Performance assessment of various graph neural network architectures for predicting yields in cross-coupling reactions

Abstract

Machine learning (ML) is revolutionizing various scientific fields, including chemistry. Among different ML techniques, graph neural networks (GNNs) have emerged as a powerful tool for capturing complex relationships in data by representing datapoints as graphs. Applying GNNs for predicting yields in chemical reactions is one of the emerging areas of focus. However, handling heterogeneous datasets remains a key challenge. In this study, we utilized diverse datasets encompassing various transition metal-catalyzed cross-coupling reactions including Suzuki, Sonogashira, Cadiot–Chodkiewicz, Ullmann-type, and Buchwald–Hartwig coupling reactions. To predict reaction yields, we implement multiple GNN architectures, including message passing neural networks (MPNN), residual graph convolutional networks (ResGCN), graph sample and aggregate (GraphSAGE), graph attention networks (GAT and GATv2), graph convolutional networks (GCN), and graph isomorphism networks (GIN). The comparative analysis of these architectures reveals that MPNN achieve the highest predictive performance, with an R² value of 0.75. Additionally, to enhance the interpretability of our models, we employed the integrated gradients method to determine the contribution of each input descriptor to the model's yield prediction. This study highlights the potential of effective and explainable graph-based models to predict the yields of chemical reactions. This work contributes to the application of machine learning in catalysis, providing valuable insights for optimizing catalytic reactions and contributing to innovations in sustainable chemistry and organic synthesis.