Understanding how diversity emerges in a single niche is not fully understood. Rugged fitness landscapes and epistasis between beneficial mutations could explain coexistence among emerging lineages. To provide an experimental test of this notion, we investigated epistasis among four pleiotropic mutations in rpoS, mglD, malT, and hfq present in two coexisting lineages that repeatedly fixed in experimental populations of Escherichia coli. The mutations were transferred into the ancestral background individually or in combination of double or triple alleles. The combined competitive fitness of two or three beneficial mutations from the same lineage was consistently lower than the sum of the competitive fitness of single mutants—a clear indication of negative epistasis within lineages. We also found sign epistasis (i.e., the combined fitness of two beneficial mutations lower than the ancestor), not only from two different lineages (i.e., hfq and rpoS) but also from the same lineage (i.e., mglD and malT). The sign epistasis between loci of different lineages indeed indicated a rugged fitness landscape, providing an epistatic explanation for the coexistence of distinct rpoS and hfq lineages in evolving populations. The negative and sign epistasis between beneficial mutations within the same lineage can further explain the order of mutation acquisition.

Understanding how diversity emerges and is maintained in a single niche is crucial to understanding evolution and the vast microbial diversity in nature. Experimental evolution studies with bacteria and yeast have demonstrated that genetic diversity can arise even in an initially homogeneous culture supported by a single resource (Maharjan et al. 2006; Gresham et al. 2008; Kinnersley et al. 2009). A recent genotype–phenotype mapping study suggested that several sources of coexistence are operating simultaneously in the same diversifying experimental population (Maharjan et al. 2012). A pertinent possibility that has not been investigated as yet is whether genetic interactions among different loci (epistasis) contribute to divergence in experimental populations; after all, epistasis is considered to be a force in reproductive isolation and speciation (Anderson et al. 2010; Presgraves 2010). Epistasis can create single or multiple adaptive peaks (fitness optima) separated by fitness minima, also known as rugged fitness landscapes (Weinreich et al. 2005; Poelwijk et al. 2007). Here, we investigate the significance of epistasis in the emergence and maintenance of coexisting lineages in experimental lab populations.

Recent studies have demonstrated epistatic interactions between beneficial mutations in bacterial and yeast model systems and their influences on evolutionary trajectory (Chou et al. 2011; Khan et al. 2011; Kvitek and Sherlock 2011; Tenaillon et al. 2012). These studies have found two important types of epistatic interaction: a widespread presence of negative epistasis among beneficial mutations leading to the diminishing returns of such mutations (Chou et al. 2011; Khan et al. 2011) and sign epistasis among some beneficial mutations leading to a rugged fitness landscape (Kvitek and Sherlock 2011; Tenaillon et al. 2012). In addition, genome sequencing of evolved clones from parallel populations have identified many mutations under various selection conditions (Barrick et al. 2009; Conrad et al. 2011; Dettman et al. 2012; Maharjan et al. 2012). In many cases, strains of bacteria repeatedly acquire mutations in a particular order, with a strong indication of association between some beneficial mutations. For example, mutations in the highly pleiotropic rpoS, mglD, and malT genes appeared consecutively in the same lineage in four long-term glucose-limited chemostat cultures of Escherichia coli (Maharjan et al. 2012). However, hfq mutations were always acquired in separate strains in the same populations (Maharjan et al. 2012). Indeed, the separate rpoS and hfq lineages make up more than 50% of the isolates that coexist over many generations (Maharjan et al. 2012). However, these studies did not follow the role of epistasis in population diversity so the question remains whether positive epistasis between strongly associated mutations or sign epistasis between strongly dissociated mutations contributes to diversifying lineages.

In this communication, we report experimental results on epistatic interaction among strongly associated mutations, rpoS, mglD, and malT and between strongly dissociated mutations, rpoS and hfq. A detailed analysis of fitness properties of single, double, and triple mutants of these four alleles permitted the identification of two types epistatic interactions, negative and sign epistasis. A combination of hfq and rpoS mutations exhibited sign epistasis, permitting a rugged fitness landscape and providing an epistatic explanation for the coexistence of distinct rpoS and hfq lineages in populations.



