Popul Ecol (2002) 44:59–70 © The Society of Population Ecology and Springer-Verlag Tokyo 2002

SPECIAL FEATURE: ORIGINAL ARTICLE

Michael Doebeli

A model for the evolutionary dynamics of cross-feeding polymorphisms

in microorganisms

Received: February 19, 2002 / Accepted: May 8, 2002

Abstract Understanding mechanisms of evolutionary

Introduction
diversification is central to evolutionary biology. Microbes
constitute promising model systems for observing processes
of diversification directly in the laboratory. One of the main Understanding the origin and evolution of diversity is one
existing paradigms for microbial diversification is the evolu- of the central problems in population biology. In particular,
tion of cross-feeding polymorphisms, in which a strain spe- understanding the processes of speciation is of fundamental
cializing on a primary resource coexists with a cross-feeding importance. Although traditional discussions of speciation
strain that specializes on a waste product resulting from are based on geographic patterns of species distributions
consumption of the primary resource. Here I propose a (Mayr 1963; Turelli et al. 2001), recent theoretical devel-
theoretical model for the evolutionary dynamics through opments have studied the ecological mechanisms that can
which cross-feeding polymorphisms can gradually emerge drive adaptive divergence between different lineages
from a single ancestral strain. The model is based on the (Geritz et al. 1998; Dieckmann and Doebeli 1999; Doebeli
framework of adaptive dynamics, which has proved to be and Dieckmann 2000). In this context, the phenomenon of
very useful for studying adaptive processes of divergence evolutionary branching is of particular importance. During
under sympatric conditions. In particular, the phenomenon this evolutionary process, frequency-dependent selection
of evolutionary branching serves as a general paradigm for drives an evolving lineage to a point in phenotype space
diversification. I show that evolutionary branching naturally where selection turns disruptive, after which the lineage
occurs in evolutionary models of cross-feeding if (1) there is splits into two diverging phenotypic branches. Evolutionary
a trade-off between uptake efficiencies on the primary and branching is a paradigm for evolutionary diversification
secondary resources, and (2) this trade-off has positive cur- emerging from the theoretical framework of adaptive dy-
vature. The model also suggests that the evolution of cross- namics (Metz et al. 1996; Dieckmann and Law 1996; Geritz
feeding should be more likely in chemostat cultures than in et al. 1998). The main conceptual idea underlying this
serial batch cultures, which conforms with empirical obser- theory for evolutionary dynamics is that the phenotypic
vations. Overall, the model provides a theoretical metaphor distribution of a resident population, as well as its ecological
for the evolution of cross-feeding polymorphisms. dynamics, are important determinants of the environment
that a mutant phenotype encounters when it first appears in
Key words Crossfeeding · Polymorphism · Adaptive dy- the population. This intuitively appealing idea is captured
namics · Evolutionary branching · Tradeoff · Frequency in the notion of the invasion fitness, which describes how
dependence the long-term growth rate of a rare mutant depends on the
resident phenotypes. Roughly speaking, adaptive dynamics
is then derived as the gradient dynamics of the invasion
fitness function, which yields, in its simplest interpretation,
a theory for gradual evolutionary change in asexual popula-
tions. Although this asexual theory has already generated
a number of interesting and useful theoretical results about
general mechanisms of diversification (e.g. Doebeli and
Ruxton 1997; Kisdi 1999; Geritz et al. 1999; Maire et al.
M. Doebeli 2001), models for evolutionary branching obviously must be
Departments of Zoology and Mathematics, University of British extended with population genetics to be meaningful for
Columbia, 6270 University Blvd., Vancouver, BC, Canada V6T 1Z4
Tel.
1-604-822-3326; Fax
1-604-822-2416 diversification in sexual populations, i.e., for speciation.
e-mail: doebeli@zoology.ubc.ca This has indeed been achieved (Dieckmann and Doebeli
60

