MMCL-CPI: A multi-modal compound-protein interaction prediction model incorporating contrastive learning pre-training

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Compound-protein interaction (CPI) prediction is vital for drug discovery.
• Traditional experimental methods for CPI are slow and costly.
• Deep learning has accelerated CPI prediction, but multimodal approaches are lacking.

Purpose of the Study:

• To develop a novel multimodal deep learning method for CPI prediction.
• To effectively combine compound sequence and image representations.
• To improve the accuracy and efficiency of CPI prediction.

Main Methods:

• Extracting compound features from both 1D SMILES (sequence) and 2D images (spatial).
• Designing a novel multimodal model integrating these diverse features.
• Implementing a multimodal pre-training strategy using contrastive learning on large unlabeled datasets.

Main Results:

• The proposed MMCL-CPI method demonstrated competitive performance on multiple datasets.
• Integrating sequence and image features enhanced compound representation for CPI tasks.
• The multimodal pre-training strategy significantly improved downstream CPI prediction accuracy.

Conclusions:

• MMCL-CPI offers a powerful new approach for accurate CPI prediction.
• Combining multimodal data sources (SMILES and images) is effective for enhancing deep learning models.
• This work paves the way for more efficient drug discovery and repositioning.