The number of protein coding genes per genome within the various strains and species of Streptococcus is relatively similar (ranging from 1,697 to 2,376; Table 1), but the gene composition of these genomes is much more variable. Based on the gene content table obtained by OrthoMCL (Additional data file 1), three strains of S. agalactiae, S. pyogenes or S. thermophilus share about 75% of their genes, and three different species of Streptococcus share only around half of their genes (Figure 1). This latter result appears to be independent of the particular strains or species involved in the comparison and of their phylogenetic affinities. Even with the inclusion of 26 genomes, the total number of possible genes - the pan-genome - of Streptococcus appears not to have been reached, as depicted in the gene accumulation curve (Figure 2), and we estimate the Streptococcus pan-genome probably surpasses 6,000 genes. A surprising 21% of the genes in the pan-genome of the genus Streptococcus (based on these 26 genome sequences), were represented in only one lineage, suggesting a remarkable degree of lateral gene transfer in shaping the genomes of each of these taxa (Figure 3). Within species, the pan-genome size also remains uncertain, although our estimates suggest that the pan-genome size of S. pyogenes is smaller, and better estimated with the currently available data, than that of S. agalactiae (Figure 2).

In contrast to the pan-genome estimates, the number of genes in common between the different species within the genus Streptococcus - the core-genome - appears to reach a plateau around 600 genes (Figures 2 and 3). Next to the genome specific genes and the genes shared by only two genomes, the genes of the core-genome were the third most common genes (11%; Figure 3), suggesting they form a coherent group. Similarly, the estimated core-genome for S. pyogenes, based on the 11 available strains, plateaus around 1,400 genes. The pattern was less clear for S. agalactiae, where the estimate of core-genome size does not level out, and appears as though it might still be influenced by the inclusion of new genome sequences. On the whole, these analyses suggest that it is possible to delineate a core-genome at both genus and species level. We analyzed four such core-genome data sets: the Streptococcus core-genome (611 genes), and the core-genomes of S. agalactiae (1,472 genes), S. pyogenes (1,376 genes) and S. thermophilus (1,487 genes). To save computation time, the Streptococcus core-genome data set was reduced to ten taxa by keeping only two strains per species for S. agalactiae, S. pyogenes, S. thermophilus (strains A909 and NEM316, MGAS9429 and M1 GAS, and CNRZ1066 and LMG 18311, respectively). After discarding clusters of genes containing paralogs (that is, clusters containing more than one gene per taxon), and alignments with uncertain site homologies, we obtained four data sets containing 260, 1,297, 1,212 and 1,365 genes representing the alignable core-genomes of Streptococcus, S. pyogenes, S. agalactiae, and S. thermophilus, respectively.

Determinations of the number of genes gained and lost on each of the lineages shows considerable variation (Figure 4) and, in agreement with earlier studies, gene gain was generally considerably greater than gene loss, as well as being particularly evident on external branches 23. The lineage in the interspecific analysis showing the greatest gene gain was S. suis, followed closely by S. pneumoniae and S. mutans. Even within a species, between strains, the numbers of genes gained and lost were very high, reaching, for example, values in excess of 150 for gene gain in S. agalactiae strain H36B. High levels of gene gain and loss were evident, even for closely related isolates of the same serotype in S. pyogenes (for example, M1 GAS/MGAS5005; SSI-1/MGAS315; MGAS9429/MGAS2096). Branch lengths of the S. pyogenes concatenated tree were much longer than those for S. agalactiae, suggesting the lineages might be much older; however, despite this there was generally more gene gain on the S. agalactiae branches than on S. pyogenes branches. Large values for duplications were also a feature of the lineage specific evolution (Figure 4). Phylogenetic analysis of several of these cases suggests this is a combination of lineage specific duplications as well as LGT events involving homologous sequences from other species of Streptococcus. When gene gain was penalized with respect to gene loss (for example, 24), not surprisingly, it globally decreased the number of gene gains and increased the number of gene losses (Additional data file 3) and, as a consequence, increased the number of genes in the pan-genomes of ancestral nodes (data not shown). Nevertheless, even with a penalty, gene gain remained in excess of gene loss on some lineages (Additional data file 3).

