A semi-supervised method for drug-target interaction prediction with consistency in networks.

Computational prediction of interactions between drugs and their target proteins is of great importance for drug discovery and design. The difficulties of developing computational methods for the prediction of such potential interactions lie in the rarity of known drug-protein interactions and no experimentally verified negative drug-target interaction sample. Furthermore, target proteins need also to be predicted for some new drugs without any known target interaction information. In this paper, a semi-supervised learning method NetCBP is presented to address this problem by using labeled and unlabeled interaction information. Assuming coherent interactions between the drugs ranked by their relevance to a query drug, and the target proteins ranked by their relevance to the hidden target proteins of the query drug, we formulate a learning framework maximizing the rank coherence with respect to the known drug-target interactions. When applied to four classes of important drug-target interaction networks, our method improves previous methods in terms of cross-validation and some strongly predicted interactions are confirmed by the publicly accessible drug target databases, which indicates the usefulness of our method. Finally, a comprehensive prediction of drug-target interactions enables us to suggest many new potential drug-target interactions for further studies.