A New Perspective on an Old Tool: Extending the Coverage of Sequence Similarity-Based Function Prediction with PFP

The ultimate aim of molecular biology is to define functional roles for all proteins. In the last decade, several methods have been developed which produce large amounts of data, including correlated expression analysis by microarrays, fast genome sequencing and large-scale proteomics screens. Computational biologists have been called upon by the experimental community to aid in the organization, analysis and interpretation of this data so that its utility may be maximized. Thus far, proposed computational gene function prediction methods can be grouped into four distinct categories: evolutionary methods, which use conserved global sequence or structure to imply homology and motifs to assign biochemical function and binding sites; genomic methods, which link proteins through domain fusion events, phylogenetic profiling, conserved gene order and common regulatory elements; cellular methods, which use large proteomics datasets to define protein-protein interaction patterns and complexes; and metabolic methods, which utilize the structured networks of biochemical pathways to match proteins to uncharacterized reactions. The most successful techniques combine clues from multiple contexts to make reliable predictions, but a major limit to function prediction is limited coverage. Even BLAST [1] searches can only cover half of the genes in a genome. In order to provide functional clues that can spark analysis of large proteomics datasets, we need a method that expands coverage by lowering prediction resolution, i.e. a method that can provide accurate (but more generalized) predictions for proteins falling outside of the coverage range for current techniques.