Historical Contingency Causes Divergence in Adaptive Expression of the lac Operon

Abstract Populations of Escherichia coli selected in constant and fluctuating environments containing lactose often adapt by substituting mutations in the lacI repressor that cause constitutive expression of the lac operon. These mutations occur at a high rate and provide a significant benefit. Despite this, eight of 24 populations evolved for 8,000 generations in environments containing lactose contained no detectable repressor mutations. We report here on the basis of this observation. We find that, given relevant mutation rates, repressor mutations are expected to have fixed in all evolved populations if they had maintained the same fitness effect they confer when introduced to the ancestor. In fact, reconstruction experiments demonstrate that repressor mutations have become neutral or deleterious in those populations in which they were not detectable. Populations not fixing repressor mutations nevertheless reached the same fitness as those that did fix them, indicating that they followed an alternative evolutionary path that made redundant the potential benefit of the repressor mutation, but involved unique mutations of equivalent benefit. We identify a mutation occurring in the promoter region of the uspB gene as a candidate for influencing the selective choice between these paths. Our results detail an example of historical contingency leading to divergent evolutionary outcomes.

Introduction "I am [speaking] of the central principle of all history-contingency. A historical explanation [rests] on an unpredictable sequence of antecedent states, where any major change in any step of the sequence would have altered the final result. This final result is therefore dependent, or contingent, upon everything that came before-the unerasable and determining signature of history." -Stephen J. Gould (1989) Wonderful Life, p. 283. "Life is just one damned thing after another!" -Lilian Bell (1909) The Concentrations of Bee, p. 241. The expression was used widely at that time.
Gould famously argued that the outcomes of biological evolution are often causally dependent on the precise series of changes that have occurred in the evolving lineages (Gould 1989;Blount et al. 2008Blount et al. , 2012Zee et al. 2014). One possible basis for such historical contingency are epistatic interactions between mutations in determining fitness. Long-term evolution can be thought of as consisting of "one damned substitution after another," to paraphrase the popular expression from 1909. If the effects of different mutations on fitness are completely independent of each other (i.e., if there is no epistasis for fitness), then later substitutions will not be causally dependent on earlier ones, and there will be no meaningful contingency. If, however, a mutation confers a benefit only in the presence of certain earlier mutations (i.e., if there is epistasis for fitness), then evolution will be historically contingent (Blount et al. 2018). In that case, the presence of a particular mutation determines the range of subsequent mutations that are likely to be favored. Epistasis for fitness means that even closely related populations might respond differently and unpredictably to similar selection pressures, depending on the substitutions they undergo, exactly as Gould envisaged (Blount et al. 2008;Wang et al. 2013;Harms and Thornton 2014;Spor et al. 2014;Zee et al. 2014;Shah et al. 2015;Phillips et al. 2016;Peng et al. 2018).
Historical contingency is difficult to examine because it requires disentangling the roles of chance and natural selection in determining evolutionary outcomes (Blount et al. 2018). One approach to this is to construct actual and alternative evolutionary pathways and determine the dependence of key phenotypes on the order in which mutations occur. This approach has been followed in several studies examining evolution of single genes, typically finding that epistasis creates a strong dependence of evolutionary outcome on the order in which mutations occur (Weinreich et al. 2006;Ogbunugafor and Hartl 2016;Starr et al. 2017). At the level of whole organisms, this approach is not generally feasible if large numbers of mutations have occurred. Instead, it is possible to use controlled laboratory experimental evolution to examine the evolutionary outcomes of replicated populations and determine statistically a signature of historical contingency (Blount et al. 2008(Blount et al. , 2012Zee et al. 2014).
Several experimental evolution studies have found that adaptive potential differs between different genotypes, often, at least in part, explained by differences in their initial fitness (Travisano et al. 1995;Barrick et al. 2010;Bedhomme et al. 2013;Kryazhimskiy et al. 2014;Perfeito et al. 2014;Wünsche et al. 2017;Johnson et al. 2019). The strong influence of fitness on mutation effect is supported by work that has directly measured the effect of adding specific mutations to diverse genetic backgrounds (Pearson et al. 2012;Wang et al. 2013;Kryazhimskiy et al. 2014;Wang et al. 2016). Nevertheless, specific genetic differences between founding strains can be influential (Blount et al. 2008(Blount et al. , 2012Zee et al. 2014). Indeed, historical contingency can affect replicate populations started from a common ancestor and evolved in a common environment (Collins and Bell 2004;Cooper and Lenski 2010;Wiser et al. 2013). In this case, contingency builds on chance differences in the origination and fixation of mutations. One example is the evolution of citrate utilization in an Escherichia coli long-term evolution experiment. In that work, one of 12 replicate populations evolved the ability to use citrate (Blount et al. 2008). Subsequent genetic analysis revealed that evolution of citrate utilization is rare because it depends on a series of prior mutational events (Quandt et al. 2014;Leon et al. 2018). Other examples of historical contingency have been inferred by finding genetically or phenotypically distinct evolutionary outcomes either dependent on early occurring mutations (Yedid and Bell 2002;Meyer et al. 2012) or from analysis of the total suite of changes occurring in evolved populations (Tenaillon et al. 2012).
We examine the basis of divergent evolution in lacI, the repressor of the lac operon genes that confer the ability to utilize lactose. Mutations in lacI fixed or rose to high frequency in 16 of 24 populations evolved for 8,000 generations in a long-term experiment that evolved populations of E. coli in environments containing combinations of glucose and lactose (Cooper and Lenski 2010;Satterwhite and Cooper 2015) ( fig. 1A and supplementary fig. S1, Supplementary Material online). Absence of repressor mutations in the remaining eight populations was surprising because lacI mutations confer a large fitness benefit in the ancestor ( fig. 1B) and occur at a high rate, due to the presence of a mutational hotspot (Farabaugh et al. 1978). Evolutionary simulations incorporating the measured lacIÀ mutation rate and fitness effect indicate that this divergence cannot be attributed to chance differences in mutation timing and success. When we added a lacIÀ mutation to strains isolated from populations that did not fix it, we found that it no longer confers any benefit, and, in fact, is often deleterious, indicating that it interacts negatively with previous substitutions. These results demonstrate that divergent adaptation in the lac operon is common and is due to effects contingent on previous adaptations.

