1 Introduction

The genomes of prokaryotes cover a broad compositional range, GC levels being approximately comprised between 25% and 75% [1, 2]. It was proposed that such range was due to a mutational bias [3, 4], a point apparently confirmed by the observation that mutator mutants of Escherichia coli could shift their GC levels [5]. Differences in GC levels were also detected between the genomes of cold- and warm-blooded vertebrates [6], previously thought to be very close in base composition. The higher GC levels of the genomes from the latter were explained as due to selection for thermodynamic stability required by DNA, RNA and proteins at their higher body temperatures [7].

At this point, two opposite explanations were competing in order to explain compositional differences in genomes: the neutralist explanation [3, 4] and the selectionist explanation [7].

The neutralist explanation was weakened by three observations: (i) that higher GC levels were found in bacteria exposed to UV light, this increase reducing the presence of TpT dinucleotides that could form dimers causing DNA replication problems [8]; (ii) that in some cases high GC levels of prokaryotes were associated with high growth temperatures [9-11] and with aerobiosis [12]; and (iii) that the compositional shifts associated with mutator mutations [5] were within the error limits of the ultracentrifugation approach used to detect them and only concerned “hot spots” [13]. Finally, when a pair of completely sequenced closely related bacteria (Corynebacterium efficiens and C. glutamicum) were compared, the positive relationship between T _opt and GC level was striking [14].

The selectionist hypothesis was (i) apparently weakened by the observation that hyperthermophiles exhibited low GC levels, a point later understood, however, as due to the avoidance of C which can be deaminated at high temperatures [15], and (ii) apparently disposed off by the investigations of Galtier and Lobry [16] and Hurst and Merchant [17] who reported no correlation between T _opt and GC (or GC₃). This was interpreted not only as a strong evidence against the thermodynamic hypothesis in prokaryotes, but also as a disproval of the same hypothesis in vertebrates [17-19].

It is important to note, however, that the papers that addressed the thermal stability hypothesis in prokaryotes [16, 17] were criticised [20] because many factors, like those quoted above, have often contrasting inputs on genome composition of such a vast array of organisms as prokaryotes, which have been diverging for at least 3.5 billion years [21].

To avoid these drawbacks, we restricted our study of co-variation between T _opt and genomic GC to the Family level for the following reasons. First, the phylogenetic relationships among prokaryotic Families are still uncertain in several cases (for a review see [22]). In contrast, the phylogenetic relationships among species within each Family are expected to be more accurate, since the times of divergence are much smaller. Second, any method aiming to estimate the correlation between T _opt and GC taking into account the phylogenetic relationships needs to infer the character states (T _opt and GC) in the internal (ancestral) nodes. Obviously, this inference would be safer for shorter times of divergence (such as within Families) than those inferences that involve nodes connecting different Families. Indeed, as we show here, the failure to detect correlations when phylogeny was taken into consideration [17] is very likely due to inaccurate inferences in deep internal nodes, which are very probably responsible for hiding the correlation at lower phylogenetic levels.

2 Materials and methods

3 Results and discussion

When T _opt is plotted against genomic GC for the 368 species belonging to the 20 Families, no trend can be detected, although there is, if any, a negative correlation between the two variables (not shown). This result is nearly identical to that reported by Galtier and Lobry [16], although these authors worked at the Genus level.

Table 1 shows the results from the 20 Families studied. Among these taxa, there are 15 which display positive trends (GC increments with T _opt), eight of these exhibiting correlation coefficients that are statistically significant. On the other hand, five Families display a negative trend, but only one shows a statistically significant correlation coefficient. The probability of obtaining such distribution of correlation coefficients by chance alone is 4.39 × 10⁻⁸. Importantly, when the analysis is extended to include those Families that have at least five members, the results remain qualitatively the same.

Two conclusions can be drawn from Table 1. First, for each Family the range of variation (Δ) for both T _opt and GC is very different. For example, ΔGC varies from 38% (Enterobacteriaceae) to 5.5% (Staphylococcaceae). ΔT _opt, on the other hand, varies from 43.5 °C (Clostridiaceae) to 5.5 °C (Acidaminococcaceae). Although some errors in the taxonomic position of certain species cannot be excluded, this variability is probably related with the time of divergence and rate of change: species which diverged from their last common ancestor more recently, and/or evolve more ‘slowly’, are expected to share more features, namely the ecological niche, T _opt, physiology, etc. This is supported by the correlation found between ΔGC and ΔT _opt: R=0.52; P=0.02.

Second, and more important, within most Families there is a link between T _opt and GC, and in the majority of cases the correlation coefficient increases (significantly in several cases) with T _opt. To sum up, we found that in 15/20 of prokaryotic Families the two variables are positively correlated (eight of them with P⩽0.05). Three examples of these correlations are displayed in Fig. 1A–C .

Although these results are clear and suggest that T _opt is a factor influencing genomic GC, we cannot rule out the effect of phylogenetic inertia (the fact that closely related species are likely to have similar GC levels), so we used the method of comparative analysis by independent contrasts [27]. This method allows us to see if the GC level shows a correlated response with the adaptation to a new thermal environment. This analysis was carried out on the species belonging to the Families listed in Table 1 for which the 16S RNA were available [24]. Our analysis found that for 17 out of 20 Families there is a positive relation between the two variables, four of them significant and two at the limit of significance. The plots for the same Families shown in Fig. 1 are presented in Fig. 2 . On the other hand, for three Families the relation was negative, yet for none of them it was significant. Following the procedure described above for the direct analysis, the probability of getting these results by chance alone is <0.01. The R values within each Family of both analyses (controlling and not controlling for phylogenetic inertia) are significantly correlated (R=0.85, P<0.0001). As can be seen in Table 1, besides the four Families for which there is a significant correlation, there are two more at the limit of significance: Flexibacteriaceae and Pseudomonadaceae. Since our hypothesis is that GC level not only changes but can increase with T _opt, it seems reasonable to apply a one-tailed test. By doing so, we found six Families displaying positive significant correlations coefficients between contrasts, while no negative correlation coefficient was statistically significant (see Table 1). The overall probability of getting by chance this group of correlations with the corresponding significance levels drops to 3.59 × 10⁻⁴.

In addition, when all independent contrasts from different Families (within each taxa) are considered together, they exhibit a positive and significant correlation coefficient (Fig. 3 ; R=0.27, P<0.0001). Moreover, the increment in T _opt was accompanied by an increment in GC in 129 independent contrasts, while 79 contrasts exhibited the opposite behaviour and 76 displayed no changes. If the two parameters were not related, the probability of obtaining this excess of double increments by chance alone is very low (P<0.001, sign test).

In conclusion, we found that T _opt and genomic GC are non-independent. In the first place, we have shown that when these two parameters are compared at the Family level they exhibit positive relations in most Families, being statistically significant in several of them. These correlations still hold when the internal phylogenetic relationships are considered. Moreover, when all Families are considered together (but excluding inter-Family comparisons) there is again a significant positive correlation between T _opt and GC. We would like to stress that the positive correlation becomes evident only when inter-Families comparisons (that are less accurate from many points of view, see Section 1) are excluded from the analysis. It is also safe to suppose that when the intra-Families comparison is performed, many variables that could affect the GC level are likely to be more similar.

Finally, we should remark that these results show not only the influence of T _opt on genomic GC in prokaryotes, but also that it is not the only one influencing genome composition, as expected from other investigations [8, 12, 30, 31]. Only when a factor becomes predominant, its effect on GC can be clearly seen. Needless to say, the results obtained in these investigations strongly support the idea that base composition is under selection in prokaryotes.