When beneficial mutations are rare, they accumulate by a series of selective sweeps. But when they are common, many beneficial mutations will occur before any can fix, so there will be many different mutant lineages in the population concurrently. In an asexual population, these different mutant lineages interfere and not all can fix simultaneously. In addition, further beneficial mutations can accumulate in mutant lineages while these are still a minority of the population. In this article, we analyze the dynamics of such multiple mutations and the interplay between multiple mutations and interference between clones. These result in substantial variation in fitness accumulating within a single asexual population. The amount of variation is determined by a balance between selection, which destroys variation, and beneficial mutations, which create more. The behavior depends in a subtle way on the population parameters: the population size, the beneficial mutation rate, and the distribution of the fitness increments of the potential beneficial mutations. The mutation–selection balance leads to a continually evolving population with a steady-state fitness variation. This variation increases logarithmically with both population size and mutation rate and sets the rate at which the population accumulates beneficial mutations, which thus also grows only logarithmically with population size and mutation rate. These results imply that mutator phenotypes are less effective in larger asexual populations. They also have consequences for the advantages (or disadvantages) of sex via the Fisher–Muller effect; these are discussed briefly.
THE vast majority of mutations are neutral or deleterious. Extensive study of such mutations has explained the genetic diversity in many populations and has been useful for inferring population parameters and histories from data. Yet beneficial mutations, despite their rarity, are what cause long-term adaptation and can also dramatically alter the genetic diversity at linked sites. Unfortunately, our understanding of their dynamics remains poor by comparison.
When beneficial mutations are rare and selection is strong, positive selection results in a succession of selective sweeps. A mutation occurs, spreads through the population due to selection, and soon fixes. Some time later, another such event may occur. This situation is sometimes called the strong-selection weak-mutation regime. To make its character clear, we refer to it as the successional-mutations regime: between sweeps, there is a single “ruling” population. In this regime, the effect of positive selection on patterns of genetic variation is reasonably well understood. A selective sweep reduces the genetic variation in regions of the genome linked, over the timescale of the sweep, to the site at which a beneficial mutation occurs: other mutations in these regions hitchhike to fixation.
Successional-mutations behavior typically occurs in small- to moderate-sized populations in which beneficial mutations are sufficiently rare. However, a different regime occurs in larger populations, in which beneficial mutations occur frequently. When beneficial mutations are common enough that many mutant lineages can be simultaneously present in the population, selective sweeps will overlap and interfere with one another (i.e., different beneficial mutations will grow in the population concurrently). If, in addition, selection is strong enough that it is not dominated by random drift (except while mutants are very rare), we have a “strong-selection strong-mutation” regime. For clarity, we refer to this as the concurrent-mutations regime. The effects of concurrent mutations in asexual populations are the focus of this article. As we will see, the concurrent-beneficial-mutations regime is not an unusual special case: many viral, bacterial, and simple eukaryotic populations likely experience evolution via multiple concurrent mutations.
In populations that contain many different beneficial mutants, there will be substantial variation in fitness within the population. This variation will be acted on by selection. But in the absence of new mutations, the variation will soon disappear. Thus the traditional approach to evolution of quantitative traits—to assume that genetic variation always exists (as for traits not subject to selection)—fails badly. New mutations are crucially needed to maintain the variation on which further selection can act. Thus to understand adaptation when multiple mutations are involved, it is essential to analyze the interplay between selection and new beneficial mutations, especially how the latter maintains the variation acted on by the former. Understanding this beneficial mutation–selection balance and the resulting dynamics is the primary goal of this article.
Both the successional- and the concurrent-mutations regimes require that selection dominates drift except while mutants are very rare. A qualitatively different regime occurs with weakly beneficial mutations: these do not sweep in the traditional sense because drift dominates their dynamics. This weakly beneficial regime most readily occurs in small populations, where selective forces cannot overcome drift, or when considering mutations of very small effect, such as those that affect synonymous codon usage (Li 1987; Comeron et al. 1999; Przeworski et al. 1999; McVean and Charlesworth 2000). In this article we are interested in beneficial mutations in moderate to large populations, so we focus exclusively on the strong-selection regimes for which drift is important for beneficial mutant lineages only while they are a tiny minority of the population.
The essential difference between the successional-mutations and concurrent-mutations regimes is presented in Figure 1, which depicts beneficial mutations in an asexual population. In a small enough population, or one whose beneficial mutation rate (Ub) is low, beneficial mutations occur rarely enough that they are well separated in time and one can sweep before another arises (Figure 1a). This is the successional-mutations regime, in which the beneficial mutations all behave independently. However, in a larger population or at higher Ub, multiple mutant populations exist concurrently and they are no longer independent (Figure 1b). Mutations that occur in different lineages cannot both fix in the absence of recombination: at least one of them must be “wasted.”
In the concurrent-mutations regime, two important effects occur. The first is when a moderately beneficial mutation occurs and begins to sweep, only to be outcompeted by a later, more strongly beneficial mutation that occurs in a wild-type individual. The first mutation is then wasted, as it is eliminated along with the then-majority type by the sweep of the stronger mutation. This effect is referred to as clonal interference; it is illustrated in Figure 1c. Note that despite earlier broader definitions we use the term “clonal interference” to refer to only this first effect, consistent with the focus of recent work on the subject (Gerrish and Lenski 1998). The second effect is when multiple mutations occur in the same lineage before the first beneficial mutation fixes. For example, a second beneficial mutation can occur in an individual that already has one beneficial mutation. The double mutant can then benefit from the combined effect of the two mutations and outcompete the single mutant as well as some other stronger single mutants that arise in the majority population. This process is illustrated in Figure 1d.
The dynamics of evolution in the concurrent-mutations regime are important to understand. At the very least, this is essential for forming sensible null expectations about experimental, observational, and genomic data from large populations. Knowing how the effects of beneficial mutations depend on mutation rate and population size is crucial for making meaningful comparisons between different populations. Most important, in our view, is developing an intuition for how large populations evolve. The simple picture of successive selective sweeps in the successional-mutations regime is a valuable guide to thinking about positive selection. Yet we have little intuitive guidance when the successional-mutations approximation does not apply. This is a serious shortcoming in our understanding of the evolution of a wide array of populations, including viruses and most unicellular organisms.
Although it is not as well understood as the successional-mutations regime, the concurrent-mutations regime has been the subject of substantial interest since the 1930s. Fisher (1930) and Muller (1932) first noted the potential importance of interference between beneficial mutations (Muller drew diagrams very similar to our Figure 1). They proposed what has come to be known as the Fisher–Muller hypothesis for the advantage of sex: sexual populations can recombine beneficial mutations in competing lineages into the same individual. This prevents mutational events from being wasted, as they often are in asexual populations.
Much subsequent work on positive selection in the concurrent-mutations regime has focused on the implications for the evolution of sex. Crow and Kimura (1965), Bodmer (1970), and Maynard Smith (1971) attempted to quantify the Fisher–Muller effect in the late 1960s and the early 1970s. However, their analysis was incomplete—it did not fully account for stochastic behavior, ignored triple and higher mutations, and did not correctly account for the effects of sex. Contemporaneously, Hill and Robertson (1966) looked at this problem from the perspective of the linkage disequilibrium generated by multiple linked beneficial mutations segregating simultaneously. This has become known as the Hill–Robertson effect. It is essentially equivalent to the Fisher–Muller effect (see Felsenstein 1974 for a detailed discussion). In recent years, Barton (1995), Otto and Barton (1997, 2001), and Barton and Otto (2005) have analyzed the Fisher–Muller effect from the Hill–Robertson perspective. Their work focuses on the buildup of linkage disequilibrium due to mutations and selection and the average effect of recombination on the variance in fitness and the destruction of disequilibrium. This provides useful insight into the effects of sex, but does not explain the full evolutionary dynamics or population genetic structure created by this type of positive selection.
In this article, we step back from the long tradition of studying the implications of concurrent mutations for the evolution of sex and focus instead on the basic dynamics shown schematically in Figure 1b. We show how an asexual population in the concurrent-mutations regime accumulates many beneficial mutations, what the fitness distribution looks like, how it develops, and how quickly selected substitutions occur via collective sweeps. We develop a framework for thinking more generally about positive selection and its effects that is applicable to large populations of asexuals or any other case where linkage between mutations is important.
We do not analyze the questions about sex or patterns of diversity in this article. However, these questions should be informed by our results; some can be studied within the framework we present in this article. For example, when recombination is rare, the average effects of sex may be irrelevant—instead all that matters is whether or not it creates rare individuals that are much more fit than the majority of the population. To study this, we must first understand the full distribution of genetic diversity within the population. Similarly, before analyzing the patterns of genetic variation exhibited by populations in which multiple linked beneficial mutations have occurred—or are occurring—one must understand the rate of beneficial substitutions and typical interference patterns between these within the linked regions.
To understand the concurrent-mutations dynamics in detail, it is essential to start with a specific model that focuses on some subset of the important effects. Features can then be added after enough understanding has been gleaned to enable predictions of which effects are model specific and which are more general. Positive selection can involve various complications, including epistasis (interactions between effects of mutations), conditionally beneficial mutations, frequency-dependent benefits, and changing environments, among others. Many different scenarios are possible. At present we have little understanding of which, if any, of these situations are biologically “typical” and which ones are unusual. In this article, we do not attempt to catalog all possible complications; this is an impossibly broad subject. Instead we look at the simplest possible situation involving positive selection of concurrent mutations. We suppose that a variety of beneficial mutations are available to a population and ask how the population acquires them. We assume these mutations interact in a simple multiplicative way (additive for the growth rates) with no epistasis, frequency dependence, or changing environment of any kind. In short, we ask how the population climbs a single smoothly sloped “hill” in fitness space.
This simple scenario is probably common. Populations often find themselves in an environment where they can accumulate quite a few different beneficial mutations that each roughly independently help them adapt. Even when this simple hill-climbing scenario does not apply, it is an important null model. Some more complex forms of positive selection may also prove tractable within the framework we describe, while others will not; these leave open many avenues for future work.
Various other authors have studied the dynamics of multiple concurrent beneficial mutations under the simple assumptions outlined above. Gerrish and Lenski (1998) analyzed clonal interference between mutations of different strengths; this has since been extended by various authors (Orr 2000; Gerrish 2001; Johnson and Barton 2002; Kim and Stephan 2003; Campos and De Oliveira 2004; Wilke 2004). This work focuses on the interference between mutations of different strengths that occur in the same lineage, while neglecting the competition between mutations that arise in different lineages—in particular multiple mutants. Yet we show below that if population parameters are such that clonal interference is important, the effects of multiple mutants are usually at least of comparable importance. Thus there is some inconsistency in focusing on clonal interference alone. Our analysis in this article starts instead with the other concurrent-mutation effect, multiple mutants, initially in a model in which clonal interference is absent. In any real situation, the two effects will both occur. We thus discuss the interplay between clonal interference and multiple mutations in a later section. Kim and Orr (2005) have also recently analyzed a model that combines some aspects of clonal interference and multiple mutations.
To focus on the effects of multiple mutants without clonal interference, two additional simplifying approximations are useful. For most of this article, we study a model in which each beneficial mutation has the same effect, s, on fitness (i.e., each step uphill is of the same size). Furthermore, to focus on the effects of positive selection, we neglect deleterious mutations in the primary analysis. Even though neither assumption will typically be true, these turn out to be reasonable approximations in many circumstances. Situations in which they are not appropriate are more complicated scenarios for positive selection, some of which, especially the effects of a distribution of fitness increments, we discuss briefly.
Remarkably, even the simplest possible model with many equal-strength beneficial mutations available is only partially understood. Kessler et al. (1997) and Ridgway et al. (1998) analyzed a similar simple model, but their initial work did not handle random drift correctly. More recently, they have developed a sophisticated although somewhat unwieldy moment-based approach (D. Kessler and H. Levine, unpublished results) from which it is unfortunately hard to understand the essential aspects of the dynamics. Rouzine et al. (2003) also studied a model similar in its essential aspects to our simplest model (although also including deleterious mutations of the same magnitude). They were concerned with viral evolution, and their results are primarily valid for very large mutation rates appropriate for many viruses; we focus instead on regimes primarily applicable to single-celled organisms (and some viruses). Nevertheless, if worked out more fully from Rouzine et al.'s analysis, several results can be obtained that are closely related to ours. But our analysis involves a less mathematically formal approach—we believe it is both clearer and a better basis for further development (some of which is included herein). We discuss the relationship between our analysis and that of Rouzine et al. (2003) in more detail below.
The outline of this article is as follows. We begin by describing in the next section a heuristic approach to the dynamics. This analysis gets the behavior roughly correct and illustrates the ideas underlying our approach. We then describe the simplest model more precisely and analyze it in the following section. We next discuss transient behavior before the population has reached its steady-state fitness distribution and address the effects of deleterious mutations. In the next section, we make comparisons between our analytic results and simulations. We then relax our assumption that all mutations have the same effect and discuss the relationship between our theory and clonal interference analysis. Finally, we summarize our results and discuss future directions.
HEURISTIC ANALYSIS AND INTUITION
In the simplest situation with multiple concurrent beneficial mutations, there are three important parameters: the population size, N, the beneficial mutation rate per individual per generation, Ub, and the fitness increase provided by each mutation, s. We refer to the basic exponential growth rate, r, of a population as its fitness (rather than its growth factor per generation w = er ≈ 1 + r). That is, we use “fitness” to mean what is sometimes called log fitness. Thus in the absence of epistasis, which we generally assume, two mutations of magnitude s1 and s2 increase fitness by s1 + s2. We call the rate of increase, d〈r〉/dt, of the average fitness of a population the speed of evolution and denote it v.
To focus on the effects of multiple mutants in a situation in which clonal interference does not occur, we initially restrict consideration to the approximation that all beneficial mutations have the same effect. A k-tuple mutant thus has fitness ks greater than the original wild type. The speed of evolution is then simply .
We begin by reviewing the successional-mutations regime where beneficial mutations are sufficiently separated in time for them to sweep independently, as in Figure 1a. Although this is exactly solvable and well known, it is instructive to consider it from a heuristic perspective. We then turn to a heuristic analysis of the more complex concurrent-mutations dynamics illustrated in Figure 1d.
Successional-mutations regime and the establishment of mutants:
Small asexual populations evolve by accumulating beneficial mutations sequentially. Beneficial mutations occur in the population at a total rate NUb. The probability that a particular mutant will survive random drift is proportional to its selective advantage s (provided ). The constant of proportionality depends on the specific model for the stochastic dynamics; for our model it is 1 and we discuss in the simplest model section below the minor modifications of our results that are needed for other stochastic dynamics. We call the process by which the lineage of a beneficial mutant that survives drift becomes large enough for the population of its descendants to grow deterministically the establishment of the mutant clone. As we show below in the section on the fate of a single mutant, a mutant population becomes established when its size reaches of order 1/s individuals. Roughly speaking, this is because a mutant lineage of size n takes n generations to change by of order n individuals due to random drift. Since selection adds on average ns individuals to the lineage per generation, in this time selection has an average effect of adding n2s individuals. So selection dominates drift provided n2s > n or . Thus the mutant lineage must reach a size before it becomes “safe” from extinction and begins to grow mostly deterministically.
We show in the section on the fate of a single mutant that if a mutant is destined to become established, it will reach this size 1/s very quickly. Thus new beneficial mutations are established at a rate roughly NUbs per generation (other mutant populations die out due to random drift), so a new mutation will become established about once every generations. Once established on reaching size of order 1/s, the mutant lineage grows roughly exponentially at rate s and hence takes of order generations to fix (we loosely call “fixed” a mutant lineage that has grown to represent a large fraction of the population; the conventional definition corresponds to fully fixed, which takes about twice as long).
When the population size or mutation rate is small enough, fixation will happen more quickly than establishment. This occurs when(1)which corresponds to . When this condition holds, we are in the successional-mutations regime, in which the establishment rate is limiting: a mutation A that arises and fixes will do so long before the next mutation destined to survive drift, B, is established. Thus mutation B occurs in a population that has already fixed A, yielding AB, and B fixes well before mutation C is established. Beneficial mutations continue to accumulate in this simple way. New mutations arise and fix at average rate NUbs, each one increasing the fitness by s. Thus fitness increases at a speed(2)linear in the product NUb. This linear mutation-limited behavior characterizes the successional-mutations regime of successional selective sweeps.
In larger populations, the behavior is more complex, as illustrated by Figure 1b. In this case, the establishment times of new mutants are shorter than their fixation times, corresponding to(3)Thus new beneficial mutations arise and become established before earlier ones can sweep, causing them to interfere with one another.
As noted in the Introduction, two types of interference are important. First, competition occurs when two mutations that have different strengths occur independently in individuals with similar initial fitness (clonal interference). We focus in the bulk of this article on the other type of interference: a mutation that arises in a fitter background (e.g., one with an earlier beneficial mutation) will outcompete another mutation of similar effect that occurs in a less fit background. In the constant-s model clonal interference is explicitly absent, and we thus initially focus exclusively on this latter effect. In this constant-s approximation, two different mutants that occur among those with the same fitness (in particular members of the same clone) will compete equally and sweep together, each becoming only partially fixed. Unless we are interested in the neutral genetic variability of the population, all subpopulations with the same fitness can be considered as a single subpopulation: we do this except in the discussion at the end of this article. Also, we postpone discussion of the interplay between clonal interference and multiple mutants (i.e., going beyond the constant-s model) to a later section below.
