The first ~10 codons are unusual in both nucleotide content and in their influence on protein levels

Before asking about possible alternative explanations, we start by asking whether there is evidence that the 5′ zone that may be considered distinct may have been misestimated. We consider trends in nucleotide usage, in RNA stability, in influence on protein expression, and on cellular growth of 5′ CDS codons.

The first 10 codons of native genes show distinct nucleotide content trends compared to downstream codons.

For completeness and the convenience of the reader, we start by illustrating AT content variation across the 5′ regions of 5,098 native E. coli genes, similar to that done previously (e.g., [39]). At all codon positions there is low GC content in the first few codons, which gradually increases up to codon 10 (Fig 1A). Beyond this region, AT content stabilizes and remains relatively uniform across nucleotide positions (Fig 1A). We note that this pattern holds true across all three codon sites, although in the 5′ domain the second site is the most AT-rich, an observation we return to later (see Discussion and S1 Text) as this is indicative of protein-level effects [76] (see S1 Fig for content by codon position relative to the start codon, by position within codons, i.e., sites 1, 2, and 3, for all four nucleotides individually).

Download:

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

Fig 1. Examination of the uniqueness of the first ~10 codons in Escherichia coli CDSs.

A. AT content by nucleotide position, divided into the three codon sites. B. Content for each of the four nucleotides by position at third sites of 4-fold degenerate codons. In both A and B, content is averaged at each nucleotide position across 5,098 native genes. The x axis represents nucleotide positions relative to the start codon (i.e., the first nucleotide of the codon after the start codon is at position 1). Error bars indicate the standard error of the mean (SEM). Dashed vertical black line marks the first 10 codons. Locally estimated scatterplot smoothing (LOESS) regression lines are included. C. Average mRNA stability by codon position. Stability was determined by two measures. First, in blue, it was computationally predicted using ViennaRNA R package [55] through a sliding window approach scanning the CDS of native genes (starting at nucleotide position −15). The x axis represents the position at the middle of each 30 bp window. Second, experimental PARS data was employed (in pink), data from Del Campo and colleagues [49]. Dashed vertical black line marks the first 10 codons. A polynomial regression of order 10 is shown for both. For ease of comparison, in both cases Z normalized data is employed with a low (negative) figure indicating lower stability. D. Relative influence of GC content at each codon position on protein per RNA levels (protein/RNA) in transgene constructs. Relaimpo analysis was conducted using the lmg model with 1,000 bootstraps, 90% bootstrap confidence intervals are shown as error bars. Codon positions on the x axis refer to absolute numbers (where start codon is 1). Transgene data from Cambray and colleagues [41]. E. Full Spearman correlations between GC content and protein per RNA (protein/RNA) levels in transgenes for each codon position. F. Partial Spearman correlations controlling for the influence of all other codon positions in the available construct sequence (codons 2–30) such that, for instance, the relation between transgene expression (as protein/RNA) and GC content at codon 2 is found independently of the relation with GC content in codons 3–30. In E, F, colored points represent rho values with a P-value ≤ 0.05, while gray points are non-significant. Locally estimated scatterplot smoothing (LOESS) regression lines are included. Codon positions on the x axis refer to absolute codon numbers (e.g., the start codon is codon 1). Dashed vertical black line marks the first 10 codons. Transgene data from Cambray and colleagues [41]. Plots D–F were repeated using Cambray protein levels only, not normalized by RNA (see S2 Fig). The data underlying this Figure can be found in htps://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g001

We also illustrate the nucleotide trends only at synonymous sites in the strictest sense by finding average content at third sites of 4-fold degenerate codons by position (Fig 1B). There is a difference in 5′ nucleotide content when compared to core in all four nucleotides, with 5′ ends characterized by high A/T and low G/C, all plateauing around codon 10. Codon usage at codon positions 2–10 is significantly different to downstream, with an over-representation of A-ending codons independent of the amino acid used [77]. As reported (e.g., [39]), we also find that the bigger differences are in A and G content, compared to T and C (Fig 1B).

Extended synonymous site conservation is not explained by gene–gene overlaps

Prior to consideration of the effects of gene overlap, we first seek to describe K_s trends using close comparators across all genes (not available to prior analyses [39]). We performed ortholog alignments between three closely related species that provide a “sweet spot” of K_s estimation (one where the amount of change is low enough that neither alignment nor saturation are an issue, but where there is enough change to be informative), these being E. coli, Escherichia fergusonii (E. fergusonii), and Salmonella enterica (S. enterica). We reconstructed via maximum likelihood the ancestral sequence of the E. coli and E. fergusonii common ancestor with S. enterica as outgroup species. To determine conservation by codon position (relative to the start codon), we separately extract each codon for each ortholog and concatenate it to the codons at that same position for all other orthologs. We do this for the first 150 codons following the start codon to give 150 codon-specific alignments, each containing sequences of at least 1,400 codons (one per ortholog that is at least 180 codons in length so as to avoid capturing 3′ effects). We exclude (rare) cases where all three species have different codons, as it is likely for there to be uncertainty in the ancestral sequence reconstruction at those positions. For each codon position, we then find the synonymous substitution rate (K_s) between the focal species and the ingroup common ancestor from the by-codon position alignment. We can also compute non-synonymous substitution rates (K_a), and the ratio between the two (K_a/K_s), the latter permitting us to address the question of the extent to which the low K_s might distort measures of protein-level selection.

We find synonymous substitution rates (K_s) to be lowest at the most 5′ codons and increase until approximately codon 60, after which they asymptote (Fig 2A). This is slightly more downstream than the plateauing around codon 30 that was previously reported from highly diverged E. coli–Salmonella sequence [39]. K_a rates do not show as clear of a trend (Fig 2B), and K_a/K_s ratios appear to be primarily driven by K_s rather than K_a levels (Fig 2C), which if anything are actually higher in the first few codons than the ones slightly 3′ (Fig 2B). While the most 5′ codons do have the lowest K_s values, there is no evident change in regression slope around or immediately after codon 10 (Fig 2A). Similar substitution rate patterns were observed when comparing the ancestor to E. fergusonii (S4 Fig). Trends for 4-fold degenerate codons (S5 Fig) are similar to those resulting from consideration of all codons (Fig 2). We also see high similarity in trends between operonic and non-operonic genes (S6 Fig). For the set of genes that have an ortholog, the 5′ GC trends are the same as the broader native gene set (S7 Fig).

Download:

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

Fig 2. Substitution rates by 5′ codon position comparing the Escherichia coli– Escherichia fergusonii ancestor to E. coli.

A. Synonymous substitution rates (K_s); B. non-synonymous substitution rates (K_a), and C. the ratio between the two (K_a/K_s). A–C plots include orthologs that are at least 180 codons in length (n = 1,443). D–F. same as A-C but only including non-overlapping orthologous genes that are at least 180 codons in length (n = 1,310). For all panels, the x axis represents absolute codon position (i.e., the start codon is codon 1). Dashed vertical black line marks the first 10 codons. Locally estimated scatterplot smoothing (LOESS) regression lines are included. Note that codon position here is by reference to the codon position in the alignment. Removal of alignment indels in the focal lineage prior to codon position categorization makes no meaningful difference (Pearson correlation between K_s with indels v K_s without = 0.99 P-value = 1.31 × 10⁻¹⁴², likewise for K₄: S8 Fig). The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g002

With this benchmark, next we consider the influence of overlapping CDSs. To account for this, we consider reference genome annotations and repeat the conservation analysis for E. coli orthologous genes only considering those that are not overlapping (meaning no other CDS is fully or partially encoded by the 5′ sequence of the focal CDS). We perform the same conservation analysis as above only for non-overlapping orthologous genes (2,027 of 2,385 orthologs) and find that, although absolute K_s values are marginally higher, the K_s trend is unaffected (Fig 2D–2F, as above we plot those that are at least 180 codons in length). This is consistent with the hypothesis that conservation at synonymous sites is not fully explained by gene overlaps. In no small part, this reflects the fact that among native genes the majority are not overlapping (669 of 4494 genes show 5′ overlaps, around 15%), and of those that are, most involve small 5′ overlaps (median overlap size = 3 bp, see S9 Fig).

While the above 15% figure is approximately as expected if 30% of genes show an overlap [65] (the others being 3′ overlaps), to determine the generality of results we ask whether the results are replicated in other species. We identified one set of Bacillus species with appropriate in and outgroup distances (B. toyonensis (focal ingroup) and B. anthracis, with B. mycoides as outgroup). This also allows us to test whether we observe the same trends in gram-positive bacteria, which differ from gram-negative bacteria (such as E. coli) in a number of ways, including lack of co-transcriptional translation in Bacillus sp. [80] and a previously reported different codon usage (specifically for Bacillus subtilis (B. subtilis) when compared to E. coli) [81].