The number of genes that showed evidence for positive selection was particularly high within the Streptococcus core-genome (between 10% and 40%; Table 3). The S. pneumoniae and S. suis lineages, and the ancestral lineage leading to these two species, exhibited the greatest proportion of the core-genome evolving under positive selection (28%, 34% and 32%, respectively; Table 3). Approximately one-third of the genes showed positive selection on only one lineage, and no gene was selected in all possible lineages (Figure 8). There were, however, many examples of genes selected on multiple lineages, including several genes selected on as many as 5 (12 genes) or 6 (4 genes) different lineages (Figures 8 and 9; see Additional data file 5 for a complete list of all genes and lineages under positive selection). A significant proportion of positively selected genes for S. suis, S. pneumoniae, and S. thermophilus was uniquely selected on each of these lineages (21%, 19%, and 24%, respectively), in contrast to that for S. agalactiae, S. pyogenes, and S. mutans, which had either no uniquely selected loci (S. agalactiae), or a very small proportion (Figure 9). Analysis of variance of genes under positive selection pressure supported a significant effect of both lineage and biochemical main role category (Table 4). Post hoc multiple comparisons showed that the main effect was due to two categories, 'DNA metabolism' and 'Transcription'. Less strongly supported, but still significant, was the interaction between lineages and main role categories (Table 4). This interaction appeared mainly due to an increase of genes under positive selection for loci involved in transcription, protein fate, protein synthesis and DNA metabolism for the S. pneumoniae-S. suis ancestral lineage and the S. suis lineage.

In addition to identifying genes and lineages under positive selection, the branch-site test also identifies sites using a Bayes empirical Bayes approach 25. For 91% of the genes under positive selection, specific sites were proposed (posterior probability >0.95). Interestingly, when a gene was independently selected on different lineages, the sites under positive selection were generally not the same across lineages, arguing for different selection pressure located at different sites. In contrast to the interspecific comparisons, positive selection was evident for only a few genes within the core-genome, across strains of the different Streptococcus species (Table 3, Additional data file 5), including a few lineages that showed slightly increased levels of positive selection relative to the rest. For S. agalactiae the exceptional lineage was COH1, for S. pyogenes the exceptional lineages were MGAS10270 and that leading to SSI-1/MGAS315, and for S. thermophilus it was LMD-9. A significant number of genes evolving under positive selection were also judged as putative recombinants (Table 5). This was particularly true for the S. pyogenes genome, where 78% of the genes under positive selection were putative recombinants. Approximately half of these genes were identified as recombinants by the substitution based recombination methods, and the other half by the phylogenetic approach.

We estimate that the pan-genome of the genus Streptococcus probably exceeds at least three times the average genome size of a typical Streptococcus species. This huge variability in gene content between species is also evident in comparisons across strains of the same species. Our prediction for the S. agalactiae pan-genome is in general agreement with that of Tettelin et al. 22. The marked difference in estimated pan-genome size for these two species may be a reflection of their habitat differences. The human oral-nasal mucosa is the primary habitat for S. pyogenes, whereas S. agalactiae was first identified as a bacteria linked to bovine mastitis, and later in humans, where it colonizes the lower gastrointestinal tract and vaginal epithelium of healthy adults. This apparent broader habitat range for S. agalactiae, and presumably, therefore, a greater available gene pool for lateral gene transfer, could explain the difference in pan-genome size of these two species.

