Drug-Protein-Disease Association Prediction and Drug Repositioning Based on Tensor Decomposition

The old paradigm "one gene, one drug, one disease" of drug discovery is challenged in many cases, where many drugs act on multiple targets and diseases rather than only one. Drug repositioning, which aims to discover new indications of known drugs, is a useful and economical strategy for drug discovery. It is also important to identify the functional clustering of target proteins, drugs and diseases, and to understand the pathological reasons for their interactions among these clusters and individuals. In this study, we propose a novel computational method to predict potential associations among drugs, proteins and diseases based on tensor decomposition. First, we collect pairwise associations between drugs, proteins and diseases, and integrate them into a three-dimensional tensor, representing the drug-protein-disease triplet associations. Then, we carry out tensor decomposition on the association tensor together with some additional information, and get three factor matrices of drugs, proteins and diseases respectively. Finally, we reconstruct the association tensor by the factor matrices to derive new predictions of triplet associations. We compare our method with some baseline methods and find our method outperforming the others. We validate our top ranked predictions by literature search and computational docking. In addition, we cluster the drugs, proteins and diseases using the factor matrices, which reflect the functional patterns of the drugs, proteins and diseases. Comparing our clustering to existing classifications/clusters, we find some agreement between them and that the factor matrices indeed reflect the functional patterns.