An improved graph representation learning method for drug-target interaction prediction over heterogeneous biological information graph

The prediction task of the relationships between drugs and targets plays a significant role in the process of new drug discovery. Computational-based strategies for predicting drug-target interactions (DTIs) are regarded as a high-efficiency way. Moreover, recent studies adopted a graph neural network (GNN) to discover underlying DTIs and achieved better performance. Although these inductive methods can straightway learn biomolecules latent representations, they have an over-smoothing phenomenon in the course of obtaining the rich neighborhood information of each node in the biological information network, which further leads to a consistent feature representation of each node. To address the above issues, a novel model, called iGRLDTI, is proposed to precisely identify new DTIs based on an improved graph representation learning strategy. Specifically, iGRLDTI first constructs a biological information graph (BIG) by calculating the biological knowledge of drugs and targets with the relationships between them. Then, an improved graph representation learning strategy is designed to capture the enriched feature representations of drugs and targets. Finally, the Gradient Boosting Decision Tree classifier is applied to predict potential DTIs. Experimental results demonstrate that iGRLDTI yields better performance by comparing it with other state-of-the-art models on the benchmark dataset. Besides, our case studies denote that iGRLDTI can successfully identify unknown DTIs according to the improved feature representations of drugs and targets.