1999; Kisdi and Geritz 1999; Doebeli and Dieckmann 2000), lites can lead to coexistence of different specialist and gen-
leading to a general theory of adaptive speciation under eralist cross-feeders (Rosenzweig et al. 1994; see also
sympatric conditions, i.e., speciation in the presence of gene Dykhuizen and Davies 1980, for coexistence between gen-
flow and in the absence of geographic isolation. eralist and specialist microbial strains).
The past years have seen a surge of empirical data It is intuitively clear that cross-feeding polymorphisms
suggesting that contrary to traditional views, speciation in require the existence of trade-offs between uptake efficien-
the presence of gene flow is not an uncommon occurrence cies on different nutrients. Such trade-offs could be the
in nature (for pertaining recent reviews, see Turelli et al. consequence of physiological constraints on catabolic path-
2001; Via 2001). However, even though there are some ways for different nutrients. Indeed, there is some empirical
nice empirical examples of recent or ongoing adaptive evidence that such trade-offs exist and that they are medi-
divergence (Schluter 1994), possibly leading to speciation ated by differential enzyme activities (Rosenzweig et al.
(Schliewen et al. 2001; Jiggins et al. 2001), many natural 1994; Turner et al. 1996), which makes it likely that these
model systems tend to be inappropriate for experimental trade-offs have a genetic basis; i.e., that for any given strain,
tests of speciation processes because generation times of the its position on the trade-off curve is genetically determined.
organisms involved, and hence the time scale of evolution- Although some distinct genetic differences between dif-
ary change, are too long. There is an important exception to ferent strains constituting a glucose–acetate cross-feeding
this, however: experimental evolution in microorganisms polymorphism have been suggested (Rosenzweig et al.
offers an escape from the conundrum of long generation 1994), it is reasonable to assume that genetic change along
times. Indeed, microorganisms appear to constitute very the trade-off curve is gradual because the metabolic path-
promising model systems for studying the evolution of ways involved are rather complicated and can be affected in
diversity (Helling et al. 1987; Rosenzweig et al. 1994; Turner many different ways. This assumption is supported by the
et al. 1996; Xu et al. 1996; Rainey and Travisano 1998; finding that a number of different morphs have been found
Treves et al. 1998; Travisano and Rainey 2000; Rozen in E. coli cross-feeding polymorphisms (Helling et al. 1987).
and Lenski 2000; Kassen et al. 2000). For example, Rainey Therefore, adaptive dynamics appears to be an appropriate
and Travisano (1998) found rapid evolution of phenotypi- theoretical framework for studying the evolution of cross-
cally and genetically distinct lineages out of a single ances- feeding.
tral Escherichia coli lineage in spatially heterogeneous The purpose of this article is to present a model for the
habitats. gradual evolution of a cross-feeding polymorphism from
The ecological mechanisms underlying the evolution and a single ancestral lineage. The model is based on classical
maintenance of this diversity are not yet well understood. Michaelis–Menten dynamics for the uptake of a resource
However, it is clear that frequency-dependent selection that is continually supplied in a chemostat. It is assumed
plays an important role, because the various phenotypes that this primary resource is limiting, and that its consump-
tend to have positive growth rates when they are rare in an tion results in the production of a secondary nutrient as a
environment consisting of populations of the other pheno- metabolic waste product. A trade-off in uptake rates of the
types (Rainey and Travisano 1998). This is precisely the primary and secondary resource then results in frequency-
situation embodied by the invasion fitness function in adap- dependent selection, because the fate of a mutant depends
tive dynamics, and it is particularly tempting to try to apply on the environment generated by the resident: if the resi-
this theoretical framework to bacterial evolution because in dent is very efficient on the primary resource, it produces a
many such systems the assumption of asexual reproduction lot of secondary metabolites, so that it may pay a mutant to
appears to be largely satisfied. be efficient on these secondary resources at the expense of
One of the best understood cases of evolutionary diver- primary efficiency; on the other hand, if the resident is not
sification in bacteria is the evolution of a cross-feeding poly- very efficient on the primary resource, then metabolites are
morphism in E. coli (Helling et al. 1987; Rosenzweig et al. not very abundant, and it may simply pay to be more effi-
1994; Turner et al. 1996; Treves et al. 1998; Rozen and cient on the primary resource. I show, in the simple model
Lenski 2000). This process can be envisaged as follows: presented, that this frequency dependence can lead to evo-
bacteria limited by a single nutrient, glucose, first evolve lutionary branching, and hence to the gradual evolutionary
to become more efficient in glucose uptake. However, this split into two distinct strains, one a specialist on the primary
increased efficiency comes at the expense of efficient up- resources, and the other a specialist on the secondary re-
take of secondary “waste” products, such as acetate, that source that is produced by the primary resource specialist.
are secreted during glucose metabolism. Therefore, mu- By yielding a stable polymorphism the model also serves
tants that are efficient metabolite consumers, but less effi- as a mathematical description of an example of coexistence
cient on the primary resource, can invade the system, of two different bacterial species on a single primary re-
leading to a polymorphism with one type being efficient in source. It has long been known that microbial organisms
the uptake of the primary resource, and the other type can violate the ecological principle of competitive exclusion
being efficient at “cross-feeding” on the products resulting in a number of ways (Stewart and Levin 1973; Thingstad et
from the resource metabolism of the first type. This descrip- al. 1996; Xu et al. 1996; see also references in Rosenzweig et
tion is of course just a simplified caricature, and in reality al. 1994). Yomo et al. (1996) have introduced an interesting
the processes involved are rather intricate (Rosenzweig et mathematical model for coexistence between two different
al. 1994). In particular, cross-feeding on different metabo- microbial strains living in chemostats in which the bacteria
 61

are provided with a single external resource. In their model, sisting of a single phenotype. Here I take this basic model to
the external resource is metabolized by the bacteria into be a Michaelis–Menten type model for microbial growth in
a secondary resource, and both resources are essential for a chemostat culture. Let n(t) be the population density of
bacterial growth. Yomo et al. (1996) showed that strains the bacteria at time t, and let c(t) be the concentration of the
with different metabolic rate constants for synthesizing the primary resource, here glucose, in the chemostat at time t.
secondary resource from the primary resource can coexist. The flow rate at which chemicals and organisms enter and
The resulting polymorphism is similar to a cross-feeding leave the chemostat is denoted by d, so that in the absence
polymorphism, except that in cross-feeding scenarios, it is of bacteria consuming the glucose, the rate of change in the
usually only the primary resource that is limiting in the glucose concentration in the chemostat is
ancestral strain (i.e., the secondary metabolite is not essen-
tial for growth of the strain that is not cross-feeding). dc
 d ◊ c
d ◊ c0 (1)
In the model of Yomo et al. (1996), coexistence requires dt
a certain degree of differentiation in the metabolic rate
constants of the two competing strains. Even though it where c0 is the glucose concentration in the ambient envi-
is reasonable to assume that single mutations affecting ronment supplying the chemostat with glucose. If consump-
enzymatic activity could lead to a considerable change in tion of secondary metabolic waste products is neglected, so
metabolic rates, this model therefore does not describe an that growth of the bacterial population is only due to con-
evolutionary process that is driven by mutations of arbi- sumption of glucose, the rate of change of the population
trarily small effects. In addition, Yomo et al. (1996) only density is given by
considered competition between two given strains differing r ◊c
dn
in their metabolic rates, and they did not address the ques-  g ◊nd◊n (2)
tion of how metabolic rates would change evolutionarily dt kg
c
due to selection in each of the competing strains. In con-
where rg and kg are parameters describing how the per
trast, the model introduced here allows for arbitrarily small
capita bacterial growth rate changes with glucose concen-
mutational effects, and for repeated occurrence of new
tration: rg is the maximal per capita growth rate, whereas kg
mutations in the strains that are already present. In this is equal to that glucose concentration at which the per
way, the model shows how arbitrarily small phenotypic capita growth rate is rg/2. Because of bacterial consumption,
changes can lead to the gradual evolution of cross-feeding glucose concentration is decreased, so that the differential
polymorphisms.
equation for the glucose concentration becomes
Perhaps the most widely cited theoretical description of
microbial coexistence on a single resource is that by Stewart dc 1 rg ◊ c
and Levin (1973), who show how a trade-off between  ◊ n  d ◊ c
d ◊ c0 (3)
dt yg kg
c
growth rates at low and high concentration of a single re-
source can lead to coexistence of different types in serial where yg is the “yield” (that is, yg units of population incre-
batch cultures, i.e., in an environment in which the concen- ment are produced from one unit of the primary resource).
tration of the resource changes temporally. In chemostat Equations 2 and 3 constitute the basic Michaelis–Menten
cultures, however, where the influx of the resource is con- dynamics for bacterial growth on a single limiting resource
stant, such a trade-off cannot lead to coexistence (Stewart in a chemostat (Edelstein-Keshet 1989).
and Levin 1973). Interestingly, when the model introduced I now extend this model to cross-feeding by assuming
here is modified to describe serial batch cultures rather than that bacteria secrete a secondary nutrient, here acetate, into
chemostats, the evolution of a cross-feeding polymorphism the chemostat as a metabolic waste product of glucose con-
through a trade-off between uptake efficiencies on primary sumption. Thus, the rate of production of this secondary
and secondary resources becomes much less likely. This nutrient is proportional to the glucose consumption rate
result conforms with empirical data suggesting that cross- 1 rg ◊ c
feeding polymorphisms in E. coli are less likely to evolve ◊n
yg kg
c . At the same time, the secondary nutrient is
when the bacteria are propagated in batch cultures than
also consumed by the bacteria, with different parameters
when they are cultured in chemostats (Rozen and Lenski
describing uptake efficiency. If a(t) denotes the concentra-
2000; M. Travisano, personal communication).
tion of the secondary nutrient in the chemostat at time t, the
bacterial growth rate therefore becomes

