Bacterial adaptation to environmental stress involves, in part, a modification of transcription patterns, with downstream impacts on physiological function. In Escherichia coli, the RNA polymerase sigma factor RpoS is a global regulator that coordinates the expression of up to 10% of the genome when the bacterium enters stationary phase or experiences stresses such as starvation, acidity or increased osmolarity [1]. Despite the importance of this protein in many environments, a functional RpoS seems to lower the ability of E. coli to scavenge for scarce nutrients [2],[3]. This cost is hypothesized to occur because there is a limiting amount of core RNA polymerase subunits in the cell, meaning that transcription of stress responsive, RpoS-dependent promoters will decrease the transcription from RpoS-independent promoters involved in nutrient acquisition and utilization [2]–[4].

The hypothesis that the nature of the RpoS regulatory network creates an inherent conflict between stress protection and nutritional competence (SPANC) [3],[4] provides a basis for predicting how natural selection acts on the global regulatory networks of E. coli. The SPANC hypothesis predicts that natural selection will modify the network in favor of nutritional ability at the expense of stress resistance in some environments by decreasing or abolishing RpoS function. Just this type of selection against RpoS activity has been documented in laboratory studies [2],[5],[6]. In addition, strains with low- or null-activity rpoS alleles have been found in natural populations of E. coli and Salmonella enterica [3],[7].

While strains without functional RpoS are favored in some environments, those same strains may do less well in other, more stressful environments where they may be less fit due to an inability to respond to new challenges. While rpoS strains could adapt by recovering or increasing their RpoS function, the mutations that abolished RpoS function may be very unlikely or impossible to reverse. An alternative mechanism involves the selection of mutations that modify the regulatory network to compensate for the loss of RpoS. These compensatory mutations would then increase the fitness of the bacterium in this new, more stressful environment. We sought to understand this type of adaptation by observing the patterns of increased fitness seen in evolving bacterial lines, and by elucidating the molecular basis of the adaptation.

Results from previous experimental studies suggest that compensation for deleterious mutations is a general phenomenon [8]–[11]. Less is known, however, about the variability of the process of compensation. Will strains that lack RpoS adapt more or less variably to a stressful environment than strains with a fully functional regulatory network? Will this involve larger or smaller increases in fitness? At a molecular level, mutations affecting other global regulators of transcription [1], or local changes at a promoter [12],[13] may permit transcription in the absence of RpoS, but we were interested in discovering which options actually are favored by natural selection. Would only a few key genes be involved in adaptation, or would adaptation involve changes in large parts of the transcriptome? Here we used experimental evolution [14],[15] to answer these questions about adaptation to the loss of RpoS.

E. coli expresses RpoS in response to stresses such as extremes of pH, temperature or osmolarity. If this regulatory pattern is important for fitness, then strains lacking RpoS should be less fit in these stressful environments. To test this hypothesis, we carried out a competition in a high osmolarity environment between a strain with wild-type rpoS and a strain with a deletion of the rpoS locus. As expected, the ΔrpoS strain was less fit than its wild-type ancestor (Table 1). This was not due to another mutation elsewhere in the genome, as other reconstructions of this same strain pairing showed the same cost (i.e. the difference in fitness) (Table 1, ANOVA, p = 0.89). The fitness cost was only slightly larger when the bacteria were grown with 0.44 M sucrose (t-test, p = 0.01), indicating that there was a general cost of osmotic stress and not only a salt-specific cost. Furthermore, the cost was specific to high osmolarity, as there was no fitness cost when the two stains were competed in the absence of NaCl stress (Table 1). Finally, the fitness cost was not due the activity of the kanamycin resistance gene used to knock out rpoS, as the same resistance cassette placed at the melB locus had no fitness cost (Table 1).

