Abstract <p>Deep learning technology has become a pivotal tool in the domain of computer-aided drug design and shows promising prospects in pharmaceutical research. Among the numerous tasks involved in drug development, drug–target affinity (DTA) prediction is of particular importance, since it expedites drug discovery and reduces resource consumption. With the fast progress of deep learning techniques, applying these techniques to enhance DTA prediction accuracy has emerged as a prominent research focus. Drugs and protein targets can be described using various representations, such as approaches grounded in structure, sequences, and graphs. We leverage the structural features of both compound molecules and proteins to generate graph representations that encode drug molecules and proteins, respectively. Building upon recent developments in graph neural networks, we employ a graph Transformer framework to extract graph representation information. In addition, we integrate SMILES based compound descriptions with textual protein features to propose a model named GraphTransDTA for effective DTA prediction. This approach fully utilizes the structural features of compound molecules and proteins, adeptly handles multi-feature nodes in heterogeneous graphs through the graph Transformer. Node level information aggregation is achieved via the multi-head attention mechanism. Moreover, the integration of multimodal data including textual features from both SMILES and protein sequences further improves the predictive accuracy and robustness of DTA models. By retaining the expressive power of deep neural networks and exploiting the integrative nature of various data modalities, the proposed method offers a novel perspective for accurate drug discovery.</p> Graphical Abstract <p></p>