"17 million compounds extracted from 14 million patents" is true and misleading in the same breath, and the gap between those two readings is where patent chemical intelligence quietly goes wrong. The extraction works. It just doesn't measure what people assume it measures. [1]
Four stages, and the error compounds
Pulling chemistry out of a patent is a pipeline, and pipelines multiply their error rates rather than averaging them. Layout analysis finds where chemistry lives — body text, tables, figures, supplementary files. Chemical named-entity recognition then locates chemical mentions in text. The mature benchmark, BioCreative CHEMDNER, sees the best systems around 87–90% F1 — but that figure is measured on clean journal abstracts, and patent full-text is a different animal: OCR noise, multi-column tables, inconsistent nomenclature, and machine-translation artifacts in non-English filings all push real-world recall below the benchmark. [2]
Name-to-structure then converts a recognized name to a molecule. OPSIN is the open, deterministic, grammar-based workhorse here — it parses systematic IUPAC nomenclature with high precision and, importantly, fails predictably: it returns nothing on a malformed or non-systematic name rather than guessing wrong. [3] Predictable failure is a feature; the trap is the stages that fail confidently.
Optical structure recognition is brittle off-distribution
The genuinely hard stage is the structure that exists only as a drawing. Modern OCSR — DECIMER, MolScribe — frames this as image-to-graph: detect atoms, bonds, and stereo-wedges and assemble a molecule. On clean, computer-generated depictions it's excellent. The problem is that patents are not clean, and the numbers fall off a cliff.
MolScribe retains over 90% on synthetic images but drops to roughly 65% once realistic images are rotated and sheared; on the Staker patent-image set it lands at 86.9%; and on the hand-drawn DECIMER-HDM benchmark its exact-match accuracy is about 11%. [4] DECIMER's own hand-drawn model reaches ~73% on that set, but rival tools sit at 7–11% — i.e., off-distribution OCSR is nowhere near solved. [5] And there's a quieter failure mode underneath: superatoms. Patents draw "Ph," "Boc," "OMe," "Ts" as shorthand, and the recognizer must expand each abbreviation to its full substructure correctly. Miss one expansion and the molecule is silently wrong — not flagged, just wrong.
The Markush wall
Now the limit that no accuracy percentage will ever capture, because it isn't an accuracy problem. Patents don't claim molecules. They claim a genus. A Markush structure specifies a scaffold with variable R-groups ("R1 = C1–C6 alkyl, optionally substituted by…"), variable attachment points, and variable counts — so a single drawn structure can legally denote thousands to billions of specific compounds, and that genus is the intellectual property. The exact exemplified compounds are almost incidental; the claimed space is the asset.
Automated extraction cannot represent that space as enumerable molecules. The proof is historical and decisive: CAS built MARPAT — a dedicated, structure-searchable database of Markush structures, launched on STN in 1989–90 — and curates it with human experts, precisely because generic-structure search could not be automated to the required fidelity. [6] Thirty-five years later, machine recognition of Markush structures is still an open research frontier: the 2025 MarkushGrapher work is among the first end-to-end attempts, and it doesn't yet even handle abbreviations. [7] So the most commercially valuable object in a chemical patent — the genus that bounds freedom-to-operate — is exactly the object an automated "compound list" omits.
What the 17 million number actually is
Put the pieces together and the headline count resolves into something specific and useful, as long as you know what it is. It is a set of specific exemplified structures extracted at ~90% per-stage accuracy on clean depictions and far lower on real ones, inflated by superatom and salt artifacts and open-valence fragment radicals, and silent about the Markush genus entirely. The error is systematic, not random — biased toward chemistries that draw cleanly and toward over-counting decorated scaffolds — so you cannot average it away with more patents. More patents bring more Markush claims and more degraded images. [8]
The discipline that follows: trust patent-extracted chemistry at the aggregate and scaffold level — "who is filing around this chemotype, and when did the activity start" survives the noise — and distrust it at the exact-molecule and genus-coverage levels, which is where the systematic errors and the un-enumerable Markush live. The intelligence isn't in the extraction model; it's in knowing which questions the extracted data can answer and which it structurally cannot. A tool that hands you a clean "list of compounds in this patent" and lets you treat it as the patent's coverage is selling you the most dangerous version of the number.
When a patent-intelligence product reports a compound count, ask three things: what was the OCSR accuracy on real images, not synthetic ones; how were superatoms and salts handled; and — the one that matters most — how does it represent the Markush genus that actually defines the claim?