To allow strains to adapt to this high-osmolarity environment, we serially cultured five rpoS⁺ lines (denoted rpoS+1 to rpoS+5) and five ΔrpoS lines (ΔrpoS−1 to ΔrpoS−5) in a high-osmolarity medium for 250 generations, and isolated a single colony from the final population. When competed against their ancestor, all 10 evolved lines showed an increase in fitness (Figure 1). This was due, at least in part, to the fact that each increased its growth rate (Table S1). Importantly, and unexpectedly, there was no significant difference in the average fitness increase of the ΔrpoS lines and the rpoS⁺ lines (t-test, p = 0.18). In addition, the variance in fitness among the ΔrpoS lines was smaller than the variance among the rpoS⁺ lines (F-test, p = 0.02).

Was the fitness increase of the ΔrpoS lines a compensation for the lack of RpoS function, or simply a general adaptation to this environment? If it was compensation for lack of RpoS function, then adaptation should be epistatic on rpoS. To test this possibility, we restored the ΔrpoS lines to rpoS⁺. While rpoS⁺ is favored over ΔrpoS in the ancestral genetic background, all five ΔrpoS lines became less fit when transduced from ΔrpoS to rpoS⁺ (Figure 2A), indicating that fitness increased by compensating for the lack of RpoS, rather than by general adaptation to the culture conditions. To see if the rpoS⁺ lines adapted in an RpoS-dependent manner, we transduced these lines from rpoS⁺ to ΔrpoS. Although ΔrpoS was costly in the ancestral genetic background, it was even more costly in three of four of the evolved rpoS⁺ genetic backgrounds (p<0.05) (Figure 2B), indicating that adaptation of these lines was via an RpoS-dependent mechanism.

We used DNA microarrays to explore the patterns of regulatory evolution underlying our observed fitness changes. To assess if our observed patterns of mean change and variance in fitness were mirrored by changes in the transcriptome, we compared the expression pattern of both the rpoS⁺ and ΔrpoS lines to that of their ancestors. For each of the 4,254 genes represented on the array, we calculated the difference of each of the 10 evolved lines from its ancestor. For each gene, we then asked whether the average difference was larger for the rpoS⁺ or the ΔrpoS populations. The rpoS⁺ lines were more different from their ancestor than the ΔrpoS lines were from their ancestor for 61% of genes, significantly more than would be expected by chance (p<10⁻¹⁶, sign test). We used the same approach to assess the variability in expression patterns, and found that the expression of 90.7% of genes was more variable in the rpoS⁺ lines than in the ΔrpoS lines (p<10⁻¹⁶, sign test). The expression level of most genes was unchanged in most strains (Table 2), so we repeated this analysis with the expression level set to the ancestral value for all measurements that did not pass our statistical threshold (FDR = 0.001). Of those genes that showed change, 91.5% changed more in the rpoS⁺ lines than in the ΔrpoS lines (p<10⁻¹⁶, sign test). Further, 92.3% of genes were more variable in the rpoS⁺ lines than the ΔrpoS lines (p<10⁻¹⁶, sign test). These results neatly paralleled our fitness results, showing that highly similar trajectories of fitness increase in ΔrpoS were underlain by similarly parallel changes in patterns of transcription.

The fitness results also showed that the ΔrpoS lines evolved to compensate for the lack of RpoS function. Did they do so by returning transcription back towards original wild-type levels, or did adaptation result in the transcriptome becoming even more different from wild type? We found that 331 genes (or 7.8% of the genome) differed significantly in expression between the ancestral wild type and the ancestral ΔrpoS strains during growth in high osmolarity media. The evolved ΔrpoS lines showed significant changes in the level of expression of between 81 and 156 genes from the ΔrpoS ancestor (Table 2). Of these, 37 were changed significantly in all five lines (Table S2). All 37 were genes that also differed between wild type and ΔrpoS. We compared the average expression level of these 37 genes in the five ΔrpoS lines with their level in their ΔrpoS ancestor, and in the rpoS⁺ progenitor. In 35 of 37 cases, the expression level of the ΔrpoS lines evolved to be more similar to rpoS⁺ than was their ΔrpoS ancestor (Figure 3). Thus, compensation for the lack of RpoS function involved partial restoration of the wild-type pattern of transcription.

