Second-Order Selection for Evolvability in a Large Escherichia coli Population

See allHide authors and affiliations

Science  18 Mar 2011:
Vol. 331, Issue 6023, pp. 1433-1436
DOI: 10.1126/science.1198914


In theory, competition between asexual lineages can lead to second-order selection for greater evolutionary potential. To test this hypothesis, we revived a frozen population of Escherichia coli from a long-term evolution experiment and compared the fitness and ultimate fates of four genetically distinct clones. Surprisingly, two clones with beneficial mutations that would eventually take over the population had significantly lower competitive fitness than two clones with mutations that later went extinct. By replaying evolution many times from these clones, we showed that the eventual winners likely prevailed because they had greater potential for further adaptation. Genetic interactions that reduce the benefit of certain regulatory mutations in the eventual losers appear to explain, at least in part, why they were outcompeted.

Organisms may vary not only in traits that determine their immediate fitness, but also in their potential to generate better-adapted descendants with new beneficial mutations. Evolutionary potential, or evolvability, can be operationally defined as the expected degree to which a lineage beginning from a particular genotype will increase in fitness after evolving for a certain time in a particular environment (1). Evolvability thus reflects a complex probabilistic integration of accessible paths in the fitness landscape influenced by mutation rates, population structure, and epistatic interactions between mutations (24). Experiments with microorganisms have shown that genotypes with elevated mutation rates have greater evolvability under certain conditions (5, 6). The evolutionary potential of microorganisms can also vary when the same mutations have different fitness effects in different genetic backgrounds due to epistatic interactions (79). The extent to which material differences in evolvability of this latter kind—reflecting genetic architecture (10), rather than mutation rates—spontaneously arise between lineages within asexual populations and play a role in ongoing evolutionary dynamics is unknown (11, 12).

We found that several genetically distinct subpopulations were already present in a 500-generation sample archived from a now >50,000-generation long-term evolution experiment with the bacterium Escherichia coli (13, 14). In particular, we characterized numerous clones sampled at 500, 1000, and 1500 generations for the presence of five previously discovered beneficial mutations (1418). Specific mutations in the rbs operon and topA, spoT, and glmUS genes fixed in the evolving population between 1000 and 1500 generations, and in pykF after 1500 generations (14). Two beneficial mutations—the ones affecting rbs and topA—were already present in many clones sampled at 500 generations (table S1). Thus, we refer to topA rbs genotypes sampled at generation 500 as eventual winners (EWs) and to other contemporaneous genotypes as eventual losers (ELs). These genotypes may have other beneficial mutations in these same or other genes, and most such mutations could not have been detected by the assays developed for the known mutations.

We further characterized two clonal isolates of each type: EW1 and EW2, each with the topA and rbs mutations, EL1 with no known mutations, and EL2 with another mutation we call the rbs1 mutation. The EW topA allele is an amino acid substitution in DNA topoisomerase I that alters chromosomal supercoiling, affects the transcription of many genes, and confers a fitness benefit of ~13% when moved into the ancestral genetic background (16). The rbs mutations are deletions of different sizes in the ribose utilization operon that occur with high frequency and cause 1 to 2% fitness gains (15). Competition experiments against the ancestral strain showed that these four representative clones were 13 to 23% more fit than the ancestor (Fig. 1). Thus, the ELs, and possibly also the EWs, had other beneficial mutations, in addition to those that could be identified by our initial genotyping.

Fig. 1

Fitness of two EW and two EL clones relative to the ancestor of the E. coli long-term evolution experiment. Error bars are 95% confidence intervals. All four clones were significantly more fit than the ancestor. Surprisingly, the EL clones were more fit than the EW clones, both as shown here and in direct competition with one another (fig. S1).

One would naïvely expect that the EWs were better adapted than the ELs at 500 generations, but the opposite was true (Fig. 1). Direct competition experiments also showed that the two representative EW clones were at a significant fitness disadvantage relative to the two representative EL clones (fig. S1). In fact, if the fitness deficit of the EWs (–6.3%) had remained constant, they would have gone extinct in another ~350 generations (19). We found no evidence of negative frequency-dependent interactions (20) between EW and EL strains that might have stabilized their continued coexistence in the long-term population (19).

