A Novel Computational Approach for Predicting Drug-Target Interactions via Network Representation Learning

Detection of drug-target interactions (DTIs) has a beneficial effect on both pathogenesis and drugs discovery. Although a huge number of DTIs have been generated recently, the number of known interactions is still very small. Thus, it is strongly needed to develop computational methods to accurately and effectively predict DTIs. In this paper, a large-scale computational method is proposed to predict potential DTIs via network representation learning. More specifically, known associations among drugs, proteins, miRNA and disease are formulated as a biomolecular association network, and the network representation method Structural Deep Network Embedding (SDNE) is used to extract network-based features of drug and target nodes. Then, the fingerprints of drug compounds and sequence information of proteins are also adopted. Finally, an ensemble Random Forest classifier is used to classify and predict DTIs. Experiment results show that the proposed method achieved a good prediction performance with an accuracy of 83.68% and AUC of 0.9052. It is anticipated that proposed model is feasible and effective to predict DTIs at a global molecule level, which is a new respective for future biomedical researches.