dn r ◊c r ◊a
The model  g ◊n
a ◊n d◊n (4)
dt kg
c ka
a

Adaptive dynamics is based on calculations of the invasion where ra and ka are Michaelis–Menten constants describing
fitness function for mutant phenotypes in the ecological bacterial growth due to consumption of the secondary nutri-
environment generated by resident phenotypes. Therefore, ent. The assumption underlying this equation is that uptake
the first step in formulating a model for the evolutionary of the two resources is decoupled, so that waste products
dynamics of cross-feeding consists of setting up a model for are first secreted into the medium and are subsequently
the ecological dynamics of a monomorphic population con- consumed in a secondary metabolic pathway. It has been
62

argued that Eq. 4 is a an appropriate model for bacterial

growth on two different resources in a chemostat (Stewart
and Levin 1973; Gottschal and Thingstad 1982).
Uptake of the secondary resource results in a decrease of
its concentration, and taking the rate of production and the
rate of loss due to outflow into account, the rate of change
of a(t) becomes

da 1 ra ◊ a e rg ◊ c
 ◊nd◊a
◊n (5)
dt ya ka
a yg kg
c

Here ya is the yield for growth on acetate, and e is a constant

of proportionality describing how glucose metabolism
translates into production of the secondary nutrient.
Because glucose uptake is not directly affected by the pro-
duction of the secondary nutrient, the dynamic equation
for the glucose concentration remains the same, and Eqs. 3–
5 therefore describe bacterial growth on a primary resource Fig. 1. Possible trade-off curves ra  g(rg) between maximal growth
in a chemostat with the addition of bacterial growth due to rate on glucose, rg, and maximal growth rate on acetate, ra. Shown are
production of a secondary nutrient during consumption of the functions ra  u0  v0r 2g with u0  3 and v0  0.1, and ra 
u0 exp[v0rg] with u0  3 and v0  0.2. Note that for the latter g  0 and
the primary resource. g  0 for all rg  0, whereas for the former g  0 and g  0 for all
Before we formulate the evolutionary dynamics emerg- rg  0 (note that the biologically feasible range of rg is bounded by
ing from this ecological setup, we must specify which of the the condition g(rg) 0)
parameters in the ecological model should be viewed as
phenotypic properties and which parameters should be
regarded as nonevolving, externally fixed quantities. For
consumption of a single resource without production of a
Thus, the two traits whose evolution is to be studied are
secondary nutrient, i.e., for the situation described by Eqs. 2
rg and ra, and to study the evolution of cross-feeding I as-
and 3, it is known that coexistence of different bacterial
sume that there is a trade-off between these two traits that
strains is not possible. Instead, if we consider the compound
constrains all possible phenotypes to a one-dimensional set
parameter J  kgd/(rg  d), which is the equilibrium concen-
in the two-dimensional rg–ra plane. Specifically, I assume
tration of the primary resource in the absence of cross-
that there is a negative trade-off of the form
feeding, then in a mixture of strains the strain with the
lowest J value will outcompete all other strains (Stewart and
Levin 1973; Hanson and Hubbell 1980). In particular, in a
( )
ra  g rg (6)
mixture of strains with equal kg values, the strain with the such that g(rg)  0 (Fig. 1). Thus, the phenotypic value of ra
highest maximal growth rate rg outcompetes strains with is uniquely determined by the value of rg (and vice versa).
lower rg. Note that coexistence of two strains is possible in Under these assumptions, the task is to study the adaptive
competition in serial batch if both the rg values and the kg dynamics of the trait rg. To do this, we consider the growth
values are different and ranked in reverse order in the two rate of a rare mutant rg in an environment determined
strains (Stewart and Levin 1973). We discuss batch culture by the ecological dynamics of a resident population that is
models at the end of this section, but for now we concen- monomorphic for a resident phenotype rg. If the ecological
trate on chemostat cultures, and for simplicity we single out system given by Eqs. 3–5 is run to equilibrium for trait
the maximal growth rates rg and ra as evolvable traits and values rg and ra  g(rg), then the resulting equilibrium con-
consider all other quantities as fixed parameters. In particu- centrations of glucose and acetate will be functions c*(rg)
lar, a glucose specialist has a high rg and a low ra, and vice and a*(rg) of the resident phenotype. So long as the mutant
versa for the acetate specialist. We note that the parameter is rare, it will not affect these concentrations. Therefore, the
e, determining how much waste acetate is secreted during growth rate of the mutant population density nmut(t) in the
glucose metabolism, is likely also an evolving trait, so that e environment set by the resident will be
is higher, i.e., acetate production increased, in strains that
are more specialized on glucose (that is, these strains are dnmut r ¢ ◊ c* r ¢ ◊ a*
very efficient at glucose uptake, but they also produce a  g ◊ nmut
a ◊ nmut  d ◊ nmut (7)
dt kg
c* ka
a*
lot of waste during glucose metabolism; Rosenzweig et al.
1994). However, to keep things simple I do not consider this where rg and ra  g(rg) are the phenotypes of the mutant,
quantity as a phenotypic trait in this article, and instead it and where kg, ka, yg, ya, and d are, as mentioned, fixed
will be considered, together with the other quantities kg, parameters.
ka, yg, and ya, as an external parameter (note that these Therefore, the per capita growth rate of the mutant rg
parameters may nevertheless influence the evolutionary in the resident rg, which we denote by f(rg,rg), is given
dynamics). by
 63

