Experimental evolution (EE) combined with whole-genome sequencing (WGS) has become a compelling approach to study the fundamental mechanisms and processes that drive evolution. Most EE-WGS studies published to date have used microbes, owing to their ease of propagation and manipulation in the laboratory and relatively small genome sizes. These experiments are particularly suited to answer long-standing questions such as: How many mutations underlie adaptive evolution, and how are they distributed across the genome and through time? Are there general rules or principles governing which genes contribute to adaptation, and are certain kinds of genes more likely to be targets than others? How common is epistasis among adaptive mutations, and what does this reveal about the variety of genetic routes to adaptation? How common is parallel evolution, where the same mutations evolve repeatedly and independently in response to similar selective pressures? Here, we summarize the significant findings of this body of work, identify important emerging trends and propose promising directions for future research. We also outline an example of a computational pipeline for use in EE-WGS studies, based on freely available bioinformatics tools.
For much of its history, evolutionary biology has been a retrospective science. Patterns of variation among contemporary populations or fragmentary data available in the fossil record are all that we have had to make inferences about the evolutionary process. This situation has begun to change with the merging of two research programs, experimental evolution (EE) and whole-genome sequencing (WGS), that together stand to provide a comprehensive view of the genomic changes underlying evolution. The long-sought link between genotype and phenotype is thus becoming more accessible and, in the process, is helping to transform evolutionary biology into a truly prospective science.
Experimental evolution aims to test evolutionary theory directly in evolving populations of laboratory organisms such as microbes. Although the technique has been around since the late nineteenth century, and was even known to Darwin through correspondence with its first practitioner, Dallinger (1887), EE has developed into an independent and vigorous research program only in the past 20 years or so (Elena & Lenski 2003; Hegreness & Kishony 2007; Buckling et al. 2009; Wichman & Brown 2010). The strength of EE lies in its ability to conduct replicated, controlled, manipulative experiments where changes in fitness can be tracked directly through time using standard laboratory techniques. When combined with rapid, affordable, high-throughput techniques for WGS, it is possible to obtain the entire genome sequences of ancestral, intermediate, and evolved strains for an entire experiment. In other words, it is now possible to provide a complete record of the genetic changes that are associated with phenotypic and fitness differences that arise during the course of evolution. This is tremendously exciting because, for the first time, it allows a far more comprehensive view of how evolution works by connecting genes directly to both phenotype and fitness.
There are now enough published EE experiments that have also used WGS that it seemed worthwhile to try to synthesize their results. Most have used microbes such as viruses, bacteria or yeast. In what follows, we have focused almost entirely on free-living, cellular microbes such as bacteria and yeast because a first review of viral experiments has recently been published (Wichman & Brown 2010). Up to now, most of the EE studies that have employed WGS have simply focused on describing what happens; they effectively constitute a natural history of how genomes change in experiments lasting hundreds or thousands of generations. The field is maturing, though, and sufficient data from multiple experiments now exist that we can begin to take a more hypothesis-driven approach. Our aim is to provide some insights into what the first wave of EE-WGS research tells us about the fundamental principles underlying adaptive evolution.
Specifically, we ask the following questions. How many mutations underlie adaptive evolution, and how are they distributed across the genome and through time? Are there general rules or principles governing which genes contribute to adaptation, and are certain kinds of genes (e.g. regulatory vs. structural) more likely to be targets than others? How common is epistasis among adaptive mutations, and what, if anything, does this reveal about the variety of genetic routes to adaptation? How common is parallel evolution, where the same mutations evolve repeatedly and independently in response to similar selective pressures? Preliminary answers to most of these questions are now possible, although in some instances major gaps exist. Identifying these gaps will, we hope, help guide the agenda for EE-WGS research into the future.
A brief guide to methods
EE is the study of the evolutionary process under controlled conditions in real time, and the advantages of this approach for testing theory in evolution are probably familiar to most readers. The repeatability of evolution can be assessed by running replicate populations founded with the same starting material under identical, strictly defined, environmental conditions, a situation that rarely or never occurs in nature. Another key benefit is that the historical environment in which mutations arise and are substituted can be known with certainty, whereas the historical conditions experienced by a natural population may not be accurately deduced by current conditions. One can follow how fitness and genetic variation, either arising de novo through mutation (e.g. Barrick et al. 2009) or being sorted from pre-existing standing genetic variation (e.g. Turner et al. 2011), change through time. Constructing experiments in the right way, which involves paying attention both to the fundamental processes one is interested in investigating and the level of replication required to distinguish signal from noise, affords direct tests of evolutionary theory.
Several technical and practical benefits unique to microbes make them ideal subjects for EE research (Elena & Lenski 2003; Hegreness & Kishony 2007; Buckling et al. 2009; Wichman & Brown 2010; Brockhurst et al. 2011). Short generation times and large population sizes allow investigators to observe hundreds or thousands of generations of evolution within a reasonable time frame. Ease of propagation makes it possible to evolve large numbers of independent populations, and clonal reproduction allows for replication at the genotypic level. Most microbes can be archived in a nonevolving state, from which they can later be revived, so ancestral, intermediate and evolved forms can be compared directly.
High-throughput, next-generation sequencing (NGS, Box 1) technologies generate entire genome sequences rapidly and at a fraction of the cost of conventional sequencing methods. It is now technically and financially feasible for individual laboratories to sequence multiple strains from large numbers of evolved populations (Brockhurst et al. 2011). Accumulated mutations can be detected comprehensively by searching for polymorphisms between ancestral and evolved genomes. This WGS approach is inherently comparative so one is not restricted to the use of organisms with previously sequenced reference genomes. Thus, even if reference-guided mapping is not possible, current technology permits the de novo assembly of any microbial genome, even those of much greater size and complexity than have traditionally been used (e.g. Gnerre et al. 2011). In practice, though, most studies to date have used microbes with sequenced genomes, such as E. coli and yeast, in part because of the wealth of molecular and physiological knowledge available for these model organisms.
