Microbial laboratory evolution in the era of genome-scale science.

Laboratory evolution studies provide fundamental biological insight through direct observation of the evolution process. They not only enable testing of evolutionary theory and principles, but also have applications to metabolic engineering and human health. Genome-scale tools are revolutionizing studies of laboratory evolution by providing complete determination of the genetic basis of adaptation and the changes in the organism's gene expression state. Here, we review studies centered on four central themes of laboratory evolution studies: (1) the genetic basis of adaptation; (2) the importance of mutations to genes that encode regulatory hubs; (3) the view of adaptive evolution as an optimization process; and (4) the dynamics with which laboratory populations evolve.

Introduction

Due to the lack of experimental approaches, the study of the process of evolution has been mostly limited to theoretical studies. Thus, this field is rich with abstract concepts that were for a time difficult or impossible to test experimentally. Recent advances in DNA sequencing, high-throughput technologies, and genetic manipulation systems have enabled empirical studies that directly characterize the molecular and genomic bases of evolution. Specifically, these tools allow the direct measurement and analysis of short-term evolutionary processes, and thus the testing of conceptual evolutionary models and theories (Wagner, 2008).

Microbes are well suited to adaptive laboratory evolution (ALE) studies for a number of reasons: fast generation times, repeatability, the ease of maintaining large population sizes, and the ability to store populations for later examination (Elena and Lenski, 2003). In the wild, evolution has diverse causes such as environmental change or isolation of small populations. ALE studies, also referred to as experimental evolution studies, seek to observe microbes under controlled scenarios in which evolution is expected to occur. While the controlled environments may not always precisely represent those in the wild, ALE studies have proven successful in grounding evolutionary theories to the actual molecular and mechanistic bases of evolution. In this review, we will highlight a few central lessons obtained through microbial ALE studies thus far. First, genomic sequencing can determine the complete set of mutations responsible for an evolved phenotype, and has led to the discovery that interactions between these mutations are very common. Second, adaptive mutations frequently target regulatory mechanisms. Third, principles of systems-level optimization underlie the genetic changes seen in adaptive evolution, and with a systems-level understanding, these optimization principles can be harnessed for the purposes of metabolic engineering. Fourth, mutant sub-populations of improved fitness invariably arise in growing populations, but their dynamics in the population are complicated due to factors such as natural selection, clonal interference, drift, and frequency-dependent selection.

Full genomic changes occurring during ALE can be directly observed

Darwin's theory of evolution preceded any knowledge of the molecular mechanisms of heredity, and the development of evolutionary theory since has frequently omitted these mechanisms. Yet, the molecular details are at the heart of the evolutionary process and, thus, of central interest. The development of next-generation sequencing technology has at last provided the means to study the molecular basis of evolution on a genome scale (Metzker, 2010; Brockhurst et al, 2011). Many recent studies have utilized next-generation technologies to find mutations on a genome-wide basis (Albert et al, 2005; Friedman et al, 2006; Herring et al, 2006; Velicer et al, 2006; Gresham et al, 2008; Barrick et al, 2009; Conrad et al, 2009; Araya et al, 2010; Atsumi et al, 2010; Charusanti et al, 2010; Kishimoto et al, 2010; Lee and Palsson, 2010; Lee et al, 2010).

Mutations found by whole-genome resequencing (WGS) of several non-mutator evolved clones from E. coli ALE experiments as compared with the sequence of the starting strain are summarized in Figure 1. In the studies summarized in Figure 1 (Herring et al, 2006; Barrick et al, 2009; Conrad et al, 2009; Charusanti et al, 2010; Kishimoto et al, 2010; Lee and Palsson, 2010), single-nucleotide mutations were the most common type of mutation observed (61% of observed mutations) and c>t/g>a mutations were disproportionally observed (48 of 111 SNPs). Deletions (29%), insertions (7%), and insertion sequence movements (3%) were also commonly observed.