The pronounced evolutionary flexibility of these bacterial genomes is further evident in the determinations of gene gain, loss and duplication on each of the respective lineages. Gene gain figures were generally higher for S. agalactiae than for S. pyogenes, despite the fact that branch lengths suggest the S. pyogenes lineages may be older, and is likely a consequence of the overall smaller pan-genome size for S. pyogenes. For some species, gene gain figures exceeded 20% of the total gene content for the organism. Our results in this regard are in general agreement with those of Hao and Golding 23, while also extending the estimates to additional taxa of Streptococcus, and lineages of S. agalactiae and S. pyogenes, and we would certainly concur with these authors that much of this gene gain likely reflects species specific adaptation. In our opinion, a plausible explanation of the discrepancy between gene gain and loss is that much of the pan-genome remains unsampled, and, therefore, we simply cannot detect many gene loss events, resulting in an underestimate of that category. In several genome reconstruction analyses, gene gain is penalized with respect to gene loss (for example, 24262728). This procedure logically results in an increase of gene loss relative to gene gain. It has the direct consequence of increasing ancestral genome sizes, and reducing the importance of LGT events to the benefit of genome reduction processes. Nevertheless, this approach is questionable when one attempts to reconstruct genome composition, and in particular, for the case of Streptococcus. First, because the genome size is relatively stable in Streptococcus, the ancestral genome sizes were arguably of the same order. Second, the number of genome specific genes is so common that there is little reason to postulate that gene gain is less probable than gene loss. Therefore, at least in the case of Streptococcus, we question the validity of an approach that favors genome reduction processes over LGT. In that regard, the development of probabilistic models will be highly valuable to estimate rates of gene gain and loss (for example, 1629).

In contrast to the pan-genome, the core-genome size of the taxonomic groups included in our analysis are much better estimated. Within Streptococcus, we characterized several core-genomes, the composition of which depended on the taxonomic level considered. Our prediction of core-genome composition for S. agalactiae was in general agreement with Tettelin et al. 22 and Brochet et al. 30. The slight differences in the absolute numbers between the three studies are due to differences in methodology used to define orthology 22, or the use of DNA microarray hybridization data 30. At the genus level, the core-genome corresponded to 25% of a typical Streptococcus genome, while at the species level it represented around 60% of the genome. Earlier studies involving other groups of bacteria have suggested that such core-genomes may be relatively free of recombination 313233. If you consider the union of both substitution based methods and phylogenetic based methods we estimate that around 18% of the core genome of the genus Streptococcus is recombinant and as much as 35% of the genome of S. pyogenes. In addition to the fact that we are analyzing a different group of taxa, and thus levels of recombination might well be expected to be different, our results differ from these earlier estimates, also because of approach. We concur with the comments of Susko et al. 34 that attempts to evaluate phylogenetic congruence of core-genes need to involve comparisons with as many relevant topologies as possible, and not just that favored by the concatenated topology, some variants of it, or particular canonical markers such as the small subunit rRNA. In doing this, however, we think it is important to keep in mind that topologies may be rejected as being congruent, even though the genes may provide little phylogenetic signal, and, thus, only that proportion of the gene trees that is rejected based on strongly supported conflicting nodes should be regarded as incongruent. This was particularly evident in our analysis of S. pyogenes, where the vast majority of topologies reject one another despite the fact that at least half of these genes have little or no phylogenetic signal. Furthermore, our assessment of core genome recombination also differs from some of these earlier studies by the inclusion of substitution based approaches to recombination detection. There was little overlap in the loci identified as putative recombinants using both the phylogenetic and the substitution based approaches, suggesting they were identifying very different types of recombination. The substitution based approaches are likely to detect homologous recombination of smaller pieces of DNA that could be missed by a phylogenetic approach. The more restricted habitat distribution for S. pyogenes may also be an explanation for the elevated amounts of recombination in the core-genome of that species. A more reduced gene pool for possible recombination would not only result in a smaller pan-genome size (as suggested above for S. pyogenes), but it would also result in the propensity for more homologous recombination of core-genome components, at least partly because the relative proportion of conspecific donor pieces of DNA are likely to be greater.

A logical surmise often made in studying pathogen evolution is that much of the host specific adaptation that a bacterial species exhibits will be associated with its species specific genes. Perhaps as a consequence of this common sense viewpoint, adaptation within the core genome has received much less attention. Our analysis reveals that during the diversification of the genus Streptococcus there has been significant amounts of positive selection pressure on core genome components and that this selection pressure has occurred disproportionately in certain lineages, and biochemical categories. Such an important positive selection signature within the core-genome of Streptococcus is perhaps somewhat surprising, as these predominately housekeeping genes might be expected to evolve under strong purifying selection. At the same time, we would argue that many of these genes are undoubtedly related to the colonization, persistence, survival, and propensity to cause disease in these organisms and thus are associated with the adaptive specifics of the bacteria.