All E. coli strains used in this study are listed in Table 1. To construct the strains with single, double, or triple mutants of rpoS, mglD, and malT, the evolved alleles of rpoS (E315stop) and mglD (S123 stop) of BW3767 were transferred into BW2952 or MC4100 by P1 transduction using the method described previously in (Notley-McRobb and Ferenci 1999a; Maharjan et al. 2010; Wang et al. 2010). To transfer the evolved malT (A236S), we first constructed aroB::bla strain using the protocol described in Yu et al. (2000). The proximity of aroB::bla locus to the malT allowed >90% cotransduction (Maharjan et al. 2013b). We also transferred evolved rpoS of BW3767 into BW5163 using the method described previously in Wang et al. (2010).

Table 1. List of strains used in this study
StrainIsolate/relavant genotypeReferences
MC4100F araD139 Δ(argF-lac)U169 rspL150 deoCl relA1 thiA ptsF25 flb5301 rbsR(Ferenci et al. 2009)
BW2952MC4100 malG::λplacMu55 ϕ(malG::lacZ)(Ferenci et al. 2009)
BW3454MC4100 metC::Tn10(Wang et al. 2010)
BW3767Chemostat evolved isolate(Maharjan et al. 2013a)
BW5163BW2952 Hfq Y25D(Maharjan et al. 2010)
BW5304BW2952 malT3767This study
BW5312BW2952 rpoS malT3767This study
BW5317BW2952 cysD::blaThis study
BW5318BW2952 rpoS3767This study
BW5320MC4100 mglD3767 aroB::blaThis study
BW5321MC4100 mglDmalT3767This study
BW5322MC4100 mglDmalT3767cysD::blaThis study
BW5323MC4100 ropSmglDSmalT3767This study
BW5324BW2952 mglDmalT3767This study
BW5325BW2952 mglD3767This study
BW5326BW2952 rpoSmglDSmalT3767This study
BW5327BW2952 rpoSmglD3767This study
BW5329BW5163 rpoS3767This study
BW5332BW5163 cysD::blaThis study
DY330W3110 ΔlacU169 gal490 λcl857 Δ(cro-bioA)(Yu et al. 2000)

Bacteria were cultured in L-broth (Miller 1972) or minimal medium A (MMA) supplemented with either 0.2% (w/v) glucose for batch cultures or 0.02% (w/v) glucose for chemostats. In both glucose-limited batch and chemostats, culture media were also supplemented with 1 μg/mL thiamine and 1 mM MgSO4. For chemostat competition, experiments 4 μg/mL methionine was also added to the medium. All growth and competition experiments were carried out at 37°C.


Fitness comparisons were made against a tetracycline-resistant derivative of BW2952 (BW3454) carrying a metC::Tn10 insertion in medium supplemented with 4 μg/mL methionine. This strain was of equal fitness to the ancestor in glucose-limited chemostats. Where possible, head-to-head competitions between mutants were also performed without selectable markers. In such cases, the proportion of competing strains was monitored on the basis of their phenotypic properties. For example, colonies of rpoS-negative strains on Luria-agar plates do not stain with iodine solution, whereas hfq strains do (light brown color). We therefore monitored their proportions in the mixed population during competitions by staining with iodine solution after plating appropriate dilutions on Luria-agar plates. Two to four chemostat competition experiments were conducted as previously described in Maharjan et al. (2010) and Wang et al. (2010). The reported relative fitness was based on the equations of Dykhuizen and Hartl (1983), measured in terms of Malthusian parameters, with the reported selection coefficient (S) determined from the slope of the linear regression of ln[p(t)/q(t)], where p(t) and q(t) represent the relative frequencies of the strains from at a given time point (Dykhuizen and Hartl 1983). At least five different time points within 24 h of competitions were used for estimating the selection coefficient. The t-test was used to evaluate whether the mean fitness differed from the other mutant strains using the two-tailed t-test assuming two-sample unequal variances.


