A GCN-Transformer Framework with Cross-Attention for Drug–Target Interaction Prediction
Huaizhou Yang, Zhongzhou Li, Chunlin Zhou · 2025
Drug–target interaction (DTI) prediction faces challenges such as insufficient feature extraction and inefficient cross-modal interaction. This paper proposes GTA-DTI, a model that integrates GCN to capture atom-level topological features from drug molecular graphs and a Transformer–DCNN hybrid encoder to extract both global dependencies and local motif patterns from protein sequences—overcoming the limitations of unimodal representations. A bidirectional cross-attention mechanism is employed to dynamically align the semantic spaces of drugs and proteins, adaptively enhancing key interaction features and replacing traditional feature concatenation. On the Davis/KIBA datasets, GTA-DTI achieves accuracy scores of 0.912/0.917 and recall scores of 0.845/0.852, significantly outperforming mainstream models. Its F1-score (0.861) is within 1.4% of the best-performing method, and its AUC improves by 1.2% across datasets—demonstrating strong adaptability to heterogeneous data and offering an innovative solution for high-accuracy, low-miss DTI prediction.