Manifold regularized matrix factorization for drug-drug interaction prediction.

Abstract

Drug-drug interaction (DDI) prediction is one of the most important tasks in drug discovery. Prediction of potential DDIs helps to reduce unexpected side effects in the lifecycle of drugs, and is important for the drug safety surveillance. Here, we formulate the drug-drug interaction prediction as a matrix completion task, and project drugs in the interaction space into a low-dimensional space. We consider drug features, i.e., substructures, targets, enzymes, transporters, pathways, indications, side effects, and off side effects, to calculate drug-drug similarities, and assume them as manifolds in feature spaces. In this paper, we present a novel computational method named "Manifold Regularized Matrix Factorization" (MRMF) to predict potential drug-drug interactions, by introducing the drug feature-based manifold regularization into the matrix factorization. In the computational experiments, the MRMF models, which utilize known drug-drug interactions and the drug feature-based manifold, produce the area under precision-recall curves (AUPR) up to 0.7963. We test manifold regularizations based on different drug features, and the MRMF models can produce robust performances. Compared with other state-of-the-art methods, the MRMF models can produce better performances in the cross validation and case study. The manifold regularization is the critical factor for the high-accuracy performances of our method. MRMF is promising and effective for the prediction of drug-drug interactions.