Globalized bipartite local model for drug-target interaction prediction

In pharmacology, it is essential to identify the interactions between drug and targets to understand its effects. Supervised learning with Bipartite Local Model (BLM) recently has been shown to be effective for prediction of drug-target interactions by first predicting target proteins of a given known drug, then predicting drugs targeting a known protein. However, this pure "local" model is inapplicable to new drug or target candidates that currently have no known interactions. In this paper, we extend the existing BLM method by integrating a strategy for handling new drug and target candidates. Based on the assumption that similar drugs and targets have similar interaction profiles, we present a simple neighbor-based training data inferring procedure and integrate it into the frame work of BLM. This globalized BLM called bipartite local model with neighbor-based inferring (BLMN) then has an extended functionality for prediction interactions between new drug candidates and new target candidates. Good performance of BLMN has been observed in the experiment of predicting interactions between drugs and four important categories of targets. For the Nuclear Receptors dataset, where there are more chances for the presented strategy to be applied, 20% improvement in terms of AUPR was achieved. This demonstrates the effectiveness of BLMN and its potential in prediction of drug-target interactions.