Mutations in lacI Occur in about Half of Evolved Populations
We previously found that loss-of-function mutations in the lac operon repressor, lacI, were common in nine of 12 populations of E. coli evolved for 2,000 generations in environments either containing lactose (Lac) only or alternating daily between lactose and glucose (G/L) (Quan et al. 2012). Here, we find that after extension for a further 6,000 generations of evolution, lacIÀ mutations had fixed or become common in two more of these populations and in five of 12 derived populations that alternated between glucose and lactose every 2,000 generations (denoted G_L and L_G where the first letter indicates the sugar initially present during selection; fig. 1A). The absence of detectable lacIÀ mutations in eight of the 24 populations was surprising because a mutational hotspot in lacI causes inactivating frameshift mutations to occur at a high rate and those mutations provide a benefit in lactose-containing environments that outweigh an associated cost in glucose (Farabaugh et al. 1978; fig. 1). Herafter, we denote evolved populations that did and did not fix a lacIÀ mutation Ev lacIÀ and Ev lacIþ , respectively.
A possible explanation for the failure of lacIÀ mutations to fix after 8,000 generations in multiple populations is that an earlier occurring substitution(s) reduced either the rate at which lacIÀ mutations occur or the benefit they confer when they do arise. We refer to this possibility as the contingency hypothesis. An alternative, noncontingent explanation is possible, however. We examine these hypotheses below.

Simulations Predict More lacIÀ Substitutions than Observed
An alternative to the contingency hypothesis is that lacIÀ mutations, although beneficial, have simply not occurred yet, or have occurred but have not yet had the time to reach high frequency. To evaluate this possibility, we simulated the evolution of populations with lacIÀ mutations available alongside a background pool of other beneficial and deleterious mutations ( fig. 2). We used an individual-based model, incorporating estimates of mutation rates and effects, and allowing for competition between mutations (see Materials and Methods for details). Our results show that, as populations become larger, the frequency of substitution (f s ) for a lacIÀ mutation rises quickly. With an effective population size of 10 5 individuals, lacIÀ fixed in most simulated populations (f s ¼ 91.5% combining Lac, G_L and L_G selection regimes). The observed frequency in experimental populations was only f s ¼ 12/24 ¼ 50% (95% CI: 31.4-68.6%) despite the fact that they were much larger (N e ¼ 3.3 Â 10 7 ) (Cooper and Lenski 2010). This difference is statistically significant (binomial test: P < 0.001). Thus, stochastic variation in mutation timing and competition cannot explain the high number of experimental populations retaining the ancestral lacI allele. Karkare et al. . doi:10.1093/molbev/msab077 MBE Evolved Clones Have Similar lacIÀ Mutation Rates to Ancestor A second hypothesis for lacIÀ mutations failing to fix in some populations is that their mutation rate was lower in those populations. Indeed, mutation at the lacI hotspot involves a strand slippage mechanism that might be affected by DNA topology, which is a common target of selection in an experiment closely related to this one (Blount et al. 2008(Blount et al. , 2012Zee et al. 2014). To the extent that a potential change in lacIÀ mutation rate depends on previously substituted mutations, this possibility represents an example of contingency. We estimate the ancestral lacIÀ mutation rate at 1.72 6 0.7 Â 10 À7 per cell per generation (95% CI), which is consistent with a previous estimate (Schaaper et al. 1986). If Ev lacIþ populations evolved lower lacIÀ mutation rates, expected fixation times would be increased. In fact, five of eight clones isolated from Ev lacIþ populations had lacIÀ mutation rates indistinguishable from the ancestor ( fig. 3). The remaining clones had an elevated lacIÀ mutation rate. In the case of the L_G1 population, this elevation was substantial, presumably due to the presence of a 27-bp deletion in the mutS gene, which is likely to result in a general increase in the genomic mutation rate (supplementary table S1, Supplementary Material online).
In summary, changes in mutation rate to lacIÀ do not explain why it has not fixed in so many evolved populations. We note, however, that this analysis does not exclude the possibility that populations might have evolved in a way causing beneficial mutations at loci other than lacI to occur at higher rates and that these mutations might outcompete mutations occurring at lacI (Gerrish and Lenski 1998). We consider this possibility unlikely, however, because it would delay but not prevent lacIÀ mutation substitutions. Moreover, results presented below support an alternative mechanism affecting lacIÀ fixation.

