Previous work on a more reduced set of sequenced genomes had determined that large genomes contain more paralogs and more gene families than smaller genomes 7. Jordan and collaborators also found a correlation between the fraction of the genome occupied by gene families and the genome size; that is, larger genomes had a larger proportion of redundant genes. However, at the time of that analysis, the sequences of genomes larger than 5 million base pairs (5 Mbp) were not available. Now, the inclusion of genomes nearly twice as large confirms both trends (Figure 1): for example, nearly 50% of the genome is occupied by paralogous genes in Streptomyces coelicolor. A closer look at these data shows that larger genomes have larger gene families, as the average family size also increases with genome size (Figure 1, inset). Thus, the higher percentage of paralogs in large genomes is partly due to the expansion of existing gene families, together with a larger number of new families. The large-genomed species at one end of the distribution, such as Streptomyces, have gene families of up to 85 members, whereas the largest gene families in middle-sized genomes such as those of E. coli or Salmonella have more moderate numbers (40-45). This is reminiscent of the situation in eukaryotes, where the number of gene families increases with the number of genes in the genome at a lower rate than in prokaryotes 6, indicating that gene families have many more members in the larger eukaryotic genomes. Also consistent with this trend, some reconstructions of prokaryotic genome evolution based on gene content conclude that gene duplication has a critical role in the expansion of genome size 15.

Exceptions to the linear correlation in this graph are interesting to consider. On one hand, Pirellula (marked as Pir in Figure 1) has an enormous genome with a surprisingly low relative number of paralogs. This is due to an overrepresentation of small gene families and the absence of large ones (the largest gene family contains 57 members; see Additional data file 1). Pirellula is a marine bacterium and the reason for the reduced gene family size might be the homogeneity of the marine environment, in contrast to other large-genomed bacteria included in the graph which have the ability to survive in many different niches or in much more heterogeneous habitats, such as soil. In agreement with this, Pirellula has a greatly reduced number of transcriptional regulators, which again might reflect a relatively constant environment 19. At the other end of the distribution, exceptions occur for three species that have small genomes with a larger-than-expected percentage of paralogs. All these species are mycoplasmas, and the high percentage of paralogs is due to a few gene families that are greatly expanded, including more than 25 members. In Mycoplasma penetrans, for example, these families include surface-exposed lipoproteins involved in antigenic variation 20, which are critical to the success of microbes exposed to the immune system of their hosts. On the other hand, the small genomes of other pathogenic bacteria correspond to intracellular parasites that do not need to evade the immune system 21, and these species show the smallest portion of paralogs. Finally, the largest gene families that we detected were those involving mobile genetic elements such as the IS elements of Shigella flexneri, where families surpassed 100 members (not included in the inset of Figure 1).

The data in Figure 1 cannot be viewed as a continuum, because small genomes are not ancestral to bigger ones. Instead, small genomes have been shown to be the result of reductive evolution, a process by which a larger-sized ancestor changes niche and undergoes a dramatic loss of DNA 2223. Both small and large genome fragments can be eliminated but the outcome of this process for gene families has not been documented. We have compared the number of members per gene family in two genomes that are undergoing rapid reductive evolution - Shigella flexneri 2a and Mycobacterium leprae TN - with larger-genomed close relatives (Figure 2). Shigella is a close relative of E. coli that has specialized in living as a human pathogen 2425. As a result of the expansion of the human population from Neolithic times a number of more generalistic or opportunistic pathogens found a new niche; Salmonella typhi might be a similar example 26. In both cases there is a clear tendency to genome reduction accompanied by expansion of IS families (314 and 46 IS elements, respectively).

In Shigella there is a clear reduction in gene family copy number (Figure 2), which seems to be higher than would be expected from the random location of IS elements, suggesting that they might insert preferentially in gene family members. Something similar is found in the case of M. leprae (Figure 3), although in this case the main mechanism for gene inactivation is the generation of pseudogenes by mutation 27. M. leprae is closely related to M. tuberculosis, with which it shares many homologous sequences. However, most gene families have been simplified in the short time period in which the leprosy bacillus has adopted its mainly intracellular lifestyle. This also illustrates the fact that, as described above, an early step in genome reduction allowed by intracellular parasitism or a narrower range of hosts is the shrinkage of gene families. It shows that the smaller percentage of paralogs in reduced genomes is probably due to simplification of existing gene families. A similar pattern was found in the small-genomed intracellular species Rickettsia and Buchnera when compared with free-living species of the same taxonomic group (see Additional data file 2). Thus, both genome expansion and reduction can be partly explained by the parallel growth or simplification, respectively, of gene families.

