The results of the present analysis of conserved gene pairs are consistent with the notion of the fluidity of prokaryotic gene order caused by extensive recombination. Only 17 invariant genes pairs were detected, all of which consists of genes for ribosomal proteins and RNA polymerase subunits. The remaining 4586 gene pairs were missing in at least one genome. The number of gene pairs represented in three, four and a greater number of genomes decayed rapidly, with highly conserved pairs forming the tail of the distribution (Fig. 1). The 95% quantile of this distribution (excluding the highly conserved pairs) was found to fit the geometric model with a high statistical significance (Fig. 1). This is compatible with random, independent loss of gene pairs during evolution suggesting that, with the caveat of horizontal transfer, the number of gene pairs shared by three genomes could reflect the evolutionary distance between them.

The number of conserved gene pairs present in individual prokaryotic genomes varied from 208 for M. genitalium to 2314 for P. aeruginosa (Table 1). Analysis of the co-occurrence of gene pairs among the prokaryotic genomes shows high values of the Jaquard coefficient, which reflect partial conservation of gene order (see legend to Table 1), for closely related species, for example, 0.32 for E. coli and H. influenzae and 0.35 for M. thermoautotrophicum and M. jannaschi (Table 1). The value of this coefficient varied from 0.16 to 0.66, with a mean of 0.26, for archaea, and from 0.04 to 0.87, with a mean of 0.16, for bacteria. In contrast, for archaeal-bacterial comparisons, the values varied from 0.04 to 0.18, with the average of 0.08 (Table 1). These observations appear to indicate that the distribution of conserved gene pairs among prokaryotic genomes carries a phylogenetic signal.

Figure 2 shows a sampling of the distributions of identity percentage between pairs of apparent orthologs identified as reciprocal best hits from a range of genome pairs separated by varying phylogenetic distances. Most of the distributions are clearly unimodal, and the distributions for pairs of phylogenetically distant genomes, such as those from different major bacterial lineages or bacteria versus archaea, have their modes within a relatively narrow range around 33% identity (Figure 2).

The use of reciprocal best hits is a conservative way to identify the set of probable orthologs between pairs of genomes because some of the orthologs are missed due to complex relationships between groups of paralogs. Nevertheless, all genome-to-genome comparisons included at least 100 (for the smallest genomes such as the mycoplasmas), and typically, a considerably greater number of protein pairs (11 and data not shown). This suggests that parameters of the distributions of the similarity level between probable orthologs identified in this fashion could potentially serve as useful measures of the evolutionary distance between genomes.

Genome trees were generated using the approaches described under Material and Methods. All the trees showed a clear separation of the two major prokaryotic domains, Bacteria and Archaea (Fig. 3,4,5). Several terminal bifurcations that reflect clustering of relatively close species, such as three mycoplasmas (M. genitalium, M. pneumoniae and U. urealiticum), two spirochetes (B. burgdorferi and T. pallidum), and H. pylori and C. jejuni, are also reproduced in all trees (Fig. 3,4,5). This retention of both the deepest and the terminal branchings shows that all types of data used for tree construction contained at least a crude phylogenetic signal. However, beyond these obvious aspects of topology, and in particular with respect to clustering of distantly related bacteria and archaea, the trees produced with different approaches showed significant differences, which appear to reflect the relative contributions of phenotypic and phylogenetic signals. A quantitative comparison of the tree topologies using the symmetric distance method showed that the presence-absence tree was most different from the trees made by the other methods (Table 2).

The topology of the parsimony tree built using this criterion appears to reflect primarily the phenotypes of the respective organisms (Fig. 3). This is most clearly manifest in the two major bacterial clusters that appear in this tree, each with a strong bootstrap support:

i) bacteria with large genomes, namely E. coli, B. subtilis, Synechocystis sp., Deinococcus radiodurans and Mycobacterium tuberculosis, and free-living bacteria with small genomes, A. aeolicus and T. maritima

ii) parasites with small genomes (mycoplasmas, spirochetes, chlamydia and rickettsia)

Parasites with moderate-sized genomes (H. influenzae, N. meningitidis, and P. multocida; H. pylori and C. jejuni) formed two distinct groups. Thus, well-established phylogenetic relationships between free-living and parasitic bacteria, such as those within the Proteobacteria (E. coli-H. influenzae-P. multocida-N. meningitidis) and within low-GC Gram-positive bacteria (B. subtilis-mycoplasmas), are not reflected accurately in this tree topology. The two free-living bacteria with small genomes, the hyperthermophiles A. aeolicus and T. maritima, did not join either the free-living or the parasitic bacterial cluster, despite their small number of genes similar to that in bacterial parasites (Fig. 3). That these bacteria do not group with the parasites despite similar genome sizes, suggests that it is not the number of genes per se, but rather the degree of genome degradation and the loss of coherent sets of genes that affect the topology of the presence-absence tree. The inclusion of the parasites M. tuberculosis and Pseudomonas aeruginosa in the cluster of bacteria with large genomes probably reflects the recent origin of parasitism in these lineages. It is further notable that, in this tree, the two representative of Crenarchaeota (A. pernix and S. solfataricus) do not comprise a sister group of the Euryarchaeota (the remaining archaeal species), but rather for am branch within the Euryarchaeal cluster (see discussion below).