Figure 1

Intragenic mutations identified in E. coli ALE studies. (A) Single-nucleotide substitutions, insertions, and deletions found within the open reading frames by whole-genome sequencing in multiple E. coli ALE studies (Herring et al, 2006; Barrick et al, 2009; Conrad et al, 2009; Charusanti et al, 2010; Kishimoto et al, 2010; Lee and Palsson, 2010) are shown on a circular representation of the E. coli chromosome. (B) The set of genes displayed on the E. coli chromosome was subjected to enrichment analysis for Gene Ontology Slim (GOslim) categories (Camon et al, 2004). Wedges that protrude outward represent statistically enriched GOslim categories (also marked by ** in the legend). (C) Genes that were mutated in multiple studies are shown. 20K=growth on glucose minimal medium for 20 000 generations (Barrick and Lenski, 2009), 45A=adaptation to high temperature (Kishimoto et al, 2010), ETM=adaptation of ethanol tolerance (Goodarzi et al, 2009), Glyc=growth on glycerol minimal medium (Herring et al, 2006), Lact=growth on lactate minimal media (Conrad et al, 2009), and PGI=growth on glucose minimal media following the deletion of pgi (Charusanti et al, 2010).

WGS of ALE strains has uncovered the single-nucleotide substitutions, small insertions and deletions (indels), large genomic duplications (Herring et al, 2006; Conrad et al, 2009), large deletions (Barrick et al, 2009; Charusanti et al, 2010), and transposable element insertions (Gabriel et al, 2006; Gresham et al, 2008; Atsumi et al, 2010) acquired during adaptation. WGS of several strains evolved in replicate shows the extent to which evolutionary genetic paths are constrained. Parallel evolution also reveals the mutations conferring the largest benefit, since only the few mutations conferring large fitness gains should be repeatedly selected for in multiple replicates (Cooper et al, 2003; Elena and Lenski, 2003; Herring et al, 2006; Conrad et al, 2009; Charusanti et al, 2010). In another approach, WGS has been applied in a longitudinal manner, in which the genome was sequenced at various time points. This approach was applied to E. coli evolved in glucose minimal medium for 40 000 generations (Barrick et al, 2009), and to E. coli adapted to growth in successively higher temperatures for 7500 generations (Kishimoto et al, 2010). In both cases, the rate of mutation appearance increased in the later phase of the adaptation due to the development of a mutator phenotype. Surprisingly, the E. coli strain grown in glucose minimal medium acquired beneficial mutations at a nearly linear rate over 20 000 generations before obtaining a mutator phenotype (Barrick et al, 2009).

Through the use of allelic replacement, the fitness conferred by a mutation can be assessed using growth rate measurements or competition experiments (Applebee et al, 2008). Allelic replacement experiments have shown that most mutations found in non-mutator populations are beneficial when introduced to the ancestral strain (Cooper et al, 2003; Herring et al, 2006; Giraud et al, 2008; Conrad et al, 2009; Goodarzi et al, 2009; Auriol et al, 2011; Tremblay et al, 2011). In addition to enhanced fitness, single allelic replacement has shown other genotype–phenotype relationships of adaptive mutations. For instance, it helped to link a mutation of a gene involved in Wsp signaling to the ‘wrinkly spreader' phenotype in Pseudomonas fluorescens (Knight et al, 2006; Bantinaki et al, 2007), to show contribution of mutations to the spoT regulatory gene to expression patterns found in evolved E. coli (Cooper et al, 2003), and to show the effect of an adaptive mutation to a Geobacter sulfurreducens gene encoding a periplasmic cytochrome on Fe2O3 reduction (Tremblay et al, 2011). Furthermore, allelic replacement has been used to iteratively reconstruct the complete multiple-mutation genotype found by WGS of the adapted strain. Complete recovery of the adapted phenotype in the reconstructed strain suggests that all relevant mutations have been identified (Herring et al, 2006; Conrad et al, 2009; Goodarzi et al, 2009; Lee and Palsson, 2010).

Allelic replacement experiments have shown that some mutations exhibit a neutral or deleterious effect on fitness when introduced into the ancestral background, but appear to confer positive fitness in the presence of other mutations acquired during adaptation. These interactions between mutations are examples of epistasis (non-additive or non-multiplicative fitness effects of multiple mutations) and appear to be a very common feature observed in ALE and genetic interaction studies (Dykhuizen and Hartl, 1980; Herring et al, 2006; Spencer et al, 2007; Applebee et al, 2008; Beaumont et al, 2009; Conrad et al, 2009; Charusanti et al, 2010; Costanzo et al, 2010; He et al, 2010; Lee and Palsson, 2010). Epistatic mutations are sometimes compensatory, providing a remedy for complications introduced to the cell by earlier adaptive mutations (Maisnier-Patin and Andersson, 2004). Alternatively, sometimes the fitness attributed to the mutation is positive only in a fast growing cell (Chou et al, 2009). Epistatic interactions can severely constrain viable evolutionary trajectories, since there may be relatively few trajectories in which fitness increases monotonically with each single mutation (Weinreich et al, 2006; Poelwijk et al, 2007).

