Pλ = 1; (b) Pλ = 0.75;
(c) Pλ = 0.50; (d) Pλ = 0.25.

satisfies the equation

1 1 ∞
p(r̂) = Pλ K(r̂, r̂ , γ) dr̂ +(1−Pλ ) K(r̂, r̂ , γ) p(r̂ ) dr̂ + K(r̂, r̂ , γ) p(r̂ ) dr̂ ,
0 0 1
(5.9)
with the boundary condition

lim p(r̂) = 0.
r̂→∞

The probability of geminate recombination is given as φ = p(α). Solving (5.9) numer-

ically, we find that the geminate recombination probability is φ = 0.12 for the first
case (Pλ = 0.5, α = 4.4258) and φ = 0.38 for the second case (Pλ = 0.25, α = 0.8887)
which is a significant difference.
Another possibility to reduce the number of algorithm parameters is by consid-
ering the values of realistic measurable parameters for a particular application. This
will be shown in the following section for the case of proteins.
6. Illustrative Brownian dynamics results. In the previous sections, we
derived relations between the algorithm parameters , σ, λ (resp. Pλ ) and the ex-
perimentally measurable quantities. In this section, we illustrate our results using a
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 13

(a) (b)
0.025 0.3
α=0

stationary distribution
α=1
0.02 0.25
Pλ
0.2
0.015
0.15
0.01
0.1

0.005
0.05

0 −4 −3 −2 −1
0
10 10 10 10 0 2 4 6 8 10
κ number of molecules of A
Fig. 4: (a) Pλ as a function of κ for α = 0 and α = 1, and γ = 1. (b) Stationary
distribution of molecules of A computed by the Brownian dynamics simulation for
Pλ = 9.95 × 10−4 and α = 1 (grey histogram). Results obtained by the Gillespie
algorithm (red dots).

simple toy problem. We will consider a cubic reactor of the size L × L × L where
L = 50 nm. In the reactor, there are molecules of three chemical species A, B and
C which are subject to the reversible reaction (1.3). The molecules diffuse inside the
reactor. The boundary of the reactor is considered to be non-reactive (reflective) and
we start with 5 molecules of each species in the domain.
Using typical diffusion constants of proteins DA = DB = DC = 10−6 cm2 s−1 ,
the reaction radius ρ̄ = 1 nm and the time step ∆t = 14 10−8 s, we obtain that the
dimensionless parameter γ defined by (5.2) is γ = 1. Considering that typical rate
constants of protein-protein interactions are about 106 M−1 s−1 , we obtain that the
dimensionless parameter κ is of the order 10−3 . In Figure 4(a), we plot the dependence
of the probability Pλ as a function of κ for α = 0 and α = 1 in the case where γ = 1.
Note that in the case α = 0, the equation (5.6) becomes
1 ∞
g(r̂) = (1 − Pλ ) K(r̂, r̂ , γ)g(r̂ ) dr̂ + K(r̂, r̂ , γ)g(r̂ ) dr̂
0 1
   1
Pλ 2 r̂2
+ exp − 2 g(z)z 2 dz.
4πγ 3 π 2γ 0

As we can see, the probability Pλ appears to be independent of α for this particular

parameter range of κ. We thus set α = 1, i.e. σ = . We use k1 = 106 M−1 s−1 and
k2 = 66.7 s−1 . Then equations (5.6)–(5.7) imply that Pλ = 9.95 × 10−4 and we can
use the steps [i]–[iii] to simulate the illustrative toy model. If the diffusive step [iii]
places a molecule outside the reactor, we return it back using mirror reflection. This
is a typical way to implement no-flux boundary conditions. For discussion of more
complicated boundary conditions, see [12].
To visualize the results of stochastic simulation, we compute the stationary dis-
tribution of the numbers of molecules of A in the whole reactor as follows. We run the
simulation for a long time and we record the number of molecules of A at equal time
intervals. The resulting (grey) histogram is plotted in Figure 4(b). Since the domain
is relatively small, we can make a direct comparison with the stationary histogram
14 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

(a) (b)
0.5 0.4
γ=1
γ=1 ∆t fixed
0.35
0.4
φ  0.3

0.3
0.25
γ>1

0.2
0.2

0.15
0.1
0.1
γ<1
0 0.05
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Pλ Pλ
Fig. 5: (a) Comparison of (5.9) with the spatial stochastic simulations algorithm for
γ = 1. (b) Dependence of  on Pλ calculated for γ = 1 (red solid line) and for
∆t = 33 × 10−4 s (blue dashed line).