Among the Bacillus species, we observe the same trends in all genes and in non-overlapping orthologous genes (2,970 of 3,305 orthologs): K_s starting low 5′ and gradually increasing until plateauing around codon 60 (Fig 3A–3F). For the distribution of overlaps at 5′ ends in native B. toyonensis genes see S10 Fig: as for E. coli, most genes are not overlapping (479 of 5229 genes show 5′ overlaps, around 9%) and overlaps are commonly short (median overlap size: 3 bp). B. toyonensis also replicates similar native GC content along the CDS as in E. coli, i.e., plateauing around codon 10 (S11 Fig), suggesting the K_s trends cannot be explained by different mononucleotide trends in orthologous genes. We further confirm that the extended conservation trend cannot be explained by gene overlaps and furthermore that it is not unique to E. coli (or gram-negative bacteria).

Download:

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

Fig 3. Substitution rates by 5′ codon position comparing the Bacillus toyonensis–Bacillus anthracis ancestor to B. toyonensis.

A. Synonymous substitution rates (K_s); B. non-synonymous substitution rates (K_a), and C. the ratio between the two (K_a/K_s). A–C plots include orthologs that are at least 180 codons in length (n = 1,933). D–F. Same as A–C but only including non-overlapping orthologs that are at least 180 codons in length (n = 1,703). For all panels, the x axis represents absolute codon position (i.e., the start codon is codon 1). Dashed vertical black line marks the first 10 codons. Locally estimated scatterplot smoothing (LOESS) regression lines are included. Note that codon position here is by reference to the codon position in the alignment. Removal of alignment indels in the focal lineage prior to codon position categorization makes no meaningful difference (Pearson correlation between K_s with indels v K_s without = 1 P-value = 2.08 × 10⁻¹⁵⁶, while for K₄ = 0.99 P-value 8.81 × 10⁻¹⁴⁶: S8 Fig). The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g003

Extended synonymous site conservation cannot be explained by a translational ramp

Having established that in Bacillus and Escherichia there is both an extended window of low K_s and that this cannot obviously be explained by overlapping genes, we sought to consider a further prominent model. Given that translationally optimal codon usage is reduced towards the 5′ end of genes in E. coli, Tuller and colleagues hypothesized that this was an adaptation to slow ribosomes to enable a more orderly initiation process, what they call a ramp [67]. To investigate this, we consider the trends in usage of unambiguously translationally optimal codons, these matching the most abundant tRNA in the transcriptome and showing classical HEG enrichment patterns.

We employed the optimal codon calls of Wei and colleagues [74] (see Methods). Plotting the usage of optimal codons against position we see, as previously seen, a tendency in E. coli to relatively underuse optimal codons in the 5′ end, this starting to plateau after about codons 10–20 (Fig 4A). It continues up to and past codon 150. In Bacillus, optimal codons tend to be A/T-ending so provide a highly informative test case. Indeed, the ramp model still predicts avoidance of optimal codons, while the low stability-A/T epiphenomenon model predicts the opposite of the slow-to-fast ramp. As predicted by the low stability-high A/T model, codon adaptation is extremely high in the 5′ end of Bacillus (Fig 4B). After codon 10–20, the trend is flat to negative. Both trends are the opposite of the predictions of the ramp model. As regards optimal codon usage and GC3, in both Bacillus and E. coli, the trends in operonic genes are the same as those in non-operonic genes (S12 Fig) indicative of effects beyond the ribosome engaging the most 5′ end of the polycistronic RNA.

Download:

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

Fig 4. Deviation in usage of optimal codon trends as a function of distance from the CDS start for A. Escherichia coli, B. Bacillus subtilis.

Linear regression and Pearson correlation with respective P-value displayed. Lines, points, and statistics in pink consider the first 10 codons (inclusive), those in blue are for all other codon positions. C. Trends in AT content at codon third sites across all native E. coli genes. D. Same as C, but for B. subtilis. Note that the x axes in all four panels match in terms of position within the CDS as codon position refers to absolute number (where the start codon is codon 1); nucleotide position 1 refers to that following the start codon. Dashed vertical black line marks the first 10 codons. The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g004

The centrality of nucleotide content above codon optimality is underscored by analysis of further species in which codon optimality can be defined by reference to tRNA levels [74]. There are four other species for which we have a good definition of optimal codons that reflect a broad range of GC contents: Leptospira interrogans GC3 ~29%, Bacteroides thetaiotaomicron GC3 ~45%, Synechocystis sp. GC ~50% and Mycobacterium tuberculosis with GC3 ~80%. In contrast to E. coli and Bacillus, these are all slow-growing [74]. We consider trends in both GC3 and optimal codon usage as a function of distance from the start codon. All four species show GC3 steeply going from low to high in the first ~10 codons (S13 Fig). For patterns of optimal codon usage there is no similar uniformity. Mycobacterium resembles E. coli with codon adaptation running from low to high, then gently further rising. Bacteroides resembles Bacillus with codon usage bias running high to low, again contra the ramp hypothesis. In the other two species, within the first 10 codons the regression line is negative (high to low) but not significantly so and codon adaptation more broadly appears to monotonically gently rise. These results further support the importance of low GC content in proximity to the initiating codon, and underscore that lack of generality of the E. coli pattern of optimal codon usage that inspired the ramp hypothesis.

The trend for AT content of the third site rather than codon optimality to be predictive of usage trends [72] is especially well reinforced when considering analysis of the trends on optimal codon usage within each synonymous block, in the first ~10 codons. Indeed, while the overall trends are opposite in E. coli and Bacillus (Fig 4A and 4B), what unifies both is that nucleotide content (for trends see Fig 4C and 4D), not codon optimality, is strongly predictive: in both species, when the optimal codon is G/C-ending there tends to be an increase in its usage moving 5′ to 3′ across the first ~10 codons, while when it is A/T-ending the trend is reversed.

Specifically, in E. coli, there are 5 of the 17 synonymous codon blocks nominated by Wei and colleagues [74] where the translationally optimal codon is A/T-ending. In all 5, the correlation between codon position and relative usage is negative, significantly so in 3 (S14 Fig). By contrast, of the 12 G/C-ending optimal codons, the trend is positive in 11, significantly so in 10. The one exception is serine’s TGC that shows no obvious trend. If we consider the trends being predicted by nucleotide content as being a “success”, then this pattern (16 successes, 1 fail) is highly significant (binomial test P-value = 0.00028). Similarly, in Bacillus, of 4 G/C-ending optimal codons, all show a positive slope, 3 significantly so (S15 Fig). Of the remaining 10 A/T-ending ones, 8 have a negative slope, 5 of which are significant (S15 Fig). The two exceptions are leucine’s 4-fold block and valine, neither of which are significant. Applying the same rules as in E. coli, this amounts to 12 in agreement with the nucleotide model and 2 against (binomial test P-value = 0.01). A simple model in which A/T-ending codons are highly used at 5′ ends, this decaying as one moves out of the 5′ domain, while G/C-ending optimal codons are under employed, this decaying as ones moves out, has overall strong support (28 in support, 3 against: P-value = 5 × 10⁻⁶), even though we take a conservative approach and consider any trend whether significant or not. Just considering cases where the trend is significant (at P < 0.05), the split is 21-0 (P-value = 9 × 10⁻⁷). Thus, codon usage trends within the first ~10 codons are predicted by nucleotide content of the synonymous base, not by translational optimality, in contradiction of the ramp model’s original claims [67].

The ramp model not only attempts to explain slow codons (non-optimal) giving way to faster codons, but it also suggests that ribosomes will gradually accelerate after the first section of the CDS [67]. There should thus be a tendency for a positive slope on optimal codon usage by the position after the initial ~10 codons. In E. coli, whose codon usage was employed to inspire (rather than test) the model, there may be such a tendency (Fig 4A). At the level of the codon block, all but 2 (of 17) show increasing usage of the optimal codon, the two exceptions both ending in A/T (S14 Fig). Bacillus data, however, strongly contradicts the model. There is no evidence for a positive trend overall and, if anything, the trend is the reverse of that predicted (Fig 4B). Indeed, considered at the codon block level (S15 Fig), the trend in the downstream section is for ribosomal deceleration (a negative slope for codon optimality) when the optimal codon is A/T-ending (9 of 10 blocks), and acceleration (positive slope) when the optimal codon is G/C-ending (3 of 4 blocks) (prediction from nucleotide content binomial test: P-value = 0.01, 12-2 split, 10 of 14 show deceleration, opposite of the prediction). This too suggests that in the downstream domains it is also nucleotide content, not optimality, that matters.

We conclude that the Bacillus data rejects all aspects of the ramp model: there is no preference for slow codons as such, no trend for early codons to be rarer than downstream and the trend after the first ~10–20 codons is not for increasing codon adaptation. As the synonymous substitution rate does nonetheless increase well beyond the limited ~10–20 codon domain, the ramp hypothesis provides no explanation for trends in Bacillus. In addition, while in E. coli K_s plateaus at codon 60, codon adaptation keeps monotonically increasing. For all these reasons, we reject the ramp as a viable explanation both for codon usage trends in 5′ domains and for the zone of reduced K_s in both species.

