The Transient Solution of the Linear Birth-Death Process with Random Spontaneous Mutation

B. G. BIRKHEAD Clinicul Operutionai Research Unit, Department of Statisticul Science, Umversity College London, Gower Street, London WC1 E 6BT. Englund Received 6 May 19%; revised 6 August 1986

ABSTRACT The transient solution of the simple linear birth-death process subject to random mutation is derived. The solution is applied to investigate the curability of cancer under drug treatment. Tumor cells are assumed to proliferate by division, and may die or be lost to the population. Mutation is from drug sensitivity to resistance, and may occur only at tumor size, division. Expressions are derived relating curability and to the delay before treatment is given. of the disease to increasing

INTRODUCTION In many biological systems, interest may be in the origin and development of mutant populations within a proliferating normal population (e.g. the survival and spread of mutant genes in organisms [l], transmutations in populations of radioactive atoms, etc.). A model which may often adequately describe such systems is that of simple linear birth-death growth with random spontaneous mutation. This model has been used in the past to describe the growth of bacterial populations [2] and malignant cell populations [3] in the presence of mutation towards resistance. The partial differential equation for the generating function of the process will be formulated, and the transient solution will be derived. The solution will be applied to investigate how the emergence of drug-resistant clones in cancerous tumors may affect the success of drug treatment (chemotherapy). A new expression for the curability of tumors of a given size will be presented, together with a simple approximate formula. The importance of the timing of treatment to its success will also be investigated. THE MODEL Without populations. loss of generality, the model will be formulated in terms of cell

MATHEMATICAL

BIOSCIENCES

82:193-200

(1986)

I93
00255564/86/$03.50

QElsevier Science Publishing Co., Inc., 1986 52 Vanderbilt Ave., New York, NY 10017

194

B. G. BIRKHEAD

ll=

(b-a lAt
FIG. 1. Cell proliferation Af.

CCAt
and the spontaneous mutation

b At
process.

d At

d At

n is the probability

in a small time interval

Suppose that the proliferation of all cells can be realistically approximated by a simple, linear birth-death process: in any small time interval At, each cell has probabilities b At and d At of respectively dividing (birth) to produce two cells in place of one, or of being lost to the population (death); b and d being time-independent constants with b > d (a birth advantage). Suppose that a random mutation process operates at constant rate LY during cell division, so that in At, the probability of a division of a normal cell to produce one normal and one mutant cell is (Y At, and a division to produce two normal cells has probability (b - a) At. Back mutation [mutant + (normal + mutant)] is assumed to have negligible probability. The process is shown diagrammatically in Figure 1.

SOLUTION The partial pendix 1) is differential equation governing the process (derived in Ap-

$={[(b-a)x-d][x-1]

+ax[y-l]}g+[by-d][y-l]i& (1)

where P(x,y, r) =~~=o~~~=o~(r,s, t)xSy is the probability generating function, and r and s correspond to the numbers of mutant and normal cells, respectively. Using the Lagrange method of solution [4], Equation (1) has auxiliary equations --= dt 1
(9

[(b-a)x-dl[~--l]+ax[y-l] (3

= [by-$&]
(iii)

= 2

TRANSIENT SOLUTION

195 constants * * of integration, (la) = e*,

(lb)

Letting

8, (i = 1,2,3) represent (iv) (i) and (iii)

P=Bi, bY - deAkO
1-Y

and (ii) and (iii) become

dx dy= b-a (by-d)(y-1)X2+ ay-(b+d) (by-d)(y-l)x+ (by-d;(y-I)

(*

which has the form of the Riccati first-order ordinary differential equation [4]. x = y is a particular solution, and the substitution x = y + l/u( y) leads to

&+(2b-dy-(b+d)u+ 4 (by-d)(y-1)
This in turn has solution

(by-bdj;y-l)

x(1-y)BP1dy+03 =

(b- a)(b+z)
(b-d)A

1
(14

1-B,l,;+l,-; y)-,

+%(by-4
where
A= b*-bd+ad b-d

-y_

B=b-d-a b-d

and

z=e,

and where F is the hypergeometric function. Initial conditions: Suppose at t = 0 there are S, normal cells and R, mutants, viz. P(x, y,O) = xSOyRo. Then elimination of the integration constants from (la)-(lc) gives qx,y,t) =( ~:~~I::j[ ~~~~1:: +[ b_b~~ar+(ze-~
1

1
so
9

(2)

196 where
(b+z)(b-a)

B. G. BIRKHEAD

J=b-d

and

4(z)=

(b_d)A

+&+I.-;).