To investigate how genetic interactions among beneficial mutations shape evolutionary diversification, we first analyzed the association and dissociation of 107 mutations that had been identified by whole genome sequencing of 27 evolved clones from previously described four long-term glucose-limited chemostat cultures (Maharjan et al. 2012). We found at least two distinct patterns of mutational accumulation in coexisting strains. First, mutations in the highly pleiotropic rpoS, mglD, and malT genes appeared consecutively in the same lineage as shown in Figure 1A. Second, hfq and rpoS mutations never appeared in the same background. The apparent dissociation between the hfq and rpoS lineages (Fisher's exact test: P = 0.002) indicates a strong repulsive interaction between them. We therefore hypothesized that these strongly dissociated mutations exhibit sign epistasis whereas the associated rpoS, mglD, and malT mutations do not.

Figure 1.

Fitness and fixation of mutations in Escherichia coli populations evolving under glucose-limited chemostats. (A) A box plot showing times of appearance of mutations in rpoS, mglD, and malT averaged over four parallel populations (Notley-McRobb et al. 2003). (B) The fitness landscape of genotypes containing the rpoS, mglD, and malT alleles in Table 1. Each node in the cube represents a genotype with evolved or wild-type alleles of rpoS, mglD, and malT. re, ge, and me represent the evolved alleles of rpoS, mglD, and malT respectively, whereas rw, gw, and mw represent the wild-type alleles of rpoS, mglD, and malT, respectively. The arrows point toward the genotypes with higher fitness. The double mutant (rwgeme) is less fit than two single mutants (rwgemw and rwgwme) and therefore not accessible from either of single mutants. The numbers at each edge of the cube are fitness values (M ± SEM) measured as the selection coefficient as described previously in Maharjan et al. (2013a). The fitness values presented were obtained from at least two independent competition experiments. All fitness values were based on competitions against the reference ancestor strain. The thick lines indicate the order of mutations based on the time line of their fixation in replicate population as shown in (A).

To test the above hypothesis, we constructed a series of bacterial mutants with single, double, or triple combinations of the beneficial mutations in rpoS, mglD, malT, and hfq (Table 1). We then estimated the fitness as the selection coefficient (S) of competing strains by competing them individually against the marked reference ancestor (BW3454), as shown in Figures 1B and 2. The mean fitness of this marked reference strain (S = 0.004 ± 0.005) against the unmarked ancestor (BW2952) was not significantly different (t = 1.69; degrees of freedom [df] = 3; 2-tailed P = 0.189). For fitness estimations, because epistatic interaction can be affected not only by genetic background but also the environment in which they are tested (Remold and Lenski 2004), competition experiments were performed under the same environmental conditions as used for the selection of evolved clones, that is glucose-limited chemostats at a dilution rate of 0.1 h−1 (Maharjan et al. 2012). Consistent with previous studies (Notley-McRobb and Ferenci 1999a,b; Maharjan et al. 2010; Wang et al. 2010), all four individual mutations resulted in significantly increased fitness against the ancestral strain (Table 2; t > 3.0, df = 3, 2-tailed P < 0.05 in each case). The mean fitness conferred by mutation in rpoS (S = 0.081 ± 0.009) was slightly higher than the mean fitness conferred by hfq (S = 0.075 ± 0.003), however the difference was not statistically significant (t = 0.62, df = 3, 2-tailed P = 0.39). The head to head competition between rpoS and hfq strains also showed a marginally higher fitness of the rpoS strain (S = 0.005 ± 0.009), but was not statistically significant (t = 1.26, df = 3, 2-tailed P = 0.289), which further demonstrated that they have similar fitness under the condition used for the selection of these mutations. However, the fitness conferred by mglD (S = 0.044 ± 0.002) and malT (S = 0.028 ± 0.010) was significantly lower than either of rpoS and hfq (Table 2; mglD: t = 11.0, df = 4, 2-tailed P = 0.0004 and malT: t = 6.8, df = 2, 2-tailed P = 0.021 against hfq). The order of fitness contributions of the four individual mutations is rpoS = hfq > mglD > malT (Fig. 2; Table 2).

Table 2. Observed and expected additive fitness of single, double and triple mutants of rpoS, mglD, malT, and hfq relative to the reference ancestral strain BW3454
Data point numberStrainMutational combinationFitness (95% CI)aObserved or expectedStandard Deviationa
  1. a

    The confidence interval and standard deviation were estimated from two to four independent assays. Expected additive fitness values for each combination of mutations were calculated from experimentally observed fitness values. N/A = not applicable.

