Use of machine learning features to detect protein-protein interaction sites at the molecular level

Abstract

Protein-protein interactions (PPI) play pivotal roles in many biological processes like hormone-receptor binding. Their disruption leads to generation of inherited diseases. Therefore prediction of PPI is a challenging task. Machine learning has been found to be an appropriate tool for predicting PPI. Machine learning features generated from a set of protein hetero-complex structures were found to be a good predictor of PPIs. These machine learning features were used as training examples to develop Support Vector Machines (SVM) and Random Forests (RF) based PPI prediction tools. Among the important features the sequence based features related to sequence conservations and structure based features like solvent accessibility were found to have the maximum predictive capability as measured by their Area Under the Receiver Operating Characteristics (ROC) curves (AUC value). The RF based predictor was found to be a better performer than the SVM based predictor for this training set.