Designing Buchwald-Hartwig Reaction Graph for Yield Prediction

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Organic Synthesis

Background:

• The Buchwald-Hartwig (B-H) reaction is a crucial cross-coupling reaction in organic synthesis.
• Predicting reaction outcomes and optimizing conditions remain challenging.
• Current computational methods may not fully capture complex reactant interactions.

Purpose of the Study:

• To develop a novel graph representation for deep learning to model B-H reactions.
• To create a Graph Neural Network (GNN) model for predicting B-H reaction yields.
• To identify key factors influencing reaction yield, particularly the role of bases.

Main Methods:

• Designed a custom reaction graph representing reactants as nodes.
• Developed a Buchwald-Hartwig Reaction Graph Neural Network (BH-RGNN) model.
• Trained the model on a high-throughput B-H reaction dataset.
• Employed perturbation-based analysis to understand model predictions.

Main Results:

• The BH-RGNN achieved a high predictive performance with an R² score of 0.971.
• The model demonstrated low computational costs.
• Perturbation analysis revealed significant insights into base-ligand interactions and their effect on reaction yields.
• Identified specific base characteristics crucial for optimizing reaction outcomes.

Conclusions:

• Tailored graph representations are effective for GNNs in chemical reaction modeling.
• The BH-RGNN offers a powerful tool for predicting reaction yield and guiding experimental design.
• This approach provides valuable guidance for base selection in B-H reactions, enhancing synthetic efficiency.