Table Box 1. . Glossary of terms, defined in the context of EE-WGS
Adaptation or adaptive evolution: Changes in trait values associated with an increase in fitness
cis-regulatory region: A region of DNA that regulates the transcription of a gene in close proximity
Clonal interference: Competition between multiple beneficial mutations that arose independently in different individuals (in an asexual population)
de novo assembly: Assembling sequence reads without using a reference genome as a guide
Epistasis: When the fitness effect of a mutation is modulated by its interactions with other genes or mutations in the genome
Genomic evolution: The accumulation of substitutions within the genome
Insertion/deletion (indel): A mutation that inserts or deletes nucleotides from a DNA sequence
Mutation: Any change in DNA sequence that arises in the population. Selection and/or drift will determine whether the mutation is lost or maintained
Negative or antagonistic epistasis: When the fitness effect of mutations in combination is less than would be expected from the sum of their independent effects
Next-generation sequencing (NGS): Any high-throughput sequencing technology that parallelizes the sequencing process, typically producing a large number of short sequence reads
Nonsynonymous mutation: A mutation in coding DNA that alters the resulting amino acid. Most of these mutations are detrimental, but some can be neutral or beneficial
Positive or synergistic epistasis: When the fitness effect of mutations in combination is greater than would be expected from the sum of their independent effects
Reference-guided mapping: Piecing together sequence reads from one genome using a previously sequenced reference genome as a guide
Sign epistasis: When the sign of the fitness effect (beneficial [+] vs. detrimental [−]) of a mutation is altered by epistatic interactions
Single nucleotide polymorphism (SNP): A mutation that changes the DNA sequence at only one nucleotide position
Substitution: A mutation that arose and eventually replaced the ancestral sequence in the population
Synonymous mutation: A mutation in coding DNA that does not alter the resulting amino acid. Most of these mutations are neutral
Most NGS methods share a common theme, although different platforms use different technologies. The basic idea is to generate an extremely large number of short sequence reads that are then pieced together using a previously sequenced reference genome as a guide (Box 2). The parameters used by the mapping algorithm and the strictness of mismatch declaration will affect the rates of false negative or positive calls. Processing the enormous amounts of raw sequence data generated from high-throughput NGS is not a trivial endeavour. Recent developments in bioinformatics associated with NGS have led to a daunting array of software and utilities. Here, we outline an example of a computational pipeline for use in EE-WGS studies, based on freely available programs, which can serve as an entry point into this subfield. The pipeline is designed to process raw sequence data files into a final list of potential mutations that are fully annotated to a reference genome (Box 2, Appendix S1, Supporting information).
Table Box 2.. Bioinformatic tools for next-generation sequencing (NGS) data analysis
With the development of increasingly higher throughput technologies, NGS capabilities need to be complemented with information technology (IT) systems for data management and analysis. There has been a flurry of recent papers dealing with subsegments of such NGS-IT systems. In this Box, we give a brief overview of the basic steps that would typically be part of a bioinformatics pipeline for the sequencing and analysis of experimentally evolved genomes (see Figure). In the accompanying Supporting Information file, we describe a pipeline in detail which can be built from freely available tools.
NGS technologies, such as Illumina, generally exhibit drastic drops in the quality of base calls as the number of cycles increases. In many cases, it may become important to evaluate these qualities and filter out reads, or parts of reads, from the data to be used in the subsequent analyses. One tool for performing this assessment is FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc/). However, there is no universal approach to filtering, and choices still need to be evaluated case-by-case.
Global read mapping
Experimental evolution data are generally associated with a previously sequenced model organism, which can serve as a reference genome to which one can align the evolved genomes. In such contexts, the well-known bioinformatics tools for sequence alignment, such as BLAST, are ill-prepared for the size and properties of the data sets to be aligned. In recent years, several new short-read alignment systems have been proposed (e.g. Li et al. 2008, 2009b; Langmead et al. 2009; Li & Durbin 2010). These novel alignment, or read mapping, systems attempt to exploit the information of pair-end mappings, which have been shown to outperform single-end mapping (Li & Homer 2010). The alignments produced by these systems can be emitted in the SAM file format (Li et al. 2009a), or its binary equivalent BAM format, which is becoming the most widely supported alignment format of NGS data.
The millions/billions of reads produced in a sequencing project are aligned to the reference genome one by one. However, the alignment of any given read can often be fine-tuned by exploiting the information of other reads that also map to the same location on the genome. Performing a multiple-sequence realignment has been shown to have a significant impact on SNP-indel detection, generally reducing the false-discovery rate (Homer & Nelson 2010). This step is likely most useful for relatively low-coverage sequencing projects, where multiple-sequence alignment is most tractable, and where making the most of each read is critical.
SNP and indel detection
The central interest of NGS in EE is to uncover simple sequence variants, namely SNPs and small insertions/deletions (indels) that have accrued over the experiment. Most of the tools for SNP and indel detection have been developed with the 1000 Genomes Project (http://www.1000genomes.org) in mind; nonetheless, they are amendable to the relatively simpler requirementsof EE. These tools are concerned with determining simple variants given the level of coverage and the error level of the sequencing platform (e.g. Li et al. 2008, 2009b). For each nucleotide position in a given data set (which at this point consists of an alignment of reads for a particular sample), simple variant detection tools evaluate the posterior probability of each possible genotype. If the state with the highest probaility is also the state of the reference, the position is considered invariant. However, if a state has a high probability and is nonreference, it is deemed to be a possible SNP/indel. In other cases, positions may be declared as variable (perhaps even without there being an obvious state to declare). Following the raw SNP/indel calls, a filtering step is generally taken to eliminate candidates with low quality or low coverage. Once again, no universal filtering criteria are yet available.
Annotating SNPs/indels is the last task of most pipelines. While most tools used for this are conceived with the most widely studied model organisms in mind, there are a few that are customizable to the species of interest, given a GenBank file or other common file formats (e.g. snpEff, http://snpeff.sourceforge.net).
There is some reassurance to viewing the assembled, aligned and sorted reads of a sequencing genome. The several visualization tools, or viewers, that have been proposed (Huang & Marth 2008; Bao et al. 2009; Li et al. 2009a; Hou et al. 2010) can differ in the particular displays available. Most viewers allow the user to create a graphical rendering of the data, navigate between different levels of detail of the data, and connect with external analysis tools or data files (such as annotation files of the reference genome). Regarding connectivity, viewers are likely to become the basic dashboard of NGS-IT systems, providing biologists user-friendly access to the bioinformatics developments that are in progress.
