MolMVC: Enhancing molecular representations for drug-related tasks through multi-view contrastive learning

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Effective molecular representation is crucial for drug development.
• Molecules require comprehensive multi-view representations (1D, 2D, 3D) for a holistic understanding.
• Existing methods may not fully capture the diverse structural aspects of molecules.

Purpose of the Study:

• To introduce MolMVC, an innovative multi-view contrastive learning framework for molecular representation.
• To develop a method that effectively integrates 1D, 2D, and 3D molecular information.
• To enhance downstream task performance in drug discovery.

Main Methods:

• Utilized a Transformer encoder for 1D sequence information and a Graph Transformer for 2D/3D structures.
• Incorporated an attention-guided augmentation scheme for tailored positive sample generation.
• Introduced an adaptive multi-view contrastive loss (AMCLoss) to align multi-view representations in latent space.
• Calculated AMCLoss at various hierarchical levels to capture molecular information intricacies.

Main Results:

• MolMVC demonstrated enhanced predictive accuracy in molecular property prediction (MPP), drug-target binding affinity (DTA) prediction, and cancer drug response (CDR) prediction.
• The framework reduced computational costs for these tasks.
• MolMVC showed efficacy in drug repositioning applications.

Conclusions:

• MolMVC provides a powerful approach for learning comprehensive molecular representations.
• The framework effectively integrates multi-view molecular data for improved drug discovery outcomes.
• The learned representations are versatile and applicable to a range of drug-related tasks.