Continuous convergent adaptation under LTSP

Our five LTSP populations, established in July 2015 maintain their viability to this day. Fig 2A depicts viability as a function of time, as estimated by plating samples and quantifying colony forming units (CFUs). We previously sequenced 413 clones from 9 time points spanning the first three years of our LTSP experiments. Here, we add to these data, resulting in whole genome sequencing data of 637 clones sampled at 12 time points, spanning the first 6 years of the experiment. A full list of the sequenced clones and their obtained sequencing coverage is given in S1 Table. It is important to note that each of the five populations was established from an individual clone and that the five ancestral clones used to establish the populations were also sequenced. Therefore, mutations identified in the evolved clones, but not in their ancestors had to have happened during the experiment. A full list of these mutations is provided in S2 Table. Focusing on 392 non-mutator clones, Fig 2B depicts the number of mutations accumulated as a function of time spent under LTSP. As can be seen, mutations continue to accumulate with time, throughout the six years examined so far.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Viability and mutation accumulation under LTSP.

(A) Viability during the first seven years under LTSP as estimated by quantification of colony forming unit (CFU) counts for each of the five populations. The 12 time points at which ~10 clones per population were fully sequenced are marked by red dots. (B) Continuous mutation accumulation—presented is the mean number of mutations acquired by non-mutator clones, as a function of time. Error bars represent standard deviations around the presented means. (C) Significant enrichment in non-synonymous mutations across examined time points–depicted are the fractions of non-synonymous mutations occurring in non-mutators across time points. The black line represents the expected fraction of non-synonymous mutations under neutral expectations (calculated based on the number of non-synonymous sites in the E. coli protein-coding genes (Materials and Methods)). (D) Convergent mutation acquisition. Depicted for each time point separately is the fraction of non-mutator mutations occurring within loci mutated across 1,2,3,4 or 5 of the populations.

https://doi.org/10.1371/journal.pgen.1011333.g002

Mutations accumulated under LTSP within non-mutator clones are significantly enriched for non-synonymous vs. synonymous mutations, relative neutral expectations (dN/dS = 3.5, P << 0.001, according to a χ² test). This indicates that mutation accumulation within non-mutators is governed by positive selection [26,34,35]. dN/dS continues to be significantly higher than 1 for mutations occurring across all examined time points (P << 0.001, Fig 2C and S3 Table). However, dN/dS for mutations occurring at or after the first year of the experiment is significantly lower than for mutations occurring at earlier time points (P << 0.001). An additional indication that many of the observed mutations are adaptive is the extremely high convergence with which mutations within non-mutator clones occur. ~88.5% of mutations acquired by non-mutator clones, fall within 265 loci that are mutated in at least two independently evolving populations (S1A Fig and S4 Table). These 265 loci constitute only ~5.7% of the E. coli genome. Mutations continue to occur more often than expected by chance in loci mutated across multiple populations, throughout the six years of the experiment. However, this trend as well seems to diminish with time (Fig 2D). Despite this apparent reduction with time, adaptation is still highly convergent even six years into our experiments—even at the last time point sampled, 84% of mutations occurring within non-mutators fall within loci that are mutated across at least two populations. Combined, these results show that our LTSP populations continuously adapt in an extremely convergent manner, up to at least six years under resource exhaustion, with a possible moderate reduction with time in the fraction of mutations subjected to positive selection.

Lineage structure and convergent lineage dynamics under LTSP

One of the most striking examples of convergence that we already reported on in our previous publications involves mutations in the RNA-polymerase core enzyme (RNAPC)-encoding genes rpoB and rpoC [27,28]. Strikingly, more than 90% of all sequenced clones carry such a mutation, with the same specific mutations often appearing in a convergent manner across populations, and with clones carrying different RNAPC adaptations often co-existing within single populations (Fig 3A). The RNAPC mutations are often the first mutations to occur, and thus, all of the descendants of the clones that initially acquired these specific rpoB and rpoC mutations also carry them, and most other mutations are acquired on the background of these RNAPC mutations. It can thus be said that the occurrence of the RNAPC mutations early on in LTSP establishes lineages that then co-exist alongside other lineages, established from clones carrying different RNAPC lineage-defining mutations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Convergent lineage structure under LTSP.