1BW5318rpoS0.081(0.066 to 0.096)Observed0.009
2BW5163hfq0.075(0.070 to 0.081)Observed0.003
3BW5325mglD0.044(0.040 to 0.048)Observed0.002
4BW5304malT0.028(0.013 to 0.044)Observed0.010
5BW5312rpoS, malT0.083(0.082 to 0.084)Observed0.000
6BW5327rpoS, mglD0.104(0.080 to 0.129)Observed0.010
7BW5326rpoS, malT, mglD0.124(0.106 to 0.142)Observed0.011
8BW5324mglD, malT−0.052(−0.081 to −0.022)Observed0.012
9BW5329rpoS, hfq−0.026(−0.044 to −0.004)Observed0.012
N/ABW5326rpoS, mglD, malT0.153(0.146 to 0.161)Expected0.002
N/ABW5327rpoS, mglD0.126(0.105 to 0.147)Expected0.009
N/ABW5312rpoS, malT0.110(0.108 to 0.111)Expected0.000
N/ABW5324mglD, malT0.071(0.044 to 0.098)Expected0.009
N/ABW5329rpoS, hfq0.158(0.120 to 0.195)Expected0.012
Figure 2.

Sign and negative (diminishing return) epistasis among rpoS, hfq, mglD, and malT beneficial mutations. To identify the nature of epistasis, fitness of single mutants and double or triple mutants were measured against the ancestral reference strain as described Methods section and plotted against the expected (additive) fitness values assuming no epistasis. If the measured fitness falls on the diagonal solid line, it indicates the interaction is additive or no epistasis. If the measured fitness falls above or below the diagonal solid line, it indicates the interaction is either positive or negative (diminishing returns). Similarly, if the measured fitness falls below the ancestral base-line (dotted) the effects are sign. Numbers in each data point represent the mutational combinations (1 = rpoS, 2 = hfq, 3 = mglD, 4 = malT, 5 = rpoS-malT, 6 = rpoS-mglD, 7 = rpoS-mglD-malT, 8 = mglD-malT, and 9 = rpoS-hfq). Y-axis bars represent standard deviations estimated from two to four independent assays. Details of mutational combinations represented by each numbers including their observed and expected fitness data and confidence intervals can be found in Table 2.

However, in all cases double and triple mutants are significantly less fit than the sum of the effects of single mutations (Table 2; rpoSmglD: t = 3.19, df = 3, 2-tailed P = 0.050; rpoSmalT: t = 83.9, df = 4, 2-tailed P < 0.0001, and rpoSmglDmalT: t = 5.1, df = 3, 2-tailed P = 0.014), which is explained by negative epistasis among these beneficial mutations (Fig. 2). Strikingly, the fitness (S = −0.026 ± 0.012) of a constructed rpoS-hfq double mutant was significantly lower than that of the ancestor (t = 5.3, df = 4, 2-tailed P = 0.006), which suggested that the combined effect of the hfq and rpoS mutations resulted in sign epistasis. In no experimental populations studied have rpoS clones harbored an hfq mutation or vice versa. These results are consistent with the conclusion that mutations in rpoS and hfq in the same lineage are mutually exclusive. The sign epistatic interaction can clearly constrain evolutionary trajectories in chemostat populations by creating a rugged fitness landscape (Weinreich et al. 2005; Poelwijk et al. 2007).

The hfq and rpoS mutations are highly pleiotropic (Maharjan et al. 2010; Wang et al. 2010), and both mutations reduce stress protection (Ferenci 2005, 2008; Maharjan et al. 2010; Wang et al. 2010), resulting in lower viability in chemostats (Notley-McRobb et al. 2002; Maharjan et al. 2013a). Although the precise molecular mechanism of the observed deleterious fitness effect of the rpoS-hfq double mutant has not been identified, we suspect that this could be due to an elevated mortality rate. In fact, we found that the cell density at 600 nm of the 18 h old glucose-limited chemostat of rpoS-hfq double mutant was only 54% of that of the ancestor, and previous studies have demonstrated that a mutation in hfq resulted in reduced cell viability possibly due to cell division errors (Tsui et al. 1994; Vecerek et al. 2008).