(
f rg¢, rg  ) ( )
r¢ ◊ a* (r )  d
rg¢ ◊ c* rg a g
solutions are not feasible, due to the complexity of the
expressions c*(rg) and a*(rg); in particular, it is also not
k
c* (r )
g k
a* (r )g a g
(8) feasible to determine analytically whether an evolutionarily
r ¢ ◊ c* (r )
g g(r ¢) ◊ a* (r )
g a g singular point r*g is actually an attractor, i.e., whether Eq. 12

d
k
c* (r )
g k
a* (r )g a g
holds. These issues therefore have to be handled numeri-
cally by using pairwise invasion plots (Geritz et al. 1998), as
where we have used the constraining relation ra  g(rg). is explained shortly.
Based on the invasion fitness function f(rg ,rg), we can define First, let us note, however, that if an evolutionary singu-
the selection gradient D(rg) as lar attractor r*g exists, then the second derivative of the
invasion function is given by

( )
D rg 
(
∂ f rg¢, rg ) (9)
rg¢  rg
∂ rg¢
(
∂ 2 f rg¢, r *g ) ( ) ( )
g¢¢ rg* ◊ a* r *g
According to the canonical equation of adaptive dynamics 2 rg¢  rg*
 (15)
∂ rg¢ ka
a* ( r * )
(Dieckmann and Law 1996), the evolutionary dynamics of g

the trait rg is then given by

Therefore, inequality 13 is satisfied if and only if g(r*g)  0,
drg i.e., if and only if the trade-off curve between rg and ra has
dt
 m ◊ D rg ( ) (10)
positive curvature at the singular point, as would necessarily
be the case if g(rg)  0 independent of rg, e.g., for trade-offs
where m is a quantity that is determined by the mutational of the form
process and controls the speed of evolution. As we are only
concerned here with qualitative features of the evolutionary
dynamics of rg, we can set m  1 without loss of generality.
( )
g rg  u0 ◊ exp v0 ◊ rg [ ] (16)
Note that so long as D(rg)  0, rg will evolve to smaller where u0 and v0 are parameters determining the exact shape
values, while the contrary is true when D(rg)  0. Equilib- of the trade-off (cf. Fig. 1). For our numerical results, we
rium points for the adaptive dynamics, which are also called shall therefore consider trade-offs of the form of Eq. 16.
evolutionarily singular points, are points r*g in phenotype If analytical tools are lacking, a good alternative for
space satisfying determining the adaptive dynamics numerically consists of
using pairwise invasion plots (Geritz et al. 1998), in which
( )
D r *g  0 (11) one plots, for each resident–mutant pair (rg,rg), the sign of
the invasion fitness function f(rg,rg). Thus, in such plots the
A singular point is an evolutionary attractor if resident trait value is on the x-axis, the mutant trait value is
on the y-axis, and a particular point (rg,rg) is marked with a
dD dot if f(rg,rg)  0, i.e., if the mutant rg can invade the resident
drg
( )
r *g  0 (12)
rg, and the point (rg,rg) is left blank if f(rg,rg)
0, i.e., if
the mutant rg cannot invade the resident. Note that by
for in that case selection tends to increase rg if rg  r*g and to definition f(rg,rg)  0, so that the diagonal is always blank.
decrease rg if rg  r*g, i.e., the evolutionary dynamics will If there is a range of resident rg values for which a region
converge toward r*g. An evolutionarily singular attractor r*g is adjacent to and above the diagonal consists of marked dots,
called a branching point if r*g is a fitness minimum (Geritz et and a region adjacent to and below the diagonal is blank,
al. 1998), i.e., if this implies that for each such resident mutants with slightly
higher trait values can invade, whereas mutants with slightly
(
∂ 2 f rg¢, rg* ) (13)
lower trait values cannot invade. Thus, in such cases the
trait value evolves to higher levels through a series of muta-
2 rg¢  rg*
0
∂ rg¢ tional steps, i.e., through a sequence of trait substitutions
(Dieckmann and Law 1996), as indicated in Fig. 2A. A
Thus, to find evolutionary branching points for the trait similar but opposite remark is true if there is region of
rg we first have to solve Eq. 11 and then check whe resident rg values for which the region adjacent to and
ther inequalities 12 and 13 are satisfied. It follows from Eq. above the diagonal is blank, while the region adjacent to
9 that and below the diagonal is marked (Fig. 2B).
Evolutionarily singular points are those points on the
( )
D rg 
( )
g¢(r ) ◊ a* (r )
c * rg g g
(14) x-axis in a pairwise invasion plot at which the sign pattern
kg
c* (r ) k
a* (r )
g a g of f(rg,rg) across the diagonal changes (Geritz et al. 1998). In
particular, evolutionary attracting singular points r*g are
Because g(rg)  0 by assumption, while all the other terms those points on the x-axis for which mutants with higher
appearing in the last equation are 0, it is possible that the values than the resident are favored for residents with rg 
equation D(r*g)  0 has a solution (note that this would not r*g (marked region above the diagonal for rg  r*g), while
be possible if g  0). However, it turns out that analytical mutants with lower values than the resident are favored for
64
Fig. 2. Schematic illustration of various A B
pairwise invasion plots. Gray areas consist
of resident mutant pairs (rg,rg) with f(rg,rg)
 0, i.e., resident–mutant pairs for which
the mutant can invade the resident (see
text). A For all resident trait values, rg
mutants with rg  rg have a positive growth
rate whereas mutants with rg  rg have a
negative growth rate. Therefore, there is
directional selection for larger trait values.
The arrow indicates a series of trait
substitutions, each of which is character-
ized by one vertical and the subsequent
horizontal segment of the path; during
each trait substitution, a mutant success-
fully invades (vertical segment) and
becomes the new resident (horizontal
segment), with the whole series of substitu-
tions leading to an evolutionary increase in
rg. B Similar to A, but now with directional
selection for small trait values. C A C D
singular point is indicated by * on the x-
axis. To the right of *, selection favors
larger trait values (gray area above
diagonal), and to the left of * selection
favors smaller trait values (gray area below
diagonal). Therefore, * is an evolutionary
attractor. In addition, the vertical line
through * is outside the gray area; hence, *
is evolutionarily stable (no mutant can
invade). D The singular point * is an
evolutionary attractor as in C but now the
vertical line through * lies entirely in the
gray area, which means that all mutants
can invade; therefore, * is a fitness
minimum, i.e., an evolutionary branching
point