What sorts of mutations brought about these parallel changes in the expression of 37 genes? To address this question, we focused first on the two genes with the most dramatic changes in expression in the evolved lines. These genes, otsB and otsA were upregulated an average of 12.5 and 9.7 fold, respectively, in the ΔrpoS lines. These genes code for the two enzymes required for trehalose biosynthesis and E. coli synthesizes trehalose under osmotic stress in order to achieve internal osmotic balance [16]. otsB and otsA form an operon (otsBA) that requires RpoS for transcription in wild-type E. coli [17], suggesting that the ΔrpoS lines had evolved RpoS-independent expression of this operon.

To determine whether this new expression pattern was via a mutation in the otsBA promoter or a mutation elsewhere in the genome, we sequenced the otsBA promoter of all five lines and found that all five contained identical IS10 insertions. This insertion was located between the +12 and +13 bases relative to the transcriptional start site determined by Becker and Hengge-Aronis [18]. IS10 contains a promoter, P_OUT, directed outward from IS10 into adjacent DNA [19],[20]. In all five ΔrpoS lines, IS10 was oriented with P_OUT reading into the otsB gene. IS10 is known to have strong sequence preferences for insertion, and while the site of this insertion at otsB resembles the preferred site, it is not optimal. There is a marked preference for the symmetric site 5′-GCTNAGC-3′, but we found insertion at 5′-GTAAAGC-3′. The presence of a thymine instead of a cytosine at the second position lowers the insertion frequency over 1,000 fold from the preferred site in another tested context [21].

Surprised to find the same mutation in all five lines, we wanted to eliminate the possibility that this mutation fixed in all five lines because it occurred once before the starting culture was split into the five separate lines. If the mutation occurred in all five lines independently, i.e. after the culture was split, then we reasoned that if we started five completely independent lines, each should acquire the same IS10 insertion into P_otsBA. To test this hypothesis, we spread an aliquot of the original ΔrpoS frozen culture onto an L agar plate, picked five separate colonies (thus each founded by a single cell) and used each of these to found a new long-term line. Diagnostic PCR of the P_otsBA promoter confirmed that each of these lines began with a wild-type promoter. These lines were evolved for 250 generations under identical conditions to those used in the first experiment. After 250 generations, all five of these lines had acquired an IS10 insertion mutation in the same location and with the same orientation as the first five lines, indicating that the repeated evolution of this particular promoter mutation is not due to the mutation having been present in the starting population.

While all five ΔrpoS lines fixed the same mutation, the dynamics of this mutational sweep need not have been uniform. To explore this, we used QPCR to determine the frequency of the IS10 insertion into P_otsBA after 80 generations. The frequency of the insertion varied over two orders of magnitude (Table S3), and the frequency in all five cultures was distinct. This could be due to the insertion occurring at distinct time points in each line, or due to the initial rise of the adaptive mutation being dominated by the stochastic dynamics of culture transfer from flask to flask.

To determine if the IS10 insertion alone was sufficient to allow for RpoS-independent expression of otsBA, we cloned the wild-type and evolved P_otsBA promoters into the promoterless gfp-fusion plasmid pZep08 [22]. In the rpoS⁺ ancestor, the wild-type promoter was expressed even in low-osmolarity MOPS minimal medium, and was upregulated upon the addition of NaCl (Figure 4A). The P_otsBA promoter remained un-expressed in both the ancestral ΔrpoS background, and in the ΔrpoS−1 line. On the other hand, the evolved (IS10 inserted) P_otsBA promoter was expressed in all three strains (Figure 4B). The ancestral ΔrpoS line and the ΔrpoS−1 line expressed at similar levels, suggesting that there was not a second mutation beyond the IS10 insertion that allowed for otsBA expression in the evolved lines. The evolved promoter was expressed at a lower level in the rpoS⁺ strain than in either ΔrpoS line, which may explain why the evolved lines became less fit when made rpoS⁺.

