Semi-template framework for retrosynthesis prediction using graph neural network

Retrosynthesis prediction, the identification of a set of reactions available to synthesize target molecules, is a crucial task in drug discovery and organic synthesis. Recently, computer-aided retrosynthesis has gained much attention. Various deep learning-based algorithms have been proposed to assist or automate retrosynthesis analysis. However, existing approaches employ a stepwise editing strategy to translate the target molecule graph into corresponding synthons and complete them into reactants. Since chemical reactions lead to alterations in the state of certain atoms and bonds between reactants and products, predicting retrosynthesis is equivalent to predicting alterations in atoms and bonds. Inspired by this view, we propose State2Edits, an end-to-end semi-template framework for retrosynthesis prediction, which sequentially edits the target molecular graph to generate the corresponding reactants. Our proposed approach involves a directed message passing neural network (D-MPNN) to predict the edit sequence. Additionally, we introduced novel motif edits to replace traditional leaving groups in synthon completion and designed a method that merges single atom and generates bond edits, sidestepping complex multi-atom edits. These approaches aim to improve prediction accuracy and diversity. Extensive experiments show that our model generates high-quality and diverse results, achieving superior performance for semi-template-based retrosynthesis with a promising 55.4% top-1 accuracy in the USPTO-50K standard benchmark data set.