J. theor. Biol.

(2002) 218, 55–70

doi:10.1006/yjtbi.3057, available online at http://www.idealibrary.com on

A Mathematical Model of a Biological Arms Race with a Dangerous Prey

Paul Waltmanw , James Braselton*z and Lorraine Braseltonz

wDepartment of Mathematics and Computer Science, Emory University, Atlanta, GA 30322, U.S.A.
and zDepartment of Mathematics and Computer Science, Georgia Southern University, P.O. Box 8093,
Statesboro, GA 30460-8093, U.S.A.

(Received on 25 January 2002, Accepted in revised form on 15 April 2002)

In a recent paper, Brodie and Brodie provide a very detailed description of advances and
counter-measures among predator–prey communities with a poisonous prey that closely
parallel an arms race in modern society. In this work, we provide a mathematical model and
simulations that provide a theory as to how this might work. The model is built on a two-
dimensional classical predator–prey model that is then adapted to account for the genetics
and random mating. The deterministic formulation for the genetics for the prey population
has been developed and used in other contexts. Adapting the model to allow for genetic
variation in the predator is much more complicated. The model allows for the evolution of
the poisonous prey and for the evolution of the resistant predator. The biological paradigm is
that of the poisonous newt and the garter snake which has been studied extensively although
the models are broad enough to cover other examples.
r 2002 Elsevier Science Ltd. All rights reserved.

1. Introduction prey relationship of the garter snake Thamnophis

This paper develops a mathematical model of a
biological arms race between a class of predators
and a class of prey where the prey is dangerous
to the predator. In their survey article, Brodie &
Brodie (1999) describe predator–prey arms races
and draw interesting parallels with the more
familiar arms races of modern society, using
as one example law enforcement and speeding
motorists. The biological arms race has also
appeared in the popular media, PBS (2001). The
most interesting example described in Brodie &
Brodie (1999) concerned the unusual predator–
*Corresponding author. Tel.: +1-912-681-0874; fax:
+1-912-681-0654.
E-mail address: waltman@mathcs.emory.edu
(P. Waltman), jimbras@gsvms2.cc.gasou.edu (J. Braselton),
loribras@gsvms2.cc.gasou.edu (L. Braselton).
sirtalis (predator) and the Oregon newt Taricha
granulosa (prey). The newt defends itself by
producing a toxin, tetrodotoxin (TTX). The
snake is the only known predator of the Oregon
newt that has developed resistance to TTX. The
TTX need not cause the death of a garter snake
directly: when a snake consumes a newt, it may
be immobilized by the TTX contained in the
newt’s skin for several hours. In this state, the
snake is susceptible to other predators and, if it
cannot move, may not be able to thermoregulate
properly, and may die, Brodie & Brodie (1999).
The newt–garter snake predator–prey rela-
tionship is a particular example of a biological
arms race where the prey is dangerous to the
predator. The prey (newt) develops a defense
against the predator by becoming poisonous
to the predator. The predator (garter snake)

0022-5193/02/$35.00/0 r 2002 Elsevier Science Ltd. All rights reserved.

56 P. WALTMAN ET AL.
develops a resistance to the prey’s toxicity. The
interactions involving dangerous prey are differ-
ent from other predator–prey relationships and
result in a co-evolutionary biological arms race.
Although the snake–newt relationship is unu-
sual, predator–prey arms races are not and have
been observed in a variety of predator–prey
relationships. Brodie & Brodie (1999) and the
references cited there provide many details and
field data. We use this newt–garter snake
relationship to guide the development of the
mathematical model using continuous models
from population genetics and standard ecologi-
cal models. Adaptation in the prey produces
selection pressure on the predator. Our approach
uses continuous models that incorporate both
genetic and ecological considerations and allow
a genotype of the prey to be lethal to some
genotypes of the predator. Although compli-
cated, numerical simulations can be easily
carried out. These simulations are presented in
a sequence of graphs. The model can also include
the more typical arms race: ‘‘the fox lineage may
evolve improved adaptations for catching rab-
bits and the rabbit lineage improved adaptations
for escaping’’ (quoted from Dawkins & Krebs,
1979).
For the prey’s growth, we use the logistic
equation, one of the building blocks of popula-
tion ecology. A derivation and the fitting of a
great deal of biological data can be found in
Hutchinson (1978, Chapter 1). The equation is
also studied in standard elementary differential
equations course and appears in many texts, for
example in Abell & Braselton (2000). We add to
logistic growth one of the standard prey capture
functions. There is a great deal of literature on
predator–prey models: Freedman (1980) devotes
a chapter to such Kolmogorov models and our
beginning, ecological model is a special case of
those considered there. We then seek to add the
genetics to the model in such a way that when
capture parameters are all equal, the model
reduces to the basic, well-established, predator–
prey equations.
A broad overview of predator–prey interac-
tions and co-evolution is given by Abrams
(2000). Deterministic genetic models were devel-
oped in a fundamental paper by Nagylaki &
Crow (1974) and have been used by Beck (1982,
1984) in a model of cystic fibrosis and in a model
of co-evolution and by Beck et al. (1982, 1984) in
infectious disease models. There is also work,
using the approach of Nagylaki and Crow for
models of growth with genotypic fertility differ-
ences, Hadeler & Lieberman (1975), Butler et al.
(1981), Hadeler & Glas (1983). So (1986) and
Josic (1997). Discrete models with fertility
differences are considered in Doebeli (1997)
and Doebeli & de Jong (1998). Freedman &
Waltman (1978). So & Freedman (1986), Freed-
man et al. (1987) and So (1990) use continuous
formulation in a model of predator–prey systems
where only the genetics of the prey is modeled. A
discussion of the general topic of an arms race
can be found in Dawkins & Krebs (1979) where
many examples are mentioned as well as Epstein
(1997). The predator–prey dynamics by their
classification is ‘‘asymmetric’’ or ‘‘attack-
defense’’ type and, of course, inter-specific. The
fact that the prey is dangerousFthe prey can
kill the predatorFdifferentiates our model of
an arms race from those involving mimicry or
physical improvements of the prey or the
predator, like those referenced above.
Although, introducing genetics into the logis-
tic equations is fairly straightforward, incorpor-
ating them into the predator systems is more
difficult because one of the basic assumption
of predator–prey systems is that growth comes
from prey capture, not just the quantity of
predators. We believe that the model of predator
growth through prey capture, when the capture
rates differ, is new.
2. The Basic Model
We begin with a standard predator–prey
equation of Kolmogorov type
 x  mxy
x0 ¼ ax 1  
K aþx
 
