Fig. 1: Machine learning approach to hybrid synthesis planning.

a Development workflow of the hybrid synthesis planner. A database of enzymatic reactions was parsed into machine-readable format. Reaction templates were algorithmically extracted from the reactions in the database. A neural network template prioritizer⁴⁷ was trained to predict the reaction template associated with each product molecule in the reaction database. The enzymatic template prioritizer and a previously trained synthetic template prioritizer²⁴ are used in tandem to predict hybrid synthesis plans. b Multi-model-guided tree search strategy used to explore the retrosynthetic search space for multi-step synthesis planning from an input molecule (yellow circle). The possible retrosynthetic reaction templates (squares) are scored using template prioritizer neural networks. Different colors correspond to different template sets (e.g., synthetic and enzymatic). (i) The leaf node from the highest-scoring path is selected. (ii) The selected retrosynthetic template is applied to the product molecule to generate the predicted precursor. The precursor is added to the search tree and the retrosynthetic templates are scored by their corresponding template prioritizer with the precursor as input. (iii) Visit counts are updated for the explored nodes. The visit counts are used in scoring to balance exploration and exploitation. Steps i, ii, and iii are repeated until a stopping criterion for the search is met. All pathways that connect the input molecule to allowed starting materials (gray circles) are returned.