First consider starting from a monoclonal population. Mutations initially give rise to a subpopulation with fitness increased by s (Figure 2a). The size of this mutant subpopulation drifts stochastically, but eventually becomes large enough, ∼1/s individuals, to become deterministic. This takes a (stochastic) establishment time, τ1. After its establishment but before its fixation, mutations can occur in the still-small mutant subpopulation to create double mutants with fitness 2s (Figure 2b). This typically happens well before the single mutants have fixed (else we are by Equation 1 in the successional-mutations regime). We assume the double mutants never arise before the single-mutant subpopulation has established; as we discuss below and in appendix g, this will be true unless mutation rates are extremely high or selection is very weak. A double-mutant population thereby becomes established a time τ2 after the establishment of the single-mutant population. Triple mutants then begin to arise and become established after an additional time τ3. This interval is typically shorter than τ2, primarily because double mutants grow faster than single mutants and hence generate more mutations and, in addition, because the triple mutants are more fit than double mutants and hence survive drift more easily (with probability 3s rather than 2s).
This process continues, accelerating at each step. Eventually, however, enough time passes that the single-mutant subpopulation (or one of the multiple-mutant subpopulations) becomes larger than the original wild type. This near fixation of the single mutants increases the mean fitness by s, which balances the accelerating front and creates a moving fitness distribution that will attain a (roughly) steady-state width with the mean fitness increasing with a steady-state average speed, v. This is a form of mutation–selection balance: as each new beneficial mutation becomes established, the mean fitness increases by s and the fitness distribution moves to higher fitness while maintaining the same shape.
It is useful to consider this process in more general terms. The key to the behavior is the balance between mutation, which increases the variation in fitness within the population, and selection, which decreases the variation by eliminating all but the fittest individuals. If we were discussing deleterious mutations, mutation would also oppose the tendency of selection to increase the mean fitness, leading to a steady-state distribution of fitness (ignoring Muller's ratchet, which for large populations only matters on extremely long timescales). This deleterious mutation–selection balance, which is independent of population size for large N, has long been understood (Gillespie 1998). In our case, the dynamics are more subtle because the important mutations are beneficial. The basic idea of mutation–selection balance, however, is unchanged. Mutations broaden the fitness distribution while selection narrows it, creating a steady-state variance around an increasing mean fitness. But unlike the deleterious case, the dynamics of the rare individuals near the most-fit tail of the fitness distribution (the “nose”) control the behavior. We show below that selection moves the distribution toward higher fitness at a rate very close to the steady-state variance in fitness—the classic result in the absence of mutations (the “fundamental theorem of natural selection”) (Fisher 1930). But new beneficial mutations at the nose are essential to maintain this variance: in their absence the fitness distribution would collapse to a narrow peak near the most-fit individual and evolution would grind to a halt.
The crucial dependence on new mutations in the nose makes the analysis of the beneficial mutation–selection balance more complex than in the deleterious case. It is now essential to account properly for random drift in the small populations near the nose. In the case of deleterious mutation–selection balance, rare new mutants are less fit than the rest of the population. They will die out soon anyway, so failing to account properly for the stochastic dynamics by which they do so has no serious consequences. Random drift is important with solely deleterious mutations only if Muller's ratchet is operating, i.e., if the most-fit individuals are rare enough that they can die out due to random drift. The beneficial mutation–selection balance is quite analogous to this Muller's ratchet case. Here too the subpopulations that are more fit than average control the long-term behavior of the population, and these are small enough that correct stochastic treatment is essential. As is the case with Muller's ratchet, infinite-N deterministic approximations are not even qualitatively correct. Indeed, with a large supply of beneficial mutations, deterministic analysis incorrectly predicts a rapid acceleration of the nose toward an infinite speed of evolution. This nonsense result is because of the creation in the deterministic approximation of (what are effectively) fractional numbers of new much fitter mutants that then grow exponentially, unhampered by drift, and dominate the behavior soon after (we describe this in more detail in appendix a).
There are two factors that determine the dependence of the speed of evolution on the population size. The first is the dynamics of already established subpopulations, which is dominated by selection. The second is the new mutations that occur in the fittest subpopulation. We define the lead of the fitness distribution, Q, as the difference between the fitness of the most-fit individual and the mean fitness of the population (more precisely, Q − s is the difference between the mean fitness and that of the most-fit established mutant class). We define q by Q = qs, so that if the lead is Q, the most-fit individuals have q more beneficial mutations than the average individual: they have a “lead” Q in the race to higher fitness. Once it is established, this fittest population grows exponentially. In the time this population took to become established, in steady state the mean fitness must have increased by s, so the newly established population will initially grow exponentially at rate (q − 1)s and later more slowly as the mean continues to advance. Growing from its establishment upon reaching size 1/qs until it reaches a large fraction of N will thus take time , since is its average growth rate during the period between establishment and fixation. In this time the mean fitness will increase by (q − 1)s. Therefore v ≈ [(q − 1)s]2/[2 ln(Nqs)]. One can show that this v is equal to the variance in fitness, as expected if mutation is indeed negligible compared to selection in the bulk (i.e., away from the nose) of the distribution, so that the fundamental theorem of natural selection applies.
The other factor is the dynamics of the nose, where mutations are essential. A more-fit mutant that moves the nose forward by s will be established some time τq after the previous most-fit mutant. Thus the nose advances at a speed v = s/〈τq〉, where 〈τq〉 is the average τq. After it is established, the fittest established population nq−1 will grow exponentially at rate (q − 1)s and produce mutants at a rate Ubnq−1 ≈ Ube(q−1)st/qs. Many new mutants will establish soon after the time τq at which becomes equal to one, so the time it takes a new mutant to establish is . This means the nose advances at rate v ≈ s/〈τq〉 ≈ (q − 1)s2/ln(s/Ub). Significantly, the behavior of the nose depends only on mutations from the most-fit subpopulation; it is almost independent of the less-fit populations and thus can depend on N only via the lead, qs. As far as the nose is concerned, the majority of the population—destined to die out shortly—is important only to ease the competition for the fittest few. Yet we argued above that the bulk of the population fixes the speed of the mean via the selection pressure: . In steady state, the speed of the mean must equal the speed of the nose—the mutation–selection balance. This implies that(4)and(5)These results are very close to the more careful calculations below. All the basic qualitative behavior follows from this intuitive reasoning.
For large NUb, we have found that v depends logarithmically on N and Ub, much slower than the linear dependence on NUb that holds for smaller populations. This reduction occurs because at large NUb, almost all beneficial mutations occur in individuals far from the nose of the fitness distribution (i.e., in a bad genetic background) and are therefore wasted, since these subpopulations are doomed to extinction. Thus increasing N does not directly increase the supply of important mutations, as these occur in the relatively few individuals at the nose. Rather, the effect of increasing N is to increase the time required for selection to move the mean fitness, which increases the lead, which makes individuals at the nose more fit relative to the mean fitness, which speeds the establishments at the nose. Similarly, increasing Ub does not directly affect the dynamics of most of the fitness distribution. Rather, it decreases the time for new mutations to occur at the nose, which means that more mutations can occur before the mean moves, which increases the lead and speeds the evolution.
This also explains why v is not a function of NUb: N directly affects only selection timescales, while Ub directly affects only the mutation supply rate, so v depends on N and Ub separately. It is not a function of the commonly used parameter θ = 2NUb. Instead, it is a function of the parameters Ns (which describes selective forces) and (which describes the strength of selection relative to mutation), and it is valid in the regime where both are large. The expression for q above is of order the basic selective timescale, divided by the basic mutation timescale, , which makes sense since the lead is set by the balance between these two forces. More generally, the two factors that determine the timescales of the multiple mutation dynamics are(6)
Although these are both logarithmic in the population parameters and thus never huge, they can be large enough to be considered as large parameters. Many of our more detailed results are valid in the limit that both L and ℓ are large, with corrections (some of which we include) smaller by powers of 1/ℓ or 1/L.
We show below that our result for v is consistent with the fundamental theorem of natural selection. Viewed in this light, our result for the speed of evolution is not in itself novel: the speed is just the variance in fitness, as usual. What our analysis does is to obtain what this variance is. In many aspects of quantitative genetics, the variance of a quantitative trait (such as fitness here) is taken as some external parameter. When the variance has accumulated during a period when it was neutral and is only starting to be selected on, this may be appropriate. But beyond that, it is surely not. Our analysis deals with the case when variance is accumulating while being selected on. That is, when variance in fitness is increasing due to mutations while at the same time it is being acted on by selection, then, even if the adaptation speed is only indirectly related to new mutations, it is essentially dependent on them: without mutations the variance will rapidly collapse to zero.
However, neither our heuristic analysis above nor our more careful work described below ever explicitly involves the fitness variance. Rather, the natural measure of the width of the fitness distribution is the lead. It is the lead, not the variance or the standard deviation, that can be most productively thought of as a balance between mutation and selection. It is true, of course, that the variance is also increased by mutation and decreased by selection. However, this is not the clearest way to understand the behavior. The increase in the variance from mutations is delayed and indirect. The new mutations that occur at the nose will only increase the variance after they have grown enough—and by then the important new mutations that will keep the variance high later are happening further out in the nose. This is not to say that a variance (and higher-moment)-based approach is impossible, but it is unwieldy and prone to hard-to-understand errors when any approximations are made. We discuss such moment-based approaches in appendix a.
We now turn away from crude (though powerful) intuitive arguments towards more rigorous analysis. We begin in this section by defining the simplest model more precisely. We consider mutation, selection, and drift within a purely asexual population of constant size N. We assume that a large number of beneficial mutations, each of which increases fitness by s, are available and define Ub to be the total mutation rate to these mutations. We consider the situation where the number of beneficial mutations fixed is small compared to the total number available so that Ub does not change appreciably over the course of the evolution (we relax this assumption in appendix c). We neglect deleterious mutations and other-strength beneficial mutations (see later sections below for a discussion of the consequences of these assumptions). These simplifications are not essential and do not change the basic behavior in many situations. Indeed, we argue that these assumptions can all be good approximations even when the situation is more complex, in particular when N or Ub are not constant, or in the presence of deleterious mutations or variable s, as we discuss in detail in subsequent sections. But, more importantly, these simplest approximations make the analysis clearer.
In addition to the more innocuous simplifications, we make two essential biological assumptions: that there is no frequency-dependent selection and that there is no epistasis, so that the fitness of an individual with k mutations is (k − ℓ)s greater than the fitness of an individual with ℓ mutations. When either of these conditions fails, the evolutionary dynamics can be very different from our predictions.
There are two primary difficulties in analyzing the multiple subpopulations that occur even in the simplest model. The first is the stochastic aspects: when a subpopulation with a given fitness is rare, stochastic drift plays a crucial role and must be handled correctly. The second is the interactions between the subpopulations: the constraint of fixed total population size means that there is effectively a frequency dependence to the growth of a subpopulation—albeit a simple one.
To model the stochastic effects, we assume that the basic process of birth and death is a continuous-time branching process. All individuals have the same constant death rate 1, which means that the average lifetime of an individual is 1 (i.e., the units of time are generations) and that the lifetimes are exponentially distributed. Each individual in the population has some number, y, of beneficial mutations. We define to be the average value of y across the population (i.e., the average number of beneficial mutations per individual). An individual with y beneficial mutations has a birth rate . This ensures that the average birth rate in the population is 1, so the population stays at a constant size N. We assume all individuals give rise to mutant offspring at rate Ub, independent of their birth rate (i.e., mutants arise at a constant rate per unit time). If mutations instead occur at a constant rate per birth event, our assumption underestimates the mutation rate for the most-fit individuals. However, we always assume for all individuals (i.e., the lead, Q ), so that the two definitions are almost equivalent.
The branching process model allows one to calculate simple analytic expressions for a number of important quantities that are not readily available in diffusion approximations of the standard Wright–Fisher model. However, branching process models cannot easily deal with the nonlinear saturation effects required to maintain a constant population size. By “saturation” effects, we refer to when a mutant subpopulation has become large enough to influence the mean fitness of the population and hence begins to compete with itself, slowing its growth: this is the essential effect of the fixed total population size. To handle the saturation effects, we make use of a simple observation: stochastic effects are important only when a subpopulation is rare, while saturation is important only when a subpopulation is common. Thus we use the stochastic branching process model, ignoring saturation effects, to describe the dynamics of a subpopulation while it is small. Conversely, when it is large, we ignore random drift and treat it with the correctly saturating deterministic equations. Our use of both deterministic and stochastic analyses requires an appropriate way of linking the two together. In this article, we describe a method for doing so. This method accounts for all of the important aspects of genetic drift and is simple and intuitive. It should be of broad applicability to related evolutionary problems.
This approach works as long as the stochastic regime and the saturation regime are different. That is, a subpopulation must become large enough to neglect random drift before it is too large to ignore saturation. We can treat a subpopulation of size n deterministically so long as . On the other hand, saturation can be ignored when . Thus to separate the stochastic and the saturating phases of growth of a subpopulation, we require . Throughout this article, we assume this condition holds. Unless s is extremely small (s ∼ Ub), a population small enough that will usually be too small for clonal interference or multiple mutation effects to matter, so this is not a serious limitation.
A situation in which there are multiple subpopulations of varying sizes is illustrated in Figure 3: this shows the logarithm of a typical fitness distribution within a steadily evolving population. Where the subpopulations are small, at the front of the distribution, stochastic analysis is necessary but nonlinearities can be ignored. When a subpopulation represents a substantial fraction of the total, nonlinear saturation is important but stochasticity is not. As long as , there is an intermediate regime where neither matters. We can thus use a nonlinear deterministic analysis in the bulk of the distribution and a linear stochastic analysis near the front and match the two in the intermediate regime in which both are valid. These approximations are fully controlled and any corrections to our results will be small for .
Relationship of our model to the Wright–Fisher model:
The deterministic limit of our model is identical to that of the Wright–Fisher model. However, the stochastic dynamics are slightly different. In the Wright–Fisher model, all individuals have a lifetime of exactly one generation, while in our model individuals have a random exponentially distributed lifetime with mean one generation. In the Wright–Fisher model, the distribution of the number of offspring per individual is approximately Poisson, while in our model the number of offspring is geometrically distributed. Both the mean lifetime and the mean number of offspring per individual are identical in the two models (hence identical deterministic dynamics), but the different distributions do lead to slight differences. In particular, although the probability a beneficial mutation of size s () will become established is proportional to s in both models, it is ≈cs with the coefficient c = 2 in the Wright–Fisher model and c = 1 in ours. Since it is likely that the population dynamics in any real population are not well represented by either of these models, there is no one “correct” model [e.g., for populations dividing by binary fission, as in many experimental studies of evolution, the establishment probability is closer to 2.8s (Johnson and Gerrish 2002)]. Fortunately, in our analysis of the behavior of large populations, these differences cause only negligible corrections in the arguments of logarithms [e.g., replacing ln(Ns) with ln(cNs) when ]. For smaller populations, however, the speed of evolution is proportional to the probability of establishment and thus does depend on more details of the model: in particular, the successional-mutation result for the speed is v ≈ cNUbs2.
It would in principle be possible to use a diffusion approximation to the Wright–Fisher model instead of our branching process model. This would have the advantage of being able to handle saturation and drift at the same time and thus cases where . Such a model could in principle treat all the different subpopulations stochastically, including all mutations between these populations. However, this would lead to a complex and difficult to analyze infinite-dimensional diffusion process. There is, however, a controlled approximation—valid for large Ns—to the full diffusion process that is exactly equivalent to ours; as it would add little, we do not discuss this explicitly here.
This section contains the primary analysis presented in this article: the accumulation of beneficial mutations in the simple model described above. We begin by looking at what happens to a single mutant individual. We then ask what happens to a mutant population that is being fed constantly by new mutations. We next couple this analysis to the behavior of the rest of the population to gain an understanding of the evolution of large asexual populations and obtain our primary results. Finally, we connect this behavior to the small-population regime.
The fate of a single mutant individual:
We begin by considering the fate of a single mutant individual. We assume that in a large clonal population of size N, at time t = 0 there is a single mutant individual with a beneficial mutation conferring fitness advantage s. We denote the size of the subpopulation carrying this beneficial mutation at time t as n(t) [by assumption, n(0) = 1]. We study the effects of selection and drift on this population by calculating the probability distribution of future n(t), , assuming that no further mutations occur. This provides an essential building block for all the subsequent analysis and also illustrates our basic approach in a simple context.
Throughout this analysis, we assume that the number of individuals with the beneficial mutation is small relative to the total population size, . Thus the mutants do not interfere with one another. Naturally, if the mutant becomes established it will supplant the wild-type population and this condition will cease to be true. By this time, however, the mutant subpopulation will be large enough that we can switch from the stochastic analysis described here to a correctly saturating deterministic analysis.
Because the mutant subpopulation is too small to affect the mean fitness, mutant individuals have a birth rate 1 + s and death rate 1. We define g(n, n0, t) to be the probability of having n descendants at time t, starting from n0 descendants at t = 0. We are interested in calculating g(n, 1, t). The probability of a birth or a death event in a unit of time dt is (2 + s)dt, and this event is a birth with probability and a death with probability . This means that(7)where δn,0 = 1 if n = 0 and is 0 otherwise. This is a standard birth–death process (Allen 2003). Assuming that individual lineages are independent and defining the generating function(8)we can rewrite Equation 7 as a differential equation for G(z, t), which we solve to find(9)
We can now determine from G(z, t). A standard inversion yields(10)valid for n > 0, and(11)
We are interested primarily in understanding the distribution of n given that the mutant population is not destined to go extinct. This is given approximately by(12)Here we have approximated the geometric factor by a simpler exponential in n that is valid for , the regime of primary interest. Note, however, that although the crucial features are more apparent in the approximate expression, all the results below follow from the exact equations.
At this stage, the above results merely reproduce classical analysis, but it is useful to pause to compare them with various intuitive predictions. We first compute the average number of mutant individuals at time t,(13)which confirms our understanding of what it means to have a beneficial mutation with advantage s. However, most of the time the mutation will die out. Conditional on not going extinct,(14)which is larger at long times by a factor of 1/s. At short times, , this is 〈n | not extinct〉 ≈ 1 + t. At long times, , the extinction probability becomes , and . Note that short times correspond to , while long times mean . (Note also that none of these expressions saturate as n approaches N; they are valid for , as discussed above.)