obtained by the (spatially-homogeneous, well-mixed) simulation of the reversible re-

action (1.3) by the Gillespie SSA [17], which is equivalent to solving the corresponding
chemical master equation. The results are plotted as red circles in Figure 4(b). As
expected, the comparison with the Brownian dynamics (spatial stochastic simulation)
is excellent.
6.1. Geminate recombination. In our second illustrative example, we use the
stochastic simulation of λ- model to directly validate our formulae for geminate
recombination. We simulate the behaviour of molecules of A, B and C in the cubic
reactor as before. Whenever two molecules of A and B are introduced in the system,
we check if they react with each other again before reacting with another molecule or
hitting the boundary of the reactor. We then approximate the geminate recombination
probability, by the ratio of geminate recombination events over the total number of
forward reactions (1.2) occurring in the simulation.
Solving (5.9) for the parameters used in Figure 4(b), we find that φ = 5.95 × 10−4
which is negligible. In order to illustrate the strength of the formula (5.9), we will
use different parameter values for which the gemination combination probability is
significant, namely DA = DB = DC = 1 µm2 sec−1 , rate constants k1 = 1 µm3 sec−1 ,
k2 = 0.005 sec−1 , L = 20 µm, α = 0, γ = 1 and different values for the probability Pλ .
In Figure 5(a), we compare the results obtained by (5.9) with the results estimated
from the Brownian dynamics simulations (red circles). The comparison is very good.
We also plot the results estimated from the same stochastic simulation showing how
often the molecules of A and B which were created from the same complex C react with
each other (blue squares). The difference between the (red) circles and (blue) squares
is that in the former we do not consider the event to be a geminate recombination
if either of the molecules of A or B has hit the domain boundary, before they react
again with each other. Thus (blue) squares give an upper estimate of the geminate
recombination given by (5.9), because we have finite number of molecules in the box
(on average only 5 molecules). In particular the blue squares would approach the
theory and the red circles for simulations of a large number of molecules.
In Figure 5(a), we fixed the value of γ as 1, since this is the value for which the
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 15

spatial stochastic simulation algorithm discussed in Section 5 is the most relevant. In

particular, every time we change the probability Pλ , we also change the time step ∆t
and the reaction radius . In Figure 5(b), we present the dependence of the binding
radius  on Pλ . Each point on this curve corresponds to a different time step. Another
option to compare the results would be to choose ∆t to be fixed for all the different
probabilities Pλ , which means that γ would have to be different in every simulation.
The dependence of the binding radius  on Pλ for fixed ∆t is also plotted in Figure
5(b) for comparison. We choose ∆t = 33 × 10−4 s, which is the value for which γ = 1,
when Pλ = 0.5. As we can see in both cases the binding radius  is a decreasing
function of the probability Pλ . When we keep ∆t fixed,  decreases slower than it
does in the case of fixed γ, which implies that γ in this case of fixed ∆t becomes
smaller than 1 as Pλ gets smaller than 0.5.

7. Discussion. Several algorithms for stochastic simulation of reaction-diffusion

processes in cell and molecular biology have been proposed in the literature. Some
of these methods are lattice-based and can be equivalently described in terms of the
reaction-diffusion master equation (RDME) [18, 20]. Approaches to simulate RDME-
based models efficiently have been recently proposed [10, 16] and the RDME methods
were generalized to unstructured meshes [11], but other open questions remain. For
example, the relation of RDME to more detailed off-lattice models [19, 13] and efficient
ways to investigate the dependence of simulation results on the model parameters, e.g.
efficient bifurcation analysis of stochastic models [29].
In this paper, we studied an alternative approach to stochastic reaction-diffusion
modelling. We presented a class of Brownian dynamics algorithms. These algorithms
are off-lattice and can, in principle, provide more details. However, they share some
problems with the RDME-based simulations, e.g. all stochastic models are usually
more computationally intensive than solving the corresponding deterministic reaction-
diffusion partial differential equations. One way to decrease the computational inten-
sity is to consider Brownian dynamics of point-like particles [5]. In [13], we presented
λ- approach which provides more flexibility in choosing the reaction radius  than
one-parameter based models. In this paper, we show that this approach can be gen-
eralized to the case of reversible reactions, addressing the criticism mentioned in the
recent paper describing the Smoldyn algorithm [4] (page 5). In particular, we show
that, in the parameter regime relevant to protein-protein simulation, it is possible to
avoid the concept of the unbinding radius σ. We illustrate that the same results can
be obtained for σ = 0 and for σ = . If we consider smaller reaction radii or larger
reaction rates, then the unbinding radius has to be taken into account. We derived
formulae for the probability φ of geminate recombination which can be used to select
the appropriate algorithm parameters. For example, given reaction rate constants k1
and k2 and parameters , ∆t and φ, we can use formulae (5.7) and (5.9) to compute
the algorithm parameters α and Pλ . In particular, we also generalized the results
of Andrews and Bray [5] (which were derived for Pλ = 1), to the case of arbitrary
reaction probability Pλ ∈ (0, 1]. It is worth noting that the RDME-based approaches
do not have special difficulties with simulating reversible reactions, because they can
be implemented as two reactions (1.2) and (1.4) in a straightforward way.
The reversible reaction (1.3) was previously investigated in the literature from
both theoretical and computational view points. In [28], Popov and Agmon studied
classical Brownian dynamics model which corresponds to the limit λ → ∞ in the
λ- model. They test different approximations for the binding probability by three-
dimensional numerical simulations. One method to increase the efficiency of their
16 J. LIPKOVÁ, K. ZYGALAKIS, J. CHAPMAN, R. ERBAN

