Bacterial strains

We obtained the WT strain of E. coli K-12 MG1655 from the Coli Genetic Stock Centre (CGSC, Yale University), streaked it on Luria Bertani (LB) agar (Miller), and chose one colony at random as the WT ancestor for subsequent experiments. We similarly obtained the mutator strains of E. coli (ΔmutT, ΔmutH, ΔmutL, ΔmutS, Δnth, Δnei, ΔmutY) from the Keio collection of gene knockouts from CGSC (BW25113 strain background [19]). These gene knockouts were made by replacing open reading frames with a Kanamycin resistance cassette, such that removing the cassette generates an in-frame deletion of the gene. The design of gene deletion primers ensured that downstream genes were not disrupted due to polar effects. For each mutator strain, we moved the knockout locus from the BW25113 background into the MG1655 (our WT) background using P1 phage transduction [20]. We then removed the kanamycin resistance marker by transforming kanamycin-resistant transductants with pCP20, a plasmid carrying the flippase recombination gene and ampicillin resistance marker. We grew ampicillin-resistant transformants at 42°C in LB broth overnight to cure pCP20 and streaked out 10 µL of these cultures on LB plates. After 24 hours, we replica-plated several colonies on both LB + kanamycin agar plates and LB + ampicillin agar plates, to screen for the loss of both kanamycin and ampicillin resistance. We PCR-sequenced the knockout locus to confirm removal of the kanamycin cassette. For generating the Δnth-nei double knockout, we first created a Δnth strain, and then moved the Δnei locus into this background using P1-phage transduction as described above. In the process of making gene knockouts, all mutator strains (except Δnth-nei) acquired background mutations (S1 Data) due to multiple generations of growth that occurred during the screening process.

Mutation accumulation (MA) experiments

We used MA experiments of varying length to obtain mutator strains carrying single mutations that would reflect the mutation spectrum of each mutator (~100 per strain). We isolated a single colony of each ancestor, suspended it in LB broth, and plated it to obtain as many colonies as were needed for each independent MA line. We used this same broth culture within 3–4 h of growth to extract DNA for whole-genome sequencing (WGS) (see next section). For mutators with very high mutation rates (∆mutH, ∆mutT, ∆mutL, and ∆mutS), each MA block had its own ancestor (since each time cells are grown up from freezer stocks, there is a very high probability that new mutations will arise) whose sequence was used to subtract the ancestral mutations from offspring lines (see below). For mutators with intermediate mutation rates and for WT, different MA blocks of a strain were started with the same ancestor.

The MA protocol minimizes the effect of selection, allowing sampling of a wide range of mutational effects, largely independent of their fitness consequences. For each MA line, every 24 h we streaked out a random colony (closest to a pre-marked spot) on a fresh LB agar plate. For MA experiments lasting more than a day, every 4–5 days we inoculated a part of the transferred colony into LB broth at 37°C for 2–3 h and froze 1mL of the growing culture with an equal amount of 60% glycerol at –80°C. For 1-day MA experiments, we similarly cultured and froze the final chosen colony. We used these freezer stocks of the MA lines for sequencing. We chose the length and number of replicate lines of the MA experiments for each strain depending on mutation rate and logistical feasibility. Since the WT had a relatively low mutation rate, we founded a small number of MA lines to make daily transfers feasible, but evolved them for many generations. For the mutators, we founded larger numbers of MA lines but evolved them for just a few generations. We also split the large number of MA lines into blocks (except ΔmutT; S1 Table) to make transfers and handling easier. Thus, most MA experiments were performed across at least two experimental blocks (S1 Table).

