Identifying functions for all gene products in all sequenced organisms is a central challenge of the post-genomic era. However, at least 30-50% of the proteins encoded by any given genome are of unknown or vaguely known function, and a large number are wrongly annotated. Many of these ‘unknown’ proteins are common to prokaryotes and plants. We set out to predict and experimentally test the functions of such proteins. Our approach to functional prediction integrates comparative genomics based mainly on microbial genomes with functional genomic data from model microorganisms and post-genomic data from plants. This approach bridges the gap between automated homology-based annotations and the classical gene discovery efforts of experimentalists, and is more powerful than purely computational approaches to identifying gene-function associations.
Among Arabidopsis genes, we focused on those (2,325 in total) that (i) are unique or belong to families with no more than three members, (ii) occur in prokaryotes, and (iii) have unknown or poorly known functions. Computer-assisted selection of promising targets for deeper analysis was based on homology-independent characteristics associated in the SEED database with the prokaryotic members of each family. In-depth comparative genomic analysis was performed for 360 top candidate families. From this pool, 78 families were connected to general areas of metabolism and, of these families, specific functional predictions were made for 41. Twenty-one predicted functions have been experimentally tested or are currently under investigation by our group in at least one prokaryotic organism (nine of them have been validated, four invalidated, and eight are in progress). Ten additional predictions have been independently validated by other groups. Discovering the function of very widespread but hitherto enigmatic proteins such as the YrdC or YgfZ families illustrates the power of our approach.
Our approach correctly predicted functions for 19 uncharacterized protein families from plants and prokaryotes; none of these functions had previously been correctly predicted by computational methods. The resulting annotations could be propagated with confidence to over six thousand homologous proteins encoded in over 900 bacterial, archaeal, and eukaryotic genomes currently available in public databases.