You are on page 1of 15

Computers and Chemical Engineering 27 (2003) 235 /249

www.elsevier.com/locate/compchemeng

Simulation of population dynamics using continuous-time finite-state


Markov chains
K. Karen Yin a,*, Hongchuan Yang b, Prodromos Daoutidis c, G. George Yin d
a
Department of Wood and Paper Science, University of Minnesota, St. Paul, MN 55108, USA
b
Department of Electrical Engineering, University of Minnesota, Minneapolis, MN 55455, USA
c
Department of Chemical Engineering and Material Science, University of Minnesota, Minneapolis, MN 55455, USA
d
Department of Mathematics, Wayne State University, Detroit, MI 48202, USA

Received 16 November 2001; accepted 23 August 2002

Abstract

This paper is concerned with the simulation of certain types of population dynamics using Markov chains. A brief review of the
Markovian property, the Chapman /Kolmogorov equation, and the forward equation together with its solution is given. The
essentials in simulating continuous-time Markov chains are provided and two case studies are presented for demonstration. The first
is a process of drug delivery for which a closed-form solution of the forward equation can also be obtained. The second one deals
with cell population dynamics, for which a Markovian model capturing birth, death, and growth is developed and simulated. The
key to such simulation studies is to specify the generator Q , which in turn requires a thorough understanding of the properties of the
underlying system. In return, one obtains not only the means and variances, but also the entire probability distributions, and the
evolution of the random processes of interest.
# 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Markov chain; Forward equation; Infinitesimal generator; Population dynamics; Particulate processes

1. Introduction Markov chains, a well-known subject introduced by


Markov in 1906, has been studied by a host of
Population dynamics are concerned with the evolu- researchers for many years (see the classical books of
tion of size and/or other properties of a population with Chung, 1960; Doob, 1953; see also Khasminskii & Yin,
time, and its dependence on the initial and environ- 1996; Yin & Zhang, 1998; Yin, Zhang, Yang & Yin,
mental conditions. The study of population dynamics 2001 for some recent results of related works). Stochas-
has been attracting growing attention in the biological, tic models and Markovian formulations have been
ecological, and medical sciences, in chemical and applied to numerous real-world systems under uncer-
biochemical engineering, and in other fields (see, e.g. tainties (see, e.g. Berthiaux, 2000; Burge & Karlin, 1997;
Ramkrishna, 2000 for a recent broad treatment of Chiang, 1980; Durbin, Eddy, Krogh & Mitchison, 1998;
population balance modeling and analysis for particu- Fan, Nassar, Hwang & Chou, 1985; Nassar, Schmidt &
late processes). In this paper, we illustrate how Markov Luebbert, 1992; Taylor & Karlin, 1998; Yin & Zhang,
chains can be used as the basis for modeling and 1998; Yin, Liu & Johnson, 2002 and the references
simulating certain types of population dynamics of therein). Their applications in genetics for finding
interest to engineers. genotype distribution of certain generation given the
parent genotype distribution, in the research of relapse,
recovery and death for cancer patients by modeling
* Corresponding author. Tel.: /1-612-624-1761; fax: /1-612-625- birth, death, and illness processes of cancer cells, in
6286
E-mail addresses: kyin@umn.edu (K.K. Yin), hyang@ece.umn.edu
computational molecular biology for gene identification
(H. Yang), daoutidi@cems.umn.edu (P. Daoutidis), gyin@math. and prediction, and in the studies of discontinuous
wayne.edu (G.G. Yin). media such as solid mixing, multiphase flow, bubble
0098-1354/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 1 3 5 4 ( 0 2 ) 0 0 1 7 9 - 5
236 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