residents with rg  r*g (marked region below the diagonal for propagated in serial batch cultures. Following Stewart and
rg  r*g). Two such situations are shown in Fig. 2C,D. In both Levin (1973), the ecological model given by Eqs. 3–5 is
cases, the point indicated by * on the x-axis is a singular changed by assuming a series of dynamic cycles represent-
point attracting the evolutionary dynamics. In Fig. 2C, the ing the single batches. The dynamics for each single batch
vertical line passing through the singular point lies entirely are obtained from Eqs. 3–5 by setting the parameter d  0
in a blank region, which means that once the resident value (no outflow), and by assuming the following initial condi-
is r*g, no mutant can invade. Therefore, this attractor for the tions for the variables c, n, and a: c(0)  c0 is set to some
adaptive dynamics is evolutionarily stable. In contrast, in (high) initial value representing the initial concentration of
Fig. 2D the vertical line passing through the singular glucose in the batch; n(0) is set to some small fraction of the
attractor r*g lies entirely in a marked region, which means endstate of the previous batch, representing the serial trans-
that every mutant can invade. In other words, in this situa- fer of a small inoculate from the previous batch to the new
tion r*g is actually a fitness minimum, and hence an evolu- batch, i.e., n(0)  p · n(tend), where p denotes the fraction
tionary branching point. If, for a given set of parameters, used for the inoculation and n(tend) is the bacterial popula-
the pairwise invasion plot looks like the one shown in Fig. tion density at the end of the previous batch dynamics;
2D, then the trait rg will first evolve to r*g, after which the finally, a(0)  0, representing the fact that all glucose, and
population will become polymorphic and consist of two hence all secondary acetate, has been used up in the previ-
phenotypic branches that diverge evolutionarily from each ous batch. With these initial conditions the dynamics are
other (see Geritz et al. 1998 for a complete theoretical then again run to tend, and the process of inoculation into a
treatment of evolutionary branching in one-dimensional new batch is repeated. This iteration is done until n(tend)
trait spaces). reaches an equilibrium, i.e., until n(tend) does not change
Therefore, in a numerical exploration we must look for further from one batch to the next.
pairwise invasion plots that look qualitatively like the one The success of a mutant is then determined by assuming
shown in Fig. 2D. The existence of such plots will imply the that the mutant is initially very rare, so that the ecological
evolution of a cross-feeding polymorphism as envisaged in dynamics in a single batch of the quantities n(t), c(t), and
this article. a(t), where n(t) is the resident population density, are the
To conclude this section, I briefly mention how the same as just described, and the mutant dynamics in a single
model can be adapted to bacterial populations that are batch are simply determined by
 65

dnmut rg¢ ◊ c r¢ ◊ a
 ◊ nmut
a ◊ nmut (17)
dt kg
c ka
a

where rg and ra  g(rg) are the mutants trait values. (Note
that this differs from Eq. 7, in which the nutrient concentra-
tions are at their equilibrium values c* and a*; here, these
concentrations undergo the batch dynamics just described.)
Starting from a very small initial condition nmut(0), one
calculates nmut(tend) in the chemical environment determined
by the dynamics of the resident during a single batch. The
inoculate size of the mutant in the next batch is then
p · nmut(tend), and the mutant can invade, i.e., the invasion
fitness f(rg,rg)  0, if and only if

nmut (0)
1 (18)
p ◊ nmut (tend )

By plotting the sign of the quantity 1  nmut(0)/[pnmut(tend)],

one can then construct pairwise invasion plots as before,
from which one can deduce the adaptive dynamics of
the trait rg in a regime of serial batch cultures. As we will see
in the next section, these dynamics are generally different
from the evolutionary dynamics in models for chemostat
cultures.

Results

Recall that a central assumption in the adaptive dynamics

model developed in the previous section is the trade-off
ra  g(rg) between the maximal growth rates on the primary
resource glucose, rg, and the maximal growth rate on the
secondary resource acetate, ra. We assumed that this trade-
off is of the form ra  u0exp[v0rg], where u0 and v0 are
parameters. For a given u0, the parameter v0 determines
the “severity” of the trade-off, i.e., how fast the value of ra
declines with an increase in rg. Numerical simulations in-
dicate that, depending on the rate e at which metabolized
glucose is transformed into secondary acetate, the following
two qualitatively different types of pairwise invasion plots
represent typical evolutionary scenarios in the model so
long as the severity of the trade-off is not too high, i.e., for
low to moderate v0.
If e is high enough, the invasion fitness function f(rg,rg)
constructed in the previous section yields pairwise invasion
plots like the one shown in Fig. 3A. This plot shows the
existence of two evolutionarily singular points, indicated by
a and b on the x-axis. Here b is an evolutionary repellor,
because for any rg  b we have f(rg,rg)  0 for rg  rg, and for
any rg  b we have f(rg,rg)  0 for rg  rg. In other words,
selection moves rg away from b on both sides of b. Exactly
the contrary is true for a: on both sides of a, selection drives Fig. 3. Pairwise invasion plots for the evolution of cross-feeding. A
rg closer to the singular value (see magnification in Fig. 3B). Tradeoff ra  u0 exp[v0rg] with u0  3 and v0  0.2 (this function is
Therefore, a is an evolutionary attractor. Moreover, a is a shown in Fig. 1). The pairwise invasion plot shows an evolutionary
branching point (a) and a repellor (b). Parameter values: kg  1, yg 
branching point, because f(rg,a)  0 for all mutants rg that 1, ka  1, ya  1, d  1, c0  10, e  1.5. B Magnification of the small
are close to a. Thus, if the initial value of rg at which the square indicated in A. C Here e  0.5, i.e., there is less efficient
evolutionary process is started lies above b, then rg will conversion of glucose into acetate. The pairwise invasion plot shows
steadily increase over evolutionary time. However, if the that there is now uniform selection for larger rg values
66