0 mx
y ¼y s : ð1Þ
aþx
The basic working assumption is that when the
genetics, introduced below, are not relevant to
the predator–prey interactions, then the system
with genetics should reduce to eqn (1). Equation
(1) often occurs with other parameters, for
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 57

y
0.8
0.5
0.6 x
0.4
0.4 0.3

0.2 0.2
y 0.1
(a) t
10 20 30 40 50 x
(b) 0 .10. 20. 30. 40. 5

Fig. 1. (a) Predator–prey time course: parameters as above with m ¼ 2:5; a ¼ 0:37; s ¼ 1:1; K ¼ 1; a ¼ 1:2: (b) Phase-
plane plot corresponding to (a).

example, a conversion constant to convert the form

captured prey to predator biomass or an extra
rate constant in front of the predator equation. 1 x2 2 a
x01 ¼ a x1 þ  x1 x;
These may be scaled out and we assume that x 2 K
scaling has been done. One could also scale the
parameter K out of the system but we choose to 1 x2  x2  a
retain it. The prey capture term is of Monod type x02 ¼ 2a x1 þ x3 þ  x2 x;
x 2 2 K
(also called a Hollings Type II response) which is
usually justified as allowing for prey-handling
1 x2 2 a
time. m reflects the difficulty of prey capture, s is x03 ¼ a x3 þ  x3 x ð3Þ
the death rate of the predator in the absence of x 2 K
prey, and a and K are discussed below when with initial conditions x1 ð0Þ ¼ x10 ; x2 ð0Þ ¼ x20 ;
considering the logistic equation. Figure 1(a) x3 ð0Þ ¼ x30 :
shows the time course of the predator–prey Built into this format is the interpretation of
system. We have selected the parameters to be the logistic equation that growth is represented
in the oscillatory range for this system. The by the linear term (ax) and (natural) death by the
corresponding phase plane plot is given in quadratic term (x2 =K) although other interpre-
Fig. 1(b). tations are possible (and will not affect the work
We begin with the growth of the prey without here). The principal result of Freedman &
a predator which, is assumed in eqn (1) to follow Waltman (1978) for eqn (3) is that the three
a basic logistic equation, genotypes evolve (have limits) in the ratio ðx1 :
 x2 : x3 Þ ¼ ðc2 : 2c : 1Þ where
x
x0 ¼ ax 1  : ð2Þ
K x10 þ 12x20
c¼ :
x30 þ 12x20
K is called the ‘‘carrying capacity’’ and reflects
the level to which the prey will grow if there This reflects the basic Hardy–Weinberg principle
are no predators. a reflects the rate at which for random mating in an asymptotic form.
the prey approaches the carrying capacity. Adding the three equations in eqn (3) with xðtÞ ¼
The prey are divided into three classes represent- x1 ðtÞ þ x2 ðtÞ þ x3 ðtÞ recovers the logistic equa-
ing three genotypes and random mating is tion (2), that is, the total grows logistically.
assumed. The usual classification of a one locus, In Fig. 2(a), we illustrate the evolution of the
two allele problem is labeled AA; Aa; aa; genotypes. The total (x) follows a typical logistic
corresponding to two choices of an allele at equation. We take the parameters (arbitrarily) to
one location. There is a deterministic formula- be a ¼ 1:2; and K ¼ 1: As to be expected, growth
tion of the evolution of the genotypes, Nagylaki occurs according to Hardy–Weinberg propor-
& Crow (1974), Butler et al. (1981) which takes tions without selection. Selection is illustrated by
58 P. WALTMAN ET AL.

1 1.2
x x
0.8 1 AA
0.8
0.6
0.6
0.4 Aa
AA 0.4
0.2 0.2 Aa
aa aa
t t
(a) 5 10 15 20 (b) 5 10 15 20

Fig. 2. (a) The evolution of three genotypes following logistic growth. The initial conditions are x1 ð0Þ ¼ 0:06; x2 ð0Þ ¼
0:1; x3 ð0Þ ¼ 0:02: (b) The evolution of three genotypes where x1 has an increase in its growth rate from 1.2 to 1.31. The

Prey Genotypic Proportions

0.2
0.175 0.8 AA
0.15
0.6
0.125
0.1 AA
0.4
0.075 Aa
0.05 0.2
0.025 Aa
aa aa
t t
(a) 10 20 30 40 50 (b) 10 20 30 40 50

Fig. 3. (a) The prey population broken into three genotypes with x1 ð0Þ ¼ 0:02; x2 ð0Þ ¼ 0:002; x3 ð0Þ ¼ 0:001; and yð0Þ ¼
0:1: (b) The genotypic frequencies corresponding to (a).

1 x2  x2  a x2 mxy
increasing the value of a in its first occurrence x02 ¼ 2a x1 þ x3 þ  x2 x  ;
in the equation for x1 : Keeping all of the other x 2 2 K x aþx
parameters the same and replacing the value a ¼
1:2 by 1:31 in its first occurrence in the equation 1 x2 2 a x3 mxy
x03 ¼ a x3 þ  x3 x  ;
for x1 yields Fig. 2(b). Clearly, the AA genotype x 2 K x aþx
is replacing the others in the mix. This is
expected and shown here to illustrate that the  
mx
continuous version of the evolution of genotypes y0 ¼ y s : ð4Þ
aþx
matches the discrete one.
We now add a predator to the system that Again, adding the prey equations produces a
preys equally on each of the three genotypes. As standard predator–prey system (1).
with all simple models we also assume that the In Fig. 3(a), the prey is broken in three
predator feeds exclusively on this prey. While genotypes, showing that each oscillates as
that is not realistic, to assume otherwise either expected. The initial conditions are x1 ð0Þ ¼
requires that one know the other prey and 0:02; x2 ð0Þ ¼ 0:002; x3 ð0Þ ¼ 0:001; and yð0Þ ¼
add them to the model or to assume that the 0:1; the system is prejudiced in favor of AA:
predator also has a logistic growth term in Since the predominant prey is AA we have
addition to the prey, which does not allow one to broken the graph at the top in order to show the
separate out the effects of this particular prey. others because aa has such small numbers it does
One hopes that the simple model captures the not show clearly in the graph. Figure 3(b), which
essence of the effect even if it is not totally is uninteresting in this context but will be
realistic in modeling the natural situation. The important in the discussion that follows, plots
equations become the evolution of the relative proportions of the
three genotypes. After a slight adjustment at
1 x2 2 a x1 mxy the beginning, the frequencies are constant.
x01 ¼ a x1 þ  x1 x  ;
x 2 K x aþx The figures illustrate that breaking the prey
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 59