behavior in the gas /solids fluidized bed, particle diffu- dated; the essentials in simulating a continuous-time
sivity and mixing coefficient in a liquid/solid fluidized Markov chain are also provided. Two illustrative
bed are a few such examples. Once a Markovian model examples are presented in the next two sections. The
is established, It can be used to characterize the mean problem formulation, numerical scheme, and simulation
behavior of the underlying system, or more importantly results are presented in Section 3 for the drug delivery
to fully explore its dynamics (this is important especially example, and in Section 4 for the cell population
for non-Gaussian stochastic processes, where the mean example. Summary and further discussion are given in
and variance alone do not contain enough information). Section 5.
The latter task requires in principle the solution of the
so-called forward equation which provides the evolution
of the probability distribution, This solution is a matrix
exponential, involving the so-called generator, which,
2. Markovian property and Kolmogorov equations
however, can be obtained in closed form only for simple
special cases. Alternatively, for continuous-time Markov
A stochastic process (see e.g. Chiang, 1980; Chung,
chains one can numerically generate sample paths of the
1960; Doob, 1953; Karlin & Taylor, 1975) can be
Markov chain (from which probability distributions and
considered as a collection of random variables {a (t,
various statistics can be obtained) based only on the
v )} (or {a (t)}) defined on a common probability space
specification of the generator and the initial conditions
and indexed by the parameter t  /T , where T is a
(see Taylor & Karlin, 1998 for more detailed descrip-
suitable set and the index t often represents time. For
tions).
a fixed t, a (t , ×/) is a random variable. For each v , a ( ×/,
In this paper, we consider two case studies for which
v ) is called a sample path or a realization of the process.
we illustrate the formulation of continuous-time finite-
Note that the parameter v is often omitted for brevity.
state Markov chains and their numerical simulation.
The state space of a (t) is the collection of all values it
The first example concerns a process of drug delivery via
may take.
the gastric and intestinal tract, for which a stochastic
Markov chains are stochastic processes which satisfy
multi-compartmental model (Chen, Chen, Hikal, Shen
the so-called Markovian property, i.e. for which given
& Fan, 1998) was developed and the master equation
the current state, the probability of the chain’s future
was derived, yielding the means and covariances of the
behavior is not altered by any additional knowledge of
random variables of interest. Because of its low dimen-
its past behavior. Suppose that a (t) is a continuous-time
sion and the special form of its generator, a closed-form
stochastic process taking values in the set M/{1, . . .,
solution of the forward equation of this process is easily
m } (i.e. a finite state space). Taking the index set T to be
obtainable, and will be presented for comparison. Our
T /[0, ), the Markov property states that for each i  /
emphasis though is on simulating the piecewise constant
M, the following conditional probability relation holds:
sample paths of the corresponding Markov chains. Such
an approach can provide us with single realizations of P(a(t)i½a(r):r5s)P(a(t)i½a(s)): (1)
the random processes as well as important statistics, and
yield the probability distributions of the random vari- Let a (t) be a Markov chain with state space M and the
ables of interest at any given time. It is particularly transition probabilities Pij (t) /P {a (t/s) /j ja (s) /i}
useful in cases where there are no closed-form solutions for all s ]/0 be independent of s; then we say that the
in the forward equation, which is typical in high- Markov chain has stationary transition probabilities,
dimensional random processes. The second case study the corresponding Markov chain is often referred to as
is such an example that concerns cell populations in an being stationary or homogeneous, and the transition
environment of changing substrate concentration. De- matrix P (t)  /Rm m can be written as:
terministic formulations of corpuscular models of such 0 1
systems (i.e. cell population balance models) lead to P1;1 (t) P1;2 (t) . . . P1;m1 (t) P1;m (t)
systems of partial-integro-differential equations whose BP2;1 (t) P2;2 (t) . . . P2;m1 (t) P2;m (t) C
P(t) B@n :: C: (2)
A
solution is a challenging task (see Mantzaris, Daoutidis n : n n
& Srienc, 2001a,b,c for recent results in this direction). Pm;1 (t) Pm;2 (t) . . . Pm;m1 (t) Pm;m (t)
In this work, we use Markov chains to model birth,
death, and growth of cells, and simulate the cell In a continuous-time and finite-state Markov chain,
population dynamics. we deal with a family of random variables {a (t):0 5/t B/
The paper is organized as follows. Section 2 briefly } whose states may be indexed by integers. For a
reviews the notion of Markov chain and Markovian finite-state Markov chain, Pij (t) ]/0, am
j1 Pij  1 for
property, the Chapman /Kolmogorov equation, and the each i  /M, and the well-known Chapman /Kolmogorov
forward equation and its solution. The need of simula- equation (see e.g. Chiang, 1980; Chung, 1960; Karlin &
tion in the study of many stochastic systems is eluci- Taylor, 1975; Taylor & Karlin, 1998) holds:
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 237

X
m Eqs. (5) /(7) are applicable to any Markov chains with
Pij (th) Pik (t)Pkj (h); (3) finite state space once the generators are identified. A
k0
thorough understanding of the physical process under
or in a more compact form study is needed to prescribe the generator Q , and to
formulate Eqs. (5) and (6). Note that Eq. (7) is merely a
P (t/h )/P (t )P (h ). representation of the solution. To solve Eq. (6) requires
That is, to move from state i to state j in an interval finding the eigenvalues of Q first and then obtaining
t/h, the chain a (t ) moves to some intermediate state k exp(Qt ). A key to this is to find the solutions of det(lI/
in the duration [0, t], and then moves from k to j in the Q )/0, where det(A ) denotes the determinant of A ,
remaining time h . If in addition we assume that which is a polynomial equation of order m . Although
limi 00 P (t)/dij where dij /1 for i/j and dij /0 this polynomial equation can be solved in some special
otherwise, then P (t) is continuous at t /0. Further- cases, there is no closed-form solution available in
more: general for m ]/5.
In this paper, we follow an alternative approach
P(t)  I (Taylor & Karlin, 1998) which numerically generates
lim Q (qij )  Rmm (4)
t00 t the sample paths of the Markov chain on the basis of the
generator Q . Note that for a continuous-time homo-
where I is the identity matrix exists, and the qij ’s satisfy:
geneous Markov chain, in view of Eqs. (5) /(7), the Q
qij (t)]/0 for j "/i and qii (t)aj"i qij (t) for t ]/0. The
‘generates’ the probabilities, so only the generator Q
matrix Q is termed an infinitesimal generator, or in
and not the transition probability matrix is needed for
short, a generator, and represents the rates of change of
the simulation (see also Chung, 1960; Taylor & Karlin,
the transitions. The transition matrix satisfies the
1998; Yin & Zhang, 1998). Specifically, suppose that
differential equation:
a (t) is a continuous-time Markov chain with state space
dP(t) M /{1, . . ., m } and generator Q /(qij ). To simulate
P(t)Q; P(0)I: (5) a (t) requires designating the state from which it starts,
dt
determining the duration of its stay and the state of its
To find the probability distribution of the Markov chain next entry, and repeating this procedure to determine
at time t, let p(t)(P(a(t)1); . . . ;/. the subsequent sojourn times and resident states. To this
def end, we need to specify the chain’s initial conditions,
P(a(t)m)) (pi (t))  R1m prescribe the parameters of interest, and to identify the
Then p (t) satisfies the following equation: infinitesimal generator Q , upon which the path structure
of the chain can be built.
dp(t)
p(t)Q; p(0)p0 ; (6)
dt The sojourn time

