Background <p>Predicting drug-protein interactions (DPI) is essential for effective and safe drug discovery. Although deep learning methods have been extensively applied to DPI prediction, effectively leveraging the multi-structural and multimodal data on drugs and proteins to enhance prediction accuracy remains a significant challenge.</p> Results <p>This study proposes CMMSCL-DPI, a cross-modal multi-structural contrastive learning model. CMMSCL-DPI applies contrastive learning to the multi-dimensional structural features of proteins and drugs separately and integrates interaction features from a DPI heterogeneous graph network to facilitate cross-modal contrastive learning. This approach effectively captures the key differences and similarities between proteins and drugs, significantly enhancing the model’s generalization capabilities for novel drug-target pairs. Experimental results across four benchmark datasets demonstrate that CMMSCL-DPI outperforms five state-of-the-art baseline models in overall performance. Additionally, the model successfully identified an unreported drug-protein interaction, which was subsequently validated through all-atom molecular dynamics simulations.</p> Conclusions <p>This case study not only confirms the predictive accuracy of CMMSCL-DPI but also underscores its potential in discovering novel protein–ligand interactions.</p>