It is useful to ignore mutations that are destined to go extinct due to drift and focus only on those that are destined to become established. We do this for the remainder of this section; all results are thus implicitly conditional on nonextinction. However, some care is required. If a mutation occurs at time t = 0 and survives drift to become established, it may seem that on average it will grow as n(t) = est, because it started from one individual at t = 0 and grows on average exponentially. However, this is incorrect. Given that it survived drift, it is likely to have grown faster than est in the early stochastic phase of its growth during which drift is faster than selection (Otto and Barton 1997; Barton 1998). This is apparent from the expressions above: for , 〈n | not extinct〉 ≈ 1 + t, which is much faster than 〈n〉 = est ≈ 1 + st. Once the population is large and stochastic effects can be neglected, it naturally grows as est. However, because it grew faster than this in the early stochastic phase, it will on average be larger than if it had grown this fast through its entire history. As is clear from the expression for the average n at long times, , the behavior can be crudely approximated by assuming that it started at size (rather than size 1) at t = 0 and then grew exponentially as est thereafter. This approximation is of course not valid during the early phase of growth. Note that the above also implies that, given that a mutation is not destined to go extinct due to drift, it will fix in a time of order , not , as is sometimes seen in the literature. For s ∼ 0.01, this is a difference of ∼500 generations. To be more precise, the fixation time is a random variable with a distribution of width 1/s and mean close to , rather than the naive .
For much of the subsequent analysis, we are concerned with the size of a subpopulation only after it is big enough to be essentially deterministic. Yet as the above discussion makes clear, the stochastic phase of growth affects the later deterministic dynamics. Thus we are interested in “summing up” the stochastic effects in terms of their impact on later deterministic growth.
Focusing only on the effects of stochasticity on later deterministic dynamics allows us to make a key simplification. Once the subpopulation is large enough to grow deterministically, but still small enough that saturation can be ignored (i.e., ), its dynamics can be described by n = νest. The value of ν is a random variable that depends on how fast the population grew in its stochastic phase. However, the only effect of this stochasticity on the later deterministic growth is to create random variation in ν. As almost all this stochasticity accumulates at short times, at large t (after the population has become deterministic) we can describe the overall effects of stochasticity in terms of a probability distribution . This is a big simplification, because the full probability distribution conditioned on nonextinction, A(n, t), depends on both n and t, while for large t is independent of t, as we show below. This simplification is possible because at large t the only time dependence is the deterministic exponential growth.
We can justify the above heuristic argument rigorously. The definition of ν is just a transformation of n, ν ≡ ne−st. This is valid in the early stochastic phase of growth as well as in the later deterministic phase. However, in the stochastic phase we do not expect that ν will be independent of t. As we have the probability distribution A(n, t), it is straightforward to transform this to the distribution . When we take the large-t limit of , it becomes independent of t. This justifies our expectation that at large t, we have , independent of time.
Rather than using the probability distribution of ν, it will prove useful to define a related variable τ by(15)
The random variable τ is simply related to ν: . Since τ is a simple transformation of n, we can immediately calculate (with the probability density as we are treating τ as a continuous variable) from A(n, t). We find(16)
As with ν, this describes the distribution of n both in the deterministic and in the stochastic phase. Since n depends on t, so does the distribution of τ. However, as expected from the previous discussion, the distribution of τ becomes independent of t for large t. We define τest as τ(t → ∞) and find(17)The average value (as well as higher moments) of τest can be easily computed from this distribution. We have(18)where γ is Euler's constant γ = 0.577216.
We see from Equation 16 that the large-t condition required for the distribution of τ to become independent of t is . This is the time at which . This indicates that our choice of as the size at which a population becomes established is appropriate. After a time , when the population on average reaches this size provided it has not gone extinct, the probability distribution of τ begins to become independent of t, indicating that the behavior of the population crosses over from mostly stochastic to mostly deterministic.
The variable τest has an intuitive interpretation: τest is the time at which n would have reached size had it always grown deterministically, as calculated by looking at n(t) at large t and extrapolating backward. This is illustrated in Figure 4a. We can therefore approximate the destined-to-be-established subpopulation as drifting randomly for a time τest, at which time it reaches size and then grows deterministically thereafter. With this simplification, the only important stochasticity is the duration of the drift period. This is the key simplification that allows us to smoothly connect the branching process with the nonlinear dynamics once the subpopulation is no longer rare. It jibes with our intuitive expectation that the subpopulation is dominated by drift when rarer than and then behaves deterministically once it exceeds this size. Note, however, that in addition to telling us nothing about n(t) before time τest, it also gives a slightly inaccurate picture immediately after τest when n(t) is ∼. The time τest is not in fact the time at which the subpopulation reaches size (see Figure 4a). Rather, it is the time at which n(t) would have reached size if we assumed that it always behaved deterministically, but it gets the large-t behavior right. In fact, some small drift does take place after reaching size ; our approximation does not ignore this drift, but rather adds up all the drift that takes place through all the time and rolls it into a change in τest. This can thus be thought of as the time at which the mutation establishes. In asking how quickly beneficial mutations accumulate, this is the most natural variable.
The caveats above illustrate why it is perfectly consistent to have τest < 0; the distribution B(τest) above shows that this is not even particularly improbable. This reflects the fact that, given that a mutant subpopulation is not going to go extinct, it is reasonably likely to grow remarkably fast in the early stochastic phase. A τest < 0 simply indicates that the mutant subpopulation grew so fast when rare that if we look at the subpopulation size much later and assume it always grew exponentially at rate s, the subpopulation would have had a size > at t = 0.
We note that , while for large t (as always, conditional on nonextinction). This may naively seem inconsistent, since for large t. However, it merely reflects the fact that 〈eX〉 ≠ e〈X〉. The difference between these two averages is in fact the essential reason that τest will prove to be such a useful variable to focus on. This is because the value of 〈n(t)〉 depends much more sensitively on the tails of than does 〈τest〉.
Mutants generated by a changing population:
The above analysis of the population size of a clone founded by a single mutant individual is an important building block. However, it does not address the full problem. We must now ask how the mutants arise in the first place. In the simplest case, we might imagine a wild-type population of size N, starting with 0 mutants at time t = 0. This population generates mutants at rate NUb. Each mutant follows the dynamics given in the above section, beginning at the time it was created, but now we have multiple such initial mutants that are created at random times.
Generally, the relevant process is even more complex. Starting from a wild-type population, a single-mutant subpopulation is generated, experiences a stochastic period, and then begins to grow deterministically. Then double mutants are created by mutation within the single-mutant population while it is still growing (i.e., before it fixes). The rate at which these double mutants are generated increases with time because the single-mutant subpopulation is growing. Later, the double mutants may themselves generate mutants before they fix (and possibly before the single mutants fix), and so on.
We therefore must tackle a more general problem: the distribution of the population size n(t) of a mutant subpopulation that starts with 0 individuals and is “fed” by mutants from a less-fit subpopulation of (growing) size f(t). If this less-fit clone is small enough that its growth is stochastic, calculating the probability distribution of the mutant subpopulation is extremely complex. Fortunately, most nonviral organisms live in parameter regimes where a clone will never generate mutants destined to establish while it is still so small that it must be treated stochastically. As we discuss in appendix g, this parameter regime is , which we will generally assume. Thus we take f(t) to be some deterministic function describing the growth of the clone from which mutants arise. Later we set the origin of time in f(t) stochastically, to reflect the stochasticity in the establishment of this feeding population.
Note that we no longer need to condition on the mutant subpopulation not being destined to go extinct. Since this subpopulation is being continuously fed with new mutations, eventually one of these mutations will survive drift. Thus at long times the mutant subpopulation will never be extinct.
Unlike in the previous section, the growth rate of the stochastic mutant population n(t) is not necessarily 1 + s. Rather, the growth rate is , where ys is the fitness of the subpopulation n(t) and is the mean fitness of the population. For convenience, we write this as 1 + rs. The death rate of this population is still 1. Since increases continuously, r is time dependent. Despite this, we approximate r as a constant. This is justified because we use the stochastic description of n(t) only during the brief period during which it is rare, and in this time r does not change significantly. We discuss this approximation in appendix h.
We define η(t − tk) to be the number of descendants at time t of a single mutant that occurred at time tk. That is, given that a mutation occurs from the “feeding” population at time tk, η(t − tk) is the number of descendants of this mutation at a later time t. Note that η is the random variable whose generating function is given by G(z, t − tk) from Equation 9 above, but with s replaced by rs. We have(19)where M is the random number of individual mutations that have occurred and Tk are the random times at which they occurred.
The number of mutations and their timings are an inhomogeneous Poisson process, fed by the population f(t). We therefore have(20)Note the lower limit of integration here represents the earliest time that mutations are allowed to occur; we have chosen this to be infinitely early. We discuss this choice of cutoff more generally in appendix e. The timings of the mutations Tk, conditional on M = m, are the ordered statistics of m independent identically distributed samples drawn from the distribution(21)This means that the joint distribution of the Tk conditional on m is given by(22)
The generating function for the distribution of the number of mutant individuals, n(t), is given by H(z, t) = 〈zn(t)〉. Note that . Conditioning on the distributions of M and the Tk given above, and using the fact that , we find(23)where the integral is over all ordered configurations of the tk. Substituting the distributions of M and the Tk above, we find(24)To understand the full probability distribution of n(t), we simply have to plug in the appropriate form f(t) and then invert this generating function.
An exponentially growing population feeding another:
In large populations, there will typically be various multiple mutants present, as illustrated in Figure 2. We can now apply the results of the previous section to this situation. As before, we define the the most-fit subpopulation that is large enough to treat deterministically to have fitness (q − 1)s above the mean fitness (note that q is not necessarily an integer). This subpopulation, nq−1, grows exponentially at rate (q − 1)s. We define the origin of time such that nq−1(t) is given by(25)Note that, analogous to the previous section, we are approximating q as constant—we discuss this further below. The reason for defining the origin of time such that at t = 0 will become clear below. We now want to understand the stochastic dynamics of the subpopulation a fitness qs above the mean [denote this population size by nq(t)]. The subpopulation nq−1 feeds mutations to nq; we therefore have f (t) = nq−1(t) in the notation of the previous section.
This problem involves one exponentially growing population, nq−1, feeding another, nq. In analyzing it, we first step back from our specific situation to study the general case of an exponentially growing population with with size feeding mutants at rate Ub to a stochastic population N2 that on average grows exponentially with rate R2. We later will substitute , R1 = (q − 1)s, and R2 = qs. We begin by plugging into Equation 24, using the obvious generalization of G(z, t) to a population that grows at rate R2. This gives us H(z, t), the generating function of the probability distribution of N2. It is convenient at this point to pass from generating functions to Laplace transforms by defining the transform variable ζ = 1 − z. For our purposes we can assume that ζ is small: this introduces errors into when N2 ∼ 1, but we will never use in this regime. We find(26)Substituting , we find(27)Assuming that ζ is small, the integral in this expression is independent of ζ and is given by . We find(28)
We can now substitute our values of ν1, R1, and R2 to find that in our case(29)This is the standard form for the Laplace transform of a one-sided Levy distribution, a well-studied special function. An integral representation of this is the inverse Laplace transform of H,(30)where the integral is over the imaginary axis. For large nq this can be integrated to give . [Note this distribution has infinite 〈nq〉, an unimportant and unbiological artifact of our choice of cutoff in the integral for H(z, t); this is discussed in appendix e.]
To understand this distribution P(nq, t), we define a variable τq similar to that described in the section above on the fate of a single mutant. We first define(31)As before, τ is time dependent, but for t → ∞ the distribution of τ is independent of t. We define τq ≡ τ(t → ∞). As before, τq is the time at which the subpopulation nq(t) would have reached size had it always grown deterministically at rate qs, as calculated by looking at the size nq(t) at large t and extrapolating backward. Unlike τest, the value of τq includes the time for the mutation (or mutations) to arise in the first place as well as time for their initial stochastic growth. This is illustrated in Figure 4b.
As in the section on the fate of a single mutant, we can think of the mutant subpopulation as drifting randomly for a time τq, at which point it reaches size and thereafter grows deterministically. We therefore sometimes refer to τq as the “establishment” time. As before, this is somewhat inaccurate in describing the dynamics right around τq (or before) when the population is around or below a size of . Again τq is not actually the time the population reaches size . This is because both future random drift and future feeding mutations, after the population reaches size , are included in the estimate of τq. However, for the purposes of understanding the dynamics of the mutant population once it becomes large compared to , it is valid to think of τq as the time it takes the population to reach size .
We often wish to use moments of τq. These are straightforward to calculate in principle, but somewhat tricky in practice. We first note that because of the definition in Equation 31, we have(32)We can therefore calculate 〈τ〉 by computing 〈ln nq〉 and plugging into this expression. Higher moments of τ are easily computed by similar expressions; these depend also on higher moments of ln nq. We can calculate 〈lnmnq(t)〉 by noting that . Using the integral representation of P(nq, t), we have(33)where the ζ-integral is over the imaginary axis and we have defined(34)We integrate this to find(35)where Γ(x) is the Gamma function.
We can now calculate derivatives of this with respect to μ to get 〈lnmnq〉 and hence the moments of τ. For large t, as expected, τ becomes independent of t. For the mean of τq, we find(36)where γ = 0.5772 is Euler's constant. The variance of τq is given by(37)Higher moments are also simple to compute if desired (and demonstrate that there is substantial skew in the distribution of τq, as τq substantially smaller than 〈τq〉 can occasionally occur, while τq substantially larger than this almost never does—this is important in understanding the fluctuations in the rate of adaptation around its steady-state value and is discussed in appendix d).
This calculation of 〈τq〉 is somewhat involved because of the need to use the integral representation of P(nq, t). We can get rough estimates (often useful in other contexts) via a simpler method. Namely, we define a typical population size , where ζ1 is defined by H(ζ1, t) = e−1. As is apparent from the definition of a Laplace transform, , for well-behaved distributions this typical value is roughly like the median of nq. We can then get a typical value from this by using the relationship between ln nq and τq. Doing this leads to a that is very close to the 〈τq〉 calculated above.
Note that the careful result for 〈τq〉 is similar to the crude result in the heuristic analysis section above, which approximated the time required for a new mutation to arise at the nose as , roughly the typical time at which the first mutant destined to establish arises. This crude expression is only weakly dependent on the lower cutoff to the integral, which is good since nq−1(t) is not given accurately by the deterministic approximation in this regime. This weak dependence appears for the same reasons in the careful calculation of τq and is discussed in more detail in appendix e. The crude and careful results do differ, however. The careful result accounts properly for the randomness in the timing of a new mutation and the fluctuations during its early drift phase. It also accounts for the fact that not only the first mutant destined to establish at the nose contributes. Rather, as we see later, of order q different mutations contribute significantly to the establishment of a new most-fit subpopulation at the nose.
The rate of evolution and maintenance of variation at large N:
We are now in a position to calculate the rate of evolution and amount of variation maintained in large populations. In the above calculations, we set t = 0 to be the time at which the population nq−1 reached size . This corresponds to the establishment time of this population. After a (stochastic) time τq, the next more-fit subpopulation, nq, establishes. For the later deterministic dynamics of nq, we can think of this as the time when nq reached size . At this point, we have reached the identical situation where we started, but with the nose of the population fitness distribution moved forward by s. In the steady state, the mean fitness of the population must also have moved forward by s in the average establishment time 〈τq〉. Thus the population at nq now has fitness only (q − 1)s ahead of the mean. It has size , but thereafter grows exponentially only at rate (q − 1)s, giving a population size .
The process now repeats itself—we can take this establishment time of the new population (nq, above) as the new t = 0, and after that this population grows as we had described for the original population nq−1. In fact, it now is the population nq−1, since the mean fitness has increased by s. Thus we can see that the mean fitness of the population and the position of the nose move forward by s in a time 〈τq〉. Thus the average rate of increase in fitness in the population is(38)Note that this discussion makes clear why, for consistency, we defined the establishment time for nq−1 to be when this population reached size , not . We also note that the population that we had originally called nq−1 is now nq−2 and its size is given by .
This change in the growth rate of the population we had originally called nq−1 raises an important point. We defined and used this expression in calculating P(nq, t), particularly for large t. Yet at this large t, our expression for f (t) is not accurate, because the mean has shifted and the population with original (relative) fitness (q − 1)s is no longer growing exponentially at rate (q − 1)s. Fortunately, the mutations that occur after the establishment of nq [when the expression f(t) becomes inaccurate] do not greatly affect its later population size, nq(t). In other words, the mutations that dominate the population nq happen early while nq−1 is still accurately given by f(t). Yet one must also ask whether these mutations happen too early when f(t) is also not a good approximation for nq−1(t) [because the definition of τq, which we used to define f(t), includes mutations and stochastic behavior that happen later]. Fortunately, the mutations that matter from nq−1 to nq do occur late enough that nq−1 is accurately described by f(t). This can be checked by studying the behavior of τ(t); we discuss this and related subtle issues in appendix g.
When q is too small, the approximations above are no longer justified. Whenever q < 2, the growth rate of the subpopulation nq−1 slows substantially during the period while the important mutations to nq are occurring. That is, nq−1 saturates while nq is becoming established. Thus our analysis in this section is valid only for q > 2. As we will see, this corresponds to large N. We discuss the q < 2 case in the next section. However, it is the large-N, q > 2 result that we are most interested in—this is where there are typically many multiple mutations at once, and the behavior differs dramatically from the successional-mutations regime.
