Prediction of hot spots in protein interfaces using extreme learning machines with the information of spatial neighbour residues

The identification of hot spots, a small subset of protein interfaces that accounts for the majority of binding free energy, is becoming increasingly important for the research on protein-protein interaction and drug design. For each interface residue or target residue to be predicted, the authors extract hybrid features which incorporate a wide range of information of the target residue and its spatial neighbor residues, that is, the nearest contact residue in the other face (mirror-contact residue) and the nearest contact residue in the same face (intra-contact residue). Here, feature selection is performed using random forests to avoid over-fitting. Thereafter, the extreme learning machine is employed to effectively integrate these hybrid features for predicting hot spots in protein interfaces. By the 5-fold cross validation in the training set, their method can achieve accuracy (ACC) of 82.1% and Matthew´s correlation coefficient (MCC) of 0.459, and outperforms some alternative machine learning methods in the comparison study. Furthermore, their method achieves ACC of 76.8% and MCC of 0.401 in the independent test set, and is more effective than the major existing hot spot predictors. Their prediction method offers a powerful tool for uncovering candidate residues in the studies of alanine scanning mutagenesis for functional protein interaction sites.