Fig. 6: The search procedure based on the AND-OR search graph.

In retrosynthetic planning, we begin with the SMILES expression of a target molecule. Here, we use a graphical representation to enhance our understanding of the molecule’s properties, and then two neural networks based on that representation will be used to guide the planning process. Before planning, we create a template library filled with a variety of reaction templates, represented by molecular expressions. These templates assist in identifying reactants that correspond to a particular product. The planning process is divided into three main iterative stages: selection, expansion, and update. In the selection stage, we choose the most promising molecule, like M4 in the diagram, for further expansion. This expansion relies on chosen reaction templates, such as R5 and R6, to perform single-step retrosynthesis, resulting in reactants like M8 and M9, which are then added to the graph. The update stage involves revising the graph to reflect new information about the synthesizability and synthesis costs of the nodes. Ultimately, this process culminates in the determination of the optimal synthesis route for the target molecule. Furthermore, a reaction template tree containing only reaction nodes can be extracted for the following analysis.