WGS and allelic replacement experiments determine the genetic basis of evolution. However, in the cases where many mutations are found (e.g., in populations with high mutation rates), it is difficult or unrealistic to iteratively construct allelic replacement collections. Two recent methods may help to overcome this limitation. First, multiplex automated genome engineering provides an inexpensive and high-throughput platform for introducing specific sequence diversity into the population (Wang et al, 2009). Under this approach, a set of synthetic oligonucleotides representing the mutations found by WGS randomly undergoes recombination with genomic DNA, creating a library of allelic replacement strains. Second, array-based discovery of adaptive mutations (ADAM) uses genetic foot printing to selectively identify mutations that provide a competitive advantage to the cell (Goodarzi et al, 2009). When reconstructing phenotypes through allelic replacement, ADAM allows the researcher to avoid the wasted effort of reconstructing non-beneficial mutations. These methods have the potential to accelerate and increase the scope of ALE studies.

Broad and systemic regulatory changes found in ALE strains

Microarrays have been used extensively to measure the global gene expression phenotype of evolved bacteria. Routine analyses exist to retrieve information from the large, complex data sets generated by these experiments (Eisen et al, 1998; Cui and Churchill, 2003). Clustering can group genes based on similar expression patterns across multiple samples. Genes showing statistically significant differential expression can further be analyzed for enrichment within functional categories or established regulons. Sometimes, the significant gene functions that come out of these analyses have obvious relevance to the strain in the adaptive growth condition, while other times the meaning of the expression changes is hard to interpret based on our current understanding of the function of the mutated gene product. Measurement of expression changes in multiple replicate evolved lines can identify the changes that are more likely to be relevant to adaptation by finding genes that are consistently differentially expressed. Likewise, fitness profiling using gene overexpression and gene knockout libraries can provide coarse-grain determination of the specific gene expression changes that are adaptive for a strain (Goodarzi et al, 2010).

Sometimes, a large degree of parallelism is observed in the gene expression changes of replicate evolved populations (Cooper et al, 2003; Fong et al, 2005b; Stoebel et al, 2009). Sequencing of these populations shows that similar regulatory genes acquire mutations within replicate populations. A single mutation to a regulatory hub can result in broad-scale changes to the expression state of cells in the mutated population (Philippe et al, 2007). In general, such mutations of large effect were expected to be deleterious (Fisher, 1930). However, ALE experiments have shown that mutations to regulatory hubs can provide a large increase to fitness (Cooper et al, 2003; Crozat et al, 2005; Herring et al, 2006; Giraud et al, 2008; Conrad et al, 2010). Due to their large benefit, mutations to regulatory hubs are common and are frequently found in replicate populations (Figure 1B and C).

One of the most central transcriptional regulatory hubs is likely the RNA polymerase (RNAP) itself (Klein-Marcuschamer et al, 2009). ALE studies subjecting the E. coli K-12 MG1655 strain to continuous exponential growth in minimal media have found that the genes encoding the RNAP core enzyme are frequently mutated (Friedman et al, 2006; Herring et al, 2006; Conrad et al, 2009, 2010; Charusanti et al, 2010; Figure 1C). Mutations to the rpoC gene, encoding the RNAP β′ subunit, were found in ∼80% of MG1655 strains that were adapted to glycerol M9 minimal media (Conrad et al, 2010). These rpoC mutations resulted in a 60% increase to growth rate in glycerol minimal medium when introduced to the ancestor strain, while simultaneously changing the expression pattern of 20–27% of genes in the genome. These genes fell into certain functional classifications, suggesting that elements of the RNAP have transcription factor-like activity. In vitro assays indicated that separate adaptive rpoC mutations resulted in similar changes in transcription kinetics of the mutants, including decreased open-complex half-life at a sensitive promoter, decreased pausing half-life, and increased elongation rate. Thus, these mutations to the core enzyme may affect broad and complex changes to the transcriptome by altering the transcription kinetics.