How did descendants of the EWs prevail over EL lineages despite their fitness deficit? The EW-derived lineage may have simply been “lucky” in this one instance of evolution; that is, they might have stochastically gained highly beneficial mutations that allowed them to overtake the EL subpopulations before they were driven extinct. Alternatively, the EW genotypes may have had a greater potential for further adaptation, such that they would reproducibly give rise to higher-fitness descendants and outcompete EL lineages before they were lost. To distinguish between these two hypotheses, we “replayed” evolution by initiating 10 replicate experimental populations from each clone isolated at 500 generations (EW1, EW2, EL1, and EL2). Each population was propagated independently under the same conditions as were used in the long-term evolution experiment for 883 generations, approximately as long as ELs and EWs coexisted in the original population.

To follow the evolutionary dynamics in more detail, we conducted these evolution experiments in a neutral marker divergence format (9, 21, 22). A variant of each of the four E. coli test strains was constructed wherein the state of a readily scored phenotypic marker (Ara), which has no effect on fitness under these culture conditions, was altered by a specific point mutation (13). Each experimental population was then started by mixing approximately equal numbers of the original test strain (Ara) and the strain with the changed marker (Ara+), with each type grown separately from a single colony to ensure that there was essentially no initial genetic variation within a population and no shared history between independent replicates. Tracking the frequency of this genetic marker in these 40 populations over time allowed us to visually follow and quantitatively analyze the first beneficial mutations that swept to high frequency within each population.

The trajectory for the Ara/Ara+ ratio in each population eventually diverged from its starting level, as bacteria with beneficial mutations linked to one marker state arose and outcompeted their ancestors and competitors with less-beneficial mutations (Fig. 2, A and B, and fig. S2, A and B). The shape of the initial divergence of the family of curves generated from evolutionary replicates of the same clone reflects its local fitness landscape. In particular, a simple model that assumes one category of beneficial mutation, with an effective mutation rate (μ) and fitness benefit (s), reproduces the salient features of these dynamics (21), provided that it includes competition between lineages with alternative beneficial mutations, i.e., clonal interference (2, 23, 24).

Fig. 2

Replay evolution experiments to measure the evolvability of the four representative 500-generation clones. (A and B) The frequencies of Ara and Ara+ versions of each test strain, initially mixed equally, were recorded at regular intervals (symbols) during 883 generations of evolution under the same conditions as were used in the long-term experiment. Marker trajectories for the replay populations initiated from EL1 and EW1 clones are shown (10 replicates each). Shifts in the Ara/Ara+ ratio occur when new beneficial mutations linked to one background arise and increase in frequency within the population. Fitting the replicate marker trajectories (lines and solid symbols) until they deviate significantly from an exponential model (open symbols) provides a distribution of empirical shape parameters for the initial divergence. (C) Effective mutation rates (μ) and fitness effects (s) for the first beneficial mutations to sweep to high frequency in a given genetic background were inferred by comparing experimental divergence parameters with those from simulated marker trajectories. Black rectangles represent maximum-likelihood estimates. Representative EW and EL isolates were grouped for this analysis (19). Figure S2 shows data for EL2 and EW2 populations and other steps in the statistical analysis.

By performing population genetic simulations to generate families of curves for many μ and s values (19), we determined the parameter combinations that agreed best with the experimental curves (Fig. 2C). The replicate EW marker ratio trajectories diverged earlier and more steeply than the EL trajectories, and the effective size of the first beneficial mutations to sweep to high frequency was significantly larger for the EWs. However, by itself, this first mutation was not sufficient for the EWs to overtake the ELs. From the 6.3% initial fitness deficit of the EWs and the 1.8% larger effect size of their first beneficial mutations, we calculate that, on average, the EWs would still be ~4.5% less fit than the ELs after the first adaptive step for each type (Fig. 3).

Fig. 3

Greater evolvability of EWs allows them to reproducibly overtake ELs. Two representative EW clones from generation 500 of the long-term evolution experiment were initially at a significant fitness disadvantage relative to two contemporary EL clones (circles). The EWs were somewhat closer in fitness to the ELs, but still lagged behind on average, after the first beneficial mutations swept to high frequency in the replay evolution experiments (triangles), as determined by the marker trajectory divergence analysis. After 883 generations, the representative EWs evolved to a higher fitness on average than the ELs in the replay populations (pentagons). Percentage differences in fitness are for pooled EWs versus ELs at the highlighted time point, and P-values indicate whether this difference was significant (19). Arrows represent presumptive mutational steps, with dashes indicating that the exact number of mutations may vary. The y axis is unlabeled for the final 883-generation replay isolates because their fitness was measured with respect to each other, not relative to the ancestor.

