MM-CPI: A multimodal fusion framework for compound-protein interaction prediction

Accurate prediction of compound-protein interactions (CPI) is critical for accelerating drug discovery and repurposing. However, existing computational methods often suffer from limited predictive accuracy and generalizability, particularly in cold-start scenarios involving unseen compounds or proteins. To address this challenge, we propose MM-CPI, a multimodal deep learning framework that integrates comprehensive compound representations, including molecular fingerprints, language model-based representation, and Graph Convolutional Networks (GCN)-derived structural features, with rich protein features derived from sequence, physicochemical properties, and protein language models. MM-CPI learns discriminative interaction patterns through effective feature fusion and classification, enabling robust performance on unseen entities. Experimental results on multiple public datasets demonstrate that MM-CPI achieves outstanding predictive performance, significantly outperforming state-of-the-art methods in accuracy, robustness, and generalization. In a case study on Fabry disease, we applied MM-CPI to drug repurposing and identified several candidate drugs, which were further validated via molecular docking. To facilitate practical use, we developed an interactive web platform (available at http://1.94.102.146:8501/MM-CPI) and released all source code and dataset at https://github.com/flyinsky6/MM-CPI.