Processing 1000+ chemical patents in minutes

Rule-based parsers choked on patent PDFs. Chemical patents aren't standardized - formatting varies by jurisdiction, time period, and filing organization. Tables appear mid-paragraph. Chemical structures interrupt text flow.

Standard OCR systems failed on chemical diagrams. Chemical structures aren't text - they're complex visual representations with bonds, angles, stereochemistry, and annotations.

Single-model approaches for chemical structure recognition plateaued at 50% accuracy. Low resolution scans, hand-drawn structures from older patents, complex multi-step reactions, overlapping labels.

The problem required specialized intelligence at every stage. We needed a multi-agent architecture where each component solved a specific sub-problem, with validation at each step.

No single image-to-SMILES model exceeded 50% accuracy on complex patent diagrams. Our solution: deploy three best-in-class models in parallel and use visual verification to select the correct output.

Each chemical image passes through three specialized models simultaneously. Each generates a SMILES string. All three outputs are converted back into chemical structure images. A vision language model then compares the three generated images against the original to identify the closest match.

Each model was trained on different datasets. Model A excelled at hand-drawn structures from older patents. Model B handled complex multi-step reactions. Model C performed best on high-resolution modern diagrams. By running all three and using visual verification, we increased accuracy from 50% to 65% - the difference between an interesting experiment and a production deployment.

Chemical compounds have multiple names and notations: IUPAC systematic names, common names, CAS registry numbers, SMILES strings, InChI identifiers, proprietary trade names. Simple keyword matching missed patents using different naming conventions. Fuzzy matching produced too many false positives.

Worse: relevance isn't just about compound presence. A patent might synthesize the target compound (highly relevant), synthesize a structurally similar compound (relevant), mention it in a comparison table (low relevance), or reference it as prior art (not relevant).

We deployed an LLM to score every patent against defined synthesis criteria. It understood that “compound 7a” in one patent might be the same as “intermediate B” in another. It recognized that a patent focusing on a related compound might contain applicable synthesis routes. It distinguished between patents that mention a compound versus patents that synthesize it.

Output: a ranked list of patents sorted by synthesis relevance, typically reducing 1,000+ candidates to 5-10 high-value documents.

Patent screening that took researchers 40-80 hours per compound now takes minutes. Automated screening, extraction, and report generation replaced manual review entirely.

Instead of searching for synthesis pathways, researchers receive curated reports with ranked options and comparative analysis. Multiple compounds process in parallel. As patent databases grow, processing scales horizontally without additional researcher time.

Every processed compound adds to the synthesis tree. The system gets smarter with each analysis, identifying patterns and connections across chemical space.