To add a new species to the COG system, the annotated protein sequences from the respective genome were compared to the proteins in the COG database by using the BLAST program and assigned to pre-existing COGs by using the COGNITOR program (and see Materials and Methods). The genomes of prokaryotes and unicellular eukaryotes that have been sequenced since the latest update of the COGs were added one at a time. At each step, the proteins that remained unassigned after manual validation of the COGNITOR results were subject to the COG construction procedure in order to identify new COGs that could be formed thanks to the addition of the analyzed genome. The resulting COG assignments for 63 prokaryotic genomes and three genomes of unicellular eukaryotes are quantified in Table 1. The addition of new species leads to incremental increase in the COG coverage for each of the included prokaryotic genomes. The highest coverage now achieved is for Buchnera sp. (99%) and the lowest coverage is for Borrelia burgdorferi (43%). Each of these organisms is a special case. Buchnera is a highly degraded ensymbiont, which evolved from a relatively recent common ancestor with E. coli but apparently lost the great majority of genes, retaining – almost exclusively – conserved, essential ones 26, whereas Borrelia has numerous plasmids that mostly encode poorly conserved genes 27. Probably more telling is the observation that, for most free-living prokaryotes, ~80% of the genes belong to COGs and there is no appreciable dependence between the number of genes in a genome and the COG coverage (Table 1). Given that most genomes encode a substantial fraction (up to 10%) of fast-evolving, non-globular proteins 28 and other poorly conserved proteins (e.g., remnants of prophages) as well, these findings seem to suggest that the COG coverage of most genomes is approaching saturation.

The COGs are accompanied by a phyletic pattern search tool, i.e., a Web-based tool that allows the user to select COGs with a desired pattern of presence-absence of species. Using the phyletic pattern search tool, one can classify the COGs by the representation of the major lineages of unicellular life forms (Fig. 1). This breakdown of the updated COGs emphasizes the important trend noticed previously915: only a minuscule fraction (~1%) of the COGs are ubiquitous and even the COGs that are present in all bacteria or in all archaea represent a small minority. Furthermore, many COGs show scattered distribution, which appears to reflect rampant lineage-specific gene loss and horizontal gene transfer, which are typical of prokaryotic evolution 293031.

Eukaryotic KOGs were constructed from annotated proteins encoded in the genomes of three animals (Homo sapiens 32, the fruit fly Drosophila melanogaster 33, and the nematode Caenorhabditis elegans) 34, the green plant Arabidopsis thaliana (thale cress) 35, two fungi (budding yeast Saccharomyces cerevisiae 36 and fission yeast Schizosaccharomyces pombe 37, and the microsporidian Encephalitozoon cuniculi 38). The basic procedure for KOG construction was the same as the procedure previously employed for prokaryotic genomes (Refs. 915 and see Materials and Methods). Given the abundance of multidomain architectures among eukaryotic proteins and the fact that apparent orthologs often differ in domain composition 3239, the protocol based on the BeT analysis was amended with domain identification using the RPS-BLAST program 40. Proteins assigned to a KOG by the initial KOG construction procedure were kept in that KOG without splitting them into individual domains if they shared a common core of domains. In addition, proteins, which consisted solely of widespread, "promiscuous" domains (e.g., SH2, SH3, WD40 repeats or TPR repeats) and did not show clear-cut orthologous relationships, were assigned to Fuzzy Orthologous Groups (FOGs). In addition to KOGs and FOGs, we also identified provisional clusters of orthologs represented in two genomes (TWOGs) by detecting bi-directional BeTs between proteins not included in KOGs or FOGs and assigning additional members by examination of the BLAST search outputs. Finally, lineage-specific expansions (LSEs) of paralogs among the proteins from each genome not included in KOGs, FOGs, and TWOGs were detected by using the clustering procedure described previously 14 accompanied by a newly developed procedure for finding tight protein clusters (BK and RLT, unpublished results). The construction of TWOGs and LSEs involved more extensive case by case evaluation than the KOG construction due to the lack of well established procedures to generate these types of clusters; nevertheless, these clusters should be considered preliminary until further validation.

Table 2 shows the assignment of the proteins from each of the analyzed eukaryotic species to KOGs. Unlike the situation with prokaryotic COGs (Table 1), the fraction of proteins assigned to KOGs tends to decrease with increasing genome size of the analyzed eukaryotic species, from the maximum of ~74% for fission yeast Schizosaccharomyces pombe, the second smallest genome (for reasons that remain unclear, the smallest genome, that of the microsporidian Encephalitozoon cuniculi, had only 61% of the proteins included in COGs) to ~49% for the largest, human genome (Table 2).