where p0 is the initial distribution. Both Eqs. (5) and (6) Assume that at time t, the chain enters certain state i.
are known as forward equations. It is straightforward Let Si be the sojourn time of a (t) in state i, which is a
that the solutions of Eqs. (5) and (6) are: random variable. To determine the residence time Si
P(t)exp(Qt)  Rmm ; and until the chain first leaves this state requires its prob-
1m
(7) ability distribution. Using Q and the Markov property,
p(t)p0 exp(Qt)  R ;
it can be shown that P {Si ]/t} /exp(qii t). Namely, Si
which are the transition probabilities and the probability follows an exponential distribution with mean /qii1.
distribution at time t , respectively. Its piecewise con-
stant behavior enables us to write the Markov chain as The transition
a(t)am i1 iIfa(t)ig ; where IA(x ) is the indicator func-
tion of the set A given by IA (x) /1 if i  /A and 0 During a period of h (with h sufficiently small) a
otherwise. It is well-known that the mean and higher transition will occur. The chain moves from i to j with
moments can be easily computed by: probability qij /o (h ). Due to the randomness of Si ,
however, the transitions occur at random times rather
X
m X
m
than at fixed time period.
Ea(t) ipi (t); and E[a(t)]l  il pi (t)
i1 i1
(8)
The path structure
for any positive integer l;
where pi (t ) is the ith component of p (t ). The variance To construct the sample path of a (t) amounts to
and other statistics of interest can also be obtained determining its sojourn time at each state and its
readily. subsequent moves. The chain sojourns in any given
238 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

state i for a random length of time, Si , which has an to find connections linking the elements of the physical
exponential distribution with parameter (/qii ). Subse- process to the Markov chains.
quently, it will enter another state. Each state j (j/1,
. . ., m , j "/i) has a probability of qij /(/qii ) being the
chain’s next residence. To simulate this process we first 3.1. Problem formulation
sample an exponential random number Si with para-
meter (/qii ) to fix the residence time in compartment i; When a capsule containing chemically-active micro-
then define a discrete random variable Xi that takes spheres/pellets is administered into the GI tract, it
values in {1, 2, . . ., i/1, i/1, . . ., m }. The chain’s post- immediately moves to the stomach that is considered
jump location is determined by the value of Xi as as the first compartment, in which it ruptures and
follows: discharges the microspheres into the slightly acidic
8 liquid environment. A portion of the microspheres is
>
> 1; if U Bqi1 =(qii ); dissolved in the stomach; the undissolved portion exits
<2; if q =(q
i1 ii )B U B(qi1 qi2 )=(qii ); from it and moves to the consecutive compartments of
Xi  (9)
>
> n small intestine, where the dissolution and transition
:m; if P q =(q )BU:
k"i;kBm ik ii processes continue. Eventually, the undissolved drug
where U is a random variable uniformly distributed in enters the colon. The drug dissolved in the GI tract is
(0, 1). The sample path is thus constructed by sampling absorbed into the blood stream for subsequent circula-
from exponential and U (0, 1) random variables alter- tion and distribution into various organs. Given the fact
nately. that the fraction of drug entering the blood from the
In what follows, we illustrate this approach by using colon is small (low bioavailability), we will follow the
two case studies. conventional treatment (Chen et al., 1998; Gibaldi &
Perrier, 1982 and the references therein) and only
consider dissolutions in the first four compartments
3. The drug delivery process (stomach, duodenum, jejunum, and ileum) that lead to
high bioavailability.
The first example concerns modeling and simulation To establish a stochastic model and to design the
of a controlled-drug delivery by the oral route. This is a simulation procedure for the system described above, we
typical random process subject to constant fluctuation. make the following assumptions:
A number of factors such as physicochemical properties (A1) the drug transition and dissolution in the system
of the drug and biological conditions of the patient will are Markovian. The process is considered stationary.
affect the drug delivery and release in the gastric and (A2) The transition and dissolution of each individual
intestinal regions. microsphere are independent; i.e. they are not
In an earlier work, Chen et al. (1998) developed a affected by the presence of other microsperes.
stochastic multi-compartmental model for the con- (A3) Drug degradation is negligible.
trolled-drug release process to describe the drug dis-
solution in four compartments, the stomach, duodenum,
jejunum, and ileum. They derived the master equation 3.2. Associated Markov chains
(van Kampen, 1992) from probabilistic population
balances, and obtained the solution by using probability Given the total number of the chemically-active
generating function, which resulted in a first-order microspheres/pellets and its initial distribution in each
partial differential equation. This partial differential compartment, we are interested in changes in the
equation was solved by means of the Lagrange method distribution with respect to time and the possible
of auxiliary equations. The solution further led to the fluctuations in the random process. We use a family of
means and covariances of the microspheres inside the continuous-time Markov chains, ak (t) (k /1, 2, . . ., k ).
system. An alternative approach to obtain the means For each Markov chain, the state space M includes all
and covariances by generating the first and second m/1 compartments along the GI tract, starting from
moments of the master equation was also presented. the stomach and ending with colon, coded 1, 2, . . ., m/
Observing the randomness and regularity involved in 1, and the blood stream as the mth state. For easy
the drug release process, we choose to model it with a comparison with results of the referenced work, we
family of continuous-time finite-state Markov chains. divide the intestine into three sections and, therefore,
Their piecewise constant sample paths constructed by m /6 hereafter. Representing k /6 random processes,
the numerical scheme simulate the evolution of the these six chains have the same state space M/{l, . . ., 6}
dynamic system. The flexibility in various choices of the and the same transition probabilities, but start from
transition and dissolution parameters allows us to different compartments/states, i.e. a (0) /k . Since we are
explore a host of phenomena. The major challenge is interested in the control in the first four compartments,
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 239