Throughout this section, we have asserted a steady state in which the mean fitness increases at the same rate as new mutations are established and have defined the lead in steady state to be qs. Yet we have not discussed the balance between mutation and selection that sets this steady state. We now turn to this question. Roughly speaking, we expect that in larger populations, elimination of less-fit clones takes longer, and more mutations can arise in this time, so the steady state q should rise.
The relationship between q and N can be obtained from τq. As we have seen, immediately after the subpopulation at q becomes established, its size is . The subpopulation at q − 1 has size , the subpopulation at q − 2 has size , and so on. All of the subpopulations must add up to size N; in practice the total is dominated by one or a few (compared to q) subpopulations so that we can equate N to the size of the largest subpopulation, the one whose fitness is closest to the mean fitness. Imposing this condition and assuming that all the τq are on average 〈τq〉, we find(39)This is a transcendental equation for q, but because of the logarithmic dependence on q on the right-hand side it is easily solved by iteration. For most purposes, even the zeroth approximation,(40)is sufficiently accurate. To get higher accuracy one can plug this into the right-hand side of Equation 39.
As expected, the value of q increases with N and also increases with Ub because when mutations happen more quickly there are more of them in the population at once. The dependence on s is more complicated, because increasing s both decreases the fixation time (leaving less time for additional mutations to occur) and increases the rate of mutations that establish (because it increases the establishment probability).
With the value of q determined self-consistently above (Equation 39), the mean fitness shifts by s in exactly the time 〈τq〉. Thus the corresponding distribution of the subpopulations is indeed a steady state (see appendix d for a discussion of fluctuations around this steady state and its stability). By plugging Equation 39 into the expression for 〈τq〉 and substituting this into , we can obtain the speed of evolution. Doing this using the lowest-order result in the iterative expansion for q (Equation 40), we find that the speed of evolution is roughly(41)valid provided q is reasonably large [basically, when 2 ln[Ns] > , which will tend to be true when ]. If a more accurate result is needed, we can simply carry the iterative expansion for q to higher order.
The calculations above confirm the intuitive picture and results described in the heuristic analysis section above. The speed of evolution is determined by two mostly independent factors. One factor is the dynamics of the nose—the feeding process from nq−1 to nq that sets τq. This process depends directly only on Ub and s; the only impact of N here is via its effect on the lead qs. The other factor is the dynamics of the already established populations. This is dominated by selection and hence depends directly only on N and s; the only role of mutation here is its role in setting q.
Our result is consistent with the fundamental theorem of natural selection, which states that the speed of evolution is equal to the variance of fitness in the population. To see this, we first note that the bulk of the fitness distribution is Gaussian. This is because a population with ℓ more (or less) mutations than the mean grows (or shrinks) as eℓst, and the mean shifts by 1 during every time interval τq. This means that at the end of an interval, the number of individuals with ℓ mutations more or less than the mean is determined by its cumulative growth or decline over all these time intervals: , a Gaussian distribution. We call the variance of this fitness distribution σ2. The number of individuals that differ from the mean by ks is then roughly Nσ exp [−(ks)2/2σ2], and the fittest established population—with k ≈ q—will have of order individuals. We therefore expect . This means that if the fundamental theorem for natural selection holds, we expect . And indeed, some algebra verifies that this yields the expression for v in Equation 41.
The fundamental theorem of natural selection should apply whenever mutation can be neglected compared to selection. Since this is true in the bulk (i.e., away from the nose) of the fitness distribution, the correspondence between our result and the theorem is reassuring. The speed of evolution is equal to the variance in fitness, as usual. Thus our calculations can be viewed as an analysis of how much variance in fitness a population can maintain while at the same time this variation is being selected on. Yet nowhere did our analysis depend on the variance in fitness. Rather, the lead proved to be a more useful measure of the width of the fitness distribution, because it is the lead that is directly affected by new mutations at the nose. The variance is of course also increased by mutations, but only as a consequence of the dynamics of the lead and only after the new mutant populations have grown to substantial numbers. The key fact that the distribution is close to Gaussian out almost to the nose, which is many standard deviations above the mean, is indicative of the small significance of the region near the mean that controls the variance.
Evolution at moderate N:
In addition to the evolution at large N, we want to understand the crossover between small-N and large-N behavior. In this subsection, we explore this crossover.
For very small N, the successional-mutations regime obtains. In the heuristic analysis section, we noted that mutations take ∼ generations to establish in this regime and then fix in a much shorter time. Thus evolution is mutation limited, and we have v ≈ NUbs2. It is instructive to redo this calculation using the machinery we developed for the large-N case. To do this, we must replace the exponential form for f(t). As before, we take the establishment time of the mutation at (q − 1)s to be t = 0. Of course, here q = 1 so (q − 1)s = 0. In this regime, each mutant fixes soon after becoming established. For the purposes of the next establishment, we can therefore approximate the population at (q − 1)s by(42)where θ(t) = 1 for t > 0 and 0 otherwise. We substitute this form of f(t) into H and integrate and take the inverse Laplace transform of the result to obtain(43)This gives , so the velocity v ≈ NUbs2, as expected.
We now turn to the intermediate regime. For NUb comparable to , the fixation time is not short compared to the establishment time. Thus we cannot use f(t) = Nθ(t). At the same time, the establishment time is not so short compared to the fixation time that saturation in the feeding population is unimportant (the large-N case we have focused on thus far). We therefore need to consider the case of a growing and saturating population feeding another. We assume that the single-mutant population always fixes before the triple-mutant population establishes, so that we have to consider only two deterministic clones and one stochastic clone in the population (i.e., q between 1 and 2). The dynamics of the single-mutant population a time t after it establishes are given by(44)Note that f(t) initially grows as e(q−1)st, with q = 2, but later slows to e(q′−1)st with q′ = 1 (i.e., it becomes approximately constant). This slowing occurs over a time interval of order 1/s, which is much smaller than the establishment times and is thus effectively a sharp transition. The behavior of the feeding population is thus roughly equivalent to having q between 1 and 2. The stochastic population that it feeds initially grows at rate qs with q = 2. The establishment of this stochastic population occurs at a time τ2 when, roughly,(45)with c of order unity. This yields(46)A more careful analysis (analogous to the earlier calculations of τq) that takes into account the distribution of τ2 yields a result that is the same as the above simple argument but with a factor of order unity inside ln Ns, which is a small correction over the whole range of validity. While in general c will depend on the detailed birth and death processes, and the speed of evolution in the successional mutations regime will be proportional to c, for the dynamics we have analyzed throughout, c = 1. We use this below. For , we obtain(47)which crosses smoothly—and simply!—over from the successional-mutations behavior for but to , which is just the result we obtain for q = 2. When NUb becomes of order unity, from the above expression we have . For the behavior is well into the multiple-mutations regime we analyzed earlier, and the results obtained for general noninteger q > 2 apply. The two sets of results match together for Ns ≈ s/Ub, up to order-unity factors inside logarithms of Ns and of s/Ub. An example of the crossover between the two regimes is shown in Figure 5a.
So far, our analysis has assumed that the mutation–selection balance has already been reached. If a population starts with an arbitrary distribution of fitnesses, it will gradually approach the steady-state distribution. A full analysis of this is beyond the scope of this article, but in this section we provide an outline of the important effects and briefly describe a method for analyzing this transient behavior. We focus on the case where the population is initially monoclonal. Other starting fitness distributions can be analyzed using similar methods. We consider the large-N concurrent-mutations regime (in the successional-mutations regime the monoclonal population is already essentially in steady state).
Starting from a monoclonal population, we can calculate the dynamics of the single-mutant subpopulation that arises by using the small-N results above, since here too the feeding population is f(t) = Nθ(t). It would now be tempting to assume that this single-mutant population just grows exponentially at rate s after first becoming established. We could then immediately import our previous results for the establishment time of the double-mutant population, τ2, triple-mutant population, τ3, and so on. We could then assume that all these populations establish in order until the qth population, at which point the steady state would be reached.
Unfortunately, this is wrong, for two reasons. First, the single-mutant population grows faster than exponentially at rate s because it is receiving mutations from the still-large wild-type population. Because of this, the double-mutant population establishes more quickly than the steady-state τ2 and then itself grows faster than exponentially with rate 2s because it is receiving more mutants from the fast-growing single-mutant population. This then affects the triple mutants, and so on. The second complication is that the mean fitness does not stay at the wild-type value until the qth mutation has established, so it takes more than q establishments to reach steady state.
Rather than attempt to find a closed-form analytical result, we discuss here an algorithmic solution to the transient dynamics. We proceed in steps. First, we calculate the lead from the current fitness distribution. On the basis of this, we calculate the next establishment time (interpolating if the lead changes during this period because of an increase in the mean fitness). We then calculate the new fitness distribution and the new lead and repeat the process.
When calculating the establishment times, we must remember that the feeding populations are not necessarily growing as simple exponentials. Earlier we used the establishment time τp to approximate the population size of np as . We noted that this is inaccurate while np ∼ , because it includes both future mutations from np−1 to np and future stochasticity. Since we have used this form of np(t) to calculate the establishment time of the next more-fit subpopulation, this approximation for np(t) must be accurate by the time the mutations that lead to the subsequent establishment occur. In the steady-state case, this holds, as shown in appendix g. However, for the transient dynamics it is not always correct.
This problem is most serious for the single-mutant population, which we consider now. The wild-type population has roughly constant size N during the period when the single-mutant population is rare. This means that the single-mutant population grows on average as(48)This reaches size after a time of order generations. However, the inferred establishment time (by extrapolating backward) is generations. This is substantially negative because mutations that occur well after the population reaches size contribute significantly to n1. The approximation we used before would be to take in calculating the establishment time of the double-mutant population τ2. But using the correct form of n1, we find that the first double mutants occur roughly at time . Thus when , double mutants do not occur until our usual approximation for n1 becomes reasonable. We can therefore use our previous calculation of the establishment time τ2 from the steady-state analysis above. All future establishment times (i.e., τ3 for the triple mutants, etc.) can similarly be imported directly from the steady-state calculations. However, when , we must use the correct form of n1 to calculate τ2 and n2. In this case, n2 will also grow faster than our usual approximation would predict. We must therefore repeat this procedure to consider whether it is reasonable to calculate τ3 on the basis of our usual approximation or whether we need to use the more complex form for n2. However, this effect is much weaker than for n1; it matters only if NUb is much larger than in the previous condition. If it does matter, we must again ask if the more complex form for n3 will be important in calculating τ4; this will matter only if NUb is larger yet. In practice, in comparing with previous experiments we have found that considering the complex form of n1 in calculating τ2 is sometimes necessary, but all future establishments can be calculated using the steady-state large-N results (Desai et al. 2007), because in these experimental situations q is never much larger than 4.
A second subtlety in the above algorithmic approach is the way in which the mean fitness changes; it does not increase in evenly spaced steps of size s as it would in steady state. For example, the double-mutant subpopulation can become established soon after the single-mutant subpopulation does. Then, as it grows twice as fast, it will outcompete the single-mutant subpopulation while both are still rare. We call such an event a “jump,” since it will lead to a jump in the mean fitness by 2s when the double mutants become the dominant subpopulation. Of course, it is also possible that the triple mutants will jump past the double mutants or that the double mutants will jump the singles, and then the quadruple mutants will jump the triples, etc. These effects can lead to complex dynamics of the mean fitness before the steady state is established. However, given the establishment times of the various populations, the time dependence of the mean fitness is straightforward to calculate from the deterministic dynamics of the competing subpopulations that are growing exponentially.
Putting all these effects together, we can construct an algorithmic solution for the transient dynamics. We calculate the first establishment time and note at what time this new subpopulation will change the mean fitness. We then calculate the next establishment time and again the implied future effects on mean fitness (modifying previous such results if jumping events will occur). We continue to repeat this process. When the mean fitness changes, we note how this changes the lead and adjust the establishment times appropriately. We iterate this process until the steady-state lead, qs, is reached. Even after that there can be some lingering effects of the transient, as the rest of the fitness distribution may not yet have reached the steady-state Gaussian profile. Yet soon thereafter the steady-state behavior is indeed reached.
Rather than using this algorithmic approach, it is also possible to use a deterministic approximation for the transient behavior. Starting from a monoclonal population, the timing of the first few establishments is given accurately by a deterministic approximation. However, this typically cannot give us the full transient dynamics, because stochastic effects at the nose become important once the fitness distribution grows to a substantial width, which usually occurs before the transient regime is over. This deterministic approach is also less versatile, as it is valid only for some starting distributions.
The transient behavior can be quite important. During the transient phase, the accumulation of beneficial mutations proceeds more slowly than in the steady state, because after the first few establishments, but before the steady state is reached, the lead will be ps with establishment interval ∼τp < τq (since p < q). Thus a clonal population will accumulate beneficial mutations slowly at first, before the rate of accumulation gradually increases to its steady-state rate. This slower transient phase lasts a substantial time—longer than it takes to accumulate q mutations once the steady state has been established, again because τp < τq for p < q (and, as noted above, in fact it can take more than q establishments to reach the steady state). While this section provides a rough sketch of the behavior, a detailed analysis of these transient effects remains an important topic for future work.
Our simplest model neglects deleterious mutations. But deleterious mutations can alter the dependence of v on the mutation rate (and on N), because increasing Ub typically comes at the cost of also increasing the deleterious mutation rate. This has proved an important consideration in clonal interference analyses (Orr 2000; Johnson and Barton 2002). In this section, we consider qualitatively and semiquantitatively various effects of deleterious mutations in the simple model in which all the beneficial mutations have the same s. The effects of deleterious mutations of size s in this model have been studied by Rouzine et al. (2003). Here we discuss briefly the effects of deleterious mutations of various sizes, but leave detailed analysis for future work.
It is useful to separate the effects of deleterious mutations into their impact on the dynamics of the bulk of the distribution (and hence the mean fitness) and their effects on the establishment of new most-fit clones at the nose. In the bulk of the distribution, deleterious mutations come to a deterministic mutation–selection balance that alters the shape of the fitness distribution and reduces the mean fitness. This effect actually speeds up the evolution: if the deleterious mutations had no effect at the nose, their impact in reducing the mean fitness would increase the lead and thus make new establishments at the front occur faster. But deleterious mutations at the nose have the opposite effect: they slow down the growth of the most-fit populations and decrease the fitness of some of these individuals, reducing the rate at which new more-fit individuals establish.
In understanding these effects, it is useful to consider large-effect and small-effect deleterious mutations separately. First we consider deleterious mutations whose cost sd > s. When a deleterious mutation with occurs at the nose, that individual is no longer at the nose. Thus the deleterious mutations reduce the effective growth rate just at the nose. If is the mutation rate to deleterious mutations with , then the growth rates of subpopulations at the nose are simply reduced by . The effect of deleterious mutations on the mean fitness is also simple, because the mean fitness of the population is dominated by the largest subpopulation (which is exponentially larger than all others). Thus in considering the effect of the deleterious mutations on the mean fitness, we can focus on their impact in this subpopulation. This remains the largest subpopulation for ∼ generations, which for sd > s is larger than . Thus it comes to a deleterious mutation–selection balance while it is largest, since this balance is obtained in generations. This means that the deleterious mutations reduce the mean fitness by (up to small corrections due to the dynamics and the other subpopulations). This reduction in the mean fitness effectively increases the lead by , which increases the growth rates at the nose by the same amount. This cancels the effect of the deleterious mutations at the nose. Thus deleterious mutations with have very little net effect on v: they do not change the rate of new establishments at the nose, up to the small corrections noted above. This is not surprising—the deleterious mutants are all doomed, so roughly speaking their effect is simply to reduce the effective fitness of all individuals equally, which has no net effect on v. But they do increase the lead qs, which changes the shape of the fitness distribution.
For weakly deleterious mutations with , which occur at mutation rate , the effects are more complicated. In this case, the fact that an individual at the nose has a deleterious mutation does not make it substantially less likely to be the source of a new nose-extending mutation. Thus the effective growth rates at the nose are unaffected by deleterious mutations. However, some nose-extending mutations will occur in individuals with one or more deleterious mutations and hence will not necessarily extend the nose by s. Instead, they will sometimes have an effect s − sd, or s − 2sd, or less. We can estimate the strength of this effect by using a deterministic approximation for the deleterious mutation accumulation at the nose. When (or, roughly, when ), we find that on average, nose-extending mutations are burdened by a deleterious load of . Thus the effect of the deleterious mutations at the nose is to reduce the effective s by the amount , which is small compared to s. This will tend to slow the evolution. An analogous calculation applies when ; here the deleterious mutations have a larger effect, but still produce an average fitness cost only at most of order sd. The effect of the deleterious mutations on the bulk of the distribution is again to reduce the mean fitness of the population. The amount of this reduction, however, does not depend only on the most-fit subpopulation as before, because . Rather, these small-effect deleterious mutations accumulate throughout the collective-sweep time, qs/v ≈ ln(s/Ub)/s, in which a subpopulation grows from being the lead population to the dominant population. We expect this effect to be largest relative to the effects of these deleterious mutations on the dominant subpopulations when 1/sd is of order the collective-sweep time. This effect reduces the mean fitness by an amount at most of order . This again speeds the evolution and partially cancels the slowing effect at the nose. Thus deleterious mutations with affect v by increasing the effective lead by of order and reducing the effective s by ∼ (when ) or by of order sd (when is larger). These effects are all small.
To analyze in more detail the quantitative effects of deleterious mutations (even in the simplest single-beneficial-s model) is beyond the scope of this article. Note in particular that the analysis in this section is invalid when the deleterious mutation rate is large enough that the deterministic approximation for their behavior at the nose becomes incorrect. In this regime—on the border between Muller's ratchet and adaptive evolution—a more careful analysis is needed. We leave this discussion, which is essential for understanding the dependence of the rate of evolution on the mutation rate when mutation rates become large, for future work.
