Review articles

Limits to natural selection

Nick Barton1* and Linda Partridge2

Summary evolved so as to facilitate further evolution, one would need to

We review the various factors that limit adaptation by show either that selection within populations favours individual
natural selection. Recent discussion of constraints on
selection and, conversely, of the factors that enhance alleles that modify the genetic system appropriately, or that
``evolvability'', have concentrated on the kinds of varia- selection between groups can overcome the intrinsically
tion that can be produced. Here, we emphasise that stronger force of selection between individuals. Such argu-
adaptation depends on how the various evolutionary ments require the greatest care.(3)
processes shape variation in populations. We survey the There have been recurrent arguments that straightforward
limits that population genetics places on adaptive evolu-
tion, and discuss the relationship between disparate natural selection is in some way unable to explain various
literatures. BioEssays 22:1075±1084, 2000. features of evolution. Following Darwin and Wallace's
ß 2000 John Wiley & Sons, Inc. proposal of adaptation by natural selection, it was argued that
there had not been enough time since the formation of the
Introduction earth for natural selection to act (Ref. 4 p206), and that sexual
Natural selection follows inevitably from inherited variation reproduction and blending inheritance would eliminate the
in the ability to reproduce. The systematic accumulation of variation on which selection depends.(5) These doubts were
chance variations is the only process that can lead to biological settled at the turn of the century by the discovery of radioactive
adaptations, and has produced impressive results: organisms decay, which established the antiquity of the earth, and by the
that can live in extreme environments and on diverse energy rediscovery of Mendelian heredity, which maintains variation
sources,(1) finely honed organs such as the ribosome, the eye from generation to generation. The early Mendelian geneti-
or the human brain, and developmental programs that can cists argued that natural selection could not generate novelty,
reliably construct such organs from the limited information since it could only sift amongst genotypes that already existed
encoded in the genome. Over much shorter times, artificial in the population (Ref. 4, p238). However, selection does not
selection has dramatically increased yields from domesticated occur in a single step: it is the accumulation of successive
plants and animals; applied to populations of RNA molecules favourable variants that produces genotypes that are, overall,
in vitro, it has produced novel catalytic activities (e.g. Ref.2). exceedingly improbable.(6) Wright(7) emphasised that selec-
In this essay, we consider what limits the power of selec- tion on a single population cannot establish variants that are
tion to produce such adaptations. We can ask how rapidly favourable in combination, but individually deleterious. He
populations respond to selection and what limits the rate of proposed that this limitation could be circumvented by a
response, whether any absolute limit is reached, and, if so, ``shifting balance'', involving selection among populations
what determines it. These are practical questions for the carrying different gene combinations. The discovery in mid-
design of artificial selection schemes, either in vivo or in vitro. century of extensive variation in protein and DNA sequence,
When applied to natural evolution, we may also ask whether both within and between species, stimulated Kimura(8) to
actual rates of adaptation are close to any limit, and whether propose that most of this variation has no effect on fitness, on
actual genetic systems have evolved so as to improve the the grounds that it could not all be maintained by selection.
response to selection. This last question is especially delicate, Finally, the striking stasis of species both across wide geo-
since species do not in general evolve so as to optimise their graphic regions, and over long geological timespans, has
own evolvability. In order to show that populations have suggested that special mechanisms such as founder events
are required for new species to form,(9) and that macroevolu-
tion is uncoupled from microevolution.(10)
1
Institute of Cell, Animal and Population Biology, University of Here, we try to make sense of these disparate arguments
Edinburgh, Scotland. by classifying and clarifying the various kinds of limits to
2
Department of Biology, University College London, UK. selection. We consider (1) constraints on variation that arise
Funding agencies: BBSRC and NERC, and the Darwin Trust of
from limitations on what can be built, (2) lack of a chain of fit
Edinburgh.
*Correspondence to: Dr. Nick Barton, Institute of Cell, Animal and
transitional forms leading to an optimal phenotype, (3) limits
Population Biology, University of Edinburgh, Kings' Buildings, inherent in the process of selection, set by the number of
Edinburgh EH9 3JT. selective deaths and by interference between linked variants,
(4) limits due to all the other evolutionary processes, which

BioEssays 22:1075±1084, ß 2000 John Wiley & Sons, Inc. BioEssays 22.12 1075
Review articles