In bacteria, sigma factors bind core RNAP, providing preference for specific promoters and assisting in open-complex formation. There are several different types of sigma factors that compete with each other for binding to free RNAP core enzyme to form the holoenzyme. Each sigma factor recognizes different sets of promoters, and thus affects the global transcription profile of the bacteria. The rpoS gene of E. coli, which encodes the stationary phase sigma factor and is needed for the global stress response (Weber et al, 2005), is frequently affected by mutations (Notley-McRobb et al, 2002; Cooper et al, 2003; Ferenci, 2008; Spira et al, 2008; Conrad et al, 2009, 2010; Charusanti et al, 2010; Wang et al, 2010). Whereas binding of the RpoS sigma factor to the RNAP core enzyme results in a global stress response, the predominant sigma factor during exponential growth is Sigma 70, which is needed for transcription of many metabolic and growth-associated genes. The competition between RpoS and Sigma 70 for RNAP core enzyme sets up a tradeoff between stress survival and robust growth that has been termed the ‘self-preservation and nutritional competence' (SPANC) balance (Ferenci, 2005). Therefore, mutations to rpoS and its regulators can serve to adjust the SPANC balance by reducing or eliminating the size of the RpoS pool that is able to bind RNAP core enzyme. Under conditions of strong growth pressures and weak environmental stress, such mutations are expected to be favorable.

These examples of regulatory hub mutations demonstrate that while the transcriptional regulatory network of E. coli is not wired to optimally respond to all environments, the network structure allows a single mutation to a hub to rewire the network into a more efficient state. While the process of discovering these mutations by WGS and introducing them into the ancestor through allelic replacement has become routine, careful characterization of the genotype–phenotype relationship of mutations to regulatory genes is difficult due to the large scale and complexity of phenotypic consequences. Understanding can be further impeded by unknown secondary functions of the mutated gene. These challenges to uncovering the genotype–phenotype relationship of mutations to regulatory hubs can to some extent be overcome using genome-scale fitness profiling (Goodarzi et al, 2010) and computational models (Lewis et al, 2010). The hard work of uncovering the genotype–phenotype relationship may be rewarded by insights into secondary and unknown functions of the mutated gene (Copley, 2003; James and Tawfik, 2003). ALE can thus serve as a discovery tool where the hypotheses about function are generated through the evolutionary process.

Cellular optimization in ALE

Models of evolution frequently invoke the concept of adaptive landscapes (Orr, 2005). These are often depicted graphically as three-dimensional landscapes in which the sequence space serves as a plane on which to map the fitness of each genotype. Adjacent points on this plane represent similar sequences. The landscape can be smooth (Figure 2A), with a single optimum, or rugged (Figure 2B), with multiple optima. In a rugged fitness landscape, accessible optima may depend on the starting point within the landscape. Thus, there may be a diversity of evolutionary outcomes. Results of ALE experiments suggest that bacteria encounter both smooth landscapes, in which the population is constrained to acquire specific mutations in an exact order (Weinreich et al, 2006; Poelwijk et al, 2007; Stoebel et al, 2009), and rough landscapes, in which WGS shows few patterns in the genes that acquire mutations among replicate populations (Gresham et al, 2008; Conrad et al, 2009; Wang et al, 2010). The continual fitness improvement of E. coli over tens of thousands of generations of adaptive evolution, albeit with diminishing returns, suggests that populations may only asymptotically approach local optima on the fitness landscape (Barrick et al, 2009).

Figure 2

Optimality principles in adaptive evolution. (A) A smooth fitness landscape consisting of a single peak. Circles represent points on the landscape and arrows indicate the pathway of genetic change through the landscape. On a smooth landscape, there is a tendency for evolutionary convergence toward the single optimum, regardless of the starting point on the landscape. (B) A rough fitness landscape consists of multiple peaks. With multiple optima, there tends to be evolutionary divergence, sometimes even when starting from the same location on the landscape. (C) A phenotypic phase plane is a representation of how two fluxes in a metabolic network relate to each other and affect in silico-predicted optimal growth. Distinct planes are represented by several colors. Here, the line of optimality (LO, yellow) defines the ratio of glycerol uptake rate to oxygen uptake rate that leads to optimal biomass production. On glycerol, wild-type E. coli initially has a phenotype that maps to a suboptimal region of the portrait. After a growing for several hundred generations on glycerol, the E. coli phenotype migrates to the line of optimality. (D) Optimality principles can used to design strains of bacteria in which growth at the maximal rate requires the secretion of a product of interest. When subjected to ALE, the designed strains increase both their growth rate and product secretion rate. The two colored regions indicate accessible flux states before and after evolution.

