Fuzzy SVM with a novel membership function for prediction of protein-protein interaction sites in Homo sapiens

Abstract

Predicting residues that participate in protein-protein interactions (PPI) helps to identify the amino acids located at the interface. In this work, experimentally verified 3-D structures of protein complexes are used for building the training model and subsequent prediction protein interactions from sequence information. Fuzzy SVM (F-SVM), which is developed on top of the classical SVM, is an effective method to solve this problem and we demonstrate that the performance of the SVM can further be improved with the use of a customdesigned fuzzy membership function. We evaluate the performances of both SVM and F-SVM on the PPI database of the Homo sapiens organism and evaluate the statistical significance of F-SVM over classical SVM. To predict interaction sites in protein complexes, local composition of amino acids together with their physico-chemical characteristics are used. The F-SVM based residues prediction method exploits the membership function for each pair sequence fragment and in all cases F-SVM improves the performances obtained by the corresponding SVM classifiers. The F-SVM performance on the test samples is measured by area under ROC curve (AUC) as 80.16% which is around 1.55% higher than the classical SVM classifier. © Springer-Verlag 2013.