To compare evolvability on a time scale that allows a lineage to accumulate multiple beneficial mutations, we isolated a random clone from each replicate population at the 883-generation endpoint. This evolved clone was either Ara or Ara+. We performed head-to-head competitions of every EL-EW pair with opposite Ara marker states to determine their relative fitness (fig. S3). We found that, on this time scale, the EWs overcame their initial fitness deficit and evolved to higher fitness than the ELs by ~2.1%, on average (Fig. 3). Thus, the EWs evidently prevailed in the original long-term population because they had greater evolvability. On average, they achieved higher fitness than the ELs after each type had equal time to evolve by multiple beneficial-mutation steps from its starting point in the fitness landscape. We stress, however, that this result is necessarily probabilistic in nature. Not every evolved EW clone was able to outcompete every evolved EL clone.

What is the genetic basis of this difference in evolvability? There are three possibilities. First, the EW genetic background may have interacted more favorably with certain potential beneficial mutations than the ancestral background with those same mutations (positive epistasis), thereby opening up additional pathways for adaptive evolution. Conversely, mutations in the ELs may have reduced the effects of otherwise beneficial mutations (negative epistasis), thereby closing off some pathways for adaptation. Finally, a mutation in the EWs may have caused an elevated mutation rate relative to that of the ELs that would allow the EWs to access rarer, more beneficial mutations.

To distinguish the salient genetic differences between the EWs and ELs, we resequenced the genomes of eight evolved E. coli isolates from generation 883 of the replay experiments (19). We chose two strains descended from each of the four 500-generation clones, so that we could reconstruct what mutations were present in the original isolates as well as sample mutations that occurred in their descendants (Fig. 4A). We found that the EWs shared only the two known mutations (topA and rbs) and that both ELs had two previously unknown base substitutions (topA1 and fadR). The EL topA1 mutation alters the amino acid (isoleucine-34 to serine) directly adjacent to the one changed by the EW topA allele (histidine-33 to tyrosine). FadR is a regulator of fatty acid and acetate metabolism (25), and the effects of this EL mutation are unknown.

Fig. 4

(A) Mutations identified by whole-genome resequencing of endpoint E. coli clones from the replay evolution experiments and inferred by parsimony in their EW and EL ancestors. Numbers in squares indicate how many mutations accumulated relative to the original long-term ancestor with mutations in key genes labeled. See tables S2 to S9 for complete lists of all mutations. (B) Fitness effects of adding the spoT mutation that fixed during the long-term experiment to the EL1 and EW2 genetic backgrounds, measured relative to the ancestral strain. Error bars are 95% confidence limits (hidden by symbols in some cases). P-values indicate the significance of the hypothesis that addition of the spoT mutation caused a fitness difference. (C) Fitness effects of the long-term spoT, EW topA, and EL topA1 mutations alone and in combination in the ancestral genetic background. Dashed lines converging on empty diamonds show the fitness predicted for each spoT and topA allele combination given independent multiplicative effects. P-values are for the hypothesis of no epistatic interactions under a multiplicative model (31). Error bars are 95% confidence limits.

From two to five mutations accumulated during the 883-generation replay experiment in the eight independently evolved isolates (tables S2 to S9). There was no evidence that EWs had an elevated mutation rate that might have contributed to their greater evolvability. The number of replay-phase mutations in the four evolved EWs (16 total) was essentially identical to that in the four evolved ELs (15 total). The only mutation in the original 500-generation clones in a gene related to DNA repair or replication (uvrB) was shared by the EL1-derived strains, contrary to the expectation if a change in mutation rate drove the difference in evolvability.

It is very likely that most of the 31 observed mutations that occurred during the replay experiments are beneficial. The rate of genomic change due to adaptive evolution greatly exceeds the rate of neutral drift in this system, and mutations in some of the same genes, operons, and pathways have been found in other isolates from the long-term experiment (14, 26). Two genes independently evolved mutations in more than one of the sequenced 883-generation EL- and EW-derived strains (Fig. 4A). Seven of the eight sequenced clones from the replays evolved mutations in pykF, which encodes the metabolic enzyme pyruvate kinase. In the original long-term population, pykF mutations were detected in EWs by 1500 generations (table S1) and in rbs1 ELs by 1000 generations (19). The long-term EW pykF mutation is highly beneficial in this environment (14), and all 12 long-term populations substituted pykF mutations by 20,000 generations (27). Two of the sequenced replay clones had point mutations in spoT, which encodes a bifunctional (p)ppGpp synthesis and degradation enzyme that is a global regulator of gene expression. The next mutation found in the long-term EW lineage, after rbs and topA, was also a spoT base substitution (table S1). This spoT mutation has been shown to affect the transcription of numerous genes and to confer a fitness benefit of ~9% when moved into the ancestral strain background (17, 28).