We founded multiple MA lines of each strain from single colonies: WT (98 lines), ΔmutT (300 lines), ΔmutH (350 lines), ΔmutL (350 lines), ΔmutS (350 lines), ΔmutY (430 lines), and Δnth-nei (300 lines) and propagated them through daily single-colony bottlenecks on LB agar plates (Table 1). We previously showed that our WT strain goes through ~27 generations in 24 h of growth on LB agar [16]. We used this estimate to calculate the number of generations elapsed in our MA experiments. We evolved WT lines in two experimental blocks: one block of 38 lines evolved for 300 days (8,250 generations, described previously in [11,16]), and a second block of 60 lines evolved for 85 days (2,295 generations) (S1 Table). We evolved mutators with intermediate mutation rates in short MA experiments: Δnth-nei in two blocks (block 1: 80 lines, 8 days, ~216 generations; block 2: 220 lines, 8 days, ~216 generations), and ΔmutY in three blocks (block 1: 300 lines, 12 days, ~314 generations, described previously in [11]; block 2: 80 lines, 5 days, ~135 generations; block 3: 50 lines, 1 day, ~27 generations; S1 Table). Finally, we evolved mutators with very high mutation rates in very short MA experiments: ΔmutS (block 1: 300 lines, 1 day, ~27 generations; block 2: 50 lines, 1 day, ~27 generations), ΔmutL (block 1: 300 lines, 1 day, ~27 generations; block 2: 50 lines, 1 day, ~27 generations), ΔmutH (block 1: 300 lines, 1 day, ~27 generations; block 2: 50 lines, 1 day, ~27 generations) and ΔmutT (block 1: 300 lines, 1 day, ~27 generations; S1 Table).

Whole-genome sequencing to identify clones with single mutations

We sequenced individual colonies from the MA experiments to identify all clones carrying a single mutation relative to their ancestor. For WT, Δnth-nei and ΔmutY, we inoculated 2 µL of the frozen stock of each evolved MA line into 2 mL LB broth, and allowed the cells to grow overnight at 37°C with shaking at 200 rpm. For ΔmutT, ΔmutH, ΔmutL, and ΔmutS, we allowed cells from frozen stocks to only grow for 3–4 h, to minimize the accumulation of additional mutations. Next, we extracted genomic DNA (GenElute Bacterial Genomic DNA kit, Sigma-Aldrich) and quantified it using the Qubit HS dsDNA assay kit (Invitrogen). We prepared paired-end libraries from each line and the respective MA ancestors, and sequenced them on an Illumina platform (either 2 × 100 bp, or 2 × 125 bp, or 2 × 250 bp). S1 Table provides details of the library preparation and sequencing methods used, as well as the sequencing depth achieved for each strain. WGS for some MA lines was unsuccessful, i.e., we obtained a very small number of reads or no reads at all; these lines were excluded from further analyses (Tables 1 and S1).

For each sample where WGS was successful, details of the number of mutations called are given in S2 Data. We aligned reads with average quality score > Q30 to the NCBI reference E. coli K-12 MG1655 genome (RefSeq accession NC_000913.2) using the Burrows-Wheeler short-read alignment tool BWA [21], and used SAMtools to further process the BWA outputs and generate pileup files [21]. Next, we used the default parameters in the VarScan package [22] to extract lists of base-pair substitutions and short indels (<10-bp length). We used Breseq with default parameters [23] to identify long indels and duplications. From these mutation lists, we only retained mutations that satisfied the following three criteria: (i) mutations represented on both the plus and the minus strand, (ii) mutations supported by at least 4 reads per strand, and (iii) mutations with frequency >80%. The first two filters would remove mutations with weak support, and the last filter would remove mutations that may have arisen either during the late stages of colony growth in MA experiments, or during the brief period of growth for stock preparation or DNA extraction. We performed this filtering using custom scripts written in R and Python. Finally, we used a custom R script to remove mutations present in the corresponding ancestor from the mutation list of each evolved line, and generated the final mutation list for each lineage (S3 Data). All scripts used for these analyses are available as Supporting information (S1–S7 Scripts).

