Three-dimensional geometric deep learning for reaction prediction with equivariant graph transformer

Organic synthesis, a critical process in drug and material development, often involves complex reactions that can be time-consuming and costly to explore experimentally. Recent advances in machine learning have shown promise in predicting reaction outcomes, but challenges remain in capturing the full complexity of molecular interactions, particularly in three-dimensional space. To this end, we propose an Equivariant Graph Transformer (termed EGT) that predicts organic reactions by learning the three-dimensional (3D) geometric characteristics of molecules. We employed the equivariant graph neural network to extract geometric spatial information and a pairwise distance fed to position embedding to capture long-range interactions, to finely delineate the spatial structure of chemical molecules, making stereochemical information of reactions learnable. To benchmark our model's performance, we conducted reaction prediction experiments on the USPTO_STEREO and USPTO_FULL datasets as well as retrosynthesis prediction on the USPTO_50k and USPTO_MIT datasets. In addition, we conducted case studies focusing on synthesis planning and reaction prediction, and compared the results with those of human evaluations. The proposed EGT model has outperforms all existing algorithms with a Top-1 accuracy of 79.4 % for forward reaction prediction on the USPTO_STEREO dataset, and excels in predicting both forward reactions and retrosynthesis. Moreover, we demonstrated the model's capability to conduct forward total synthesis planning, showcasing its reliability and accuracy in achieving high Top-1 predictions. Molecular 3D geometry learning positions our model as a leading tool in the field of organic synthesis, paving the way for more efficient and accurate drug development.