Fitness Effect of lacIÀ Is Lower in Ev lacIþ Populations
To directly test for a change in the selective benefit of lacIÀ mutations as posited by the contingency hypothesis, we performed allelic replacement experiments to measure the fitness effect of lacIÀ mutations in clones isolated from all evolved populations. If the lacIÀ mutation was prevented from fixing due to epistasis, we expected it to be less beneficial, or even deleterious, in Ev lacIþ populations.
We found that the lacIÀ mutation was beneficial in the 12 populations in which it fixed ( fig. 4). Among clones isolated from populations selected in Lac, G_L, and L_G environments, the grand mean fitness effect of the lacIÀ mutation measured in the lactose environment is 9.66% (62.29% [95% CI]), which is not significantly different from its effect in the ancestor (two-tailed t-test, P ¼ 0.53). Among clones isolated from the G/L environment, the lacIÀ mutation conferred a significantly greater benefit than it did in the ancestor measured in that environment (8.50% vs. 4.05%, two-tailed t-test: P ¼ 0.02). By contrast, the lacIÀ mutation conferred no benefit or was deleterious when introduced into clones isolated from the eight Ev lacIþ populations, imposing a mean cost of 7% (64.81% [95% CI]) when measured in lactose (compared Simulated frequency of lacIÀ substitution (fs). Simulations were modeled in Lac and extrapolated to apply to G_L and L_G 2,000-generation fluctuating environments as detailed in Materials and Methods. Symbols indicate the proportion of replicate simulations (n ¼ 100 except at I e ¼ 5 Â 10 4 [n ¼ 50] and N e ¼ 10 5 [n ¼ 20]) in which lacIÀ fixed. Error bars are 95% CI. Means and errors were calculated using 1,000 boot-strapped samples. In the Lac treatment, populations were simulated for 8,000 generations. For G_L and L_G 2,000 generation fluctuating environments, the same simulated populations were sampled at 2,000 generations and frequency of lacIÀ substitution (fs) was calculated as detailed in Materials and Methods. The dashed line indicates the overall frequency of lacIÀ fixation at 8,000 generations among 24 experimentally evolved populations. Note that the two treatments are plotted with a small offset to avoid overplotting.
Divergence in Adaptive Expression of the lac Operon . doi:10.1093/molbev/msab077 MBE with its effect in the ancestor, two-tailed t-test: P < 0.001) ( fig. 4). These results reveal the presence of strong negative epistatic interactions between the lacIÀ mutation and one or more of the mutations that accumulated in the Ev lacIþ populations, in agreement with the contingency hypothesis.