An unexpected sign epistasis was also found between mglD and malT, despite being always found in the same lineage. Mutational combinations harboring only mglD and malT mutations are thus unlikely to persist in our selection environment and genetic context. However, when an rpoS mutation was introduced into the mglD-malT double mutant, the fitness of the resulting triple mutant was significantly higher than that of any other combination (Table 2). The rpoS mutation not only alleviated the deleterious effect but also turned the triple mutant into a highly beneficial genotype. Thus, the observed pathway in chemostats (Fig. 1A), with rpoS arising first then mglD and malT, is consistent with these epistatic effects. Despite the high fitness of the triple mutant, the slight negative epistasis between the mutations resulted in cumulative diminishing returns among these beneficial mutations (Fig. 2). Another interesting observation is the lag between the fixation time of mglD and malT (Fig. 1A). This is most likely due to differences in the size of the mutational target sites in these genes. The observed mutations in malT were always gain-of-function or constitutive mutations at a limited number of sites (Notley-McRobb and Ferenci 1999b; Maharjan et al. 2012), whereas mutations in rpoS and mglD were much more common loss-of-function mutations throughout the gene (Maharjan et al. 2012).

The reason why malT and mglD mutations in combination result in negative fitness is not readily explained by the known roles of these genes. Both the mal-genes and the mgl genes regulated by the two transcriptional regulators, respectively, have easily explained roles under glucose limitation; both mutations increase glucose uptake (Ferenci 1996). The unexpected result that the combination is deleterious indicates our lack of a full understanding of E. coli physiology and regulation. The defect cannot be in transport because the rpoS mutation relieves the defect, so it is more likely to be due to unknown transcriptional-regulatory effects of the malT and mglD mutations in combination.

The observed sign epistasis between coexisting lineages and within the same lineage in the evolving populations not only provides direct evidence for the importance of epistasis, but also has implications for our understanding of evolutionary biology (Wright 1988; Whitlock and Phillips 1995). Earlier, it was believed that it is impossible to prove empirically that multiple fitness peaks exist (Whitlock and Phillips 1995), because intermediate or modifier mutations should be able to form ridges connecting the valleys and thereby providing accessible paths between two peaks (Gavrilets 1999; Lunzer et al. 2010). To our knowledge, only two empirical studies have shown the existence of ridges between two fitness peaks (Lunzer et al. 2010). In chemostats there appears to be no connection between the rpoS and hfq fitness peaks.

In exploring the epistatic interactions among the beneficial mutations, it is worth considering whether the fitness characteristics of the mutations indeed explain the coexistence of the two major lineages in approximately the same proportions for at least 2 weeks in glucose-limited chemostats (Maharjan et al. 2012). The initial rpoS and hfq mutations have approximately the same fitness (Fig. 2), but the additional mutations in the rpoS lineage give a fitness advantage over hfq alone. The hfq lineage is competitive because at the end of the experiment, genome resequencing of multiple clones of hfq isolates from the 37-day-old chemostat showed additional, as yet undefined mutations that potentially maintain the relative the fitness of the hfq lineage (result not shown). Hence once established, the rpoS and hfq lineages are maintained through fitness equivalence rather than additional, nontransitive interactions.

In conclusion, the results in Figures 1 and 2 demonstrate that epistatic interactions have a significant role in evolutionary pathways, because both the order of mutation acquisition and the distinct pathways were found to be at least partly dependent on epistasis. We observed no positive epistatic effects, but negative and sign interactions instead. The negative epistasis among beneficial mutations in chemostat evolution was consistent with previous observations on diminishing returns of new mutations in other selection environments (Chou et al. 2011; Khan et al. 2011; Kvitek and Sherlock 2011; Tenaillon et al. 2012). More importantly, the sign interaction between beneficial mutations from different lineages resulted in a rugged fitness landscape offering a general explanation of divergence in a single niche.

Associate Editor: I. Gordo


The authors thank the Australian Research Council for funding support and two anonymous reviewers for their helpful comments on the earlier version of this paper. The authors declare that no competing interests exist.