Major efforts have been made to bring together the numerous bioinformatics tools addressing the different subsegments of NGS-IT systems mentioned here into a single package, library or development environment. The Genome Analysis Toolkit (McKenna et al. 2010) is the most well-known example of this. The Galaxy framework (Blankenberg et al. 2010; Goecks et al. 2010) is another integrative effort consisting of ‘wrappers’ around existing tools. There are a number of benefits to centralizing developments, such as end users need not worry about performing multiple disparate updates of individual tools. On the other hand, integrative systems can sometimes force the user into a predefined path of analysis and make substitutions of some tools within the system very difficult or impossible. Most IT-savvy users will want to have direct control over the primary systems, and an example of this type of pipeline is detailed in the Supporting Information file.
Other approaches to identifying mutations have been used, in particular, candidate gene sequencing and microarray-based polymorphism detection, but neither is as comprehensive as WGS. Candidate gene sequencing relies on prior knowledge, and a good amount of luck, to screen the potential physiological and metabolic targets of selection for mutations. This approach was often successful to the extent that mutations were usually found in at least some of the analysed loci (Treves et al. 1998; Notley-McRobb & Ferenci 1999; Cooper et al. 2003; Crozat et al. 2005, 2010; Woods et al. 2006; Ostrowski et al. 2008; Kinnersley et al. 2009). However, these mutations rarely accounted completely for the observed changes in fitness implying that other, undetected, mutations remained to be found.
Microarray-based systems, on the other hand, allow investigators to screen the entire genome for mutations using DNA–DNA hybridization patterns between a reference and an evolved genome (Gresham et al. 2006). Mutation detection relies on the expectation that mutated DNA hybridizes with array probes less well than nonmutated DNA. Hybridization strength of an oligonucleotide is quantitatively sensitive to the number of mismatches, but the position and type of mismatch effects hybridization in a less consistent manner (Gresham et al. 2010). Microarray-based methods can identify regions of several nucleotides that may contain mutations, but exclusion of false positives and identification of the specific mutation site still require significant amounts of follow-up manual sequencing (Herring et al. 2006; Gresham et al. 2008). A further drawback of microarray-based methods is that they have been shown to miss up to 30% of the mutations found through WGS (Herring et al. 2006; Herring & Palsson 2007; Gresham et al. 2008; Kao & Sherlock 2008; Araya et al. 2010; Kvitek & Sherlock 2011).
WGS not only detects more mutations but also detects them more accurately. Moreover, the inclusion of more primary WGS data decreases false positive rates while simultaneously increasing true positive rates. WGS is undeniably the most comprehensive mutation detection method and, for this reason, we focus our efforts on those studies to which this technology has been applied.
Mutations accumulated during adaptation
A sufficient number of EE-WGS studies have now been published that we can begin to ask at what rate are mutations accumulated during adaptation. We collated data on the number of substitutions uncovered in EE-WGS experiments for which appropriate data were available, as shown in Table 1. Note that the short read-lengths of high-throughput WGS hamper its ability to identify larger structural variation in the genome, so some insertion/deletions (indels), gene expansion/contractions and chromosomal rearrangements may be missed. High-throughput WGS excels at detecting single nucleotide polymorphisms (SNPs), so we have restricted most of our attention to this form of genomic mutation.
Table 1. Summary of relevant data from selected EE-WGS studies
For all organisms and selective environments reported in Table 1, an average of 0.97 SNPs was substituted in an adapting genome for every 100 generations of evolution. The distribution is somewhat skewed, with most values being ≤1.0 (Fig. 1). When small indels (≤10 bp) and SNPs are combined, an average of 1.14 substitutions per 100 generations of evolution was observed. The mean for E. coli (1.06/100 generations) is 24% less than that of S. cerevisiae (1.39/100 generations), but this difference probably can be explained by a difference in gene number. The E. coli genome has approximately 24% less genes than the yeast genome, making the rates of substitution per gene the same for the two taxa. Note that we have reported rates of genomic substitution, not total numbers. This is because experiments differ greatly in the number of generations of evolution (Table 1), so some standardization is necessary to compare across studies. The number of new mutations that arise during an experiment cannot be determined because many mutations will be lost from the population without ever being detected. Only mutations that increase to appreciable frequencies, or become fixed in the population (i.e. substitutions), are likely to be identified. Nevertheless, remembering that the neutral theory predicts a per generation rate of substitution that is equal to the mutation rate, we can at least say that the observed rates of substitution are much too high to be purely neutral.
Ideally, we would like to use data on the number and rate of substitution to address one of the oldest questions in biology, whether evolutionary change occurs through many or few mutations. The prevailing view has been that adaptation proceeds gradually, through the substitution of many mutations with exceedingly small fitness effects (Fisher 1930). If, however, mutations are highly variable in their fitness effects, then adaptation may proceed more rapidly by substituting fewer mutations of larger effects. Support for the latter view has been provided by theoretical work that models adaptation as a sequence of steps in phenotype or sequence space towards an optimum (reviewed in Orr 2005a), and by several non-WGS empirical studies (Lenski et al. 2003; Zeyl 2005; Schoustra et al. 2009). Deciding between these alternatives requires not just estimates of the rates of substitution but the total number of mutations fixed during an adaptive walk to evolutionary equilibrium (that is, where the rate of change of fitness is effectively zero) and their fitness effects in the genetic background in which they first arose. Efforts along these lines have begun (Chou et al. 2011; Khan et al. 2011) but this is clearly an avenue for future work.
Adaptive and nonadaptive substitutions
Mutations are the raw material for adaptive evolution, but not all substitutions are adaptive. A common finding is that only a subset of the total substitutions identified by WGS contributes to the fitness gains that are observed in the endpoint populations. More surprising, perhaps, is that some substitutions in bacterial experiments have no effect on fitness at all (Barrick et al. 2009; Conrad et al. 2009; Lee & Palsson 2010; Wang et al. 2010; Minty et al. 2011) or even show detrimental effects on fitness despite evidence of strong parallelism and a signature of positive selection (Conrad et al. 2009; Crozat et al. 2010). Similar results are found in yeast, where two studies systematically tested for the adaptive effects of all observed substitutions (Anderson et al. 2010; Kvitek & Sherlock 2011) and both found that only 40–43% of SNPs conferred a significant fitness advantage. The small numbers of observed substitutions, and the fact that only a proportion of them is responsible for increasing fitness, are consistent with the assertion that the bulk of adaptation is attributable to a few substitutions of large effect.