population into three genotypes preserves the We take m1 ¼ m2 ¼ 2:5 as in the previous
expected predator–prey behavior. In what fol- computations but reduce the capture rate for
lows we will disturb this basic predator–prey the aa genotype (x3 ) by 10% to 2:25%: This
relations to formulate the models of the arms means that x3 is more difficult or less desirable to
race. capture corresponding to a genetic trait as noted
above. For example, x3 may be faster or quicker
3. Elusive and Poisonous Prey at turning when being pursued making it more
difficult to capture; x3 may exhibit coloring or
In this section, we let the prey develop a a marking that causes the prey y to find it
defense against the predator. Defenses are undesirable; x3 may taste bad so that when y
evolutionary traits that can be physical (faster, captures it, y ‘‘spits it’’ out leaving it unharmed,
quicker turning, etc.), passive (camouflage), which has been observed in some newt–snake
offensive (poisonous) or a combination of these interactions; x3 may secrete a chemical that
and give the prey a survival advantage. We also causes it to smell bad to y and so on. We plot the
consider the possibility that a poisonous prey evolution of the predator and the three prey
is able to alter the predator capture rate. We genotypes in separate graphs. To make this
assume that the trait is genetic and by conven- evolution more dramatic we also plot the
tion we let the ‘‘special’’ prey be of the aa type, evolution of the relative frequencies in Fig. 4(c).
denoted by x3 ðtÞ: Complete dominance of A is The elusive prey has become the established
assumed so that neither AA nor Aa are elusive or prey; contrast Fig. 4(c) with Fig. 3(b). The
dangerous (produce a toxin). We illustrate how predator survives but at a lower level.
these different strategies affect the evolution of A more serious capture avoidance can lead to
the genotypes. We first relabel the m parameter the extinction of the predator. If instead of
in eqn (4) to be m1 ; m2 ; m3 ; respectively, to reducing the capture rate by 10%; the improve-
produce the system ment in the prey’s ability to avoid capture
reduces the capture rate to 50%; then the
1 x2 2 a x1 m1 xy predator becomes extinct. The prey again is
x01 ¼ a x1 þ  x1 x  ;
x 2 K x aþx dominated by aa although the route is not
smooth as in the previous case. Figures 5(a) and
(b) show the predator and prey time courses and
1 x2  x2  a x2 m2 xy
x02 ¼ 2a x1 þ x3 þ  x2 x  ; Fig. 5(c) shows the evolution of the relative
x 2 2 K x aþx frequencies.
Of course, a 50% improvement represents a
1 x2 2 a x3 m3 xy drastic step. Although we do not consider the
x03 ¼ a x3 þ  x3 x  ; case here, one might want to model the
x 2 K x aþx
improvement as a separate process, allowing
gradual improvement in avoiding capture. This
  could probably be done with a multi-locus model
0 m1 x1 þ m2 x2 þ m3 x3
y ¼y s : ð5Þ which would introduce considerable complexity.
aþx

y Predator Prey Genotypic Proportions

0.7 1
0.4 0.6 AA
0.8
0.3 0.5 aa
0.4 AA aa 0.6
0.2 0.3 0.4 Aa
0.1 0.2
0.1 Aa 0.2
t t t
(a) 100 200 300 400 (b) 100 200 300 400 (c) 100 200 300 400

Fig. 4. Evolution of the (a) predator and (b) prey with an elusive prey and m1 ¼ m2 ¼ 2:5; m3 ¼ 2:25: (c) Evolution of
the relative frequencies of an elusive prey.
60 P. WALTMAN ET AL.

y Predator Prey Genotypic Proportions

1
0.4 0.8 aa 0.8 aa
0.3 0.6 0.6
0.2 0.4 0.4
0.1 0.2 Aa 0.2 Aa
AA AA
t t t
(a) 20 40 60 80 100 (b) 20 40 60 80 100 (c) 20 40 60 80 100

Fig. 5. Evolution of the (a) predator and (b) prey with an elusive prey (aa) and with m1 ¼ m2 ¼ 2:5; m3 ¼ 1:125: (c)
Evolution of the relative frequencies of an elusive prey.

y Predator Prey Genotypic Proportions

0.2
0.4 0.175 0.8 AA
0.15
0.3 0.125 0.6
AA
0.1 Aa 0.4
0.2 0.075
0.1 0.05 0.2 Aa
0.025 aa
t aa
t t
(a) 20 40 60 80 100 (b) 20 40 60 80 100 (c) 20 40 60 80 100

Fig. 6. (a) Evolution of the (a) predator and (b) prey with a poisonous prey with low initial density. (c) Evolution of the
relative frequencies of a poisonous prey with low initial density.

