Multi-modal contrastive learning based on molecular and textual data for drug response prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in medicine

Background:

• Drug response prediction (DRP) is vital for precision medicine.
• Current deep learning DRP methods neglect biomedical knowledge from text and inadequately integrate feature representations.
• Limitations hinder optimization of therapeutic strategies.

Purpose of the Study:

• To develop a multi-modal contrastive learning method for enhanced DRP.
• To integrate complementary biological insights from textual corpora.
• To effectively combine attribute-based and structure-based molecular representations.

Main Methods:

• Incorporated textual descriptions of biological pathways and mechanisms as a complementary modality.
• Extracted attribute-based representations from multi-omics profiles and drug graphs.
• Captured structure-based representations using heterogeneous graph neural networks for drug-cell interactions.
• Employed a contrastive learning module for multi-modal representation alignment.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art methods on GDSC and CCLE datasets.
• Successfully integrated diverse data modalities for improved DRP accuracy.
• Established a unified semantic space for enhanced feature integration.

Conclusions:

• The multi-modal contrastive learning approach significantly advances drug response prediction.
• Integrating biomedical knowledge and dual-view molecular features is effective for DRP.
• The method offers a promising direction for precision medicine and therapeutic strategy optimization.