necessarily tend to oppose adaptation by natural selection, divergence of 20 phenotypic standard deviations. There was
and (5) limits caused by conflicting natural selection. no sign of a decrease in response over time, suggesting that
Most discussion has tended to be of the first kind of limit: the response is due to variants of slight effect. Such sustained
throughout the history of evolutionary biology, there has been responses conflict with the frequent identification of major
a tension over whether the course of evolution is determined ``quantitative trait loci'' (QTL) responsible for divergence.
primarily by selection among slight variants, or by constraints However, statistical bias leads QTL effects to be over-
on the kinds of major variant that can be produced. This is estimated, and so there may well be no conflict between these
apparent in Jenkin's(5) concerns over blending inheritance, in different lines of evidence.(19) Artificial selection on growth rate
the bitter conflict between biometricians and Mendelians at the has been continued for over 20,000 generations on large
turn of the century, and on through to current debates over the asexual populations of bacteria, and has led to a plateau.(20)
evolutionary consequences of developmental and genetic This may reflect a genuine limit to selection. However, these
mechanisms. ``Developmental constraints'' have been much results cannot be compared directly with those from sexual
discussed elsewhere (e.g. Refs 3,11,12) and we do not dis- populations, which respond much more effectively to selection
cuss these further. In this article, we take a population-genetic (see below).
view, and survey a variety of limits on the evolution of whole It could be argued that artificial selection is successful
populations. Selection acts on variation that has accumulated because it is strong. If alleles that influence the selected trait
in populations and that is influenced by a variety of processes: have deleterious side-effects on fitness, then selection on the
it is not simply a matter of whether or not a particular favourable trait may need to be above some threshold for there to be any
variant can be produced. response (see below). However, selection in nature is also
intense,(21,22) and so the ready response of domesticated and
Limits to variation laboratory populations remains the strongest evidence for the
The most obvious limit to natural selection is that suitable power of natural selection.
variation may not be available. This may be because certain The limited set of feasible phenotypes could imply that
phenotypes cannot be built, being ruled out either by physical there are ``trade-offs'' between different components of
law or by the properties of biological materials. Some fitness; the constraint on the phenotype would then arise from
examples of such limitations are: (1) enzyme catalysis by the constraints on trait combinations, rather than on single traits.
rate of diffusion of substrates to the active site, (2) mutation One of the most frequently observed trade-offs is that between
rate by the thermodynamics of base pairing, (3) reaction times reproductive rate and survivalÐthe ``cost of reproduction''.(23)
by the speed of propagation of nerve impulses, and (4) the Imagine that each phenotype is described by components of
acuity of the eye by spherical aberration. These limits depend fitness, for example, survival, mating success and fecundity,
on assumptions about what is possible, and may be circ- measured across various ages and environments. If one
umvented by the ingenuity of natural selection. For example, component increases then, if trade-offs are present, the
(1) sparse molecules diffusing in three dimensions may be maximum possible contribution of other components must
concentrated by absorption onto two- or one-dimensional necessarily decrease (Fig. 1). This has spawned a research
surfaces, as on a moth's antenna,(13) (2) mutation rates may program that measures genetic correlations between fitness
be reduced by successive proof-reading mechanisms,(14) and components, and infers a trade-off, and hence a constraint, if
(3) spherical aberration can be abolished by a pupil that closes such correlations are negative. However, there are two
down to the diffraction limit. Nevertheless, we can suppose problems with this approach.(23 ±25) First, a negative genetic
that there is a set of possible phenotypes, while bearing in mind correlation does not necessarily imply an absolute constraint:
that this set occupies a space with more dimensions than we the response to selection may be reduced by the conflict
can imagine. There is necessarily no variation across the between different components, but not eliminated. No em-
boundary of this set, and hence no response to selection for pirical measurements can establish that there is absolutely no
phenotypes that cannot be realised (Ref. 3 Chapters 6,11). variation in the desired direction. Second, a trade-off only
Long-term artificial selection is a powerful way of exploring describes variation along the boundary of feasible pheno-
the space of possible phenotypes. Darwin's strongest evi- types. Mutations may in principle affect only single traits, and
dence for the power of natural selection was by analogy with involve no trade-off. In addition, standing variation may lie
the dramatic success of artificial selection,(Ref. 15, Chapter 1) inside the feasible set, and involve primarily positive corre-
and studies since Darwin's time have confirmed his view. What lations among fitness components (Fig. 1). The genetic
is remarkable is that almost all traits respond to selection, correlation does predict the immediate direction in which a
and that selection on large sexual populations causes a population will respond to selection, but cannot safely be
sustained response over many generations.(16,17) For exam- extrapolated over evolutionary timespans.(26) The response to
ple, Weber(18) selected on the angle between two wing veins in artificial selection, this time on combinations of traits, there-
Drosophila melanogaster for 20 generations, and achieved fore, gives a better indication of the presence of trade-offs.

1076 BioEssays 22.12

 Review articles

strongly on mutation rates. Instead, the response will be

limited by a conflict between directional selection on the trait of
interest, and whatever form of balancing selection maintains
variation. Despite considerable efforts, we do not at present
know if the ``classical'' or the ``balance'' view is more nearly
correct. Discriminating tests have proved hard to devise.
Are observed rates of evolution limited by rates of
environmental change, or by levels of genetic variability? If
the latter, is the immediate limitation set by mutation rate?
First, consider morphological change. Average rates of
change seen in the fossil record are extremely slow, and even
the fastest are easily explained by the levels of standing
variation seen in present-day populations: genetic variability
Figure 1. The relation between trade-offs and genetic
does not seem limiting.(22,27) In artificially selected popula-
correlations. The two axes represent two components of tions, genetic variation necessarily determines the initial rate
fitnessÐsay, juvenile and adult survival. Each female of response. However, the increase in additive genetic
produces one daughter at each of two ages; the chance of variance due to mutation is  10 ÿ 3 of the standing variation,
survival from birth to age 1 is J, and from age 1 to age 2 is A. which makes a significant contribution after  50 generations
The heavy curve delimits the set of possible viabilities, and
represents the trade-off between adult and juvenile survival
of selection.(19,28)
(J ‡ A4  1): if either component increases, the maximum Any one gene mutates much more rarely: m  10 ÿ 9 per
possible value of the other must decrease. The light curves base per generation, say.(29) Nevertheless, in a large popula-
show contours of fitness, spaced at intervals Dr ˆ 0.1. The tion of N > 109 diploid individuals, any favourable single base
optimal life history is where the furthest contour from the change will be established quickly: a mutation with selective
origin just touches the trade-off curve (base of arrow). The
arrow shows the reduction of survival below the optimal value
advantage s has a chance 2s of being established,(30) and so
caused by deleterious mutations at a rate U ˆ 0.1, on the the net rate of substitution is 2N 2s m ˆ 4Nms. This is reflected
assumption that mutations have a uniformly distributed effect in the rapid evolution of pesticide and antibiotic resistance, and
on both fitness components. The ellipse indicates the by the frequent multiple mutational origin of such adaptations.
distribution of fitnesses generated by such mutations. The Genetic variation and, more specifically, lack of mutations
step-like series of horizontal and vertical arrows shows that, in
principle, evolution can occur via a series of mutations with
should not and, it seems, does not limit at least straightforward
effects only on one or other component, and therefore selection response.
involving no trade-off. The evidence that, in the short term, the direction and rate
of evolution are not limited by mutation must be reconciled with
other evidence that, over the longer term, evolution may be
The evidence from artificial selection, then, suggests that limited by lack of the appropriate variability. On the broadest
most (but not all) traits are capable of responding to artificial scale, the contrast between the metabolic diversity of
selection, but that limits on trait combinations may be impor- prokaryotes and the morphological diversity of the eukaryotes,
tant in limiting the response to selection. presumably reflects fundamental constraints on the type of
variation that can arise in these groups (perhaps due to the
Rates of evolution mechanism of replication, or the structure of the cell wall).(1)
Almost all populations contain abundant genetic variation, at Dramatic adaptive radiations (for example, of metazoans at
all levels: DNA sequence, protein sequence, and phenotypic the beginning of the Cambrian or of Drosophila on the
traits. Thus, at least the initial response to selection may Hawaiian islands) involve the diversification of a single lineage
depend on standing variation, rather than new mutations. If, on into very different morphologies and behaviours. However,
one hand, this standing variation is mainly due to what were none of the suggestions for the limiting factors±from the wait
originally neutral or deleterious mutations (the ``classical'' view for a single appropriate variant, as was supposed by the
of genetic variation; in which standing genetic variation is mutationist theories from de Vries and Bateson, through to
maintained by mutation, in opposition to drift or selection), recent suggestions that multiple duplications of the Hox genes
then, in theory, the longer-term rate of evolution will still be triggered vertebrate evolution (for review see Ref. 31)±are
proportional to mutation rate. If, on the other hand, most satisfactory. There are two more plausible explanations. The
standing variation is due to balancing selection on different radiation may simply have awaited an appropriate environ-
alleles (either heterozygote advantage, or, more likely, some ment,(32,33) for example, rising oxygen levels in the PreCam-
kind of frequency-dependent selection that favours rare brian,(1) or an island lacking competitiors.(34) Alternatively,
variants), then long-term rates of evolution need not depend major shifts may await a combination or sequence of histori-