Ev lacIþ Populations Have an Alternative Mechanism of Reducing Lag Time
To examine the basis of the different fitness effects of the lacIÀ mutation, we compared its effect on growth dynamics in the ancestor, and in Ev lacIÀ and Ev lacIþ clones. In the ancestor, the most striking effect of the lacIÀ mutation was a decrease in the time taken for the population to resume growth following transfer to fresh medium (i.e., its lag time) from 5.00 to 3.87 h (t-test: P < 0.001; fig. 5A  The absence of lacIÀ mutation benefit may be because Ev lacIþ clones had substituted an alternative means of shortening lag time such that the effect of the lacIÀ mutation has become redundant. To test this possibility, we compared the lag times of Ev lacIþ and Ev lacIÀ clones. We found that lag times were indistinguishable, consistent with Ev lacIþ populations having evolved an alternative means of shortening lag time (two-tailed t-test P ¼ 0.96; fig. 5D).
Ev lacIþ Populations Have Increased Sensitivity to a lac Operon Inducer As a first step in examining the basis of the reduction of lag time among Ev lacIþ populations, we measured changes in the expression of lac genes in environments with increasing concentrations of IPTG, a synthetic inducer. We found that clones from all Ev lacIþ populations were more sensitive in responding to IPTG, reaching half maximum expression of LacZ, a gene product controlled by the LacI repressor, at a lower concentration of IPTG than the ancestor (supplementary fig. S4, Supplementary Material online). Qualitatively similar results were seen when we used an alternative inducer, TMG, and compared expression of a P lac -gfp reporter at 4 and 6 h following induction (supplementary fig. S5, Supplementary Material online). These findings are consistent with a higher sensitivity to IPTG and TMG inducers in the Ev lacIþ strains.
An increase in inducer sensitivity could cause reduced lag times in Ev lacIþ populations by allowing them to more quickly express lac genes following transfer to a fresh lactose supplemented environment. One mechanism through which this could occur is if basal lac expression was higher in Ev lacIþ populations, which increases the probability that a cell will turn on lac operon expression and resume growth (Chu and Barnes 2016). However, in both the IPTG and TMG induction data experiments, basal lac expression appeared unchanged in Ev lacIþ populations relative to the ancestor. We also considered the possibility of some change in the nature of lactose uptake. None of the evolved strains had mutations in the LacY permease protein so, to test the possibility that some other inducer import mechanism had evolved, we determined the reliance of lac induction on the LacY permease. Mean per locus mutation rates are calculated from n ! 10 replicates, and errors are 95% CIs. Asterisks indicate populations with a lacI mutation rate significantly different than the ancestor (t-test P < 0.05). . We also found that TMG, but not lactose, was able to weakly induce lac genes independent of LacY, but that ability was not different between the ancestor and the Lac5 strain.

Reduced Benefit of lacI Mutations Is Caused by Mutations in uspB
To examine the nature of the mutations that interact with lacIÀ to reduce its benefit, we estimated the fitness effect of a lacIÀ mutation introduced into clones isolated at intervals along the evolutionary path of the eight Ev lacIþ populations ( fig. 6A). We reasoned that steep declines in lacI mutation effect would indicate the substitution of a negatively interacting mutation in the population. Among the tested populations, the steepest change was seen in the Lac5 population. Focusing on the beginning part of the evolution of this population, the effect of the lacIÀ mutation declined from conferring an 8.3% benefit in the ancestor to a mean cost of 0.9% among four independent clones isolated from this population at 500 generations of selection ( fig. 6B). To determine the genetic basis of this change, the 500 generation evolved clones were sequenced and mutational changes relative to their ancestor identified (table 1). All four clones shared an IS150 insertion occurring 24-bp upstream of the uspB gene. The same mutation is present in clones isolated from this population after 4,000 and 8,000 generations of evolution.
To determine the interaction between lacIÀ and uspB mutations, we focused on a Lac5 8,000 generation evolved clone. The lacIÀ mutation confers a cost in this clone measured in the lactose environment (Lac5 lacIÀ vs. Lac5: relative fitness ¼ 0.89 6 0.09 [95% CI], P ¼ 0.03). When we reverted the evolved uspB allele, the lacIÀ mutation conferred a 26% benefit, comparable with its effect in Ev lacIÀ clones and greater than its effect in the ancestor (Lac5 uspB Anc lacIÀ vs. Lac5 uspB Anc : relative fitness ¼ 1.26 6 0.09 [95% CI], P ¼ 0.006). This result identifies the uspB mutation as being one cause of the historical contingency we see, acting to reduce the benefit conferred by lacIÀ. We note, however, that the situation is complex and might depend on higher order interactions. Insertion mutations upstream or in uspB are present in all of seven sequenced clones isolated from Ev lacIþ populations, but also in clones isolated from four Ev lacIÀ population (G_L3, G/L5, L_G6, and Lac1) (supplementary tables S1 and S2, Supplementary Material online). This pattern indicates that clones without the lacIÀ mutation are significantly more likely to have the uspB mutation, consistent with a negative genetic interaction (one-tailed Fisher's exact test: P ¼ 0.007). Nevertheless, that uspB and lacIÀ are sometimes found together suggests that other mutations might influence the relationship between them.

Divergent lac Regulation Strategies Do Not Promote Further Genetic Divergence
To test if the historical contingency influencing the effect of the lacI mutation extends to promoting continued divergence of Ev lacIþ and Ev lacIÀ populations along different evolutionary paths, we used Dice's coefficient of similarity (s) to compare the sets of mutations found among each population group (supplementary tables S1 and S2, Supplementary Material online). This metric is bounded between 0, indicating completely dissimilar mutation sets, and 1, indicating identical mutation sets. We found no signal of higher mutation   fig. S7A, Supplementary Material online). Moreover, no similarity estimate was significantly different to a null expectation based on randomly assigning sequenced clones to the Ev lacIþ and Ev lacIÀ group labels. A limitation to this analysis is that it groups populations that were selected in different environments. Although it would be ideal to consider the effect of lacIÀ on subsequent mutation accumulation separately for each selection environment, only one of four environments contains multiple populations of each lacI type, limiting our ability to distinguish repeated from random effects. We note, however, that, with one exception, clones from the different selection environments did not detectably differ in the sets of mutations they accumulated (supplementary fig. S7B, Supplementary Material online).