only k/2 /4 chains are monitored in the following 7) mm1 /0; mi /0 for i /1, . . ., m/2.
simulation examples.
Let y (t )/[y1(t), y2(t), . . ., yk (t)]?, where the symbol Note that m /6 in our simulation and that the fifth
‘?’ denotes transpose. We define yi (t) to be the total and the sixth states correspond to the colon and the
number of microspheres at time t that originated from blood stream, respectively. Assumptions 1/3 and 6
division i (at t/0). Note that y(0) /[y1(0), y2(0), . . ., describe the probabilities of the microspheres’ transiting
yk (0)]?, where yi (0) are the number of microspheres in from one compartment to the next, dissolving and
division i at t/0. Since yi (t) are the number of those at entering the blood stream from each of the first m/2
i originated from the ith division, the dynamics of this compartment, and retaining in the current compartment
group is governed by the Markov chain starting from during a very short period of time h . Assumption 4 is the
division i . This relates the number of microspheres in initial condition. In this work, we are not concerned
each division i to the Markov chains ai (t). with the subsequent circulation and distribution from
Let N (t) be the total number of microspheres in the the blood stream that is the m th division in the model.
system at t : Therefore, it is treated as an absorbing state in
Assumption 5. The drug’s dissolution in the colon is
X
k
considered negligible and hence mm1 /0 as stated in
N(t) yi (t)N(0)N0 :
Assumption 7. Consequently, colon is also an absorbing
i1
state. Organizing all the assumptions above yield the
Thus N (t ) is also related to the Markov chains. infinitesimal generator Q /(qij ).

0 1
(l1 m1 ) l1 0 0  0 m1
B0 (l2 m2 ) l2 0  0 m2 C
B C
B0 0 (l3 m3 ) l3  0 m3 C
B :: C
QB
Bn n n : n n n C
C (10)
B0 0 0  (lm2 mm2 ) lm2 mm2 C
B C
@0 0 0  0 0 0 A
0 0 0  0 0 0

Consequently, simulating the dynamics of the con- Observe that qij (t) ]/0 for j "/i and qii (t)aj"i qij (t)/
trolled-drug release process amounts to keeping track for t ]/0. The last column in the Q matrix indicates the
of each Markov chain ai (t) as explained in the following dissolution in the blood stream from each compartment.
sections. Note that there are two absorbing states in this
problem */the blood stream and the colon. Rather
than a 6/6 matrix, an m /m matrix Q is used for
3.3. Infinitesimal generator generalization.

Our approach requires the infinitesimal generator Q .


3.4. Markovian drug delivery process and the closed-form
Suppose all k chains have the same transition prob-
solution
abilities. For notation simplicity, we drop the designator
k of the chain ak (t). Assume that for any fixed a (t )/i The transition and dissolution rates were computed
(i/1, . . ., m ), Pij (h ) for small h satisfies the following
based on the transition and dissolution intensity func-
conditions:
tions provided by Chen et al. Using their estimated
1) Pi ,i1(h )/li h/o(h ) as h 0/0, m /i ]/1 where li is quantities, we calculate the qij ’s and construct the
called the infinitesimal transition rate at division i.
2) Pi ,m (h) /mi h/o (h) as h 0/0, m/2 ]/i ]/1 where mi Table 1
Elements of the generator
is the infinitesimal dissolution rate at division i .
3) Pi ,i (h) /1/(li /mi )h/o (h ) as h 0/0, m /i ]/1. Compartment Index qii //(li /mi ) qi ,i1 /li qi 6 /mi
4) Pi ,j (0) /dij ; the Kronecker delta dij /0 for i "/j,
Stomach 1 /2.249 1.114 1.114
and dij /1 for i/j . Duodenum 2 /0.154 0.149 0.005
5) Pm ,j (t)/0 and Pm ,m (t) /1 for m/1 ]/j ]/1, i.e. m Jejunum 3 /1.059 0.807 0.252
is an absorbing state. Ileum 4 /1.419 0.452 0.967
6) li /0 for i/1, 2, . . ., m/1.
240 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

generator Q defined in Eq. (10), whose elements are conditions as that in the work of Chen et al., where
tabulated in Table 1. y (0) /[100 0 0 0 0 0]?. Namely, at t/0, there were a
Note that the Q matrix has a simple upper triangular total of 100 chemically-active pellets, all in the stomach.
form. It is readily seen that the eigenvalues of Q are qii Fig. 1 presents the result of a typical simulation run,
for i /1, . . ., 4 and 0. From elementary linear algebra, it in which (a) /(d) demonstrate the changes of number of
is also easy to obtain that V 1QV /D (i.e. Q is microspheres in the stomach, duodenum, jejunum, and
diagonalizable), where: ileum along time. The spread of the pellets in the system