BioEssays 22.12 1077

Review articles

cal circumstances (including a sequence of mutations Orr(40,41) further extended Fisher's model to follow a sequence
and appropriate environments). In the next section, we of substitutions with additive effects, which take the population
consider obstacles to the assembly of multiple evolutionary closer and closer to the optimum. Orr showed that regardless
changes. of the distribution of mutational effects, the distribution of
effects of the substitutions that actually occur is exponential,
p
Absence of fit intermediates with mean inversely proportional to n. The initial substitution
``If it could be demonstrated that any complex organ existed, can be quite large, relative to the distance from the optimum,
which could not possibly have been formed by numerous, and tends to have deleterious side effects. Later substitutions
successive, slight modifications'' then Darwin's theory of tend to have smaller effects, which compensate for the
evolution by natural selection ``would absolutely break down'' deleterious side-effects of the earlier mutations. (Such
(Ref. 15, Chapter 6). The difficulty in establishing multiple compensation for the deleterious consequences of major
changes, all of which are apparently required to give a selec- mutations is seen in the evolution of antibiotic and pesticide
tive advantage, has long been seen as a key objection to resistance, Refs. 42,43).
adaptation by simple natural selection, and has stimulated The Fisher/Orr model suggests that adaptation of complex
elaborate alternative explanations. For an asexual organism, organs (i.e., of high dimension, n) is likely to be slow, and to
the problem can be seen as one of lack of variation: the depend on predominantly minor variations. (We consider
probability that two or more specific mutations occur in the below how the effective dimensionality of an organism affects
same individual is extremely low, and so individual favourable its evolvability). It is much oversimplified, in that it supposes
mutations must be established in sequence.(35,36) In a sexual that evolution is based on a sequence of allelic substitutions at
species, recombination breaks up adaptive gene combina- single genes, rather than on standing variation at multiple loci,
tions even if they do arise at low frequency, and again prevents and in that it assumes a smooth ``fitness landscape''. If there
establishment of all but immediately favourable variants. are many fitness peaks, then the population will be trapped at
Darwin's response to this difficulty remains convincing: the nearest. In a broad sense, this is clearly why different
some sequence of changes, each individually advantageous, lineages are constrained to particular ways of life: almost all of
can always be found. Only a single sequence of ancestral the multiple changes needed to allow a pig to fly would be
forms, each fit in at least some environment, is needed. The deleterious. However, there is a sequence of ancestral
existence of such a sequence is more plausible when one phenotypes that connects pigs with bats and birds, and a
realises the very large number of directions in which a combination of appropriate environments and appropriate
population might evolve. To take Darwin's original example, genetic variability has taken organisms along this path.
eyes capable of focusing an image have evolved indepen-
dently at least 15 times, through a series of gradations ranging The cost of natural selection
up from single photoreceptive cells.(37) Plausible assumptions Natural selection establishes fitter genes because those
about genetic variation in each component part lead to genes reproduced, whilst others did not. Haldane(44) argued
reasonably rapid evolution of the whole system.(38) The that selection is therefore costly, relative to the ideal alternative
requirement that natural selection can only produce changes in which all genes are instantaneously changed to the currently
that each give an immediate fitness advantage has been optimal allele. For example, suppose that conditions change
termed a ``selective constraint'', in contrast with ``develop- suddenly, so that most genes have viability 1 ÿ s, whilst a rare
mental constraints'', which limit what kind of organisms can be allele at frequency p0 is certain to survive. That allele will take
produced.(11)  (1/s)ln(1/p0) generations to reach appreciable frequency.
Selective constraints are most easily satisfied if an During each of those generations, a fraction ~s of the
organism evolves through a series of slight adjustments. population dies because it does not carry the optimal allele-,
Fisher(35) proposed a simple quantitative model of Darwin's and, overall, a total of ln(1/p0) selective deaths are required,
gradualist argument. He assumed that fitness depends on a per head of population. This ``substitution load'' is twice as
large number, n, of continuously varying traits, and increases great for a diploid population, if heterozygotes have inter-
towards some optimal combination of phenotype. Mutations of mediate viability. A single substitution requires at least ln(1/p0)
small effect have roughly equal chances of taking the selective deaths; unlike other arguments based on ``genetic
population towards the optimum or away from it, whereas load'', this is true even if fitnesses vary over time, or depend on
mutations of large effect are certainly deleterious. (Imagine a gene frequency,(45) and even if selection is ``soft'' (i.e., does
population lying at a point in n dimensional space, some not influence population size). There have been on average
distance r from the optimum; mutations of effect larger than 2r  1.3 amino-acid substitutions per generation since humans
must take it further from the optimum). Kimura (Ref. 39, diverged from chimpanzees.(46) Haldane's argument implies
Chapter 6) extended this argument to allow for the greater that no more than a small fraction of these could have been
chance that a strongly selected allele will be fixed, and established by selection.