What is the fitness effect of this IS10 insertion? When the IS10 insertion was moved into the ancestral ΔrpoS background, we found that the newly constructed strain had a fitness of 1.25 when competed against the ΔrpoS strain. The fitness advantage due solely to the IS10 insertion was not different from the fitness of the five evolved ΔrpoS lines (ANOVA, p>0.05), suggesting that the IS10 insertion was responsible for all of the adaptation. To complement these experiments, we transduced all five of the ΔrpoS evolved lines to a wild-type otsBA promoter. These strains were then competed against their ΔrpoS ancestor, and four of the five were now found not to be different to their ancestor (Figure 5). Only the ΔrpoS−3 line was significantly fitter than its ancestor when transduced to wild-type otsBA (fitness = 1.04, p = 0.003), suggesting that it has a second mutation beyond the IS10 insertion at that locus. There is no evidence that the other four strains have any other mutation that affects fitness in the high-osmolarity environment.

Did all five of the lines that contained the same IS10 do so because this was the only way to upregulate otsBA, or do other routes exist? To answer this question, we selected mutants that could upregulate an otsB-lacZY transcriptional fusion in a ΔrpoS background. Of 21 independent mutants, 19 possessed the same IS10 insertion as we recovered from our experimental evolution, while another had an IS10 insertion between bases −11 and −12 in the promoter (Figure 6). Finally, one mutant had a 6-bp deletion overlapping the −35 box of the promoter [18]. While the wild-type −35 (TGGCGA) box of the P_otsBA promoter differs strongly from the RpoD −35 consensus (TTGACA) [23], a sequence closer to the consensus (TTGCAA) lies just upstream in the wild-type promoter. The 6-bp deletion moved this other sequence into position to serve as a −35, presumably allowing RpoD-dependent transcription. These results demonstrate that other mutational routes to upregulation exist, but the IS10 insertion observed in our experimental evolution is the most likely to occur.

The nature of the regulatory network in E. coli that governs its response to stress creates trade-offs between fitness in high and low-stress environments. If selection cannot increase the RpoS-dependent response to stress, then novel regulatory mechanisms may evolve to compensate. We have experimentally evolved both rpoS⁺ and ΔrpoS lines in an osmotically stressful environment and explored how these populations adapted to this stress. We were surprised to find that the ΔrpoS populations evolved via the same mutation in each line, which did not result in a larger increase in fitness than the rpoS⁺ lines. We had expected that because they were less fit, the ΔrpoS lines would fix mutations causing a larger average fitness increase, as has been found in other cases of experimental evolution for compensation of deleterious mutations [9],[24]. The expectation of larger fitness increases is supported by models of the genetics of adaptation that follow from Fisher [25],[26]. These models also predict that less fit genotypes will have more possible mutations resulting in more variation in fitness increases. We found the opposite: there was less variation in the fitness increase of ΔrpoS lines, due to the fact that the same mutation fixed in all five populations.

One possible explanation for our observation is that the deletion of rpoS causes the strain to cross a fitness valley, and places it on a smaller peak. On this new, smaller peak, adaptive mutations would be of smaller size, and there would be fewer of them, resulting in more parallel evolution. In biological terms, this implies that it is possible that trehalose biosynthesis is so critical that mutations upregulating otsBA will be much more strongly favored than any other adaptive mutation. If this is the case, then recovery of the same mutation in replicate lines may reflect the limited number of possible adaptive mutations that can upregulate otsBA. Alternatively, it may reflect the fact that the site in the promoter is a hotspot for IS10 insertion. This latter possibility is supported by our observation that 90% of the mutants up-regulating otsBA that we recovered in our screen contained the same IS10 insertion as found in our experimental evolution.