The concept of traversing the fitness landscape depends on the assumption that cells will modify their genotype to optimize their fitness in the given environment. Therefore, insight into cellular optimization is gained through the use of mathematical models. When the organism's metabolic network and biomass composition are known, flux-balance analysis (FBA; Orth et al, 2010) and other constraint-based modeling methods can be used to predict optimal phenotypes and their corresponding flux distribution based on measured nutrient uptake rates. Essentially, FBA calculates the maximum rate that biomass can be produced given the uptake rates of nutrients. FBA results can be presented on phenotypic phase-plane (PhPP) plots of model-predicted optimal biomass production (growth rate) versus carbon source uptake rate (SUR) and oxygen uptake rate (OUR). The graph of optimal OUR as a function of SUR is linear under physiological constraints. When mapped on to the PhPP, this line defines the line of optimality (LO) describing the most efficient ratio of SUR and OUR for biomass synthesis. Under several conditions, the experimentally measured E. coli phenotype corresponds to the LO of the PhPP (Edwards et al, 2001). When E. coli growth is not consistent with the LO, populations migrate toward the LO through adaptive evolution (Ibarra et al, 2002; Figure 2C). Migration toward the model-predicted optimal phenotype has also been observed in Lactobacillus plantarum adapting to growth in glycerol (Teusink et al, 2009). These studies show that cells will optimize their metabolic efficiency consistent with the demands from the selective pressure.

Achieving an optimal metabolic flux state requires sufficient expression of enzymes, yet proteins are costly for the cell to express (Kurland and Dong, 1996; Scott et al, 2010) due to limited quantities of free RNAP enzymes and ribosomes in the cell (Stoebel et al, 2008). Therefore, there is expected to be an optimal protein expression pattern. The principle of optimization of protein expression is exemplified by a cost-benefit analysis of Lac operon expression in the presence of lactose (Dekel and Alon, 2005). The mathematical model developed in this study shows that the optimal Lac expression level varies with the lactose concentration in the media. ALE experiments revealed that strains that initially expressed suboptimal levels of LacZ for a given lactose concentration rapidly evolved to produce the predicted optimal amount of LacZ, demonstrating optimization of gene expression for a single operon. Efforts have also been made to show economy of protein expression on the genome scale. Constraint-based modeling analyses of ALE outcomes have shown that evolved strains exhibit a general pattern of increased expression of genes and proteins associated with the optimal flux distribution, and decreased expression of genes and proteins associated with unused pathways (Becker et al, 2007; Lewis et al, 2010). Thus, the optimization of expression levels is a significant goal of adaptation.

The concept of evolution toward optimal phenotypes has a practical application in metabolic engineering. Strain engineering aims to coerce a cell to produce large amounts of product. However, this often inhibits growth. As a consequence, production strains often lose their ability to excrete product as the cellular fitness improves. Algorithms have been developed to predict gene deletions that couple fast growth with metabolite excretion (Burgard et al, 2003; Pharkya et al, 2004; Patil et al, 2005; Lun et al, 2009; Kim and Reed, 2010). In these growth-coupled designs, the metabolic pathways of an organism are altered such that the predicted optimal growth rate requires the secretion of the desired product (Figure 2D). Thus, adaptation for a faster growth rate can result in higher product yields, and the production strains become more stable. Many growth-coupled designs have been predicted in silico (Burgard et al, 2003; Bro et al, 2006; Feist et al, 2010) and improved product yield following the adaptation has been observed in designs for lactate and ethanol producers in the laboratory (Fong et al, 2005a; Trinh and Srienc, 2009).

Currently, the most extensive and detailed optimization models used to assess ALE experiments are constraint-based models (Feist et al, 2009). However, these focus primarily on metabolism, which accounts for less than one third of the genes in E. coli. Therefore, mechanistic conclusions drawn from metabolic models alone might be incomplete, since they do not directly include important biochemical mechanisms such as regulation, cellular maintenance and protection, and other non-metabolic functions. Efforts are being made to incorporate the functions of transcriptional regulation (Gianchandani et al, 2006; Chandrasekaran and Price, 2010) and the transcription/translation machinery (Thiele et al, 2010), thereby increasing the genome coverage by these models. In addition, alternative modeling frameworks are also being developed to model multiple cellular processes beyond metabolism (Covert et al, 2008; Molenaar et al, 2009). However, mutations in regulatory proteins still will pose a challenge regardless of the modeling framework. Mutations may alter the binding kinetics and motifs of these proteins, thereby changing the topology of the regulatory network. Despite this potential challenge, the assessment of high-throughput data in the model context may allow for the elucidation of the topological changes through an analysis of genes and proteins that failed to behave as predicted by the optimization models. Therefore, the use of optimization models in the assessment of ALE experiments will complement the analysis and provide hypotheses to possibly guide the difficult assessment of individual mutations.