Discussion
We found that the absence of potentially beneficial lacI mutations in Ev lacIþ populations was due to epistasis and is, therefore, a case of historical contingency. In other words, populations followed at least two distinct evolutionary paths contingent on the presence of some earlier mutational change. This early mutation caused the fitness effect of subsequent lacIÀ mutations to be neutral or deleterious in all Ev lacIþ populations, instead of beneficial as in the ancestor and Ev lacIÀ populations. We found no evidence to support alternative hypotheses of insufficient time or changes in mutation rate explaining the absence of lacIÀ mutations.
In the context of replicate evolved populations started from a common ancestor, as examined here, historical contingency stems from negative genetic interactions between early mutations and potentially beneficial mutations. At the physiological level, these interactions reflect that otherwise beneficial mutations are either redundant to, or interact antagonistically with, earlier substituting mutations. Redundancy could occur if there were two alternative mutational targets that cause the same phenotypic change. For example, mutations affecting either repressor proteins or operator binding sites (Quan et al. 2012). Antagonism could reflect, for example, alternative adaptive strategies that share some metabolite such that the evolution of one strategy changes the cellular physiological environment in a way that means the second is no longer beneficial (reviewed in De Visser et al. [2011]). Whereas redundancy allows for convergence in adaptive traits, even as populations might diverge genetically, antagonism between adaptive mechanisms predicts eventual phenotypic divergence as new mutations come to depend on the different cell environments caused by earlier ones (Szamecz et al. 2014).
We find that the shortened lag time conferred by lacIÀ mutations in the Ev lacIÀ populations is matched in Ev lacIþ populations, consistent with Ev lacIþ populations substituting some alternative mutation(s) that makes lacIÀ redundant. This possibility is also supported by our findings that Ev lacIþ and Ev lacIÀ populations had indistinguishable growth curves, that Ev lacIþ populations increased in their sensitivity to lac operon induction, and that the overall set of substituted mutations did not differ between the population groups. Indeed, it seems that "evolutionary routes are many, but the destinations are limited." (Blount et al. 2008(Blount et al. , 2012Zee et al. 2014). A purely redundant mechanism of interaction does not, however, easily explain our finding that adding lacIÀ decreases fitness in five of eight Ev lacIþ strains more than can be explained by only the cost of constitutive lac expression. These high costs indicate some direct antagonism  MBE between lacIÀ and alternative adaptations. To the extent that this antagonism reflects a distinct mechanistic basis of adaptation, albeit with similar immediate phenotypic effects, it might eventually lead to diverging selective opportunities and, therefore, evolutionary paths. The mechanism by which the lacIÀ mutation shortens lag time likely depends on the lack of functional repressor causing cells to produce higher levels of lac gene products through stationary phase so that they can more quickly import and catabolize lactose in fresh medium (Quan et al. 2012;Chu and Barnes 2016). The most obvious candidate for a mutation interacting with lacIÀ, therefore, would be one causing the same higher lac activity. We did not find any candidate mutations in the canonical lac regulon of Ev lacIþ clones, and, indeed, we did not detect any increase in basal lac expression among these clones. However, we did identify an IS150 insertion mutation occurring upstream of uspB as interacting with lacI in at least one population. The insertion occurs just upstream of uspB between the promoter and the start of the gene. It therefore seems likely to reduce or eliminate that gene's expression, though IS150 has an outward-facing promoter and could also upregulate uspB (Schwartz et al. 1988). In either case, a link between UspB and lactose metabolism is unclear. UspB is an integral membrane protein that is upregulated in stressful conditions, including starvation, and might play a role in mediating changes in membrane composition either directly or as a result of disrupted proteinprotein interactions (Farewell et al. 1998;Liu et al. 2019). In principle, such changes could affect stationary phase membrane composition to facilitate lactose uptake although lactose induction of lac expression still depended on the LacY permease so any change in uptake must have been minor (supplementary fig. S6, Supplementary Material online). We did not rule out the possibility that uspB interacts indirectly with lactose metabolism, though the benefit of lacIÀ was lost in strains containing only uspB and one other mutation, a deletion of the rbs operon. The rbs deletion mutation occurred in almost all of our evolved populations, regardless of their lacI status, so it seems likely the interaction is, in fact, direct. Finally, we note that although the uspB mutation was found in all Ev lacIþ clones, it was also found in one Ev lacIÀ population, indicating that the interaction between uspB and lacI mutations might be influenced by other mutations.
Both Ev lacIþ and Ev lacIÀ populations are found in three of the four lactose-containing environments we consider, but at different frequencies. The ratio of lacIÀ absence to fixation (7:3) is substantially higher among the 2,000 generation fluctuating G_L and L_G selected populations than among those selected in the daily alternating G/L environment (0:5) (Fisher's exact test: P ¼ 0.026), even though populations spent the same total time in each environmental component. The lacIÀ mutation was also less likely to fix among G_L and L_G populations than among Lac populations, although this difference was not significant (Fisher's exact test: P ¼ 0.119). One explanation for these differences is that mutations selected in the glucose portion of the selection regime can interact negatively with lacIÀ, decreasing its benefit in lactose. Such mutations would not be expected to reach high frequency in the G/L daily fluctuating environments where they would compete directly with lacIÀ mutations for fixation.
How the history of a population affects its future evolution remains a major question in evolutionary biology (reviewed in Blount et al. [2018]). Does natural selection generally converge on high-fitness genotypes, such that evolutionary outcomes, even if not paths, are deterministic, or do antagonistic interactions cause chance differences in mutation occurrence and success to be built on, causing evolutionary outcomes to be contingent, and, therefore, unpredictable? Contingency leading to divergent fitness outcomes has been described in several controlled laboratory evolution experiments, even when comparing populations started from a common ancestor, and therefore having relatively few differences among which contingency can arise (Collins and Bell 2004;Blount et al. 2008;Beaumont et al. 2009;Cooper and Lenski 2010;Wiser et al. 2013). Divergence is commonly seen at the genetic level in these kinds of experiments, though the basis of differences is not usually examined. For example, whether distinct mutational pathways are likely to be equivalent or to affect available subsequent adaptive paths, and therefore likely to lead to eventual fitness differences. One example of the difficulty in evaluating this possibility comes from an experiment demonstrating a reproducibly different adaptive potential of genotypes differing crucially in mutations causing different residues in a gene affecting DNA supercoiling (Woods et al. 2011).
In summary, we describe an example of historical contingency affecting selection of lac regulation. Whereas approximately half of a set of replicate populations selected in lactose-containing environments quickly substituted a mutation in lacI that provided a benefit by decreasing lag time, remaining populations substituted other mutations that interacted negatively with lacI causing them to follow alternative evolutionary paths. At the point in our evolution experiment that was considered here, these paths are marked by distinct genetic, but not fitness, outcomes.