0 1
1:0000 0:4695 0:1503 0:1421 0:1049 0:4750
B0 0:8829 0:1606 0:1059 02118 0:4157C
B C
B0 0 0:9755 0:8988 0:2189 0:4114C
V B B0
C;
B 0 0 0:4010 02873 0:3702C C
@0 0 0 0 0:9020 0 A
0 0 0 0 0 0:5433
0 1
1:0000 0:05317 0:0666 0:0647 0:0130 0:4738
B0 1:1326 0:1865 0:1190 0:3491 1:0889C
B C
B 0 0 1:0251 2:2980 0:9808 2:3423C
1
V B B C
B0 0 0 2:4941 0:7944 1:6996C C
@0 0 0 0 1:1086 0 A
0 0 0 0 0 1:8406

and D is a diagonal matrix with diagonal entries / at any given time and its progress along time can also be
2.249, /0.154, /1.059, /1.419, 0, and 0. Thus, in easily obtained. Each simulation run provides us with a
view of Eq. (7), the transition matrix P (t) and the single realization of this random process. Repeating the
probability distribution p (t ) (with initial distribution p0) same simulation procedure a finite number of times and
are given by: collecting the results for a fixed time t will yield
P(t)V exp(Dt)V 1 ; and probability distribution of the said random variable
1
(11) a (t, v) at that particular time. Fig. 2 shows the
p(t)p0 V exp(Dt)V ; distribution of the number of the pellets in each
respectively. A closed-form solution Eq. (11) is thus compartment at the end of the second hour. It shows
obtained. that we can expect that, with above 95% confidence,
Due to the low dimension and the special form of the there will be about 30 /50% drug left in the stomach, and
generator of this process, solving the forward equation about 2 /18% in the jejunum, etc.
Eqs. (5) and (6) is straightforward. With P (t ) and p (t) Repeating the numerical experiment a number of
available, statistics such as mean, covariance, and higher times, collecting the results at each time instant and then
moments can be computed easily from Eq. (8). How- taking the average give us the sample mean of the pellets
ever, it should be emphasized that although formulas in each compartment. Similar procedure can be used for
are available for solving second-, third- and fourth-oder the sample variance or standard deviation (S.D.). Fig. 3
polynomial equations, there is no closed-form solution displays the means and S.D. obtained from 200 simula-
available in general, while the simulation approach tion runs. They are in good agreement with those
provides a viable alternative. obtained from the solutions of the master equation as
well as their governing equations (see Figs. 6 /9 in Chen
3.5. Simulation results et al., 1998). Also observe the similarity of the trends
between the same compartments of Figs. 1 and 3.
The simulation is relatively simple owing to its low Our results, e.g. the mean values of 41, 1.5, 11 and 7 at
dimension. It includes three components: generating t /2 h are consistent with those obtained in the work of
sample paths of the Markov chains, examining the Chen et al. Indicative of the fluctuations around the
dynamic behaviors of each individual chain and the means, the variances or S.D. in all compartments can
combined effects of all chains, and studying the prob- also be computed easily. It has been observed that
ability distribution in each state/division at a fixed time. compartments retaining larger numbers of particles tend
For easy comparison, we have adopted the same initial to have larger variances. Nevertheless, their relative
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 241

Fig. 1. A single realization of the controlled drug release process-changes of the numbers of pellets in (a) the stomach; (b) the deuodenum; (c) the
jejunum; and (d) the ileum.

Fig. 2. Population distribution at t /2 h in (a) the stomach; (b) the deuodenum; (c) the jejunum; and (d) the ileum.
242 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

Fig. 3. Calculated temporal variation of the mean number of pellets and its deviations for (a) the stomach; (b) the deuodenum; (c) the jejunum; and
(d) the ileum, ( */) mean; (  ) /S.D.; ( /) /S.D.

Fig. 4. A single realization of the controlled-drug release process when the second dose is taken 4 h after the first */changes of the numbers of pellets
in (a) the stomach; (b) the deuodenum; (c) the jejunum; and (b) the ileum.
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 243

