A Drug-Target Interaction Prediction Method Based on Attention Perception and Modality Fusion

[Purposes] Predicting drug-target interactions is a critical step in drug discovery. The existing prediction methods based on graph neural networks have achieved good results, but there are still two challenges: first, how to extract the deep features and rich semantic information in drugs and targets more efficiently; and second, how to explicitly model and learn the interactions between drugs and targets for better prediction and interpretation. In order to solve the above problems, a drug-target interaction prediction method based on attention perception and modality fusion is proposed in this paper. [Methods] For drug branches, Graph Transformer and Graph Convolutional Neural Network were used to jointly characterize the global structures and biochemical information of drug molecules. For protein branches, a separate encoding module was employed to extract the global and local features of the protein. In order to model the relationship between drugs and targets more accurately, a Transformer-based module was designed to fuse drug and target features. [Results] Finally, the experimental results on the three public datasets show that the main evaluation indicators have been improved, and the method can predict the interaction between drugs and targets more accurately.