Our analysis involves a number of approximations. While we have analyzed their validity above and in the appendixes, we also used computer simulations to test our results. In this section, we describe these simulations and the comparisons to our results.
We started our computer simulations with a clonal population with a birth and death rate of 1 and a mutation rate of Ub. We arbitrarily defined this population to have fitness 0. We divided time into small increments. At each increment, we first calculated the average fitness and then produced births, deaths, and mutations with the appropriate probabilities. The birth rate of individuals at fitness y was set to be (with always small compared to unity), their death rate 1, and the mutation rate Ub. We then repeated this process to simulate the population dynamics, providing a full stochastic simulation of the simplest constant-s, beneficial-only model analyzed above. We recorded the mean fitness and lead as a function of time and, for each set of parameters, measured the average v and q once past the initial transient regime.
We carried out these simulations at a variety of different parameter values. The match between simulations and our theoretical results was good, provided the conditions for the validity of the concurrent-mutations regime obtained. Examples of these comparisons are shown in Figures 5 and 6. In Figure 5, we show the theoretical predictions for the average speed of adaptation (using the lowest-order iterative result for v presented in Equation 41) compared to simulation results as a function of N, Ub, and s. In Figure 6, we show similar comparisons for the average lead q (again using the lowest-order iterative result for our theoretical predictions). The agreement is good in both cases, although our theory slightly underestimates both v and q. This may be due to the effects of fluctuations in τq (described in appendix d) slightly increasing the mean v and q because of their nonlinear effects or to other factors arising from ln(s/Ub) not being sufficiently large for the asymptotic results to obtain to this accuracy.
DISTRIBUTIONS OF s, AND RELATIONSHIP TO CLONAL INTERFERENCE ANALYSES
The simple model we have analyzed assumes that all beneficial mutations confer the same advantage s. But in most natural situations different beneficial mutations will have different fitness effects. This does not change the basic dynamics of adaptation in large asexual populations: many beneficial mutations still occur before earlier ones have fixed and these can help or interfere with each other's fixation (Figure 1b). And the successful mutant lineages are likely to have had multiple beneficial mutations before they fix, while many other mutations will be wasted when other lineages outcompete them.
Thus far we have focused on how beneficial mutations are wasted because they occur in individuals who are not very fit (i.e., away from the nose) and are therefore handicapped by their poor genetic background. But when beneficial mutations have a variety of different effects, there is another way they can be wasted: small-effect mutations can be outcompeted by larger mutations that occur in the same or a similar genetic background. We refer to this latter process as “clonal interference.” As before, we use the term clonal interference to refer to this latter effect only (despite some broader definitions in the literature), consistent with the focus of recent work on the subject. This can occur only when not all mutations have the same fitness increment and is thus absent in the simple constant-s model.
Recent work by Gerrish and Lenski (1998) and others (Orr 2000; Gerrish 2001; Johnson and Barton 2002; Kim and Stephan 2003; Campos and De Oliveira 2004; Wilke 2004; Kim and Orr 2005) has taken the opposite approach to the multiple constant-s mutations approximation and focused instead on the effects of clonal interference, while ignoring multiple mutations. In this section, we first summarize the conclusions of such analyses, which assume all mutations occur on the same genetic background. We then consider the effects of including both clonal interference and multiple mutations. As we will argue, whenever the former plays a significant role, so does the latter.
The now-conventional clonal interference analysis considers how small-effect mutations can be outcompeted by larger mutations. Specifically, if a mutation A with fitness sA becomes established, one considers the probability that another mutation B, with effect sB > sA, will also become established before mutation A has fixed. If this happens, mutation B drives A to extinction and mutation A is thus wasted. Of course, it is also possible that mutation B is subsequently outcompeted by a still fitter mutation C, and so on. The key approximation is that the largest mutation that occurs and is not outcompeted by a still larger one fixes, becomes the new wild type, and the process then repeats. Additional mutations that might occur in a lineage that already has mutation A, B, or C are ignored. For any fixed population size, there is some selective advantage, sci, such that sufficiently large mutations, those with s > sci, are rare enough that they are unlikely to occur before some less-fit mutation arises and fixes. In the conventional clonal interference analysis, it is assumed that a mutation of size around sci will thereby fix before any others, and the process will then repeat. This is equivalent to successional-mutation behavior with a set of mutations each with the same strength, sci. Since sci increases with the population size, more mutations are wasted in larger populations, implying that v increases less than linearly with NUb.
Before discussing the problems with the basic successional-fixation assumption, we consider how the characteristic sci depends on N and on the distribution of selective advantages, ρ(s)ds. Because only beneficial mutations with substantial s matter for large N, the total Ub itself is not important. It is more convenient to use the mutation rate per generation for mutations in a range ds about s:(49)We assume that large-effect beneficial mutations are typically much less common than small-effect ones, so that μ(s) is small and decreases rapidly with s. Since μ(s) = Ubρ(s) is dimensionless, it is convenient to define Λ(s) by(50)Note that increases with s. Mutations with effect of order s occur at an overall rate of order sμ(s) = se−Λ(s), so Λ(s) roughly plays the role that ln(s/Ub) does in the single-s case.
The basic clonal interference analysis is simple: in the time that a mutation of size sA will take to fix, , some mutation of larger size s will have time to occur and become established as long as the total establishment rate for mutations larger than sA is sufficiently large:(51)This will no longer be true above some critical sci, where . We can estimate this sci by noting that since μ(s) decreases rapidly with s, . We find(52)Using the definition of Λ, we see that sci(N) is the value of s at which in the whole population there is of order one mutation per generation. Further, because μ(s) = Ubρ(s), we see that sci depends only on the product NUb, with the functional form determined by ρ(s). In the successional clonal interference analysis approximation, the speed of evolution is assumed to be the size of these mutations, sci, times the rate at which mutations of order this effect occur, sciμ(sci), times the probability that they become established, sci. This yields(53)where C is a factor of order unity that is not really obtainable from clonal interference analysis, as it depends on the details of further approximations. (Note that the details of how we define fixation do not make much difference in the clonal interference result. We have also ignored other factors inside logarithms, since .) At this point we should note that various potential improvements are possible. In particular, it is not at all clear why the establishment time rather than the fixation time should be used to obtain the accumulation rate of the sci mutations. As we shall see below, if the latter rather ad hoc assumption is made instead, the clonal interference analysis gives closer to the correct results for certain distributions: those with long tails in ρ(s). But with or without such improvements, some of the predictions of clonal interference analysis are qualitatively wrong—in particular, the prediction that as the overall beneficial mutation rate increases, the typical size of the mutations that fix (predicted to be sci) also increases. As we shall see, the opposite is true.
The above clonal interference analysis makes a crucial approximation that is essentially never valid: that double mutants can be ignored even when mutations are common enough that they often interfere. This is manifest in the assumption that the important mutations occur only in the majority (wild-type) population. The basic problem is that even if a more-fit mutation B occurs before an earlier but less-fit mutation A fixes, A may still survive. An individual with A can get another mutation D such that the A–D double mutant is fitter than B. If this happens, mutation A (along with D) can fix after all. Indeed, such events should be expected: any population large enough for clonal interference to matter is also large enough for double mutants to routinely appear even for s ∼ sci. This is because clonal interference can affect the fixation of a mutation of size s only when the establishment rate of mutations stronger than s, which is at least , is large compared to the rate at which the mutation of size s fixes, . But when this occurs, we have . Thus, from our analysis of the single-s model, whenever clonal interference occurs, multiple mutations also play a role.
The single-s model, in contrast, is unrealistic because it explicitly excludes competition between mutations of different effects. Thus the conclusions from this model and the clonal interference analysis are each only part of the story. In the remainder of this section, we outline the behavior for more general distributions of beneficial mutations, taking into account both clonal interference and multiple mutations. Fortunately, as we shall see, for many forms of μ(s), the single-s approximation can implicitly account surprisingly well for the effects of clonal interference. Detailed analysis will be published elsewhere.
Let us first consider starting from a clonal population (although this is an oversimplification that misses important aspects of the dynamics; see below). Depending on N and Ub, various different mutants will arise, as well as double mutants, etc. One of these will be the fittest mutant that is established in the wild-type population before any other mutation or combination of mutations fixes. All the other mutations that have already occurred will be driven to extinction and thus do not matter for the long-term evolution. For a given N, Ub, and ρ(s), there is a typical fitness effect (call this ) of the beneficial mutations that create—singly or in combination—this fittest mutant. We call mutations of roughly this magnitude predominant mutations and define—crudely at this point— as the mutation rate to these mutations. Clonal-interference-like competition determines the predominant range of mutations. Unfortunately, however, we cannot simply lift the definition of from clonal interference theory. Except at very short times, the population will not be monoclonal but will include various single and multiple mutants with a distribution of overall fitnesses. This means that is determined by a delicate balance between clonal interference and multiple mutation effects. Given an , however, the predominant mutations accumulate via a process similar to that described by our analysis of the constant-s model, with population size N and the effective parameters , and .
Why should there be a predominant range of s? The basic argument is simple. Mutations significantly smaller than occur frequently. But, by definition, these mutations are routinely outcompeted by predominant mutants. Thus these mutations do not interfere with the accumulation of the predominant mutants. In contrast, larger-than- mutations do interfere with others when they occur. But, by definition, these must be rare enough that it is unlikely that such a mutation will arise in the time it takes a predominant mutant—or a combination of predominant mutations—to fix (else the larger mutation would be the predominant mutant). Thus the population will primarily evolve via the accumulation of mutations with s in some range around . Our previous analysis does not predict , but given a value of it determines how these mutations accumulate (see below for more details). This is a slight oversimplification, as mutations of both smaller effect and larger effect than will play some role. These considerations affect the appropriate definition of and the range of s around that is important.
What we must now address is the crucial fact that (and ) depend on N and Ub. As we increase N or Ub, more mutations occur before others fix: this suggests will thereby change. Clonal interference analyses consider part of this process and predict that the analog of (sci) increases slowly with both N and Ub (Gerrish and Lenski 1998; Wilke 2004). But these approximations oversuppress smaller mutations by ignoring multiple mutations, which are more likely to involve the common smaller mutations. Thus we expect that should increase even more slowly with N and Ub than clonal interference models suggest. Nevertheless, even a slow increase in could be important, since in the single-s model, v increases with s2 but only increases slowly with N and Ub. As we now show, the form of ρ(s) qualitatively affects the behavior.
In the extreme case in which ρ(s) decreases very slowly with s [ or slower], the largest mutation that can typically occur and establish in a given time always dominates the cumulative evolution up until that time. Thus a predominant does not even exist and neither our analysis nor clonal interference describes the dynamics: they are controlled by successional fixations—but with no steady-state speed—no matter how large the population. We do not discuss this seemingly unlikely situation further.
Whenever ρ(s) falls off faster than , the basic single-s behavior obtains, with a narrow range of s (roughly a factor of two or less) around some predominant , with the effective mutation rate crudely being that for mutations in this range. But even though one could then simply plug the appropriate and into our earlier expressions for the speed, v, the single-s forms for the dependence on N and Ub may not be accurate, because and themselves depend on N and Ub. There are two possibilities. The first is that and depend weakly enough on N and Ub that our expressions are roughly accurate. Another possibility is that the evolution is dominated by larger and larger mutations as the population size increases, as found in the clonal interference analysis. Again mutations in some restricted range will control the behavior (and some degree of multiple mutations will still be involved), but will increase markedly with N. We shall see that both these behaviors can occur, depending on the form of the distribution of mutations ρ(s)ds.
A simple approximation that might be expected to be valid if a sufficiently narrow range of s dominates is to ignore all the mutations except those in some narrow range about s, compute the evolution speed v(s) from the single-s analysis, and then maximize this over s to obtain the predominant s, , and an approximation for the actual speed,(54)which defines . In the above expression, we define v(s) to be the speed of evolution to mutations with effect of order s, as calculated from our single-s analysis with the appropriate Ub being the total mutation rate for these mutations. We call this approximation the predominant-s approximation, as it ignores the question of how wide a range of s is important. We can then make a conservative check of our assumption that a narrow range of s dominates by computing how quickly v(s) falls off away from , because mutations at other s cannot increase the actual velocity by more than their v(s).
For concreteness, we consider a class of distributions μ(s) parameterized by three quantities: a characteristic selective advantage, σ, a parameter ℓ that controls the overall mutation rate, Ub ∝ e−ℓ, and a parameter β that characterizes the shape of the distribution of rare large mutations. We thus write(55)For convenience we use the shorthand notation(56)We will see that the behavior depends qualitatively on whether β is larger or smaller than 1. For β > 1, the distribution falls off faster than exponentially, and we refer to this as a “short-tailed” μ(s). The exponential case is exactly marginal. For β < 1, the distribution falls off more slowly than exponentially. We refer to this as the “long-tailed” μ(s) case.
We begin by considering the case of β > 1, that is, a distribution that falls off at least exponentially. The behavior is simplest when the population size is large enough that 2L/Λ(s) is substantially greater than unity. In this regime we have from Equation 41 that(57)where we have used Λ(s) in place of ln(s/Ub) [valid because the total mutation rate to mutations with effect of order s is sμ(s)]. This v(s) has a maximum at given by(58)Plugging in , we find(59)and thus in the predominant-s approximation we have(60)with the coefficient Cβ = (β − 1)2−2/β/β2, valid for β > 1. We have used the large-q single-s results in making these calculations. We can check the consistency of this by noting that the value of q for the predominant mutations is , which yields(61)This is large when 2L/ℓ is large, unless β − 1 is small (i.e., the tail is becoming long), so our results are indeed consistent.
Note that our result for is roughly independent of N in this regime, but (in contrast to clonal interference analysis) decreases as the overall beneficial mutation rate increases (i.e., as ℓ decreases). In other words, does not depend strongly on N, but does decrease as Ub increases. This makes sense: as Ub grows, multiple small mutations become more important compared to single larger mutations. Because of this, the dependence of v on N is very similar to our single-s approximation, but the dependence on the mutation rate is weaker.
The behavior for β > 1 can also be analyzed when L is not so large. As L decreases, the predominant s decreases—i.e., it begins to depend on N. The resulting expressions are more complicated, but can be computed from Equation 41 in a similar way. However, they are of questionable validity, since only some of the significant s will be in the multiple-mutation regime, while others will be in the crossover regime of , so our use of the large-q results becomes inconsistent. As we have seen in a previous section, this crossover is complicated even for the single-s model; it will be even more so with a distribution of s.
For distributions that fall off more slowly than a simple exponential—i.e., β < 1—the behavior is rather different. This is apparent even in the crude predominant-s approximation. Again, we begin by considering the simpler large 2L/ℓ limit. We have(62)with Λ(s) = ℓ + (s/σ)β, which we maximize to find(63)Plugging this into μ(s) = e−Λ(s), we find the corresponding effective mutation rate(64)and the predominant s approximation(65)with coefficient Aβ = β(2 − 2β)2/β−2(2 − β)1−2/β. In this case, we see that v grows faster than linearly with ln N. Surprisingly, the dependence on the mutation rate in this regime is negligible: Ub determines only how large N has to be to be in this regime. The smaller the mutation rate, the larger the N needed. But in contrast to the short-tail case, here(66)is not large, so that even for very large N, the important multiple mutants still involve only of the predominant mutations. The fact that q never becomes particularly large for long-tailed μ(s) is because in this case increases substantially with N: in the short-tailed case, many small mutations contribute, while in the long-tailed case, fewer larger mutations are involved. But we must be careful with the above results for the long-tailed case, as they are not valid if the inferred q < 2: below this the crossover from successional- to concurrent-mutations behavior will apply, and our use of Equation 62 becomes inconsistent. We need to distinguish two cases.
If , q > 2 and the above results apply. The corresponding effective mutation rate decreases with a power of 1/N less than unity, so that the total mutation supply rate for the predominant mutations, , grows with N as N(3−2β)/(2−β). (Of course, many of these are wasted as multiple mutants outcompete the single mutants and control the dynamics, as described by our single-s theory.)
If , then the above analysis would give q < 2 and , which indicates a breakdown of the approximations. In this case q sticks at 2, and the dynamics are basically successional, with the predominant mutants being those for which the total rate . This means that , and we expect(67)Note that the coefficient coincides with the earlier expression at β = . The steady state is at the upper end of the crossover between the successional- and multiple-mutation behavior as discussed in the Evolution at moderate N section.
For , the clonal interference-only approximation agrees with the predominant-s approximation, as the total mutation rate to the predominant mutants is of order unity so that . In contrast, for the intermediate case with , clonal interference analysis yields sci ≈ σL1/β. This is still the correct behavior, but the numerical coefficient is wrong: as noted above, the total mutation rate for the predominant mutants grows as a power of N, in contrast to the clonal interference approximation in which it assumed to be independent of N. For the speed of evolution, naive application of the clonal interference analysis gives v ∼ ∼ L2/β, which is not even the correct scaling with L. But if, instead, the fixation rate rather than the establishment rate is used to give an improved (though it is not a priori clear why this should improve the result) clonal interference estimate of v, the correct scaling with L can be obtained.
At this point, it is not clear how good the predominant-s approximation is for the long-tailed distributions or how wide a range of s around its predominant value is important. A more sophisticated analysis is needed for this, as well as for understanding the crossover from the successional to the large-q regime analyzed above; these are topics for future work.
The width of the important range of s around s̃:
We now turn to a discussion of the basic assumption of the predominant-s approximation: that a narrow range of s around dominates the evolution.
