Rate and molecular spectrum of spontaneous mutations in the bacterium Escherichia coli as determined by whole-genome sequencing

The major conclusions of the study presented here are (i) the best estimate of the spontaneous mutation rate of a wild-type E. coli strain growing on rich medium is 1 × 10⁻³ per genome per generation; (ii) in the absence of MMR, mutations are dominated by A:T > G:C transitions; (iii) A:T > G:C transitions occur preferentially with A templating the lagging strand and T templating the leading strand, whereas G:C > A:T transitions occur preferentially with C templating the lagging strand and G templating the leading strand; and (iv) there is a strong bias for transitions to occur at 5′ApC3′/3′TpG5′ sites and, to a lesser degree, at 5′GpC3′/3′CpG5′ sites.

In addition to these major points, our results identify GATC sites as hotspots for A:T transversion mutations. We extend to the whole genome the previous findings that CCAGG and CCTGG sites are hotspots for G:C transitions (41) and that indels occur preferentially at sequence repeats (40). However, although the rate of indels at repeats is high and increases exponentially with the length of a sequence run, the overall genomic rate of indels is only 1/10th that of BPSs.

Our finding that the mutation rate of this model prokaryote is one third of Drake’s estimate based on analyses of specific-locus experiments (1, 21) was unexpected. As pointed out by Drake (4), different estimates obtained from MA experiments and specific-locus methods may reflect selection against the mutations that accumulate during the former protocol, which, by the use of a carefully chosen reporter gene, are minimized in the latter. However, by several criteria the MA protocol followed here minimized selection: (i) the ratio of synonymous to nonsynonymous BPSs in the wild-type strain did not differ significantly from that expected by chance; (ii) the wild-type mutation rate based solely on synonymous BPSs differs little from that calculated based on all BPSs (see below); (iii) most synonymous BPSs resulted in less commonly used codons, indicating that these mutations were not selected against; and (iv) nonsense mutations were recovered at the expected frequency, indicating that these mutations also were not selected against.

Alternative explanations for the discrepancy are that the specific loci used by Drake (1, 21) to estimate the genomic mutation rate were not representative of the genome as a whole or that some of the approximations used were not appropriate for E. coli. In addition, the growth conditions used in the two types of experiments are very different. Most obviously, during MA experiments the cells grow in colonies, which can become microaerobic, whereas for specific-locus experiments cells typically are grown in well-aerated liquid cultures. Given these differences, perhaps it is remarkable that the two estimates are within threefold.

Recently Wielgoss et al. (3) estimated the E. coli genomic mutation rate based on synonymous BPSs occurring during long-term evolution experiments. Their value, 0.41 × 10⁻³ per genome per generation, is less than half the rate that we observed. The mutation rate based solely on the 55 synonymous BPSs in our wild-type dataset is 1.1 x10⁻³ per genome per generation. This value is close to the rate we obtained considering all the BPSs, and is nearly threefold higher than the value calculated by Wielgoss et al. (3). Drake (4) has argued that selection against synonymous codon changes in the long-term evolution experiments was sufficient to account for the low rate obtained by Wielgoss et al. We believe that the higher mutation rate that we obtained reflects the fact that the MA protocol truly minimizes selective pressure against mutations. The lowest mutation rate estimate for E. coli, 0.1–0.2 × 10⁻³ per genome per generation, comes from comparative genomics (42). Although the various contributions of selection, mutational bias, and experimental conditions appear sufficient to explain differences in experimentally determined mutation rates, there is as yet no consistent explanation for the difference between estimates based on comparative genomics and those derived from experimental evidence (2).

Although whole-genome sequencing can give an unparalleled view of mutational events across the entire genome, it cannot supersede the results obtained with more traditional methods. For example, except for highly frequent events, such as indels at repeat sequences, sequencing unselected genomes reveals only single mutational events. Cataloging events at similar sites in the genome allows aspects of mutagenic specificity to be deduced (see below) but does not identify true hotspots, i.e., specific sites that are particularly mutation-prone. Thus, the mutational topography revealed by single-locus studies is a necessary component of our understanding of mutational processes.