We used several measures to maximize and estimate the accuracy and reliability of mutation-calling. To identify ancestral mutations, we sequenced ancestral MA clones at higher depth (S1 Table), and relaxed the 80% frequency filter such that we captured ancestral mutations segregating at lower frequency. This was especially important for strains with very high mutation rate. We confirmed that all offspring MA lines seeded by a given ancestral clone showed the expected set of ancestral mutations at very high frequency. To quantify the potential effects of secondary low-frequency mutations on our fitness measurements (and therefore the DFE), for each evolved MA line, we generated a separate list of mutations after relaxing the 80% frequency filter. We used the number and frequency of these secondary mutations to estimate the robustness of our results (described in the Results section). To determine the false negative rate of our WGS pipeline, we measured the recall rate of two known mutations in our WT ancestor (the progenitor of all our mutators) relative to the NCBI reference genome (RefSeq accession NC_000913.2) [16]: a G → A SNP at position 2845011 and a 2-bp CG insertion at position 4296380, expecting these mutations to be called at 100% frequency in all evolved MA lines. Finally, to determine the false positive rate, we measured the fitness of a subset of single-mutation clones of WT across two growth cycles, expecting a strong positive correlation if the identified mutations were real.

Estimating mutation rate, spectra, beneficial supply, and deleterious load

We estimated the mutation rate (μ) and mutation bias for each mutator and WT using mutations called from all sequenced isolates (Table 1, row “MA lines successfully sequenced”). We calculated μ (per bp per generation) as:

The number of mutations are given in Table 1, and the genome size is 4.6 × 10⁻⁶. The number of generations was calculated as:

We tested the observed frequency distribution of the number of mutations called in each lineage, against the expected Poisson distribution for random mutations. We calculated mutational biases from the different types of mutations observed in the MA-evolved lines (Table 1) as described in [11]. For instance, we calculated the Tv bias as:

We estimated 95% confidence intervals as 1.96 times the standard deviation of calculated bias. We estimated the confidence intervals for mutation rate as the known mean mutation rate ± the margin of error for a t-distribution with known mean and unknown standard deviation.

For each strain, we used the estimated mutation rates and the fractions of beneficial and deleterious mutations (f_b and f_d, see below) to calculate the predicted total supply of beneficial mutations as:

and the total genetic load due to deleterious mutations as

We estimated S_b and L_d using either the WT DFE (i.e., assuming that DFEs were invariant across strains), or using strain-specific DFEs measured in each environment. In each case, we also estimated confidence intervals as 1.96 times the standard deviation of S_b or L_d. We then compared these values for each strain, to quantify the impact of mutation bias on the supply of beneficial mutations and the deleterious load.

Growth rate measurements to construct single-mutation DFEs

From the complete set of all MA-evolved lines with WGS, we focused on those that had a single new mutation compared to the respective ancestor. This included 91 clones of ∆mutS, 97 ∆mutL, 100 ∆mutH, 102 ∆nth-nei, 94 WT (80 from block 1 [16] and 14 from block 2), 113 ∆mutY (79 from block 1 [11], 26 from block 2, and 8 from block 3), and 97 ∆mutT MA lines (Table 1, row “MA lines with single mutations”). We performed all fitness assays and subsequent analyses (described below) with these 694 isolates. We measured growth rates of each evolved MA line with a single mutation and its respective ancestor in two liquid culture media: LB broth (Miller, Difco) or M9 minimal salts (Difco) + 5 mM glucose. We inoculated each isolate from its freezer stock into either LB broth or M9 minimal salts medium with 5mM glucose, and allowed it to grow at 37°C with shaking at 200 rpm for 14–16 h. We inoculated 2 µL of this culture into 200 µL growth media in 96-well plates (Costar) and incubated the well plate in a Tecan F200 Multimode plate reader at 37°C with orbital shaking at 185 rpm for 16–18 h. Every 15 min, the plate reader measured the optical density (OD600) for all wells. In each plate, we included the MA ancestors relevant for the evolved MA isolates to enable fitness calculations, a reference strain (the parent WT strain) to enable checks for consistency across plate reader runs, and blank control wells to check for media sterility. For each evolved isolate, we used the average growth rate of three technical replicates to calculate the relative growth rate as:

For WT MA lines, we used the WT ancestor; and for mutator MA lines, we used the mutator ancestor. We estimated maximum growth rate, obtained from a linear fit to log OD600 versus time curves, using the Curve Fitter software [24]. The fitness effect of each mutation (s) was then calculated as:

We used s values of mutations to construct strain- and environment-specific distributions of fitness effects (DFE). Importantly, we corrected our DFEs for the expected selection bias in bacterial colonies in MA experiments as described before [11,17]. Briefly, the correction involves binning the measured selection coefficients into discrete bins, and then down-weighting the beneficial fraction of the DFE and over-weighting the deleterious fraction to account for the slightly higher probability of finding beneficial mutations. The bias-corrected versus uncorrected DFEs are shown in S7 and S8 Figs; note that the bias correction procedure leads to discretized bins in the corrected frequency distributions (DFEs).

