A meta-learning approach for selectivity prediction in asymmetric catalysis

Area of Science:

• Organic Chemistry
• Catalysis
• Machine Learning

Background:

• Asymmetric catalysis is crucial for synthesizing chiral molecules.
• Machine learning (ML) accelerates catalyst development but requires extensive data.
• Data scarcity hinders ML application in discovering new catalytic protocols.

Purpose of the Study:

• To develop a meta-learning workflow for predicting reaction outcomes with limited data.
• To leverage literature-derived data for feature extraction in catalysis.
• To improve the efficiency of identifying promising new reactions.

Main Methods:

• A meta-learning workflow utilizing prototypical networks was designed.
• The model predicts enantioselectivity in asymmetric hydrogenation of olefins.
• Literature data was mined to extract shared reaction features.

Main Results:

• The meta-learning model outperformed random forests and graph neural networks.
• Performance was validated across varying training dataset sizes, showing utility with limited data.
• Strong performance on an out-of-sample test set confirmed general applicability.

Conclusions:

• Meta-learning offers a powerful solution for data-limited reaction development.
• This approach accelerates the discovery of novel transition metal-catalyzed reactions.
• The workflow enables efficient prediction of enantioselectivity in early-stage research.