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

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Accurate compound-protein interaction (CPI) prediction is vital for drug discovery.
• Existing methods struggle with generalizability, particularly for new compounds or proteins (cold-start problem).

Purpose of the Study:

• To develop a robust deep learning framework, MM-CPI, for accurate CPI prediction.
• To improve generalizability and performance in cold-start scenarios.

Main Methods:

• MM-CPI integrates multimodal compound features (fingerprints, language models, GCN) and protein features (sequence, physicochemical properties, language models).
• A deep learning approach with effective feature fusion and classification is employed.

Main Results:

• MM-CPI demonstrates outstanding predictive performance across multiple datasets.
• Significantly outperforms state-of-the-art methods in accuracy, robustness, and generalization.
• Identified potential drug candidates for Fabry disease through drug repurposing, validated by molecular docking.

Conclusions:

• MM-CPI offers a powerful tool for accurate CPI prediction, addressing limitations of current computational methods.
• The framework shows significant promise for accelerating drug discovery and repurposing, particularly in challenging cold-start scenarios.
• An interactive web platform and open-source code are available for practical application.