A mutational model for E. coli is not falsified

Although selection was the first explanation considered for the low 5′ synonymous substitution rate [39], spontaneous mutation may yet explain variation in rates. We start by employing data from MA experiments in E. coli [56]. We divide all CDSs in windows and for each window determine the sum number of mutations observed in that window across all genes and the sum number of base pairs within such windows. We can then determine the density of mutations per window (i.e., mutations per kb of sequence across all MA lines). We find that the first 10 codons have a density of 0.115 per kb which compares with 0.231 for the gene body (post-codon 70), the former being 50% of the latter (Fig 5A). We note that this is quantitatively similar to the ratio of the mean K₄ (i.e., K_s at 4-fold degenerate sites) for the first 10 codons (excluding the start codon) and the mean of those post-70 codons (for K₄, ratio = 0.0853/0.185 = 0.46). The first 60 codons have a positive slope on the line while the mutational density after that point is not significant (Fig 5A), mirroring that which is seen for K₄ (Fig 5B). The same trend is seen when considering genes without overlaps (S16 Fig). We have replicated the low 5′ mutation rate using Zhang and colleagues’s [57] higher resolution spontaneous mutation data generated with Duplex Sequencing that employs 3.5 × 10⁴–3.8 × 10⁴ reads (but may be biased [59]), and MA data from Foster and colleagues [58] (S17 Fig). The same trend is not, however, so in evidence in Zhang and colleagues’s samples sequences at “lower” depth (<1,500 reads) for reasons unclear (S17D–S17F Fig). Nonetheless, these results ensure that we cannot reject the hypothesis that the K_s trends running until codon 60 is a consequence of low mutation rate. Low K_s therefore need not be indicative of purifying selection.

Download:

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

Fig 5. Mutation profile as a function of distance from the gene’s start.

A. Mutation density (mutations per kilobase, kb) as a function of within gene position. The amount of sequence with each genic window, across all CDS, was determined, the density then being the number of mutations per bp, here scaled to kb. The blue line is a polynomial regression of degree 4. Yellow dashed line and yellow statistic is for the first 60 codons, dark purple dashed line and dark purple statistics is for the rest of the gene. Pearson correlation provided. B. Comparison of K₄ values by codon and mutation density from wild-type (WT) lines. Mutation density is in blue with positions specified by middle position of the window. K₄ data per codon is in pink. Lines reflect polynomial regression of degree 4. To determine pseudo-significance, we interpolate values for each codon by fitting to the blue polynomial line. These values are then correlated against the observed K₄ values (Pearson correlation shown). C. Deviation from null (O − E)/E for WT (alternative metric for data in panel A) in blue and from MA lines that have MMR deleted in pink. The first 60 codons are positively correlated for the WT data (statistics as panel A), but the MMR deletion data is not (Pearson correlation r = 0.78, P-value = 0.06). Dark purple dashed line is regression for data post-60 codons for MMR-deficient data, yellow dashed line for data within 60 codons. The pink line is the polynomial regression (of order 4) for MMR-deficient, the blue for WT. The horizontal gray line marks (O − E)/E = 0. In all panels mutational data from Wei and colleagues [56]. The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g005

Mismatch repair does not fully explain reduced 5′ mutation rates

Given the low 5′ mutation rate, the next question is what might be causing it? It is classically reported in E. coli that mismatch repair (MMR) is directed to the genes [58,82], this giving intergenic sequence a higher density of observed mutations (see S18C Fig). Might something similar explain the especially low rate at 5′ ends? MMR via mutS and mutL are transcriptionally coupled [83]. In transcription-coupled repair (not involving MMR), a stalled RNA polymerase is the signal for the recruitment of repair enzymes (reviewed in [84]). If we suppose there to be some constant rate at which an RNA polymerase will prematurely abort (e.g., encounter with a DNA polymerase), thus not signal 3′ errors, then it is possible that 5′ mismatches are more likely to be repaired, rather than resolved as mutations (post-replication).

To address this issue, we consider the mutational profile in MA lines that have MMR deleted [56]. This data was provided in the same experiment as our original wild-type (WT) MA data [56], thus controlling for numerous potential confounding variables. In this data, the first 10 codons have a mutational density of 10.6 per kb, which compares with 11.7 in the gene body (post-codon 60), i.e., about 90% of the rate. While the contrast between 50% (as above in the WT) and this 90% figure seems at first sight indicative of a role for MMR, caution is required in this interpretation. A statistical consequence of an absolutely higher rate is that the difference in relative mutational density between 5′ and gene body regions could trivially be lesser. Consider, for instance, that in the WT the 5′ end has a mutation rate x (per unit of sequence), the gene body then having a rate 2x. Now imagine that loss of MMR forces all sites to have 100x more mutations (approximately as seen). The 5′ end will then have 100x + x mutations and the gene body 100x + 2x mutations. The relative mutation rates in the 5′ end are now not 50% that of the gene body but 101/102 the rate. Thus the transition from WT having 50% of the rate to the MMR knockout having over 90% of the rate could be explained by nothing more than an equal increase in the absolute number of mutations, but with the 5′ having a lower—in absolute terms—mutation rate (the difference between 5′ and gene body would be x in both cases). That the ratio is ~90% not 99% could, nonetheless, be consistent with MMR operating differentially across gene bodies.

That this naïve model is incomplete is suggested by the fact that in the WT the absolute difference in mutation rate is of the order of 0.1 per kb, while in MMR-deficient lines it is ~1 per kb, i.e., not a simple addition. As the absolute difference is in fact higher in the absence of MMR, one could conclude that MMR if anything suppresses the differences between 5′ and gene body. We suggest, however, that close analysis of the dynamics of MMR would be needed to address these issues. We thus are agnostic as to effects of MMR on the 5′ v gene body difference.

What, however, is clearer is that MMR alone is not adequate to explain all the deviation seen in mutation rates. To enable comparison with the WT lines, we convert data to deviation from expected, where in each genome we consider the total number of mutations in the genome as a whole and the amount of sequence in each gene window summed across all genes. From the proportion of sequence in any given window we then derive the expected number of mutations as the product of the proportion of sequence within a window, multiplied by the total number of mutations. From this, we compute (Observed − Expected)/Expected ((O − E)/E), this being relatively insensitive to sample sizes. As can be seen, MMR-deficient lines have (O − E)/E values close to zero across the gene body (Fig 5C). The WT data is a transposition of that in Fig 5A with deviation scores that are highly negative in gene bodies, this reflecting MMR’s activity in gene bodies (and the consequential higher intergenic rate). While the trend in the MMR-deficient condition in the first 60 codons is marginally not significant (Pearson correlation r = 0.78, P-value = 0.07) there are significantly fewer mutations in the first window than expected by chance. There are 1,398 of 48,141 mutations that occur in CDS in the first 10 codons. These same 30 bp are 3.227% of all CDS and as such we would expect 1556 mutations if the distribution were random. The observed number is lower than this (χ² = 16.583, P-value <0.00005, df = 1). If we sum all mutations across the 5′ end, only when we get to the window ending at codon 60 does the 5′ end not have fewer mutations than expected (at P-value<0.05), allowing for multiple testing.

Compositional variation likely explains some of the variance in evolutionary rates

The above evidence suggests that even if MMR is more efficient in the 5′ domain, even allowing for this, there is still a lower mutational density than expected under a model of mutations being randomly distributed within genes. What else might affect this? One possibility is GC-associated mutation bias. To appraise this, in the first instance we predict, given the observed genomic mutational frequencies for each nucleotide, the mutation rates at each codon position normalized by the nucleotide content at the same position across all E. coli genes. Our expectation is that this positional trend is unlikely to replicate the K_s trend (that plateaus at codon 60) as it is simply going to follow nucleotide content trends (that alter at codon ~10–15) and represent their known mutability. To determine the per nucleotide rates, we determine the number of mutations for each nucleotide genomically, dividing it per occurrence of the ancestral nucleotide in the E. coli reference genome. We then scan each reference CDS and find the expected by-position mutation rates by multiplying nucleotide counts at each position across all genes by the observed genomic mutational frequencies previously calculated. Predictably, we find expected mutational frequencies to follow nucleotide trends of native genes, representing the known nucleotide mutagenicity (S19 Fig). Indeed, as C is known to be highly mutagenic and A to be lowly mutagenic [85], and that native genes are A-rich at 5′ ends especially in their first ~10 codons (Fig 1), the predicted mutation rates follow the 10-codon effect (S19 Fig). We see the same trend at all three codon sites (S19A–S19C Fig). This hints at why the 5′ domain might have a low K_s. For potentially better resolution of mutational properties, we also consider the larger mutational sample of Zhang and colleagues [57] and observe similar trends (S19D–S19F Fig).