1078 BioEssays 22.12

 Review articles

The cost of natural selection can be seen as limiting the rate balance is reached between selection and deleterious muta-
of accumulation of genetic information. An allele with no effect tion, which reduces mean fitness. This limits the net mutation
on fitness has a chance of eventually fixing in the population, rate, and hence the size of functional genome that can be
which is just equal to its initial frequency. Natural selection maintained by selection. With asexual reproduction, a simple
makes this unlikely event certain (at least, provided that the argument shows that deleterious mutation reduces mean
initial frequency is high enough for random drift to be fitness by a factor e ÿ U, where U is the genomic deleterious
negligible). The information added by some signal is defined mutation rate. The population can only sustain itself if those
as the log of the ratio between the probability of events, given fittest individuals, which carry no deleterious mutations,
the signal, and the probability without it: it is a measure of the produce the same number of intact offspring. Since the
concentration of probability around certain outcomes. Thus, chance that an offspring escapes mutation is e ÿ U (assuming a
the information added by natural selection in this example is Poisson distribution of mutant number), the fittest class must
log(1/p0), and can be no greater than the number of selective reproduce at a rate eU faster than the average. With sexual
deaths that occurred. Kimura(47) argued that only a limited reproduction, the fittest class can be regenerated by recombi-
amount of genetic information could have been accumulated nation between genomes carrying different mutations; the load
by natural selection in any one lineage since the Cambrian. can be greatly reduced if there is ``negative epistasis'', such
Gene duplication complicates matters, since it increases that fitness decreases faster than multiplicatively with number
genome size for the same cost as substitution of a single base of mutations (50,51) (Fig. 2). In asexual bacteria, mutation rates
pair. A duplication in itself, however, increases genetic infor- are low enough that the mutation load is negligible, even in
mation by a small amount: it is the subsequent divergence of absence of recombination.(14,29) In contrast, RNA viruses are
the duplicate gene that requires an additional cost. More close to the maximum mutation limit,(52) and in vitro replication
selective deaths, however, are presumably due to fluctuating experiments show that genome size is limited by the mutation
selection rather than to substitutions; the former does not load. In eukaryotes, genomic mutation rates are not well
increase the amount of functional genetic information. While it known, but may be high enough to require that negative
is not easy to define the amount of information in the genome, epistasis and recombination combine to alleviate the load.(53)
Kimura's argument does seem to put a strong upper bound on In particular, examination of sequence conservation between
the size of the functional genome. humans and primates implies that  38% of coding sites are
The most serious difficulty with Haldane's argument is that maintained by selection, and that the net mutation rate is high:
the effects of different loci are assumed to be independent. In U  4.2 per generation.(46) At present, it is not understood
contrast, if just those individuals with the largest number of whether or how genome size might evolve so as to approach
favourable alleles reproduce, then very many more alleles can the upper limit set by mutation load: presumably, there is a
be substituted for the same reduction in mean fitness.(48,49) balance between the evolution of novel gene function, and loss
This can be demonstrated very simply. Most of the substitution of existing functions under mutation pressure.
load accrues while favourable alleles are rare, and in different
individuals. If a fraction w < 1 of the population survives in each Random drift
generation, and if this fraction includes all those individuals Since different genes produce different numbers of offspring
that carry rare favourable alleles, then the frequency of each by chance, allele frequencies inevitably fluctuate. This process
such allele increases by 1/w. Hence, each takes log(1/p0)/ of random genetic drift interferes with selection in several
log(1/w) generations to become common, and the overall ways. Even in a large population, most favourable mutations
mean log fitness is  log(1/p0). On this argument, alleles at are lost: the chance of a single copy fixing in the population is
any number of loci can be substituted with the same fitness only twice its selective advantage.(30) This fixation probability
loss as at one. Of course, these theoretical arguments leave is further reduced by selection on linked loci, which in effect
open two questions: whether interactions between genes are amplifies the random influence of genetic drift.(54) In the
in fact such as to reduce the ``cost of natural selection'', and if extreme case of an asexual population, favourable mutations
so, why interactions have evolved to be of this form (Refs that occur in different lineages cannot be brought together in
3,39,Chapter 2). We return to this issue below. one individual. Thus, a slightly favourable allele can only be
fixed if it arises in a genetic background that is free of
Mutation deleterious mutations.(35) Such perfect backgrounds may be
Natural selection apart, all evolutionary processes are random extremely rare.
with respect to adaptation, and therefore tend to degrade it. Even in a sexual population, selection at one locus is
The other three evolutionary forces, mutation, genetic drift and impeded by any kind of selection at linked loci. Weakly
gene flow can all work against adaptation by natural selection. selected alleles are especially sensitive: an allele that
The most important of these is mutation, which is the inevitable increases fitness by less than some critical value, proportional
consequence of imperfect replication. In a large population, a to the rate of selective sweeps per map unit, has essentially no

BioEssays 22.12 1079

Review articles

sence of recombination is shown by the degeneration of Y

chromosomes.(58)
In a finite population (N diploid individuals, say), random
drift can overwhelm selection weaker than s  1/2N. More-
over, many factors (linkage to selected genes among them)
reduce the effective population size, Ne, to considerably less
than N. In particular, slightly deleterious mutations have a
probability  exp( ÿ Nes) of fixing, so that functions cannot be
maintained in the face of mutation when selection is weaker
than  1/2Ne. In many species, this is a stronger constraint
than the requirement that selection be stronger than the
mutation rate: in eukaryotes, mutation rates might be  10 ÿ 6
per gene, and  10 ÿ 9 per base, per generation, while Ne may
be smaller than 106. It may be that most amino acid changes
cause fitness differences large enough to counterbalance
mutation and drift, allowing protein function to be maintained
by selection even in less abundant species. However,
selection on synonymous changes (i.e., those that do not
affect protein sequence) may be much weaker. For example,
codon usage is strongly biased in E. coli, and yeast, and
somewhat less biased in Drosophila; there is no evidence of
bias in mammals.(59) A plausible explanation is that such bias
can only be maintained in sufficiently abundant species.
Similar evidence comes from an apparent relaxation of
constraints on transfer RNAs coded by mitochondrial gen-
omes, which have little recombination, and a smaller effective
Figure 2. The balance between deleterious mutation and population size.(60) It is unclear how serious this difficulty is for
selection. A: The heavy line shows the decrease in fitness as functions that impose no strong constraint on sequence, and
the number of deleterious alleles, k, increases. (W ˆ (1 ÿ s)k,
which therefore may involve very weak selection coefficients.
where s ˆ 0.02). With U ˆ 2 mutations per genome per
generation, the number of deleterious mutations is distributed Although loss of function at any one weakly selected site would
around a mean of k ˆ 100 (light curve), and the mean fitness have trivial effects, the cumulative effect could in principle
is (1 ÿ s)U/s ˆ 0.135 (dashed line). B: If the deleterious effect be serious.(61,62) Of course, in a very rare species, strongly
of each extra allele increases with the number already present selected functions can be lost under the combined pressure of
2
(i.e. negative epistasis; W ˆ (1 ÿ a)k(1 ÿ b)k , a ˆ 0.0005,
drift and mutation and, conversely, selection may be unable to
b ˆ 0.00005), the mean fitness can be substantially increased
(W ˆ 0.355; dashed horizontal line). This is because, with track a changing environment. In the medium term, this may be
loose linkage, the mean number of deleterious mutations is U/ an important cause of extinction.(53)
s , where s is the decrease in log(W) due to one additional Random drift could play a positive role in evolution. In
mutation. To a good approximation, s is minus the gradient of Sewall Wright's ``shifting balance'' theory of evolution,(7) sets
the heavy curve in B, evaluated at k. Therefore, the intercept
of genes that are favoured only in combination are established
of the tangent at is at U ‡ log(W) (dashed lines); with negative
epistasis, the mutation load (defined as ÿ log(W)) is in local populations by random drift, despite their individually
necessarily smaller than U. deleterious effects. Favourable combinations can then spread
through the whole species by various means, including
selection between local groups. Wright's theory has been
influential, and is feasible in principle. However, there is little
chance of fixation in a large population.(55) In Drosophila, evidence that it makes a significant contribution to adaptation,
regions of low recombination show less neutral diversity,(56) and it faces substantial theoretical obstacles.(63) It is more
and a weaker bias towards use of the optimal codon.(57) It is not parsimonious to suppose that selection acts primarily on the
known whether the selection responsible for this arises from effects of individual genes, averaged over the range of
the elimination of linked deleterious mutations, or ``selective environments and genetic backgrounds that they encounter.
sweeps'' of favourable mutations. Whichever the cause, this is
good evidence that selection is limited by linkage even in an Gene flow
outcrossing sexual species. More striking evidence of the Gene flow, the movement of gametes or individuals, can lead
failure of natural selection to maintain function in the ab- to loss of adaptation if selection pressures vary from place to

