Abstract
Protein-protein interaction hot spots, as revealed by alanine scanning mutagenesis, make dominant contributions to the free energy of binding. Since mutagenesis experiments are expensive and time-consuming, the development of computational methods to identify hot spots is becoming increasingly important. In this study, by using a new combination of sequence, structure and energy features, we propose an iterative semi-supervised algorithm, SemiHS, to incorporate unlabeled data to improve the accuracy of hot spots prediction when sufficient training data is un-available and to overcome the imbalanced data problem. We evaluate the predictive power of SemiHS on a labeled set of 265 alaninemutated interface residues in 17 complexes and a large unlabeled set of 2465 interface residues with 10-fold cross validation, and get an AUC score of 0.85, with a sensitivity of 0.70 and a specificity of 0.87, which are better than those of the existing methods. Moreover, we validate the proposed method by an independent test and obtain encouraging results.
Keywords: protein-protein interaction, hot spots, semi-supervised, SVM, apoptosis, protein engineering, Systematic mutagenesis, protein interfaces, double water exclusion, desolvation energy, Semi-Supervised Learning, SemiHS, jackknife test, Human Growth Hormone, ML methods, mutagenesis, Bayes networkprotein-protein interaction, hot spots, semi-supervised, SVM, apoptosis, protein engineering, Systematic mutagenesis, protein interfaces, double water exclusion, desolvation energy, Semi-Supervised Learning, SemiHS, jackknife test, Human Growth Hormone, ML methods, mutagenesis, Bayes network