Energy-Guided Denoising Contrastive Learning for Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Self-supervised learning for 3D molecular representations is promising for property prediction.
• Existing methods using random perturbations lack chemical environment awareness and can produce unrealistic structures.

Purpose of the Study:

• To develop an improved self-supervised framework for learning 3D molecular representations.
• To enhance the physical realism and predictive accuracy of molecular property prediction models.

Main Methods:

• Proposed an energy-guided denoising contrastive learning framework with an adaptive noise generator.
• The generator creates structure-aware, atom-specific perturbations based on local chemical environments.
• Integrated equivariant denoising, contrastive alignment, and self-supervised energy-gap prediction.

Main Results:

• Achieved state-of-the-art performance on multiple molecular property prediction benchmarks.
• Demonstrated interpretability and successful learning of chemical-energy relationships in case studies.

Conclusions:

• The proposed energy-guided framework effectively learns meaningful 3D molecular representations.
• This approach advances self-supervised learning for accurate and physically realistic molecular modeling.