A Graph Convolution-Transformer Neural Network for Drug-Target Interaction Prediction

Identifying the ligand-binding affinity toward a target is crucial to ensure the drug s effects. The wet experiment is demanding when measuring massive drug-target interactions (DTI). Thus, many researchers applied machine learning to accelerate drug design. This study practised transformer and graph convolution neural network (GCN) to predict DTI. The molecules in the neural networkwere represented as graphs and then updated features by convolution. The Self-attention machine, transformer, can seek connections between source and target subjects. Thus, they were used to learn unique features of drugs and targets and then mathematically regress the interactions. The GCN-Transformer was trained on widely used benchmark DTI datasets. The best model classification performance was approximately 0.85 AUC-ROC and 0.86 AUPR. However, for the biased kinome selectivity dataset, the model s AUPR dropped to 0.75. The attention weights were also visualised, highlighting the compounds. The focused atoms by the weights contribute to the critical binding chemical bonds and conformation stability, showing reasonable explainability. This model might help insight into the DTI mechanism.