GATv2-TransDTI: A graph and sequence hybrid model for fine-grained drug-target interaction prediction

Accurately predicting drug-target interaction (DTI) is critical for drug discovery and development. Existing methods typically rely on the atomic structure or molecular graphs of drugs, along with amino acid sequences of target proteins, to extract features using graph neural networks or sequence modeling techniques, followed by binary classification. However, these methods often ignore the fine-grained local interactions between drug substructures and protein binding sites, and generally lack the ability to interpret the prediction results at the biological or pharmacological level. In this study, we propose GATv2-TransDTI, a novel multi-module fusion model designed to enhance the modeling of both local and global interaction features. The model utilizes the GATv2 graphical attention mechanism to encode drugs to capture key drug molecule substructure information. Protein sequence encoding, on the other hand, uses a module that mixes convolution with a self-constructed Transformer, thus extracting local motifs and long-range dependencies. For deeper integration of drug and protein features, we employ a multi-head fusion mechanism that allows the model to focus on the most informative local interaction regions through an attentional mechanism, thus enabling accurate learning of fine-grained interactions between drug targets. Extensive experiments are performed on BindingDB, BioSNAP and Human datasets. The results show that GATv2-TransDTI outperforms the comparative state-of-the-art models. The AUROC scores of GATv2-TransDTI are 0.967 and 0.919, and the AUPRC scores are 0.954 and 0.923, respectively.