Abstract
Single-step retrosynthesis is a crucial task in organic chemistry and drugdesign, requiring the identification of required reactants to synthesize aspecific compound. with the advent of computer-aided synthesis planning, thereis growing interest in using machine-learning techniques to facilitate theprocess. Existing template-free machine learning-based models typically utilizetransformer structures and represent molecules as ID sequences. However, thesemethods often face challenges in fully leveraging the extensive topologicalinformation of the molecule and aligning atoms between the production andreactants, leading to results that are not as competitive as those ofsemi-template models. Our proposed method, Node-Aligned Graph-to-Graph (NAG2G),also serves as a transformer-based template-free model but utilizes 2Dmolecular graphs and 3D conformation information. Furthermore, our approachsimplifies the incorporation of production-reactant atom mapping alignment byleveraging node alignment to determine a specific order for node generation andgenerating molecular graphs in an auto-regressive manner node-by-node. Thismethod ensures that the node generation order coincides with the node order inthe input graph, overcoming the difficulty of determining a specific nodegeneration order in an auto-regressive manner. Our extensive benchmarkingresults demonstrate that the proposed NAG2G can outperform the previousstate-of-the-art baselines in various metrics.