Abstract

Accurate determination of functional interactions among proteins at the genome level remains a challenge for genomic research. Here we introduce a genome-scale approach to functional protein annotation-phylogenomic mapping-that requires only sequence data, can be applied equally well to both finished and unfinished genomes, and can be extended beyond single genomes to annotate multiple genomes simultaneously. We have developed and applied it to more than 200 sequenced bacterial genomes. Proteins with similar evolutionary histories were grouped together, placed on a three dimensional map and visualized as a topographical landscape. The resulting phylogenomic maps display thousands of proteins clustered in mountains on the basis of coinheritance, a strong indicator of shared function. In addition to systematic computational validation, we have experimentally confirmed the ability of phylogenomic maps to predict both mutant phenotype and gene function in the delta proteobacterium Myxococcus xanthus. The computational detection of interacting sets of proteins 1 is a topic of considerable interest, and several algorithms for this purpose have been developed in the last five years. These methods include the use of gene fusions as 'Rosetta Stones' 2 , the tracking of correlated mutations to infer interactions 3,4 , the enumeration of conserved operons 5 and gene neighbors 6 , and the calculation of phylogenetic profiles 7 . During the same period a host of approaches for the visualization of multidimensional biological data sets were described, including hierarchical clustering 8 , planar graph drawing 9 , singular value decomposition 10 and variants of multidimensional scaling We began with the premise that pairs of genes whose products are dissimilar, but which are consistently coinherited in the same sets of organisms are likely to be functionally linked 14 . Though the determination of coinheritance becomes more accurate as the number of sequenced genomes increases, the visualization of coinheritance patterns with clustered heatmaps 8 becomes impractical as the number of genomes approaches the number of proteins. Furthermore, although invaluable for the analysis of small sets of proteins 15 , heatmaps display too much information in each row to allow global visualization of similarity relationships. Phylogenomic mapping addresses both of these issues by providing a global visualization of coinheritance that can readily scale to incorporate the thousands of genome sequences that will soon be available. Phylogenomic map construction A phylogenomic map is a topographical visualization 11 of the similarity structure of a phylogenomic matrix 2 A phylogenomic map with labeled mountains for the genome of M. xanthus is pictured in