Based on currently available phylogenies for the genus Streptococcus 35, there appears to be a rough correlation with the relative divergence of taxa and the level of positive selection detected in different lineages. For example, S. pyogenes and S. agalactiae are both members of the Pyogenes taxonomic group and there are fewer numbers of genes selected on these lineages than for taxa from different taxonomic groups. Several genomes also tend to resemble one another in relation to the genes that were positively selected, while others, such as S. suis and S. thermophilus, exhibit higher levels of specific adaptation. S. suis was also the species with the largest number of positively selected genes in its core-genome, relative to the other lineages, and the genome that had the greatest amount of gene gain and loss incurred since the separation of the S. pneumoniae-S. suis common ancestor. This suggests a lineage that has been under strong selection pressure, both with regard to acquiring new genes and with regard to the sequence characteristics of the core genome components. This selection pressure is undoubtedly correlated with particular characteristics of the host species, swine. Current phylogenies support S. acidominimus as the sister group to S. suis 35, a species associated with bovine vagina, the skin of calves, and raw milk, suggesting the possibility of host divergence in one or another (or possibly both) lineage, subsequent to their split from a common ancestor. S. suis is also known as an occasional zoonotic pathogen of humans, causing septicemia, meningitis, endocarditis 36, and, most recently, streptococcal toxic shock syndrome 15. Our analysis suggests an apparent evolutionary flexibility of the S. suis genome that could perhaps be related to this propensity for host jumping.

In addition to the lineage effect in the interspecific selection analysis, there was also an effect of biochemical roles, with DNA metabolism and transcription significantly different from the majority of the other categories and correlated with higher incidence of positive selection. An excess of positive selection in the genes related to transcription is perhaps surprising, as these genes are generally well conserved and are known to be particularly recalcitrant to recombination. Furthermore, the S. suis and S. pneumoniae-S. suis ancestral lineages showed a disproportionate amount of positively selected genes for several biochemical categories. Thus, the extent of adaptive molecular evolution varies across lineages and gene roles, undoubtedly reflecting the habitat heterogeneity of the genus Streptococcus. Our results also reveal that positive selection was partly linked to recombination. Recombination can create artifactual positive selection results 37, particularly for genes showing evidence of intragenic recombination. This is less probable for genes identified as recombinant using the phylogenetic recombination detection procedure, because these loci were analyzed for positive selection with their own phylogenies, thereby taking into account possible LGT events. It is possible that many LGT loci could also be positively selected because of strong selective forces that might be expected to act on newly acquired genes. Marri et al. 16 found that many species specific genes within the Streptococcus genus showed evidence of adaptive evolution, and concluded that LGT played an important adaptive role.

Although a good deal of positive selection pressure was evident in the analysis of different lineages within the genus Streptococcus, there was much less evidence for positive selection between the different lineages of S. agalactiae, S. pyogenes, and S. thermophilus. Nevertheless, these absolute numbers of genes under positive selection are not corrected for depth of ancestry. In other words, they do not reflect rates of adaptative mutations for specific lineages; instead, they represent the core-genome fraction that has participated in the adaptation of a specific taxon. There were, however, several lineages in each of these species that had slightly elevated levels of selection relative to the rest. In S. pyogenes, for example, the lineage leading to the M3 serotype had nine genes under positive selection, while the majority of other lineages had two or less (see Additional data file 5 for a complete description of these genes). Compared to other M types, serotype M3 strains cause more cases of invasive disease, such as necrotizing fasciitis, bacteremia, and streptococcal toxic shock syndrome 3839404142, a higher rate of lethal infections 414243, and exhibit occasional epidemic tendency 38. The nine genes we identified as positively selected included loci implicated as virulence determinants in other species of Streptococcus, such as adenylosuccinate lyase, a homotetramer that catalyzes two discrete reactions in the de novo synthesis of purines, and has recently been implicated as a virulence factor for infective endocarditis, a serious endovascular infection caused by Streptococcus sanguinis 44, as well as genes involved in cell envelope, such as UDP-N-acetylmuramoylalanyl-D-glutamyl-2,6-diaminopimelate-D-alanyl-D-alanyl ligase, demonstrated to be integral to peptidoglycan biosynthesis and cell growth 45. In addition to the evaluation of genes unique to M3 strains (for example, 46), we suggest that these nine core-genome loci uniquely under positive selection pressure in this M3 lineage should be considered as putatively important in the unique pathogenic features of M3 strains.

