Improving Protein–Protein Interaction Prediction Based on Phylogenetic Information Using a Least‐Squares Support Vector Machine

Abstract:  Predicting protein–protein interactions has become a key step of reverse‐engineering biological networks to better understand cellular functions. The experimental methods in determining protein–protein interactions are time‐consuming and costly, which has motivated vigorous development of computational approaches for predicting protein–protein interactions. A set of recently developed bioinformatics methods utilizes coevolutionary information of the interacting partners (e.g., as exhibited in the form of correlations between distance matrices, where, for each protein, a matrix stores the pairwise distances between the protein and its orthologs in a group of reference genomes). We proposed a novel method to account for the intra‐matrix correlations in improving predictive accuracy. The distance matrices for a pair of proteins are transformed and concatenated into a phylogenetic vector. A least‐squares support vector machine is trained and tested on pairs of proteins, represented as phylogenetic vectors, whose interactions are known. The intra‐matrix correlations are accounted for by introducing a weighted linear kernel, which determines the dot product of two phylogenetic vectors. The performance, measured as receiver operator characteristic (ROC) score in cross‐validation experiments, shows significant improvement of our method (ROC score 0.928) over that obtained by Pearson correlations (0.659).