Accuracy of mutation calling and fitness measurements

As described above, our main goal in this study was to construct and compare single-mutation DFEs across strains with distinct mutation biases. To maximize the accuracy of mutation calling, we used stringent sequencing quality filters (see Methods), e.g., only using lineages with high sequencing depth (mean ~ 40× across all lines, S1 Table). We also conducted several analyses to test the accuracy and reliability of our WGS pipeline. We first confirmed that two known mutations in our WT ancestor (the parent of all our strains) were called in every single evolved MA line with high frequency (SNP at ~100% allele frequency, indel with >93% read support; S1 Fig), indicating a very low false negative rate. Second, in all strains, the observed frequency distribution of the number of mutations called in each lineage was indistinguishable from the expected Poisson distribution for random mutations (S2 Fig), suggesting that non-random processes or mutation calling protocols did not significantly influence the outcome of our MA experiments. Third, we identified one-mutation clones only when they had a single mutation at >80% frequency; on relaxing this filter we found either no secondary mutations or secondary mutations segregating at low frequencies (S3 Fig). Together, these results suggest high accuracy of mutation-calling in our study.

Next, we tested the repeatability and reliability of our fitness estimates, which entailed measuring exponential growth rates of each of the single-mutation clones. In previous work, we had found high repeatability of growth rates [11,16]. We re-confirmed this for the current study: for a subset of isolates, relative fitness measured by two different experimenters in different years was strongly positively correlated (S4 Fig), as were fitness estimates in 48- versus 96-well plates performed in two different years (S5 Fig). These analyses gave us confidence that our measured fitness values are generally robust. Finally, for a subset of clones of WT, fitness values across two successive growth cycles were also strongly positively correlated (S6 Fig). Together, these results indicate that the single mutations that we called were “real”, and that our fitness measurements were reliable.

Transversion-biased strains have a right-shifted distribution of fitness effects (DFE) with a higher proportion of beneficial mutations and lower deleterious load

To test our hypothesis that transversion-biased E. coli strains can access more beneficial mutations, we constructed single-mutation DFEs for strains with different transition/transversion biases (Table 1). Each strain acquired distinct mutations, with no shared mutations (S3 Data). For each strain, we characterized the fitness effects of a total of ~100 single mutations in two growth media: one rich (LB) and one relatively poor (minimal medium with glucose), using exponential growth rate as a measure of fitness. Using these fitness estimates, we first constructed raw DFEs, and then corrected them to account for selection bias during MA (S7 and S8 Figs). Note that the bias correction does not alter the selection coefficients measured for each mutation, but directly modifies the DFE to down-weight the proportion of beneficial mutations (see Methods). The median selection coefficient of single mutations (estimated from the corrected DFEs) varied from −0.175 to +0.025 (i.e., a 17.5% reduction to 2.5% increase in growth rate) in LB, and from −0.013 to +0.063 (i.e., 1.3% reduction to 6.3% increase in growth rate) in glucose.