(A) Convergent pattern of RNAPC lineage-defining mutation acquisition across populations. (B) Muller plots depicting the relative abundance of the different lineages, as a function of time. Lineages are colored according to their lineage defining mutations, allowing for the use of the same colors in both panels A and B.

https://doi.org/10.1371/journal.pgen.1011333.g003

By using our WGS data to build phylogenies within each population, we characterized each population’s lineage structure. Obtained results were visualized as Muller plots depicting the relative frequencies of lineages as a function of time (Fig 3B and Materials and Methods). To build our phylogenies and characterize our lineages we focused on mutations that reached at least 30% frequency during at least one time point. For a lineage to be defined, it was required that the phylogeny of a potential lineage lasted within our population for at least two time points, and had been founded by a mutation that, at some time point reached at least 30% frequency within its population. We were able to assign ~95.3% of all sequenced clones to a lineage. A total of 18 lineages were identified across our five populations. While this was not an a-priori requirement, for all lineages the founding mutation was one of the RNAPC mutations summarized in Fig 3A. We observed continuous mutation accumulation occurring with time within lineages (S2 Fig). This means that clones from a specific lineage sampled at a given time point tended to accumulate additional mutations, compared to clones sampled from the same lineage at earlier time points.

Interestingly, lineages that share the same lineage-defining mutation, but belong to different populations often behave similarly. The most striking example of this involves lineages defined by the RpoC K334Q mutation, which emerge early on across all five populations, but then also reduce in frequency below our detection ability by day 127, across all populations (Fig 3B). In contrast, lineages defined by RpoC T428S also establish early in all five independent populations, but persist at high frequencies over the duration of the experiment so far, in all three populations that were not taken over by mutators.

Mutations within the same populations occurring within clones belonging to different lineages have necessarily occurred independently. We were therefore able to examine convergence not only between populations, but also within them. Our five populations contain 18 independently evolving lineages. ~89.3% of non-mutator mutations fall within 268 loci that are mutated in two or more of these lineages (S1B Fig and S5 Table). Other than the RNAPC genes rpoB and rpoC, the most convergently mutated locus is the gene encoding the global transcription regulator CRP, which is mutated in 11 of 18 independently evolving lineages. Mutations occurring within genes mutated in two or more lineages are significantly enriched for non-synonymous vs. synonymous mutations relative random expectations (dN/dS = 5.5, P << 0.001, according to a χ² test), while, in contrast mutations within loci that are mutated in only a single lineage are enriched for synonymous mutations (dN/dS = 0.4, P << 0.001, according to a χ² test). This suggests that loci mutated in two or more lineages are accumulating LTSP mutations as a result of positive selection [26,34,35]. Such convergently mutated loci are significantly enriched for genes involved in regulation of gene expression and transport (see Materials and Methods).

Frequent association between the occurrence of mutations across loci

Focusing on all 268 loci that are mutated in at least two independently evolving lineages or populations (excluding rpoB and rpoC), we extracted all possible loci pairs for which there existed clones that carried mutations in both loci in at least two independently evolving lineages or populations (see Materials and Methods). For each such pair of loci, we then asked in how many lineages did we see clones that carried mutations to both loci, clones that carried mutations to neither locus, and clones that carried mutations in only one or only the other of the two loci. Using these counts, we could then calculate the strength of association between the presence of mutations in one locus and the presence of mutations in the other across all lineages. We calculated three coefficients of association which range in their strictness from low to high: Yule’s association coefficient Q [36], Yule’s coefficient of colligation Y and Yule’s phi coefficient of contingency φ [37]. Depending on the coefficient used we identified between 360 and 664 instances of statistically significant association (after correcting for multiple testing, see Materials and methods and S6 Table). In total, we found that 91–94 out of the 268 (34.0–35.1%) convergently mutated loci are involved in at least one significantly associated loci pair. It can thus be said that for more than a third of convergently mutated loci the presence of mutations is significantly associated with the presence of mutations within at least one additional locus. To further support the biological relevance of the predicted contingent loci pairs, we examined whether they tended to be functionally related, more often than expected by chance, according to annotations contained within the STRING database [38]. Depending on the coefficient of association used, between 18.1% and 21.7% of putatively contingent loci pairs had known functional relationships. Based on 1000 repeated randomizations of our networks of putatively contingent loci pairs, we found that for all three coefficients, this proportion was significantly larger than expected by chance (see Materials and Methods and S7 Table).