One key weakness with the above approach is that we have determined observed and predicted mutational rates at each codon position only considering mononucleotide context. However, as mutation rates are influenced by immediate flanking nucleotides [54,86], we also perform a trinucleotide mutation analysis considering all instances where a trinucleotide N₁N₂N₃ mutates to N₁M₂N₃, M being a point mutation. N₁, N₂, and N₃ can be any one the four nucleotides and need not be the same as each other. We analyze all 64 possible trinucleotides across the genome.

We first determine which trinucleotides are most prone to mutation by finding their genomic mutation frequencies normalized to their ancestral (pre-mutational) occurrence (Fig 6). We find that the E. coli genomic mutational profile is largely dominated by GC-rich trinucleotides (Fig 6A), consistent with previous findings (see, e.g., [57,87]). The most commonly mutating trinucleotide for Wei and colleagues data is GCC (Fig 6A), and trinucleotide mutation frequency is significantly positively correlated to trinucleotide GC content (Pearson correlation r = 0.42, P-value = 0.0005, Fig 6B). Considering Zhang and colleagues data, the most mutating trinucleotides are also GC-rich and especially CpG-containing, CpG indicating the dinucleotide (S20A Fig). The mutation frequency distribution is less homogeneous compared to that of Wei and colleagues data. Indeed, the four most commonly mutating trinucleotides (per occurrence of the trinucleotide) in Zhang and colleagues are CpG-containing and account for 61% of normalized mutation rates (S20 Fig). The two most mutating are GCG and CGC (accounting for 48%) which are more likely to be creating further CpG dinucleotides if flanked by G or C (S20A Fig). Here, we expand the analysis to Bacillus as well, and compute observed trinucleotide mutation rates for B. subtilis MA data [62]. We find that the tendency for GC richness among the most commonly mutating trinucleotides persists (Fig 6F and 6G).

Download:

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

Fig 6. The influence of trinucleotide context on mutation and substitution.

A. Observed genomic trinucleotide mutational frequencies rank-ordered from most to least frequent in E. coli. Mutational data for E. coli from Wei and colleagues [56]. Mutation frequency refers to mutation count per occurrence of ancestor trinucleotide. Trinucleotide mutations are such that the middle base is the mutated base. Trinucleotides on the x axis are rank-ordered by frequency and bars are color-coded by trinucleotide GC content. B. The same genomic mutation frequencies as in A, plotted against trinucleotide GC content. Line represents linear regression and Pearson correlation with respective P-value is also shown. C. Expected trinucleotide mutation rates by codon position, predicted by trinucleotide genomic mutational rates and genomic trinucleotide content. Trinucleotide mutations are such that the middle base is the mutated base, and it occurs at third sites in 4-fold degenerate codons. D. Comparison of expected mutational rates in C with E. coli conservation trends by codon position at 4-fold degenerate sites (4-fold K_s, K₄, as seen in S5 Fig). Both metrics are normalized by Z score. E. Comparison of expected trinucleotide mutation rates and K₄ by position without Z transformation. Pearson correlation data is shown. Line is the orthogonal (major axes) regression line. F–J as A–E but for Bacillus sp., with mutational data for B. subtilis from Sung and colleagues [62], and conservation trends at 4-fold sites in B. toyonensis. For C–D and H–I position on the x axis refers to absolute number of codons (where the start codon is position 1), and the dashed vertical black line marks the first 10 codons. Locally estimated scatterplot smoothing (LOESS) regression lines are also shown. The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g006

We then ask how trinucleotide mutation rates translate in terms of expected by-position mutation rates as done above (S19 Fig) for mononucleotides. To capture solely synonymous mutation trends, we only consider cases where the mutated base occurs at a third site of 4-fold degenerate codons. We find the resulting trend (Fig 6C) to replicate the form of the K_s trend, with expected trinucleotide mutational rates starting low at the 5′ end and increasing until plateauing around codon 60 (Fig 6D and 6E). We observe the same whether we consider Zhang and colleagues’s [57] E. coli samples sequenced at higher (S20A–S20E Fig) or lower (S20F–S20J Fig) depth. The trends also match when considering K_s and 4-fold trinucleotide mutability in Bacillus sp. (Fig 6F–6J). In both species, the observed K_s and trinucleotide predicted rates are strongly correlated (Fig 6E and 6J). Observing the difference in expected trinucleotide mutation frequencies by position at 5′ ends and gene body, we see that the average mutation frequency in codons 2–10 is around 91% that of codons 60+ in E. coli MA data (~95% for Bacillus). Although this does not replicate the 50% difference seen in the observed trinucleotide mutation rates trends (Fig 5), this suggests we can use underlying trinucleotide trends to explain at least in part the lower mutational density of 5′ windows. A GC effect likely explains why in MMR-deficient lines the 5′ end still has a low mutation rate.

Our approach, extrapolating from well-resolved genome-scale data, comes with a caveat in that we have assumed that the 5′ end and CDS core have the same mutational properties, and that they are also the same as what we see genomically. We can test this in several ways, but only in E. coli mutational data from Zhang and colleagues as this uniquely has a high enough number of mutational calls. First, we consider the observed mutational rates at 5′ and core in E. coli, separating the observed mutations in the two regions (i.e., not by taking the global sample). Given the yet limited mutational sample size, we are unable to consider mutations only in the first 10 codons and thus consider the lowest round number of codons from genes’ start that allows analysis: the first 20 codons, while we consider gene cores as the rest of the CDS. We find the two mutational profiles to be significantly positively correlated (S21D Fig, Spearman rho = 0.8, P-value = 3.5 × 10⁻¹³). The correlation is also positive for Wei and colleagues [56] data though, likely due to lack of data resolution as mentioned above, not significant (S21A Fig, Spearman rho = 0.15, P-value = 0.44). Moreover, as above for genomic frequencies, we find the observed trinucleotide mutational frequencies given their ancestral occurrence in each gene region, and also find that the higher mutation rates tend to be in GC-rich trinucleotides (S21E and S21F Fig). We consider 5′ and core mutations at the level of the dinucleotide and consider the predicted mutational equilibrium (i.e., if the only activity was mutational bias and neutral evolution). For this, we see no difference between the 5′, the core and the global genomic predicted equilibria (P > 0.2 in all pairwise comparisons, S3 Table; we return to this below). We thus find no evidence for different normalized mutational profiles in the 5′ and core. The observed differences in expected mutation rates are thus better explained as owing to differences in relative frequencies of different trinucleotides, not the mutational properties of those trinucleotides.

No robust evidence for selection for high AT from mutational equilibrium analysis

While we can account for the trend in K_s as a function of spontaneous mutation rates, can we exclude the hypothesis that synonymous mutations at the 5′ end are under selection? The transgene data strongly suggest that synonymous mutations can have a meaningful effect on protein levels (see [41–43] and Fig 1D and 1E) and the commonality of GC3 running from low to high in the first 10 codons in species with both high and low GC content (S13 Fig) is indicative of a selectively crafted GC profile.

An alternative way to ascertain whether there is selection (or biased gene conversion) acting at 5′ ends in favor of AT mutations is to determine the AT content expected at neutral mutational equilibrium and compare it to the AT content observed in native genes. Assuming the species is at nucleotide content equilibrium, an observed AT content higher than that predicted at mutational equilibrium would be consistent with selection on G/C->A/T mutations, if lower it would be consistent with selection favoring A/T->G/C mutations, while if the equilibrium is approximately the same as the observed composition then there is no need to evoke selection [54,87,88,89]. Long and colleagues [54] determined from MA data a genomic AT content at mutational equilibrium (AT*) of 0.52 for E. coli and therefore lower than natively observed at 5′ ends, this being consistent with selection favoring a higher AT content, the canonical model, as we discussed in the Introduction. With more data, we can now return to this issue. While we find that few mutational data sets are consistent with selection favoring G/C->A/T mutations (the canonical model), we unfortunately discover that different sets predict strikingly different equilibrium nucleotide contents, rendering strong conclusions impossible.

In E. coli, we estimate the equilibrium AT content to range from 0.55 to 0.89 depending on the experiment (Table 1). The highest figures (0.86 and 0.89) are derived from Duplex Sequencing approaches, but even the MA line data estimates (0.55 and 0.65) are significantly different one from another (Table 1), and all higher than the prior estimate of 0.52 [54]. The average observed AT3 content at the 5′ end is 0.55 with a maximum at any codon position of 0.62 (Table 1). If the lowest MA estimate (AT* = 0.55) is correct, then selection on G/C->A/T mutations could be evoked to account for the codon position with 0.62 AT content. If we instead believe the higher estimates, these would be indicative of selection being directed A/T->G/C, otherwise the observed AT content in E. coli is close to neutral on the average (i.e., around 0.55). In B. subtilis, MA experiments with more limited data estimate the equilibrium to be around 0.56 [54], while AT3 in the first 10 codons is a bit higher (Table 1), also consistent with weak selection favoring G/C->A/T mutations.

