Fig. 2: Performance of retrosynthetic planning of a single molecule.

a The comparison of the success rate between our method, greedy Depth First Search (Greedy DFS), DFPN-E¹⁰, MCTS-rollout⁹, Retro*¹¹, Retro*+¹⁵, EG-MCTS¹⁶ and PDVN⁴⁰ on the Retro*-190 dataset. The success rate is defined as the ratio of the number of molecules for which a synthesis route is successfully found to the size of the dataset. Experiments are run over 10 random seeds, and the average results are reported; the same applies to other panels in this figure. The iteration limit is set to 500; the lower and upper whiskers represent the minimum and maximum of the ten data points, respectively; the same applies to Panels (d–f). b The comparison of the average success rate under different search iteration limits. The shaded area represents the range between the maximum and minimum values. c The comparison of the time used to solve 190 retrosynthetic planning tasks across methods and under different iteration limits. The box plots show the medians as center lines, interquartiles as hinges, and up to 1.5 times the interquartile ranges as whiskers (outliers are not shown). d The average number of molecule nodes, when planning terminates, was compared between our method and other baselines. e The average number of reaction nodes, when planning terminates, was compared between the methods. f The comparison of the average number of shortest routes found by our method and baselines. Shorter synthesis routes usually mean lower cost and higher validity. g The comparison of the average success rate on the larger and more challenging dataset (20,000 molecules) was plotted in the same way as Panel (b). h The comparison of the time used to solve all tasks between our method and baselines on the larger dataset; plotted in the same way as Panel (c). Source data are provided as a Source Data file.