Abstract
A well-known pitfall of molecular generative models is that they are notguaranteed to generate synthesizable molecules. There have been considerableattempts to address this problem, but given the exponentially largecombinatorial space of synthesizable molecules, existing methods have shownlimited coverage of the space and poor molecular optimization performance. Totackle these problems, we introduce ReaSyn, a generative framework forsynthesizable projection where the model explores the neighborhood of givenmolecules in the synthesizable space by generating pathways that result insynthesizable analogs. To fully utilize the chemical knowledge contained in thesynthetic pathways, we propose a novel perspective that views syntheticpathways akin to reasoning paths in large language models (LLMs). Specifically,inspired by chain-of-thought (CoT) reasoning in LLMs, we introduce thechain-of-reaction (CoR) notation that explicitly states reactants, reactiontypes, and intermediate products for each step in a pathway. With the CoRnotation, ReaSyn can get dense supervision in every reaction step to explicitlylearn chemical reaction rules during supervised training and performstep-by-step reasoning. In addition, to further enhance the reasoningcapability of ReaSyn, we propose reinforcement learning (RL)-based finetuningand goal-directed test-time compute scaling tailored for synthesizableprojection. ReaSyn achieves the highest reconstruction rate and pathwaydiversity in synthesizable molecule reconstruction and the highest optimizationperformance in synthesizable goal-directed molecular optimization, andsignificantly outperforms previous synthesizable projection methods insynthesizable hit expansion. These results highlight ReaSyn's superior abilityto navigate combinatorially-large synthesizable chemical space.