In the case of S. agalactiae, the lineage that stood out from the rest with regard to levels of positive selection pressure was COH1, which is serotype III, ST17, significantly associated with neonatal invasive disease 47, and is hypothesized to have recently arisen from a bovine ancestor 48. The seven genes selected on this lineage include a number of loci either already implicated in virulence in other bacteria, or for which there is some reason to suspect them as candidate virulence loci (see Additional data file 5 for a full description of these genes). For example, adenylosuccinate lyase, discussed above with regard to S. pyogenes, was also positively selected in this lineage. Phosphate acetyltransferase was uniquely selected along this COH1 lineage, and has recently been implicated in virulence in Salmonella enterica 49. Also selected were a protein involved in the cell envelope, as well as two different ABC transporters. The cell envelope is a key overall component of virulence in S. agalactiae 50. ABC transporters have been known for some time to be efflux mechanisms of drug resistance, although such efflux pumps are now also known to have physiological roles, conferring resistance to natural substances produced by the host, as well as possible roles in pathogenicity 51. Perhaps the proposed recent shift in host preference from bovine to human for this lineage 48 is facilitated by molecular adaptation of ABC transporters that confer resistance to natural substances of the new host.

Twenty six genomes of Streptococcus were downloaded from GenBank (Table 1), representing six different species. Coding sequences were extracted from GenBank files, and orthologs were determined using OrthoMCL 43. This program first makes an all-against-all BLASTp, and then defines putative pairs of orthologs or recent paralogs based on reciprocal BLAST. Recent paralogs are identified as genes within the same genome that are reciprocally more similar to each other than any sequence from another genome. OrthoMCL then converts the reciprocal BLAST p values to a normalized similarity matrix that is analyzed by a Markov Cluster algorithm (MCL) 55. In return, the MCL yields a set of clusters, with each cluster containing a set of orthologs and/or recent paralogs. OrthoMCL was run with a BLAST E-value cut-off of 1e-5, and an inflation parameter of 1.5. We used the OrthoMCL output to construct a table describing genome gene content (Additional data file 1). Genes that were not included in a cluster were considered taxon specific genes only if they were at least 50 amino acids long and had no BLAST hit with any other protein (E-value ≤ 1e-10). Preliminary analysis indicated that many truncated proteins found at the ends of contigs of the incomplete genomes, although exhibiting clear evidence of homology, were not included in any cluster because they had weak or no reciprocal BLAST hit. This table was used to plot venn diagrams with R 2.2.1 56 and to construct four core-genome data-sets corresponding to the following taxa: genus Streptococcus, S. agalactiae, S. pyogenes, and S. thermophilus.

Gene loss, acquisition, and duplication were determined on all branches of trees involving these taxa using the parsimony criterion with the DelTran option, implemented in Paup 4.0b10 57. Ancestral gene state reconstruction was run with two different step-matrices: the first one assesses gene gain, loss and duplication as being equally likely (for example, 23), while the second penalizes gene gain by assuming a double cost compared to loss and duplication (for example, 24). The gene content table was also used to perform gene accumulation curves using R, which describe the number of new genes and genes in common, with the addition of new comparative genomes. The procedure was repeated 100 times by randomly modifying genome insertion order to obtain means and standard errors.

Orthologs were first aligned at the DNA level with ClustalW 1.82 58. To ensure homology, alignments that contained less than 35% conserved sites for the Streptococcus genus data set, and 50% for the Streptococcus species data sets, were discarded. A preliminary analysis of the data revealed that many sequences identified as under positive selection contained frameshifts that disrupted the reading frame (a single insertion or deletion that modified the reading frame), resulting in high non-synonymous substitution rates. Unfortunately, it is not possible to accurately discriminate sequencing errors from actual insertions or deletions; however, most of these frameshifts were found within the unclosed genomes and appeared at the beginning or end of the contigs, where sequencing errors are more probable. As described by Perrodou et al. 59, most of these frameshifts are probably sequencing errors, although it is possible that some are not 60. We chose the conservative approach of removing all codons appearing before or after the frameshift when located at the beginning or end of the coding sequence, respectively. For that purpose, Perl scripts were developed to find frameshifts on the DNA alignments, and the sequences were edited manually. A second alignment step was then used to refine all alignments by translating sequences to amino acids, aligning them with ClustalW, and then back-translating to DNA, using the script transAlign 61. Finally, amino acid alignments showing a low percentage of conserved sites were manually inspected, and removed if the alignment was ambiguous.

