Support Vector Machines for Predicting Protein-Protein Interactions

Abstract

Protein-protein interactions play a crucial role in cellular processes. Although these interactions should be determined by stringent experiments and succeeded assays, these steps are time-consuming and labor-intensive. For this reason, many efforts have been made to predict unknown protein-protein interactions in silico from genomic and/or other information. One of the problems we have faced in this context is the accuracy of their prediction, which has great influences on the understanding of biological meanings inferred from their results. In this sense there is the need for further research to improve the accuracy of protein-protein interaction prediction. It is known that some of conserved domains physically interact each other, and some predictive methods have been devised based on this fact. However, since it is not guaranteed that an interaction between two proteins solely depends on the existence of particular domains within the proteins, it is expected that protein-protein interactions are more precisely predicted utilizing additional information. Support Vector Machines (SVM) has demonstrated high classification ability in the field of prediction of protein-protein interaction, functional classification of proteins, protein fold recognition, and prediction of subcellular location. For predicting protein-protein interactions, SVM is very useful in that domain information and several protein features including amino acid composition, sequential amino acid usage, and localization, whether they are continuous or discrete values, can be easily combined into a feature vector, and take them all into account. In this study we investigate the availability of SVM for predicting interactions within Saccharomyces cerevisiae using domain information and protein features, and show that the prediction accuracy is improved by the addition of protein features.