KGAT: Predicting Drug-Target Interaction Based on Knowledge Graph Attention Network

Prediction of Drug-target interaction (DTI) is an important topic in bioinformatics which plays an important role in the process of drug discovery. Although many machine learning methods have been successfully applied to DTI prediction, traditional approaches mostly utilize single chemical structure information or construct heterogeneous graphs that integrate multiple data sources for DTI prediction, while these methods ignore the interaction relationships among sample entities (e.g., drug-drug pairs). The knowledge graph attention network (KGAT) uses biomedical knowledge bases and entity interaction relationships to construct knowledge graph and transforms the DTI problem into a linkage prediction problem for nodes in the knowledge graph. KGAT distinguishes the importance of features by assigning attention weights to neighborhood nodes and learns vector representations by aggregating neighborhood nodes. Then feature vectors are fed into the prediction model for training, at the same time, the parameters of prediction model update by gradient descent. The experiment results show the effectiveness of KGAT.