numerical algorithm is by using the “dynamic time step”. This means that trajectories
of molecules which are not surrounded by other reactants can be simulated over longer
time steps. Considering an isolated pair of reaction molecules, the corresponding
Green function can be calculated explicitly [21, 22] and can be used to guide the
selection of the next time steps. This idea is also used in Green’s function reaction
dynamics [34]. In this manuscript, we focussed on algorithms with fixed time steps
(but with different values of λ). We leave the dynamic time step models for future
investigations.
Bimolecular reactions are very common in cell biology [25, 1] and therefore, it
is important to study their correct implementation in the computational algorithms
[13]. However, there are several other issues which needs to be considered in order
to simulate realistic spatially-distributed reaction-diffusion systems [33]. Brownian
dynamics require extra attention when simulating reactive boundaries (e.g. reactions
on the plasma membrane) [12, 3] and one should also have in mind steric interactions,
i.e. the consequences of macro-molecular crowding inside the cytoplasm [1]. We will
address this issue in a future publication.

Acknowledgements. This publication is based on work (JL, KZ, RE) supported

by Award No. KUK-C1-013-04, made by King Abdullah University of Science and
Technology (KAUST). RE was also supported by the European Research Council
Starting Independent Researcher Grant and by Somerville College, Oxford.

REFERENCES

[1] B. Alberts, A. Johnson, J. Lewis, M. Raff, K. Roberts, and P. Walter, Molecular Biology
of the Cell, Garland Science, New York, 2002.
[2] S. Andrews, Serial rebinding of ligands to clustered receptors as exemplified by bacterial
chemotaxis, Physical Biology, 2 (2005), pp. 111–122.
[3] , Accurate particle-based simulation of adsorption, desorption and partial transmission,
Physical Biology, 6 (2009), p. 046015.
[4] S. Andrews, N. Addy, R. Brent, and A. Arkin, Detailed simulations of cell biology with
smoldyn 2.1, PLOS Computational Biology, 6 (2010), p. e1000705.
[5] S. Andrews and D. Bray, Stochastic simulation of chemical reactions with spatial resolution
and single molecule detail, Physical Biology, 1 (2004), pp. 137–151.
[6] H. Berg, Random Walks in Biology, Princeton University Press, 1983.
[7] D. Brune and S. Kim, Predicting protein diffusion coefficients, Proceedings of the National
Academy of Sciences USA, 90 (1993), pp. 3835–3839.
[8] B. Corry, S. Kuyucak, and S. Chung, Test of continuum theories as models of ion chan-
nels. II. Poisson-Nernst-Planck theory versu Brownian dynamics, Biophysical Journal, 78
(2000), pp. 2364–2381.
[9] M. Doi, Stochastic theory of diffusion-controlled reaction, Journal of Physics A: Mathematical
and General, 9 (1976), pp. 1479–1495.
[10] B. Drawert, M. Lawson, L. Petzold, and M. Khammash, The diffusive finite state projec-
tion algorithm for efficient simulation of the stochastic reaction-diffusion master equation,
Journal of Chemical Physics, 132 (2010), p. 074101.
[11] S. Engblom, L. Ferm, A. Hellander, and P. Lötstedt, Simulation of stochastic reaction-
diffusion processes on unstructured meshes, SIAM Journal on Scientific Computing, 31
(2009), pp. 1774–1797.
[12] R. Erban and S. J. Chapman, Reactive boundary conditions for stochastic simulations of
reaction-diffusion processes, Physical Biology, 4 (2007), pp. 16–28.
[13] , Stochastic modelling of reaction-diffusion processes: algorithms for bimolecular reac-
tions, Physical Biology, 6 (2009), p. 046001.
[14] R. Erban, S. J. Chapman, and P. Maini, A practical guide to stochastic simulations of
reaction-diffusion processes. 35 pages, available as [Link] 2007.
[15] D. Ermak and J. McCammon, Brownian dynamics with hydrodynamic interactions, Journal
of Chemical Physics, 69 (1978), pp. 1352–1360.
 BROWNIAN DYNAMICS OF REVERSIBLE BIMOLECULAR REACTIONS 17