Compared to prokaryotes, a considerably smaller fraction of eukaryotic genes could be included into KOGs (Tables 1 and 2). Thus, the apparent difference in coverage with highly conserved clusters of orthologs (C/KOGs) between prokaryotes and eukaryotes, particularly complex ones, is probably due to the relatively small number of eukaryotic genomes included in this analysis and is expected to level off with the growth of the eukaryotic genome collection. This view is compatible with the observed dependence of the KOG coverage on the number of genes (Table 1), which suggests that the KOGs are still far from saturation.

Examination of the phyletic patterns of KOGs points to the existence of a conserved eukaryotic gene core as well as substantial diversity (Fig. 2); this clearly resembles the evolutionary pattern seen previously during the analysis of archaeal COGs 41. The genes represented in each of the 7 analyzed genomes comprise ~20% of the KOG set and approximately the same number of KOGs includes 6 species, with the exception of the microsporidian. The prevalence of the latter pattern is not surprising given that microsporidia are intracellular parasites with minimal metabolic capabilities and a dramatically reduced genome 38. The next largest group consists of animal-specific COGs, which, again, could be expected because animals are the only lineage of complex eukaryotes that is represented by more than one species in the analyzed set of genomes. However, a notable observation is that ~30% of the KOGs had "odd" phyletic patterns, e.g., are represented in one animal, one plant and one fungal species (Fig. 2).

To illustrate the typical composition of a KOG, some of the problems that tend to emerge with their construction, and possible biological implications, we briefly discuss here KOG3378, which includes proteins already mentioned above as a typical case of paralogy and orthology, namely, the globins (Fig. 3). Globins are small (typically, between 140 and 150 amino acid residues) and relatively poorly conserved proteins. As a consequence, the initial, automatic procedure for KOG construction produced a candidate KOG consisting of only 3 proteins from 3 species: S. cerevisiae YGR234w, its ortholog from S. pombe SPAC869.02c, and human neuroglobin Hs10864065. The remaining proteins were brought into the KOG manually, as the result of examination of BLAST search outputs, focused on the conservation of the globin-specific sequence motifs. The final KOG is represented in 6 of the 7 analyzed eukaryotic species, with the sole exception of E. cuniculi (Fig. 3). The most remarkable aspect of this KOG is the apparent independent proliferation of genes for globins and globin-like proteins in vertebrates (represented here by humans): 11 paralogs, and nematodes (C. elegans): 24 paralogs (CE23430 and CE23431 are parts of the same gene). Strictly speaking, to demonstrate that these expansions are, indeed, independent, rather than ancestral, complete phylogenetic analysis is required, which is a difficult task given the low sequence conservation in many members of the KOGs. However, the presence of only one globin homolog in D. melanogaster is best compatible with hypothesis of lineage-specific expansion because, regardless of the exact topology of the animal phylogenetic tree 42, the alternative to this hypothesis would involve massive loss of globin-like genes in insects. Furthermore, this hypothesis is also compatible with the topology of the crude similarity dendrogram, which accompanies the KOG and in which the majority of human and nematode members form distinct clusters (Fig. 3). Thus, at this stage, the most likely, conservative interpretation of the evolutionary relationship between vertebrates and nematode globins is that they comprise co-orthologous sets and are legitimately included in the same KOG. Similarly, the two paralogous leghemoglobins from A. thaliana should be considered co-orthologous to the human and C. elegans paralogous sets.

The functions of human globins and globin homologs, primarily in oxygen delivery to different tissues, at different developmental stages have been studied in great detail 43. In contrast, the dramatic proliferation of globin-like proteins in the nematode C. elegans, while noticed, in part, in previous work 44, is not well understood. To our knowledge, KOG3378 is the most complete current representation of this lineage-specific expansion of globin-like paralogs; the experimental study of these genes is expected to reveal novel aspects of invertebrate physiology.

Another notable observation comes from the analysis of the yeast members of KOG3378. A BLAST search of the non-redundant protein sequence database (NCBI, NIH, Bethesda) and examination of the domain composition of the S. cerevisiae protein YGR234w shows that this protein (named flavohemoglobin) consists of the globin domain fused to a flavodoxin reductase domain and is highly similar to a variety of oxidoreductases from several bacterial species and some lower eukaryotes (e.g., slime molds and other protists), which have the same domain composition (45 and data not shown). The S. pombe flavohemoglobin belongs to the same protein family but is not the closest relative of the S. cerevisiae flavohemoglobin (data not shown). These observations strongly suggest that the yeast flavohemoglobin genes have been acquired from bacteria via horizontal gene transfer and hence have an evolutionary history that is distinct and independent from those of plant and animal globins. Notably, the second member of this KOG from S. cerevisie YNL234w is not at all a close paralog of the flavohemoglobins. The only identifiable domain in this large protein is the globin domain, which is most similar to vertebrate neuroglobins. These observations illustrate an important general point to be kept in mind when perusing the KOGs: although a given set of proteins may have been legitimately brought together in the same KOG in the context of eukaryotic genome comparison, on some occasions, different KOG members have different evolutionary trajectories.

