Thanks to the availability of an increasing number of complete bacterial genomes with a high G+C content, there was a representative distribution of G+C content in the dataset (Table 1), and the values of G+C content in third codon positions (P₃, as previously defined in [11]; see also Materials and methods), in first and second codon positions averaged (P₁₂), and in intergenic spaces (GC_IGR), were found to be consistent with previously known results [10,11] both in terms of their range of distribution and the correlation between them (Figure 4). Intragenomic P₃ distributions of all species examined (all histograms are given in the tables in Additional data files under the 'GCorder' flag) are unimodal and narrow for a major class of genes comprising more than 90% of the total genes, although average P₃ values differ widely between species. Some species show scattering of P₃ distribution. In particular, a previous study of 254 genes in Pseudomonas [7] revealed a sizeable minor class of genes (28%) forming a trail toward lower G+C content (genes with P₃ < 0.75). In our present study of 5,255 genes, the minor class of genes (with P₃ < 0.75) is only 8% of the total, indicating that the previous sample was biased towards the minor class. Moreover, the width of the distribution (standard deviation) of the majority of genes is narrower in species with both extreme ranges of P₃ than in those with middle ranges. These features of the P₃ distribution are expected from the theory of bidirectional mutation rates and their equilibrium [8]. In this theory, mutation rates, u (GC → AT) and v (AT → GC), will lead to a G+C content (P₃) at equilibrium of v/(u+v). The expected standard deviation is

where b is an average of the number of third codon positions per gene [8].

Statistically highly significant differences (p < 0.01) in P₃ between coding sequences located on the leading and lagging strands were found in 23 out of 43 chromosomes. In most cases (21 out of the 23) genes on the leading strand were found to have a lower P₃ content than those on the lagging one. (Two exceptions were Mycoplasma genitalium, which is known to have a peculiar behavior with respect to G+C content [5,6], and Caulobacter crescentus.) As far as we know, this is the first time that such a systematic difference has been reported, but note that the differences were always very small. The maximum was only 0.035, for the difference in P₃ content of leading and lagging strands of Vibrio cholerae chromosome 2, as compared to the 0.768 average P₃ difference between Halobacterium sp. and Ureaplasma urealyticum, or the 0.050 accepted range of genomic G+C content polymorphism [57]. There was no significant correlation between the average P₃ and the difference in P₃ between the two groups of genes. A highly significant difference in G+C content between the leading and the lagging sequences was found in only five chromosomes for large intergenic regions, and three chromosomes for small intergenic spaces.

Out of 43 chromosomes, the differences in PR2-biases between the leading and lagging groups were highly significant (p < 0.01) in 39 chromosomes for G₃/(G₃+C₃) and in 34 for A₃/(A₃+T₃) and in 32 chromosomes for simultaneously G₃/(G₃+C₃) and A₃/(A₃+T₃) (see Additional data files). In the PR2-bias-plot analysis, we found a general pattern of x₁ >x₂ and y₁ <y₂, as in Figure 2b, meaning that coding sequences in the leading strand had higher G₃/(G₃+C₃) values and lower A₃/(A₃+T₃) values than those on the lagging strand: leading coding sequences were enriched in keto-bases (G and T) in the third codon position compared to lagging coding sequences. There were, however, exceptions to this general trend: out of the 39 chromosomes that were highly significant for G₃/(G₃+C₃), all had higher values in the leading strand, but out of the 34 chromosomes that were highly significant for A₃/(A₃+T₃), three chromosomes did not follow the general trend (Lactococcus lactis and Staphylococcus aureus strains Mu50 and N315). Leading and lagging coding sequences are separated in PR2 plots as expected under replication-associated effects, and the leading group is almost always down right of the lagging group. However, the extent to which the groups are separated differs between species. The two most extreme species are B. burgdorferi, where the two groups of genes are completely resolved, and M. genitalium, where the two groups are almost completely overlapping (Figure 5). Other species with a spectacular difference in PR2 plots were Chlamydia muridarum, Chlamydia trachomatis and Treponema pallidum. Out of 43 chromosomes, the residual bias B_II was in most cases oriented as in Figure 2c, with x_c < 0.5 (all except Pyrococcus abyssi and T. pallidum) and y_c < 0.5 (all except Thermotoga maritima). The relative contribution of replication-coupled bias (B_I) to the total bias (B_I plus B_II) ranged from 6% (Mycoplasma pulmonis) to 86% (B. burgdorferi) with an average value of 52%.