1080 BioEssays 22.12

 Review articles

place. Such spatially varying selection is common, as a result rate of reproduction and disrupt the genome of their host.(71)
of changes in the physical and biological environment. Allele There is therefore conflicting selection on transposition rate,
frequency can track changing conditions only over distances with the transposable element selected to increase its own
p
greater than  s/ s, where s is the distance moved by a gene reproductive rate and the genes of the host selected to
in each generation, and s is the selection coefficient.(64) suppress the element. Several cellular processes can be
Quantitative traits subject to stabilising selection of strength s understood as host adaptations of this kind.(68,72,73) Sexual
towards an optimum can track changes in the optimum over reproduction and outbreeding by the host exacerbate this kind
p
distances greater than ~s/ s Va .(65) In many organisms, of conflict, because in an asexual organism the transposable
dispersal ranges are short, so that genetic divergence can elements cannot disobey Mendel's laws, their own evolu-
occur over short distances if selection is sufficiently strong. A tionary future becomes identical with that of their host clone,
classic example is the adaptation of the grass Agrostis tenuis and conflicting selection ceases. To the extent that transposi-
to tolerate heavy metals concentrated in mine waste only a few tion is a cause of mutation, we might therefore expect to find
tens of metres across.(66) Burt(67) pointed out that the overall variation in mutation rate between organisms with different
degree of local adaptation can be measured by finding how kinds of breeding system. The data are at present too sparse
much mean fitness is reduced when individuals are trans- for any conclusion to be reached.
planted by one dispersal range. The limited data available from Evolutionary conflict can occur between sex chromosomes
plant populations suggest that selection must raise mean and autosomes, nuclear and cytoplasmic genes, females and
fitness by a few percent each generation in order to counter males and between individual and group interests.(68) Long-
immigration.(67) term persistence of populations or of whole species may be
best served by reproductive restraint of individuals, if exhaus-
Fluctuating selection tion of local food supplies can lead to extinction. However,
Selection pressures can very greatly with time, so that the genes increasing reproductive success of individuals are
phenotype favoured by natural selection changes. Even within under selection to increase within local groups. Which side in
a species held under apparently constant conditions, genetic such an evolutionary conflict will win? In general, selection acts
conflicts can lead to varying selection and continuing evolution more strongly at lower levels.(3) Selection on genes increasing
(Ref. 68 and see below). A single gene can only respond to reproductive success of individuals within groups is stronger
selection over time scales inversely proportional to the than selection for individual restraint as a result of extinction of
selection coefficient, or longer if genetic variation is limiting: whole groups. The reason is that, for a system of reproductive
this leads to a ``lag load'', equivalent to the ``cost of natural restraint to be evolutionarily stable, every allele that increases
selection'' discussed above.(69) Similarly, selection can individual reproductive success that is established by mutation
change a quantitative trait at a rate proportional to its additive or immigration must be balanced by one group extinction.
genetic variance, Va. In principle, arbitrarily rapid fluctuations Arguments that characteristics of organisms have evolved
can be tracked, provided that Va is large enough. However, ``for the good of the species'' should therefore be treated with
quantitative traits are usually subject to selection against scepticism. It has, for instance, recently been argued (but
extreme phenotypes (``stabilising selection''), and so high echoing back to a much earlier debate, Ref. 74), that ageing
genetic variance reduces fitness. There is an optimal has evolved to allow faster turnaround of generations and
value of genetic variance, set by a trade-off between hence evolution, because this is good for the persistence of
directional and stabilising selection. Even with this optimal species on the long term. Such an argument overlooks the
variance, a population will go extinct if selection fluctuates much stronger force of natural selection on individuals to
too much.(70) maintain survival and reproduction so as to maximise lifetime
reproduction.(3) Recent suggestions that mutases(75) and
Conflicts between levels of selection systems for co-ordinate expression of groups of genes(76)
Any entity that has the capacity for replication and variation have evolved because they increase the potential rate of
can be subject to natural selection. Because more than one evolution suffer from a similar difficulty (see Ref. 14).
kind of such an entity can be present simultaneously,
conflicting selection can occur. Such conflicts are most Evolvability
obvious in interactions between species, but also occur We have discussed a variety of limits to natural selection,
between genes that are carried within one organism, but are arising from (1) what organisms can be constructed, (2) the
inherited in different ways. fitness of those organisms and the lineages that connect them,
Selection can act at different levels. Transposable genetic (3) the nature of selection itself, (4) the interaction between
elements, present in the genomes of most living organisms, selection and other evolutionary processes, and (5) conflicting
often behave as genomic parasites. By replicative transposi- selection. We now reverse the question: instead of asking
tion, they disobey Mendel's laws, and both increase their own what limits evolution, we ask what features of life have

