AI Agents Automate Chemical Reaction Rule Generation

Computer-assisted synthesis planning relies heavily on extensive libraries of reaction rules to break down target molecules into simpler precursors. However, the vast and ever-evolving landscape of chemistry makes manual rule encoding impractical, and existing tools often use static rulesets that cannot adapt to new discoveries. This research introduces a fully automated pipeline that addresses these challenges.The pipeline utilizes a multi-agent framework of large language models (LLMs) to not only classify chemical reactions but also to generate the reaction rules themselves. Each rule is created under a rigorous verification loop, testing it against a large corpus of patent reactions. This system successfully expanded a standard chemical taxonomy from 68 to over 14,000 classes without human intervention, demonstrating its ability to adapt and grow. The resulting living reactivity database can classify new reactions with high accuracy, matching leading proprietary classifiers while offering finer chemical resolution and on-demand extensibility.