ReaSyn: Generative AI with Chain-of-Reaction Reasoning for Predicting Synthesizable Molecular Pathways

A major challenge in molecular design—whether for pharmaceuticals, chemicals, or materials—is ensuring that newly proposed molecules can actually be synthesized. This requires identifying a viable synthesis pathway: a sequence of chemical reactions that transform starting materials into the target molecule. Traditional methods rely heavily on expert knowledge and trial-and-error, but recent advances in generative AI are changing the landscape. NVIDIA’s ReaSyn is a new generative model designed specifically to predict molecular synthesis pathways with greater accuracy and reliability by applying reasoning strategies inspired by large language models. One key insight driving ReaSyn is the importance of chain-of-thought (CoT) reasoning in AI. In natural language tasks, CoT enables models to break down complex problems into intermediate steps, improving both transparency and performance. ReaSyn adapts this idea to chemistry through a novel chain-of-reaction (CoR) notation. In CoR, a synthetic pathway is represented as a linear sequence where each step includes the reactants, the reaction rule applied, and the resulting product. Reactants and products are encoded as SMILES strings with special boundary markers, while reactions are represented by reaction-class tokens. This structure mirrors how chemists reason through synthesis, step by step, and allows the model to receive feedback at each stage—enabling more accurate learning of chemical rules. ReaSyn operates as an autoregressive generative model, building pathways incrementally from simple building blocks to the final target. This bottom-up approach ensures that each intermediate product is chemically valid and consistent with known reaction patterns. The model uses a reaction executor like RDKi to verify and generate products at each step, providing real-time feedback that enhances training and improves pathway reliability. To further improve performance, ReaSyn incorporates advanced reasoning techniques from LLMs. It uses outcome-based reinforcement learning (RL) finetuning via GRPO, where the reward is based solely on the final molecular similarity to the target. This encourages the model to explore diverse pathways without being constrained by intermediate steps, increasing its ability to find novel and effective routes. Additionally, ReaSyn applies goal-directed test-time search using beam search with a reward function tailored to the task—whether it’s maximizing similarity to a target molecule or optimizing a desired chemical property. Performance benchmarks show ReaSyn outperforms existing methods. In retrosynthesis planning, ReaSyn achieves a success rate of 76.8% on the Enamine dataset, compared to 25.2% for SynNet and 63.5% for SynFormer. On synthesizable molecular optimization tasks, ReaSyn combined with Graph GA achieves an average optimization score of 0.638, surpassing prior methods. It also excels in synthesizable hit expansion, generating diverse, viable analogs of promising molecules by exploring their neighborhoods in synthesizable chemical space. By treating synthesis as a reasoning process, ReaSyn bridges the gap between theoretical molecular design and practical synthesis. It enables researchers to generate molecules that are not only functionally promising but also feasible to produce, accelerating drug discovery and materials development. With its foundation in advanced reasoning and real-world validation, ReaSyn represents a significant step forward in making generative AI a practical tool for chemistry. The full paper is available on arXiv, and the code is hosted on GitHub.