Why are so many substitutions apparently ‘nonadaptive’? There are at least three possible reasons:
1The substitutions are neutral and either hitch-hike to high frequency with beneficial mutations or drift to high frequency by chance. This explanation seems unlikely because most experiments involve very large population sizes where selection is likely to be much more important than drift. Moreover, there is often strong evidence in these experiments to suggest that most substitutions were positively selected (see Distribution of substitutions across the genome section).
2The substitutions are adaptive, but our methods of fitness estimation are not sensitive enough to detect the fitness increase. There are two reasons why this might be so. First, some experimenters report fitness measures based on single components of total fitness (e.g. maximum growth rate, population density) that may not capture the phenotype through which the adaptive advantage operates. Typically, the preferred method is head-to-head competitive fitness assays, as these integrate all components of the growth cycle into a single measure. Alternatively, even if competitive fitness assays are used, they may not accurately reflect the abiotic or biotic selection environment in which a given mutation arose. For example, if fitness depends on the frequency or characteristics of other genotypes in the population (Rainey & Travisano 1998; Friesen et al. 2004; Lang et al. 2011) or sequential substitutions have nontransitive effects on fitness relative to the ancestor (Paquin & Adams 1983).
3The fitness effects of mutations are epistatic, that is, they are modulated by their interactions with other mutations in the genome (Segre et al. 2005; Phillips 2008; He et al. 2010), and this can have important consequences on the adaptive trajectory taken by a lineage (Lang et al. 2011; Woods et al. 2011). Epistasis seems to be a very common occurrence in EE, although the precise nature of the epistatic interactions can be quite variable. A number of studies have identified instances of synergistic (positive) epistasis, where the fitness effect of mutations in combination is greater than would be expected from the sum of their independent effects (Herring et al. 2006; Applebee et al. 2008; Conrad et al. 2009; Anderson et al. 2010; Lee & Palsson 2010; Minty et al. 2011). In propanediol-adapted E. coli (Lee & Palsson 2010), for example, two mutations had no adaptive benefit when occurring alone but accounted for over half of the adaptive fitness increases observed in the evolved line when introduced together into the ancestral genome. Evolved alleles may be beneficial only in conjunction with other co-acquired evolved mutations (conditionally beneficial), explaining why single mutations may appear nonadaptive.
There are also examples of sign epistasis, where the sign of mutation’s fitness effect (beneficial [+] vs. detrimental [−]) is determined epistatically. In lines of E. coli adapting to lactate, for example, a mutation in the kdtA gene evolved in many parallel replicate lines suggesting it was under strong selection. However, when this mutation was tested in the ancestral genetic background, it was severely deleterious, as it did not grow at all (Conrad et al. 2009). When in combination with co-acquired mutations in other genes, the effect was reversed and the kdtA mutation was then significantly beneficial. A second example comes from glucose-limited yeast (Kvitek & Sherlock 2011) where two mutations were singly adaptive yet the double mutant was actually less fit than the ancestor.
As we discuss in more detail later in this study, there are also examples of antagonistic (negative) epistasis, where combinations of mutations produce fitness benefits that are less than the additive expectation (Chou et al. 2011; Khan et al. 2011). However, it is more difficult to see how this form of epistasis can lead to apparently nonadaptive mutations, because the effect of each mutation singly is clearly beneficial. Nevertheless what is clear from this survey is that epistasis is pervasive in EE, and we still do not understand its role in adaptive evolution. More studies are needed, especially those that comprehensively test the multiplicity of epistatic interactions among co-acquired mutations (Chou et al. 2011; Khan et al. 2011), and how epistasis can affect evolutionary outcomes (Woods et al. 2011).
Distribution of substitutions across the genome
How are substitutions distributed in evolved genomes? In the absence of selection, the proportion of substitutions within open reading frames would be equivalent to the gene density of the genome. If mutations in coding DNA are more likely to have fitness consequences, selection should lead to an over-representation of substitutions in coding regions. Bacterial genomes typically have high gene densities (e.g. E. coli∼86%), which constrains the assessment of deviations from neutrality. S. cerevisiae has a greater proportion of noncoding regulatory elements and lower gene density (∼71%), but 82% of SNPs were found within open reading frames. This over-representation is consistent with positive selection operating on functional coding DNA during adaptation.
Within genes, a similar argument can be made for synonymous vs. nonsynonymous mutations: nonsynonymous mutations alter the amino acid in the protein and are more likely to have fitness consequences than synonymous (silent) mutations. The exact proportion of possible mutations that are nonsynonymous depends on the details of codon composition for each genome but typically ranges between 75% and 80%. As shown in Table 1, the vast majority of EE-WGS strains have 100% nonsynonymous mutations, with an overall average of 92%. When small indels in coding regions are also included, 94% of mutations are amino acid altering. This over-representation of amino acid-altering mutations indicates that positive selection was the main driving force for substituted mutations, and neutral drift played only a minor role.
The timing of adaptive and genomic evolution
A common observation in long-term EE is that the rate of adaptation decelerates over time (see Barrick et al. 2009 for example). Large and rapid increases in fitness are usually seen in the early phase of adaptation, followed by more gradual fitness increases as time goes on. This pattern suggests that either more mutations tend to be fixed early in adaptive evolution than later or the fitness effects of mutations that fix early are larger than those that fix later. These alternative explanations are not mutually exclusive, and EE-WGS data have provided some support for both.
Most EE-WGS studies in which the number and timing of adaptive mutation fixation was estimated find that mutations tend to occur in the early stages of adaptive evolution, tracking the fitness trajectories (e.g. Conrad et al. 2009; Lee & Palsson 2010). Further evidence for the bulk of mutations fixing early is provided by the significant negative relationship between SNP/100 generations and the total number of evolved generations (Fig. 2). As experiments are run longer, the rate at which new mutations are fixed decelerates and lowers the overall rate for the experiment.
