Predicting Protein Function and Nworks on a Genomewide Scale

Abstract

Perhaps the most significant finding from the more than 80 genomes that have been sequenced as of 2002 has been the extent of our ignorance about the constituents of cells. In virtually every genome sequenced, the majority of genes have never been studied directly. In spite of this, for about half of the genes at least one near or distant relative has been studied, so we glean our knowledge from the activities of these relatives. Until recently such methods for extending information to proteins with similar sequences or structures (homology-based methods) have been the only form of inference about protein function. Homology-based annotation, with algorithms such as BLAST (Alt-schul et al., 1997; http://www.ncbi.nlm.nih.gov/BLAST), has been wildly successful in extending knowledge from the small set of experimentally characterized proteins to the tens of thousands of proteins found in genome sequencing projects. However, these methods perform as one might expect: they provide information only for proteins with very closely related functions. They reveal little about proteins that work together but typically have unrelated sequences or structures. Thus, the homology-based methods cannot be used to reconstruct metabolic or signaling pathways or other protein interaction networks. That such a bias exists shows that there are different aspects to protein function; methods that reveal one aspect do not necessarily reveal others. The two most important aspects of protein function, defined in figure 9.1, will be referred to as the molecular function and the cellular function of proteins. The homology-based methods tend to find only the molec