Predicting Enantioselectivity of Ruthenium-Catalyzed Ketone Hydrogenation with 3D Structure-Based Deep Learning

Area of Science:

• Catalysis
• Machine Learning
• Computational Chemistry

Background:

• Predicting enantioselectivity in catalytic hydrogenation is crucial for synthesizing chiral molecules.
• Existing predictive models often struggle with accuracy and generalization due to data limitations and representation choices.

Purpose of the Study:

• To develop highly accurate machine learning models for predicting the enantioselectivity of ketone hydrogenation reactions.
• To evaluate the effectiveness of 3D graph neural networks compared to other modeling approaches.

Main Methods:

• Combined literature data with over 1,000 high-throughput experimental results to create a robust dataset.
• Developed and trained 3D graph neural networks using representations of the enantiodetermining transition state.
• Compared performance against models using 2D graphs, molecular fingerprints, and quantum mechanical descriptors.

Main Results:

• 3D graph neural networks significantly outperformed other methods in predicting enantioselectivity.
• Models trained on 3D representations demonstrated strong extrapolation capabilities to novel catalyst-substrate structures.
• The study mitigated bias by integrating curated literature data with extensive experimental data.

Conclusions:

• Three-dimensional structural representations capturing catalyst-substrate interactions are key for accurate enantioselectivity prediction.
• Machine learning models, particularly 3D graph neural networks, offer a powerful tool for catalyst design and optimization.
• This approach enables reliable prediction of enantioselectivity, facilitating the development of efficient asymmetric hydrogenation processes.