Bacterial Strains and Growth Conditions
Bacterial strains were isolated from populations evolved for 8,000 generations as part of a long-term evolution experiment. This experiment evolved replicate populations in seven different environmental treatments, four of which are considered here: lactose only (Lac), a combination of glucose and lactose fluctuating daily (G/L), or the same combination fluctuating every 2,000 generations with one treatment starting with glucose (G_L) and another with lactose (L_G) (Satterwhite and Cooper 2015) (supplementary fig. S1, Supplementary Material online). Evolution environments comprised these sugars added to base Davis-Mignoli (DM) medium at a concentration of 175 lg/ml (glucose) or 210 lg/ ml (lactose). These concentrations support approximately equal densities of stationary phase bacteria ($3.5 Â 10 8 cfu/ ml) (Cooper and Lenski 2010). Populations were propagated in 1 ml of medium in 96 Â 2 ml well blocks for 8,000 Divergence in Adaptive Expression of the lac Operon . doi:10.1093/molbev/msab077 MBE generations, using a daily 1:100 serial transfer. Each population was initially homogeneous, such that de novo mutation was the only source of genetic variation. Samples were frozen at À80 C, with glycerol as a cryoprotectant, every 500 generations. Six replicate populations were evolved in each environment, with three started from REL606 and three from REL607, a spontaneous Araþ derivative (Lenski et al. 1991). The different ara markers allow strains to be differentiated by plating on tetrazolium-arabinose plates, on which araþ strains form white colonies and araÀ red colonies. The Ara marker does not affect fitness measurements in any of the assay environments (Cooper and Lenski 2010). Lysogeny broth (LB) was used for nonselective culturing, whereas the respective evolution environment of each clone was used for fitness assays and growth measurements.

Identification of lacI Mutations
To determine fixation of lacI mutations within evolved populations, we plated 8,000-generation population samples (1,000-3,000 cells) on TGX indicator medium (Blount et al. 2008(Blount et al. , 2012Schenk et al. 2012;Zee et al. 2014). This medium consists of 10 g/l tryptone, 2.5 g/l sodium chloride, 5 g/l glucose, and 30 mg/ml of the LacZ substrate 5-bromo-4-chloro-3-indolyl-beta-D-galactopyranoside (X-gal). Colony color on this medium indicates the level of LacZ activity in the absence of any recognized inducer, and thus distinguishes between ancestral repressed (white colonies) and mutant derepressed (blue colonies) lac expressing strains. Four populations had a mix of blue and white colonies and were not considered here. Strains were isolated from populations having >95% blue colonies and sequenced to confirm that derepression was due to a mutation in the lacI repressor. Evolved strains from populations that fixed a lacI mutation are denoted as Ev lacIÀ and those from populations that maintained the ancestral allele are Ev lacIþ .