fluctuations with respect to the means tend to be Contrary to the traditional stochastic approach which
smaller. On the contrary, retention in the smallest assumes that cells in a bio-system are all the same, we
compartment, duodenum, is usually the smallest among take growth into consideration and distinguish indivi-
the four. However, it experiences the largest relative dual cells according to their mass, Rather than a birth/
fluctuation as compared with other compartments. death process, we treat birth /growth /death processes.
We have also calculated the mean residence time of Observing the randomness and regularity in cells’ birth,
the drug pellets in the four compartments. They are death, and growth process, we model it with continuous-
2.43, 0.13, 1.16 and 1.38 h, respectively. They are close time and finite-state-space Markov chains. Our formu-
to the mean sojourn times /qi ,i with lation allows both constant and variable birth, growth,
q1;1  2:25; q2;2  0:15; q3;3  1:06; and death rates. In addition, it enables us to describe
(12) population dynamics at different ‘stages’ by incorporat-
q4;4 1:42 ing age and other related factors into the model.
as expected. The results are consistent with those in The major challenge we face herein is that the cells are
Chen et al. (1998) and are useful in determining the deemed different. As a result, a number of Markov
dosage etc. Fig. 4 presents the situation when another processes are involved rather than a single one as in the
dose is taken 4 h after the first one. Results similar to traditional model. To relate elements of the bioprocess
those presented in Fig. 2 can also be readily obtained. to the Markov chains, we define the state space of the
Markov chains on the basis of the mass of individual
cells, and formulate the infinitesimal generators of the
3.6. Remarks underlying Markov processes with the birth /growth/
death rates of the bio-system. The developed stochastic
In this work, we modeled the controlled-drug release model is used as the basis for simulating the cell
process using continuous-time and m -state Markov population dynamics in representative case studies.
chains, which simulate the rnicrospheres’ move along Issues including specification of the initial conditions,
time. All the chains have the same state space {1, . . ., construction of the sample path structure, and determi-
m }. The generator Q used is similar to a birth /death nation of the sojourn times are addressed.
process. Nevertheless, in a birth /death process, if the
birth and death rates li /0 and mi /0 for all i ]/1, then
it is irreducible and all states are recurrent. Conse- 4.1. Problem formulation
quently, we can obtain its stationary distribution (Karlin
& Taylor, 1975). On the contrary, in the drug delivery To establish a stochastic model for the growth, split/
model there are two absorbing states (colon and blood birth and death process, we make the following assump-
stream) while all other states are transient; thus even- tions:
tually there will be no pellets left in the first four
compartments. (B1) The population is well mixed; each cell acts
independently.
(B2) The cells of a population differ in size, and
individual cells of different sizes have different
4. Cell population dynamics growth, death, and reproduction rates, respectively.
This is the major difference between the current
Numerous investigations have been made of the model and the traditional birth and death process
continuous propagation of cells with various modeling model.
methods. Among the earliest works, Tsuchiya, Fredrick- (B3) We assume equal split; namely, one mother cell
son and Aris (1966) provided a comprehensive review of splits into two identical daughter cells.
the mathematical modeling of population dynamics.
They considered ‘pure birth processes’ of indistinguish-
able individual cells. Based on the assumption that the
proliferation of cells is a Markov process, the differ-
ential equation of the transition probability was estab-
lished; its solution under a constant growth rate was
obtained; subsequently, the generating function was
obtained and used to estimate the mean and variance
of the distribution. It was pointed out in that work that
if we choose to include cellular structure in the model,
the distribution of structures in the cells must be Fig. 5. State space of the Markov chains for the mass structured cell
considered. birth /growth /death process.
244 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

(B4) There are sufficient nutrition supplies, therefore, Assume that there are k independent, continuous-
the effects of substrate concentration can be ne- time finite-state Markov chains having the same state
glected. space M /{1, . . ., m }, however, each starting from
(B5) The age differences can be neglected. different states, i.e. ak (0) /k , (1 B/k B/k ). Assume the
(B6) The properties of growth, birth, and death do transition probabilities of each chain ak (t) are station-
not vary with time. ary, i.e. Pkij (t)/P {ak (t/s)/jjak (s)/i} for all s /0
for the k th chain. Thus we can model the situation with
We assume equal split for simplicity. In fact, unequal a family of k members continuous-time, m -state Mar-
division can be treated by a slight modification of the kov chains where each state corresponds to the cell
model presented in the following sections. The substrate mass. This model is illustrated in Fig. 5.
condition specified in assumption (B4) can also be
relaxed by including the nutrition effect into the split
4.3. Infinitesimal generator
and growth parameters. We assume the process of
interest is stationary, therefore (B6) is required. In order
To specify the generator Q in the structured cell
to be able to distinguish the age difference we may
population problem, we consider all of the k chains
consider the bio-process as a multistage process and
have the same transition probabilities. For notation
apply the model to different stages using different model
simplicity, we drop the subscript k . Assume that for
parameters.
each fixed i and j/1, . . ., m , Pij (h ) for sufficiently small
h satisfies the following conditions:
4.2. Associated Markov chains
1) Pi ,i1(h )/li h/o (h ) as h 0/0, m /i ]/1 where li is
the growth rate of cells in division i .
First, we need to decide what the states of the Markov
2) Pi ,m (h )/mi h/o(h) as h 0/0, m /i ]/2 where mi is
chains in the model correspond to and how many states
the death rate of cells in division i.
there should be in the model. Let w (t) denote the cell
3) Pi,l (h )/gi h/o (h ) as h 0/0, m /i ]/3; where l is
mass at time t, w(t)  /(0, wmax], where wmax is the
the division satisfying wl1 5/wi /2 B/wl and gi , is the
maximum cell mass. To construct a mass-structured
split rate of cells in division i .
model we divide the entire population into m sections
4) Pi ,i (h )/1/(li /mi /gi )h/o (h) as h 0/0, m /i ]/
according to the mass of the individual cells; i.e.
1.
partition (0, wmax] into m subintervals [wi1, wi ]
5) Pi ,j (0) /dij , dij /0 for i "/j and dij /1 for i/j.
where:
6) Pm ,j (t) /0, Pm ,m (t) /1; i.e. m is an absorbing state.
wmax 7) li /0 for i/1, 2, . . ., m/2; lm1 /lm /0.
wi wi1  for i 1; 2; . . . ; m:
m 8) mi /0 for i/1, 2; mi /0 for i/3, . . ., m .
9) gi /0 for i/1, m ; gi /0 for i/2, 3, . . ., m/1.
Assume that cells in the same division are all the same;
i.e. they have the same growth, split, and death rates. Assumptions 1 /3 describe the probabilities of the
The cells in division i may grow and enter the next chain’s growth, death, and split in a very short period of
division, i/1; they may split and become two identical time h if the chain starts in state i . Assumption 6
cells leaving division i then entering the division /wi /2 /, indicates that the division m is the so-called absorbing
where /A / denotes the integer part of A . The cells in state. Organizing all the assumptions above gives us the
any division 1B/i 5/m/1 may die and enter the m th infinitesimal generator Q /(qij ) of a (t).

