Fragment-Aware Contrastive Learning Framework for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Artificial intelligence (AI) accelerates drug discovery through molecular property prediction.
• Self-supervised learning (SSL), particularly contrastive learning, is used for molecular representation learning (MRL).
• Challenges exist in creating meaningful contrastive pairs due to molecular structural integrity, leading to suboptimal representations.

Purpose of the Study:

• To develop a novel contrastive pair construction strategy for improved MRL.
• To enhance the quality of molecular embeddings without external prior knowledge.
• To improve downstream task performance in drug discovery.

Main Methods:

• Introduced a contrastive pair construction strategy based on molecular fragment contributions.
• Utilized information bottleneck theory to evaluate fragment importance.
• Implemented a contrastive learning framework with an improved quadruplet loss function.

Main Results:

• Achieved outstanding performance on the MoleculeNet benchmark.
• Demonstrated promising results in predicting pharmacokinetic (PK) and toxicity properties.
• Learned a higher-quality embedding space by effectively capturing fine-grained molecular similarities.

Conclusions:

• The novel fragment-based contrastive strategy significantly improves MRL for drug discovery.
• The enhanced framework shows potential for real-world applications in predicting critical molecular properties.
• This approach addresses limitations in current SSL methods for molecular data.