Designing Buchwald-Hartwig Reaction Graph for Yield Prediction

The Buchwald-Hartwig (B-H) reaction graph, a novel graph for deep learning models, is designed to simulate the interactions among multiple chemical components in the B-H reaction by representing each reactant as an individual node within a custom-designed reaction graph, thereby capturing both single-molecule and intermolecular relationship features. Trained on a high-throughput B-H reaction data set, B-H Reaction Graph Neural Network (BH-RGNN) achieves near-state-of-the-art performance with an R² score of 0.971 while maintaining low computational costs. Using a perturbation-based analysis, the model reveals significant insights into the relationship between bases and reaction yields, identifying key characteristics of bases influencing yields, which provides valuable guidance for base selection in the B-H reaction. This work demonstrates the effectiveness of tailored graph representations in advancing GNN applications for chemical reaction modeling, offering a promising tool for predicting reaction yield and identifying key reaction factors.