0 1
l1 l1 0   0 0 0
B0 (l2 m2 ) l2 0   0 m2 C
B C
Bg3 0 (l3 m3 g3 ) l3 0  0 m3 C
B C
QB
B 0 g4 0 (l4 m4 g4 ) l4 0  m4 C C (13)
Bn n n n n n n n C
B C
@0 0  0 gm1 0 (mm1 gm1 ) mm1 A
0 0 0 0 0 0 0 0

division. All the cells in division m are dead cells that Observe that qij (t) ]/0 for j "/i and qii (t)aj"i qij (t)
will not grow or split and will stay there forever. for t ]/0. Note the position of gi , in the matrix is
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 245

Fig. 6. Simulation example of the Markov model for the mass structured cell birth /growth /death process-comparison of different linear growth
rates: l/1.0w in (a and b) vs. l /5.0w in (c and d). Population distributions shown in (c and d) are for t/0 (solid lines) and t /200 (dashed lines),
respectively.

symbolic and it should be determined by the cell mass in Let N (t) be the size of the entire population at t.
the ith division.
X
m1
N(t) yi (t):
i1

4.4. Markovian cell population dynamics Thus N (t) is also related to the Markov chains.
Consequently, simulating the dynamics of the popula-
Given the initial population size and its distribution tion amounts to keeping track of each Markov chain
with respect to cell mass, the population dynamics can ai (t) and the offsprings resulting from its move as
be modeled by a family of continuous time Markov explained in the following sections.
chains ai (t ), for i /1, 2, . . ., m/1. These m/1 chains
have the same state space, M/{1, . . ., m }, and the 4.5. Simulation results
same transition probabilities, but starting from different
initial states, ai (0) /i. Note that k /m/1 in this case. Several simulation examples performed according to
Let y(t)/[y1(t ), y2(t), . . ., ym1(0)]? where the the above scheme are presented in this section. Each
symbol ‘?’ denotes transpose. We define yi (t ) to be the simulation started from the same initial mass distribu-
total number of offsprings at time t that are originated tion and initial population size. We have tried different
from division i (at t /0). Note that y(0) /[y1(0), y2(0), numbers of divisions in our study. As can be expected,
. . ., ym1(0)]?, where yi (0) are the number of cells in more divisions can provide smoother profile, however,
division i at t/0. Since yi (t) are the number of require more computation time. In the examples to
offsprings at t originated from the ith division, the follow, the entire population was divided into m /20
dynamics of this group is governed by the Markov chain sections according to the mass of individual cells. Under
starting from division i. Equivalently, the number of conditions of sufficient substrate concentration as well
cells in each division i is related to the Markov chains as equal partitioning, we studied population dynamics
ai (t). for constant, linear, and quadratic growth rates with or
246 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

Fig. 7. Simulation example of the Markov model for the mass structured cell birth /growth /death process-comparison of linear (a and b) vs.
quadratic (c and d) growth rates. Population distributions shown in (c and d) are for t/0 (solid lines) and t /200 (dashed lines), respectively.

Fig. 8. Simulation example of the Markov model for the mass structured cell birth /growth /death process-comparison of constant (a and b) and
linear split rates (c and d). Population distributions shown in (c and d) are for t/0 (solid lines) and t /200 (dashed lines), respectively.
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 247

Fig. 9. Simulation example of the Markov model for the mass structured cell birth /growth /death process-comparison of cases with (a and b) and
without (c and d) death. Population distributions shown in (c and d) arc for t /0 (solid lines) and t/200 (dashed lines), respectively.

Fig. 10. Simulation example of the Markov model for the mass structured cell birth /growth /death process-changes in biomass composition for t /
0 /500.
248 K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249