As predicted, in both environments, strains that reinforced the WT (ancestral) mutation bias (i.e., were strongly Ts-biased: ΔmutS, ΔmutL, ΔmutH, and Δnth- nei) had left-shifted DFEs relative to WT (Figs 2 and 3). In contrast, the DFEs of strains that opposed the WT mutation bias (i.e., had a strong Tv bias, ΔmutY and ΔmutT) had relatively right-shifted DFEs. The DFE differences across strains were reflected in the proportion of beneficial mutations (f_b), which was significantly higher in Tv-biased versus Ts-biased strains (Fig 4A; S2 and S3 Tables). Concomitantly, the fraction of deleterious mutations (f_d) was significantly lower in Tv-biased strains (Fig 4A; S2 and S3 Tables). These patterns did not change when we imposed more stringent conditions for single-mutation calling, e.g., if we constructed DFEs using only clones with no secondary mutations or with low-frequency secondary mutations, or clones with two mutations (S9 Fig). Similarly, reducing sample sizes commensurate with the stringent filtering did not alter the patterns (S10 Fig). Thus, the results are robust, and support our main prediction (Fig 1) that mutation bias shifts that reinforce the Ts bias of WT E. coli will have left-shifted DFEs and those that reverse the bias (i.e., are transversion biased) will have right-shifted DFEs (Fig 4A). A related prediction is that the magnitude of the DFE shift is correlated with the magnitude of the bias shift; our results are also consistent with this prediction (S11A and S11B Fig, f_b increases significantly with increasing Tv bias; S11C and S11D Fig, f_d tends to reduce with Tv bias, but not significantly so). However, we caution that given the limited number of bias-shifted strains in our dataset, additional work is required to adequately test this correlation. Despite the overall trends described above, we note some exceptions. In LB, ΔmutY had a much higher f_b compared to ΔmutT, despite a slightly weaker Tv bias (Fig 4A and 4B; S2 Table). In glucose, ΔmutL and ΔmutH had higher f_b values despite the same transition bias as ΔmutS and a stronger Ts bias than Δnth-nei (S3 Table). We explore these exceptions in more detail in the Discussion section.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Distribution of fitness effects of new mutations in a rich medium (Luria Bertani broth, LB) for Escherichia coli strains with varying mutation bias.

The distribution of fitness effects of single randomly occurring mutations in each strain (sample size in parentheses), calculated as maximum growth rate relative to the respective ancestor (x-axis). The schematic at bottom right shows the transversion (Tv) bias of each strain. The WT strain is shown in cyan; strains that reinforce the WT transition bias are in purple; strains that reduce or reverse the WT bias (i.e., have a transversion bias) are in pink. Each DFE was corrected for selection bias that could occur during MA; raw (uncorrected) DFEs are shown in S7 Fig; all raw fitness values are provided in S3 Data. Gray areas indicate neutral mutations (s = 0 ± 0.05 to account for experimental measurement error); bold lines and numbers indicate median values of s. Data underlying this figure are given in S4 Data.

https://doi.org/10.1371/journal.pbio.3003282.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of fitness effects of new mutations in a minimal medium (M9 broth with 5 mM glucose) for Escherichia coli strains with varying mutation bias.

The distribution of fitness effects of single randomly occurring mutations in each strain (sample size in parentheses), calculated as maximum growth rate relative to the respective ancestor (x-axis). The schematic at bottom right is identical to the schematic in Fig 2, and shows the transversion (Tv) bias of each strain. The WT strain is shown in cyan; strains that reinforce the WT strain’s transition bias are in purple; strains that reduce or reverse the WT bias (i.e., have a transversion bias) are in pink. Each DFE was corrected for selection bias that could occur during MA; raw (uncorrected) DFEs are shown in S8 Fig; raw fitness values are provided in S3 Data. Gray areas indicate neutral mutations (s = 0 ± 0.025 to account for experimental measurement error); bold lines and numbers indicate median values of s. Data underlying this figure are given in S5 Data.

https://doi.org/10.1371/journal.pbio.3003282.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Differences in DFEs of strains are most strongly associated with transversion bias.