We observed pykF mutations in both EW- and EL-derived strains. However, both spoT mutations arose in EW-derived lines (Fig. 4A). This association with genetic background raised the possibility that spoT mutations might be involved in epistatic interactions that underlie differences in EW-EL evolvability, especially given the potential for widespread pleiotropic effects caused by mutations in this global regulator (29). To increase the statistical power for detecting an association, we sequenced the complete spoT reading frame in all 40 EW and EL endpoint clones isolated from the replay experiments. We found spoT mutations in 6 of the 20 evolved EW clones (table S10). Notably, we found no spoT mutations in the 20 EL clones, a difference unlikely by chance (two-tailed Fisher’s exact test, P = 0.0202).

To test directly for epistatic interactions, we measured the fitness effects of adding the spoT allele that arose during the long-term experiment to the EW2 and EL1 strain backgrounds. We found that this spoT allele confers a large fitness benefit in the EW background, but has no significant effect in the EL background (Fig. 4B). Because changes in chromosomal supercoiling also may have widespread pleiotropic effects, it seemed likely that interactions with the topA alleles specific to the EW and EL backgrounds might explain these epistatic effects. To test this hypothesis, we constructed otherwise isogenic strains carrying only the EW topA or EL topA1 alleles on the ancestral background and then measured the fitness effects of adding the spoT mutation that arose during the long-term experiment to each strain (Fig. 4C). The EL topA1 allele is beneficial on its own, though less so than the EW topA mutation. The spoT mutation is also highly beneficial on its own and has essentially the same fitness effect in the presence of the EW topA allele. By contrast, this spoT mutation is neutral, or at least much less beneficial, in combination with the EL topA1 mutation.

These results therefore support the hypothesis of negative epistasis between EL mutations and later mutations that arise in spoT, and they contradict the hypothesis of positive epistasis between the EW and that spoT allele. Highly beneficial spoT mutations are evidently at the leading edge of the adaptive mutations that are accessible to the EW subpopulation, which is why they often evolve in this background. The alternative topA1 mutation that arose in the EL subpopulation evidently renders those spoT mutations neutral, leaving other less-beneficial mutations the best available. In essence, the ELs followed a trajectory in the fitness landscape that allowed more rapid improvement early on, but which shut the door on at least one important avenue for further improvement. By contrast, the EWs followed a path that did not preclude this option, giving them a better than otherwise expected chance of overtaking the ELs. Because spoT mutations evolved in only 6 of the 20 EW replays, and because it took multiple mutations for the EWs to overtake the ELs, it is likely that epistatic interactions with other beneficial mutations also contributed to their differences in evolutionary potential.

We have demonstrated in detail a case in which epistatic interactions between beneficial mutations caused differences in bacterial evolvability that appear to have played a pivotal role in the evolution of a population. Similar cases are expected in any population of asexual organisms that evolve on a rugged fitness landscape with substantial epistasis, as long as the population is large enough that multiple beneficial mutations accumulate in contending lineages before any one mutation can sweep to fixation (2, 11, 2224, 30). This scenario thus provides a general mechanism for the evolution of evolvability by “second-order selection” (3) on genetic architecture. Our results also suggest that studying the interactions among regulatory networks could lead to a deeper understanding of how genetic changes in those networks might either promote or impede evolvability at a systems level.

Supporting Online Material

Materials and Methods

Figs. S1 to S3

Tables S1 to S10

References and Notes

References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. This work was supported by the NSF (DEB1019989 to R.E.L.), Defense Advanced Research Projects Agency (HR0011-09-1-0055 to R.E.L. and T.F.C), the NIH (K99GM087550 to J.E.B.), and the McDonnell Foundation (220020174 to T.F.C). We thank N. Hajela for laboratory assistance, staff at the Michigan State University (MSU) Research Technology Support Facility for assistance with genome sequencing, and the MSU High-Performance Computing Cluster for computational support. Genome sequencing data have been deposited in the NCBI Short Read Archive (SRA024331).

Stay Connected to Science

Navigate This Article