The Prediction of Protein-Protein Interaction Sites Based on RBF Classifier Improved by SMOTE

Protein-protein interaction sites are the basis of biomolecule interactions, which are widely used in drug target identification and new drug discovery. Traditional site predictors of protein-protein interaction mostly based on unbalanced datasets, the classification results tend to negative class, resulting in a lower predictive accuracy for positive class. A method called RBFIS (radial basis function improved by SMOTE) is presented in the paper to address the problem. The intelligent algorithm SMOTE is used to artificially synthesize the imbalanced datasets of negative sample classes. Simultaneously, KNN algorithm is utilized to interpolate values between the minority class samples to generate new samples, making the sample data tend to balance as much as possible. Then, RBF classifier is used to construct the site predictor of protein-protein interaction based on the processed quasi-equilibrium sample sets. The results of experiments indicated that the method had an improvement on recall and f-measure of positive class compared with traditional methods by 12% and 25%. Moreover, many rounds of experiments were performed for different combinations of features. It was observed that the key combination of different multiple features can better efficiently improve the prediction performance. In conclusion, the studies we have performed show that the proposed method is better for dealing with the imbalanced protein interaction sites.