In the successional-mutations regime, the speed of evolution is . This means that s of order σ dominates [as long as μ(s) falls off faster than 1/s3]. That is, v(s) falls off quickly enough away from its maximum that a range of s within a factor of 2 or so of the typical value dominates the evolution. In the multiple-mutations regime, the maximization of the single-s speed v(s) over s gives a predominant ≫ σ, but no direct information on the range of s that contributes. To estimate this range, we look at how quickly v(s) falls off as a function of . A natural estimate is the range over which v(s) is not lower than by more than, say, a factor of 2. Using these criteria, the width of the range is comparable to itself: that is, mutations with effects between and matter. This confirms our assertion that the single-s model gives at least a good qualitative picture of the dynamics. Since all the important mutations are of order , “leapfrogging” (by which, for example, a double mutant gets a mutation that makes it more fit than an existing quintuple mutant) does not have a large effect on the evolution. We can thus indeed consider the basic dynamics to be the accumulation of mutations of roughly size according to the single-s description given above.
However, our calculation of the range of that matters calls into question the predominant-s approximation: Why should the actual v be as we have defined it thus far rather than, e.g., v(s) averaged (or some other weighted integration) over s? A more sophisticated analysis, which will be described elsewhere, shows that for short-tailed distributions (β > 1), both and v are given correctly by the predominant-s approximation in the large-L/ℓ limit—up to only differing factors inside logarithms and other small corrections. But the range of s that significantly affects v is much smaller than that guessed from the predominant-s approximation. This should perhaps not be surprising, as the predominant-s approximation assumes that all s contribute to v as if different-sized mutations did not interfere. But interference will in fact tend to suppress the contribution to v from s away from . We find that for short-tailed distributions, in fact only of order are important. In terms of the mutation rate to mutations with , this range has width . That this difference does not invalidate the predominant-s approximation result for v can be understood by considering the weak dependence of v on the mutation rate in the single-s model. As v depends only logarithmically on Ub, replacing Ub either by an effective that includes a substantial range around or by one that includes only a narrow range will alter factors only inside the logarithms and thus have little effect on the inferred v. Since the fuller analysis finds that an even narrower range around matters, it strengthens our contention that there is a predominant s (albeit one that depends on Ub) and that the full dynamics are very similar to those of the single-s case analyzed in detail in this article. The exception to this is the intermediate-N regime in which the crossover from successive to multiple mutations occurs and the effective q < 2 or so: we do not discuss this complicated crossover regime further here, although it may be relevant in many experimental situations.
We have seen that the predominant-s approximation does well for the primary quantities of interest, and v, although it overestimates the range of s that plays a role. In contrast, the clonal-interference-only analysis yields the incorrect behavior for short-tailed distributions. For the model distributions, Λ(s) = ℓ + (s/σ)β, the clonal interference analysis yields(68)
For the short-tail case, this is much larger than the predominant value, . Indeed it is qualitatively wrong: sci increases with increasing Ub, while decreases. Using sci instead of leads to incorrect predictions of v; in particular, clonal interference predicts that v grows only sublinearly with ln N. This problem stems from the fact that clonal interference analyses have the wrong basic picture of the dynamics. The evolution is not in fact dominated by the rare very large mutations that occur only once per generation in the full population, as the clonal interference approximation implicitly assumes. Rather, the evolution is actually controlled by multiple mutations of smaller (though still larger than average) fitness that occur frequently even in the much smaller subpopulations that exist in the nose of the fitness distribution of the steady-state evolving population. Because the multiple-mutation effects depend on there being sufficiently large rates for the predominant mutations, increasing the overall mutation rate allows multiple smaller mutations to beat larger ones. Thus increasing Ub results in decreasing —in contrast to the increase of sci with Ub.
A simple example:
A concrete (albeit artificial) example is useful to illustrate the points made above. We consider a simple model with three classes of mutations, each with a single s: weak mutations with a small ss, intermediate ones with a medium sm, and strong mutations with a large sl; each class has its own mutation rate. Specifically, we consider ss = 10−3, sm = 10−2, and sl = 10−1, with mutation rates Us = 9 × 10−6, Um = 4 × 10−6, and Ul = 5 × 10−10, crudely approximating an exponential distribution of beneficial mutations (with, in terms of the family of distributions discussed above, β = 1, σ = 10 × 10−2, ℓ ≈ 7).
For small population sizes, the successional regime obtains and(69)which is dominated by the medium mutations. As N increases, we expect multiple mutations to start to play a role when NUm ∼ 1/ln(sm/Um) ≈ , corresponding to crossover out of the successional-fixations regime for N ∼ 3 × 104.
To understand the behavior for larger N, we first analyze the three types of mutations separately, similar in spirit to the predominant-s approximation. That is, we consider three submodels, each of which has only one of the three types of mutation. The corresponding rates of evolution, vs, vm, and vl must all be less than vtot, that of the full model, because the full model has more beneficial mutations than any of the three submodels. Conversely, we expect vtot ≤ vs + vm + vl because, at best, the different mutations can accumulate independently; in practice, they will tend to interfere (although multiple mutants with combinations of the different types can matter and contribute to the actual speed). Each of the three submodels has only one type of mutation, so our single-s results can be used directly to obtain vs, vm, and vl.
For a population of size N = 105—just into the multiple-mutations regime—we find vs = 3.5 × 10−7, vm = 1.5 × 10−5, and vl = 5 × 10−7. The leads of the corresponding fitness distributions—the number of multiple mutants above the mean that exist at one time—are qs = 2.7, qm = 2.2, and ql = 1. Thus the small and medium mutations accumulate primarily as double and triple mutants, while the large mutations (alone) would be in the successional-mutations regime. For this moderate-size population, the mutations with effect sm are the predominant mutants. They clearly dominate the full model, since vtot will be in the very narrow range between vm and vm + vs + vl. Although the small mutations are common, they do not matter because even triple-small mutants—as occur in the small-only model—will be routinely outcompeted by single medium mutations. The medium mutations occur frequently in the fixation time of the triple-small mutants and thus routinely leapfrog them. The small mutants never interfere with medium mutations, and those that fix do so only because they happen to be linked to medium mutants. The large mutations, in contrast, do interfere with the medium mutations, but occur so rarely that they are not important for the overall evolution rate. In this example, a few hundred medium mutations fix for each large mutation that establishes, so almost all medium mutations fix without being affected by a large mutation. Thus the accumulation of mutations is very well approximated by the process that our single-s analysis describes, provided we choose and .
As the population size is increased, vl will increase faster than vm or vs because it is not yet in the regime with logarithmic N dependence. For N = 106, vs = 5 × 10−7, vm = 2 × 10−5, and vl = 5 × 10−6. The medium mutations still predominate, but less strongly than before. By N = 107, we have vs = 6 × 10−7, vm = 3 × 10−5, and vl = 5 × 10−5, so the large mutations begin to dominate. For larger N, they will do so even more strongly. This shows how increases with N. With this discrete ρ(s), changes quite rapidly in a small range of ln N, but for a continuous fitness distribution the increase will be smooth (of course, continuous distributions present additional complications involving the proper weighting of mutations near ).
We could also apply clonal interference analysis to this three-class model. From these analyses, for a beneficial mutation to fix, it must establish and then not be interfered with by a more-fit mutation before it fixes. The probability that a mutation of size s will be interfered with is . Thus the putative distribution of beneficial mutations that fix will be ρF(s) = Kse−λ(s)ρ(s), where K is a normalizing constant. The average effect of a fixed beneficial mutation—effectively sci—would be the mean, 〈s〉F, of this ρF(s). These mutations arise at average rate 〈k〉F = NUbPfix, where Pfix is the average probability of fixation, . Clonal interference analysis yields v = 〈s〉F〈k〉F. For our three-class example, with N = 105, this gives vtot ≈ 4 × 10−5, ∼3 times higher than the maximum possible as calculated from vtot = vs + vm + vl. For N = 106, clonal interference predicts vtot ≈ 4 × 10−4, ∼20 times too high. The problem is easy to diagnose. For both values of N, the clonal interference theory correctly predicts 〈s〉F ≈ sm. However, implicit in the calculation of 〈k〉F is the incorrect assumption that these medium mutations accumulate singly. Conversely, the predominant mutation approach is to choose sm as the single value of s and then analyze how the multiple-mutation process sets the rate at which this class of mutations accumulate.
Beneficial mutations are often assumed to be rare, and adaptation therefore to be mutation limited. This is the basis for the picture of successional selective sweeps and the conclusion that mutations arise and fix at a rate proportional to NUbs. This picture of successional sweeps underlies the strong-selection weak-mutation assumption that is essential to many conclusions in population genetics and evolutionary theory. This assumption is likely to be correct for the evolution of some strongly selected characters in complex multicellular organisms. But most unicellular organisms and viruses tend to live at much larger population sizes and can have larger mutation rates. For such populations, much of one's intuition from the rare-mutations picture will often be wrong. This makes it important to go beyond the successional-mutations regime and to develop an understanding of evolutionary dynamics when beneficial mutations are common.
This is a very broad subject. In this article, we have focused on the concurrent-mutations regime in which there are strong selection and strong mutation. By strong mutation, we mean that the total beneficial mutation production rate NUb is sufficiently large that the time to establish a mutant population is less that the time it will take to sweep to fixation. As the establishment time is 1/(NUbs) and the sweep time is , the condition to be in the concurrent mutations regime is , so that multiple beneficial mutations are present in the population and tend to interfere. By strong selection, we mean both and . The former condition is what is commonly meant by strong selection and is required to ensure that selection is strong compared to drift except when subpopulations are rare. The latter constraint makes the analysis simpler, because it ensures that only one population at a time needs to be treated stochastically, but is not essential for the general picture.
The concurrent-mutations regime that we analyze is likely to be quite common in nature. Even if there are only 10 or so beneficial point mutations available to a population that has a per base pair mutation rate of order 10−9, this gives Ub ∼ 10−8. To have , we therefore need only population sizes of order 107 (1/ln[Ns] will typically be ∼ for any reasonable values of s in such large populations). In other words, if there are even a few mutations of effect available, a population as small as 107 individuals will experience the multiple-concurrent-mutation effects. These sizes are well within normal ranges for many populations, including, for example, Escherichia coli in a single human gut, cells in an evolving cancer, pathogens within a single host, and many others. Moreover, this is a very conservative estimate. Viral and certain bacterial populations, or mutator strains in any organism, often have much higher overall mutation rates. Organisms with more beneficial mutations available will also have much larger Ub. In recent experiments in Saccharomyces cerevisiae adapting to low glucose, we have inferred a beneficial mutation rate of Ub = 10−5.5 in nonmutator strains and an order of magnitude higher in mutators (Desai et al. 2007). Such values are not atypical (Joseph and Hall 2004). For these values of Ub and s of order 1% or a fraction of 1%, a mutator population of N ∼ 107 will have q ∼ 4, so that quadruple mutants will be present and sweep collectively (for nonmutators, q ∼ 3). With these parameters, each factor of 10 increase in N will increase q by ∼1. In general, we see that the concurrent-mutations regime is surely relevant for many microbial populations.
Within the concurrent-mutations regime, we have explored how a population accumulates beneficial mutations and maintains variation in fitness. The fundamental theorem of natural selection states that the rate of increase in the mean fitness of a population equals the variance in fitness (Fisher 1930). This remains true. Our work demonstrates how the variance is itself determined: how fitness variation accumulates while it is being selected on. The key here is the balance between selection narrowing the fitness distribution and mutation broadening it. This is an unusual type of mutation–selection balance, very different from the deleterious case. Only mutations at the nose of the distribution matter. Others inherit a less good genetic background and do not contribute to the long-term evolution of the population: they are destined to be outcompeted by new mutations at the nose. The dynamics at the nose, where subpopulation sizes are small, dominate the behavior. This means that the natural measure of the width of the distribution is the lead, not the variance—in contrast to conventional treatments. It also means that random drift and finite N effects are crucial, even for arbitrarily large N, as long as there are more than a few beneficial mutations to be acquired. Thus for any treatment of evolution in fitness “landscapes,” these effects need to be taken into account whenever the population is not localized around a fitness peak: in contrast to quasi-species equilibria near fitness peaks, deterministic approximations give nonsense.
By matching the speed of advance of the nose with the speed of advance of the bulk of the distribution, we have shown that the lead depends logarithmically on N and Ub according to the formula(70)This leads to a speed of evolution that is also logarithmic in N and Ub,(71)
Our work extends and complements earlier work on the concurrent-mutations regime. Kessler et al. (1997) and Ridgway et al. (1998) studied a model like ours, although their initial work did not properly account for all stochastic effects. Recently they have developed a moment-based approach that provides results qualitatively similar to ours in certain regimes (D. Kessler and H. Levine, unpublished results). This is a potentially useful technique, although, as discussed in appendix a, it quickly becomes unwieldy as more moments need to be kept, and numerical analysis is required.
Rouzine et al. (2003) also studied a model similar to ours in the context of human immunodeficiency virus evolution. Their analysis also involves a separation between deterministic and stochastic behavior, but treats the stochasticity at the nose in a different and less explicit manner. To couple this to the deterministic results, Rouzine et al. (2003) appear to require a smoothness in the fitness distribution that would obtain only when it is broad. Thus their analysis is strictly valid only at what we would call very high speeds: . But, because they treat only one population stochastically at a time, their analysis also requires , so their results are valid only at enormous population sizes [and very large ]. This regime is likely to be relevant for certain viral populations, which was their main focus. Nevertheless, the results of Rouzine et al. (2003) are similar to ours, in that they involve logarithms of Ns and in similar ways (though they do differ substantially—in the regimes we have considered their results lead to errors typically ranging from ±50 to 250%). This is unsurprising, since the simple beneficial mutation–selection balance arguments (in our heuristic analysis section) apply to the very fast regime of Rouzine et al. (2003) as well and lead generally to logarithms of Ns and . Further analysis shows, if some algebraic errors are corrected in their work, and the large Ns and s/Ub asymptotics are worked out, that our result for v can be recovered up to somewhat different factors inside large logarithms (I. Rouzine, personal communication).
Various studies have been carried out on clonal interference—the other effect that occurs when there are concurrent mutations (i.e., in the strong-selection strong-mutation regime) (Gerrish and Lenski 1998; Orr 2000; Gerrish 2001; Johnson and Barton 2002; Kim and Stephan 2003; Campos and De Oliveira 2004; Wilke 2004). We have discussed the relationship between this work and ours and analyzed a model with a distribution of beneficial mutations that includes both clonal interference and multiple-mutation effects. Kim and Orr (2005) have also analyzed some of the interplay between these effects. Clonal interference analysis by itself makes qualitatively similar predictions to our work about the rate of accumulation of beneficial mutations. Both predict that v grows much less than linearly in N and Ub (as do the analyses of D. Kessler and H. Levine, unpublished results, and Rouzine et al. 2003), although the quantitative predictions differ. The major qualitative differences are in the mechanisms by which the evolution takes place. In clonal interference analysis large mutations that occur in individuals that have roughly the mean fitness—i.e., in the majority subpopulation—dominate the evolution. Thus one would expect to see strong selective sweeps and a population that is typically either nearly clonal or in the midst of such a sweep (except occasionally when a smaller mutation becomes transiently very common before being outcompeted by a larger one). By contrast, except when there is a long tail to the distribution of s, we have shown that the evolution is dominated by multiple mutations of intermediate effect, so the selective sweeps are much less pronounced and the population always maintains substantial variation in fitness. And we have shown that even when the distribution of s does have a long tail, some of the quantitative predictions for the speed of the evolution are different from clonal interference predictions. We find that the mutations that dominate the evolution in the concurrent-mutations regime have effects in a narrow range around some predominant value, . The simple single-s model is thus surprisingly good, provided we use and Ub equal to the mutation rate to beneficial mutations of this magnitude.
Over the past few years, much experimental evidence has accumulated that supports the prediction that v grows less than linearly in N and Ub (de Visser et al. 1999; Miralles et al. 1999, 2000; Colegrave 2002; de Visser and Rozen 2005). This has often been interpreted as support for the clonal interference picture. However, the experimental data on the quantitative details of the dependence of v on N and Ub cannot distinguish between clonal interference analysis and our results. Thus these experiments also support our theory.
We have recently (in collaboration with Andrew Murray) conducted experiments on asexual evolution of yeast in low glucose. For a range of different N and Ub we measured the distributions of fitnesses within the evolving populations and the dynamics (v(t)) by which the fitness increased (Desai et al. 2007). Since, unlike earlier work, these experiments measured the widths of the fitness distributions and the strengths of selective sweeps, we were able to distinguish between our analysis and clonal interference acting alone. The experimental data support the multiple-mutation theory, with both v and the leads of the fitness distributions depending on N and Ub consistent with our predictions. Clonal interference analysis, on the other hand, would predict that populations maintain less variation in fitness and that this variation would not scale with N and Ub as we predict. We also measured how the populations increased in fitness over time, finding smooth increases suggestive of multiple mutations of intermediate size fixing together. This was again consistent with our theory and inconsistent with clonal interference alone, which would suggest that rare larger mutations dominate the evolution. Combining all these data, we found that clonal interference was ruled out unless several parameters were finely tuned. Thus in the only experimental test able to distinguish the two effects, multiple mutation effects explain the data better than clonal interference alone. This represents only one set of experiments in one organism in one selective condition, so it is quite possible that in other circumstances the reverse will be true. Yet, even if clonal interference is found to better characterize the dynamics in some situations, we have shown that to understand this properly one needs to analyze the interplay between this and multiple beneficial mutations on the same genome.
Despite being consistent with one experimental test, the model we have analyzed surely has many shortcomings. We have analyzed one of the simplest possible situations for positive selection. Violations of certain simplifying assumptions, such as neglecting deleterious mutations and assuming a single effect s of beneficial mutations, may well, as we have argued, have relatively minor effects beyond modifying the effective parameters Ub and s of the model. Furthermore, the neglect of interactions between effects of mutations (epistasis) may not invalidate the overall results. The key assumption is that the distribution of the magnitudes of available beneficial mutations is roughly independent of the genetic background even though the actual set of these mutations varies. That is, after each uphill-fitness step is taken, the distribution of possible next steps is similar, although they may now be in different “directions.”