Strain Construction
To determine the fitness effect of lacIÀ mutations in evolved strains, we reverted them using the suicide plasmid pDS132::lacIþ, as described previously (Satterwhite and Cooper 2015). Potential engineered revertants were screened for expected lacIþ phenotypes first on TGX medium. White colonies were streaked onto minimal medium supplemented with P-Gal. Clones that reverted to lacIþ cannot grow on this medium (Nghiem et al. 1988). Clones isolated from populations that retained the ancestral lacI allele were converted to lacIÀ by isolating spontaneous mutants selected on minimal medium supplemented with P-Gal as the only potential carbon source and with X-gal as a means to increase visual contrast of small colonies, making them easier to detect. In all cases, mutant isolation was confirmed by sequencing of the lacI gene. This also allowed us to confirm that selected lacIÀ strains had the same 4-base insertion/deletion mutations that dominated the mutations that fixed in the evolution experiment (Quan et al. 2012).
We used a combined CRISPR-recombineering approach to replace the evolved uspB allele (uspB::IS150) with the ancestral allele (Jiang et al. 2015). A uspB targeting sgRNA pTarget plasmid (pTarget-uspB) was constructed by using the NEBuilder HiFi DNA Assembly kit to assemble DNA fragments containing a donor region of the ancestral uspB locus, the pTarget vector backbone including a uspB N20 sequence, and two linking oligonucleotides (details in supplementary table S1, Supplementary Material online). The uspB donor DNA encoded by the resulting plasmid does not alter the N20 target sequence in the edited strain. To prevent persistent editing after DNA repair, we modified the procedure used by Jiang et al. (2015) by simultaneously treating cells with arabinose and lactose for 1.5 h following transformation of pCas containing cells with pTarget-uspB. This combination activates the red recombination system and cuts the pTarget plasmid, causing it to be lost from cells. Transformants were plated on kanamycin plates to select pCas carrying cells and then grown at 42 C to select for loss of that plasmid. The candidate edited cells were screened and confirmed by colony PCR.

Mutation Rate Estimates
The rate of mutation to constitutive lac expression was estimated using a fluctuation test. For each clone, the freezer stock was inoculated into LB, grown overnight at 37 C, and then diluted 1:1,000 into ten fresh 1 ml LB cultures. After overnight growth, a 100-ll sample from each replicate population was plated onto DM agar supplemented with Xgal and P-gal. A diluted sample was also plated onto LB agar plates to estimate total cell density. Plates were incubated at 37 C for 48 h prior to counting.
To assess the plating efficiency of lacIÀ mutants on the Xgal þ P-Gal selection plates, we compared the number of colonies growing when a small number of lacIÀ cells were plated by themselves on to permissive LB medium and when they were plated onto P-gal medium with an excess of a REL606 strain that had a lac operon deletion that made it unable to mutate to grow on P-Gal. The plating efficiency of lacIÀ cells was calculated as p eff ¼ m c /m e , where m c is the number of lacIÀ mutants able to form visible colonies on P-Gal in combination with the deletion strain, and m e is the number of mutants that form colonies when plated alone onto LB. Colony counts in the fluctuation test were multiplied by 1/p eff to obtain corrected mutation counts. Following this correction, mutation rate analysis was carried out using the bz-rates estimator (http://www.lcqb.upmc.fr/bzrates) (Gillet-Markowska et al. 2015).

