Supervised prediction of drug-target interactions using bipartite local models

Abstract

Motivation: In silico prediction of drug-target interactions from heterogeneous biological data is critical in the search for drugs for known diseases. This problem is currently being attacked from many different points of view, a strong indication of its current importance. Precisely, being able to predict new drug-target interactions with both high precision and accuracy is the holy grail, a fundamental requirement for in silico methods to be useful in a biological setting. This, however, remains extremely challenging due to, amongst other things, the rarity of known drug-target interactions.

Results: We propose a novel supervised inference method to predict unknown drug-target interactions, represented as a bipartite graph. We use this method, known as bipartite local models to first predict target proteins of a given drug, then to predict drugs targeting a given protein. This gives two independent predictions for each putative drug-target interaction, which we show can be combined to give a definitive prediction for each interaction. We demonstrate the excellent performance of the proposed method in the prediction of four classes of drug-target interaction networks involving enzymes, ion channels, G protein-coupled receptors (GPCRs) and nuclear receptors in human. This enables us to suggest a number of new potential drug-target interactions.

Availability: An implementation of the proposed algorithm is available upon request from the authors. Datasets and all prediction results are available at http://cbio.ensmp.fr/~yyamanishi/bipartitelocal/.

Fig. 1.

PR curves for predicted drug–target interactions using BLMs on four benchmark datasets: (a) enzyme, (b) ion channel, (c) GPCR and (d) nuclear receptor. The solid line is for leave-one-out on potential drugs (row 2 of Tables 1–4), the dashed line for leave-one-out on potential target proteins (row 5 of Tables 1–4) and the dotted line for aggregating the two scores for each putative drug–target interaction (row 8 of Tables 1–4). In the benchmark experiments (a), (c) and (d), the aggregated curve mimics or gives a significant improvement over the other two curves. For ion channels (b), leave-one-out on potential target proteins (dashed line) perform slightly better overall than aggregation (dotted line), but both curves represent extremely strong results.

Fig. 2.

Part of the predicted interaction network for the nuclear receptor data. Circles indicate drugs and squares target proteins. Solid edges represent known interactions and dashed ones show some of the 20 highest scoring predicted interactions. Dashed edges with asterisks represent compound–protein interactions now annotated in the SuperTarget database or confirmed in the literature.