BioEssays 22.12 1081

Review articles

facilitated its extraordinary diversification, and why those different genetic backgrounds, reducing the constraints on
features are as they are. selectable variation; conversely, the striking redundancy
Recent interest in ``evolvability'' began in computer observed for most eukaryotic genes makes organisms robust
science, where genetic algorithms are designed to evolve to loss of function. This kind of gene interaction has important
efficiently towards some optimum.(77) The term has a technical consequences at the population level. Both the mutation load,
meaning in evolutionary biology, as a dimensionless measure and the ``cost of natural selection'' may be unsupportable
of quantitative genetic variance.(78) However, it is now used in unless there is ``synergistic epistasis'', such that the marginal
a broader sense by molecular and developmental biologists. effect of deleterious alleles increases with the number already
Kirschner and Gerhart(12) define ``evolvability'' as ``the capacity present, and conversely, the benefits of favourable alleles
to generate heritable, selectable phenotypic variation'', a decreases as they accumulate.(51)
definition that relates to both ``selective'' and ``developmental'' Note that it is unlikely that any genes are strictly redundant:
constraints (see Refs. 11,79). any gene that is not maintained by natural selection will be
Kirschner and Gerhart(12) discuss several features of life disabled by mutation and genetic drift. The pseudogenes
that facilitate the generation of variation that can be the basis of found in eukaryotic genomes are presumably gene duplica-
adaptive evolution. A key requirement is that a change to one tions that have suffered just this fate. On a larger scale, there is
part should not disrupt the whole system. This is achieved by extensive loss of gene function following duplication of the
the following: (1) versatile proteins, which contain several entire genome by polyploidy.(83) The frequently undetectable
functional sites connected by allosteric regulation, and which effects of gene knockouts only imply that such genes are
can bind to a variety of targets,(80) (2) the division of the embryo maintained by weak selection(84): even in microorganisms, no
into separate compartments, each with independent patterns feasible experiments can detect fitness differences smaller
of gene expression, (3) ``weak linkage'' of regulatory networks than  0.5%.(85,86) Moreover, absence of deleterious effects
(for example, where gene expression depends on subtle in the laboratory cannot necessarily be extrapolated to nature.
interactions among many transcription factors), and (4) If it is the case that genes interact in such a way that genetic
``exploratory mechanisms'' which adapt to circumstances load is reduced, and the opportunity for future adaptation
rather than following a predetermined program (for example, increased, we must ask why that should be. The simplest
in growth of neurons towards their targets, foraging behaviour explanation is that evolvability is a simple side effect of
of ants, or generation of diversity by the vertebrate immune selection for robustness of individuals to environmental or
system). It is this issue that is addressed by Fisher's genetic perturbations. At the phenotypic level, selection
geometrical argument: one can see many of these features reduces both environmental and genetic variation around the
as reducing the effective dimensionality of the organism, optimal phenotype (``genetic canalisation'' Ref. 87). Extensive
allowing evolution by a wider range of larger steps. overlap in the function of genes can be maintained by selection
Discussions of ``evolvability'' concentrate on limits on the if, for instance, the fidelity or reliability of a biological function is
production of ``selectable variation''.(79) Since adaptation thereby enhanced.(88,89) Genes that overlap in function are
involves the accumulation of variations in populations, the often the product of gene duplication and, in principle, genes
factors identified by Kirschner and Gerhart(12) influence the with identical functions can be maintained, by selection for a
long-term fate of variants, as well as their initial production and functional copy of one in individuals that are mutant for the
selective advantage. In addition, many other factors are other.(90) However, such selection is weak (of order mutation
important: effective population size, breeding system, recom- rate) and, in yeast, gene duplicates are no more redundant
bination rate, and so on. There is a substantial population than unique genes.(91)
genetic literature on the evolution of genetic systems, which is Selection can also act on evolvability itself, if modifier
essentially concerned with ``evolvability'', but which is not alleles become associated with the favourable alleles that they
usually associated with that term (e.g. Ref. 81). For example, produce. Evolution of mutation rates are the most straightfor-
Maynard Smith and Szathmary(82) demonstrate the impor- ward example, at least theoretically.(14) Rates of sex and
tance of discrete and unlimited heredity, such as that allowed recombination can also evolve in this way. For example, if a
by DNA sequence and by human language in allowing trait is under stabilising selection, then negative associations
continued biological or cultural evolution. Maynard Smith develop between the underlying genes ( ‡ ‡ ÿ ÿ , ÿ ‡ ÿ ‡
and Szathmary(82) also emphasise the importance of factors etc), which reduce the genetic variance. Sex and recombina-
that suppress conflicts between different replicators, which tion break up these associations, and hence increase the
has allowed cooperation of genes within cells, and of cells genetic variance. With a constant optimum for the trait,
within organisms. recombination tends to decrease, but, with a moving optimum,
An example that has received much recent attention in alleles that increase recombination become associated with
disparate literatures is the nature of gene interaction. A the favourable ‡ ‡ gene combinations which they gener-
modular organisation may allow novel alleles to function in ate.(92) Such mechanisms can generate short-term selection