We now turn to another improvement in the The Monod term, formerly reflecting the added
development of the prey, a poisonous genotype. growth of the predator by capturing x3 ; now no
The snake and newt system discussed in the longer does so and, in addition, the capture of x3
Introduction is the prime example of such a contributes to the death rate of the predator. We
system. We again assume that the poisonous use the same parameters as before with initial
prey is represented by the aa genotype and that conditions x1 ð0Þ ¼ 0:02; x2 ð0Þ ¼ 0:002; x3 ð0Þ ¼
its consumption is fatal to the predator. As 0:001; which represents a rare, poisonous prey.
noted in the Introduction, this is an extreme We plot the predator evolution, prey evolution,
assumption because most newts may only render and the evolution of the relative frequencies
the snake immobile for a while and the snake is in Fig. 6. Although the predator is diminished
subject to other forces while in this state. A slightly, almost nothing changes from the
‘‘correction’’ factor could be entered in the original model (4). From the standpoint of
removal term for the predator but the value of the predator, this is an acceptable ecosystem.
such a correction factor seems unlikely to be The poisonous prey is providing a type of
known. The equations for a poisonous prey take ‘‘group defense’’ with little effect.
the form However, if the initial density of the poisonous
prey is high, the results are disastrous for the
1 x2 2 a x1 mxy predator. The same three plots follow in Fig. 7
x01 ¼ a x1 þ  x1 x  ;
x 2 K x aþx except that now the initial conditions are x1 ð0Þ ¼
0:02; x2 ð0Þ ¼ 0:002; and x3 ð0Þ ¼ 0:03:
1 x2  x2  These figures illustrate what could happen if
x02 ¼ 2a x1 þ x3 þ the density of the poisonous prey becomes high.
x 2 2
However, if the genetic event occurs from a
a x2 mxy random mutation that makes the prey lethal to
 x2 x  ;
K x aþx the predator but does not give the prey a survival
advantage, Fig. 6 shows that the poisonous
1 x2 2 a x3 mxy prey will not achieve high enough densities to
x03 ¼ a x3 þ  x3 x  ; eliminate the predator.
x 2 K xaþx
Now suppose the poisonous prey has a slight
  advantage, like those described earlier, that
mðx1 þ x2 Þ mx3 y makes it less susceptible to being captured. The
y0 ¼ y s  : ð6Þ
aþx aþx model is adjusted to take this into consideration
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 61

y Predator Prey Genotypic Proportions

1 1
0.5 Aa
0.8 0.8
0.4 aa
0.6 0.6 Aa 0.3
0.4 0.4 aa 0.2 AA
0.2 0.2 AA 0.1
t t t
(a) 5 10 15 20 (b) 5 10 15 20 (c) 5 10 15 20

Fig. 7. (a) Evolution of the (a) predator and (b) prey where aa corresponds to the poisonous prey with high initial
density. (c) Evolution of the relative frequencies of a poisonous prey with high initial density.

y Predator Prey Genotypic Proportions

1 1
0.8
0.8 0.8
0.6
0.6 0.6 Aa
Aa 0.4 aa
0.4 0.4
AA
0.2 0.2 0.2 AA
aa
t t t
(a) 50 100 150 200 (b) 50 100 150 200 (c) 50 100 150 200

Fig. 8. Evolution of the (a) predator and (b) prey with a poisonous prey low initial density and a slight advantage in
capture avoidance. aa corresponds to the poisonous prey with low initial density and a slight advantage in capture
avoidance. (c) Evolution of the relative frequencies of a poisonous prey with low initial density and a slight advantage in
capture avoidance.

and becomes to dominate, causing extinction of the predator.

(cf. Fig. 8 with Fig. 6).
1 x2 2 a x1 m1 xy If the poisonous prey becomes established at a
x01 ¼ a x1 þ  x1 x  ;
x 2 K x aþx high level it eliminates the predator. If it has an
advantage in avoiding capture (or detection),
1 x2  x2  it first out-competes its rivals because of the
x02 ¼ 2a x1 þ x3 þ advantage of a lower capture rate, and thus
x 2 2
moves from an extremely low level to a
a x2 m2 xy significant level, eliminating the predator.
 x2 x  ; Snakes have been observed ‘‘spitting out’’ the
K x aþx
newt (Brodie, private comm., 2001) which would
give it a slightly diminished capture rate, i.e.,
1 x2 2 a x3 m3 xy m3 omaxfm1 ; m2 g:
x03 ¼ a x3 þ  x3 x  ;
x 2 K x aþx If the predator does not respond with a genetic
alteration, it will become extinct. Hence, one has
  the next step in the arms race.
0 m1 x1 þ m2 x2 m3 x 3 y
y ¼y s  : ð7Þ
aþx aþx 4. The Arms Race
We take m1 ¼ m2 ¼ 2:5 and m3 ¼ 2:25 and We now let the predator, y; evolve with
initial conditions x1 ð0Þ ¼ 0:02; x2 ð0Þ ¼ 0:002; immunity to the dangerous prey, x: (We have
and x3 ð0Þ ¼ 0:001: This reflects genetic change in mind the example of the poisonous newt and
by giving the poisonous prey a slight advantage the garter snake discussed in the Introduction
in capture avoidance but also a very low initial where the garter snake acquires resistance to the
size, as would be expected after an advantageous toxin produced by the newt, but the model will
mutation has occurred. Figure 8 shows an have wider applicability.) At the key locus we
opening salvo in the biological arms race: even denote the genotype for the predator as BB; Bb;
though x3 has a very low initial size its and bb and label the concentrations of each by
advantage in avoiding the predator allows x3 y1 ; y2 ; and y3 ; respectively. We assume that bb is
62 P. WALTMAN ET AL.

the resistant genotype. The redistribution of the mx

genotypes due to random mating is much more y02 ¼ 2
ða þ xÞðy1 þ y2 þ y3 Þ
delicate than that of the prey discussed pre-
 y2  y2 
viously. In the case of the model of the prey, the y1 þ y3 þ  sy2 ;
growth rate is constant (a) so the increase in prey 2 2
depends on the numbers in each class, but in the
case of the model of the predator, growth mx  y2 2
follows from prey capture. y03 ¼ y3 þ sy3 : ð9Þ
ða þ xÞðy1 þ y2 þ y3 Þ 2
Since the model is somewhat complicated, we
develop it in stages in hopes of achieving greater If one adds the last three equations in the system,
clarity for the form of the final model. If x then using y ¼ y1 þ y2 þ y3 reproduces the basic
denotes the concentration of the prey (we ignore predator–prey equations (8).
genotypes at first) and y; the predator, the We now rewrite the system with the full prey
increase in the concentration of the predator is genotypes, borrowing from the prey equations
driven by the Monod term developed in the previous section. The model
takes the form (where we are expansive in the
mxy notation to illustrate the effect of incorporating
:
aþx the genotypes for both predator and prey)

