Wendan decoction is one of the classic prescriptions of Traditional Chinese medicine (TCM). It is widely used in the treatment of the diseases of cardiovascular and cerebrovascular, digestive system, mental and nervous system, but its modern clinical position is not clear. Recently, TCM has been receiving more and more attention due to its unique pathological characteristics and its potential as a supplement or substitute for modern medicine. However, the complex composition of TCM makes it very difficult to study its pharmacology. The identification and prediction of drug-target interaction (DTI) plays a vital role in various areas of drug development, and plays a positive role in pharmacological research. While traditional DTI prediction methods still have various limitations, recently, prediction methods based on knowledge graph have begun to enter the vision of researchers. However, there are few TCM researches using knowledge mapping at present. In this study, we constructed the Knowledge graph of Wendan Decoction (WDKG) by using multiple data sources, and proposed a DTI prediction framework, SGRN-Trans, based on knowledge graph and attention model. The prediction framework firstly uses graph neural network (GNN) to obtain the low-dimensional representation of each entity in the knowledge graph, then introduces the structural features of drugs and targets, finally integrates the attention model, Transformer, to predict DTI. We evaluate each part of SGRN-Trans in detail, and compare it with other kinds of high-performance algorithms. Experimental results show that SGRN-Trans has significantly superior performance in DTI prediction tasks. All in all, this study is a valuable attempt to combine knowledge graph and attention model and apply it to DTI prediction, which provides a new insight for TCM research.
 

Knowledge graph；TCM；DTI； GNN；Transformer