The rapid accumulation of genome sequences is a major challenge to researchers attempting to extract the maximum functional and evolutionary information from the new genomes. To avoid informational overflow from the constant influx of new genome sequences, a comprehensive evolutionary classification of the genes from all sequenced genomes is required. Such classifications are based on two fundamental notions from evolutionary biology: orthology and paralogy, which describe the two fundamentally different types of homologous relationships between genes [1
]. Orthologs are homologous genes derived by vertical descent from a single ancestral gene in the last common ancestor of the compared species. Paralogs, in contrast, are homologous genes, which, at some stage of evolution of the respective gene family, have evolved by duplication of an ancestral gene. The notions of orthology and paralogy are intimately linked because, if a duplication (s) occurred after the speciation event that separated the compared species, orthology becomes a relationship between sets of paralogs (co-orthologs), rather than individual genes. A classic case of the interplay between orthologous and paralogous relationships is seen in the globin family: all animal globins, including myoglobin, are paralogs, but they are all co-orthologs of the plant leghemoglobin(s) [5
Deciphering orthologous and paralogous relationships among genes is critical for both the functional and the evolutionary aspects of comparative genomics [4
]. Orthologs typically occupy the same functional niche in different species, whereas paralogs tend to evolve toward functional diversification. Therefore, robustness of genome annotation depends on accurate identification of orthologs. Similarly, knowing which homologous genes are orthologs and which are paralogs is required for constructing evolutionary scenarios involving, along with vertical inheritance, lineage-specific gene loss and horizontal gene transfer.
In principle, identification of orthologs requires phylogenetic analysis of entire families of homologous proteins, which is expected to isolate orthologous protein sets in distinct clades [6
]. However, on the scale of complete genomes, such analysis is both extremely labor-intensive and error-prone due to the inherent artifacts of phylogenetic tree construction. Therefore shortcuts have been developed by introducing the notion of a genome-specific best hit (BeT). A BeT is the protein in a target genome, which is most similar to a given protein from the query genome [9
]. The underlying premise is that orthologs are more similar to each other than they are to any other protein from the respective genomes. In multiple-genome comparisons, pairs of potential orthologs identified via BeTs can be joined to form clusters of orthologs represented in all or a subset of the analyzed genomes [9
]. This approach to the identification of orthologous protein sets meets with two obvious complications. Firstly, many proteins belong to lineage-specific expansions, i.e., have evolved via duplication(s) after the divergence of the compared species [12
]. In these cases, deciphering (co)orthologous relationships can be a hard task and clusters of orthologs that include such expansions should be treated with particular caution. The second complication is caused by the fact that many proteins exist in multidomain forms encoded by a single gene in some species and as products of two or more stand-alone genes in others. In protein clustering, multidomain proteins may connect distinct clusters of orthologs resulting in artifactual lumping.
The approach to the identification of orthologous protein sets based on clustering of consistent BeTs has been implemented in the collection of Clusters of Orthologous Groups (COGs) of proteins [9
]. The COG construction protocol included an automatic procedure for detecting candidate sets of orthologs, manual splitting of multidomain proteins into the component domains, and subsequent manual curation and annotation. The COGs started with 6 prokaryotic genomes and one genome of a unicellular eukaryote, yeast Saccharomyces cerevisiae
]. Subsequent updates increased the number of prokaryotic genomes in the COGs to 43 [15
]. The procedure for COG construction required that each COG included proteins from at least three sufficiently distant species. This conservative approach notwithstanding, ~60 to ~85% of the proteins encoded in prokaryotic genomes were included in the COGs.
The COG system, which includes the COGNITOR program for adding new members to COGs (RLT, unpublished results), has become a widely used tool for computational genomics. The most important applications of the COGs are functional annotation of newly sequenced genomes [16
] and genome-wide evolutionary analyses [21
Here, we present a major update to the COGs, with over 63 sequenced prokaryotic genomes and three genomes of unicellular prokaryotes now included. Furthermore, the COG system is extended to complex, multicellular eukaryotes by constructing clusters of probable orthologs, which we named KOGs (eukaryotic orthologous groups) for 7 sequenced genomes of animals, fungi, microsporidia, and plants.