However, breakdown of some of our assumptions will surely be crucial. For example, certain nonmultiplicative (epistatic) effects of beneficial mutations, as well as frequency-dependent selection, can lead to very different behavior. But our results should serve as a null model, useful in forming baseline predictions. Departures from the main results—especially the scalings with population size and mutation rates—indicate the presence of one or more complicating factors.
Even within the context of our simple model, however, many important questions remain. One of these is the expected genetic variation. We have calculated the expected variation in fitness, but individuals with the same fitness will often have different sets of beneficial mutations. Thus the true genetic diversity at the positively selected sites can be substantially greater than the variation in fitness. Although sometimes the first new mutant to establish will dominate the lead population, typically around q different beneficial mutations will occur and contribute to extending the nose during one establishment. Subsequent mutations that further extend the nose will occur at random among these different backgrounds, thus changing (and typically reducing) this diversity, even as the diversity of the new mutations is created. Eventually particular beneficial mutations do sweep, but these sweeps are not necessarily uniform. Instead, frequencies typically go up and down depending on which backgrounds future mutations occur in. Understanding this diversity is important if one is to look for the signature of this type of selection in sequence data. It is also important to understand the potential benefits of sex, as we discuss below.
In addition to the diversity at the positively selected sites, we also want to understand the expected patterns of variation at linked neutral and deleterious sites. These will have a very different character than in neutral evolution or in the successional-mutations picture of positive selection and may also help us detect concurrent-mutations evolution in sequence data. The neutral, deleterious, and beneficial diversity is also important in understanding the role of epistasis. If potential beneficial mutations have epistatic interactions with other mutations, the typical variation in the presence of these other mutations is crucial.
Another important question is the effect of sex or recombination in a population in the concurrent-mutations regime. According to the Fisher–Muller hypothesis, sex should reduce interference effects and hence prevent the wasting of beneficial mutations. This allows sexual populations to accumulate beneficial mutations faster than asexual ones. Crow and Kimura (1965), Bodmer (1970), and Maynard Smith (1971) attempted to calculate the strength of this effect by comparing the v in an asexual population to the v in a population with free recombination. They defined the advantage of sex to be the difference between these quantities. However, their calculation of the asexual v assumed that only two beneficial mutations were possible—thus ignoring triple and higher mutants and not properly accounting for the competing effects of mutations and selection. With our calculation of the asexual v, however, one can make this comparison. In the completely free recombination case, all beneficial mutations behave independently: there is no interference between them or collective behavior among them. Thus with free recombination, vfr = NUbs2, as in the successional-mutations regime. The difference between our calculated asexual v and vfr thus predicts a potentially huge Fisher–Muller advantage to sex, which is zero in small populations and grows rapidly as N or Ub increases.
However, the above analysis is not directly applicable to the evolution of sex, since sex and completely free recombination are certainly not synonymous. Rather, sex may occur only occasionally or recombination might be infrequent, so that linkage persists for some time. An interesting situation is when sex and recombination are relatively rare. We want to understand whether or not a small amount of sex in an otherwise asexually evolving population would be advantageous (and hence be likely to become more common). To do so within the simplest model for the asexual evolution, we must first calculate the true genetic diversity among beneficial mutations within all the subpopulations at different fitnesses. Given this, we can then calculate the probability that sex between any two individuals will produce more-fit offspring.
This is a subtle question, because the average effect of sex on the variance in fitness or in the tendency to bring together good mutations more than it breaks them up is largely irrelevant. Rather, what is important is the rate at which recombination generates (or eliminates) anomalously fit individuals—that is, its effect on the nose. Sex will tend to break up beneficial mutations at the nose and hence tend to destroy some of the most-fit individuals. At the same time, however, it will occasionally mix two less-fit individuals in just the right way to create an offspring that is more fit than the current nose. It is the competition between these two effects that determines the advantage of sex. Even if sex on average tends to increase the variance in fitness, this will not increase the speed of evolution in the long term if it does not also extend the nose. Rather, the increased variance from sex will be balanced by the actions of selection and mutation (in the end, the mean fitness cannot advance any faster than the nose), and the rate of adaptation will be largely unchanged. On the other hand, if sex does extend the nose it will tend to speed up the evolution even if it has little effect on the variance. In this case, these occasional sex-driven expansions of the nose would act like extra mutations, which modify the mutation–selection balance and cause an increase in the steady-state variance via increasing the lead—even though sex has no direct effect on the variance.
In recent years, Otto and Barton have made substantial progress in understanding the effects of sex, short of completely free recombination, in the Fisher–Muller picture (Barton 1995; Otto and Barton 1997, 2001; Barton and Otto 2005). This work takes the Hill–Robertson perspective and does not include the full dynamics of the asexual population that we have worked out here. As far as we are aware, it is not clear whether the effect of sex at the nose within our calculated population structure is the same as the effect of sex in Otto and Barton's analysis. Future work is needed to unify these perspectives and understand the effects of sex even within the simplest models.
To summarize our work, we have explored evolutionary dynamics when beneficial mutations are common and there are many present concurrently. We have laid out an analytical and conceptual framework for understanding how asexual populations accumulate beneficial mutations—the dynamics of adaptation in this extremely basic situation. Using this framework, we have demonstrated that the rate at which a population accumulates beneficial mutations does increase only slowly with population size or mutation rate beyond a certain point. Although we have focused on the effects of multiple mutations, we have also analyzed the interplay between this and clonal interference between mutations of different strengths. Our results have implications for comparing evolution between different populations and for designing experiments to investigate various aspects of evolution in the laboratory. Statistical tests that can distinguish, on the basis of sequence data, between various scenarios for ongoing evolution are needed: our results provide a step in this direction. More generally, our results provide a framework for starting to address the effects of sex, of mutators, and of epistatic interactions in large populations.
APPENDIX A: DETERMINISTIC AND MOMENT-BASED APPROACHES
There are a variety of other possible approaches to studying the problem we have analyzed. In this appendix, we briefly discuss two of these: deterministic approximations and moment-based approaches. Both of these methods start by considering the distribution of fitnesses within the population as some function w(x, t), which describes the number of individuals at fitness x at time t. As long as s is small, w can be treated as continuous: this is equivalent to the conventional “diffusion approximation.” The forces of mutation, selection, and random drift then lead to a stochastic differential equation that describes the time evolution of this distribution w(x, t),(A1)where is the population mean fitness and ξ is a Gaussian random term but with subtle correlations needed to ensure that the fluctuations do not change the total population size . Studying this equation can then lead to predictions of the speed of evolution, maintenance of variation, and other interesting quantities.
The simplest possible approach is to neglect genetic drift and attempt an “infinite-N” solution to the problem. This deterministic approach is extremely useful in many situations, including in understanding deleterious mutation–selection balance. However, when considering beneficial mutations, it is essential to account for genetic drift and, crucially, the discrete nature of individuals. Fractional numbers of deleterious mutations, implicit in the deterministic mathematical analyses that are often appropriate for large populations, are of little consequence because they are selected against. But allowing fractional numbers of beneficial mutants at the nose yields nonsense because fractional individuals that are highly fit multiply and take over the population. Thus even for very large populations, the population size, which determines the smallest fraction of the total population that represents at least one individual, plays a crucial role. Infinite-N deterministic approximations are not even qualitatively correct.
The problems with the simple deterministic approximation to Equation A1 are revealed by analyzing the resulting behavior. This shows that the deterministic solution does not support a steady state v—rather, it predicts that the speed of evolution accelerates without bound. This is clearly unbiological, as it involves a concomitant exponentially increasing width of the distribution and thus smaller and smaller numbers in the nose. Except for very short times (roughly until the nose develops in the correct analysis), the deterministic approximation is thus drastically wrong even for very large N. The source of the problem is that each more-fit population grows faster than the one before. Thus early mutants into a new more-fit fitness class at fitness x + s grow faster than the population at fitness x. This means that even tiny fractions of an individual—certainly nonbiological!—will later give rise to a large population even without further mutations. Indeed, it is the “descendants” of these early fractional mutants that will later dominate the population of individuals at fitness x + s, despite the fact that there are more mutants occurring from fitness x. These descendants then produce fractional mutants to fitness x + 2s, and the unrealistic aspects are further exacerbated.
An alternative way to study Equation A1 is to use a moment-based approach. We can can multiply Equation A1 by x and integrate to find the rate of change of the first moment of the fitness distribution (the speed of evolution) in terms of the second moment (the variance). In the limit that mutation is negligible compared to selection in the bulk of the fitness distribution, d〈x〉/dt ≈ var(x), simply the fundamental theorem of natural selection. One can easily work out that the time derivative of the second moment (the variance) involves the third moment. The time derivative of the third moment involves the fourth moment, and so on. This moment hierarchy does not close. Even so, this approach can yield accurate results for short timescales. The more moments that are kept, the longer the results will be accurate for, and if enough are kept the steady-state speed of evolution can be calculated accurately. The lowest-order version of this is familiar—it corresponds to assuming that the variance is given by its value at t = 0 and does not change and that the speed of evolution is equal to that.
D. Kessler and H. Levine (unpublished results) carried out a sophisticated analysis using a moment-based approach; their work contains a more detailed analysis of the issues involved. Accounting properly for the effects of mutations, stochasticity, discreteness in population number, and fixed total population size is very difficult. Thus far, this analysis involves complex moment equations that unfortunately provide little intuition and no simple analytic results.
The problems with moment equations are unsurprising on the basis of our analysis. As we have noted, it is the lead qs, not the variance or another moment that is most naturally thought of as being maintained by the balance between mutation and selection. This lead is not a moment of the fitness distribution—it is instead a measure of its nose, near to which the discreteness in population number is crucial. The lead thus represents some combination of high moments of the fitness distribution, with the order of the moments that matter depending on N: to capture the effects of the sharp nose of the distribution, at least of order 2 ln Ns moments are needed, and such high-order moments may be dominated by rare fluctuations of the lead. It is hardly surprising that getting at the dynamics of the lead with a moment expansion is very cumbersome. Our approach, in contrast, handles the stochastic issues at the nose in a natural way while simply tracking the effects of selection that dominate in the bulk of the distribution.
APPENDIX B: VARIABLE N AND EFFECTIVE POPULATION SIZE
We have thus far assumed that the population size is constant. We now consider what happens when we relax this assumption.
If changes in N are rapid compared to the changes in the mean fitness, then we can define a constant effective population size Ne. The definition of Ne can be complicated—it is not necessarily the geometric mean of the actual population sizes. Rather, Ne is the value of the constant population size in our model that gives the same dynamics as the changing-N situation averaged over a timescale long compared to the shifts in N. In practice, this means that if our variable-N population were clonal, NeUbs must be the time-averaged rate at which beneficial mutations would establish. Our theory at constant N is then correct provided we use N = Ne. Strictly speaking, this Ne must also describe the time selection takes to operate, which can mean that a single effective population size does not exist—but since the timescale for selection depends only weakly on N, this can often be neglected. Serial dilution protocols are one case relevant to many experimental situations. Here, a population grows exponentially for G generations, is diluted back to its original size Nb, and then this cycle is repeated. The effective population size in this scenario was calculated by Wahl and Gerrish (2001), who found Ne = NbG ln(2).
In the opposite regime where the changes in N are much slower than changes in the mean fitness, the lead and fitness distribution adjust quickly enough that the correct steady-state behavior for the current N always obtains. This means that we can simply replace the N in our results with the time-dependent N(t).
If the changes in N occur on comparable timescales to the changes in the mean fitness, the situation is much more complicated. We cannot define an effective population size, because the changes in N are too slow to be “averaged” over. On the other hand, the changes in N are too fast to allow the population to continuously adjust and stay in steady state. Rather, the population will often be in a transient regime with a complex dependence on past values of N. We do not analyze this case. Though it is an interesting subject for future work, it is a special situation that is unlikely to have general importance.
APPENDIX C: RUNNING OUT OF BENEFICIAL MUTATIONS
We have taken the beneficial mutation rate Ub to be a constant. However, each beneficial mutation that establishes is likely to change the total number of beneficial mutations that are available. Clearly once an individual has a beneficial mutation, that particular mutation is no longer available. But it is also possible that one mutation may open up or close off other possibilities. Thus the beneficial mutation rate Ub may change in complicated ways.
In many cases, Ub will change slowly with each mutation. Our theory predicts that the steady-state value of q at a given Ub is . Provided that the change in q(Ub) over q mutations (after which the fitness distribution has moved through its full width) is small, then the population is always approximately in the steady state and our theory still holds—we simply replace Ub everywhere with the appropriately varying Ub(t). This condition holds provided the change in Ub from a single mutation is small enough that q(Ub) changes by ≪1.
When Ub changes rapidly enough with each mutation that this condition is violated, the population fitness distribution does not adjust quickly enough to stay in steady state. In this case, the population will often be in a transient regime with a complex dependence on past values of Ub. This situation can be analyzed with the algorithmic methods described in the section on transient behavior.
One type of change in Ub is of particular interest: when each mutation that establishes is no longer available, but does not open up or close off any other possibilities. We assume that there are initially k beneficial mutations, each of which occurs at a rate μ. After i such mutations have been established, there are ℓ = k − i left, and the mutation rate is Ub = ℓμ. This situation has been analyzed in great detail by Rouzine et al. (2003). We can get a sense of the behavior by substituting Ub = ℓμ into our formula for q to calculate how much q changes after a single establishment. If this is ≪1, our steady-state theory is a good description of the dynamics; we simply use the appropriate (changing) value of Ub. Otherwise, the population will often be in a more complicated transient regime. This condition corresponds to(C1)where qℓ is the value of q corresponding to Ub = ℓμ. Since we have assumed that , this condition will almost always be satisfied, even for very small values of ℓ (i.e., when the population has almost reached the fitness “peak”). The only potential complication is that if , then our assumption may break down for small values of ℓ.
APPENDIX D: FLUCTUATIONS IN τ, VARIATIONS IN v, AND STABILITY OF THE STEADY STATE
The establishment time τq is a random variable. Above we calculated the steady state assuming that each establishment takes the average establishment time 〈τq〉. However, there are stochastic variations in this establishment time that lead to fluctuations in the speed of evolution. These variations could also affect the average v, because the average v is really determined by the average effect of variable τq, not the effect of the average τq as we have assumed thus far.
The full distribution P(τq) is a special function—a change of variable in the one-sided Levy distribution P(nq). However, we can calculate arbitrary moments 〈〉. The second moment is(D1)From this we can calculate the variance in τq,(D2)The relative variation in τq is thus(D3)For small , this is small even for q = 2 and decreases as for large q. Thus the total fluctuations in the lead (and the speed of evolution) are small, and ignoring them in calculating the average v is reasonable.
From these fluctuations in τq, we want to calculate the expected fluctuations in v. This would explain how much variation in adaptation we should expect between different populations experiencing the same conditions (for example, geographically distinct subpopulations or different experimental lines). Unfortunately, however, this is a difficult problem. This is because successive establishment times are not independent. A shorter than average τq immediately increases the lead. This tends to make subsequent establishments shorter as well. The opposite is true for longer than average τq. Thus the lead is unstable to fluctuations in the short term—increasing the lead due to a short τq creates a tendency to further increase the lead, and vice versa. This effect is enhanced because a shorter than average τq means that the population is less influenced by subsequent mutations, so its size earlier is slightly bigger than usual [i.e., τ(2τq) is closer to τq than usual]. Again, the opposite is true for longer than average τq. This short-term instability is checked at later times. A subpopulation with a short τq is more fit relative to the mean than it would be with an average τq. It thus becomes the dominant subpopulation, increasing the mean fitness, more quickly. When this happens, the lead is decreased—roughly q establishments after the short τq. Thus the various τq are correlated in a complicated way: a short τq tends to favor further short τq, until roughly q establishments later when it favors longer τq, and the opposite is true for longer than average τq.
To understand these complications, it is important to consider more carefully the form of the distribution of τq, especially for large q. Since ℓ ≡ ln s/Ub is large, it is convenient to define(D4)with Δ having both average value and stochastic fluctuations of order unity (and thus small compared to ℓ). For small q, the characteristic magnitude of the fluctuations is correctly captured by the variance. The behavior for large q is somewhat more subtle. In this limit, the mean value of Δ is of order 1/q, but its distribution has an interesting form: Δ is typically ≪1/q and is rarely negative, but with probability of order 1/q it is positive of order unity. The variance of Δ is thus of order 1/q as can be seen from the above result for Var(τq), but, in contrast to what one might expect, all higher moments are also of order 1/q. The strongly asymmetric form of the distribution of τq has a simple origin: there is some chance that an establishment occurs anomalously early, but as the feeding population is producing mutants at an exponentially growing rate, it is highly unlikely that the establishment will be anomalously late.
For large q the form of the distribution of τq has implications for the distribution of the “sweep” time, ts, until new mutants dominate the population. This is tsqτq ≈ qℓ/(q − 1)s ≈ ℓ/s on average for large q. The variations in ts will arise from two sources. The first is the sum of the variations of q successive τq's. From the above discussion, the sum of q Δ's will have a distribution with typical and average value both of order unity. This will give rise to fractional variations of ts of order 1/qs, which is smaller than the mean ts by a factor of 1/qℓ.
