Agentic generation of verifiable rules for deterministic, self-expanding reaction classification

Technical Summary: Agentic Generation of Verifiable Rules for Deterministic, Self-Expanding Reaction Classification

Problem Statement

Computer-assisted synthesis planning (CASP) relies on large libraries of reaction rules to decompose target molecules into precursors. However, chemical reactions follow a power-law distribution, where the "long tail" of rare transformations makes exhaustive manual encoding intractable. Existing classification tools (e.g., NameRXN, Rxn-INSIGHT) rely on fixed, manually curated ontologies and rule sets. These systems face two critical limitations:

1. Static Coverage: They cannot adapt to new chemistries or transformations absent from their original design, leading to uninformative "unclassified" tags or mis-assignments.
2. Scalability: Manually encoding thousands of reaction types, including necessary stereochemical constraints and functional group compatibilities, is a labor-intensive bottleneck that limits the exploration of novel chemical space.

While template-free neural models have shown promise in predicting reaction outcomes, they often generate physically invalid molecules and lack the deterministic, interpretable labels required for reliable synthesis planning. Conversely, rule-based matching offers unambiguous assignment but requires a comprehensive, validated library that current automated methods have failed to produce at scale.

Methodology

The authors present a fully automated, multi-agent pipeline that leverages Large Language Models (LLMs) to classify reactions and generate verifiable reaction rules (SMIRKS) across a corpus of 665,901 USPTO patent reactions. The system operates through three core components:

1. Multi-Agent Classification and Dynamic Taxonomy Expansion
The pipeline decomposes the classification task into five specialized LLM agents, utilizing a "Chain-of-Verification" paradigm to minimize hallucination:

• Cohort Grouping: Reactions are grouped by shared retrosynthetic templates (using RDChiral) to ensure mechanistically equivalent transformations receive identical labels.
• Hierarchical Agents: A Hierarchy Agent maps cohorts to 68 seed sub-classes (based on the RXNO ontology and Carey et al.). A Detailed Agent refines these to the finest applicable level.
• Verification and Generation: A Verifier Agent audits assignments. If a reaction does not fit existing categories, a Generator Agent proposes new taxonomy entries (code, name, hierarchy). An Aggregator Agent deduplicates and integrates these proposals into a "living" hierarchy.
• Result: The taxonomy expands organically from 68 seed classes to 14,073 distinct classes without human curation.

2. Autonomous SMIRKS Generalization
For each class, an LLM generates generalized reaction patterns (SMIRKS) through an iterative refinement loop:

• Drafting: The LLM creates initial, broad patterns (low specificity, high recall).
• Validation: Patterns are tested against the corpus. False positives (patterns matching incorrect reaction classes) are identified.
• Refinement: The LLM applies chemically grounded mutations (e.g., tightening hydrogen counts, adding context atoms, merging halide lists) to eliminate false positives while maintaining recall.
• Outcome: This process yields 4,964 generalized, high-specificity SMIRKS patterns.

3. Deterministic Classification Architecture
The final system, ReactionClassifier, combines the LLM-derived hierarchy and the SMIRKS database with two inference modes:

• Hybrid Strict: A lightweight Multilayer Perceptron (MLP) trained on reaction difference fingerprints (MDP) predicts a high-level class. This prediction gates a subset of the exact template library, ensuring rapid, deterministic matching with high precision.
• Generalised SMIRKS: A global first-match-wins strategy using the 4,964 generalized patterns ordered by a false-positive-minimizing algorithm (based on Tarjan's algorithm for strongly connected components).
• LLM Fallback: Reactions that fail the deterministic layer (no template match) are routed back to the multi-agent pipeline to be classified and potentially expand the taxonomy further.

Key Results

• Taxonomy Expansion: The system expanded the reaction taxonomy from 68 to 14,073 classes, capturing granular distinctions (e.g., specific reagents or conditions) that manual tools miss.
• Accuracy:
  • The Hybrid Strict mode achieved 97.9% overall accuracy at the third hierarchical tier on a frequency-weighted USPTO split and 94.1% on a template-balanced split.
  • The Generalised SMIRKS mode achieved 85.9% overall accuracy at tier 3.
  • Label noise estimation via confident learning and LLM-as-judge adjudication suggests a true mislabeling rate of approximately 0.33%, comparable to human-curated baselines (NameRXN).
• Out-of-Distribution (OOD) Performance:
  • On the CRD-2025 corpus (academic reactions from 2025+), the Generalised SMIRKS mode achieved 68.3% coverage, outperforming Rxn-INSIGHT (64.8%) and approaching NameRXN (89.3%), despite using only ~10% of the rule count.
  • On the RingBreaker corpus (ring-forming reactions), the system achieved 80.6% coverage, significantly outperforming NameRXN (73.7%) and Rxn-INSIGHT (6.9%).
• Latency: The deterministic modes operate at 6–9 ms per reaction, roughly twice as fast as Rxn-INSIGHT.
• Self-Expansion: When applied to unclassified academic chemistry, the system successfully added 1,942 new taxonomy entries, including established methodologies underrepresented in patent data (e.g., C–H functionalization, reductive cross-coupling).

Significance and Claims

The paper claims to demonstrate that LLMs can not only understand chemical language but also "write its grammar" by generating reliable, deterministic symbolic rules at scale. The primary significance lies in resolving the tension between determinism (required for reliable synthesis planning) and adaptability (required for evolving chemical knowledge).

By separating the tasks into a fast, deterministic symbolic layer and a generative language-model layer, the system creates a "living reactivity database." This approach:

1. Automates Curation: It replaces years of expert manual encoding with an on-demand, automated pipeline.
2. Enables Fine-Grained Control: The expanded taxonomy allows generative models to be conditioned on specific named transformations rather than broad categories, enabling more precise exploration of chemical space.
3. Generalizes Beyond Training Data: The system can identify and encode novel chemistry outside its original training distribution, bridging the gap between industrial patent data and academic methodology.

The authors conclude that this architecture provides a general route for turning generative models into reliable, self-expanding symbolic systems, potentially allowing future CASP tools to approach the synthesis of complex molecules that currently require bespoke methodology development.