While these data are strikingly contradictory, we can exclude the possibility that the discrepancies are owing to problems in the methods to estimate the equilibria. Above we perform the simplest model that considers the ratio of G/C->A/T mutations per G/C to the sum of G/C->A/T mutations per G/C and A/T->G/C mutations per A/T (as performed by Long and colleagues [54], see Methods). For larger data sets, we can also consider a more complex mononucleotide-based method considering all 12 possible mutational classes and solving the relevant simultaneous equations (see Methods), as well as a dinucleotide model with a 16 x 16 mutational matrix with 256 parameters (mutations from one dinucleotide to another, per ancestral occurrence of the focal dinucleotide, see Methods). All methods applied to Zhang and colleagues’s mutational data [57] agree that the AT*~85% (for full genomic data from the samples sequenced at higher depth, AT* by simpler G/C<->A/T method = 0.86; AT* by full mononucleotide method = 0.86, AT* by dinucleotide method = 0.85, see S3 Table and S2 Text).

While some AT* estimates above (Table 1) suggest a model of selection in favor of A/T->G/C mutations, albeit weak, this comes with the caveat that we assume the mutational biases are identical in all gene regions. We can also employ Zhang and colleagues’s higher resolution spontaneous mutation data [57] to ask whether there might be possible differences in mutational spectra in different genomic/gene regions. To this end, we derive AT* content by considering a) the global genomic mutational profile, b) the profile of mutations seen in the first 20 codons of genes, and c) the profile of mutations seen in the gene cores (everything in CDS after the first 20 codons). For the genomic AT*, we estimate a value of 0.85 ± 0.015 (standard deviation of 1,000 bootstraps), for the gene cores we estimate AT* = 0.84 ± 0.027 and for the 5′ domain AT* = 0.87 ± 0.12. All three (Fig 7A) are considerably higher than observed AT contents (Fig 7B) and none significantly different from any other (see bootstrap error bars in Fig 7A). As an additional check, we also perform the mutational equilibria test considering samples sequenced at a lower depth, and find that the predicted AT content also remains high (AT* = 0.88 ± 0.028 genomically, 0.93 ± 0.013 at 5′ ends and 0.88 ± 0.016 at cores, P-value = 0.54 for test of difference between the two, see S3 Table). We thus see no reason to suppose that we cannot extrapolate from genomic mutational profiles.

Download:

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

Fig 7. AT content within and across genomes and its influence on transgene activity.

A. Estimated AT content expected at mutational equilibrium determined from rates of spontaneous mutation at 5′ ends, at gene cores and overall genomically in Zhang and colleagues data [57]. Nucleotide content determined using a dinucleotide approach for both samples sequenced at low and high depth, see Methods. Error bars represent standard deviation of the mutational equilibria calculation for 95% bootstrap bounds of 1,000 re-samplings. B. Observed AT in E. coli native genes. Error bars represent standard error of the mean (SEM). In both panels A and B 5′ ends are taken to be the first 20 codons and gene core is the rest of the CDS, while genomic refers to the whole genome (including non-protein-coding sequences). Legend describing gene regions applies to panels A and B exclusively. C. GC3 trends at 5′ ends and gene core in 1,355 bacterial species. The linear pink line represents a line of slope 1 and intercept 0 (i.e., perfect correlation between the two gene regions). The blue line is a quadratic fit to the plotted values. Error bars represent SEM. D–F. Expression levels of transgenes as a function of 5′ GC4 content for D. protein per RNA E. protein and F. RNA. GC4 refers to GC content at third sites of 4-fold degenerate codons. We take the 5′ end to mean the first 10 codons following the start codon. Each bin contains either 15,522 or 15,523 transgenes. Error bars represent SEM. Transgene data retrieved from Cambray and colleagues [41]. The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g007

There are multiple possible explanations as to why AT* estimates might be so variable, none of which we can satisfactorily resolve (S2 Text). We note, however, that estimates of mutational equilibria from MA lines are relatively close to the observed AT content of intergenic sequence (E. coli AT* ~0.6, intergenic AT=0.58; Bacillus AT* ~0.56, intergenic AT=0.64; Mycobacterium smegmatis AT* = 0.43, intergenic AT = 0.39; Table 1). This inclines us to suppose that the MA data is more likely to be nearer the truth. Nonetheless, given the large variation in mutation equilibrium estimates, even between MA line estimates (0.55 and 0.65), the safer conclusions are that a) in this instance, the deviation from equilibrium test appears to be too highly contingent on some details of experimental protocol to make for robust inference and b) we see no robust evidence for G/C->A/T mutations being favored by selection, with alternative estimates consistent with neutrality or with selection favoring AT->GC mutations.

High 5′ GC3 in AT-rich organisms supports selection for raised GC

If the AT content is actually lower than expected, as some values of the mutational equilibrium estimates would suggest, this would be consistent with selection in favor of A/T->G/C mutations (the opposite of the canonical model). This could mean that the process is so heavily biased at equilibrium for AT that, while higher AT and lower stability are favored at 5′ ends compared with gene cores, 5′ ends nonetheless are under selection for a non-minimized stability and non-maximized AT. Were such a model correct, we might also expect that bacteria with extreme AT pressure might evolve 5′ ends with a higher GC3 content than the gene core (most bacteria have GC3 of the 5′ end lower than that of the core [40]). We thus determine trends in average GC3 across 1,355 bacteria species (see S4 Table for species list). We confirm this expectation observing that GC3 at the 5′ end is higher than that at the core when core GC3 is lower than about 35% (Fig 7C). We note that the same effect was tested for, but not found, by Allert and colleagues, who report high 5′ AT in AT-poor genomes using a more limited sample of genomes (n = 816) [45].

Selection for higher GC content is not expected to increase protein production

The next natural question is then whether higher GC content at 5′ ends might influence gene expression. Perhaps maximal gene expression is also found when the AT content is high, but not exceptionally high? To answer this, we consider again the transgene data by Cambray and colleagues [41] and ask whether those constructs with very low levels of 5′ GC at 4-fold degenerate sites produce lower levels of protein per RNA than those transgenes with slightly higher GC. We find that they do not (Fig 7D). Rather, transgene expression (protein per RNA) steadily decreases as 5′ GC4 content increases. Interestingly, this trend seems to be driven by protein levels rather than RNA, as GC4 poor constructs tend to produce the highest levels of protein (Fig 7E), though the lowest of RNA (Fig 7F). Nonetheless, our results suggest that if there is selection acting on 5′ ends towards GC, it is not simply favoring higher protein expression per RNA. Other factors such as ribosomal usage efficiency or noise reduction may also be playing a role, as previously discussed for both bacteria and eukaryotes [90–92]. This observation could also be pointing to an alternative explanation to selection for GC, as lower AT content could also simply be a result of biased gene conversion directed towards GC [87].

Analysis of Mycobacterium smegmatis suggests that K_s is more likely to mislead when mutation and selection operate in the same direction

In E. coli (and Bacillus sp.), it seems likely that part of the lower rate of mutation at 5′ ends is owing to compositional differences. This is in no small part because the 5′ end is AT-rich and AT residues in these two species are less mutagenic than GC-rich residues. Thus, in the case of E. coli, as mutational bias at the 5′ end is in the same direction as putative selection for low stability (G/C->A/T), by Ocham’s razor we should not interpret the low K_s as unambiguous evidence for selection on synonymous mutations. As seen above, however, at extremes of nucleotide content the conclusions drawn from E. coli need not apply: in very AT-rich genomes the 5′ ends are more GC-rich than the gene cores (Fig 7C). Potentially the most interesting case comes when the mutation bias is anticorrelated to the direction of selection. To consider what might happen in this instance, we consider an unusual species in which the mutation bias is the reverse of the common GC->AT bias. M. smegmatis is exceptional in that it has lost mutL and mutS and has a mutation bias that is highly AT->GC biased [61]. While it has lost two key MMR enzymes, it has gained a replacement in NucS [60]. The unusual mutation bias is likely reflected in part in it having an exceptionally high GC content (around 80% GC, Table 1). We find that its 5′ CDS is also GC-rich, just not as GC-rich as the gene core (Fig 8A). Thus, like E. coli and B. subtilis, it too has accelerating GC content (and hence decelerating AT content) moving 5′ to 3′, consistent with selection for reduced mRNA stability.

Download:

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

Fig 8. Mycobacterium smegmatis mutation and conservation analysis.