Mutational dynamics and diversity within evolving populations

When placed into a challenging environment, natural selection acts on bacteria to select for populations harboring mutations that provide improved growth or survival. ALE studies have addressed the dynamics with which these mutations are acquired and propagated through a population. In many cases, there are expansions of diversity as additional beneficial mutations appear in the population, followed by contractions in which a sub-population of superior fitness eliminates the inferior alleles. In other cases to be discussed, mechanisms for stable population diversity have been observed.

Mutations provide the raw material for population diversity and adaptation. Thus, the mutation rate has profound effects on population diversity and evolution dynamics. A low mutation rate will result in slow discovery of adaptive mutations, while a fast mutation rate will increase the rate of occurrence of both deleterious and beneficial alleles, and result in more diverse populations. Mutator phenotypes frequently appear during adaptation and are characterized by an increased mutation rate, including a higher fraction of synonymous mutations (Sniegowski et al, 1997; Gresham et al, 2008; Barrick et al, 2009; Kishimoto et al, 2010). Mutator phenotypes may be transiently useful for accelerated discovery of a beneficial mutation, allowing the mutator allele to ‘hitchhike' during an adaptive sweep. However, they may be harmful over the long term due to accumulation of deleterious mutations (increased genetic load) (Giraud et al, 2001; Elena and Lenski, 2003). Furthermore, the mutation rate varies with environment and exposure to antibiotics (Notley-McRobb et al, 2003; Sakai et al, 2006; Kohanski et al, 2010). Frequent DNA damage or oxidative stress due to environmental stresses could provide a mechanism by which the mutation rate increases during a stressful condition (Notley-McRobb and Ferenci, 1999b; Notley-McRobb et al, 2003).

The mutation rate of a bacterial strain can be measured by a fluctuation test (Luria and Delbruck, 1943), in which the mutation rate is approximated by frequency of phage resistance or antibiotic resistance in an initially sensitive population, or by a mutation accumulation experiment (Kibota and Lynch, 1996; Lind and Andersson, 2008), where serial passage of single cell bottlenecks is used to measure the rate of appearance of mutations in a lineage over time. However, rates measured by either of these methods greatly over-approximate the rate at which mutations appear at a detectable frequency in laboratory evolution cultures (Lenski and Travisano, 1994). This discrepancy points to the importance of mutational dynamics within the population. First, many mutations are deleterious and will often be removed by selective pressures from the culture before reaching detectable frequency. Second, in large bacterial populations, multiple sub-populations can exist that possess differing beneficial alleles. When one sub-population gains and maintains a substantial fitness advantage over the rest of the population, it can increase its fraction in the population and even sweep the population (Muller, 1932), eliminating other sub-populations completely or to undetectable levels in a phenomenon known as clonal interference. Clonal interference may also impose a ‘speed limit' on adaptation, in which an increase in the mutation only results in an increase in adaptation rate to a certain point, after which selection, not the supply of beneficial mutations, becomes rate limiting (Arjan et al, 1999). Finally, many laboratory evolution experiments include ‘bottlenecks' in which a random fraction of the population is transferred to fresh medium to maintain exponential rates of growth. In serial passage experiments, the fraction of the population discarded tends to be very high. Thus, even beneficial mutations are routinely lost to drift. In chemostat experiments, cells are discarded continuously as quickly as new cells replicate. Therefore, the experiments are free of bottlenecks, with higher numbers of mutants remaining in the population.