This equation is similar to that developed by Goldie and Coldman [5] in the context of drug resistance, with an additional component representing the growth of mutant cells present at t = 0. In theory, albeit with tedious algebra, expressions for the coefficients p(r,s,t), for all r,s, cm be found from (2) by repeated differentiation, and other expressions of physical importance can be deduced in the usual way. For example, the probability of there being no mutants at time t after the start of the process is given by P(l,O, t), and so on. APPLICATIONS TO DRUG RESISTANCE IN CANCER THERAPY

Goldie and Coldman have recently described a mathematical model of the effects of spontaneous mutation of cells, from drug sensitivity towards phenotypic resistance, on the curability of primary human tumors [3,6]. Their most recent model [3], whose solution they have approximated, corresponds to the stochastic birth-death process dealt with in this paper, with normal cells corresponding to drug-sensitive cells, and mutant cells to resistant. New expressions for curability, based on the exact solution above, will now be derived, and a useful approximate formula will be presented. An application to the treatment of residual tumors with cytotoxic drugs, after an alternative mode of therapy has already been used, will also be discussed. When a tumor is treated with drugs (chemotherapy), the best that can be achieved is complete eradication of the population of sensitive cells; the probability of cure is then equivalent to the probability of eventual extinction of the remaining resistant cells under the stochastic process.
A. TREATMENT OF THE PRIMARY TUMOR

Suppose at time t during its development, a tumor has s cells sensitive to the drug of choice, and that optimal treatment is then applied. (For moderate mutation rates [3] (a < 10e5, say), s will approximate the overall size of the tumor at the time of treatment.) Then the expected probability of cure (as defined above) is given by

c(s,t) =
where theconditionalprobabilityp(r]s,

2 p(rls,t)($)
r=O

(3)
t); (d/b)

t) =p(r,s,

t)/p(s,

is the

TRANSIENT

SOLUTION

197

probability of eventual extinction of r resistant cells in a simple linear birth-death process, and p(s, t) is the probability of there being s sensitive cells at time t. In terms of the process generating function, C(s,t) and from (2) (d/b)[(b-a)x-b]-b(x-d/b)G(t) [(b-cu)x-b]-(x-d/b)(b-a)G(t) = coefficient of xs in P( x, d/b, P(S, t) t)

(4)

P(x,d/b,t)

1[I
so ;

R,

where G(t) = exp{[db - ad - b2]t/b}. If, for simplicity, we assume that the tumor has developed from a single sensitive cell (Se = 1, R, = 0), recursive differentiation of (5) gives coefficient of xs in P x, b-(d/b)(b-a) i b-(d/b)(b-a)G(t) that

d
T; ,

)I l-G(t) b/(b-a)-(d/b)G(t) s-l i (@

=G(t)

and it is known

p(st)=

(Oe - d) 19 [d-(b-(Y)e@r]2

(7) i

where 0 = b - a - d, since the sensitive cell population itself proliferates under a simple linear birth-death process with parameters b - a and d. C(s, t) is then given by the expression (6) divided by the expression (7). A tumor with birth and death (b and d) parameter values 1 and 0.98 respectively will have an expected doubling time of approximately 35 days. Suppose that optimal treatment is applied to such a tumor at 1000 days into its development. The stochastic nature of its growth will mean that the actual size at this time may assume a range of possible values, each with some probability. C( s, 1000) represents the probability of cure given that its size is approximately s at this time. C(s, 1000) versus s is plotted in Figure 2 for a range of different mutation rates. The curves, which can be shown to be robust to increases in t, are of the same shape and scale as those published by Goldie and Coldman, and indicate a rapid loss of curability over a relatively short period of tumor growth. The insensitivity of these curves to the death parameter d, found by the same authors, can now easily be

198
1

B. G. BIRKHEAD

0.8 -

0.6 -

0.4 c ti 0

O2 Id IO

Id

IO

S
FIG. 2.

Curability against sensitive-cell compartment size (b = 1, d = 0.98).

confirmed,

since for large t and small by the expression

a, it can be shown

that C(s, t) is very

well approximated

(-1
b
B. THE USE OF ADJUVANT

b-&

S+l .

