FGMDTI: Fine-grained Multimodal Learning for Drug-Target Interaction Prediction via Bidirectional Cross-Attention

Drug-target interaction (DTI) prediction is critical for drug discovery and repositioning. Existing methods often rely on coarse-grained representations and fail to capture the reciprocal nature of molecular recognition. We propose FGMDTI, a fine-grained multimodal framework that models bidirectional interactions between drugs and proteins represented as functional units, each encoded through a dual-view feature representation that integrates structural features from pre-trained foundation models and semantic features from large language model (LLM)-based text embeddings. Results on five benchmark datasets show that FGMDTI achieves the highest accuracy, precision, AUC, and AUPR among the compared methods on all five datasets, and the highest F1-score on three datasets. It also exhibits strong generalization under cold-start settings involving unseen drugs, targets, and interaction pairs. Visualization of cross-attention patterns shows that the model highlights interaction-relevant functional units, providing qualitative insights into the learned interaction patterns. These observations support the effectiveness and generalization capability of fine-grained multimodal representation learning for DTI prediction.