without death. The paths of the Markov chains were cell growth. In this work, we have partitioned the entire
constructed by keeping track of their sojourn times and population into a number of groups according to cell
residence states as illustrated in Section 2. mass, and constructed the model that consists of a
Figs. 6 and 7 compare two situations where the cells’ family of continuous-time finite-state Markov chains. It
split rates in each mass division are the same but their should be noted that such an approach is not restrictive
growth rates are different. Both figures show the to cell mass. It can be generalized to other cases where
population dynamics along with population distribu- groups in a bio-system are characterized by certain
tions at t/0 and t/200. Fig. 6(c) and (d) present the attributes.
case of which the growth rates in each mass division,
li /kli wi (for i/1, . . ., m/1), are five times those
represented in (a) /(b) of Fig. 6. Fig. 7 displays a
comparison of the case of linear growth rates (a)/(b) 5. Summary and discussion
versus the case where the growth rates are quadratic
functions of mass (c) /(d), namely. li /kqi w2i . This paper is concerned with stochastic modeling and
Fig. 8 compares two situations that have different simulation using continuous-time and finite-state Mar-
split rates for the same mass division but their growth kov chains. Some of the theoretical aspects involved in
rates are the same. It was assumed in the above Markov chains, the Chapman /Kolmogorov equation
examples that the cell death is negligible. Fig. 9 and the forward equations were reviewed briefly, along
demonstrates the death effect. In Fig. 9(a) and (b) the with the necessity of numerical simulation in certain
death rates were assumed to be zero whereas a constant cases. Two illustrative case studies, a drug delivery
death rate was assumed for each division and presented process and a cell growth process, were modeled and
in Fig. 9(c) and (d). their dynamics were simulated within the Markovian
Fig. 10 shows the change in mass composition along framework. The approach followed requires integrating
time, where the abscissae are the time axes, covering the the information of the randomness with the inherent
periods of 0 B/t B/100 in (a), 200 B/t B/300 in (b), and regularity of the underlying physical process, expressing
400 B/t B/500 in (c), and the ordinates denote the 20 them in a probability framework, and relating them to
divisions of this system. Different shades represent the corresponding Markov chains. The key is to identify
different compositions. It can be seen that the popula- the generator, which requires a thorough understanding
tion distribution moves from lower mass to higher mass of the basic properties of the underlying systems. Both
as time goes on. application examples belong to biological systems.
It should be noted that the meaning of linear/ Nevertheless, such an approach is not restricted to
quadratic growth rates used herein are different from bio-systems. Rather, it is applicable to other participate
those used in the deterministic models or those used in processes that may be described via finite-state Markov
classical birth and death processes. The linear or chains. Stochastic modeling and simulation have be-
quadratic in the latter cases refer to the functional come frequently used and powerful tools in quantifying
relationship between rates and the number of cells or the dynamic relations of sequences of random events and
population size. In our case, however, we relate the uncertainties. Considering the complex and random
growth rates to the cell mass rather than the population natures of many chemical processes, it is conceivable
size. that these methods will play an increasingly important
Birth, death, and growth processes can be generalized role in chemical engineering and related fields.
to birth, growth, illness, and death processes in which
illness is potentially concurrent, repetitive, and reversi-
ble. Further complexity can also be introduced by
competition of various causes. Models similar to the References
one above can be established and all the health, illness,
and death transition probabilities as well as the expected Berthiaux, H. (2000). Analysis of grinding processes by Markov
duration of stay in each state can be calculated. These chains. Chemical Engineering and Science 55 , 4117 /4127.
quantities, in turn, are useful in estimating other Burge, C., & Karlin, S. (1997). Prediction of complete gene structures
in human genomic DNA. Journal of Molecular Biology 268 , 78 /
important parameters of interest such as the population 84.
size in each state and the duration of survival time. Chen, X., Chen, W. Y., Hikal, A. H., Shen, B. C., & Fan, L. T. (1998).
These can be further expanded to the study of various Stochastic modeling of controlled-drug release. Biochemical and
factors and their impacts on the transition probabilities Engineering Journal 2 , 161 /177.
and the time of staying in each state, hence their effects Chiang, C. L. (1980). In E. Robert (Ed.), An introduction to stochastic
processes and their applications . New York: Krieger Publishing.
on the survival time, and population size in different Chung, K. L. (1960). Markov chain with stationary transition prob-
stages etc. More importantly, we will be able to develop abilities . Berlin: Springer.
effective measures to create a favorable environment for Doob, J. L. (1953). Stochastic processes . New York: Wiley.
K.K. Yin et al. / Computers and Chemical Engineering 27 (2003) 235 /249 249

Durbin, R., Eddy, S., Krogh, A., & Mitchison, G. (1998). Biological Mantzaris, N. V., Daoutidis, P., & Srienc, F. (2001c). Numerical
sequence analysis */probabilistic models of proteins and nucleic solution of multi-variable cell population balance models III. Finite
acids . Cambridge: Cambridge University Press. element methods. Computers and Chemical Engineering 25 (11 /
Fan, L. T., Nassar, R., Hwang, S. H., & Chou, S. T. (1985). Analysis 12), 1463 /1481.
of deep bed filtration data: modeling as a birth /death process. Nassar, R., Schmidt, J., & Luebbert, A. (1992). A stochastic dispersion
American Institute of Chemical and Engineering Journal 31 (11), model in gas /liquid flow systems. Chemical and Engineering
1781 /1790. Science 47 , 3657 /3664.
Gibaldi, M., & Perrier, D. (1982). Pharmacokinetics (2nd ed.). New Ramkrishna, D. (2000). Population balances. Theory and applications
York: Marcel Dekker. to particulate systems in engineering . New York: Academic Press.
van Kampen, N. G. (1992). Stochastic process in physics and chemistry Taylor, H. M., & Karlin, S. (1998). An introduction to stochastic
(2nd ed.). Amsterdam: North-Holland. modeling . Boston: Academic Press.
Karlin, S., & Taylor, H. M. (1975). A first course in stochastic Tsuchiya, H. M., Fredrickson, A. G., & Aris, R. (1966). Dynamics of
processes . Boston: Academic Press. microbial cell populations. Advnace in Chemical Engineering 6 ,
Khasminskii, R. Z., & Yin, G. (1996). Asymptotic series for singularly 125 /206.
perturbed Kolmogorov Fokker Planck equations. SIAM Journal of Yin, G., & Zhang, Q. (1998). Continuous-time Markov chains and
Applied Mathematics 56 , 1766 /1793. applications. A singular perturbation approach . New York:
Mantzaris, N. V., Daoutidis, P., & Srienc, F. (2001a). Numerical Springer.
solution of multi-variable cell population balance models I. Finite Yin, G., Zhang, Q., Yang, H., & Yin, K. (2001). Discrete /time
difference methods. Computers and Chemical Engineering 25 (11 / dynamic systems arising from singularly perturbed Markov chains.
12), 1411 /1440. Nonlinear Analysis of Theory Methods and Applications 47 , 4763 /
Mantzaris, N. V., Daoutidis, P., & Srienc, F. (2001b). Numerical 4774.
solution of multi-variable cell population balance models II. Yin, K., Liu, H., & Johnson, N. (2002). Markovian inventory policy
Spectral methods. Computers and Chemical Engineering 25 (11 / with application to the paper industry. Computers and Chemical
12), 1441 /1462. Engineering 26 (10), 1399 /1413.

You might also like