RCsearcher: Reaction center identification in retrosynthesis via deep Q-learning

The reaction center consists of atoms in the product with local properties that differ from those in the reactants. Previous studies focused on identifying the reaction center using semi-templated retrosynthesis methods, which are limited to single reaction center identification. In reality, however, many reaction centers involve multiple bonds or atoms, referred to as multiple reaction centers. This paper introduces RCsearcher, a unified framework that exploits the strengths of graph neural networks and deep reinforcement learning for identifying both single and multiple reaction centers. The key insight of the framework is that the single or multiple reaction center must be a node-induced subgraph of the molecular product graph. At each step, RCsearcher selects a node in the molecular product graph and adds it to the explored node-induced subgraph as an action. Comprehensive experiments demonstrate that RCsearcher consistently outperforms other baselines, and is able to identify reaction center patterns not present in the training set. Ablation experiments confirm the effectiveness of individual components of RCsearcher, including the beam search and the one-hop constraint of the action space.