Individual-Based Simulations
The frequency of substitution (f s ) for a lacIÀ mutation is a function of mutation rate, population size, and mutation dynamics. These dynamics, such as clonal interference and hitchhiking, can significantly affect fixation times. To estimate the expected fixation time for lacIÀ mutants in the long-term evolution experiment, we conducted individual-based simulations using a Wright-Fisher regime with a genome-wide mutation rate of U ¼ 7 Â 10 À4 , the mean genomic rate as calculated from relevant estimates (Wielgoss et al. 2011;Lee et al. 2012;Long et al. 2016), and a lacIþ to lacIÀ mutation rate of 1.72 Â 10 À7 (measured in the ancestor, see below). Mutations to lacIÀ conferred a fitness increase as measured in this work in the ancestor (8.31% in Lac, 4.05% in G/L; fig. 1B). The effective population size prevailing in the evolution experiment (N ¼ 3.3 Â 10 7 ) (Cooper and Lenski 2010) was too large for us to replicate simulations in a reasonable time frame. We therefore simulated population sizes up to 10 5 . Fixation times are expected to be decreased at larger population sizes, therefore results at N ¼ 10 5 provide a conservative estimate for the probability of a lacIÀ mutation fixing in the experimental populations. Populations were allowed to evolve for 8,000 generations or until the lacIÀ mutation fixed (frequency above 95%). For each population size, we simulated 100 replicate populations and recorded the proportion of these replicates that fixed the lacIÀ mutation. We then bootstrapped the simulation results 1,000 times to estimate the 95% CI for f s . For the 2,000-generation fluctuating environments, we simulated only the two 2,000 generations of evolution in lactose, and calculated f s as where f 2k is the frequency of lacIÀ fixation during the first period of 2,000-generations of lactose selection. Assuming that lacIÀ mutations that did not fix during lactose selection periods were lost during any subsequent period of glucose selection, where they were costly, the second term gives the number of those populations expected to fix lacIÀ in the second period of lactose selection.
We did not simulate evolution in the G/L environment, as this requires estimation of the correlated effects of background mutations across glucose and lactose environments, which are ignored in the more slowly fluctuating 2,000 generation environments.

Fitness Assays
The relative fitness of a given strain was assayed relative to a reference with the opposite Lac (i.e., lacIþ vs. lacIÀ) or Ara (i.e., araþ vs. araÀ) marker. Fitness assays were carried out in identical conditions to a tested strain's evolution environment. Prior to each assay, competitors were independently preconditioned to the competition environment for 1 day, except for the G/L environment, which was preconditioned over 2 days. For the G/L environment, cultures were diluted 100-fold from glucose to lactose for day two (for both preconditioning and competition steps). Following preconditioning, competitors were mixed at a 1:1 volume ratio and transferred with a 100-fold dilution into the competition environment. A sample was immediately plated on indicator agar (TGX or TA) in order to determine the starting frequency of the two competitors. At the end of the competition, samples were again plated on indicator agar. Absolute fitness of competitor, a, was calculated as w a ¼ ln(100 d Â(N a (f)/N a (i)), where d is the number of competition days, and N is the number of colony-forming units at the initial (i) and final (f) time points. The relative fitness of competitor a relative to competitor b is then w a/b ¼ w a /w b , and the selective advantage of a over b is s ¼ w a/b À1.

Growth Rate Estimates and Virtual Competitions
Each clone was inoculated from a freezer stock into 1 ml of LB media and grown overnight. The following day, 1 ll of each culture was transferred into 1 ml of DM supplemented with 210 lg/ml lactose and incubated at 37 C for 24 h. On day three, 2 ll of the preconditioned culture was transferred into 198 ll of the same media in a 96-well polystyrene plate. This plate was incubated in a VersaMax microplate reader (Molecular Dynamics, CA) and grown at 37 C until cells reached stationary phase. The culture's optical density at 450 nm (OD450) was measured every 5 min during growth. Growth curves were analyzed using an extension of the logistic model (Baranyi and Roberts 1994;Ram et al. 2019) as implemented at: https://multi-choice-comparison.shinyapps. io/growth_curves/. To check that estimated growth parameters were meaningful estimates of fitness components, we implemented a double-strain Baranyi-Roberts model to perform virtual competitions as mediated only through growth parameters separately estimated for each competitor (Ram et al. 2019). Virtual competition results were compared with direct competition fitness estimates. Briefly, virtual competitions model a common nutrient pool that is independently used for growth by two competitors depending on their measured growth parameters. The change in the density of each competitor was used to estimate a virtual relative fitness.

Genome Sequencing
Genomic DNA was isolated and purified using the Wizard Genomic DNA Purification Kit (Promega) following the protocol for Gram-negative bacteria at one-third volume. Double-stranded DNA was quantified using SYBR Green I Nucleic Acid Stain (Invitrogen) in a SpectraMax M5 Fluorescence Microplate Reader (Molecular Devices). Libraries were created following the Nextera XT DNA Library Prep Kit protocol, at one-quarter volume, with Nextera XT Index Kit v2 adapters (Illumina). Libraries were individually quantified using the Qubit dsDNA High Sensitivity Assay with a Qubit 2.0 Flourometer (ThermoFisher) and fragment size determined using the Agilent 2100 BioAnalyzer with a High Sensitivity DNA Analysis Kits (Agilent). Libraries were pooled and sequenced on an Illumina NextSeq, producing 150 bp, paired-end reads. The Breseq computational pipeline was used to align reads to the reference sequence and identify mutations ). Statistics R3.5.3 used for plotting and mutation similarity analysis (R Core Team 2017). The mclustcomp function in package mclustcomp was used to estimate Dice's coefficient of similarity comparing mutations fixed in clones isolated from different evolved populations. Simulations were performed using a custom Python script, which is available on request.