In panels A—F, bars are colored by the mutation bias of each strain, as indicated in Figs 2 and 3. (A) Stacked bar plots show the total fraction of neutral, deleterious, and beneficial mutations observed in the DFEs of each strain in each growth medium, extracted from the DFEs shown in Figs 2 and 3. Percentage deleterious and beneficial mutations are indicated at the top and bottom of each bar, respectively. Panels B–G show different aspects of the mutation spectra of strains, with darker vs. lighter shades indicating mutational classes. (B) Tv bias (C) Indel bias (D) Noncoding mutation bias (E) Synonymous mutation bias (F) GC→AT bias (note that mutations that do not affect GC→AT bias, i.e., AT→TA and GC→CG mutations, are not shown here) and (G) Types of base-pair substitutions (BPS). Data underlying this figure are given in S6 Data.

https://doi.org/10.1371/journal.pbio.3003282.g004

An important consequence of increased genomic mutation rates is that mutators enjoy an increased total beneficial mutation supply (S_b), but must also contend with higher deleterious genetic load (L_d)—both important factors in determining their fates during adaptation [26–28]. We first calculated S_b and L_d for all strains assuming WT f_b and f_d (i.e., assuming similar DFEs across strains) [29]. As expected, these values scale linearly with the genomic mutation rate, with strains with 100× higher mutation rates predicted to have a 10- to 250-fold greater beneficial supply and lower deleterious load compared to WT (S12 Fig, S4 and S5 Tables). However, on accounting for the altered f_b and f_d values of mutator strains due to their DFE shifts, we observed deviations from this relationship. Ts biased strains typically had lower S_b while Tv biased strains had higher S_b than expected (S12A and S12B Fig, compare open circles versus filled circles; S4 Table). In the case of L_d, accounting for the observed DFEs caused relatively small changes for Ts biased strains, whereas Tv biased strains had a substantially lower L_d than expected based on mutation rate alone (S12C and S12D Fig, S5 Table). The effect of using the empirically observed DFEs was strongest for ΔmutT, where the beneficial supply relative to WT increased from ~250-fold to ~650- and ~400-fold in LB and glucose, respectively (S4 Table), and the deleterious load reduced from ~250-fold to ~53-fold in LB and ~58-fold in glucose (S5 Table). Thus, mutation bias shifts could have very large effects on the evolutionary fate of mutators.

DFE shifts and fraction of beneficial mutations are strongly associated with reversal of Tv bias, with some unexplained variation

The strains used in our analysis differed in several aspects of their mutation spectra (Table 1), so we examined variation in each aspect of the spectrum in more detail (Fig 4B–4G). The increase in f_b across strains was positively correlated with the magnitude of Tv bias (i.e., stronger bias reversal) (S11A and S11B Fig; also compare Fig 4A and 4B). In contrast, no other axis of variation in mutation spectrum was correlated with variation in f_b across strains (p > 0.05 in each case; compare Fig 4A with Fig 4C–4F). In each of these cases, either the range of variation in mutation bias across strains was very small (indel bias, synonymous bias, non-coding bias; Fig 4C–4E, Table 1) or there was no consistent pattern (noncoding bias and GC→AT bias; Fig 4D and 4F, respectively). Even when we considered each type of base substitution separately, the fitness effects and type of mutation were not associated (Fig 4G). For instance, ΔmutY and ΔmutT each sample distinct types of transversion mutations, yet both have the highest f_b values in both environments. Finally, pooling data across all strains in our dataset, transversion mutations were significantly more beneficial than transitions (Fig 5A and 5B), whereas no such fitness difference was observed for any other aspect of the mutation spectrum (except BPS versus indels in glucose; S13 Fig). Together, these results pointed to transversion bias as the dominant cause of DFE differences across strains.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Transversion mutations are more beneficial than transition mutations.

Fitness effects of mutations in (A) LB and (B) glucose; sample sizes are shown in x-axis labels in panel A. Note that y-axis ranges differ across panels. In each panel, the first two boxplots compare the effects of all transition vs. transversion mutations, pooled across all strains (asterisks indicate significant differences in Wilcoxon’s rank-sum tests; LB: p = 7.7E−10; Glucose: p = 5.1E−10). Following this, boxplots show the fitness effects of only Ts mutations from transition-biased strains and only Tv mutations from transversion-biased strains. Strains with the same letter have similar fitness effects based on pairwise Wilcoxon’s rank sum tests with Benjamini–Hochberg correction (i.e., all strains marked “a” have similar fitness, and all are significantly different from strains marked with “b” or “c”. Strains marked “bc” are similar to both “b” and “c”). Comparisons across all other types of mutations are shown in S13 Fig. Data underlying this figure are given in S7 Data.

