A strategy for large-scale comparison of evolutionary- and reaction-based classifications of enzyme function

Determining the molecular function of enzymes discovered by genome sequencing represents a primary foundation for understanding many aspects of biology. Historically, classification of enzyme reactions has used the enzyme nomenclature system developed to describe the overall reactions performed by biochemically characterized enzymes, irrespective of their associated sequences. In contrast, functional classification and assignment for the millions of protein sequences of unknown function now available is largely done in two computational steps, first by similarity-based assignment of newly obtained sequences to homologous groups, followed by transferring to them the known functions of similar biochemically characterized homologs. Due to the fundamental differences in their etiologies and practice, `how' these chemistry- and evolution-centric functional classification systems relate to each other has been difficult to explore on a large scale. To investigate this issue in a new way, we integrated two published ontologies that had previously described each of these classification systems independently. The resulting infrastructure was then used to compare the functional assignments obtained from each classification system for the well-studied and functionally diverse enolase superfamily. Mapping these function assignments to protein structure and reaction similarity networks shows a profound and complex disconnect between the homology- and chemistry-based classification systems. This conclusion mirrors previous observations suggesting that except for closely related sequences, facile annotation transfer from small numbers of characterized enzymes to the huge number uncharacterized homologs to which they are related is problematic. Our extension of these comparisons to large enzyme superfamilies in a computationally intelligent manner provides a foundation for new directions in protein function prediction for the huge proportion of sequences of unknown function represented in major databases. Interactive sequence, reaction, substrate and product similarity networks computed for this work for the enolase and two other superfamilies are freely available for download from the Structure Function Linkage Database Archive (http://sfld.rbvi.ucsf.edu).