All of the evolution in our ΔrpoS lines was mediated by IS10. While this constancy may be unusual, IS elements have been frequently found as the causes of adaptive mutations in experimental evolution [27]–[29]. In addition, comparative genomics suggests that IS elements play an important role in genome evolution [30]–[32]. While the specific role of IS elements in regulatory evolution has been less thoroughly explored, a wide range of IS elements are known to activate transcription by insertion into promoters [12]. Further, IS elements are found in almost all strains of E. coli. In their survey of a representative collection of E. coli isolates, Sawyer et al. [33] found that 97% contained at least one of the six IS elements for which they screened. They did not screen for IS10, but genomic DNA sequencing reveals that it is not rare: it is present in 16% of E. coli and Shigella genomes in GenBank. Thus, it is not improbable that IS elements play an important role in the evolution of the RpoS regulon.

The IS10 insertion in P_otsBA affected the transcription of a number of genes beyond only otsBA. The process of compensation for a fitness cost might result in two sorts of patterns of change of transcription. In the first, deletion of RpoS changes the level of transcription of a gene, and compensatory adaptation brings the level of transcription back towards the wild-type level. An alternative is that the process of compensation results in more change in the level of transcription. For example, a gene that is somewhat down-regulated with the loss of RpoS becomes even further down-regulated, or a gene that was not initially affected by the loss of RpoS is upregulated to compensate for the lack of transcription of some other gene. We found that of the 35 genes other than otsBA with changed pattern of transcription across all 5 lines, 33 fell into the former category of (partial) restoration of transcription levels. OtsB and OtsA are biosynthetic enzymes, not transcription factors, but the small molecule they synthesize, trehalose, does stimulate expression of genes involved in maltose transport [34], and the malK-lamB-malM operon is up-regulated in the evolved lines. Other changes in transcription may be responses to other restored physiological states, such as growth rate (the gene rmf [35]), or osmotic balance (proP [36], pspABC [37]). A general implication of these observations is that regulatory systems may be structured such that in the absence of reversion at the genetic level, compensatory mutations can lead to restoration of wild-type patterns of transcription.

What are the consequences for future evolution of compensating for the lack of RpoS with an IS10 insertion? The SPANC model put forth by Ferenci and co-workers [3],[4] proposes that selection will favor decreased levels of RpoS activity in nutrient-poor environments, and increased levels in stressful environments. If null rpoS alleles that cannot be reverted are fixed in nutrient-poor environments, then strains may compensate in stressful environments with mutations elsewhere in the regulatory network. Because epistasis is generated in this compensation, subsequent re-acquisition of RpoS function will decrease fitness. Strains that have taken the first step of adapting in an RpoS-independent manner will continue to do so, leading to even further divergence of regulatory networks.

A mechanism to explain the origins of the observed epistasis, sigma factor competition [38],[39], suggests that epistasis is likely to be a general phenomenon. Sigma factor competition is hypothesized to occur because amounts of core RNA polymerase are limiting for transcription. Thus, if an RpoS molecule interacts with a core subunit to promote transcription from one promoter, there is one less core subunit available for any other sigma factor to interact with and promote transcription. Since the P_OUT promoter of IS10 is not an RpoS-dependent promoter, the presence of functional RpoS protein will reduce the levels of transcription from promoters like P_OUT. Because this cause of epistasis is embedded in the fundamental process of transcription, it suggests that mutations recruiting non-RpoS-dependent promoters to compensate for the loss of RpoS will generally be epistatic on the absence of RpoS.

The RpoS regulatory network is a major target for selection because it cannot handle environments that are both physically stressful and nutrient-poor. The SPANC model posits that differing levels of RpoS activity will be selected in different environments. Our work has shown an alternative to increasing RpoS levels in a stressful environment. Strains adapt with a repeatable pattern by up-regulating a single pair of genes on the periphery of a regulatory network, suggesting that regulatory networks may evolve novel structures in a rapid and predictable manner.