Graph-DTI: A new Model for Drug-target Interaction Prediction Based on Heterogenous Network Graph Embedding

Abstract Background Accurate prediction of drug-target interactions (DTIs) can guide the drug discovery process and thus facilitate drug development. Most existing computational models for machine learning tend to focus on integrating multiple data sources and combining them with popular embedding methods. However, researchers have paid less attention to the correlation between drugs and target proteins. In addition, recent studies have employed heterogeneous network graphs for DTI prediction, but there are limitations in obtaining rich neighborhood information among nodes in heterogeneous network graphs. Results Inspired by recent years of graph embedding and knowledge representation learning, we develop a new end-to-end learning model, called Graph-DTI, which integrates various information from heterogeneous network data and automatically learns topology-preserving representations of drugs and targets to facilitate DTI prediction. Our framework consists of three main building blocks. First, we integrate multiple data sources of drugs and target proteins and build a heterogeneous network from a collection of datasets. Second, the heterogeneous network is formed by extracting higher-order structural information using a GCN-inspired graph autoencoder to learn the nodes (drugs, proteins) and their topological neighborhood representations. The last part is to predict the potential DTIs and then send the trained samples to the classifier for binary classification. Conclusions The substantial improvement in prediction performance compared to other baseline DTI prediction methods demonstrates the superior predictive power of Graph-DTI. Moreover, the proposed framework has been successful in ranking drugs corresponding to different targets and vice versa. All these results suggest that Graph-DTI can provide a powerful tool for drug research, development and repositioning.