Recent experimental evidence suggests that negative epistasis may cause the fitness effects of mutations that fix early to be larger than those that fix later. Two studies have found that the proportional fitness advantage of a beneficial mutation decreases as the fitness of the genetic background increases, independent of the order in which those mutations actually fixed (Chou et al. 2011; Khan et al. 2011). This means that pervasive negative epistasis will tend to cause the effect of the mutation to be largest early in adaptation and smaller later on. Other studies have found results consistent with this mechanism: the magnitude of the fitness effect tends to decrease with each subsequent fixation event (Betancourt 2009; Schoustra et al. 2009; Gifford et al. 2011), and the proportion of all mutations that are beneficial also decreases as a population approaches its fitness optimum (Schoustra et al. 2009). An alternate explanation relies on standard population genetics: beneficial mutations with large effect sizes are under stronger selection, so their probability and rate of substitution will be greater than those for low-effect mutations. This assumes that mutation supply is not limiting and most beneficial mutations are available to selection early in the adaptive process. We should interpret the results from these studies with some caution, however, as the relative contribution of mutation supply and epistatic fitness effects to this relationship cannot be explicitly determined with the available data. Deciding which of these factors is most important in governing the decelerating rates of adaptation is a major challenge for the future.
There is a notable exception to the common observation that the rate of genomic evolution declines with time. Genome sequencing of a long-term E. coli line (Barrick et al. 2009) revealed that it accumulated substitutions at a fairly constant rate for 20 000 generations, while its rate of adaptive fitness gain decelerated. Constant genomic evolution could be explained simply by the random fixation of neutral mutations through drift. The strong over-representation of nonsynonymous mutations and the occurrence of parallelism both suggest that most substitutions were beneficial and under positive selection, so this hypothesis can be rejected. It is possible that, after the large effect mutations have all been substituted, numerous small effect substitutions steadily contribute to the marginal fitness gains observed in the later stages of adaptation. An alternate explanation is provided by a study of population-level diversity (Barrick & Lenski 2009) which found that two deep branching lineages within this population were competing with each other. Such clonal interference can delay the fixation of beneficial mutations, thereby slowing the early accumulation of substitutions and leading to an overall pattern of constant genomic evolution.
Part of what makes this E. coli line unique is the large number of generations (20 000) studied relative to all other EE work. This raises the issue about what timescale is most appropriate for analysing the adaptive dynamics of a population. Adaptation clearly has distinct phases: it is rapid early on and much slower later. It might be that we need different population genetic models to describe the dynamics of substitution in each phase. Early in adaptation, a model that takes into account multiple mutations that compete for fixation (clonal interference; Kao & Sherlock 2008; Lang et al. 2011) may be more appropriate than one that assumes that beneficial mutations are fixed sequentially, as in the strong-selection, weak-mutation assumption used in many models of adaptation (Orr 2005a). From an empirical perspective, this means we have to be mindful of how and when we sample a population: if the point is to understand the early stages of adaptation, then frequent sampling of multiple isolates or whole-population sequencing will be necessary. Indeed, many of the discrepancies between previous studies may be a product of the different timescales of the experiments. Detailed comparisons of the rates of genomic and adaptive evolution are still limited in scope, so additional, highly replicated EE-WGS studies with intensive sampling at appropriate intervals are needed.
Targets of selection
Combining the comprehensive mutation detection of WGS with the wealth of data that have accumulated over the last 20–30 years on molecular genetics of microbes means that we can now begin testing some of the predictions of functional biology. Molecular geneticists have noted that some features of genes and genome architecture can guide predictions about the likely targets of selection. It is now possible to begin asking, for example, whether there are any general patterns in the kinds of genetic targets that respond to selection, at least in laboratory experiments. What functions do these genes perform? Do adaptive mutations occur more frequently in certain functional classes of genes than others? Based on our current knowledge, do the target genes play an obvious or predicted role in the focal phenotype?
Regulatory vs. nonregulatory changes
There is some debate over whether adaptive evolution and phenotypic change typically occur through mutations in cis-regulatory or protein-coding regions of the genome (Hoekstra & Coyne 2007; Wray 2007; Stern & Orgogozo 2008). In principle, this topic could be addressed with EE-WGS data but in practice this is challenging. Most microbes have streamlined genomes with a relatively small percentage of noncoding DNA (e.g. E. coli∼14% noncoding), and we already know that most SNPs are in coding regions. Unlike coding DNA, determining what proportion of noncoding DNA is potentially cis-regulatory is not usually possible, so simply tallying the numbers of substitutions in coding and noncoding DNA will tell us very little.
What is more important to consider is the function of the target (regulatory vs. nonregulatory) and the effect the mutation has on it. While it may be true that mutations that have effects on gene regulation are likely to be associated with adaptation, at least in its early stages, there are real challenges associated with identifying a given gene as ‘regulatory’ or ‘nonregulatory’. The reason is that cellular networks are highly interconnected and seemingly nonregulatory genes can indirectly regulate other genes or pathways. As systems biology continues to integrate these networks into a more complete understanding of cellular function, we should be able to make more accurate classifications of a gene’s functional effects.
Another caveat is that the importance of regulatory vs. nonregulatory mutations may depend on the mechanism of the particular adaptive response, which in turn may depend on the selective environment. As an example, consider microbial experiments that have examined adaptation to nutrient limitation vs. those to novel substrates. With nutrient limitation, the organism already possesses pathways for nutrient uptake and utilization, so simply modifying their regulation would be a likely adaptive solution (e.g. quantitative changes in pathway induction). To deal with novel substrates, the evolution of a qualitatively new function may be required, suggesting that mutations in nonregulatory or structural genes would be favoured. Do the general mechanisms of adaptive response differ between these two classes?
There is some evidence to support this distinction, at least when we examine genes with known functions. In glucose- or phosphate-limited E. coli strains (Woods et al. 2006; Barrick et al. 2009; Maharjan et al. 2010; Wang et al. 2010), examples of such targets are a regulator of stress response (spoT), a starvation-specific transcription initiation factor (rpoS), an RNA chaperone (hfq) and a transcriptional repressor (nadR). Similar results have been seen in S. cerevisiae strains evolved under glucose-limitation (Anderson et al. 2010; Kvitek & Sherlock 2011). In contrast, mutations in nonregulatory or structural genes are often observed in strains experimentally adapted to novel substrates. Examples include mutations in a multi-drug efflux pump implicated in antibiotic and solvent tolerance (acrA) found in E. coli strains adapted to isobutanol (Minty et al. 2011) and an H+ efflux pump (PMA1) responsible for maintaining cytoplasmic pH and plasma membrane potential in high-salt adapted S. cerevisiae (Anderson et al. 2010). Nevertheless, a strong test of this hypothesis has not been conducted and may not be possible until we better understand the manifold functions of genes and their interactions.
