Predicting Drug–Gene–Disease Associations by Tensor Decomposition for Network-Based Computational Drug Repositioning

Abstract

Drug repositioning offers the significant advantage of greatly reducing the cost and time of drug discovery by identifying new therapeutic indications for existing drugs. In particular, computational approaches using networks in drug repositioning have attracted attention for inferring potential associations between drugs and diseases efficiently based on the network connectivity. In this article, we proposed a network-based drug repositioning method to construct a drug–gene–disease tensor by integrating drug–disease, drug–gene, and disease–gene associations and predict drug–gene–disease triple associations through tensor decomposition. The proposed method, which ensembles generalized tensor decomposition (GTD) and multi-layer perceptron (MLP), models drug–gene–disease associations through GTD and learns the features of drugs, genes, and diseases through MLP, providing more flexibility and non-linearity than conventional tensor decomposition. We experimented with drug–gene–disease association prediction using two distinct networks created by chemical structures and ATC codes as drug features. Moreover, we leveraged drug, gene, and disease latent vectors obtained from the predicted triple associations to predict drug–disease, drug–gene, and disease–gene pairwise associations. Our experimental results revealed that the proposed ensemble method was superior for triple association prediction. The ensemble model achieved an AUC of 0.96 in predicting triple associations for new drugs, resulting in an approximately 7% improvement over the performance of existing models. It also showed competitive accuracy for pairwise association prediction compared with previous methods. This study demonstrated that incorporating genetic information leads to notable advancements in drug repositioning.