The two sets of orthologous genes overlap because the three species of unicellular eukaryotes were included in both sets; the proteins from these species obviously form connections between prokaryotic orthologous sets (COGs) and eukaryotic orthologous sets (KOGs). Such connections, suggestive of orthologous relationships, were established between 1253 COGs, each of which included at least one protein from a unicellular eukaryote (not counting COGs that consisted exclusively of eukaryotic proteins), and 2000 eukaryotic KOGs. The greater number of eukaryotic KOGs involved in this relationship is due to the fact that, on many occasions, several proteins from unicellular eukaryotes that are part of the same COG have their distinct orthologs in other eukaryotes and, accordingly, belong to several KOGs. Only relatively small fractions of the prokaryotic COGs (27% of the COGs that include at least one prokaryotic species) and eukaryotic KOGs (34% of the KOGs and TWOGs) comprised sets of putative orthologs represented in both prokaryotes and eukaryotes. This emphasizes the distinction between the repertoires of genes that are conserved in prokaryotes and in eukaryotes and the considerable amount of innovation in both groups of organisms. However, these numbers give the low bound of the shared clusters of orthologs because some of the KOGs are not represented in the relatively small genomes of unicellular eukaryotes, primarily due to gene loss in the latter, but have prokaryotic counterparts.

Functional annotation of the detected orthologous clusters is one of the crucial and most labor-consuming aspects of the C/KOG analysis. Given the well-known inaccuracy of the currently available schemes for automatic annotation (e.g., Refs. 5,18, and references therein), no attempt was made to fully automate the C/KOG annotation; instead, assignments were made on a case by case basis through a combination of published data on C/KOG members and their homologs, protein domain analysis and different types of context analysis, particularly phyletic patterns and, in prokaryotes, conservation of gene strings which comprise putative operons 464748. Figure 4 shows the distribution of known and predicted protein functions for the prokaryotic COGs (i.e., the subset of the COGs obtained by subtraction from the COG collection of those COGs that included solely unicellular eukaryotes were) and the eukaryotic KOGs. The difference between prokaryotic and eukaryotic clusters of orthologs is obvious in that the latter are substantially enriched in proteins involved in signal transduction and intracellular trafficking; certain functional categories, such as cytoskeleton formation and chromatin dynamics were unique to eukaryotes. In contrast, metabolic and transport functions were relatively more prominent among the prokaryotic COGs (Fig. 4).

Phyletic pattern search can be employed for preliminary assessment of specific functional and evolutionary hypotheses. With the increased number of included genomes and enhanced capabilities of the phyletic pattern search tool, this analysis becomes particularly informative. Below we discuss straightforward examples of its use. Figure 5a shows the results of querying the COG database for COGs that are represented in the microsporidian parasite Encephalitozoon cuniculi but not in the two yeast species. Given the dramatic genome reduction seen in the microsporidium 38, it is not unexpected that the query retrieves only a small set of 13 COGs. This phyletic pattern can be explained either by loss of ancestral eukaryotic genes in yeast or by acquisition of genes by E. cuniculi via horizontal gene transfer. At least in some cases, further examination of the phyletic patterns of the retrieved COGs suggests the most likely scenario. Thus, COGs 1078, 1258, 1690, and 2263 are represented in all archaea, but are either missing in bacteria or are present in a minority of species (Fig. 5a). Therefore these COGs most likely are part of the ancestral archaeo-eukaryotic heritage 49 and might have been lost in yeasts; the respective proteins are known or predicted to be involved in translation or RNA modification, which is compatible with this evolutionary scenario. In contrast, COG3202 seems to be a likely case of horizontal gene transfer. Remarkably, the proteins in this COG are ADP/ATP translocases, which seem to be a hallmark of intracellular parasitism (or symbiosis) allowing the respective organisms to tap into the ATP supplies of the host cell 5051. Indeed, this COG is shared by the eukaryotic (E. cuniculi) and bacterial (Chlamydia and Rickettsia) intracellular parasites, the only exception being a diverged member of the COG found in the plant pathogenic bacterium Xylella fastidiosum (Fig. 5a).