Out of 43 chromosomes, the differences in PR2-biases between the leading and lagging groups were highly significant (p < 0.01) in 36 chromosomes for G/(G+C) and in 34 for A/(A+T) and in 32 chromosomes for simultaneously G/(G+C) and A/(A+T) (see table in Additional data files). All the 36 chromosomes highly significant for G/(G+C) had higher values in the leading strand (that is, x₁ >x₂ as in coding sequences). Out of the 34 chromosomes highly significant for A/(A+T), 28 had lower values in the leading strand (that is, y₁ <y₂ as in coding sequences). Exceptions were Bacillus subtilis, Bacillus halodurans, L. lactis, S. aureus strains Mu50 and N315, and Streptococcus pyogenes. The distribution of intergenic regions in a PR2-plot was as expected under replication-associated effects (Figure 3a), and is particularly visible in the case of B. burgdorferi (Figure 5c). Unlike the third codon position, the leading and lagging groups clustered symmetrically around the center point of the PR2-bias plot. In intergenic regions, both x_c and y_c were always near 0.5, so that the residual bias B_II was always close to zero. The relative contribution of replication-coupled bias (B_I) to the total bias (B_I plus B_II) was therefore very high, on average 90%, and ranged from 88% (Helicobacter pylori) to 99.6% (B. subtilis) in the intergenic regions of species with significant B_I values.

Out of 43 chromosomes, the differences in PR2-biases between the leading and lagging small intergenic spaces among co-oriented genes were highly significant (p < 0.01) in 26 chromosomes for G/(G+C) and in 7 chromosomes for A/(A+T) and in 7 chromosomes for both G/(G+C) and A/(A+T) together (see table in Additional data files). All the 26 chromosomes with highly significant difference in G/(G+C) showed the same pattern as in third codon positions in coding sequences with x₁ >x₂. Of the seven chromosomes highly significant for A/(A+T), all showed the same pattern as in third codon positions in coding sequences with y₁ <y₂, except for B. subtilis and S. pyogenes. Out of 43 chromosomes, the orientation of the residual bias B_II showed no clear tendency (only 31 chromosomes had x_c > 0.5 and 23 y_c > 0.5; the trend, if any, would be in the opposite direction, as in third codon positions). The relative contribution of replication-coupled bias (B_I) to the total bias (B_I plus B_II) ranged from 10% (H. pylori strain 26695) to 93% (Thermoplasma acidophilum) with an average value of 58%.

The differences between PR2-biases in intergenic regions between the leading and lagging strands (B_I) were significantly and highly correlated with the PR2-biases in the third codon position as is evident from the regression coefficient of approximately 0.6 and correlation coefficient (r² = 0.77) among 43 chromosomes (Figure 6). The correlation was still significant when the extreme B. burgdorferi was removed from analysis (r² = 0.68). The 95% confidence interval for the value of the slope of the regression line [0.50-0.71] is less than 1, meaning that replication-associated biases were significantly smaller in intergenic regions than in third codon positions. This slope value was very close to the one obtained for the regression line between G+C content in intergenic regions versus third codon positions (Figure 4).

If differences in the G+C content between species were dictated by asymmetric replication-associated mutation pressure, one would expect a correlation between the extent of PR2-bias and G+C content; however, no significant correlation between B_I in coding sequences and P₃ were found (Figure 7).

A general feature observed in the present study is that, whenever there is a difference in base composition between the leading and the lagging strand, the G and T bases are enriched in the leading strand, and the G+C content is somewhat lower in the leading strand (see Results). The direction of observed biases are universal in bacteria [28], although not every species is biased [58]. This pattern is not restricted to the bacterial world, as GT-enriched leading strands have been documented in mitochondria [59,60,61], in the chloroplast genome of Euglena gracilis [62], in viruses [45,46,47], and in the telomeric ends of the Saccharomyces cerevisiae genome [63]. Not all sequences are biased, as shown by the well-documented β-globin region in the Homo sapiens genome [37]. Remarkably, absolutely no exceptions to the relative enrichment in G of the leading strand were found here, either in third codon positions or intergenic positions. However, the sole general rule in biology is that there are no general rules, and during the time of revision of this manuscript the first exception was documented in the chromosome of Streptomyces coelicolor [64]. The relative enrichment in T of the leading strand suffers from some exceptions that, interestingly, are all found here in the low-G+C group of Gram-positive bacteria (namely B. subtilis, B. halodurans, L. lactis, S. aureus, and S. pyogenes).

Because of the strong correlation in direction and intensity of PR2 deviation between third codon positions and intergenic spaces (Figure 6), the most likely explanation for the GT enrichment of the leading strand is a replication-induced asymmetric mutation pressure between the leading and lagging strand.

The universal excess of G and T in the leading strand is common to distantly related species, suggesting that asymmetric directional mutation pressure has a chemical basis. This does not mean that the mechanism is simple: there are 12 substitution rates under asymmetric molecular evolution, but only four base frequencies; many different substitution scenarios can explain the observed biases. Here we discuss only simple hypotheses in which only one of the 12 substitution rates is assumed to differ between the two strands: for example, only an excess of A → G or of C → T in the leading strand is able to produce a GT-rich leading strand. Because the leading strand spends more time in the single-stranded state than the lagging strand during DNA replication [65], it is natural to look for mutations that are increased in the single-stranded state as possible explanations.