A striking result from our MA experiments is that, in the absence of MMR, the BPS spectrum of cells growing on rich medium is dominated by A:T > G:C transitions (Table 3 and Fig. 2). Because of individual idiosyncrasies of each mutational target, this bias had not been seen in many previous studies of MMR-defective microbes. Indeed, the observed bias in MMR-defective E. coli often was toward G:C > A:T mutations (43–45). Only when collections of mutations were sequenced, as in the mnt gene (46), the lacI gene (47), the rpoB gene (22), and the gyrA gene (23), was an A:T > G:C bias in MMR-defective strains revealed. Our data confirm and extend these results to the entire genome.

Mutational data from a variety of sources have confirmed that in almost all organisms spontaneous mutations are dominated by G:C > A:T transitions, a bias that tends to drive genomes toward greater A:T content (28). However, the G:C content of genomes varies widely, so some selective pressure or nonadaptive mechanism must drive genomes back toward G:C-richness. Because MMR corrects replication errors, the mutational spectrum in its absence reveals that errors made during replication would, if uncorrected, increase the genomic G:C content. Thus, the MMR system serves not only to keep mutation rates low but also to prevent genomes from drifting toward ever-higher G:C content. This observation suggests that the G:C/A:T balance of the genomes of different organisms might be determined, at least in part, by the functioning of their MMR system. In a recent survey of 699 bacterial genomes, about 20% lacked one or more genes in the MMR pathway, but there was no correlation with the G:C content of the genomes (48). Two caveats apply to these results: (i) other repair activities in the cell may compensate for the absence of MMR; and (ii), the presence of the MMR genes does not mean that the system is functioning to the same extent in each organism. Our results also have a forensic implication. Because the activity of the MMR varies by growth phase and is influenced by environmental conditions (19, 20), an organism’s mutational spectrum might be a fingerprint of its recent history.

Loss of MMR caused a larger increase in mutations in coding than in noncoding DNA (Table 3), suggesting that the MMR system is more active on the DNA that encodes genes. This difference could reflect the participation of the MMR proteins, including MutL, in transcription-coupled repair (49). However, contrary to a recent report (39), we found no bias for any specific mutational event to occur on the transcribed versus the nontranscribed DNA strand (Table S3); such a bias could reflect the activity of transcription-coupled repair. Nor did our data support previous conclusions that highly expressed genes are either vulnerable or resistant to mutation (37–39).

Our results shed light on the origins of spontaneous mutations. Deamination of cytosine to uracil frequently is cited as a cause of G:C > A:T transitions. The mutation rate of G:C > A:T transitions in the MutL⁻ strain was 0.035 per genome per generation, which is 60 times the estimated deamination rate of cytosine in dsDNA and 10–20 times the estimated deamination rate of cytosine in ssDNA (Table S5). Adding to the low level of deamination are the activities of uracil glycosylase, which removes uracil from DNA, and mismatch uracil DNA glycosylase, which removes uracils paired with guanines. Thus, unless the rate of cytosine deamination is greatly underestimated, it cannot be responsible for the number of G:C > A:T mutations observed. However, the creation of thymines by deamination of 5meC is important. In the wild-type strain, 10% of the G:C > A:T transitions occurred at the C that is methylated by the Dcm methylase. Thus, our results support and extend to the whole genome the prescient finding that Dcm sites are mutational hotspots (41).

Mutations at Dcm sites can be prevented by very short patch (VSP) repair, an enzyme system that removes the T mispaired with C at these sites (reviewed in ref. 50). MutL participates in VSP repair, so it was surprising that the rate of G:C > A:T mutations at Dcm sites was increased only threefold in the MutL⁻ strain. However, MutL facilitates but is not required for the relatively inefficient VSP repair (51, 52). Because G:C >A:T mutations were increased 100-fold overall in the MutL⁻ strain, other sources of G:C > A:T mutations must occur at much higher frequencies than deamination of 5meC. Because MMR normally prevents these G:C > A:T transitions, mutations at Dcm sites become prevalent in wild-type cells.

