Agentic generation of verifiable rules for determin... | AI Research

Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
Computer-assisted synthesis planning relies on libraries of reaction rules to break down complex molecules into simpler precursors. Traditionally, these rules are manually encoded by experts, a process that is slow, labor-intensive, and unable to adapt to the vast, "long-tailed" nature of chemical reactions. This paper introduces a fully automated, multi-agent pipeline that uses Large Language Models (LLMs) to classify chemical reactions and generate the necessary rules without human intervention. By creating a self-expanding system that verifies its own output against a massive corpus of patent data, the researchers have developed a living reactivity database that can adapt to new chemical discoveries on demand.

A Multi-Agent Approach to Classification

The researchers addressed the challenge of classifying hundreds of thousands of reactions by decomposing the task into a multi-agent framework. Instead of asking a single model to process the entire dataset, the system uses specialized agents to handle different stages of the process: high-level classification, detailed refinement, and verification. When the system encounters a reaction that does not fit into the existing taxonomy, a "generator" agent proposes a new category. This allows the taxonomy to grow organically, expanding from 68 initial seed classes to 14,073 distinct classes. This structure ensures that the system can handle both common and rare chemical transformations with high precision.

Automating Rule Generation

A major bottleneck in cheminformatics is the creation of SMIRKS—text-based patterns that define how atoms and bonds transform during a reaction. Manually writing these is difficult because they require precise constraints on stereochemistry and functional group compatibility. The authors developed an iterative, LLM-driven loop to automate this. The LLM drafts initial, broad rules that are then tested against a large corpus of reactions. If a rule produces a false positive (incorrectly labeling a reaction), the system feeds that failure back to the LLM, which then refines the rule to improve its specificity. This process creates a robust library of generalized templates that are both accurate and computationally efficient.

Performance and Reliability

To test the system, the researchers compared their automated taxonomy against established expert-curated systems like NameRXN. The results show that the LLM-derived hierarchy is highly reliable, achieving a 97.7% accuracy rate on unseen reactions. The study found that the LLM-generated taxonomy often resolves chemistry at a finer level of detail than existing tools, effectively capturing nuances that manual systems might miss. By combining this LLM-based classification with a lightweight, fast-matching neural classifier, the pipeline provides a deterministic, high-speed way to categorize reactions, making it suitable for real-time integration into modern drug discovery and molecular design workflows.