A initial value of rg lies between 0 and b, then rg will first

converge to a, after which the population will become di-
morphic. Figure 4 shows an example of the diverging evolu-
tionary dynamics that occur after the population has
converged to the branching point a. In one of the two
branches rg evolves to ever higher values, and in the other
branch rg evolves to ever lower values. In the example
shown, both these processes would only stop because of
externally imposed constraints on the range of physiologi-
cally possible efficiencies of resource uptake. What is clear,
however, is that the evolving polymorphism consists of
one phenotype with a high rg and a correspondingly low ra,
representing a glucose specialist, and one phenotype with a
low rg and a correspondingly high ra, representing an acetate
specialist. Figure 4B shows the ecological dynamics of a
coexisting pair consisting of a glucose specialist and an ac-
etate crossfeeder. It is important to note that these coexist-
ing specialists gradually evolved from a single ancestor
B through the process of evolutionary branching.
Technically speaking, the branching point a and the
repellor b in Fig. 3A arise because the one-dimensional
subset in the resident–mutant plane defined by f(rg,rg)  0
has the shape carved out by the marked regions in Fig. 3A.
This set consists of the diagonal [as f(rg,rg)  0 in any case],
as well as of a second curve that arches across the diagonal
before it bends back toward the x-axis, enclosing the blank
area above the diagonal in the diagram. As the rate of
production of acetate during glucose metabolism, given by
e, is decreased, this blank area becomes smaller, until for
small enough e the curved part of the set defined by f(rg,rg)
 0 no longer intersects the diagonal. If this is the case, the
whole region above the diagonal is marked [i.e., f(rg,rg)  0
for any rg  rg], indicating that there is now uniform selec-
tion for increased rg (cf. Fig. 2A). Note that for a range of
intermediate resident rg values there are mutants with lower
rg values that can invade (as is indicated by the black region
Fig. 4. A Evolutionary dynamics of a cross-feeding polymorphism
after evolutionary branching. To simulate evolution in two coexisting below the diagonal) and subsequently coexist with the resi-
strains, the basic ecological model 3–5 was extended in a straight- dent (Fig. 3C). However, this can only occur if mutations
forward way to competition between two bacterial strains. A given toward lower rg values exceed a certain minimal size (which
ecological system comprising two strains was first run to equilibrium,
and then a mutant occurred with probability 0.1 in either strain. If a is given by the vertical distance between the current resi-
mutation occurred, its rg value was chosen from a normal distribution dent value and the black area below the diagonal). Such a
with mean the resident rg value and variance 5% of the mean. If the scenario may still be relevant for cross-feeding polymor-
mutant had a positive growth rate on the equilibrium glucose and phism in real systems, where mutations may have large
acetate concentrations determined by the two resident strains (which
necessarily have zero growth rates on these concentrations, because effects. Strictly speaking, however, this would not be evolu-
they are at ecological equilibrium), then the mutant was assumed to tionary branching as defined in adaptive dynamics, because
replace the resident strain in which it appeared (otherwise the resident in the case shown in Fig. 3C diversification would not occur
strain remained unchanged), and the new system was again run to
ecological equilibrium. This process was started with two initial resi-
for arbitrarily small mutations.
dent strains with rg values that were very close to and on opposite sides The number of parameters in the model is quite large,
of the branching point shown in Fig. 3B, representing a situation just and I have not fully explored the whole range of param-
after the initial polymorphism is established at the branching point. eters. Note that it is possible to rescale the ecological equa-
The figure illustrates diverging evolution in the two branches. Param-
eter values as for Fig. 3A. B Ecological dynamics of coexistence be- tions 3–5 so that of the ten original parameters only six
tween glucose specialist and acetate cross-feeder. The two coexisting remain, two of which correspond to the two maximal
strains correspond to the polymorphism present at time t  1000 in A. growth rates, and the remaining four correspond, respec-
One of the strains is a glucose specialist (high rg  5.78, low ra  0.94),
tively, to the ratios yg/ya, kg/ka, c0/kg, and to e. Because the
the other is an acetate cross-feeder (low rg  1.21 and high ra  2.36)
(note that the y-axis does not have units because the absolute magni- maximal growth rates are used in the trade-off, there are
tude of densities and concentration is arbitrary) therefore four external parameters that can potentially
affect the evolutionary dynamics resulting from a particular
choice for the trade-off function. Numerical exploration
indicate that none of the parameters yg/ya, kg/ka, and c0/kg
 67

Fig. 5. Example of a pairwise invasion plot for the tradeoff ra  u0 

v0 · r 2g with u0  3 and v0  0.1 shown in Fig. 1. In this case, an evolution-
arily singular attractor r*g exists at which the tradeoff curve has
negative curvature, i.e., with g(r*g)  0 (cf. Fig. 1). Therefore, the
evolutionary attractor is also evolutionarily stable (cf. Fig. 2C), and no
evolutionary branching occurs. Other parameter values as in Fig. 3A

qualitatively affects the finding that a decrease in e results in

the transition from Fig. 3A to 3C (so long as v0 is not too
high). What is affected by changes in the external param-
eters, as well as by the parameter u0 in the trade-off func-
tion, is the point of transition between the two regimes.
However, further and more detailed explorations may re-
veal regimes not reported here.
In particular, it is important to note that different
trade-off functions can lead to different evolutionary
dynamics. For example, as already noted (cf. Eq. 16), a
trade-off function with negative curvature cannot lead
to evolutionary branching, and instead often leads to the
existence of an evolutionarily stable singular point r*g. If
such a singular point is an attractor for the adaptive dynam-
ics, an example of which is shown in Fig. 5, then it represents Fig. 6. Pairwise invasion plots for the serial batch culture model. A
the final outcome of the evolutionary process, and branch- Same parameter values as in the chemostat model whose pairwise
ing, i.e., the evolution of a polymorphism, does not occur invasion plot is shown in Fig. 3A,B (note that the dilution rate d is not
in such a model. This result reiterates the fact that posi- a parameter in the serial batch culture model, and c0 now denotes the
initial glucose concentration in a single batch). The area above the
tive curvature of the trade-off function at an evolutionarily diagonal is dark everywhere, indicating that mutants with higher trait
singular point is a necessary condition for evolutionary values than the resident can always invade. There is also a dark area
branching. below the diagonal that touches the diagonal, indicating that for a
To end this section, it is interesting to note that even with range of resident values mutants with lower trait values can invade
provided the difference between resident and mutant exceeds a certain
a trade-off with positive curvature, evolutionary branching threshold. The point on the x-axis above which the two dark areas
seems to occur much less frequently in the evolutionary below and above the diagonal come together is a fitness minimum, but
dynamics of populations that are propagated in serial batch it is not a branching point, because it is not an evolutionary attractor (to
the right of this point, selection favors larger rg values). B Evolutionary
cultures. For example, if the uptake parameters from Fig. branching point in the serial batch culture model for higher conversion
3A are used in the serial batch culture model, the pairwise rate from glucose to acetate. The trade-off between maximal growth
invasion plot shown in Fig. 6A is obtained, in which the rates on glucose and acetate was of the form ra  u0 exp[v0rg] with
curved segment of the set f(rg,rg)  0 barely touches u0  5 and v0  0.1; other parameter values were kg  1, yg  1, ka  1,
ya  1, c0  10 (initial glucose concentration in a single batch), and
the diagonal. As mentioned, such cases can still give rise to e5
the evolution of cross-feeding polymorphisms, especially if
mutations have large effects. However, to obtain clear-cut
cases of evolutionary branching in the serial batch culture
68