Predictability of targets
Based on what we know about the genetic pathways involved in growth in different environments, it is often possible to make fairly good predictions as to what genes are likely to be involved in adaptation. For example, reducing the intracellular concentration of a stressor would be a predictable adaptive solution. In yeast, this is exactly what happened through the amplification of a cluster of salt efflux genes (ENA) in a salt-adapted strain (Anderson et al. 2010). Increasing the uptake of a limiting nutrient would also be a logical response, as has happened through the duplication of high affinity glucose transporters (HXT6/7) in glucose-limited strains (Kvitek & Sherlock 2011).
Despite a few successes, other experiments examining the evolution of more complex phenotypes often report causal mutations in inexplicable target loci. In the socially cooperative bacterium Myxococcus xanthus, a transition from a incompetent cheater to a superior cooperator was caused by just a single mutation in the promoter of a GNAT-family acetyltransferase gene (Velicer et al. 2006). In Pseudomonas fluorescens, the evolution of rapid, stochastic colony-morphology switching (i.e. bet-hedging, Beaumont et al. 2009) was caused by a single nucleotide change in the large subunit of carbamoyl-phosphate synthetase. How these mutations cause these phenotypes is completely unknown.
In general, the targets of selection during adaptive evolution are very difficult to predict from prior knowledge. This may be because we know much more about the components of genetic pathways than their regulation, and if much of adaptive evolution involves changes in regulation of gene expression, these would not be identified as potential targets. Probably, the most important insight that EE-WGS has provided is how much we do not know about the link between genotype and phenotype. But uncovering these unexpected targets of selection demonstrates how this approach can identify novel associations between genotype and phenotype and provides new information on the genes involved in particular physiological processes, metabolic pathways or cellular subsystems. Finding new genes that are integral to the function or regulation of specific pathways is particularly useful for practical applications such as discovering novel targets for drug development, preventing the evolution of antibiotic resistance and genetic engineering of desired phenotypes.
Parallel evolution and the variety of genetic routes to adaptation
How reproducible is the outcome of evolution? Parallel evolution is considered de facto evidence for natural selection, but in nature, typically only a single outcome can be observed for a specific evolutionary scenario. By analysing multiple independent bouts of evolution under identical conditions, EE is the most rigorous approach to distinguishing the contributions of genetic drift and natural selection to evolution. For EE-WGS studies with multiple replicates, we can ask whether parallelism occurs, to what extent, and at what biological level? Was selection strong enough to repeatedly target the same locus, or even nucleotide, in independent evolutionary realizations?
Level of biological organization
Intuition tells us that the extent of parallelism should decrease with the level of biological organization. Fitness, the character of ultimate concern in evolution, should show the most parallelism because, provided selection is capable of doing its job (that is, drift and mutation are weak relative to selection), fitness will inevitably increase. In most organisms and in most environments, there are probably many different combinations of phenotypes and even more combinations of genotypes that can generate high fitness. So we might expect that the extent of parallelism should decrease in a predictable order: phenotype > locus > nucleotide.
A number of studies have provided evidence for parallel phenotypic evolution (e.g. Lenski & Travisano 1994) and for parallel systematic changes in gene expression (e.g. Ferea et al. 1999; Cooper et al. 2003; Gresham et al. 2008) and protein profiles (Pelosi et al. 2006). Others have suggested the opposite: despite similar increases in fitness in response to the same selection environment, global gene expression patterns can be quite different (Fong et al. 2005). This last result is further supported by the observation that replicate selection lines that have all increased in fitness in the same environment often display substantial genotype-by-environment interaction variance for fitness across novel environments, implying that different loci are responsible for adaptation in different replicate lines (Travisano et al. 1995; Melnyk & Kassen 2011).
Candidate gene sequencing (Crozat et al. 2005, 2010; Woods et al. 2006; Ostrowski et al. 2008; Kinnersley et al. 2009) suggested that parallelism at the locus level was fairly common. In the one WGS study for which there is sufficient data, a much different picture emerges for parallelism at the locus level. In 11 lactate-adapted lines of E. coli (Conrad et al. 2009), 72% of the target loci were unique to a single line, and the most common target locus had mutations shared by only 7 of 11 lines. Similar results are seen in microarray-based studies (Herring et al. 2006; Gresham et al. 2008). This suggests that locus-level parallelism may be much less common than candidate gene approaches led us to believe. The likely reason for this is that most candidate genes are chosen because they are expected targets of selection, making any candidate list a biased sample of total targets. There may also be reporting bias because instances of parallelism typically are given more attention than cases of nonparallelism.
Parallelism at the lowest scale, the same nucleotide, is very rare. Models of adaptive evolution at the sequence level suggest that the probability of parallel evolution is 2/(n + 1), where n is the number of beneficial mutations available to a genotype (Orr 2005b). As n is usually very difficult, if not impossible, to know with precision, this prediction remains untested. Empirically, it is notable that even if a gene is under strong selection, mutations in different replicates typically occur at different residues within the same gene. We found only four cases of nucleotide parallelism in bacteria (pykF and nadR,Woods et al. 2006; kdtA and hfq, Conrad et al. 2010). Some of these loci showed high parallelism but the mutated residue was shared by, at the most, only three lines. We were unable to find any examples of nucleotide parallelism in yeast.
Magnitude of adaptive effect
One might expect that the degree of parallelism to be positively correlated with the magnitude of the adaptive advantage conferred by the mutation (i.e. mutations with larger fitness benefits would be under stronger selection and become substituted in more replicate lines, Orr 2005b). Data from studies that report ranges of parallelism and fitness estimates of individual mutations (Herring et al. 2006; Barrick et al. 2009; Conrad et al. 2010) provide mixed support for this prediction. While the targets with the most parallelism do tend to confer higher fitness benefits (Barrick et al. 2009), in some cases, the opposite is true. Conrad et al. (2009), for example, found that the two most common SNPs actually provided the lowest fitness increases of all the single SNPs tested. This observation is likely explained by epistasis; the mutations require the presence of other co-acquired mutations and therefore are only conditionally beneficial.
