ChiGNN: Interpretable Algorithm Framework of Molecular Chiral Knowledge-Embedding and Stereosensitive Property...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Molecular chirality presents significant challenges in materials machine learning (ML) due to subtle enantiomeric differences.
• Effective steric molecular descriptions and chiral knowledge integration are crucial for improving ML model accuracy and interpretability.

Purpose of the Study:

• To develop a deep learning framework for enhanced molecular chirality tasks.
• To improve the accuracy and interpretability of machine learning models in predicting chiral properties.

Main Methods:

• Proposed a Chiral Graph Neural Network (CGNN) incorporating chiral physicochemical knowledge.
• Utilized Trinity Graph and stereosensitive Message Aggregation encoding.
• Combined with quantile regression for retention time prediction.

Main Results:

• Achieved state-of-the-art accuracy in chiral chromatographic retention time prediction.
• Developed Trinity Mask and Contribution Splitting for multilevel model interpretation (atomic, functional group, molecular levels).

Conclusions:

• The CGNN framework offers significant advancements in handling molecular chirality in ML.
• The multilevel interpretation provides scientific and practical insights into chiral chromatography and stationary phase selection.
• The framework serves as an extensible template for future stereosensitive ML tasks.