1082 BioEssays 22.12

 Review articles

for ``evolvability'', and are supported by the observation of 3. Williams GC. Natural Selection: Domains, Levels and Challenges. Oxford:
increased recombination in artificially selected populations.(92) Oxford University Press. 1992.
4. Bowler PJ. Evolution: the History of an Idea. Berkeley, USA: Univ
However, recombination rates and genetic variance do not in California Press. 1989.
general evolve to any optimal value, and selection on modifiers 5. Jenkin F. ``The Origin of Species''. North Brit Rev 46;277±318.
of the genetic system tends to be weak, and dependent on 6. Dawkins R. The Blind Watchmaker. London: Longman. 1986.
7. Wright S. Evolution in Mendelian populations. Genetics 1931;16:97±159.
arbitrary factors such as the linkage relations of the modifying 8. Kimura M. Evolutionary rate at the molecular level. Nature 1968;217:
genes. While there is a good theoretical understanding of how 624±626.
selection may act on genetic systems, there is no consensus 9. Mayr E. Systematics and the Origin of Species. New York: Columbia
University Press. 1942.
over the extent to which they have been shaped by adaptive 10. Gould SJ, Eldredge N. Punctuated equilibria: the tempo and mode of
selection.(81) evolution reconsidered. Paleobiology 1977;3:115±151.
11. Maynard Smith J, et al. Developmental constraints and evolution. Quart
Rev Biol 1985;60;265±287.
Conclusions 12. Kirschner M, Gerhart J. Evolvability. Proc Natl Acad Sci USA;95:
Although much of the emphasis in the recent literature has 8420±8427.
been on limits to selection set by the nature of the phenotypic 13. Murray JD. Lectures on Nonlinear-Differential-Equation Models in
Biology. Oxford: Oxford Univ Press. 1977.
variation that can be generated (``developmental constraints''), 14. Sniegowski P, Gerrish PJ, Johnson T et al. The evolution of mutation rates:
population genetic constraints may be just as, or even more, separating causes from consequences. Bioessays 2000;12:1057±1066.
important. There is clear evidence that the rate and extent of 15. Darwin C. On the Origin of Species by Means of Natural Selection.
London: John Murray. 1859.
evolutionary change can be constrained by (1) lack of fit 16. Barton NH, Turelli M. Evolutionary quantitative genetics: how little do we
intermediaries leading to an optimal phenotype, (2) limits know? Ann Rev Genet 1989;23:337±370.
inherent in the process of selection, set by the number of 17. Falconer DS, Mackay TFC. Introduction to Quantitative Genetics. London:
Longman. 1995.
selective deaths and by interference between linked variants, 18. Weber K, Eisman R, Morey L, Patty A, Sparks J, Tausek M, Zeng ZB. An
(3) limits due to evolutionary forces such as drift, mutation and analysis of polygenes affecting wing shape on chromosome 3 in
gene flow, and (4) limits caused by conflicting natural Drosophila melanogaster. Genetics 1999;153:773±786.
19. Lynch M, Walsh JB. Genetics and Analysis of Quantitative Traits. Sun-
selection. derland, Massachussetts: Sinauer Press. 1998.
A particular kind of gene interaction is needed to allow 20. Lenski RE, Mongold JA, Sniegowski PD, Travisano M, Vasi F, Gerrish PJ,
species to counter the deleterious mutations that degrade Schmidt TM. Evolution of competitive fitness in experimental populations
of Escherischia coli: What makes one genotype a better competitor than
large genomes, and to allow simultaneous adaptation of many another? Antonie Van Leeuwenhoek Intl J Gen Mol Microbiol 1998;73:
genes; sex and recombination are needed to allow selection to 35±47.
act independently on separate functions; and strict genetic 21. Endler JA. Natural Selection in the Wild. Princeton: Princeton Univ Press.
1986.
rules are required to prevent conflicts between different 22. Hendry AP, Kinnison MT. The pace of modern life: measuring rates of
replicators.(82) In principle, such features can be selected contemporary microevolution. Evolution 1999;53:1637±1653.
because they facilitate evolution. However, such selection is 23. Partridge L, Barton NH. Optimality, mutation and the evolution of ageing.
Nature 1993;362:305±311.
weak, and the simpler alternative is that genetic systems are 24. Houle D. Genetic covariance of fitness correlates: what genetic correla-
primarily determined by immediate functional requirements. tions are made of and why it matters. Evolution 1991;45:630±648.
These kinds of question cannot be resolved by either 25. Charlesworth B. Optimization models, quantitative genetics and muta-
tion. Evolution 1990;44:520±538.
population genetics or molecular biology alone: the former 26. Turelli M. Phenotypic evolution, constant covariances, and the main-
merely calculates the consequences of specified fitness tenance of additive variance. Evolution 1988;42:1342±1348.
differences, while the latter specifies the components of life, 27. Charlesworth B. The cost of phenotypic evolution. Paleobiology 1984;10:
319±327.
without showing how or why they came to be as they are. A 28. Hill WG. Predictions of response to artificial selection from new muta-
major challenge for the future is to bring together our detailed tions. Genet Res 1982;40:255±278.
mechanistic knowledge of living organisms with an under- 29. Drake JW, Charlesworth B, Charlesworth D, Crow JF. Rates of spon-
taneous mutation. Genetics 1998;148:1667±1686.
standing of how they function as whole systems, and how they 30. Fisher RA. On the dominance ratio. Proc Roy Soc Edin 1922;42:321±431.
evolve in populations. 31. Holland PWH. The future of evolutionary developmental biology. Nature
1999;402 supp:C41±C44.
32. Simpson GG. The Major Features of Evolution. New York: Columbia Univ.
Acknowledgments
Press. 1953.
We thank Brian Charlesworth, Toby Johnson, and two 33. Schluter D. Ecological causes of adaptive radiation. Am Nat 1996;148:
anonymous referees for helpful discussions and criticism of S40±S64.
34. Grant PR. Evolution on Islands. Oxford: Oxford University Press. 1988.
the manuscript.
35. Fisher RA. The Genetical Theory of Natural Selection. Oxford: Oxford
University Press. 1930.
References 36. Muller HJ. Some genetic aspects of sex. Am Nat 1932;66:118±138.
1. Nealson KH, Conrad PG. Life: past, present and future. Phil Trans Roy 37. Futuyma DJ. Evolutionary Biology (3rd edn). Sunderland, Massachus-
Soc B 1999;354:1923±1939. setts: Sinauer Press. 1998.
2. Lehman NGF. Evolution in vitro: analysis of a lineage of ribozymes. Curr 38. Nilsson DE, Pelger S. A pessimistic estimate of the time required for an
Biol 1993;3:723±734. eye to evolve. Proc Roy Soc Lond B 1994;256:59±65.