a  x2 2 ax1 ðx1 þ x2 þ x3 Þ
We seek to incorporate the distribution of
x01 ¼ x1 þ 
genotypes using this term. As long as there are x1 þ x2 þ x3 2 K
no genotypic differences affecting predator–prey
reactions, a basic hypothesis is that one must be mx1 ðy1 þ y2 þ y3 Þ
 ;
able to recombine the genotypes into the basic a þ x1 þ x2 þ x3
predator–prey system which we have assumed
from the beginning to be of the form 
2a x2  x2 
 x02 ¼ x1 þ x3 þ
0 x  mxy x1 þ x2 þ x3 2 2
x ¼ ax 1   ;
K aþx
ax2 ðx1 þ x2 þ x3 Þ mx2 ðy1 þ y2 þ y3 Þ
    ;
0 mx K a þ x1 þ x2 þ x3
y ¼y s : ð8Þ
aþx
a  x2 2 ax3 ðx1 þ x2 þ x3 Þ
x03 ¼ x3 þ 
We assume that the prey captured by yi is the x1 þ x2 þ x3 2 K
fraction of the total catch that yi represents in
the population: mx3 ðy1 þ y2 þ y3 Þ
 ;
a þ x1 þ x2 þ x3
yi mxy
:
y aþx
mðx1 þ x2 þ x3 Þ
(The y terms cancel.) The basic assumption of y01 ¼
ða þ x1 þ x2 þ x3 Þðy1 þ y2 þ y3 Þ
predator–prey models is that the captured prey
translates to growth of the predator. For a single  y2 2
prey, x; this leads to the equations y1 þ sy1 ;
2
 x  mxðy1 þ y2 þ y3 Þ
x0 ¼ ax 1   ; mðx1 þ x2 þ x3 Þ
K aþx y02 ¼ 2
ða þ x1 þ x2 þ x3 Þðy1 þ y2 þ y3 Þ
 y2 2 
y01 ¼
mx
y1 þ sy1 ; y2  y2 
ða þ xÞðy1 þ y2 þ y3 Þ 2 y1 þ y3 þ  sy2 ;
2 2
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 63

mðx1 þ x2 þ x3 Þ The variables in this model do not add to

y03 ¼ reproduce eqn (8) unless m1 ¼ m2 ¼ m3 : How-
ða þ x1 þ x2 þ x3 Þðy1 þ y2 þ y3 Þ
ever, eqn (10) could be used to model the
 y2 2 ‘‘elusive’’ prey case discussed previously.
y3 þ sy3 : We now turn to the main development, a
2
model for the poisonous prey and the resistant
Again, if one adds the first three equations and predator. We remind the reader that we take the
the last three equations, the basic predator–prey poisonous prey to be the aa genotype (denoted
equation (8) is reproduced. At this point the by x3 ) and the resistant predator to be the bb
model with three genotypes for each has not genotype (denoted by y3 ). The consumption of
changed. However, suppose that each prey has a x3 by y1 or y2 does not lead to added growth (so
different capture rate which we denote by m1 ; the term must be subtracted from the preceding
m2 ; and m3 ; respectively. Then we produce a new model) and does lead to increased death (so a
model for predator–prey interactions. Note the term must be added to the intrinsic death rate).
assumption that the capture rate is dependent on However, the consumption of x3 by y3 does lead
the x genotype and not on the y genotype, that to increased growth. Thus, the distribution of
is, all y genotypes capture a given x genotype genotypes will not be as convenient as that
equally. To conserve notation, we return to expressed in eqn (9). For y1 ; the growth term
using x ¼ x1 þ x2 þ x3 and y ¼ y1 þ y2 þ y3 for becomes [corresponds to eqn (9)]
the respective sum of genotypes if there is no
 
multiplication by an mi : For brevity, we let Tðx1 ; x2 ; 0Þy1 1 Tðx1 ; x2 ; 0Þy2 2
þ
Tðx1 ; x2 ; x3 Þ ¼ m1 x1 þ m2 x2 þ m3 x3 : aþx 2 aþx
Tðx1 ; x2 ; x3 Þy
The model takes the form aþx
a x2 2 ax1 x m1 x1 y
x01 ¼ x1 þ   ; 
x 2 K aþx Tðx1 ; x2 ; 0Þ2 y2 2
¼ y1 þ :
ða þ xÞTðx1 ; x2 ; x3 Þy 2
2a x2  x2  ax2 x m2 x2 y
x02 ¼ x1 þ x3 þ   ;
x 2 2 K aþx Similarly, the term for y2 becomes
  
Tðx1 ; x2 ; 0Þy1 1 Tðx1 ; x2 ; 0Þy2 Tðx1 ; x2 ; x3 Þy3 1 T ðx1 ; x2 ; 0Þy2
þ þ
aþx 2 aþx aþx 2 aþx
Tðx1 ; x2 ; x3 Þy
aþx
1 1
ðTðx1 ; x2 ; 0Þy1 þ Tðx1 ; x2 ; 0Þy2 ÞðTðx1 ; x2 ; x3 Þy3 þ Tðx1 ; x2 ; 0Þy2 Þ
¼ 2 2
Tðx1 ; x2 ; x3 Þyða þ xÞ

a x2 2 ax3 x m3 x3 y while that of y3 becomes

x03 ¼ x3 þ   ;
x 2 K aþx  
Tðx1 ; x2 ; x3 Þy3 1Tðx1 ; x2 ; 0Þy2 2
T ðx1 ; x2 ; x3 Þ y2 2 þ
y01 ¼ y1 þ sy1 ; aþx 2 aþx
ða þ xÞy 2
Tðx1 ; x2 ; x3 Þy3
T ðx1 ; x2 ; x3 Þ y2  y2  aþx
y02 ¼ 2 y1 þ y3 þ  sy2 ;
ða þ xÞy 2 2
1
T ðx1 ; x2 ; x3 Þ y2 2 ðTðx1 ; x2 ; x3 Þy3 þ Tðx1 ; x2 ; 0Þy2 Þ2
y03 ¼ y3 þ sy3 : ¼ 2
ða þ xÞy 2 Tðx1 ; x2 ; x3 Þða þ xÞy
64 P. WALTMAN ET AL.

