Machine Learning for Reaction Performance Prediction in Allylic Substitution Enhanced by Automatic Extraction of a...

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Machine learning (ML) shows promise in organic synthesis but struggles with accurate reaction outcome prediction.
• A key challenge is the development of efficient molecular descriptors that retain essential predictive information.

Purpose of the Study:

• To develop and validate a novel descriptor for predicting performance in iridium-catalyzed allylic substitution reactions.
• To address the limitations of existing descriptors in balancing predictive accuracy and computational efficiency.

Main Methods:

• Introduction of SubA, a substrate-aware descriptor incorporating graph matching and DFT-derived properties.
• Evaluation of SubA against four mainstream descriptors using random and scaffold splitting.
• Analysis of SubA's interpretability to understand its focus on reaction-driving features.

Main Results:

• SubA achieved reduced dimensionality and improved prediction accuracy, with over 2% mean absolute error reduction.
• The descriptor demonstrated superior generalization capabilities on novel substrate combinations.
• Interpretable analysis revealed SubA's focus on critical atomic and molecular features.

Conclusions:

• SubA represents a significant advancement in descriptor development for ML-driven organic synthesis.
• The substrate-aware approach enhances prediction accuracy and interpretability in catalytic reaction modeling.
• This work facilitates more reliable ML applications for predicting reaction outcomes and understanding mechanisms.