Large margin methods for partner specific prediction of interfaces in protein complexes

Abstract

The study of protein interfaces and binding sites is a very important domain of research in bioinformatics. Information about the interfaces between proteins can be used not only in understanding protein function but can also be directly employed in drug design and protein engineering. However, the experimental determination of protein interfaces is cumbersome, expensive and not possible in some cases with today's technology. As a consequence, the computational prediction of protein interfaces from sequence and structure has emerged as a very active research area. A number of machine learning based techniques have been proposed for the solution to this problem. However, the prediction accuracy of most such schemes is very low. In this dissertation we present large-margin classification approaches that have been designed to directly model different aspects of protein complex formation as well as the characteristics of available data. Most existing machine learning techniques for this task are partner-independent in nature, i.e., they ignore the fact that the binding propensity of a protein to bind to another protein is dependent upon characteristics of residues in both proteins. We have developed a pairwise support vector machine classifier called PAIRpred to predict protein interfaces in a partner-specific fashion. Due to its more detailed model of the problem, PAIRpred offers state of the art accuracy in predicting both binding sites at the protein level as well as inter-protein residue contacts at the complex level. PAIRpred uses sequence and structure conservation, local structural similarity and surface geometry, residue solvent exposure and template based features derived from the unbound structures of proteins forming a protein complex. We have investigated the impact of explicitly modeling the inter-dependencies between residues that are imposed by the overall structure of a protein during the formation of a protein complex through transductive and semi-supervised learning models. We also present a novel multiple instance learning scheme called MI-1 that explicitly models imprecision in sequence-level annotations of binding sites in proteins that bind calmodulin to achieve state of the art prediction accuracy for this task.