Fig. 3: Analyses of using the search graph to perform retrosynthetic planning of a group of molecules.

a The comparison of the iterations for the group planning of 20 similar molecules between different methods. The baselines used here are Retro*¹¹ and EG-MCTS¹⁶. “Ours-Graph” and “Ours-Shared Graph” indicate the planning of each molecule using a separate search graph or solving a group of molecules together using a shared search graph. Experiments are run over 10 random seeds, and the average results are reported; the same applies to Panels (b, c). The lower and upper whiskers represent the minimum and maximum of the ten data points, respectively. b The comparison between the planning based on graph and the shared graph for a group of molecules with different similarity levels, in terms of the benefits gained from reducing the number of nodes compared to tree-structured retrosynthetic planning. The similarity is defined by the Tanimoto distance between two molecules. Experiments are run across 200 groups. The shaded area indicates the 95% confidence intervals around the mean. c The comparison of different group sizes and varying levels of similarity on the benefits of shared graph search. The shaded area represents the range between the maximum and minimum values. d The illustration of molecules with different similarities (the differences are highlighted). The changes in peripheral groups have a minor impact on overall similarity, while modifications in the crucial central ring result in significant dissimilarity. e An example of group retrosynthetic planning for three similar molecules that employs a shared intermediate molecule (indicated by the black box) based on the shared search graph. This helps to save the number of search iterations. Arrows and boxes of different colors are used to distinguish the synthesis pathways of the three molecules. Source data are provided as a Source Data file.