Deep contrastive learning of molecular conformation for efficient property prediction

Area of Science:

• Computational Chemistry
• Machine Learning
• Deep Learning

Background:

• Data-driven deep learning models accurately predict molecular properties but are limited by geometric relaxation constraints.
• Domain-shift issues arise when using cost-effective conformation generation methods, reducing prediction accuracy.

Purpose of the Study:

• To develop a domain-adaptation method to bridge the gap between different geometric conformations in molecular property prediction.
• To enable deep learning models to maintain accuracy across varying levels of geometric relaxation.

Main Methods:

• Proposed a deep contrastive learning-based domain-adaptation method named Local Atomic environment Contrastive Learning (LACL).
• LACL learns a domain-agnostic latent space by comparing different conformation-generation methods to capture local atomic environment semantics.

Main Results:

• LACL effectively alleviates distribution disparities between different geometric conformations.
• Achieved quantum-chemical accuracy without the geometric relaxation bottleneck.
• Demonstrated generalizability across small organic molecules, biological chains, and pharmacological molecules.

Conclusions:

• LACL enables accurate molecular property prediction independent of geometric relaxation constraints.
• The method facilitates broader applications like inverse molecular engineering and large-scale screening.
• LACL offers a generalizable solution for diverse molecular structures in computational chemistry.