https://doi.org/10.1371/journal.pbio.3003282.g005

Next, we tested whether the observed associations between transition/transversion bias and fitness effects of mutations are confounded by strain background (i.e., specifically which DNA repair genes were deleted and what other mutations occurred during the genetic manipulations (S1 Data). We first tested whether higher initial fitness of each strain was associated with lower f_b, reflecting the expected pattern of diminishing beneficial mutations with increasing background fitness [30,31]. Although the fitness of the original deletion strains varied significantly in both media, it was not correlated with f_b (S14 Fig). We then conducted pairwise comparisons between strains, considering only Ts and Tv mutations. We expected that Tv mutations should be generally more beneficial than Ts mutations regardless of strain background, and the same type of mutation (Ts or Tv) should have similar fitness effects in all strain backgrounds. These predictions are broadly borne out in both media (Fig 5A and 5B): Tv mutations in ΔmutY and ΔmutT had similar (and higher) s values than Ts mutations, though in LB the fitness effects of Tv in ΔmutY were similar to Ts in WT and in Δnth-nei.

Notably, in some cases Ts mutations in different Ts-biased strains had significantly different fitness effects (Fig 5A and 5B). In LB, Ts mutations in WT and Δnth-nei strains were more beneficial than other Ts-biased strains, and in glucose, Ts in Δnth-nei and ΔmutL were more beneficial. These exceptions could potentially be explained by aspects of the mutation spectra other than Ts/Tv bias, but these strains do not stand out as exceptional along other axes of the mutation spectrum (Fig 4B–4G, Table 1). Further, comparing Ts-biased strains in each growth medium, we found very few and inconsistent differences in fitness effects along other axes of mutation bias (e.g., coding versus non-coding, synonymous versus non-synonymous; S6 Table), indicating that they cannot explain the differences in fitness effects across Ts-biased strains. To summarize, the broad patterns of DFE variation that we observed are explained by Ts/Tv bias, but there is additional variation among Ts-biased strains that remains unexplained.

Reversal of the GC/AT bias does not alter DFEs

Prior simulations had predicted that the beneficial effects of sampling unexplored mutational space should extend to any axis of the mutation spectrum, including the GC→AT bias [11]. However, here we did not observe the predicted impacts of reversing the GC→AT bias. The WT has a slight GC→AT bias relative to the unbiased expectation of 0.5, and all three MMR strains as well as ΔmutT reverse this bias, whereas ΔmutY and Δnth-nei strongly reinforce the bias (Table 1, Fig 4F). However, as discussed above, the DFEs of the strains were not correlated with the magnitude of GC→AT bias, and overall, the fitness effects of GC→AT mutations versus AT→GC mutations were not distinguishable (Fig 6A and 6B). Since the GC→AT bias varies substantially across genes (S15A and S15B Fig), we hypothesized that local rather than global bias reversal may be more relevant. However, mutational effects were not correlated with gene GC content regardless of the direction of mutation bias (Fig 6C and 6D; also see S15C–S15F Fig). Thus, in contrast to the effects of reversing the Tv bias and contrary to our previous simulations, neither local nor global GC bias reversal altered the fitness effects of new mutations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Reversal of the GC/AT bias does not impact fitness effects of mutations.

Boxplots show fitness effects of AT→GC vs. GC→AT mutations in (A) LB and (B) Glucose. In each plot, data are pooled across all strains; sample sizes (total number of single mutations tested) are shown in the LB panel. Nine mutations that did not alter GC content (i.e., GC→CG or AT→TA mutations) are not shown here. Mutation type did not change fitness effects in either panel (Wilcoxon’s rank-sum test; LB: p = 0.4; Glucose: p = 0.24. (C–D) Scatter plots show the relationship between gene GC content and fitness effects of either AT→GC mutations (maroon points) and GC→AT mutations (black points) in (C) LB and (D) Glucose. Data underlying this figure are given in S8 Data.

https://doi.org/10.1371/journal.pbio.3003282.g006