MolFCL: predicting molecular properties through chemistry-guided contrastive and prompt learning

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Accurate molecular property prediction is vital in molecular machine learning.
• Effective molecular representations are key to generalizing across the chemical space.
• Existing contrastive learning methods often neglect the original molecular environment and chemical prior information.

Purpose of the Study:

• To develop a novel molecular property prediction framework, MolFCL.
• To address limitations in existing contrastive learning models by preserving molecular environments and incorporating prior chemical knowledge.
• To improve the accuracy and interpretability of molecular property predictions.

Main Methods:

• Proposed a fragment-based contrastive learning approach incorporating fragment-fragment interactions.
• Designed a fragment-based augmented molecular graph to preserve the original chemical environment and reactions.
• Introduced functional group-based prompt learning during fine-tuning, integrating functional group knowledge and atomic signals.

Main Results:

• MolFCL outperformed state-of-the-art baseline models on 23 molecular property prediction datasets.
• Visualizations demonstrated MolFCL's ability to learn representations that distinguish chemical properties.
• The model exhibited interpretability by assigning higher weights to chemically relevant functional groups.

Conclusions:

• MolFCL offers a practically useful tool for molecular property prediction.
• The framework enhances molecular representations and prediction accuracy.
• MolFCL assists drug scientists in more effective drug design by providing interpretable insights.