112. Structure from sequence: A view based on a global organization of the protein space

Introduction An essential step in any large-scale structural and functional genomic projects is to assign information to new un-annotated sequences. The flood of sequences accumulating from genome projects argues that a pressing need in structural and functional prediction efforts are automatic methods to bridge sequence and structural information. An unbiased set of structural representatives is used herein for determining the limitation of sequence-based information towards structural similarity [1]. It was shown that pairs of protein having >30-35% sequence identity (over long sequence), almost always share structural similarity. Below that ‘safe’ level the noise dominates, and for all protein pairs sharing at most 25% sequence identity only ~10% are indeed structural homologues. Below the 20% sequence identity, sequence based methods shown to perform poorly. Comparing advanced search programs for sequence-based structural predictions evaluated the limits of accurate identifications for proteins with low sequence significance scores [2]. Herein, we show that a hierarchical clustering of the entire protein space, based on sequence alone, manages to enrich the proportion of those pairs of proteins that are indeed structural homologues from within a set of protein pairs that share < 25% sequence identity.