To demonstrate the effect of drift and clonal interference in a serial passage experiment, consider E. coli grown in glycerol M9 minimal medium where cultures are maintained in continuous exponential growth by transferring a small amount of the culture to fresh media once each day before late exponential phase is reached (Herring et al, 2006). If there is an average of 2.5 × 10−3 mutations across the genome per cell division (Drake, 1991) and 1011 cells in a 250-ml culture of E. coli at the time of passage, there will be ∼2.5 × 108 mutations appearing during a single passage cycle. Because only ∼108 cells or fewer are passaged, only a small fraction of the mutants avoids extinction by drift each passage. However, mutants with a growth advantage will replicate more often before the next passage, giving a higher probability of surviving subsequent passages. For example, during adaptation in glycerol M9 minimal medium, an rpoC mutant was identified that increased the growth rate by 60%. Because the rpoC mutant sub-population grows much faster than the rest of the population, a single rpoC mutant cell passaged among 2 × 108 cells of wild-type growth rate will increase to represent on average 46 of the 2 × 108 cells of the subsequent passage. Within 144 h, >98% of the population is expected to have the mutant rpoC allele (Figure 3A), and eventually the small, shrinking sub-population with the wild-type allele will be lost to drift. As an example of clonal interference, the population sweep of rpoC mutants will be slowed down if it enters the population in the presence of another mutant sub-population that grows at a rate intermediate between the wild-type and rpoC mutant growth rates (Figure 3B). As in the previous scenario, the rpoC mutant eventually sweeps the culture, this time eliminating a lesser beneficial allele from the population.

Figure 3

Simulation of the evolution dynamics of an rpoC mutant. Exponential growth in the number of cells N at time t is given by the function N(t)=N0e where N0 represents N at t=0 and k represents the growth rate. (A) The fraction of the population represented by an rpoC mutant (k=0.43) in an otherwise wild-type population (kwt=0.27) and which initially exists at a ratio of one rpoC mutant cell per 2 × 108 wild-type cells is shown. (B) The scenario in this graph is the same as before, except initially the one rpoC mutant (blue) cell exists in a population of 108 cells that is 95% wild-type and 5% glpK mutant (red, kmut=0.35). This situation results in a situation of clonal interference between two beneficial alleles.

As the above examples suggest, homogeneous populations are often found in serial passage experiments, where strong selective pressures favor sweeps and small bottlenecks purge diversity (Herring et al, 2006; Conrad et al, 2009; Lee and Palsson, 2010). In contrast, sweeps are more rarely observed in chemostat ALE experiments (Kao and Sherlock, 2008). High levels of diversity within a single niche rapidly appear in chemostat conditions (Notley-McRobb and Ferenci, 1999a, 1999b; Maharjan et al, 2006; Wang et al, 2010). Diversity within chemostat cultures is unintuitive, since, as the competitive exclusion principle (Hardin, 1960) states, if two mutants are competing for the same resource, one mutant should win out if it has even a small edge of the other mutant. However, diversity may exist given the large number of generations needed to edge out a slightly less fit competitor, especially when even a nearly down-and-out competitor can make a comeback in the population through acquiring additional mutations (Kao and Sherlock, 2008). High population diversity may also be due to elevated mutation rates in slow-growing chemostat cultures (Notley-McRobb et al, 2003) and a lack of bottlenecks.

The stable coexistence of two clonal sub-populations on a single resource has been witnessed in ALE experiments. A number of phenomena can explain these populations. Evolution of cross-feeding commensalisms (Pfeiffer and Bonhoeffer, 2004) allow for a first sub-population to consume a primary nutrient (i.e., glucose) and excrete a secondary compound (i.e., acetate) that is metabolized by a second sub-population (Helling et al, 1987; Rosenzweig et al, 1994; Rozen and Lenski, 2000; Kinnersley et al, 2009). A related example of niche diversity is a ‘fast switch' (FS) phenotype that was found to evolve in minimal media containing both glucose and acetate (Friesen et al, 2004; Spencer et al, 2007, 2008). The FS phenotype switches to metabolizing acetate faster than the ancestor when glucose is exhausted (diauxic switch). However, this comes at the cost of slower growth in glucose. In this scenario, the FS strain is able to stably coexist with a non-FS phenotype through frequency-dependent selection. Finally, ALE resulted in the development of stability of a syntrophic mutualism in which Methanococcus maripaludis fed on hydrogen produced in the fermentation of lactate by Desulfovibrio vulgaris. The consumption of hydrogen by M. maripaludis benefited D. vulgaris by providing the thermodynamic driving force for the fermentation of lactate that is necessary for growth (Hillesland and Stahl, 2010).

