Information-theoretic multi-scale geometric pre-training for enhanced molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning
• Nanotechnology

Background:

• Effective molecular representation learning requires maximizing information transfer across structural scales.
• Current graph neural networks struggle with multi-scale molecular geometry, hindering information propagation.
• This limits the ability to capture both local and global structural features.

Purpose of the Study:

• To introduce Multi-Scale Geometric Pre-training (MSG-Pre), an information-theoretic framework for molecular representation learning.
• To address the limitations of existing methods in capturing multi-scale molecular information.
• To enhance molecular understanding and prediction capabilities for drug discovery and nanomaterial design.

Main Methods:

• Developed an information-theoretic framework (MSG-Pre) integrating atomic, functional group, and conformer levels.
• Employed entropy-guided mechanisms, including scale-adaptive attention and hierarchical contrastive learning.
• Utilized a geometric regularization strategy to preserve conformational properties.

Main Results:

• Achieved state-of-the-art performance on 14 molecular benchmark datasets, with improvements up to 5.2%.
• Significantly enhanced information extraction for nanomedicine applications like nanoparticle-protein interactions.
• Demonstrated effective maximization of cross-scale mutual information and minimization of intra-scale redundancy.

Conclusions:

• MSG-Pre establishes an information-theoretic foundation for geometric pre-training.
• The framework optimizes information-entropy balance in molecular representations.
• MSG-Pre improves molecular understanding and prediction for drug discovery and nanomaterial design.