BioEssays 22.12 1083

Review articles

39. Kimura M. The Neutral Theory of Molecular Evolution. Cambridge: 65. Slatkin M. Spatial patterns in the distribution of polygenic characters. J
Cambridge Univ Press. 1983. Theor Biol 1978;70:213±228.
40. Orr HA. The population genetics of adaptation: The distribution of factors 66. MacNair MR. Heavy metal tolerance in plants: a model evolutionary
fixed during adaptive evolution. Evolution 1998;52:935±949. system. Trends Ecol Evol 1987;2:354±359.
41. Orr HA. Adaptation and the cost of complexity. Evolution 2000;54: 67. Burt A. The evolution of fitness. Evolution 1995;49:1±8.
13±20. 68. Partridge L, Hurst LD. Sex and conflict. Science 1998;281:2003±2008.
42. Levin BR, Perrot V, Walker N. Compensatory mutations, antibiotic resis- 69. Maynard Smith J. What determines the rate of evolution? Amer Nat
tance and the population genetics of adaptive evolution in bacteria. 1976;110:331±338.
Genetics 2000;154:985±997. 70. Lande R, Shannon S. The role of genetic variation in adaptation and
43. McKenzie JA, Batterham P. The genetic, molecular and phenotypic population persistence in a changing environment. Evolution 1996;50:
consequences of selection for insecticide resistance. Trends Ecol Evol 434±437.
1994;9:166±170. 71. Charlesworth B, Langley CH. Population genetics of transposable
44. Haldane JBS. The cost of natural selection. J Genet 1998;55:511±524. elements. In Selander RK, Clark AG and Whittam TS, eds; Evolution at
45. Worden RP. A speed limit for evolution. J Theor Biol 1995;176:137±152. the Molecular Level. Sunderland, Massachussetts: Sinauer Press. 1991.
46. Eyre-Walker A, Keightley PD. High genomic deleterious mutation rates in p 150±176.
hominids. Nature 1999;397:344±347. 72. Selker EU. Premeiotic instability of repeated sequences in Neurospora
47. Kimura M. Natural selection as the process of accumulating genetic crassa. Ann Rev Genet 1992;24:579±613.
information in adaptive evolution. Genet Res 1961; 2:127±140. 73. Lin R, Avery L. Policing rogue genes. Nature 1999;402:128±129.
48. Sved JA, Reed TE, Bodmer WF. The number of balanced polymorphisms 74. Kirkwood TBL, Kremer T. Cytogerontology since 1881: a reappraisal of
that can be maintained by natural selection. Genetics 1967;55:469±481. August Weissmann and a review of recent progress. Hum Genet 1982;
49. Maynard Smith J. ``Haldane's dilemma'' and the rate of evolution. Nature 60:101±121.
1968;219:1114±1116. 75. Radman M. Enzymes of evolutionary change. Nature 1999;401:866±869.
50. Kimura M, Maruyama T. Mutational load with epistatic gene interactions 76. Niehrs C, Pollet N. Synexpression groups in eukaryotes. Nature 1999;
in fitness. Genetics 1966;54:1337±1351. 402:483±488.
51. Kondrashov AS. Deleterious mutations and the evolution of sexual repro- 77. Forrest S. Genetic algorithmsÐprinciples of natural selection applied to
duction. Nature 1988;336:435±441. computation. Science 1993;261:872±878.
52. Drake JW, Holland JJ. Mutation rates among RNA viruses. Proc Natl 78. Houle D. Comparing evolvability and variability of quantitative traits.
Acad Sci USA 1999;96:13910±13913. Genetics 1992;130:195±204.
53. Lynch M, Blanchard J, Houle D, Kibota T, Schultz S, Vassilieva L, Willis J. 79. West-Eberhard MJ. Evolution in the light of developmental and cell
Perspective: Spontaneous deleterious mutation. Evolution 1999;53: biology, and vice versa. Proc Natl Acad Sci USA 1998;95:8417±8419.
645±663. 80. Ptashne M, Gann A. Imposing specificity by localization: mechanism and
54. Hill WG, Robertson A. The effect of linkage on limits to artificial selection. evolvability. Curr Biol 1998;8:R812±R822.
Genet Res 1966;8:269±294. 81. Barton NH, Charlesworth B. Why sex and recombination? Science 1998;
55. Barton NH. The reduction in fixation probability caused by substitutions 281:1986±1990.
at linked loci. Genet Res 1994;64:199±208. 82. Maynard Smith J, Szathmary E. The major transitions in evolution.
56. Charlesworth B. Background selection and patterns of genetic diversity Oxford: WH Freeman. 1995.
in Drosophila melanogaster. Genet Res 1999;68:131±150. 83. Wendel JF. Genome evolution in polyploids. Plant Mol Biol 2000;42:
57. Comeron JM, Kreitman M, Aguade M. Natural selection on synonymous 225±249.
sites is correlated with gene length and recombination in Drosophila. 84. Brookfield JFY. Genetic redundancy. Adv Genet 1997;36:137±155.
Genetics 1999;151:239±249. 85. Smith V, Chou KN, Lashkari D, Botstein D, Borwn PO. Functional analysis
58. Charlesworth B. The evolution of chromosomal sex determination and of the genes of yeast chromosome V by genetic footprinting. Science
dosage compensation. Curr Biol 1996;6:149±162. 1996;274:2069±2074.
59. Akashi H, Eyre-Walker A. Translational selection and molecular evolution. 86. Dykhuizen DE. Experimental studies of natural selection in bacteria. Ann
Curr Opin Genes Dev 1998;8:688±693. Rev Ecol Syst 1990;21:373±398.
60. Lynch M. Mutation accumulation in transfer RNAs: molecular evidence 87. Scharloo W. Canalization: genetic and developmental aspects. Ann Rev
for Muller's Ratchet in mitochondrial genomes. Mol Biol Evol 1996;13: Ecol Syst 1991;22:65±94.
209±230. 88. Tautz D. Redundancies, development and the flow of information.
61. McVean GAT, Charlesworth B. The effects of Hill-Robertson interference Bioessays 1992;14:263±266.
between weakly selected sites on patterns of molecular evolution and 89. Thomas JH. Thinking about genetic redundancy. Trends Genet 1993;9;
variation. Genetics 2000;155:929±944. 395±399.
62. Kondrashov AS. Contamination of the genome by very slightly dele- 90. Nowak MA, Boerlijst MC, Cooke J, Maynard Smith J. Evolution of genetic
terious mutations: why have we not died 100 times over? J Theor Biol redundancy. Nature 1997;388;167±171.
1995;175:583±594. 91. Wagner A. Robustness against mutations in genetic networks of yeast.
63. Coyne JA, Barton NH, Turelli M. A critique of Wright's shifting balance Nat Gen 2000;24;355±361.
theory of evolution. Evolution 1999;51:643±671. 92. Charlesworth B. Directional selection and the evolution of sex and
64. Slatkin M. Gene flow and selection in a cline. Genetics 1973;75:733±756. recombination. Genet Res 1993;61;205±224.

1084 BioEssays 22.12