CHEMOTHERAPY

Although chemotherapy is commonly used as first treatment for cancer, in many situations drugs are given as an adjuvant after some (perhaps different) form of treatment such as surgery or radiotherapy has already been used. The chemocurability of residual tumors at this stage is likely to be very different from when chemotherapy is the first and only treatment used. It will now depend on the size and composition of the tumor following the earlier treatment, and may depend critically on the time allowed to elapse before the adjuvant therapy is given. Both these factors can be investigated using Equation (2). Let t now denote the time between the end of first-line treatment and the administration of adjuvant chemotherapy. R, and SO will now correspond to the numbers of residual resistant and sensitive cells surviving the earlier treatment. Retaining the assumption that the drug acts optimally and instantaneously eradicates all sensitive cells, the expected (over r and s)

TRANSIENT SOLUTION

199 of the delay before chemotherapy is then

probability

of cure as a function

CCP(rJJ> ()=P(l,%,l)
b

=K(t),

say.

So from (5),

K(t)=

(d/b)a+(b-d)G(t) (Y+(1-_d/b)(b-a)G(t)

so ;
10

RcJ
(8)

Equation (8) can be used to investigate the circumstances in which the timing of adjuvant chemotherapy is important to the chances of effecting cure. DISCUSSION The transient solution of a biological system in which growth of normal and mutant populations can be approximated by a simple linear birth-death process with random forward mutation has been derived. The model approximates such processes as cellular proliferation by smoothing out over the cell cycle the discrete events of division and death, so that for each cell, an event has constant probability per unit time. The larger the total number of cells present, the more realistic this assumption will be. The model also treats each cell independently with regard to these events. The presence of the hypergeometric function in the transient solution means that it will be of closed or polynomial form only for certain sets of parameter values, and useful approximations will need to be investigated. In the application given, a reexamination of a previously discussed aspect of cancer treatment has been touched upon, and it has been shown how different but related problems may now be amenable to analysis using the time-dependent solution. A fuller treatment will be the subject of a future paper. In some physical situations, explicit time dependence may not be required per se. It may be, for example, that while the total population size is known or is measurable, the time since the process started is not. Then the probability of there being a mutant clone of a given size conditional only on total population size will be more directly useful. A model reformulation in these terms is the subject of current research. APPENDIX. PARTIAL DIFFERENTIAL EQUATION

Let p(r, s, t) denote the probability of there being r mutant and s normal cells at time t. Then, if At is a time interval small enough so that at most one event (birth or death) can occur, the forward equations of the

200

B. G. BIRKHEAD

process are given by p(r,s,t+At) =p(r-1, s,t).(r-l)bAt+p(r-l,s,t)asAt

+P(r,s-l,t).(b-a)(S-l)At+p(r,s+l,t).d(s+l)At +p(r+l,s,t).d(r+l).At +p(r,s,t)[l-(r+s)(b+d)At+O(At*)], P(o,l, t + At) =p(O,2, r)2dAt +p(l,l, t)dAt r,s>l,

+p(O,l,t)[l-(b+d)At+@At*)], p(l,O,t+At) =p(l,l,t)dAt+p(2,0,t)2dAt +p(l,O,t)[l-(b+d)At+O(At*)], ~(O,O,t+Ar)=P(0,1,t)dAt+p(1,0,t)dAt+p(O,O,r). Multiplying these equations by xSyr, adding, and letting the partial differential equation ,={[(b-a)x-d][x-l]+ax[y-l]}~+[hy-d][y-l]~, where
P= E E p(r,s,t)xsyr r=Os=O

At -+ 0, we obtain

ap

is

the probability

generating

function.
of Health and Social Security,

This work was funded by the Department Great Britain.

REFERENCES W. Feller, An Introduction and to Probability

Theoty and Its Applications, of bacteria from virus

Vol. 1, Wiley, New sensitivity to virus

York, 1951. S. E. Luria

M. Delbrtick,

Mutations

resistance, Generics 28:491 (1943). J. H. Goldie and A. J. Coldman, Quantitative model for multiple levels of drug resistance in clinical tumours, Cancer Treatment Rep. 67:923-931 (1983). H. T. H. Piaggio, An Elementary Treatise on Differential Equations and Thew Applications, Bell, London, 1950. J. H. Goldie and A. J. Coldman, The effect of cellular differentiation on the development of permanent drug resistance, Math. Biosci. 74:177-198 (1985). J. H. Goldie and A. J. Coldman, A mathematic model for relating the drug sensitivity of tumours to their spontaneous mutation rate, Cancer Treatment Rep. 63(11):1727-1733 (1979).