ALE studies have demonstrated that evolving populations often exhibit dynamics that are more complicated than the clonal replacement model suggests would occur. Furthermore, the discovery of stable diverse bacterial populations within a single niche due to frequency-dependent selection or mutualistic interactions has challenged the competitive exclusion principle (Hardin, 1960). While further characterization of diversity within chemostat and serial passage populations is of interest, most ALE studies using WGS have focused on sequencing clonal samples of the population. Recent work has described a detailed tracking of genetic diversity in a serial passage population throughout adaptation (Barrick and Lenski, 2009), and advances in sequencing technology may soon allow high resolution measurements of population diversity on a genome scale.

Conclusions

Experience from microbial ALE experiments under a variety of selective pressures is accumulating rapidly. Here, we have reviewed what has been learned from ALE experiments about the genetic basis of adaptation, regulatory changes, optimality in evolving populations, and population dynamics. Importantly (1) The genetic basis of adaptation can now be determined by next-generation sequencing technologies and can be reconstructed using allelic replacement. Epistasis is frequently observed between multiple reconstructed mutants and may indicate either compensatory effects for the earlier mutation or a dependence of the fitness benefit of the allele on the cellular growth rate. (2) Broad-acting mutations to regulatory hubs, including the RNAP itself, are often highly beneficial. (3) The target phenotype of an evolving population is likely guided by principles of optimality. Models based on methods such as FBA and cost-benefit analysis can predict aspects of these target phenotypes. (4) Diversity in evolving populations is related to the mutation rate, the fitness effects of beneficial mutations, and the size of bottleneck involved. The minimization of diversity through clonal replacement is more common in serial passage experiments than in chemostat experiments.

Mutations selected for in ALE experiments that cause broad expression changes can be thought of as systemic perturbation variables. For instance, a mutation that is found to increase tolerance to a toxic end product represents a particular change in the DNA that has a very particular systemic effect on network functions that is selected for. Biological function can be studied through combinations of environmental and genetic perturbations. The field typically considers perturbation variables to be environmental (chemical) or genetic (gene knockout). Dual perturbation experiments are used to discover gene function (Nichols et al, 2011). A collection of systemic perturbation variables resulting from ALE, that Figure 1 basically represents, gives the field of systems biology its own perturbation variables to work with. These perturbation variables, singly or in combination, confer some optimality property on network functions that become known through the selection process implemented in ALE. Thus, they may be used to study gene function or as raw material for ‘accelerated evolution' (Wang et al, 2009).

The frequent mutation of transcriptional regulators is consistent with recent evidence showing that regulatory networks evolve faster than other networks, such as genetic networks, protein interaction networks, and metabolic networks (Shou et al, 2011). Therefore, tools to improve understanding of changes to the regulatory network are of high priority. Omics measurements have shown great value for providing complete systems-level determination of changes in the state of the regulatory network. Yet, these data sets are impossible to analyze fully without a genome-scale context. As discussed earlier, use of genome-scale metabolic reconstructions as a scaffold has yielded insight into changes in the expression state of evolved populations (Lewis et al, 2010). Currently, progress is being made toward a comprehensive transcription regulatory network (TRN) for E. coli (Cho et al, 2008) reconstructed from genome-scale data sets. The comprehensive TRN will enable detailed analysis of transcription changes. Another form of genome-scale context is provided by fitness profiling (Goodarzi et al, 2010). This method determines the fitness change associated with a large increase or decrease in expression of a single gene, therefore, providing clues to changes in gene expression that are important for the mutant phenotype.

In addition to academic interest, ALE studies have a number of important practical applications. In the field of metabolic engineering, ALE is useful both for improved product yields in growth-coupled designs and for the study of tolerance to the strenuous conditions often implicit in industrial-scale growth (heat, pH, and toxic levels of alcohols) (Hughes et al, 2007; Atsumi et al, 2010; Horinouchi et al, 2010; Kishimoto et al, 2010). ALE studies can simulate the development of antibiotic resistance and drug sensitivity within the laboratory, thus indicating mechanisms by which dangerous drug-resistant bacterial strains can arise (Friedman et al, 2006). Other studies of adaptive evolution are moving beyond the laboratory and studying the evolution of bacteria within a human host (Zdziarski et al, 2010). Therefore, the coming years promise practical knowledge and applications to come from ALE experiments, in addition to important biological insights. Some of these insights will come through systems biology and the identification of systemic perturbation variables.