Both phylogenetic and substitution pattern methods were use to detect recombination events. The phylogenetic methods rely on the examination of phylogenetic congruence among genes (for example, 6263). If a gene has been laterally transferred, the phylogenetic relationships depicted by this gene will be different from the species tree. We first applied the method suggested by Susko et al. 34, which tests for the rejection of a set of topologies by a set of orthologous genes using the AU test 64. When possible (that is, the number of taxa ≤ 7), the trees tested are all the possible unrooted trees (for example, 65). When a gene rejects a tree that is supported by the majority of the other genes, this gene is considered to have been laterally transferred. We applied this approach to our data sets (except for S. thermophilus, which contains only three taxa), by using as tested topologies the individual gene trees obtained by phyML (general time reversible (GTR) +Γ4+I model of evolution with a BIONJ starting tree) 66, with the addition of the tree obtained with Paup (GTR+Γ4+I model of evolution, neighbor-joining (NJ) starting tree, and a tree-bisection-reconnection (TBR) branch-swapping algorithm) reconstructed from the concatenation of all genes. The site likelihood of each tree was than computed by the program baseml (PAML package) 67 using a GTR+Γ4 model of evolution. The AU test was then applied using Consel 68.

As suggested by Susko et al. 34, results (p values for the rejection of each tree) were plotted using heatmaps obtained with R. This approach has the disadvantage of been developed with a test (the AU test) that assesses the rejection of a tree by a gene and not for its acceptance 34. As a result it is not possible to say if a tree is not rejected because it is not significantly different or because it is simply unresolved. This sort of situation is particularly expected with weakly divergent alignments, and at the opposite spectrum, with saturated alignments. We thus developed and performed a second set of analyses to complement the Susko approach and intended to quantify the amount of supported and incongruent phylogenetic signal between two gene trees. This approach relies on the discovery of well supported conflicting bipartitions (that is, branches that can not be observed in the same tree), as measured by non-parametric bootstrap analysis 69, thus revealing incongruence between gene histories. Support for each bipartition was obtained by bootstrapping a maximum likelihood (ML) tree search using Paup (GTR+Γ4+I model of evolution). Custom-made scripts were then used to find and count well supported (≥90% bootstrap support) conflicting bipartitions between gene trees. Additional to phylogenetic recombination detection, we employed methods specifically developed to detect homologous intragenic recombination. We used the compatibility approach between site histories, based on the pairwise homoplasy index (PHI), developed by Bruen et al. 70 and implemented within the program PhiPack 71. Bruen et al. have suggested PHI is a more robust and sensitive method than many of the earlier approaches. Additional to the PHI statistic, for comparative purposes, we computed MaxChi 72 and NSS 73p values for recombination using PhiPack (employing 1,000 permutations).

We employed the branch-site test of Yang and Nielsen 2074 implemented in the program Codeml of the package PAML 67 to assess positive selection at particular sites and lineages. Briefly, the likelihood of a model that does not allow positive selection is compared to one allowing positive selection on some specified lineages. The model allowing positive selection is tested using a likelihood ratio test (LRT) that is compared to a χ² statistic with two degrees of freedom. Likelihoods were estimated on the genes or species trees. For the Streptococcus data set each lineage leading to the six different species was tested. For the species analyses each of the lineages corresponding with the different strains was tested, as well as several internal branches supported by the majority of genes. Finally, p values were corrected for multiple hypothesis testing using the Benjamini and Yekutieli method 75. The effect of lineages, and genes' TIGR main role categories, and their interaction were tested using an analysis of variance on the LRT, using R. Role categories containing less than ten genes were merged. If F-statistic was significant, Tukey's 'honest significant difference' multi-comparison method was used to discriminate lineages and role categories associated with different LRTs.