We now incorporate these ideas into the model oscillatory case and the initial conditions
of a biological arms race: reflect zero poisonous prey and resistant preda-
a x2 2 ax1 x m1 x1 y tors but a very small number of heterozygotes
x01 ¼ x1 þ   ; carrying one copy of the respective alleles. An
x 2 K aþx
even lower number of heterozygotes (reflecting a
2a x2  x2  ax2 x m1 x2 y random perturbation) would present the same
x02 ¼ x1 þ x3 þ   ;
x 2 2 K aþx result but with a longer time-scale. An intuitive
a x2 2 ax3 x m3 x3 y explanation begins with the fact that x1 and y1
x03 ¼ x3 þ   ; dominate the initial configuration that would be
x 2 K aþx
in a oscillatory regime, if they were the only
Tðx1 ; x2 ; 0Þ2  y2 2 organisms present. Gradually, because of the
y01 ¼ y1 þ lower capture rate, x3 ; the poisonous prey, out-
ða þ xÞTðx1 ; x2 ; x3 Þy 2
competes x1 and x2 and lowers the predator
m3 x 3 y 1 pressure by increasing the death rate of y1 and y2 :
  sy1 ;
aþx This allows for the emergence of y3 ; the resistant
predator. Finally, y3 and x3 coexist in an
1 oscillatory regime. This result is more dramati-
Tðx1 ; x2 ; 0Þy1 þ Tðx1 ; x2 ; 0Þy2
y02 ¼2 2 cally portrayed in Fig. 10 which plots total prey
Tðx1 ; x2 ; x3 Þy x ¼ x1 þ x2 þ x3 and total predators, y ¼ y1 þ
1 y2 þ y3 against time in the middle part of the
Tðx1 ; x2 ; x3 Þy3 þ Tðx1 ; x2 ; 0Þy2 evolution.

2
aþx The reader is reminded that x and y are sums
m3 x3 y2 of components of a system of differential
  sy2 ;
aþx equations and do not satisfy a two-dimensional
system as Fig. 10 might suggest. However, if one
1
ðTðx1 ; x2 ; x3 Þy3 þ Tðx1 ; x2 ; 0Þy2 Þ2 accepts that the functions x1 ðtÞ; x2 ðtÞ; y1 ðtÞ; and
y03 ¼ 2  sy3 : y2 ðtÞ all tend to zero as t tends to infinity, as
Tðx1 ; x2 ; x3 Þða þ xÞy
the computations suggest, then eqn (11) is an
ð11Þ
asymptotically autonomous system with limiting
Figure 9 shows a typical arms race. The equations of the form (8) with m ¼ m3 and the
parameters have been selected to show an other parameters as specified. Of course, the

Prey
0.6
0.5
0.4 x1
0.3
0.2
0.1 x2 x3
t
1000 2000 3000 4000 5000

Predator
0.6
0.5
0.4
0.3 y1
0.2
0.1 y2 y3
t
1000 2000 3000 4000 5000

Fig. 9. Co-evolution of predator and prey using parameter values m1 ¼ m2 ¼ 2:5; m3 ¼ 2:45; a ¼ 0:37; s ¼ 1:1; a ¼ 1:2
and initial conditions x1 ð0Þ ¼ 0:6; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 0; y1 ð0Þ ¼ 0:6; y2 ð0Þ ¼ 0:02; and y3 ð0Þ ¼ 0:01:
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 65

0.3

3250 0.25
y
3500 0.2
0

3750 0.5
0.4
t 4000
0.3 x
0.2
4250

Fig. 10. Co-evolution of the total predator population, y ¼ y1 þ y2 þ y3 ; and the total prey population, x ¼ x1 þ x2 þ
x3 ; using the same parameter values and initial conditions as in Fig. 9.

Prey
0.8
0.6 x1 x2 x3
0.4
0.2
t
200 400 600

Predator
0.6
0.5
0.4 y1 y3
0.3
0.2
0.1
y2
t
200 400 600

Fig. 11. Co-evolution of predator and prey using parameter values m1 ¼ m2 ¼ 2:5; m3 ¼ 2; a ¼ 0:37; s ¼ 1:1; a ¼ 1:2
and initial conditions x1 ð0Þ ¼ 0:6; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 0; y1 ð0Þ ¼ 0:6; y2 ð0Þ ¼ 0:02; and y3 ð0Þ ¼ 0:

predator–prey relationship need not be oscilla- An important question remains: should

tory and Figs 11 and 12 show a similar evolution
but with the parameters chosen so that the
limiting system is in a steady state.
If the parameter m3 is lowered even
farther, it is possible for the prey to cause
extinction of the predator as illustrated in
Fig. 13. Essentially, x3 out-competes x1 and x2
but y3 cannot exist on x3 alone and thus becomes
extinct. The levels are so low that they do not
really show on the graph; however, the prey sum
is tending to K ¼ 1 so no predators will be
present.
the capture rate for x3 by y3 be the same as that
by the others? It is possible to alter the model
to allow the capture rate to be dependent on
both the prey and the predator genotypes by
replacing mi by mij : This is a major increase
in complexity and, in addition, it is unlikely that
such parameters could be realistically determined.
However, to answer the question as to
whether the dominant conclusion is due to the
low capture rate, we make a final alteration
to the model to allow the capture of x3 (only) to
be different for y3 : To avoid an unnecessarily
66 P. WALTMAN ET AL.

0.6
0
0.4
y
0.2
0
200
0
1
t 0.75
400
0.5
0.25 x

600 0

Fig. 12. Co-evolution of predator y ¼ y1 þ y2 þ y3 and prey x ¼ x1 þ x2 þ x3 using the same parameter values and
initial conditions as in Fig. 11.

