Prediction of Protein-Protein Interacting Sites by Combining SVM Algorithm with Bayesian Method

The ability to identity protein-protein binding sites has important implications for drug design and understanding cell activity. This paper presents a method that can predict protein binding sites of transient protein-protein interactions using protein residue conservation and evolution information, i.e., spatial sequence profile, sequence information entropy and evolution rate. A two-stage predictor is constructed to predict surface residues are participated into protein-protein interface. The first stage consists of three predictors based on support vector machines (SVM) algorithm. Bayesian discrimination is used at the second stage by considering the predicted labels of spatial neighbor residues. The improvement of prediction performances exploits that binding site tend to form spatial cluster. Our proposed approach is promising which can be verified by its better prediction performance based on a non-redundant data set of transient protein- protein heterodimers.