ReactionReasoner: Towards Reasoning LLM for Chemical Reaction Prediction

Chemical reaction prediction involves predicting reaction outcomes from given molecules, inferring the required starting materials from a given product, and identifying reagents that bring about the transformation. It is a complex problem that demands explicit reasoning such as functional group recognition, reaction mechanism analysis. Recent advances in general-purpose large language models (LLMs) have led to improved performance on reaction prediction tasks and can generate some reasoning traces. However, due to a lack of domain-specialized training, they consistently struggle with reaction-specific reasoning and consequently exhibit poor accuracy. In parallel, molecular LLMs have also been recently developed, yet such models typically predict only the molecules without providing reasoning and thus their performance on complex reaction prediction tasks remains limited. To address this limitation and move toward reasoning-capable LLMs for chemical reaction prediction, we present SyntheticReact, a synthetic reasoning data generation method for chemical reactions, mirroring practicing chemists’ strategies, and ReactionReasoner, an LLM-based reaction reasoning model trained on data produced by SyntheticReact. In particular, given reaction SMILES (RXN SMILES), SyntheticReact collects reaction documents via web scraping, extracts information about human chemists’ strategies, and uses an LLM to structure it into reasoning data. Using reasoning data generated by SyntheticReact, we train ReactionReasoner through supervised fine-tuning. In addition, we apply a self-bootstrapping approach: reasoning data that lead to correct answers are used for an additional supervised training, while those that fail are used to generate reflection data, capturing why the reasoning is unsuccessful. Through our experiments, we show that detailed, step-by-step reasoning, similar to how human chemists approach problems, is more valuable than reasoning that only provides general explanations.Moreover, when proper reasoning is given, ReactionReasoner significantly outperforms models that attempt predictions without reasoning.