The second case in point that we consider here is a search for COGs, which are represented in the causative agent of plague, Yersinia pestis 52, but not in other Proteobacteria (the taxon to which Y. pestis belongs) or eukaryotes; this query retrieves 7 COGs (Fig. 5b). These genes probably have been acquired by Y. pestis via horizontal gene transfer. On a more practical note, some of these genes could be potential targets for highly selective anti-bacterial agents. It is noticeable that three of these genes are predicted to be involved in cell wall metabolism (COGs 2152, 2401, and 3867), whereas the functions of others remain uncharacterized.

The protein sets for all newly included bacterial and archaeal genomes, the yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe, the microsporidian Encaephalitozoon cuniculi, the thale cress Arabidopsis thaliana, and the fruit fly Drosophila melanogaster were extracted from the Genome division of the (NCBI, NIH, Bethesda). The protein sequences for the nematode Caenorhabditis elegans were from the WormPep67 database, the sequences for Homo sapiens were from the NCBI build 30.

The new genomes were added to the COGs by using the COGNITOR program, with the results validated manually, essentially as described previously 915. After the completion of the validation process, the remaining proteins were subject to the COG construction procedure, in order to detect new COGs that could not be formed without the added genomes; the validation and annotation steps were repeated with the newly detected COGs.

The construction of KOGs followed the previously outlined strategy based on sets of consistent BeTs 915, but included additional steps that reflected specific features of eukaryotic proteins. Briefly, the procedure was as follows. 1. Detection and masking of widespread, typically repetitive domains, which was performed by using the RPS-BLAST program and the PSSMs for the respective domains from the CDD collection 40. These domains, namely, PPR (pfam01535), WD40 (pfam00400), IG (pfam00047), IGc1, Igv, IG_like, RRM (pfam00076), ANK (pfam00023), myosin tail (pfam01576), Fn3 (pfam00041), CA, (IG), ANK, kelch (pfam01344), OAD_kelch, SH3 (pfam00018), intermediate filaments (pfam00038), C2H2 finger (pfam00096), PDZ (pfam00595), POZ (pfam00651), PH (pfam00169), ZnF-C4 (pfam00105), spectrin (pfam00435), Sushi (pfam00084), TPR (pfam00017), BTB, LRR_CC, LY, ARM, SH2, and CH, were detected and masked prior to applying the COG construction procedure. Masking these domains was required to ensure the robust classification of the eukaryotic orthologous clusters with the KOG detection procedure because hits between these common, "promiscuous" domains resulted in spurious lumping of numerous non-orthologous proteins. 2. All-against-all comparison of protein sequences from the analyzed genomes by using the gapped BLAST program 58, with filtering for low sequence complexity regions performed using the SEG program 59. 3. Detection of triangles of mutually consistent, genome-specific best hits (BeTs). 4. Merging triangles with a common side to form crude, preliminary KOGs. 5. Case by case analysis of each candidate KOG. This analysis serves to eliminate the false-positives that are incorporated in the KOGs during the automatic steps and included, primarily, examination of the domain composition of KOG members, which was determined using the RPS-BLAST program and the CDD collection of position-specific scoring matrices (PSSMs) for individual domains 40. Generally, proteins were kept in the same KOG when they shared a conserved core domain architecture. However, in cases when KOGs were artificially bridged by multidomain proteins, the latter were split into individual domains (or arrays of domains) and steps (1)-(4) were repeated with these sequences; this results in the assignment of individual domains to KOGs in accordance with their distinct evolutionary affinities. 6. Assignment of proteins containing promiscuous domains. In cases when a sequence assigned to a KOG contained one or more masked promiscuous domains, these domains were restored and became part of the respective KOG. Proteins containing promiscuous domains but not assigned to any KOG were classified in Fuzzy Orthologous Groups (FOGs) named after the respective domains. 7. Examination of large KOGs, which included multiple members from all or several of the compared genomes by using phylogenetic trees, cluster analysis with the BLASTCLUST program ftp://ftp.ncbi.nih.gov/blast/, comparison of domain architectures, and visual inspection of alignments; as a result, some of these protein sets were split into two or more smaller ones that were included in the final set of KOGs.

The KOGs were annotated on the basis of the annotations available through GenBank and other public databases, which were critically assessed against the primary literature. For proteins that are currently annotated as "hypothetical" or "unknown", iterative sequence similarity searches with the PSI-BLAST program 58, the results of the RPS-BLAST searches, additional domain architecture analysis performed by using the SMART system 60, and comparison to the COG database by using the COGNITOR program (RLT, unpublished results) were employed to identify distant homologs with experimentally characterized functions and/or structures. The known and predicted functions of KOGs were classified into 23 categories (see legend to Fig. 4); these were modified from the functional classification previously employed for prokaryotic COGs 15 by including several specific eukaryotic categories.