Prey
0.6
0.5
0.4 x2
0.3 x1
0.2 x3
0.1
t
100 200 300

Predator

0.6
0.5
0.4 y1
0.3
0.2
0.1 y2
t
100 200 300

Fig. 13. Co-evolution of predator and prey using parameter values m1 ¼ m2 ¼ 2:5; m3 ¼ 1:7; a ¼ 0:37; s ¼ 1:1; a ¼ 1:2
and initial conditions x1 ð0Þ ¼ 0:6; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 0; y1 ð0Þ ¼ 0:6; y2 ð0Þ ¼ 0:02; and y3 ð0Þ ¼ 0:

complex model, we allow just two capture We incorporate this into model (11) to obtain
rates: m is the capture rate of x1 and x2 by all
predators and of x3 by y3 : We retain the capture a x2 2 ax1 x mx1 y
x01 ¼ x1 þ   ;
rate notation m3 for the capture of x3 by y1 x 2 K aþx
and y2 : The total capture of prey by predators is
given by 2a x2  x2  a x2 x mx2 y
x02 ¼ x1 þ x3 þ   ;
x 2 2 K aþx
mxy3 þ ðy1 þ y2 ÞTðx1 ; x2 ; x3 Þ
aþx
; a x2 2 ax3 x x3 ðm3 ðy1 þ y2 Þ þ my3 Þ
x03 ¼ x3 þ   ;
x 2 K aþx
where
m2 ðx1 þ x2 Þ2
y01 ¼
Tðx1 ; x2 ; x3 Þ ¼ mx1 þ mx2 þ m3 x3 : ða þ xÞðmxy3 þ ðy1 þ y2 ÞTðx1 ; x2 ; x3 ÞÞ
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE 67

Prey
0.6
0.5 x1
0.4
0.3
0.2
0.1 x2 x3
t
1000 2000 3000 4000 5000

Predator
0.6
0.5
0.4
0.3 y1
0.2
0.1 y3
y2
t
1000 2000 3000 4000 5000

Fig. 14. Co-evolution of predator and prey using parameter values m ¼ 2:5; m3 ¼ 2:45; a ¼ 0:37; s ¼ 1:1; a ¼ 1:2 and
initial conditions x1 ð0Þ ¼ 0:6; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 0; y1 ð0Þ ¼ 0:6; y2 ð0Þ ¼ 0:02; and y3 ð0Þ ¼ 0:01:

0.3
3000
0.25
0.2
0 y
3500
0.15
0
0.5
4000 0.4
t
0.3 x
0.2
4500

Fig. 15. Co-evolution of predator y ¼ y1 þ y2 þ y3 and prey x ¼ x1 þ x2 þ x3 using the same parameter values and
initial conditions as in Fig. 14.

 y2 2 m3 x3 y1

y1 þ   sy1 ;
2 aþx 0.35
Predator and Prey

1
mðx1 þ x2 Þðy1 þ y2 Þ 0.3 y3
0
y2 ¼ 2 2
mxy3 þ ðy1 þ y2 ÞTðx1 ; x2 ; x3 Þ 0.25

1 0.2 x1
ðmxy3 þ mðx1 þ x2 Þy2 m x y x2
2 
3 3 2
 sy2 Þ;
0.15

aþx aþx 0.1

1
ðmxy3 þ mðx1 þ x2 Þy2 Þ2 0.05
x3
y03 ¼ 2  sy3 :
ðmxy3 þ ðy1 þ y2 ÞT ðx1 ; x2 ; x3 ÞÞða þ xÞ 9960 9970 9980 9990 10000
t

ð12Þ Fig. 16. Figure 14 for 9950ptp10 000:

68 P. WALTMAN ET AL.

Prey
0.6
0.5
x2
0.4
x1
0.3
0.2 x3
0.1
t
100 200 300 400 500

Predator
0.6
0.5
0.4
0.3 y1 y3
0.2
0.1 y2
t
100 200 300 400 500

Fig. 17. Co-evolution of predator and prey using parameter values m ¼ 2:1; m3 ¼ 1:9; a ¼ 0:37; s ¼ 1:1; a ¼ 1:2 and
initial conditions x1 ð0Þ ¼ 0:6; x2 ð0Þ ¼ 0:02; x3 ð0Þ ¼ 0; y1 ð0Þ ¼ 0:6; y2 ð0Þ ¼ 0:02; and y3 ð0Þ ¼ 0:01:

1
0.8
0
0.6
200 0.4
0 y
0.2
0
400
0
600 0.8
0.6
t 0.4 x
800
0.2
0
1000

Fig. 18. Co-evolution of the total prey and total predator using parameter values as in Fig. 17.