A. AT content by nucleotide position in native M. smegmatis genes. The x axis represents nucleotide positions relative to the start codon (i.e., the first nucleotide of the codon after the start codon is at position 1). Error bars indicate the standard error of the mean (SEM). Dashed vertical black line marks the first 10 codons, and a locally estimated scatterplot smoothing (LOESS) regression line is shown. B. Observed genomic trinucleotide mutational frequencies rank-ordered from most to least frequent in M. smegmatis. Mutational data from Kucukyildirim and colleagues [61]. Mutation frequency refers to mutation count per occurrence of ancestor base. Trinucleotide mutations are such that the middle base is the mutated base. Trinucleotides on the x axis are rank-ordered by frequency and bars are color-coded by trinucleotide GC content. C. Estimated AT content expected at mutational equilibrium determined from rates of spontaneous mutation overall genomically in MA data from Kucukyildirim and colleagues [61]. Nucleotide content determined for WT samples using a simultaneous equations approach for mononucleotide changes, see Methods. Error bars represent standard deviation of the mutational equilibria calculation for 95% bootstrap bounds of 1,000 re-samplings. D. Observed AT content in the M. smegmatis native genome. Error bars represent SEM. 5′ ends are taken to be the first 20 codons and gene core is the rest of the CDS, while genomic refers to the whole genome (including non-protein-coding sequences). Legend color-coding gene regions applies to panels C and D exclusively. E. Expected trinucleotide mutation rates by codon position, predicted by trinucleotide genomic mutational rates and genomic trinucleotide content. Trinucleotide mutations are such that the middle base is the mutated base, and it occurs at third sites in 4-fold degenerate codons. F. Comparison of expected mutational rates in C with M. smegmatis conservation trends (obtained from three-way analysis with M. goodii and M. septicum as outgroup) by codon position at 4-fold degenerate sites (4-fold K_s, K₄, as seen in S5 Fig). Both metrics are normalized by Z score. For panels E and F, position on the x axis refers to absolute number of codons (where the start codon is position 1), and the dashed vertical black line marks the first 10 codons. Locally estimated scatterplot smoothing (LOESS) regression lines are also shown. G. Comparison of expected trinucleotide mutation rates and K₄ by position without Z transformation. Pearson correlation data is shown. Line is the orthogonal (major axes) regression line. The data underlying this Figure can be found in https://doi.org/10.5281/zenodo.17378284.

https://doi.org/10.1371/journal.pbio.3003569.g008

With MA mutational data [60,61] we have also determined the mutability of all trinucleotides and unusually, the GC-rich ones are not the most mutagenic (Fig 8B, Pearson correlation for trinucleotide GC content v trinucleotide mutability: r = 0.04, P-value = 0.756). The predicted mutational equilibrium of this species is an AT* of 0.43 (Fig 8C and 8D, Table 1, see also Discussion; Long and colleagues report AT* = 0.42 [54]). Employing the observed 5′ CDS, we can then predict the expected profile of substitutions under a mutation bias model (Fig 8E). Exceptionally, this then predicts a higher rate of mutation at the 5′ end, it being less GC-rich. This is supported by MA data [60,61], although the trend is much less clear than in E. coli (S22 Fig). Despite this, the first 60 codons have a mutation rate higher than the rest of the gene body (χ² = 8.96, P-value = 0. 0027, df = 1). Thus, unlike E. coli, the 5′ domain appears to have a higher mutation rate owing to the reversed mutation bias, but a preference for AT richness (relative to the gene body).

If in this species K_s reflects the mutation process alone, then it should be higher at the 5′ end. By contrast, if there is selection for higher 5′ relative AT content, then selection (low K_s) should be especially in evidence as it runs counter the mutation bias. We find that the profile of substitutions (K_s) shows a reduced K_s in the 5′ end (Figs 8F, 8G, and S23), consistent with selection opposing the mutation bias. However, while the predicted mutational equilibrium of this species is an AT* of 0.43, through all parts of gene bodies, including 5′ ends, the observed AT is much lower than this (Table 1). This is consistent with selection, but selection (or biased gene conversion) in the direction of the mutation bias favoring A/T->G/C mutations, the strength of the effect is just weaker in the 5′ end than on the gene body. Why this species has such a dramatically high GC content in the CDS is unclear. Considering MA lines with NucS deleted, we find that variation in mutation rate across the gene body disappears, implicating mutational repair bias in the intragenic variation in mutation rate (χ² = 0.36, P-value = 0.548, df = 1; S22 Fig).

Nucleotide content trends by codon position in native genes

Reference genomes were retrieved from RefSeq NCBI [109]. For E. coli, native gene analysis were performed on the reference genome downloaded on 13th September 2023 with accession GCF_000008865.2. For B. subtilis, the reference genome was downloaded on 16th May 2025 with accession GCF_000186085.1, and for B. toyonensis on 16th January 2025 with accession GCF_016605985.1. For M. smegmatis, the reference genomes were downloaded with accession GCF_000283295.1.

All genes underwent the necessary sequence checks: starting with a start codon (any allowed by translation table 11 [110]), ending in a stop codon, being in multiple of three, containing only canonical bases and no internal stops. Average nucleotide content was then determined for each codon position along the CDS, with division into the three codon sites. For 4-fold content, the average content for each nucleotide was determined as a proportion of all 4-fold degenerate sites.

mRNA stability prediction for native genes

mRNA stability was predicted using the ViennaRNA R package v 2.0 [55] using a sliding window approach with windows of 30 bp and shifts of 3 bp (to give a by codon mRNA stability prediction). The analysis was conducted from base pair -15 until codon position 150. This was conducted for all E. coli native genes and an average per codon position was determined.

Evaluating impact of codon position on transgene expression levels

All protein and RNA transgene expression data, as well as transgene sequences, was retrieved from Supplementary Data 15 of Cambray and colleagues [41]. PNI was used as protein measure and RNA_SS as RNA measure. The R package relaimpo (v 2.26) [78] was used to assess the relative influence of GC content on PNI per RNA_SS (protein/RNA) for each 5′ codon position until 20 codons after the start codon (with lmg model protein/RNA ~ GC_codon_2 + GC_codon_3, etc.). This approach quantifies the contribution of each predictor variable while accounting for covariance among them. The analysis was also repeated using protein only rather than protein/RNA.

Correlation and partial correlation analysis

The Spearman correlation between PNI/RNA_SS and GC content was found by codon position for the first 32 codons following the start codon (maximum length of sequence provided by Cambray and colleagues for each of their 244,000 constructs) using the rcorr function in the R package Hmisc [111]. A partial correlation was also determined to examine the effect, while accounting for the influence of all other codon positions, using the R package ppcor [112]. Similarly, a correlation and partial correlation analysis was performed between GC content and cell fitness (Cambray and colleagues’s W_RI [41]), and CAI values taken to represent the enrichment of a codon relative to its synonyms in the 10 codons at genic cores in highly compared to lowly expressed native genes (nHEGs v nLEGs) according to amassed protein abundance data [113]. CAI is represented through log odds ratios such that a higher value means higher usage in nHEGs.

The same packages were used to determine Spearman correlation and partial correlation between predicted mRNA stability and transgene expression. mRNA stability was predicted using the ViennaRNA R package v 2.0 [55]. Correlation with expression was separately found between stability of three transgene sequence regions (codons 2–11, 12–21, and 22–31), provided by Cambray and colleagues [41].

Calculation of optimal codon enrichment trends

Wei and colleagues [74] (their Table 1) provide a list, for E. coli and B. subtilis separately, of translationally optimal codons for any given block of codons coding for the same amino acid. Unusually, they determine the optimal codon by reference to the abundance of iso-acceptor tRNAs within the transcriptome (as opposed to copy number [75]) and require that the nominated optimal codon confirm with classical patterns expected of optimal codons (HEG enrichment, etc.). Their metric is thus comparable to tAI [75,114,115] but outperforms prior measures. They nominate 17 such optimal codons in E. coli and 14 in B. subtilis. As in E. coli, they could define one for the 3-fold degenerate isoleucine block, their codons belong to groups that are either 2-, 3-, or 4-fold degenerate. In B. subtilis, groups are either two- or 4-fold degenerate. We make the classical presumption, as assumed by the supporters of the ramp hypothesis [67], that codon optimality equates to faster processing.

To define the degree of usage of optimal codons in any set of codons (e.g., in a CDS), we implement the following approach. For every codon nominated as optimal, we recover both its identity (from their Table 1 [74]) and the degree of redundancy of the block of synonyms. We add to a list of optimal codons the nominated optimal codon and add to a list of qualifying codons all codons within in the same block (i.e., that code for the same amino acid). For example, if TGC is nominated as optimal for cysteine, TGC is added to the list of optimal codons, and both TGC and TGT are added to the list of qualifying codons, and both classed as belonging to the 2-fold degenerate class of codons. Blocks with no nominated optimal codon are ignored.