The different coefficients used vary greatly in their stringency. As a result, many more putative contingent pairs were identified for the less strict coefficients. We aimed to set cutoffs for each of the coefficients that would allow us to select the most reliable putative contingencies, without discarding too many meaningful pairs. To do so, for each of the two less strict coefficients, we selected association strengths at which the highest percentage of predicted associated pairs carry a known functional association (S8 Table). For Yule’s association coefficient Q the cutoff was 0.8, while for Yule’s coefficient of colligation Y it was 0.5. Interestingly, this resulted in us obtaining the same set of 352 putatively contingent gene pairs for both the Q and Y coefficients. We did not set a threshold of significant contingencies for the most stringent phi coefficient of contingency φ, which identified 360 putatively contingent loci pairs, including 351 of the 352 pairs identified by the other two coefficients (above the corresponding threshold). We thus consider the 351 pairs agreed upon by all three metrics as our most reliable set of putatively contingent loci pairs.

When we consider only these 351 most reliable pairs, we still find that 91 out of our 268 (34.0%) convergently mutated loci are involved in at least one putative contingency. This demonstrates that our finding of more than a third of convergently mutated loci being involved in at least in one putative contingency is insensitive to the thresholds applied for our identification of putatively contingent loci pairs.

Directionality of contingency can be determined for a majority of the most reliable putatively contingent loci pairs

When mutations within one member of a putatively contingent loci pair always occur prior to mutations within the second member, it implies that mutations within the second locus are contingent on mutations within the first one. This logic enabled us to identify the likely directionality of contingency for 294 out of the 351 (83.8%) most reliable putatively contingent loci pairs (see Materials and Methods and Fig 4A and S9 Table). For an additional 30 pairs, we could not identify the directionality, because mutations within the two loci are always seen together (S9 Table). In the remaining 27 putatively contingent loci pairs, directionality could not be assigned because in some instances mutations were first observed in one member of the pair, while in others they were first observed in the second member (S9 Table). That for a majority of putative contingencies directionality is consistent and can thus be determined also adds credibility to our predictions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Identified putatively contingent loci pairs.

(A) Presented are the Yule’s Q coefficient strengths of association of the 351 most reliable putatively contingent loci pairs. These data are represented in a semi-directed adjacency matrix heatmap, in which only in cases where the directionality could not be determined the association is shown as undirected, while for the majority of cases the association is shown as directed. Both rows and columns represent the 91 loci involved in at least one out of the 351 putatively contingent loci pairs. Rows and columns are sorted in the same order. The rows (y axis) represent the loci that were found to be mutated first in each contingency pair, while columns (x axis) represent the loci that were mutated second. (B) For each locus that participates in eight or more putatively contingent loci pairs we present a summary of the directionality identified for the putative contingencies in which that locus participates.

https://doi.org/10.1371/journal.pgen.1011333.g004

LTSP contingencies with known effect

We identified two putative contingencies for which we already know the biological meaning from our previous studies [28,39]. The first contingency involves mutations within the gene dnaQ, which encodes a proofreading component of the DNA polymerase. Such mutations to dnaQ occur within two of our populations and occur only on the background of mutator clones. We have previously shown that the combination of a mutation to a mismatch repair gene, with a mutation to dnaQ significantly increased mutation rates beyond those of “normal” mutators [28]. The second example involves mutations within the genes atoC or atoS, that together encode a two-component system, which regulates fatty acid metabolism. Such mutations occur exclusively on the background of mutations to the gene encoding FadR—the E. coli master regulator of fatty acid metabolism. We have previously shown that this mutation combination allows E. coli to metabolize the short chain fatty acid butyrate, which E. coli itself produces during the first 24 hours of growth within fresh LB, but which wildtype E. coli cannot metabolize [39].

