MVR-DTI: A Multimodal Molecular Visual Representation Learning for Drug–Target Interaction Prediction

Drug-target interaction (DTI) prediction is a pivotal step in modern drug discovery. However, most existing DTI prediction methods still exhibit limited capability in representing multiple molecular modalities, and the application of molecular visual representation learning to DTI prediction remains largely unexplored. To address these challenges, we propose a Multimodal Molecular Visual Representation for DTI prediction (termed MVR-DTI), which integrates molecular visual representations into multimodal representation learning. Specifically, MVR-DTI employs a vision transformer to extract structure-aware visual features that capture molecular spatial information. Furthermore, these visual features are further aligned with traditional molecular descriptors, protein sequences, and knowledge graph information embeddings through a cross-modal fusion module based on contrastive learning and attention mechanisms. Experimental results show that MVR-DTI consistently outperforms competitive baselines across multiple evaluation metrics.These findings highlight the potential of multimodal visual representation learning for improving DTI prediction and advancing the computational drug discovery.