Transitions arise from A:C and G:T mismatches, both of which are well corrected by the E. coli MMR system in vitro (53) and in vivo (54, 55). If both these mismatches occurred with equal frequency and irrespective of DNA strand, then loss of MMR would elevate both A:T > G:C and G:C > A:T transitions to the same extent. Therefore, the observed bias for A:T transitions in the MutL⁻ strain implies either that the mismatches that lead to A:T transitions are replication errors that occur more frequently than those that lead to G:C transitions or that a separate error-correcting pathway efficiently and preferentially prevents G:C transitions. One candidate for this error-correcting function is the proofreading activity of the E. coli replicase DNA polymerase III, which resides in a separate subunit, epsilon. Whether error correction by epsilon is biased has been debated for years, but a recent study (56) showed a bias toward correcting G:C > A:T mutations that would account for our results. Another candidate is MutY, which can remove an A mispaired with C, preventing G:C > A:T mutations if C is the template (and promoting A:T > G:C mutations if A is the template) (57). Future MA studies will reveal if these enzymes contribute to mutational specificity in the absence of MMR.

Transitions arise when a purine mispairs with the wrong pyrimidine during replication. Long ago, such mispairings were postulated to result from tautomeric shifts in the bases (58, 59). Others have proposed mispairs involving ionized bases (60). Crucial to these schemes is that the postulated mispair match the geometry of the normal base pair, allowing it to fit within the DNA polymerase-active site for catalysis. Recently both A:C and the G:T mispairs have been captured in the active site of a crystallized polymerase (61, 62); the former involved amino:imino tautomers, and the latter probably involved ionized bases. In both cases the mismatched base pair had the canonical Watson–Crick geometry of the cognate base pair. One of the striking aspects of the MutL⁻ data is that 79% of the A:T transitions occurred at sites with the sequence 5′ApC3′/3′TpG5′ (Table S2), a preference that has been observed in previous studies (22, 63). We favor the hypothesis that these transitions are templated by the A for the following reason: because polymerase approaches the template base from the 3′ side, a C:G base pair would be established first and, because of its strong base-pairing and stacking interactions, could stabilize the A:C mispair. In the alternative configuration, requiring a T:G mispair, the stabilizing G:C base pair would not be established until after the T:G mispair formed. The strong preference for C, not G, 3′ to the A, further suggests that the stabilizing base pair must have the correct orientation.

Another striking aspect of the MutL⁻ data is the 2:1 bias for A:T transitions occurring with A in the top strand on the right replichore and T in the top strand in the left replichore. The purine:pyrimidine strand-bias for G:C transitions was the reverse: they were twice as likely to occur with C in the top strand on the right replichore and G in the top strand in the left replichore. This mutational pattern would tend to produce the well-known “keto-skew” of the E. coli chromosome (64); i.e., in the first half of the chromosome the top strand has more Gs and Ts, the bases with keto groups, and in the second half of the chromosome, the top strand has more As and Cs, the bases with amino groups. A similar mutational bias and keto-skew near replication origins has been described recently in S. cerevisiae (65). However, the bias in itself does not tell us which base gives rise to the mutation.

If A is the templating base for A:T transitions, as hypothesized above, the A/T strand bias would require that the mutational event be A templating the insertion of a C during lagging-strand synthesis. If the G/C strand bias reflects the same molecular constraints, then the G:C transitions would be produced by C templating insertion of an A during lagging-strand synthesis. Arguing against this hypothesis is the fact that G:C transitions did not show a strong bias for a particular base 3′ to the C. However, 74% of the G:C transitions occurred at sites with either C or G 3′ to the C (Table S2), suggesting that the C:A mispair could be stabilized by the formation of a 3′G:C base pair regardless of its strand orientation.

Although we favor the hypothesis that A and C are the templating bases and that the A:C mispair forms during lagging-strand synthesis, the alternative is that T and G template the G:T mispair during leading-strand synthesis. Then the mismatch would be stabilized by the G:C base pair formed after the mismatch. In support of this hypothesis, extension past a mismatch by only 4 nucleotides protects the mismatch from exonucleolytic proofreading (66). The hypothesis that the mismatch occurs during leading-strand synthesis also would agree with the often-stated conclusion that leading-strand synthesis is less accurate than lagging-strand DNA synthesis (67). However, although the fidelity of synthesis of the two strands may not be equal, the conclusion that synthesis of the leading strand is the less accurate one is based on primer-extension reactions in vitro and may not pertain in vivo.