From inspection of our data, we can reject the simple model in which a unique asymmetric mutation pressure causes variation in G+C content between species. For instance, if only C → T mutations were the underlying phenomenon, we would expect low G+C in species with the highest B_I. Figure 7 shows that this is clearly not the case. Moreover, if asymmetric directional mutation pressures were the major origin of variation in G+C content between species, we would expect a much more important difference in term of G+C content between the genes on the leading strand and those on the lagging strand.

If we exclude as an explanation that many unannotated genes remain to be discovered, or that the boundaries of annotated genes should be greatly extended, so that some intergenic regions under analysis are not actually intergenic, then, there are two non-exclusive interpretations of the fact that B_I is smaller in intergenic regions than in third codon positions. On the one hand, one may postulate that intergenic regions are under weaker selection pressure against inversions, so that deviations from PR2 are cancelled. On the other hand, one may assume that some positions are not free to deviate from PR2 because of selective pressure for some function. For instance, there are 100 copies per enterobacterial genome of palindromic sequences in intergenic spaces [80], which may be under selective pressure to preserve their palindromic character and therefore follow PR2 (as pure palindromic sequences are effectively base-paired). This interpretation is compatible with the fact that interspecific G+C content variation is higher at the third codon position than in intergenic spaces [10,11], as shown here in Figure 4. Moreover, the surprisingly low level of polymorphism in large intergenic regions among natural strains of E. coli [81] is in favor of selective constraints.

Francino and Ochman [22,81] pointed out that transcription is likely to cause more mutations in the sense strand than in the anti-sense strand because of its transient single-stranded state during transcription. We found no evidence for this in intergenic spaces (> 100 bp), which could be interpreted either as an absence of transcription-induced mutation pressure or by a low proportion of transcribed fragments in these regions. Third codon positions are known to be transcribed, so that we expect an effect of transcription-induced mutation pressure on the residual deviation from PR2 (B_II). However, selection for translation optimization interferes, so that we cannot interpret unambiguously the almost universal orientation of B_II (down left in PR2-plot, as in Figure 2c) as the result of transcription-induced mutation pressure. If we interpret results this way, we would have to postulate a different mechanism for replication-induced and transcription-induced mutation pressures, as B_I and B_II have different orientations. Moreover, in small intergenic spaces, which are potentially transcribed untranslated regions, no general pattern is evident for the residual deviation from PR2, so that no universal transcription-induced mutation pressure was found here as a strong contributor to the deviation from the PR2 state.

A chirochore is a contiguous segment of the chromosome homogeneous with respect to its deviation from PR2 [44]. A strong asymmetric mutation pressure induces a chirochore structure, as observed in B. burgdorferi [82], but the converse need not to be true: a weak asymmetric mutation pressure does not imply an absence of chirochore. For example in M. genitalium, replication-associated biases, B_I/(B_I + B_II), are the weakest among species under study (B_I = 12%), but there is still a strong chirochore structure [83] because of a large excess (80%) of coding sequences in the leading strand. The deviation from PR2 in the first and second codon positions, because of selection at the amino-acid level (for example [84]), is then a major factor. It is not known whether chirochores could be adaptative, but recent results suggest a functional polarization of the E. coli chromosome [85,86] and may be relevant to this problem.

A prerequisite for using the simple graphical representations and quantifications used here is that the coding sequences and intergenic spaces can be divided into two groups: the leading and the lagging group. This is not necessarily always possible when, for instance, all coding sequences are in the leading group (this situation has never been encountered in bacteria but could be problematic in small genomes such as those of mitochondria). Another limitation is encountered when there are genes in both groups but we are unable to assign sequences to the correct group because of limited information about the location of origin and terminus of replication used during the course of evolution. Although an estimation of B_I is not possible in this latter case, an oblong distribution of points in PR2-plots may suggest a non-zero value for B_I. When more than one strain was sequenced for a species (that is, Chlamydophila pneumoniae, E. coli, H. pylori, Mycobacterium tuberculosis, Neisseria meningitidis, and S. aureus) B_I and B_II values in third codon position are extremely stable within a species (see table in the Additional data files), but the maximum number of three strains available here for one species does not allow us to estimate their within-species variability, and therefore their taxonomic utility.

Data on 51 complete bacterial chromosomes were retrieved from the EMGLib server [87,88] and used to extract annotated coding sequences and intergenic spaces (Figure 1). These 51 chromosomes were from 49 different strains (there are two chromosomes in Deinococcus radiodurans and Vibrio cholerae), from 41 different species (there were three strains of C. pneumoniae, E. coli; two strains of H. pylori, M. tuberculosis, N. meningitidis, S. aureus) as detailed in Table 1.