We repeat the first two simulations above for
eqn (12) with the same parameters and initial
conditions; the results are shown in Figs 14
and 15.
We next plot a short time period to illustrate
the periodic nature of the final outcome in
Fig. 16.
In this case, all three prey genotypes survive
but only the resistant predator, y3 ; survives. The
arms race ends as it had begun but with only
the resistant predator surviving. This is to be
expected since now y3 feeds equally on all prey
genotypes, so x3 cannot eliminate its competitors
although it does eliminate the non-resistant
predators. However, it does increase its relative
frequency during the time that y1 and y2
dominate the mix. Thus its final proportion is
much higher than if the same problem was
simulated with only the resistant predator
present (and the same prey initial conditions).
The choice between eqns (11) and (12) could be
decided by observable data. If a territory could
be found where the resistant predator dominates
but the non-poisonous prey survives in quantity,
 MATHEMATICAL MODEL OF A BIOLOGICAL ARMS RACE
then eqn (12) is supported. Figure 15 shows the
three-dimensional plot of total prey and total
predators against time.
The result need not be oscillatory as Figs 17
and 18 illustrate.
5. Conclusion
We have provided a model of a biological
arms race motivated by a predator–prey system
where the prey develops the ability to produce a
toxin against the predator and then the predator
responds with resistance to the toxin. We have as
a biological model that of the newt–garter-snake
relationship studied by Brodie and Brodie (1999).
Our model and simulations seem to capture most
of the points discussed there. The principal
modeling difficulty was to expand the determi-
nistic genetic modeling to the predator dynamics
where growth depends on prey capture.
One thinks of the genetic change as occurring
by a random mutation that we model by taking a
very small initial condition in the differential
equations. The simulations seem to show that the
poisonous prey cannot become established in
large enough numbers to affect the system with-
out having an advantage. Then we assumed that
the special prey has an advantage with respect to
prey capture. This is also observed in the newt–
snake system: the poisonous prey sometimes
escapes the predator’s grasp alive. The size of
the advantage determines how rapidly the system
evolves but any advantage will lead to establish-
ment. Once the poisonous prey is established in
sufficient quantities, the non-resistant predators
face extinction (although this is an artifact of our
assumption that it lives only on this prey) and the
resistant predator does not have this added death
rate and thus thrives. The arms race ends as it
began except with slightly altered players. The
next step in the arms race requires a new
mutationFperhaps an altered poison.
The model suffers the usual deficiencies of
predator–prey models in that it assumes the
predator lives exclusively on the prey. It also
presumes that the change in the genetic trait is at
one locus whereas major alterations usually
reflect multiple loci. Additional loci can be
included in the model at a considerable increase
in complexity but well within the reach of
69
modern computers. In the model of So (1990)
two loci are considered and the number of
equations increase from three to nine. Addi-
tional alleles can also be included. However, the
assumptions here are no worse than those
usually associated with such systems. Genetic
improvement often comes at a cost, usually
reflected in a lower reproduction rate. For the
poisonous newt this seems not to be the case (or
it is negligible). That is not the case with the
garter snake as Brodie & Brodie (1999) show
that the resistant snake has a lower sprint
velocity that would be modeled in our system
by a change in capture rate. We feel that this can
be incorporated into a more general model,
alluded to in the main text, by making the
capture rate dependent on both predator and
prey genotypesFintroducing mij instead of mi :
We hope to do this in a later study.
All of the differential equations were solved with,
and all of the figures created with, Mathematica 4,
Wolfram Research, Inc., 1999. (see Wolfram, 1999).
The authors wish to thank Edmund D. Brodie III for
communicating his insights regarding predator–prey
arms races and to thank Martha Abell for assistance
with the graphics.
The research of Paul Waltman was supported by
National Science Foundation Grant, DMS-9801622.
REFERENCES
Abell, M. L. & Braselton, J. P. (2000). Modern
Differential Equations, 2nd Edn, Fort Worth, TX:
Harcourt College Publishers.
Abrams, P. A. (2000). The evolution of predator–prey
interactions: theory and evidence. Ann. Rev. Ecol. Syst.
31, 79–105.
Beck, K. (1982). A model of population genetics of cystic
fibrosis in the United States. Math. Biosci. 58, 243–257.
Beck, K. (1984). Coevolution: mathematical analysis of
host–parasite interactions. J. Math. Biol. 19, 63–67.
Beck, K., Keener, J. P. & Ricciardi, P. (1982). Influence
of infectious disease on the growth of a population with
three genotypes. In: Lecture Notes in Biomathematics,
Vol. 52. New York: Springer.
Beck, K., Keener, J. P. & Ricciardi, P. (1984). The effect
of epidemics on genetic evolution. J. Math. Biol. 19,
79–94.
Brodie, E. D. III & Brodie, E. D. Jr (1999). Predator prey
arms races: asymmetrical selection on predators and prey
may be reduced when prey are dangerous. BioScience 49,
557–568.
Butler, G. J., Freedman, H. I. & Waltman, P. (1981).
Global dynamics of a selection model for the growth of
a population with genotypic fertility differences. J. Math.
Biol. 12, 25–35.
70 P. WALTMAN ET AL.
Dawkins, R. & Krebs, J. R. (1979). Arms race between
and within species. Proceedings of the Royal Society of
London. Series, Biological Sciences. 205, 489–511.
Doebeli, M. (1997). Genetic variation and the persistence
of predator–prey interactions in the Nicholson–Bailey
model. J. theor. Biol. 188, 109–120.
Doebeli, M. & de Jong, G. (1998). A simple genetic model
with non-equilibrium dynamics. J. Math. Biol. 36,
550–536.
Epstein, J. M. (1997). Nonlinear Dynamics: Mathematical
Biology and Social Sciences, Santa Fe Institute Studies
in the Sciences of Complexity. Reading, MA: Addison-
Wesley.
Freedman, H. I. (1980). Deterministic Mathematical
Models in Population Ecology. New York: Marcel Dekker.
Freedman, H. I. & Waltman, P. (1978). Predator influence
on the growth of a population with three genotypes.
J. Math. Biol. 6, 367–374.
Freedman, H. I., So, J. W.-H. & Waltman, P. (1987).
Predator influence on the growth of a population with
three genotypes III. J. Math. Anal. Appl. 128, 287–304.
Hadeler, K. P. & Glas, D. (1983). Quasimonotone
systems and convergence to equilibrium in a population
genetic model. J. Math. Anal. Appl. 95, 297–303.
Hadeler, K. P. & Lieberman, V. (1975). Selection
models with fertility differences. J. Math. Biol. 2,
19–32.
Hutchinson, G. E. (1978). An Introduction to Population
Ecology. New Haven, CT: Yale University Press.
Josic, K. (1997). Local bifurcations in the symmetric model
of selection with fertility differences. J. theor. Biol. 189,
291–295.
Nagylaki, T. & Crow, J. F. (1974). Continuous selection
models. J. Theor. Popul. Biol. 5, 267–283.
PBS video (2001). Nature: Triumph of Life: The Eternal Arms
Race (video). Available at: http://www.shoppbs.com.
So, J. W.-H. (1986). Analysis of a continuous one-locus,
two-allele genetic model with genotypic fertility differ-
ences. Math. Biosci. 82, 221–233.
So, J. W.-H. (1990). Persistence and extinction in a
predator prey model consisting of nine prey genotypes.
J. Austr. Math. Soc. 31, 347–365.
So, J. W.-H. & Freedman, H. I. (1986). Persistence and
global stability in a predator prey model consisting of
three prey genotypes with fertility differences. Bull. Math.
Biol. 48, 469–484.
Wolfram, S. (1999). The Mathematica Book, 2nd Edn.
Cambridge: Cambridge University Press.