Adaptation is constrained within the limits of an organism’s biological functions, so the number of possible routes, and their accessibility, may be governed in part by particular characteristics of the organism itself. For example, the frequency of parallelism, both at the locus and nucleotide level, appears to scale inversely with organismal complexity (virus < bacteria < yeast). While very common in viruses, it is less common in bacteria and even rarer in yeast. If a viral genome has only a handful of genes, the paucity of adaptive targets will lead to high levels of parallelism, as documented even at the nucleotide level (Wichman & Brown 2010; Miller et al. 2011). As biological complexity increases, so does genome size and the number of potential adaptive targets. More importantly, increases in genome complexity and the interconnectedness of regulatory and metabolic networks are hypothesized to increase the potential for pleiotropy and epistasis (Sanjuan & Elena 2006; Sanjuan & Nebot 2008), which can translate into more adaptive possibilities. Yeast may show less parallelism simply because more possible adaptive solutions to a given problem are available (i.e. more options = less parallelism).
Does this pattern extend beyond unicellular microbes? It is too early to tell. Some evolution experiments have been performed with complex, multi-cellular macrobes (e.g. Burke et al. 2010; Denver et al. 2010), but not enough to compare relative levels of parallelism. Moreover, these experiments all start with large amounts of standing genetic variance rather than relying on mutations arising de novo during the experiment, involve much smaller population sizes, and are often obligately sexual. These differences make direct comparisons with the results of microbial experiments challenging. As additional evolution experiments are performed with higher complexity organisms, it will be interesting to see how the conclusions compare with those from the microbial body of work.
General adaptive responses
Occasionally, parallel locus evolution is observed among different selective environments, indicating the mutations confer a general adaptive benefit that is not specific to any particular environment. For example, mutations in the rph-pyrE operon were observed in lines of E. coli adapted to lactate (Conrad et al. 2009), glycerol (Herring et al. 2006) and isobutanol (Minty et al. 2011). Other examples of cross-environment locus parallelism in E. coli were observed with an outer-membrane porin (ompF; Shendure et al. 2005; Barrick et al. 2009) and glycerol kinase (glpK; Herring et al. 2006; Kinnersley et al. 2009). By comparing the distribution of mutations among highly replicated lines in different environments, within- and between-environment components of variation in adaptive evolution could be quantified, helping to distinguish between environment-specific and general adaptive factors. It is likely that adaptation to laboratory culture conditions is driving at least some nonspecific parallel evolution, but other cases may represent a common adaptive response to a variety of different environments (see next section).
Does something similar happen in eukaryotic microbes? To answer this question, we surveyed EE target loci in yeast (Anderson et al. 2010; Araya et al. 2010; Kvitek & Sherlock 2011) and found several genes that are clearly associated with the Spt-Ada-Gcn5-acetyltransferase (SAGA)-mediated stress response pathway. In yeast, the SAGA transcription factor complex is functionally similar to bacterial σS (Huisinga & Pugh 2004). Some target genes encode subunits of the SAGA complex itself (SGF73, TAF5), while others help recruit SAGA to promoters (CYC8). Other stress-related targets function to sense and relay nutrient and stress signals (CCR4-NOT) or coordinate the arrest of cell growth upon nutrient depletion (WHI2). We also found examples of target loci that can regulate eukaryotic gene expression in the chromatin-remodelling, transcriptional, post-transcriptional, translational, and post-translational stages.
These results suggest that mutations in stress response genes may be a very general feature of adaptive evolution in laboratory populations of microbes. Why might this be? Microbes typically live in dynamic environments where physical stressors and nutrient availability fluctuate through time, and responding rapidly to these changes is critical to survival. Under stressful conditions, stress response genes are activated at the expense of general housekeeping and growth functions because transcriptional cofactors such as σS and SAGA are preferentially recruited by RNA polymerases. When the stress is relieved, stress response genes are de-activated and housekeeping and growth functions can resume. There is thus a shifting balance between stress response and growth under ‘typical’ conditions for microbes in nature.
Laboratory conditions typically represent stressful environments, so the physiological response of most microbes is to activate stress responses. But because the environment remains fairly constant from one generation to the next, this stress response becomes overly costly, and any mutation that ratchets it down, and so frees up energy for growth, is likely to be favoured. This explains why mutations that reduce stress responses, which may at first seem counterintuitive, seem to be common in most microbial evolution experiments (Philippe et al. 2007; Conrad et al. 2010; Wang et al. 2010). It also suggests that selection in an environment that alternates between stress and no stress would lead to many fewer mutations in stress response genes compared to what we would see under constant, stressful conditions.
Relevance to nature and comparative genomics
Experimental evolution is sometimes criticized for being too far removed from the natural world, which is too complex to be modelled accurately by examining adaptation in test tubes. Although we are inclined to disagree, it is perhaps more diplomatic to say that the issue is an empirical one. WGS can take us some way towards providing an answer.
To the extent that the components of selection in the laboratory and in nature overlap, and the experimental conditions successfully model the important parameters of the natural environment, then the resulting adaptive mutations are likely to be ecologically relevant. We can then ask whether the mutations we observe to be selected in the laboratory are also seen segregating in natural populations. There is an important caveat here, though, just because a mutation segregates in natural populations does not mean it has been under selection. Comparative genomics of naturally occurring strains provides a catalogue of existing mutations (Liti et al. 2009; Schacherer et al. 2009; Croucher et al. 2011; Luo et al. 2011) but their evolutionary significance can only be inferred because knowledge of the environmental, ecological and selective histories of extant lineages will invariably be incomplete.
There are a number of examples where laboratory and natural populations have similar or the same mutations. The bacterial starvation sigma factor gene (rpoS), for example, is a common experimental target of selection and shows similar polymorphism in natural isolates (Notley-McRobb et al. 2002; Dong et al. 2009). In yeast, premature stop codon mutations in the MTH1 and RIM15 genes were observed in both experimentally derived and naturally occurring strains (Liti et al. 2009; Kvitek & Sherlock 2011). Anderson et al. (2010) also found that an experimentally evolved adaptive mutation in the MKT1 gene was identical to a nucleotide polymorphism found in a range of wild strains (Liti et al. 2009) and large-scale functional genomic studies analysing differences in quantitative traits and gene expression among genetically diverse yeast strains identified MKT1 as major determinant of phenotypic and gene expression variation (Deutschbauer & Davis 2005; Sinha et al. 2006; Zhu et al. 2008; Lee et al. 2009). Polymorphisms in MKT1 thus seem to significantly impact fitness in both experimental and natural strains.