Another feature we could detect in the evolution of gene families was that large families were more divergent (Figure 3). This could partly be due to a side-effect of the higher variability of a larger sample size or to misidentification of family members at low sequence identity levels. However, given the observed similarity of functions in these large families (428 and R.P., A.M. and F.R-V., unpublished results), a substantial proportion must be true paralogous genes. Thus, this relationship can be interpreted as older (more divergent) families containing more members. Smaller families range from those with very similar members to those in which the members are very different. The latter probably represent either old families in which new members have not evolved because new duplications do not confer a selective advantage, or more recent incomplete xenologous replacements.

The sequencing of several strains of a single species is now common in bacterial genomics. One of the most remarkable findings has been the different gene pools carried by strains that are highly similar if their housekeeping genes only are compared. For example, different virotypes of E. coli were shown to contain very different gene complements, with large pools of genes characteristic of each virotype 10. Obvious candidates to vary would be multigene families. Thus, the comparison of the numbers of members within a single species might shed light in their origin. If the members of a gene family are frequently acquired by HGT from outside, the numbers should be expected to vary broadly in different lineages of the species (as a result of different acquisitions). On the other hand, if the numbers are similar, that would indicate that the families were already present in the common ancestor and represent a relatively stable feature of the genome.

We selected distinct prokaryotic taxa in which three or more strains have been fully sequenced (Escherichia coli, Streptococcus pyogenes, Staphylococcus aureus and Chlamydophila pneumoniae) and for each taxon established a list of homologous genes common to all strains. The gene family to which each homolog belonged was determined for each strain, and the number of family members compared for equivalent families (Figure 4). In all four species considered, the different strains showed a remarkably similar pattern of gene family size distributions: large gene families in one strain were also expanded in the others; small families were small, regardless of strain or virotype. Caution has to be exercised when examining these plots, as a gene can be a member of more than one gene family. However, although some of the gene families in Figure 4 are redundant, the parallel size pattern of gene families across strains is remarkably clear and seems to reflect a stable feature of the genome. Thus, the majority of gene families were most likely to have been formed by ancestral gene duplications or ancient gene transfers common to all strains. In addition, the preservation of gene family size in different strains strongly suggests that most family members have a high value for survival; redundant copies would otherwise be quickly eliminated.

We have obviously not excluded the possibility that nonhomologous gene families add to the differences among the compared genomes. For example, in a pairwise comparison between E. coli K12 and E. coli O157:H7, 186 genes belonging to paralog families were unique to K12 and 788 to O157:H7, versus 403 singletons (single-copy genes not belonging to families) unique to K12 and 883 to O157:H7. Thus, K12 keeps the same standard proportion of 30% paralogs for the differential gene pool. In O157:H7, on the other hand, paralogs account for 47% of the set of unique genes. The interpretation might be that the large islands that characterize the genome of the enterohemorrhagic virotype tend to carry a bigger proportion of families than the rest of the genome. Thus, it is possible that in some strains, HGT may contribute to expand and generate gene families that do not appear as homologs in closely related genomes. For example, 146 genes belonging to families of 10 or more members were detected in the O157:H7 differential pool, including three whole families of 14, 17 and 20 members with a G+C content of 57, 54 and 53%, respectively (the average G+C content in E. coli O157:H7 is 50.6%). The largest differential family in K12 had 11 members, which were not present in the enterohemorrhagic strain, and had a G+C content of 54.1% (the average G+C content of E. coli K12 is 50.5%).

To investigate whether the conservation in the size of homologous families was maintained across more divergent genomes, gene family plots were performed between species. A representative case for a Gram-negative (Pseudomonas) and a Gram-positive (Bacillus) comparison is illustrated in Figure 5. The preservation of family size was still remarkable, although, in the case of Pseudomonas, the number of orthologous genes is considerable smaller. The overall pattern of family sizes is preserved across these species. The two Bacillus species considered have the same genome size and one species contains larger numbers in some families but fewer in others (Figure 5b). The same trend was found in comparisons between species of Staphylococcus, Streptococcus, Salmonella and Mycoplasma (data not shown). It is also interesting to analyze the variation detected. Part of it can be attributed to differences in genome size. Pseudomonas syringae is approximately 200 kb larger than its other sequenced partners, which have mostly smaller gene families. However, part of the variability is also due to intrinsic differences between the species. For example, P. syringae contains some large gene families involved in invasion of the plant host and in pathogenesis 29. One way to examine whether this variation can underlie the phenotypic/ecological characteristics of a given species is to visualize the size difference of each paralog group for some representative cases.

