A Graph-to-Sequence Method for C–H Activation Retrosynthesis

C-H activation reactions are widely used in organic synthesis but pose significant challenges for retrosynthesis due to the high bond energy, chemical inertness, and complex regioselectivity of C-H bonds. Traditional models struggle to identify reactive centers and predict plausible disconnection paths in such systems. We propose TransGraphEdit, a template-free framework that combines a graph neural network (GNN) encoder and a Transformer decoder to model stepwise structural edits based on the product's molecular graph. The model captures long-range dependencies and synergistic effects between distant functional groups, enabling interpretable and mechanism-aware retrosynthesis predictions. To improve generalization on low-resource C-H activation reactions, we applied a SMILES augmentation strategy to expand the training space. TransGraphEdit achieves 58.06% Top-1 accuracy on a curated C-H Arylation data set and 53.8% on USPTO-50K, demonstrating its robustness and domain adaptability.