In previous studies that employed similar approaches to genome-tree building, phylogenetically reasonable clades were observed after a simple omission of parasitic species 89. Such an operation could be applied to the tree shown in Fig. 3, indeed resulting in the correct recovery of the proteobacterial and Gram-positive bacterial lineages. However, it seems that, because known natural groups could be reproduced by this approach only after omission of certain species on the basis of independent prior knowledge, this method hardly can be useful for delineating new, phylogenetically sound clades.

The topology of the tree based on gene pair conservation seems to carry a stronger phylogenetic signal than the gene presence-absence tree because it correctly groups together related free-living and parasitic bacteria despite major differences in gene repertoires (Fig. 4). The bacterial side of this tree consists of three major clades: i) proteobacterial clade that, in addition to bona fide Proteobacteria, includes also A. aeolicus, M. tuberculosis, D. radiodurans, and Synechocystis sp, ii) Gram-positive clade that additionally includes T. maritima, and iii) an unexpected clade that unites spirochetes and chlamydia. In the archaeal domain, the two species of the Crenarchaeota did not form a clade, but instead were present as separate branches interspersed with euryarchaeal species. To further assess the robustness of the obtained tree, we varied the parameters of the included conserved pairs by allowing distances between the genes comprising a pair from 0 to 5 and changing the minimal number of genomes, in which a conserved gene pair had to be present, from 2 to 4. These changes did not significantly affect the tree topology (data not shown). The topology of a neighbor-joining tree constructed by using the number of gene pairs shared by two genomes to calculate the evolutionary distance between them was similar to the topology of the maximum parsimony tree (Table 2 and data not shown).

At least some unusual aspects of this tree's the topology could be explained by horizontal transfer of operons between particular bacterial and archaeal lineages. Specifically, it has been noticed previously that T. maritima shares a considerable number of genes and operons with Gram-positive bacteria, to the exclusion of other bacteria 21; this seems to be compatible with the position of T. maritima with the Gram-positive cluster. Similarly, considerable horizontal gene transfer appear to have occurred between the Sulfolobus and Thermoplasma lineages, which cluster together in the archaeal part of this tree. The presence of extra species in the proteobacterial cluster is more surprising because no obvious trend for operon transfer between these bacteria and bona fide Proteobacteria has been noticed during systematic genome comparisons; however, a considerable number of shared gene pairs was detected during the present analysis (Table 1). Artifacts of tree construction could also contribute to these associations. In contrast, the spirochete-chlamydia clade might reflect a deep phylogenetic relationship (see discussion below).

Different characteristics of the distributions of percent identity between the probable orthologs, such as the mean, the median, the mode and various quantiles, were used to calculate distances between genomes and construct phylogenetic trees. Trees built with different cut-off values for symmetrical best hits, four different formulas for the evolutionary distance calculation (see Materials and Methods) and different parameters of the distributions showed essentially the same topology, with strong bootstrap support for most of the clades (Fig. 5 and data not shown). The complete proteobacterial and Gram-positive bacterial clusters were recovered in this tree as well as the unexpected grouping of chlamydia with spirochete noticed above in the tree based on conserved gene pairs (Fig. 4,5). Also similarly to the previous two trees, the Crenarchaea grouped with Thermoplasma within the archaeal part of the tree. Beyond these groupings, the tree appeared conservative in the sense that the unassigned bacterial species formed separate branches near the root of the bacterial subtree. The closest to the root were the two hyperthermophilic species, A. aeolicus and T. maritima, which is compatible with the standard view of their phylogenetic position 126.

The above three approaches involve construction of genome trees "par excellence", i.e. based on integral characteristics of genomes (or, more precisely, gene sets) that are not directly related to more traditional, alignment-based measures, which are usually employed for calculating evolutionary distances or for parsimony analysis. These genome tree raise several interesting phylogenetic questions, for example, do spirochetes and chlamydia indeed share a common ancestor, and are Euryarchaeota, in fact, a paraphyletic group with respect to the Crenarchaeota. However, the reliability of the conclusions drawn from the topology of these trees remains uncertain. Therefore we decided to complement these genome-oriented approaches with more traditional ones applied on a large scale.

