Retrosynthesis is the process of recursively decomposing target molecules into available building blocks. It plays an important role in solving problems in organic synthesis planning. To automate or assist in the retrosynthesis analysis, various retrosynthesis prediction algorithms have been proposed. However, most of them are cumbersome and lack interpretability about their predictions. In this paper, we devise a novel template-free algorithm for automatic retrosynthetic expansion inspired by how chemists approach retrosynthesis prediction. Our method disassembles retrosynthesis into two steps: i) identify the potential reaction center of the target molecule through a novel graph neural network and generate intermediate synthons, and ii) generate the reactants associated with synthons via a robust reactant generation model. While outperforming the state-of-the-art baselines by a significant margin, our model also provides chemically reasonable interpretation.

Clarity: The language throughout the paper could be improved (set notation used when referring to sets, poor grammar in many sentences, spelling out "ground truth" instead of GT). The work also does not have a real discussion section or conclusion, likely due to length restrictions. I have some pickier comments about the contextualization of the work: - L28: Only one of the previous methods the authors cite used hand-coded reaction templates; the others algorithmically extract them. The authors could revise the discussion of template-based methods to more clearly separate inherent limitations in template-based methods from limitations due to manual encoding (L34-35). The use of "total number of reactions" is a little confusing because this is one-step retrosynthesis where there is, by definition, one reaction.