We then consider any relevant list of codons and score optimal usage within this list. For example, to determine 5′ to 3′ codon optimality trends in native genes, we employ all second site codons, all third site codons, etc. up to codon 150 for genes longer than 450 bases. In this instance, all second site codons are added to one list, all third site codons their own list, etc. For each such list, we consider each codon in turn and ask whether it exists within the list of qualifying codons for that species. If it does, we add one to the count of codons of that codon’s block degeneracy (d) giving us a count of the total number of qualifying codons of each relevant degeneracy (T_d). If it belongs to the optimal codon group, we add one to the count of optimal codons for the relevant degeneracy (O_d). For example, if, as in E. coli, TGC is optimal for cysteine, each time we encounter TGC we add one to the optimal codon count of 2-fold degeneracy and one to the list of 2-fold qualifying codons. Occurrence of TGT increments the latter count and not the former. At the end of the process, T_d is a count of the total number of qualifying codons of degeneracy d and O_d the count of optimal codons of degeneracy d.

The meaning of counts of O_d of different degeneracy has different meanings dependent on degeneracy: a null expectation is that 4-fold optimal codons will be proportionally employed (O_d/T_d), just less than 2-fold ones because there are more alternatives for 4-fold codons. Thus, for each degeneracy block of class d, we calculate deviation, D_d = (O_d − (T_d/d))/(T_d/d). In principle, unlike χ², this metric should not be biased by sample sizes. As a simple example, if we record 100 usages of the optimal codon in all 4-fold degenerate groups with 400 instances of usages of any 4-fold degenerate codon, the O_d = 100, T_d = 400, d = 4, so we record D₄ = 0. To determine the mean deviation, we considered the weighted mean D value across all applicable degeneracy classes (d = 2,3,4 for E. coli, d = 2,4 for B. subtilis). Thus, for B. subtilis, weighted mean (D) = (D₄.T₄ + D₂.T₂)/(T₄ + T₂). Employment of weighted means is desirable as the proportion of 4-fold degenerate codons varies by codon position. The weighted mean provides our metric of the usage of optimal codons for any given list of codons.

We repeat these analysis for the four others species whose optimal codons are defined using the genome accession numbers as provided by Wei and colleagues [74] (their Table 2) i.e., Mycobacterium tuberculosis (NC_000962), Synechocystis sp. (NC_017277), Bacteroides thetaiotaomicron (NC_004663), and Leptospira interrogans (AE016823).

Orthologous database retrieval

To perform a conservation analysis, RefSeq genomic coding and protein sequences were firstly downloaded from NCBI [109]. For gram-negative bacteria: Escherichia coli (GCF_000005845.2), Escherichia fergusonii (GCF_020097475.1), and Salmonella enterica (GCF_000006945.2) downloaded on 18th October 2024. For gram-positive bacteria: Bacillus toyonensis (accession: GCF_016605985.1), Bacillus anthracis str. ‘Ames Ancestor’ (accession: GCF_000008445.1), Bacillus mycoides (accession: GCF_000832605.1) downloaded on 16th January 2025. For Mycobacteria: Mycobacterium smegmatis (accession: GCF_000283295.1), Mycobacterium goodii (accession: GCF_022370755.2), Mycobacterium septicum (accession: GCF_046506965.2). Bacterial proteomes were processed with OrthoFinder (v 2.5.5) under standard configurations [116]. The species tree utilized for posterior K_a/K_s calculations was directly obtained in this step from the OrthoFinder output [117].

Alignment of orthologous sequences

Putative orthologous protein sequences were matched to the corresponding RefSeq CDS using their protein IDs. Subsequent three-way alignments were performed for each orthogroup corresponding to single-copy genes using MAFFT (v 7.526) under standard configurations for nucleotide and protein sequences [118]. Codon-based alignments were performed using PAL2NAL (v 14) software [119]. All aligned orthologous genes then underwent necessary sequence checks: starting with a start codon (any allowed by translation table 11 [110]), ending in a stop codon, being in multiple of three, containing only canonical bases and no internal stops. We also limited the analysis to orthologous genes that were at least 180 codons long (to see trends in the first 150 codons without capturing 3′ effects). The orthologs that passed checks were then used to reconstruct the ancestral state between E. coli and E. fergusonii (with S. enterica as outgroup), and between B. toyonensis and B. anthracis (with B. mycoides as outgroup) for gram-negative and gram-positive bacteria, respectively. We used a codon-based maximum likelihood approach implemented in iqtree2 [120] under -asr mode. We repeat the analysis for M. smegmatis, reconstructing the ancestral state between it and M. goodii (with M. septicum as outgroup).

Conservation analysis by codon position

For our conservation analysis, we consider E. coli and the ancestral state, B. toyonensis and the ancestral state, and M. smegmatis and the ancestral state. To determine conservation by codon position, we separately extract each codon for each ortholog and concatenate it to the codons at that same position for all other orthologs. We do this for the first 150 codons following the start codon. This results in 150 codon-specific files. We exclude cases where all three species have different codons, as it is likely for there to be uncertainty in the ancestral sequence reconstruction at those positions. We repeat the process extracting only those codons coding for 4-fold degenerate amino acids. Since not all orthologous genes have 4-fold degenerate codons at every position, the resulting codon-specific files vary in sequence lengths. We then determine the synonymous (K_s) and non-synonymous (K_a) nucleotide substitution rates, and the ratio between the two (K_a/K_s) using the CODEML program in PAML (v 4.10.7) [121]. Performing the process for each codon-specific file allows us to then assess substitution rates per codon position.

Conservation analysis for non-overlapping genes

We repeat the conservation analysis considering only genes that do not present overlaps. We consider reference genome annotations and repeat the same conservation analysis as described above for orthologous genes only considering those that are not overlapping.

Conservation analysis of operonic and non-operonic genes

Annotation of operonic genes for E. coli is from Mao and colleagues [122] with the data file downloaded from Supplement File 1. E. coli annotation is from GCF_000005845.2 to correspond. Seven genes are called in the operon file that aren’t in the reference genome. By manual inspection, these were present in NC_000913.3 but have been discontinued. B. subtilis operon annotation was from Geissler and colleagues [123] and downloaded from https://zenodo.org/records/4305872. From file BSGatlas-v1.1.xlsx, we employed the operon sheet of the Excel file and converted to csv. We obtained the corresponding genome file (CDS) for ASM904v1 from NCBI [109] along with the annotation GFF file.

Mutation accumulation and spontaneous mutation data retrieval

We employ different sources of MA data for E. coli. Our main analysis employs that of Wei and colleagues [56], which contains mutation records for both WT and MMR-deficient E. coli (RefSeq accession GCF_000005845.2). We also employ MA data of Foster and colleagues [58] merging mutation records from three E. coli strains PFM2, ED1a, and IAI1 (all considered WT and referred back to the same reference genome, GCF_000005845.2). We obtain spontaneous mutation data from Zhang and colleagues [57], who use E. coli strain ATCC 8,739 (RefSeq accession GCF_026016785.1). Their experimental design involves parallel samples grown from separate single colonies which are then sequenced via Duplex Sequencing at very high depth, allowing them to output a large catalog of spontaneous mutations across the whole E. coli genome. Zhang and colleagues data contains 12 independent samples, all of which are sequenced with “lower” depth (>1,500) and two of which are also sequenced at “higher” depth (3.5 × 10⁴–3.8 × 10⁴) [57]. Except for where it is indicated, e.g., predicting AT content at mutational equilibrium, we perform our main analysis with those higher depth samples as deeper sequencing allows to capture even the rarest mutations before selection had any filters on them. We employ another set of spontaneous mutation data recorded via Duplex Sequencing by Bhawsinghka and colleagues (for E. coli GCF_013166975.1) [59].

We additionally perform mutational data analysis for B. subtilis (GCF_000186085.1) employing MA data from Sung and colleagues [62]. For M. smegmatis we combine WT MA data from Kucukyildirim and colleagues [61] and Castañeda-Garcia and colleagues [60] (GCF_000283295.1). From Castañeda-Garcia and colleagues we also employ MA data for MMR-deficient M. smegmatis.

Analysis of mutation trends along the CDS

To determine observed mutational trends along the CDS, we divide all CDSs into fixed-size windows. For each window, we find its mutation density as the total number of mutations observed across all genes normalized by the total number of base pairs within that window (i.e., mutations per kb of sequence). Where possible, we do this both for WT and MMR-deficient mutation data (see above for data retrieval and RefSeq accessions of reference data downloaded from NCBI [109]).

To enable better comparison with the WT lines, we also convert data to represent deviations from expected values. For each genome, we first consider the total number of observed mutations in the entire genome and the total sequence length of each gene window summed across all genes. We then calculate the expected number of mutations in a given window as the product of the window’s proportion of the total sequence and the total number of mutations. We subsequently compute the deviation as (Observed − Expected)/Expected (i.e., (O − E)/E). As this metric expresses relative deviation, it is largely independent of differences in total mutation counts or sample sizes across genomes.