[16] L. Ferm, A. Hellander, and P. Lötstedt, An adaptive algorithm for simulation of stochastic
reaction-diffusion processes, Journal of Computational Physics, 229 (2010), pp. 343–360.
[17] D. Gillespie, Exact stochastic simulation of coupled chemical reactions, Journal of Physical
Chemistry, 81 (1977), pp. 2340–2361.
[18] J. Hattne, D. Fange, and J. Elf, Stochastic reaction-diffusion simulation with MesoRD,
Bioinformatics, 21 (2005), pp. 2923–2924.
[19] S. Isaacson, The reaction-diffusion master equation as an asymptotic approximation of diffu-
sion to a small target, SIAM Journal on Applied Mathematics, 70 (2009), pp. 77–111.
[20] S. Isaacson and C. Peskin, Incorporating diffusion in complex geometries into stochastic
chemical kinetics simulations, SIAM Journal on Scientific Computing, 28 (2006), pp. 47–
74.
[21] H. Kim and K. J. Shin, Exact solution of the reversible diffusion-influenced reaction for an
isolated pair in three dimensions, Phys. Rev. Lett., 82 (1999), pp. 1578–1581.
[22] H. Kim, M. Yang, and K. J. Shin, Dynamic correlation effect in reversible diffusion-influenced
reactions: Brownian dynamics simulation in three dimensions, The Journal of Chemical
Physics, 111 (1999), pp. 1068–1075.
[23] R. Larson, H. Hu, D. Smith, and S. Chu, Brownian dynamics simulations of a DNA molecule
in an extensional flow field, Journal of Rheology, 43 (1999), pp. 267–303.
[24] K. Lipkow, S. Andrews, and D. Bray, Simulated diffusion of phosphorylated CheY through
the cytoplasm of Escherichia coli, Journal of Bacteriology, 187 (2005), pp. 45–53.
[25] H. Marianayagam, M. Sunde, and J. Matthews, The power of two: protein dimerization in
biology, Trends in Biochemical Sciences, 29 (2004), pp. 618–625.
[26] S. Northrup and H. Erickson, Kinetics of protein-protein association explained by Brownian
dynamics computer simulation, Proceedings of the National Academy of Sciences USA, 89
(1992), pp. 3338–3342.
[27] E. Platen, An introduction to numerical methods for stochastic differential equations, Acta
Numerica, 8 (1999), pp. 197–246.
[28] A. V. Popov and N. Agmon, Three-dimensional simulations of reversible bimolecular reac-
tions: The simple target problem, Journal of Chemical Physics, 115 (2001), pp. 8921–8932.
[29] L. Qiao, R. Erban, C. Kelley, and I. Kevrekidis, Spatially distributed stochastic sys-
tems: Equation-free and equation-assisted preconditioned computation, Journal of Chemi-
cal Physics, 125 (2006), p. 204108.
[30] C. Siettos, M. Graham, and I. Kevrekidis, Coarse Brownian dynamics for nematic liquid
crystals: Bifurcation, projective integration, and control via stochastic simulation, Journal
of Chemical Physics, 118 (2003), pp. 10149–10156.
[31] M. Smoluchowski, Versuch einer mathematischen Theorie der Koagulationskinetik kolloider
Lösungen, Zeitschrift für physikalische Chemie, 92 (1917), pp. 129–168.
[32] E. Teramoto and N. Shigesada, Theory of bimolecular reaction processes in liquids, Progress
of Theoretical Physics, 37 (1967), pp. 29–51.
[33] F. Tostevin, P. ten Wolde, and M. Howard, Fundamental limits to position determination
by concentration gradients, PLOS Computational Biology, 3 (2007), pp. 763–771.
[34] J. van Zon and P. ten Wolde, Green’s-function reaction dynamics: a particle-based approach
for simulating biochemical networks in time and space, Journal of Chemical Physics, 123
(2005), p. 234910.