Mutations to many loci are contingent on previous mutations to the gene encoding the cAMP-activated global transcriptional regulator CRP

We identified some loci that serve as hubs of contingency, meaning that they are involved in many putative contingent loci pairs. When directionality was assigned, it became clear that mutations to the other loci tended to more often be contingent on mutations to these hubs, rather than vice versa (Fig 4B and S10 Table). The most striking example of this involves the gene crp, which encodes the cAMP-activated global transcriptional regulator CRP and is involved in a total of 52 of the most reliable contingent loci pairs. In 51 of these pairs, mutations to the other locus involved in the pair are putatively contingent on mutations to crp, while for the remaining pair the directionality could not be assigned.

We identified 11 putatively contingent loci pairs, involving a transcription factor (TF) and its known target. In 8 of these 11 pairs, mutations within the gene regulated by the TF were contingent on the occurrence of previous mutations within their regulating TF. For seven of these eight pairs the regulating TF was CRP. For six of these seven pairs in which mutations to genes regulated by CRP were contingent on previous mutations to crp, the mutations within the regulated loci were never seen in the absence of mutations to crp. In the remaining locus mutations are seen in the absence of mutations to crp in only one lineage, while they are seen together with mutations to crp in six lineages.

Loci that are regulated by CRP, and for which mutations are contingent on previous mutations to crp include: (1) the glpF gene and its promoter—both mutations to the glpF gene and to its promoter are strongly contingent on previous mutations to crp (Fig 5, Edges 1 and 2). glpF encodes a glycerol uptake gene, while CRP upregulates glpF [40]. Glycerol can be used as an energy-poor carbon source by E. coli and the downstream metabolites of glycerol can participate in glycolytic as well as gluconeogenic reactions [41]; (2) ompF–is upregulated by CRP (Fig 5, Edge 3) and encodes the outer membrane porin F, which is responsible for small molecule diffusion across the membrane [42]. (3) acs—encodes an acetyl-coenzyme A synthetase and is upregulated by CRP (Fig 5, Edge 4) [43]; (4) pck—encodes a phosphoenolpyruvate carboxykinase and is involved in gluconeogenesis. pck can be both upregulated or downregulated by CRP (Fig 5, Edge 5) [44]. (5) hfq—is downregulated by CRP (Fig 5, Edge 6) [45] and encodes a key chaperone of RNA folding that is involved in the regulation of interactions between small non-coding RNAs and their targets. As an important chaperone, hfq is involved in response to a large variety of conditions and stresses [46,47]; (6) cycA–encodes a D-serine/D-alanine/glycine transporter that is upregulated by CRP (Fig 5, Edge 7). (7) glcC–encodes a TF and is upregulated by CRP (Fig 5, Edge 8) [48]. GlcC itself regulates the glcA gene (Fig 5, Edge 9), which is responsible for the uptake of glycolate [49]. Glycolate can be used by E. coli as a carbon source and oxidized to carbon dioxide [50]. Mutations to glcA are contingent on previous mutations to glcC (Fig 5, Edge 9). This is therefore an example of a chain of contingencies involving regulated genes and their TFs–mutations to glcA, which is regulated by GlcC, are contingent on mutations to glcC, while mutations to glcC, which is regulated by CRP, are contingent on mutations to crp.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Schematic representation of putative historical contingencies involving regulatory interactions.

