Protein–protein interaction sites prediction by ensembling SVM and sample-weighted random forests

Predicting protein–protein interaction (PPI) sites from protein sequences is still a challenge task in computational biology. There exists a severe class imbalance phenomenon in predicting PPI sites, which leads to a decrease in overall performance for traditional statistical machine-learning-based classifiers, such as SVM and random forests. In this study, an ensemble of SVM and sample-weighted random forests (SSWRF) was proposed to deal with class imbalance. An SVM classifier was trained and applied to estimate the weights of training samples. Then, the training samples with estimated weights were utilized to train a sample-weighted random forests (SWRF). In addition, a lower-dimensional feature representation method, which consists of evolutionary conservation, hydrophobic property, solvent accessibility features derived from a target residue and its neighbors, was developed to improve the discriminative capability for PPI sites prediction. The analysis of feature importance shows that the proposed feature representation method is an effective representation for predicting PPI sites. The proposed SSWRF achieved 22.4% and 35.1% in MCC and F-measure, respectively, on independent validation dataset Dtestset72, and achieved 15.2% and 36.5% in MCC and F-measure, respectively, on PDBtestset164. Computational comparisons between existing PPI sites predictors on benchmark datasets demonstrated that the proposed SSWRF is effective for PPI sites prediction and outperforms the state-of-the-art sequence-based method (i.e., LORIS) released most recently. The benchmark datasets used in this study and the source codes of the proposed method are publicly available at http://csbio.njust.edu.cn/bioinf/SSWRF for academic use.