Conclusions

The evolution of cross-feeding is one of the main empirical

paradigms for evolutionary diversification and subsequent
ecological coexistence of different microbial species on a
single limiting resource. Various types of cross-feeding
polymorphisms have been observed in the lab (Helling et al.
1987; Turner et al. 1996; Treves et al. 1998; Rozen and
Lenski 2000), but the main mechanism appears to be that
populations propagated from a single ancestral strain in-
crease uptake efficiency on the primary resource, often glu-
cose, which increases the amount of a secondary nutrient
that results as a waste product of the metabolism of the
primary resource. This in turn generates the conditions for
a specialist on this secondary resource to be sustained, i.e.,
the conditions for the appearance of a cross-feeder. Al-
Fig. 7. Ecological dynamics of a bacterial population during a single
batch of a serial batch culture (model as described in the text). Acetate though conditions for coexistence of different microbial
has a significant concentration only for a relatively short period of time. species have attracted interest from theoreticians (Stewart
Note that the bacterial population equilibrates because, for simplicity, and Levin 1973; Dykhuizen and Davies 1980; Gottschal and
no death is assumed to occur after all the glucose has been used up in
the batch. Parameter values as for Fig. 6A. (Note that the y-axis does Thingstad 1982; Thingstad et al. 1996; Yomo et al. 1996),
not have units because the absolute magnitude of densities and concen- no theory appears to be available describing the gradual
tration is arbitrary) evolution of a monomorphic population into a cross-
feeding polymorphism. The aim of the present article was to
take the first step in filling this theoretical gap by applying
models, the rate e at which acetate is produced from the the theoretical framework of adaptive dynamics to an ideal-
glucose metabolism must be much higher. An example of ized model for evolution in bacteria that are cultured on a
an evolutionary branching point in serial batch culture is single primary resource.
shown in Fig. 6B. Note that branching only occurs at a very The model is based on an ecological description of the
high resident rg value. Unfortunately, no analytical explana- dynamics of a bacterial population consuming a primary
tion is available at present why evolutionary branching resource that is continually supplied in a chemostat, as well
occurs for a smaller range of parameters, and in particular as a secondary resource that is produced during consump-
requires higher conversion rates e, when bacterial popula- tion of the primary nutrient. The basic assumption I made is
tions are propagated in serial batch cultures. Nevertheless, that there is a trade-off between uptake efficiencies on the
Fig. 7 provides at least some intuitive arguments for this two resources (see Fig. 1). Thus, a phenotype is character-
phenomenon. The figure shows the ecological dynamics of a ized by a single trait, its uptake efficiency on the primary
monomorphic population and of the concentrations of the resource (or, alternatively, its uptake efficiency on the sec-
primary and secondary nutrients during a single batch. The ondary resource, because the two efficiencies determine
concentration of the primary nutrient starts off at a high each other through the trade-off function). Because equilib-
level and decreases to 0 as it is consumed by the bacteria, rium concentrations of both the primary and the secondary
whose population eventually stops growing. The concentra- resource depend on the phenotype of a given (monomor-
tion of the secondary nutrient starts off at 0 in every single phic) resident population, the growth rate of a rare mutant
batch (as explained in the previous section), then increases also depends on which resident phenotype is present when
due to metabolism of the primary resource, and then de- the mutant appears. In other words, the growth rate of
creases again, due to consumption by the bacteria, as its mutants is frequency-dependent, which sets the stage for
production slows down because of a decrease in the con- evolutionary branching, a process by which a population
sumption of the primary resource. Figure 7 illustrates that first converges to a point in phenotype space at which selec-
the period of time during which the secondary nutrient is at tion turns disruptive due to frequency dependence, so that
high concentrations is relatively short; this is in contrast to the population subsequently splits into two diverging phe-
the conditions in the chemostat, where the concentration of notypic clusters representing the two resource specialists
the secondary nutrient attains a (positive) equilibrium coexisting in a cross-feeding polymorphism (see Fig. 4).
state. This indicates that conditions for cross-feeders, i.e., Besides showing how a cross-feeding polymorphism
specialists on the secondary resource, appear to be gener- can gradually evolve from a single ancestral strain through
ally harsher in serial batch cultures than in chemostat evolutionary branching caused by frequency-dependent
cultures. selection, the theoretical analysis presented here makes two
predictions. First, for cross-feeding to evolve in the models
analyzed here, the trade-off between efficiencies in uptake
of the primary and the secondary resource should have a
positive curvature (at least at the singular point). Although
 69

