Abstract

In-silico prediction of interactions between drugs and proteins has become a crucial step in pharmaceutical sciences to reduce the time and cost required for drug discovery and repositioning. Even if the problem may be approached using standard recommendation algorithms, the accurate prediction of unknown drug-target interactions has shown to be very challenging due to the relatively small number of drugs with information of their target proteins and viceversa. This issue has been recently circumvent using regularization methods that actively exploit prior knowledge regarding drug similarities and target similarities. In this paper, we show that an additional improvement in terms of accuracy can be obtained using an ensemble approach which learns to combine multiple regularized filters for prediction. Our experiments on eight drug-protein interaction datasets show that most of the time this method outperforms a single predictor and other recommender systems based on multiple filters but not specialized to the drug-target interaction prediction task.