The database of Clusters of Orthologous Groups of proteins (COGs) was used as the source of information on orthologous genes in prokaryotic genomes 3536. Briefly, the COGs were constructed from the results of all-against-all BLAST 37 comparison of proteins encoded in complete genomes by detecting consistent groups of genome-specific best hits (BeTs). The COG construction procedure does not rely on any preconceived phylogenetic tree of the included species except that certain obviously related genomes (for example, two species of mycoplasmas or pyrococci) were grouped prior to the analysis, to eliminate strong dependence between BeTs. In order to avoid spurious occurrence of the same gene pair, only gene pairs conserved in three or more genomes were considered. A pair of genes from two COGs was considered to be conserved if the respective genes were adjacent in at least one genome and were separated by no more than two genes in at least two additional genomes. This relaxed definition of a conserved gene pair was adopted to take into account the high level of recombination in prokaryotic genomes. From the data on the presence-absence of each conserved gene pair in the analyzed genomes (excluding pairs of closely related species: E. coli-Buchnera sp., H. influenzae-P. multocida, C. trachomatis-C. pneumoniae, P. horikoshii-P. abyssi, M. genitalium-M. pneumoniae-U. urealyticum, H. pyroli – C. jejuni, T. acidophilum-T. volcanium), a 0/1 matrix analogous to the one used for the presence-absence of individual genes was constructed, and a tree was built using Dollo parsimony 38. A parsimony method was chosen for this analysis because the presence-absence of a conserved gene pair in a genome can be naturally treated in terms of character states. The Dollo model is based on the assumption that each derived character state (in this case, the presence of a gene pair) originates only once, and homoplasies exist only in the form of reversals to the ancestral condition (absence of a gene pair) 38. In other words, parallel or convergent gains of the derived condition are assumed to be highly unlikely. The Dollo parsimony method is not sensitive to gene loss which is extremely common in evolution of prokaryotes, but the results can be affected by independent acquisition of the same gene pair by different genome via horizontal gene transfer. Phylogenetic analysis was performed by using the PAUP 4.0 program 39, with 1000 bootstrap replicates performed to assess the reliability of the tree topology. In addition, the tree topology was analyzed using the neighbor-joining method 40.

The information on orthologous genes in prokaryotic genomes and the yeast genome was derived from the COGs as in the previous approach, and the orthology data were similarly represented as a 0/1 matrix of presence-absence of the analyzed genomes in the COGs. A Dollo parsimony tree was constructed and the reliability of its topology was assessed using the bootstrap method as described above.

The sequences of all proteins encoded in the analyzed genomes were compared to each other using the gapped BLASTP program 37. Reciprocal, genome-specific BeTs were collected at different expectation (E) value cutoffs (0.01, 0.001, 0.0001, 0.00001). This method for identification of probable orthologs is, in principle, similar to the method employed in COG construction, but differs in that there is no requirement for the formation of triangles of consistent BeTs. The result of this procedure is a conservative selection of orthologous pairs because the cases of lineage-specific duplication that result in non-symmetrical BeTs are excluded and so are orthologous pairs with very low sequence similarity. However, the limitation of the COG system, namely the requirement that each orthologous group is represented in at least three genomes, is avoided. The distributions of identity percentage among the reciprocal best hits were derived for each pair of species. The mean, mode, median and different quantiles of the identity percentage distributions were used for estimating evolutionary distances. Four distance measures were used, namely: i) P-distances calculated as the fraction of different residues: d = 1-q, ii) Poisson distances d = -1nu, iii) geometric distances calculated using the formula d = 1/u-1, and iv) logarithmic distances found as a solution of the equation u = ln(1+2d)/(2d), where d is the evolutionary distance, q is percent identity, and u = (q-0.05)/0.95 414243. Trees were constructed from the distance matrices obtained with the above distance estimates using the neighbor-joining method 40 as implemented in the NEIGHBOR program of the PHYLIP package 44. Bootstrap values were estimated by resampling the set of orthologs identified for each pair of genomes 1000 times and reconstructing trees from the distributions of the distances from these resampled sets.

Sets of orthologous ribosomal proteins were extracted from the COG database, and their amino acid sequences were aligned using the T-Coffee program 45, with subsequent manual validation and removal of poorly aligned regions. The alignments are available upon request. Pairwise evolutionary distances between the sequences in concatenated alignments were calculated using the Dayhoff PAM model as implemented in the PROTDIST program of the PHYLIP package 44. A distance tree was constructed from the resulting distance matrix by using the least-square 46 method as implemented in the FITCH program of PHYLIP 44. The maximum likelihood tree was constructed with the JTT-F model of amino acid substitutions 47, as implemented in the ProtML program of the MOLPHY package 48, by optimizing the least squares tree with local rearrangements. Alternative topologies were created manually by modifications of the original tree and directly compared by ProtML. Bootstrap analysis was performed by using the Resampling of Estimated Log-Likelihoods (RELL) method as implemented in ProtML 4849.

The representative families were selected from the COG database according to the following criteria: i) at least 30 species are represented; ii) no more than two paralogs in any of the species; iii) no more than 1.2 paralogs per genome on average; iv) at least 100 positions in the alignment containing less than 30% of gaps. This selection procedure resulted in a set of 132 families (COGs). Alignments and ML trees were constructed for these families as described above for the concatenated ribosomal proteins.

To compare tree topologies quantitatively, the symmetric distance between trees 50 was computed using the TREEDIST program of the PHYLIP package (version 3.6a). Briefly, each of the two compared trees is divided by each internal branch into two partitions. The symmetric distance is the number of partitions that are found in one tree but not the other.