Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
06:00 · July 2, 2026 · arXiv cs.AI RSS
Computer-assisted synthesis planning breaks target molecules into accessible precursors using large libraries of reaction rules that assign each transformation a deterministic, interpretable label. But chemistry is long-tailed, making manual encoding intractable, and existing tools rely on fixed rulesets that cannot adapt to new chemistries. Here we present a fully automated pipeline in which a multi-agent framework of large language models (LLMs) classifies reactions and writes the rules themselves across 665,901 US patent reactions, generating each rule under a verification loop that tests it against the corpus. It expands a standard taxonomy from 68 to 14,073 classes without human curation. With a lightweight fingerprint classifier, it classifies 97.7\% of unseen reactions, matching a leading proprietary classifier while resolving chemistry more finely and extending on demand to chemistry outside its training distribution. The result is a living reactivity database and a general route to turning generative models into reliable, self-expanding symbolic systems.
Summary
Computer-assisted synthesis planning relies on large libraries of reaction rules that label each transformation with a deterministic, interpretable class. Existing systems depend on fixed, manually curated ontologies such as the RXNO hierarchy, which cannot accommodate the long-tailed distribution of chemical reactions observed in practice. The authors address this limitation with a fully automated, multi-agent LLM pipeline that processes 665,901 reactions extracted from United States patents.
The framework begins with the RXNO seed taxonomy of 68 classes. Sequential classification agents assign reactions to existing categories or flag those that fall outside the current hierarchy. When gaps are detected, a label-generation agent proposes new classes, which a verifier agent tests against the corpus through an iterative refinement loop. For each validated class the system also produces generalized SMIRKS patterns that encode the reaction centre and its immediate structural environment. These patterns are refined automatically to reduce false positives while preserving recall, yielding a compact, machine-readable rule set.
The resulting library expands the taxonomy to 14,073 classes without human intervention. A lightweight fingerprint-based classifier built on the distilled rules then assigns unseen reactions to their correct position in the hierarchy at 97.7 percent strict-match accuracy at the third level, matching the performance of a leading proprietary tool while providing finer resolution and the ability to extend coverage on demand. The approach therefore converts the generative capabilities of large language models into a living, verifiable symbolic database suitable for integration into real-time synthesis-planning workflows.
Why it matters
This research is highly relevant for Dutch AI researchers and the robust local chemical and biotech industries, offering a novel neuro-symbolic approach to computer-assisted synthesis planning. The use of LLM agents with a verification loop aligns with the Netherlands' strategic focus on transparent, reliable, and verifiable AI systems.