it is often difficult to measure the curvature of trade- Dieckmann U, Doebeli M (1999) On the origin of species by sympatric
speciation. Nature (Lond) 400:354–357
offs, this prediction appears to be testable, e.g., by isolating
Doebeli M, Dieckmann U (2000) Evolutionary branching and sympa-
strains at different times during an experiment that leads tric speciation caused by different types of ecological interactions.
to the evolution of a cross-feeding polymorphism. In ad- Am Nat 156:S77–S101
dition, if it were possible to manipulate trade-off curves Doebeli M, Ruxton GD (1997) Evolution of dispersal rates in
metapopulation models: branching and cyclic dynamics in pheno-
experimentally (e.g., by using different types of media type space. Evolution 51:1730–1741
for bacterial growth), one could then test whether evolu- Dykhuizen D, Davies M (1980) An experimental model: bacterial
tionary branching becomes less likely when the curvature specialization and generalists competing in chemostats. Ecology
of trade-off curves becomes negative, as predicted by the 61:1213–1227
Edelstein-Keshet L (1989) Mathematical models in biology. McGraw-
model. Hill, New York
The second prediction of the model is that cross-feeding Geritz SAH, Kisdi E (2000) Adaptive dynamics in diploid, sexual
polymorphisms should be more likely in chemostat cultures populations and the evolution of reproductive isolation. Proc R Soc
Lond B 267:1671–1678
than in serial batch culture, because in the models for the Geritz SAH, Kisdi E, Meszéna G, Metz JAJ (1998) Evolutionarily
latter case evolutionary branching appears to require very singular strategies and the adaptive growth and branching of the
high rates of production of the secondary nutrient during evolutionary tree. Evol Ecol 12:35–57
primary resource metabolism (see Fig. 6). This result con- Geritz SAH, van der Meijden E, Metz JAJ (1999) Evolutionary dy-
namics of seed size and seedling competitive ability. Theor Popul
forms with some empirical data suggesting that, indeed, Biol 55:324–343
cross-feeding polymorphisms are less likely to be found in Gottschal JC, Thingstad TF (1982) Mathematical description of com-
evolving bacterial populations that are propagated in serial petition between two and three bacterial species under dual sub-
strate limitation in the chemostat: a comparison with experimental
batch cultures (Rozen and Lenski 2000). The suggested
data. Biotechnol Bioeng 24:1403–1418
mechanism responsible for this is that in the single batches Hansen SR, Hubbell SP (1980) Single-nutrient microbial competition:
the concentration of the secondary nutrient is at appre- quantitative agreement between experimental and theoretical fore-
ciable levels only for a relatively short period of time, thus cast outcomes. Science 207:1491–1493
Helling RB, Varga CN, Adams J (1987). Evolution of Escherichia coli
making conditions harsher for specialists on this secondary during growth in a constant environment. Genetics 116:349–358
resource. Again, this is supported by the models presented Jiggins CD, Naisbit RE, Coe RL, Mallet J (2001) Reproductive isola-
here (see Fig. 7). tion caused by colour pattern mimicry. Nature (Lond) 411:302–305
Clearly, however, these models are but a first step to- Kassen R, Buckling A, Bell G, Rainey PB (2000) Diversity peaks at
intermediate productivity in a laboratory microcosm. Nature (Lond)
ward a more thorough understanding of the evolution of 406:508–512
cross-feeding polymorphisms. Future work should include Kisdi E (1999) Evolutionary branching under asymmetric competition.
the analysis of alternative trade-offs, such as between the J Theor Biol 197:149–162
Kisdi E, Geritz SAH (1999) Adaptive dynamics in allele space: evolu-
maximal growth rate rg attained when the resource is not tion of genetic polymorphism by small mutations in a heterogeneous
limiting and the “scavenging” growth rate kg attained when environment. Evolution 53:993–1008
the limiting resource is at very low concentrations (Hanson Maire N, Ackermann M, Doebeli M (2001) Evolutionary branching
and Hubbell 1980), or trade-offs between efficiency e and and the evolution of anisogamy. Selection 2:117–129
Mayr E (1963) Animal species and evolution. Harvard University
yield yg (Pfeiffer et al. 2001). In addition, it will be interest- Press, Cambridge
ing to consider multiple waste products from the primary Metz JAJ, Geritz SAH, Meszéna G, Jacobs FJA, van Heerwaarden JS
metabolism, which could result in the coexistence of mul- (1996) In van Strien SJ, Verduyn Lunel SM (eds) Stochastic and
spatial structures of dynamical systems. North Holland, Amsterdam,
tiple strains exhibiting various degrees of specialization
pp 183–231
on the different resources. For example, in addition to the Pfeiffer T, Schuster S, Bonhoeffer S (2001) Cooperation and competi-
glucose and acetate specialists, Rosenzweig et al. (1994) tion in the evolution of ATP-producing pathways. Science 293:1436–
found a third bacterial species that cross-feeds also on glyc- 1436
Rainey PB, Travisano M (1998) Adaptive radiation in a heterogeneous
erol, another waste product, and that exhibits properties of environment. Nature (Lond) 394:69–72
a generalist feeding on all three resources. In general, it is Rainey PB, Buckling A, Kassen R, Travisano M (2000) The emergence
important to further reconcile theoretical results about the and maintenance of diversity: insights from experimental bacterial
evolution of cross-feeding polymorphisms with empirical populations. Trends Ecol Evol 15:243–247
Rosenzweig RF, Sharp RR, Treves DS, Adams J (1994) Microbial
studies of this fundamental process of evolutionary evolution in a simple unstructured environment: genetic differentia-
diversification. tion in Escherichia coli. Genetics 137:903–917
Rozen DE, Lenski RE (2000) Long-term experimental evolution in
Acknowledgments I would like to thank Y. Toquenaga, M. Escherichia coli. VIII. Dynamics of a balanced polymorphism. Am
Kawata and M. Shimada for their great hospitality, and M. Travisano Nat 155:24–35
for discussions. Schliewen U, Rassmann K, Markmann M, Markert J, Kocher T, Tautz
D (2001) Genetic and ecological divergence of a monophyletic
cichlid species pair under fully sympatric conditions in Lake
Ejagham, Cameroon. Mol Ecol 10:1471–1488
Schluter D (1994) Experimental evidence that competition promotes
divergence in adaptive radiation. Science 266:798–801
References Stewart FM, Levin BR (1973) Partitioning of resources and the out-
come of interspecific competition: a model and some general con-
siderations. Am Nat 107:171–198
Dieckmann U, Law R (1996) The dynamical theory of coevolution: a Thingstad TF, Havskum H, Garde K, Riemann B (1996) On the strat-
derivation from stochastic ecological processes. J Math Biol 34:579– egy of “eating your competitor”: a mathematical analysis of algal
612 mixotrophy. Ecology 77:2108–2118
70
Travisano M, Rainey PB (2000) Studies of adaptive radiation using Via S (2001) Sympatric speciation in animals: the ugly duckling grows
model microbial systems. Am Nat 156:S35–S44 up. Trends Ecol Evol 16:381–390
Treves DS, Manning S, Adams J (1998) Repeated evolution of Xu W, Kashiwagi A, Yomo T, Urabe I (1996) Fate of a mutant
an acetate-cross-feeding polymorphism in long-term populations of emerging at the initial stage of evolution. Res Popul Ecol 38:231–
Escherichia coli. Mol Biol Evol 15:789–797 237
Turelli M, Barton NH, Coyne JA (2001) Theory and speciation. Trends Yomo T, Xu W, Urabe I (1996) Mathematical model allowing the
Ecol Evol 16:330–343 coexistence of closely related competitors at the initial stage of evo-
Turner PE, Souza V, Lenski RE (1996) Tests of ecological mechanisms lution. Res Popul Ecol 38:239–247
promoting the stable coexistence of two bacterial genotypes. Ecol-
ogy 77:2119–2129