The historical contingencies described in our text are presented graphically as a network. The edges are numbered according to the order of appearance in the main text.

https://doi.org/10.1371/journal.pgen.1011333.g005

Additional putatively contingent loci pairs involving loci with known regulatory interactions

There are several examples of putatively contingent loci pairs that do not involve mutations to crp, but nevertheless have known regulatory interactions. Such putatively contingent loci pairs that involve a TF and its target include: (1) the gcvT-gcvA loci pair—mutations to the gene encoding the GcvA TF are contingent on previous mutations occurring within its regulated gene gcvT (Fig 5, Edge 10), which encodes a protein that functions in the glycine cleavage system [51]. For both gcvA and gcvT mutations occur within their promoter regions, rather than within the genes themselves. (2) putA and putP–PutA is a TF that downregulates putP (Fig 5, Edge 11), which encodes a sodium/proline symporter [52]. No directionality could be assigned to the putative contingency involving these two genes. Mutations were observed within both the putP promoter region and the gene itself, in both cases with strong association to mutations in the putA gene (Fig 5, Edges 12 and 11).

In addition to putative contingencies involving genes and their TFs, we also observed four instances in which mutations to a certain gene were associated with mutations to the same gene’s promoter region: (1) dctA and its promoter—mutations within the dctA gene, which encodes a C4-dicarboxylate transport protein, and its promoter are only ever seen together (Fig 5, Edge 13). (2) SstT is a serine/threonine transporter. Mutations in the promoter of sstT are contingent on previous mutations to the sstT gene itself (Fig 5, Edge 14). (3) Mutations to the putP promoter region are contingent on previous mutations occurring within the putP gene itself (Fig 5, Edge 15). As previously described, putP encodes a sodium/proline symporter. (4) Mutations to the cycA gene, which, as previously described, encodes a D-serine/D-alanine/glycine transporter, only occur on the background of mutations to its promoter (Fig 5, Edge 16). Interestingly, all four examples of putative contingent loci pairs involving a gene and its promoter involve transporter genes (although, as already stated, there is a general enrichment in genes related to transport among the loci that are convergently mutated in our experiments).

Finally, a putatively contingent pair in which both genes co-regulate the same operon was observed. DeoR and CytR co-regulate the deo operon, which is involved in nucleoside and deoxy-ribonucleoside catabolism [53]. Interestingly deoR and cytR are almost always mutated together (Fig 5, Edge 17). Clones that carry mutations to both deoR and cytR arise early, across all five populations and always belong to the minority of clones that cannot be assigned to any of the lineages, which are all defined by RpoB and RpoC mutations. Clones carrying different mutation combinations to deoR and cytR can often be seen within a single population (S2 Table), suggesting that this mutation combination occurs very frequently. However, they can never establish high frequency lineages, that persist across multiple time points. Combined, this suggests that clones carrying this putatively contingent mutation combination may have a very strong fitness advantage early on under LTSP, but are not able to persist at frequencies above our detection level.

Mutations within some loci are contingent on specific RNAPC mutations

Once we determined the lineage structure of our populations and assigned lineages according to the RNAPC mutation that established them, it became evident that mutations within certain loci occurred predominantly within lineages established by a specific RNAPC mutation. To identify instances of significant association between the occurrence of mutations within certain loci and the occurrence of specific RNAPC mutations, we calculated the association coefficients, as previously done for convergently mutated loci pairs (Materials and Methods). A summary of the obtained association strengths is given in S11 Table.

The most striking example of such an association is that mutator clones always belong to the lineages defined by the RpoB E1272G mutation. The RpoB E1272G mutation established lineages in three of our five populations, and across all these populations, all clones sequenced belonging to these lineages also had a mutator mutation, within a mismatch repair gene (S2 Table). In two of the populations (populations #2 and #3) mutators appear to have emerged only once, as all mutator clones within these populations carried the exact same mismatch repair gene mutation. In the third population (population #4) two different types of mutators evolved. However, both evolved on the background of population 4’s RpoB E1272G lineage. No other lineages ever included any such mutator clones. When considering all mismatch repair genes in which mutations occurred as a single entity, the association between the occurrence of mismatch repair mutations and the RpoB E1272G mutation is statistically significant (S11 Table). Interestingly, the pattern by which mutators occur only on the background of lineages established by the RpoB E1272G mutation was also recapitulated in a second study, in which six additional LTSP populations were established [39]. An additional example of a locus that is mutated exclusively within lineages defined by a specific RNAPC mutation is the emergence of mutations in both the dctA dicarboxylate transport gene and its promoter on the background of three out of five lineages defined by the RpoC K334Q mutation, and never on the background of any other lineage-defining mutations (S11 Table). This is especially impressive considering the fact that the RpoC K334Q lineages are always very short lived.