The evolution of antibiotic resistance is a prime example of where in vitro selection can target genes that are also targeted in natural isolates. Surveys of clinical strains of several bacterial species have found that resistance to quinolone antibiotics is commonly conferred by mutations in one of the topoisomerase genes (e.g. gyrA, gyrB, parC; Piddock 1999; Wong & Kassen 2011). Mutations similar, or even identical, to these naturally occurring mutations are frequently found in short-term, in vitro experiments that selected for resistance to quinolones (De Vecchi et al. 2009; Spigaglia et al. 2009; Drago et al. 2010; Zhang et al. 2011). It should be noted that these topoisomerase genes are known target sites for this class of antibiotics, so this represents a special case with an over-simplified adaptive landscape. Nonetheless, it still exemplifies parallel evolution between in vitro and natural settings and provides cross-validation for both approaches.
The merging of EE with WGS has provided, as we have shown, a number of new insights into the molecular evolution of adaptation. To our minds, there are four main avenues of investigation that require further study.
The first is to understand in more detail the dynamics of adaptation. What has become increasingly clear from theoretical and empirical studies in microbial evolution over the past 10 years or so is that adaptation, even in the relatively simple environments of a laboratory microcosm, is a far more complex process than we had originally thought. The classic model of adaptation sees a population as genetically uniform most of the time, and polymorphism, when it does exist, is transient, occurring only as the mutation is substituted. This model is almost surely wrong. Genetic variation is much more common in evolving populations because mutation supply rates (the combination of mutation rate and population size) are often high (Lang et al. 2011), there may be multiple genetic routes to the same adaptive endpoint, or polymorphism may even be actively maintained by diversifying or balancing selection (Rozen & Lenski 2000; Maharjan et al. 2006; Kinnersley et al. 2009; Rozen et al. 2009; Wang et al. 2010; Kvitek & Sherlock 2011).
The bulk of EE-WGS studies, however, have applied their sequencing efforts to single, endpoint clones that emerge victorious from the selected population. This black box approach (genome in → genome out) focuses only on the outcome of evolution and does little to address all the underlying population-level dynamics. Future studies should be designed to sample this variation explicitly, for example, by WGS of entire populations rather than individual clones, and on much finer timescales. Sequencing the entire population as a whole is the only practical approach to characterizing the amount of population-level genome diversity and how it changes through time. This approach has been employed in only a few studies (Barrick & Lenski 2009; Paterson et al. 2010), and additional controlled proof-of-principle experiments would be useful for establishing how these methods perform under various conditions. As for all current NGS technologies, issues regarding sequencing error, sequencing bias, sensitivity of rare allele detection and accuracy of allele frequency estimation need further characterization (Harismendy et al. 2009; Shen et al. 2010).
We feel that the way forward will be to combine EE-WGS data with the integrative powers of systems biology (Papp et al. 2011). Genes should not be viewed as independent units, but rather as components of complex genetic networks with finely regulated and coordinated levels of expression. Genome-level genetic interaction maps (Costanzo et al. 2010; Szappanos et al. 2011) demonstrate how most cellular bioprocesses are functionally interconnected and may help to reveal how mutational perturbations propagate through the molecular networks of the cell. In the process, we may be able to make stronger statements about how the genetic architecture of an organism impacts its evolutionary potential, for example, by asking whether the placement and position of a gene in a network affects its evolvability. Even for the most-studied organisms, however, we are far from a system-level understanding of how epistasis, pleiotropy, modularity and redundancy all interact to determine the balance between robustness and evolvability (Wagner 2008).
Third, we need to understand better how important adaptive evolution is in structuring natural populations of microbes. What we learn in the laboratory also has practical implications with respect to drug resistance and evolution of pathogenic microbes. For example, recent studies of chronic infections in the lungs of cystic fibrosis patients have revealed that the bulk of early mutations appear to be adaptive (Smith et al. 2006; Yang et al. 2011). It may also help us understand global climate change scenarios by allowing us to ask whether and how populations might adapt to changing environmental conditions (Bell & Collins 2008). Integrating the results of laboratory experiments with WGS of natural populations can take us further towards developing good predictive models for addressing these more practical problems.
Lastly, there is plenty of scope for incorporating more environmental and ecological complexity into our laboratory models, and seeing how this effects genome evolution. How does the strength of selection, and fluctuations in selective types and pressures, affect adaptive dynamics? What characteristics of the environment are most relevant for predicting adaptive outcomes? Ecological complexity can be added by co-evolving multiple species together (see Paterson et al. 2010). Further experiments will reveal how interspecific interactions such as competition, predation and parasitism change the evolutionary dynamics of the system at a molecular level.
WGS of EE strains will likely become, if it has not already, commonplace as the rate and affordability of high-throughput genome sequencing continue to increase. As the raw sequencing power of NGS technologies advances, so will its impact on most areas on modern biological research. Given the massive amount of sequence data generated and the associated computational challenges, entering into this field can be a daunting prospect. We have included a methodological outline that will hopefully prove useful to those research groups interested in employing the EE-WGS approach (Box 2, Appendix S1, Supporting Information) and contributing to this rapidly expanding field.
EE-WGS studies have already provided much insight into the dynamics and processes of evolution, but the field is still in its infancy. The specifics of evolutionary dynamics likely vary with the environment and organism under study so much more work is needed before drawing firm conclusions regarding the general rules of adaptive evolution. Nonetheless, the principal patterns of genomic substitution we have extracted from studies of short-term adaptation of microbes are still meaningful and will provide a solid framework for future research. One challenge that awaits is to translate and apply the lessons learned from these laboratory EE-WGS studies to patterns of genomic and phenotypic variation found in naturally occurring microbes. It also will be interesting to see if the fundamentals of microbial evolution can be extrapolated to adaptation of more morphologically and physiologically complex organisms.
Funding was provided by the Natural Sciences and Engineering Research Council of Canada. N.R. was also supported by a post-doctoral fellowship from the Quebec Centre for Biodiversity Science.
J.R.D. studies the evolutionary genetics and genomics of microbes. N.R. is a computational biologist interested in molecular evolution, with a focus on protein-coding genes. A.H.M. is an evolutionary biologist interested in the genetics underlying adaptation and speciation. A.W. is an evolutionary biologist interested in the genetic basis of adaptation. S.F.B. studies how ecological processes, such as competition, drive adaptation and diversification. R.K. is an evolutionary biologist with interests in the genetics of adaptation and diversification.