To classify sequences into the leading and the lagging groups (Figure 1), we must be able to locate them relative to the origin and terminus of replication. Out of 51 bacterial chromosomes, the locations of the origin and the terminus were known from experimental evidence or were estimated from chirochore boundaries [44] in 43 chromosomes with the program Oriloc [89,90]. Details of assignments are given in EMGLib database under a '/strand=' qualifier. To discriminate between the origin and the terminus of replication, we used the assumption that organisms tend to avoid head-on collisions between the DNA replication apparatus and the RNA polymerase transcription complex [54,55,56]. We also assumed that coding sequences were more abundant on the leading than on the lagging strand. As far as we know, there is no exception to this rule, and the report that the trend is exacerbated in highly expressed genes supports this assumption [48]. Moreover, the comparison of the maps of closely related bacterial chromosomes shows that observed chromosome rearrangements tend to preserve an excess of genes in the leading strand [91,92,93,94,95,96,97,98]. Our figures show sequences on the leading strand as filled circles, and those on the lagging strand as open circles.

The nucleotide composition of coding sequences was always calculated from the sense strand (Figure 1) for each codon position. The sense strand (or anti-template strand) is the DNA strand opposite to the template strand for transcription. The average G+C content of the third codon position (P₃) and that of the first and second codon positions (P₁₂) were computed to draw neutrality plots relative to P₃ as previously described [11]: stop codons and codons from odd-numbered sets ATG (Met), TGG (Trp) and ATA (Ile) were excluded from the calculations of P-values below to avoid an extra cause of potential deviation from PR2. The coding sequences after position 1,950 kb on D. radiodurans chromosome 1 were discarded because of a frameshift error in the annotations. Some of the apparent coding sequences identified in genome sequences could be artifacts that do not correspond to actually expressible genes [99]. To overcome this problem, coding sequences of less than 300 bp were removed from the analysis (less than 10% of sequences were discarded). As the probability of obtaining an open reading frame with more than 300 bp by chance is low, most (> 99%) sequences analyzed should correspond to actual coding sequences [100].

GC_IGR denotes the G+C content, and deviations from PR2 were characterized by G/(G+C) and A/(A+T) to draw PR2-plots in a way similar to coding-sequence analyses. Intergenic sequences were split into two non-overlapping groups (Figure 1). The first group contained small intergenic sequences among co-oriented genes that are potentially transcribed untranslated spacers. The strand used to compute nucleotide composition - chosen to allow direct comparisons with coding-sequence analyses - is the same as the anti-template strand for transcription of the two flanking genes (CDS, tRNA, rRNA). Therefore, when the two flanking genes are coding sequences, the strand is the same as the strand used for the analysis of coding sequences, that is the sense strand (Figure 1). Their minimum size was set to 50 bp, a trade-off between the likelihood of being transcribed and the avoidance of too much scattering in PR2-plots. Their maximum size was set to 100 bp to avoid data overlap with large intergenic regions. The second group contained large intergenic regions that are potentially untranscribed untranslated regions. Their minimum size was set to 100 bp, which is an amount of information equivalent to the third codon position of 100 codons, the minium size required in coding-sequence analyses. The analyzed strand cannot be defined unambiguously with respect to the flanking genes when they are convergently or divergently transcribed. The nucleotide composition was therefore always calculated from the published strand (Figure 1).

The significance of the analyses was assessed using unpaired t-test implemented in Statview 5.0 (SAS Institute Inc., North Carolina, USA) to compare the mean between two groups - here the leading and the lagging group. The null hypothesis is the equality of the mean value between the two groups. The critical value p is the probability of taking the wrong decision of rejecting the null hypothesis when the null hypothesis is true. Two notes of caution should be introduced about the interpretation of results in the present context. First, we made runs of 43 comparisons so that the odds of getting at least one wrong significant result, 1 - (1 - p)⁴³, is 0.89, 0.35, 0.04 for p = 0.05, p = 0.01, p = 0.001 respectively. We selected the highly significant critical level (p < 0.01) so that odds of getting one, two, or three wrong decisions were 0.28, 0.06, and 0.01, respectively, for each run. Therefore, a simple figure to keep in mind when looking at the results is that the odds of getting by chance three or more highly significant differences in a run were less than 1 in 100. The second note of caution is that we were working with large data sets, typically one thousand values within each group, so that the power of the test, that is its ability to reject the null hypothesis, was high. As a matter of consequence, even a small difference between the mean of the two groups was considered, correctly from the statistical test point of view, as significant. However, statistically significant does not mean automatically biologically significant because a second-order or a side effect could easily be evidenced this way. To help interpretation, the simple graphical representations used here, very close to raw data, were connected through hyperlinks to the tabulated test results. By this means, readers can quickly form their own opinions by viewing the original data.