LTSP evolutionary experiments

The five LTSP populations investigated here were established in July 2015, as described in Avrani et al. [27]. Briefly, each population was initiated from a separate clone of E. coli K12 strain MG1655 (following streaking of bacteria on LB plates and random selection of five single colonies). Approximately five million cells per ml were inoculated into 400 ml of Lysogeny broth (LB) within 2 liter polycarbonate breathing Erlenmeyer flasks. The five flasks were placed in an incubator set at 37 °C, set to shaking at 225 rpm. No external nutrients or resources were added to the cultures with time, except for sterile water that is added to compensate for evaporation every 10–15 days, according to the weight lost during that time period.

Sequencing of LTSP clones and mutation calling

One milliliter of each of the five LTSP populations, established in July 2015, was sampled at days 0,1, 2, 3, 4, 5, 7, 8, 11, 22, 32, 42, 53, 64, 94, 127, 154, 184, 241, 308, 374, 436, 557, 730, 1095, 1178, 1460, 1612, 1816 and 2191. Sample dilutions were used to estimate viability by quantification of colony-forming units (CFUs) with the help of robotic plate reader. Samples were preserved in 50% glycerol at -80°C for further analysis. Dilutions of samples from days 11, 22, 32, 42, 64, 127, 374, 730, 1095, 1460, 1816 and 2191 were plated and cultured overnight. ~10 colonies from each time point were inoculated in 4 ml of media and were subsequently cultured until the optical density of 1 was reached. A similar procedure was followed to sequence the five ancestral clones used to establish the populations. DNA was extracted with the Macherey-Nagel NucleoSpin 96 Tissue Kit from pelleted cells, resulting from the centrifugation of one milliliter of each culture at 10 000 g for 5 minutes. The remainder of each culture was frozen in 50% glycerol at -80°C. Libraries were prepared according to the protocol described in [59]. 150 bp paired end sequencing was carried out on an Illumina HiSeq 2500 machine at Admera Health. Clones from all five populations from days 1460, 1816 and 2191 as well as clones from time points 22, 32 and 42 from populations 3 and 5 were sequenced during the current study, while clones from all other time points and the ancestral clones were sequenced as part of our previous studies (S1 Table) [27,28]. Obtained reads were aligned to the E. coli K12 MG1655 reference genome (accession NC_000913) and mutations were called, using Breseq [60], which enabled us to identify SNPs as well as small and large indels (S2 Table).

Raw short read sequencing data obtained for this study were deposited to the Sequence Read Archive (SRA) under accession number PRJNA1049885. Short read data obtained as part of our previous studies can be found under accessions BioProject ID: PRJNA380864 and BioProject ID: PRJNA674613.

dN/dS calculations

Numbers of synonymous and non-synonymous sites of all protein-coding sequences of E. coli were calculated as in [27]. Briefly, The protein-coding gene sequences of E. coli K12 MG1655 were obtained from the NCBI database. For each DNA position of each protein coding gene, we calculated the likelihood that a mutation within that site would be synonymous or non-synonymous. These likelihoods were then summed up across all protein coding sequences. dN/dS ratios were then calculated for non-mutator clones as .

Building phylogenies within each LTSP population

To characterize the lineage structure of each LTSP population, we created a database of the mutations found within each of its sequenced clones. Only mutations that rose to a frequency of 30% and higher at any time point throughout the experiment were included in the databases. Using these data, the genetic distances between clones were estimated using the Levenshtein distance metric, while perceiving mutations as single characters and lists of mutations as strings (the Levenshtein distance is a measure of the difference between two strings that is estimated as the minimal number of character replacements needed in order to convert one string to the another [61]). The obtained distances were utilized to determine the "parent" for each clone and the resulting ancestor-descendant pairs were used to build phylogenies for each of the five populations and create Muller plots (Fig 3B) with the help of the R package MullerPlot [62].