We next consider mutational trends relative to the sequences’ underlying nucleotide content. We first determine observed trinucleotide mutation frequencies, where the mutated base is the middle base of the trinucleotide. This allows to control for flanking nucleotides. For each of the 64 possible trinucleotides, we find the mutation counts and divide them per occurrence of the trinucleotide to give the trinucleotide mutation frequencies. Given the limited mutational sample size, we are unable to find observed frequencies for trinucleotides by codon position or codon site. For Wei and colleagues [56] and Zhang and colleagues [57] data, we instead differentiate between mutations at 5′ ends (first 20 codons following the start codon) and those gene cores (rest of the CDS). We also consider the total genomic trinucleotide mutation frequencies.

Although higher data resolution would be required to directly assess mutation frequencies by codon positions or codon site, we can infer positional trends along the CDS by estimating the expected mutation frequencies at each codon position based on the observed genome-wide mutation frequencies. In order to do this, we first determine the genomic frequencies by finding the number of mutations for each mononucleotide and dividing it per occurrence of the ancestral base in the E. coli reference genome. We then scan each reference CDS and find the expected by-position mutation rates by multiplying nucleotide counts at each position by the observed genomic mutational frequencies previously calculated. We also perform the same test for trinucleotides where the mutated (central) base is at third sites of 4-fold degenerate codons.

As repair is directed to the genes, we also ask whether the mutation rate of any given central base of each of the 64 possible trinucleotides is higher in intergenic sequence compared to annotated genes. We consider anything annotated within the full E. coli reference genome as a possible gene, including RNA, and define intergenic as the regions between these (extracted using Gff-Ex v2.3 [124]). For both classes of sequences, we determine using bedtools (v 2.31) [125] (via pybedtools v 0.12.0 [126]) whether a mutation was intergenic or genic, and compute for each trinucleotide their rate of mutation per occurrence of that trinucleotide in the relevant sequence class. We thus derive two 64-element vectors (each entry being a trinucleotide), one for normalized mutation rates for genic and one for intergenic sequence. We then compare via paired t test.

We determine trends in observed and expected mutation frequencies for E. coli, B. subtilis, and M. smegmatis (see above for data set information).

Native trends in trinucleotides

Analyses on native trinucleotide usage were performed on the reference genome downloaded for E. coli ATCC 8,739 on 28th November 2024 from RefSeq NCBI [109] with accession GCF_026016785.1, as linked to mutational data from Zhang and colleagues [57]. For each codon position, we find the summed content of trinucleotides ANT, ANA, TNT, TNA (where N is any of the four nucleotides). We consider the instances where the middle base of the trinucleotide is located at codon third sites, and we retain the codon position of that base to find trends along the CDS.

Analysis of amino acid chemical properties by codon position

We determine the average score at each codon position across eight amino acid chemical properties, four that Jin and colleagues [76] have previously found to be associated to second sites (hydropathy, chemical composition of the side chain, molecular volume, polarity), and four that supposedly aren’t associated (molecular weight, melting point, isoelectric point, refractivity). We first retrieved the reference scale for each property: hydropathy scores determined according to the Kyte–Doolittle scale where low values represent hydrophilic amino acids (see Table 2 in [127]). Scales were retrieved from [128] for isoelectric point, defined as the pH at which the amino acid loses the electric charge, melting point, polarity, molecular volume, and chemical composition of the sidechain, defined as the atomic weight ratio of hetero (noncarbon) elements in end groups or rings to carbons in the side chain. For molecular weight, scale retrieved from [129]. Finally, for refractivity, defined as the amount of refraction per gram of amino acid, the scale was retrieved from McMeekin and colleagues [130] in Jones [131].

For each chemical property, we find the average index for each codon position taking into account the observed amino acid frequencies at that position across native genes of E. coli (reference genome downloaded on 25th November 2024 from RefSeq NCBI [109], accession: GCF_000008865.2). We also find the average index for each codon position taking into account the amino acid frequencies expected from mononucleotide, calculated as probabilities based on the frequencies of mononucleotides found at the codon third site of 4-fold degenerate amino acids. A comparison between observed and expected allows to determine whether observed trends are present just by considering the codon distribution that occurs by chance (expected).

Predicting AT content at mutational equilibrium

The neutral AT equilibrium can be estimated using a simple method which involves finding the relative mutation rates of G/C to A/T and A/T to G/C. However, we also apply a more comprehensive approach also used by Rice and colleagues [132] that treats each base as an independent state and allows to determine the equilibrium frequencies of all four nucleotides while also accounting for nucleotide skews [133]. The method can be illustrated by consideration of the mononucleotide model. We for instance define G as the frequency of G and T as the frequency of T. The mutation rate from G to T is denoted as G2T, expressed per occurrence of the ancestral base. We describe frequencies for each nucleotide N after a given period (N′) and, to determine equilibrium frequencies, we solve for conditions where N′ = N, leading to the following equations:

G (1 − G2T − G2C − G2A) = A (A2G) + T (T2G) + C (C2G)

C (1 − C2T − C2G − C2A) = A (A2C) + T (T2C) + G (G2C)

A (1 − A2T − A2C − A2G) = G (G2A) + T (T2A) + C (C2A)

T (1 − T2G − T2C − T2A) = A (A2T) + G (G2T) + C (C2T)

Here, the left-hand side of each equation represents the loss rate given the current nucleotide abundance, while the right-hand side represents the gain rate at equilibrium (i.e., we solve for the state where gain = loss). The 12 flux parameters (i.e., G2T, C2G, etc.) are derived from the mutation profile, calculated as the observed number of mutations per occurrence of the ancestral nucleotide. For any given mutational matrix, we solve for four simultaneous equations, ensuring that one nucleotide frequency is determined as 1 minus the sum of the other three (e.g., T = 1 − A − C − G). The equations are solved using NumPy [134]. To resolve the ambiguity of which strand the mutation is happening on, we consider A + T and G + C expected content at equilibrium rather than the four nucleotides separately.

To estimate confidence intervals, we perform a bootstrap resampling procedure, drawing different sets of mutations with replacement from the original set of mutations. We repeat this 1,000 times. For each resampled data set, equilibrium frequencies are recalculated, allowing us to establish confidence bounds.

For the central analysis, we expand this approach and consider 16 × 16 mutational matrixes including each dinucleotide mutating to each other, rather than mononucleotides. This results in 15 simultaneous equations which we solve for. For equations and means to solve them see scripts at https://doi.org/10.5281/zenodo.17378284.

We repeat the whole process for 5′ ends (which, due to data sample limits, we take to be first 20 codons following the start codon, i.e., the lowest round number that allows analysis), gene cores (the rest of the CDS), intergenic regions (mutations with annotated positions that fall in between CDSs), as well as genomically (including everything, also non-protein-coding sequences). For completeness, we perform the analysis for both those samples in Zhang and colleagues [57] that were sequenced with a higher and lower depth (S3 Table). The other mutation data sets described above lack sufficient resolution for a by region analysis.

As a sanity check, we also perform the test through the simpler G/C<->A/T approach considering mutations that were called with different significance (that Zhang and colleagues [57] report). We divide the mutations by quartiles of mutational calling P-value (Q1 meaning lowest P-value, i.e., higher significance), and find the estimate neutral AT content for each mutation group separately (S2 Text).

We perform a further test repeating the analysis employing a separate data set by Bhawsinghka and colleagues [59], who collect spontaneous mutation data with Duplex Sequencing in WT E. coli as well as mutants defective in stress response-related mechanisms (mutL and mutT). For this data, we compute mutation equilibrium AT content using the full simultaneous equations method considering mononucleotides, as explained above.

We estimate mutational equilibrium nucleotide content through the simplest mononucleotide method for E. coli MA data [56,58], as well as E. coli spontaneous mutation data recorded via Duplex Sequencing [57,59]. For B. subtilis, using data from Sung and colleagues [62], and for M. smegmatis, combining data from Kucukyildirim and colleagues [61] and Castañeda-Garcia and colleagues [60]. See Table 1 for all AT* results computed using the simple method. For M. smegmatis merged data, we also compute AT* using the full simultaneous equations method based on mononucleotide changes (Fig 8C).

Cross-species nucleotide content by codon position

For the analysis of 5′ and core GC3 content across bacteria species, we firstly retrieved all reference CDSs for 1,355 bacterial species (see S4 Table for full list of species names and accession). Sequences were downloaded on 13th September 2023 from RefSeq NCBI [109]. These represent a selection of all the available reference genomes on RefSeq NCBI, keeping one species per genus. For each species, we then found the average GC content at codon third sites across all genes for 5′ ends (first 10 codons following the start codon) and gene cores (rest of the CDS).

Comparison between transgene GC4 content and expression

For the analysis of 5′ GC4 content and transgene expression, we used data from Cambray and colleagues [41], as above. For each transgene, we found the average GC content at third sites of 4-fold degenerate codons for the first 10 codons following the start codon. We also retrieved an expression measure as protein per RNA where PNI was used as protein measure and RNA_SS as RNA measure (see Supplementary Data 15 in Cambray and colleagues [41]).

Z score normalization

For value normalization, we use a Z score approach, i.e., (observed − mean)/standard deviation.