Transversions made up 44% of the mutations in the wild-type strain. Because loss of MMR increased transversions only about 10-fold, compared with 200-fold for transitions, transversions were only a minor component of the MutL⁻ spectrum. This result means that MMR is relatively inefficient in correcting the mismatches that lead to transversions and/or that other repair pathways are active in preventing transversions. For example, mutations that inactivate the enzymes that deal with oxidized guanines are powerful mutators, greatly increasing both A:T > C:G and G:C > T:A mutations (68). These two transversions accounted for 29% of the BPSs recovered in the wild-type strain, but their rates were increased only sixfold in the MutL⁻ strain, suggesting that MMR is relatively blind to the mismatches that cause these mutations. As mentioned above, A:T transversions occurred preferentially at GATC sites where the A is methylated at the 6 position by the Dam methylase. GATC sites also were hotspots in the mutational spectrum of the rpoB gene (69). Depurinated bases lead to transversion mutations because for many DNA polymerases the preferred order of base insertion opposite a missing base is A followed by G, although this order is far from strict (33, 70). The estimated rate of depurination of 6meA is about 3 × 10⁻³ per E. coli genome per generation (Table S5), which is sufficient to account for the observed rate of A:T transversions at GATC sites in both the wild-type and the MutL⁻ strains (0.09 × 10⁻³ and 2.4 × 10⁻³ per genome per generation, respectively).

Our data show that the rate of indel formation at mononucleotide runs increases exponentially with the length of the run and that the ability of MMR to correct these indels may increase with run length (Fig. 3A). Although we identified indel hotspots based on repeated occurrence, any run of 4 nucleotides or more is a potential hotspot (Fig. 3B). Because an indel is likely to inactivate a gene, there should be selection against runs in genes encoding important cell functions. On the other hand, genes that have significant runs could be dispensable; indeed, such genes may be “contingency loci,” able to be activated and inactivated by frequently occurring mutations (71). We inspected the list of genes in which indels were recovered to find those whose activation or inactivation might have a selective advantage under stressful conditions. Interestingly, we recovered two indels in dinB and one in umuC, the genes that encode E. coli’s two error-prone polymerases. The two runs in the dinB gene, of four and six Gs respectively, occur early in the coding sequence and thus indels in these runs certainly inactivate the gene. These runs are conserved in most E. coli and Shigella genomes in the database but are not conserved in Salmonella. The mutated run in the umuC gene, consisting of five As late in the coding sequence, likewise is conserved in most E. coli and Shigella genomes but not in Salmonella. Surprisingly, an indel was isolated midway in the mutT gene in a run of six Cs that is not well conserved even among E. coli strains. The phenotype of mutT mutants is an increase in G:C > TA mutations (68); because the MA strain carrying this mutation did not accumulate any G:C > TA mutations, the indel must have occurred late in the MA experiment. Other mutated genes of interest are xthA, which encodes an exonuclease involved in base-excision repair, and alkB, which encodes an enzyme involved in the repair of alkylation DNA damage. A recent survey of bacterial genomes found that hundreds of species had repeat sequences in MMR genes in which indels could act as genetic switches to change the mutation rates of these microbes (72). Thus, in addition to mechanisms that modulate the activity of DNA repair enzymes, microbes also may have adaptive or serendipitous mechanisms that genetically alter their DNA repair capabilities, and thus their mutation rates, allowing them to respond to changes in environmental conditions.

The authors declare no conflict of interest.

We thank D. Osiecki, D. Simon, K. Storvik, J. Townes, N. Yahaya, and A. Ying Yi Tang for valuable technical assistance; members of the Lynch laboratory for useful discussions; and E. Eisenstadt, S. Finkel, A. Hanson, M. Lynch, and the anonymous reviewers for helpful comments on this manuscript. This research was supported by Multidisciplinary University Research Initiative Award W911NF-09-1-0444 from the US Army Research Office (to P.L.F.).