Figure 5c shows the difference in gene family size in the interspecific comparison of E. coli K12 and S. typhimurium LT2. Both strains have similarly sized genomes (S. typhimurium is 218 kb larger) and a relatively high level of homology (3,026 orthologous genes). Of these, there are 572 homologs belonging to families that differ in size between the two genomes, and 435 belonging to families having the same number of members in both species. The rest are single-copy genes in both genomes. Forty-eight families were significantly larger (two or more extra copies) in E. coli, while 53 were larger in Salmonella. These differences can be taken as an example of the evolution of gene families in two diverging groups. Although the natural history of these model bacteria is not as well known as might be expected, it is generally believed that both Salmonella and Escherichia are mostly saprophytic facultative anaerobes that inhabit the intestine of vertebrates. The divergence between these two microbes arose after the origin of mammals around 120 million years ago. E. coli specialized as a commensal and an opportunistic pathogen of mammals, as witnessed, for example, by its ability to degrade lactose. On the other hand, Salmonella remains as a commensal in reptiles, with some serotypes colonizing mammals, but as a pathogen rather than a commensal and after developing strategies for intracellular invasion of the host 3031. Accepting this scenario, the fact that many gene families (and the number of members of each family) are preserved reflects a significant involvement in the saprophytic intestinal lifestyle, preserved over many millions of years. On the other hand, significant differences are starting to arise between the two species, perhaps reflecting their specialization in different hosts and lifestyles 32. A dramatic example is the potG gene family, which has 13 more members in S. typhimurium than in E. coli (Figure 5). This is an ATP-binding component of spermidine/putrescine transport and for some reason its amplification has been selected in this species. Proteins involved in the transport of spermidine and putrescine have been shown to be involved in attachment to host cells and virulence 33. Therefore, the size of this gene family might reflect the more pathogenic lifestyle of Salmonella.

Do certain functions predispose genes to form families? Do single genes that do not form families belong to a different category? To address these questions, extended gene families were identified, where a gene was not allowed to belong to more than one family. Thus, if gene A matched gene B, and gene B matched C, but A did not match C, all three were considered part of the same family, as it is likely that they are all evolutionarily derived from each other 34. This method of transitive assembly of paralogs has been confirmed to include, in most cases, genes with related functions 4. We found that, for all 127 sequenced species, singletons (genes in a single copy in a given genome) were massively overrepresented by genes with an unknown or hypothetical function. When only genes with a known or predicted function were included, these single genes without paralogs appeared equally distributed among the different functional categories. However, when genes belonging to families, especially large ones, were considered, a significant fraction had particular functions, such as transport of metabolites (data not shown). These data are, however, probably unrealistic because they represent the distribution of genes in sequenced genomes only, and certain species are overrepresented. In addition, larger genomes will also weigh more in this comparison than small genomes, as will species with several sequenced strains. We did, however, find relatively uniform results for individual genomes. Figure 6 shows such a distribution for two species, one Gram-negative (E. coli K12) and the other Gram-positive (B. subtilis). For E. coli, in which a large proportion of genes has been allocated a function, families with more than five members contain fewer unknown or hypothetical genes than do smaller families, and the distribution of functions among categories is unequal, with certain categories being overrepresented. Among these, genes involved in transport of different metabolites predominate (39% of the total), followed by those with transcription and replication/repair functions. In genes that do not belong to a family, however, most functional categories are equally represented and a large proportion of these singletons have an unknown function. The overrepresentation of unknown or hypothetical open reading frames (ORFs) could, in part, be due to many of these singletons not being real genes, as supported by their shorter length when compared to genes belonging to families. In the gamma-proteobacteria, for example, average singleton length is 127 nucleotides less than in genes belonging to families. It is also interesting to note that the phylogenetic distribution of these unknown singletons is not different from that of unknown paralogs (see Additional data file 4). In conclusion, some functions do appear to be more prone to develop families, although the functions overrepresented in a particular species may depend on its lifestyle.