Predicting unknown interactions between known drugs and targets via matrix completion

Drug-target interactions map patterns, associations and relationships between drugs and target proteins. Identifying interactions between drug and target is critical in drug discovery, but biochemically validating these interactions are both laborious and expensive. In this paper, we propose a novel interaction profiles based method to predict potential drug-target interactions by using matrix completion. Our method first arranges the drug-target interactions in a matrix, whose entries include interaction pairs, non-interaction pairs and undetermined pairs, and finds its approximation matrix which contains the predicted values at undetermined positions. Then our method learns an approximation matrix by minimizing the distance between the drug-target interaction matrix and its approximation subject that the values in the observed positions equal to the known interactions at the corresponding positions. As a consequence, our method can directly predict new potential interactions according to the high values at the undetermined positions. We evaluated our method by comparing against five counterpart methods on “gold standard” datasets. Our method outperforms the counterparts, and achieves high AUC and F₁-score on enzyme, ion channel, GPCR, nuclear receptor and integrated datasets, respectively. We showed the intelligibility of our method by validating some predicted interactions in both DrugBank and KEGG databases.