But another factor needs to be taken into account: a short τq will increase the lead and thus make the next establishment likely to happen somewhat sooner, thereby making subsequent ones likely to be even earlier. Until the mean population feels the effects of the series of new mutant subpopulations, the lead is thus exponentially unstable. But this effect is not large: the deviation from average, u(t), of the speed of the lead, grows proportionally to the increase, λ(t), of the lead from qs with(D5)Thus dλ/dt = λs/ℓ, so that for large q in a time ts ≈ ℓ/s, an anomalously large lead will grow further only by a factor of e. This means that the effects of the exponential instability of the lead are only beginning to be felt before they are counteracted by a sooner than typical advance of the mean fitness. The above estimate from a sum of roughly independent Δ's thus correctly gives the rough magnitude of the small variations in ts. But the correlations between successive τq's mean that the velocity fluctuations are correlated over times of order ts.
On timescales ≫ts, the mean fitness will grow, with the mean speed and diffusive fluctuations around this described by(D6)with the diffusion coefficient inferred from the above to be(D7)
APPENDIX E: ON THE CUTOFF IN THE INTEGRAL IN H AND THE PATHOLOGIES OF 〈nq(t)〉
One initially surprising property of the distribution P(nq, t) is that it has infinite mean: that is, 〈nq〉 = ∞. The infinity arises because we have allowed mutations from nq−1 to nq to occur arbitrarily far back in the past—even before the establishment of the q − 1 population (as described in the main text, this was implicit in using −∞ as the lower limit of integration in the expression for H). Naively, it seems that this is a serious problem and that the solution is to impose a realistic cutoff in time before which mutations are disallowed. That is, we could say that before t = ti there is a negligible chance of mutations occurring and therefore set the lower limit of integration in H to be ti. This does remove the infinite 〈nq〉. However, it does nothing to address the underlying issue. Rather than being infinite, we would then have 〈nq〉 depending very strongly on ti. This is biologically unreasonable, since the population nq arises from mutations that tend to occur only after nq−1 reaches a relatively large size (naively, of order ). Certainly the important properties of nq(t) should be independent of whether we consider only mutations that occur after nq−1 reaches one individual vs. two individuals, for example. Indeed, since our expression for nq−1(t) is not valid at these small subpopulation sizes anyway, for our results to be valid, they had better not depend on such early times.
The solution to this apparent dilemma lies in the fact that the average nq(t) is not an important property of the distribution of nq(t). Rather, 〈nq〉 is dominated by events so rare that they will never actually occur in practice—namely, when a mutation occurs in the subpopulation nq−1 while nq−1 is extremely small. The reason for the resulting large 〈nq〉 is that even though mutations are very rare far back in time when nq−1 is small, they have a huge effect on the future nq when they do occur and establish. Since the subpopulation nq grows faster than the subpopulation nq−1, the very early mutations dominate over later ones. This can be seen explicitly. The probability of a mutation from the population nq−1 at a time t0 is , and if a mutation occurs at that time it will on average lead to a lineage that at later time t is of size . Thus the contribution to 〈nq(t)〉 from mutations at time t0 is of order . The dependence on t is as expected. However, the dependence on t0 is such that the smaller t0 is (especially at large negative t0), the larger the contribution to 〈nq〉. This average nq is thus dominated by mutations that happened very early. The essential point is that although the probability of a mutation decreases exponentially at rate (q − 1)s as we decrease the initial time t0, its effect on nq increases exponentially at the faster rate qs.
But the lower limit of the mutation times is important only for determining the very-large-nq form of P(nq, t). This part of P(nq, t) contains extremely small probabilities of extremely large nq, in such a way that all integer moments of nq depend crucially on this choice. However, this high-nq part of P(nq, t) represents such a small total probability that it would not occur in any real population. Thus getting P(nq, t) correct for this high nq cannot matter. To get the quantities of interest—in particular 〈ln nq(t)〉—we can therefore use any cutoff we choose, and −∞ is a convenient choice.
The problems with 〈nq(t)〉 all stem from the fact that the population grows exponentially once mutations occur. Thus it is natural to “factor out” this deterministic exponential growth in defining aspects of the distribution P(nq, t) and then focus on the distribution of ln nq(t) − qst. This is what our definition of τq accomplishes. The variable τq, as we have seen, has none of the problems of 〈nq(t)〉 and its distribution is independent of the cutoff we choose (except for a tiny and irrelevant tail for very anomalously small τq). As described in the section on the fate of a single mutant, the essential point here is the difference between 〈eX〉 and e〈X〉. The former (analogous to 〈nq〉) is very sensitive to the tails of P(X), while the latter (analogous to ) is not. And it is the latter that will determine the mean speed and fluctuations around this.
APPENDIX F: MULTIPLE STOCHASTIC CLONES AND Ub/s
Our analysis rests on a separation between deterministic and stochastic dynamics, which we used to overcome the limitations of branching process models. Such a separation is always possible for , as noted above, because nonlinear effects are not important when stochastic effects are, and vice versa. However, we have made a stronger assumption: that the separation is possible right at the nose, so that only the most-fit subpopulation must be treated stochastically but that all other subpopulations are deterministic. This is an important assumption, as a full stochastic treatment would involve, for example, a double-mutant subpopulation whose size is a random variable sending mutations into a triple-mutant subpopulation whose size is also a random variable, and so on. These multiply random processes are difficult to understand analytically.
Fortunately, there is a broad parameter regime in which only the most-fit subpopulation is small enough to require stochastic analysis. Two conditions must be met. First, the most-fit subpopulation at the nose cannot generate new mutations that are destined to fix until it has become large enough that the stochastic effects are negligible. Implicit in this condition is the assumption that the most-fit subpopulation can generate mutations that are destined to go extinct due to drift. This naively seems reasonable, as mutations destined to go extinct due to drift should not matter in the long term. This leads to the second condition: a population destined to go extinct due to drift cannot itself generate a mutation that will become established—otherwise it does matter after all. Here we consider this latter condition. In appendix g, we consider the former.
We begin by studying the dynamics of the lineage founded by a single mutant. Thus we are concerned with a stochastic subpopulation with a fitness s (or some ps) greater than the mean fitness of the population, evolving by our branching process model starting from 1 individual at t = 0 and with no further mutations. We denote the size of this subpopulation at time t by n(t). We have already calculated P(n, t), but this quantity offers no straightforward ways to understand whether mutations can arise while n is still stochastic.
The expected number of mutations that arise from the mutant lineage is . Inspired by this, we define(F1)as the “weight” of the mutant lineage. If the lineage becomes established, W will be infinite (the nonlinear saturation effects are not part of the branching process). However, if the lineage goes extinct due to drift, W is the overall integrated population size. The expected number of mutations destined to survive drift, k, that arise from this lineage is therefore k = WUbs.
We can exploit the independence between stochastic lineages (valid because ) to calculate W. The initial mutant that founds the lineage will either die [with probability ] or give birth [with probability ]. The time T until this happens is exponentially distributed with rate 2 + s [i.e., ]. If it dies, W is simply T. If it gives birth, W is T plus the W of each of the two offspring. We therefore have(F2)Converting to Laplace transforms, we can solve for W to find(F3)where W(z) is the Laplace transform of . Note that , not 1. This is because there is a finite probability (roughly s) that the lineage becomes established and thus has infinite weight. To focus on the lineages that do go extinct, we simply ignore this weight at infinity.
This form of W(z) is impossible to invert analytically for general w. However, the small-z behavior controls the dynamics at large w. For we have that falls off at least as fast as(F4)Values of w ≳ are exponentially suppressed. Integrating this result, we find that less than a fraction s of the lineages have a weight >, and almost all of these are right at . This makes intuitive sense. The largest size a lineage can reach without establishing is ∼. If it does so, it takes ∼ generations to get to this size and another generations to then go extinct. This is because the dynamics are an approximately neutral process while the lineage size is < (drift dominates selection in this regime), so the classical neutral result applies. During this period its average size is ∼, so the maximum value w can take should indeed be ∼. The chance of the lineage reaching size is also ∼s (again by analogy to the classical neutral result) and once there it is about as likely to establish as to eventually go extinct. So our result that w takes on this maximum value roughly a fraction s of the time also makes intuitive sense.
To assume that mutations destined to establish never arise from a subpopulation destined to go extinct, we require . Note that the right-hand side of this expression is s because lineages go extinct for every one that establishes and mutations destined to fix must be much more likely to arise from lineages that establish. Since the maximum value of w is ∼ and this occurs a fraction s of the time, this translates to the condition . (Values of w < are more common, but in sum are still less likely to produce a mutation.) Thus we can ignore mutations from stochastic lineages destined to go extinct provided(F5)
We have not yet considered whether mutations can arise in the stochastic period of lineages destined to survive. We address this question in more detail in appendix g. However, below a size the lineages that establish behave similarly to the lineages that are destined to reach size and then go extinct, and above this size the surviving lineages quickly become deterministic. Thus we expect that whenever mutations never arise from lineages that go extinct, they will also never arise during the stochastic period of lineages destined to survive.
APPENDIX G: THE τ(t) APPROXIMATION
Our method of linking the deterministic behavior of the bulk of the population to the stochastic behavior at the nose hinges on our definition of τq. We defined τq as τ(t → ∞), where τ(t) is defined by(G1)The variable τ(t) is just a change of variable from nq(t). From its definition, we see that τ(t) is the time at which the subpopulation would have reached size had it always grown exponentially at rate qs until reaching size nq(t) at time t. Thus τ(t) accounts for all the incoming mutations and stochastic behavior up to time t and allows us to summarize it by saying nq reached size at time τ(t) and was deterministic thereafter. The definition of τq as τ(t → ∞) thus summarizes all the random behavior and all incoming mutations into a time the subpopulation would have reached size . Yet this is not actually the time the subpopulation reached size (Figure 4). It could, for example, have reached earlier than this but by chance have grown slower than eqst for a while thereafter. Despite this, we have assumed that the subpopulation did in fact reach size at its establishment time in defining its size thereafter. That is, we have written , defining t = 0 to be the establishment time of this population. And we use this form of nq−1(t) in calculating how many mutations this subpopulation generates.
For this to be reasonable, our form of nq−1(t) must be accurate once this population becomes large enough that it starts generating mutants. This happens ∼τq generations after nq−1 became established (by definition, it takes ∼τq generations for the next mutations to occur, because τq is dominated by the waiting time for the first mutation to occur). Thus for our result to be accurate, τ(2τq) must be ∼τq (to be precise, we require ). That is, there must not be much stochasticity after the population is large enough to generate mutations (and additional incoming mutations must be negligible). Looked at another way, this means that the population cannot generate mutations while it is stochastic.
To calculate τ(2τq), we return to our solution H(ζ, t) for the Laplace transform of P(nq, t). The time dependence of τ is hidden in Equation 27—our assumption that ζ is small here assumes we are interested only in larger nq and is thus equivalent to taking t → ∞. We can do this integral more carefully; the result involves hypergeometric functions. These can be expanded for but nonzero, corresponding to values of nq(t) > but before this subpopulation generates mutations. We find(G2)Unfortunately, this form of H is more complex and we cannot exactly compute τ(t). However, we can find typical values of τ(t) and τq from this by the same methods as before. We can also compare the size of the second term in H [which gives the time dependence in τ(t)] to the first for values of , which corresponds to , the time this subpopulation begins to generate new mutations. Both calculations demonstrate that our approximation is valid provided that(G3)This result can be confirmed with a deterministic analysis. About τq generations after becoming established, a subpopulation has a size . Once it has reached this size, selection dominates drift and mutations. Thus subsequent random or mutational events will not significantly affect n, so τ(2τq) and τq = τ(∞) are similar.
Thus whenever , our method of linking together stochastic and deterministic dynamics is reasonable. Populations never generate mutations while they are stochastic, and hence we are justified in using a deterministic approximation for all but the most-fit population. When this condition fails, we must treat multiple populations stochastically and the analysis becomes much more complex. We could still divide up the population into a nonlinear deterministic part and a linear stochastic part (provided only that ), but the stochastic part would have to include multiple subpopulations.
APPENDIX H: APPROXIMATIONS IN THE BEHAVIOR OF q
In our analysis to this point, we have assumed that the mean fitness changes abruptly, increasing by s every τq generations. We used this assumption in calculating q and it is the reason why we have a constant q. In this appendix, we discuss this approximation.
Two important timescales determine the relative sharpness of the changeovers from one dominant population to the next and, concomitantly, from the lead population growing with rate qs to rate (q − 1)s. Because the second largest population grows at rate s, the timescale for this changeover is 1/s. But the time between such changeovers is τq. The ratio of these is(H1)which is 1/ln(s/Ub) and thus small at the crossover from the successional- to the multiple- mutations regimes. Indeed as long as , the changeover is relatively sharp on the scale of τq and it is a good approximation to consider it abrupt, as we have done.
We can make this more precise by computing the actual behavior of the mean fitness as a function of time. Assume (for convenience) that at t = 0, the mean fitness is at , i.e., in the middle of a changeover. The subpopulations at y = −1 and at y = 0 are equal in size, and those at other values of y are smaller by a factor of . For small w, the one or two largest subpopulations strongly dominate, as these factors are all very small. This is because the variance of the fitness in the population in the multiple-mutations regime is simply v, since the dynamics of the bulk of the population are controlled by selection. Thus the standard deviation is <s, making the other subpopulations far smaller than the dominant one. The parameter w is simply the variance in units of s2.
At future times, the subpopulations all grow (or shrink) exponentially at a rate ys reduced by the mean fitness (but we can neglect the mean fitness in this calculation because it affects all subpopulations equally). To keep the total population fixed thus requires that the mean fitness be(H2)We can perform these Gaussian sums by the Poisson resummation formula to yield(H3)If w were large, the k = 0 term would dominate, and the k = ±1 would yield relative variations in the speed(H4)and corresponding variations in that are a smaller by a factor of 1/(2π). Thus in practice the parameter that needs to be large for to increase smoothly is 2π2w. Only for w < 0.2 do the variations in v become more than a factor of 2, and substantial deviations of from smooth occur only for w < 0.1. Above this, our abrupt-transition approximation is not valid, but despite this our earlier results are still good; we discuss this below.
The parameter that we have taken to be small throughout is 1/ln( s/Ub). This is the value of w at the crossover from successional- to multiple-mutations regimes. Strictly speaking, this means that w is small until ln Ns ∼ (ln s/Ub)2. For even larger population sizes, the behavior near the nose changes somewhat, as discussed in the main text.
When w is small enough that the shifts in are abrupt, the dynamics can be worked out more generally than we have done in the main text. There, we approximated the most-fit deterministic subpopulation to be growing as e(q−1)st for τq generations, after which increases by 1 and the subpopulation growth slows to e(q−2)st, and so on. This is strictly valid only for integer q. When the naive value for q, 2L/ℓ, is noninteger, the populations shift between growth rates some fraction of the way between one establishment and the next. The effects of this can be taken into account straightforwardly as long as the shift between growth rates is indeed abrupt on the scale of τq: i.e., that w is small. Here we ignore factors inside logarithms: to get these one would need to use the fuller analysis of the feeding and lead population dynamics used in the text. For our purposes here, the heuristic derivation of the establishment times is sufficient.
It is convenient to keep q an integer, with qs the growth rate of the lead population when it first becomes established. We then define a noninteger generalization of q to be with(H5)the greatest integer . It is that is simply related to the population parameters via(H6)i.e., what was previously found for q. The dimensionless speed is found to be(H7)which is equal to the result in the text, , for integer . The difference between these is small for large , with the fractional error of the simple result (which is an overestimate) largest at , where it is only 1/4q(q − 1) and thus small even for the worst case .
In the opposite case where w is not small, the approximation of abrupt shifts in is not valid. In this case, we can make the opposite approximation that the mean fitness increases at a uniform rate: from the above discussion, this is valid unless w is quite small (although strictly speaking this is not true in the limit that ℓ is large with fixed q). In the constant mean-speed approximation, one obtains(H8)which is an underestimate that is worst at integer ; for large the worst fractional error is 1/4(q − 1)2. Since this is small compared to the speed, the approximation in the main text is reasonable.
We can get an intuitive understanding of why this approximation of abrupt shifts in gives reasonable results, even when actually increases smoothly. First we consider the deterministic dynamics of the bulk of the fitness distribution. Here the shape of the distribution (and hence the identity of the most common subpopulation) depends only on the relative growth rates of the subpopulations, so assumptions about are irrelevant. For the stochastic behavior at the nose, our assumption is more problematic. When the mean fitness in fact increases steadily, rather than jumping by s every time an establishment occurs, our calculated lead q gives the correct average mean fitness over the stochastic period. This means we calculate the stochastic dynamics assuming the correct average mean fitness, but this is slightly different from the stochastic dynamics given the changing mean fitness. Essentially, we have used q = 3.4, for example, as an interpolation for the correct behavior when the lead is just below 4 immediately before an establishment, declining gradually to below 3 shortly before the next establishment. Rather than calculate τ3.4 from the stochastic behavior while the lead shifts correctly, in the main text we have calculated it on the basis of a constant lead of 3.4. As we have seen above, however, the difference is small.
We conclude with some comments on the stochastic aspects of the speed of the nose. These make the above analysis questionable because of the assumption of deterministic establishments of the lead populations. But, as we discuss in appendix d, the variations in the establishment times are at worst of order 1/qs compared to the mean τq that is of order (for large q they are even smaller than this, as discussed in appendix d). Thus the variations in the time intervals between takeovers of the population by new dominant subpopulations are small compared to the time intervals themselves. Hence the deterministic approximation for the increase of the mean fitness is good, at least as far as its effects on the dynamics of the lead populations.
We thank John Wakeley, Igor Rouzine, Dan Weinreich, and especially Andrew Murray for useful discussions. This work was supported in part by the National Institutes of Health via grant P50-GM068763-01, by the National Science Foundation via grant DMR-0229243, and by the Merck Foundation.
↵2 Present address: Department of Applied Physics, Stanford University, Stanford, CA 94305.
Communicating editor: M. W. Feldman
- Received November 1, 2006.
- Accepted April 19, 2007.
- Copyright © 2007 by the Genetics Society of America