Gene ontology enrichment analysis

To test whether loci mutated across two or more lineages are enriched for particular biological processes, we used the Gene Ontology’s tool “GO Enrichment Analysis” [63–65]. The list of the genes used in the analyses can be found in S5 Table. Statistically significant enrichment was deduced when the FDR corrected P value was smaller than 0.05.

Screening for putatively contingent loci pairs

We first focused on 268 loci that were mutated in at least two independent lineages or populations, excluding rpoB and rpoC. From this list of loci we generated a list of pairs of loci that were mutated together within a non-mutator clone in at least two lineages or populations. For each such pair of loci we then generated contingency tables by counting in how many lineages did we see clones that carried mutations to both loci, clones that carried mutations to neither locus, and clones that carried mutations in only one or only the other of the two loci. It should be noted that the same lineage can be counted towards different categories simultaneously, meaning that for example we can see within a specific lineage clones in which both loci are mutated, and others in which only one of the two is mutated. In such a case, for this pair of loci, that lineage will be counted under both categories. In order to account for contingencies that occur within the rare clones that could not be assigned to any lineage, we assigned such clones to "pseudo-lineages" (one for each population in which such clones existed). In total, after removing the three mutator lineages, there remained a total of 20 lineages and pseudo lineages.

The strength of association between mutations within specific loci and specific RNAPC mutations was estimated similarly to what is described above. The only difference was that instead of creating contingency tables for the occurrence of mutations within two convergently mutated loci, we created the tables for the occurrence of mutations within each specific locus and the occurrence of a specific RNAPC mutation. We initially excluded the RpoB E1272G mutation from consideration, as all clones carrying this mutation are mutators. In order to estimate the strength of apparent association between the RpoB E1272G lineage defining mutation and the mutator trait, which, in our experiment was driven by mutations to three different mismatch repair genes (mutS, mutL and mutH), we generated a contingency table in which mutations in all three DNA-mismatch genes were counted as one entity.

Using the obtained contingency tables, we calculated three metrices of association: Yule’s association coefficient Q [36], Yule’s coefficient of colligation Y and Yule’s phi coefficient of contingency φ [37]. To ascertain statistical significance, we performed a one-sided t-test with α = 0.05 followed up with FDR multiple testing correction (S6 Table).

Functional association analyses

We used the STRING database [38] to quantify the fraction of putatively contingent loci pairs that are already known to be functionally associated. To examine whether this number was higher than expected by chance, we performed 1000 randomizations where we selected at random loci pairs from the list of 268 convergently mutated loci. For each randomization we could then quantify what fraction of randomly selected loci pairs were functionally associated according to the STRING database. These numbers were then used to generate a random distribution and calculate a P-value for the observed numbers of putatively contingent loci pairs that were also functionally associated. 1000 randomizations were carried out separately for each of the three metrics of association used, and in each instance the number of loci pairs selected per randomization matched the number of putatively contingent loci pairs identified using that metric.

Estimation of contingency directionality

In order to estimate the directionality of the contingency for each of the most reliable 351 putatively contingent loci pairs, we considered only the lineages in which mutations to both members of the pair were found within single clones. For such clones the phylogenies were built by tracking the chain of ancestor-descendant pairs of clones with the smallest Levenshtein distance, while representing mutations as single characters and all mutations in the clones as strings. The obtained data was used to examine whether the predicted contingent mutations always occurred together in the same clones of the lineage or whether one of the mutations could occur separately from another. In cases where the mutations did not always occur together, the directionality was determined by identifying which of the pair of mutations was found earlier in an ancestor of the clone that eventually carried both putatively contingent mutations. The resulting data was then combined across lineages to determine the ultimate trend (e.g. that across lineages one of the mutations was always seen prior to the other (S9 Table).