Chemical Structure Recognition & Correction
Turn the structure diagrams in patents, literature, and lab records back into searchable, trainable structural data—covering full compound names, Markush formulas, non-specific structures, and complex substituents. This is one of the highest-cost-to-get-wrong tasks in chemical data annotation, and the project type we have delivered the longest.
Why This Can't Rely on Tools Alone
The hard part isn't recognition—it's judgmentThere is no shortage of tools that convert structure diagrams into machine-readable formats, and their performance in public benchmarks is not bad. But in real production environments, the gap shows up entirely in the "irregular" portion of the data.
A benchmark covering patents and papers shows: for tabular patents, structure and identifier recognition accuracy is about 95%; for non-tabular patents, structure recognition is about 80% and identifier recognition about 75%. For paper-type documents, identifier extraction exceeds 99%, while structure recognition falls at 78–80%.
There is an even more insidious class of problem: the model may output structures that are syntactically valid but chemically impossible in valence, such as pentavalent carbon. Such results do not throw exceptions—they quietly enter the training set and are only found retrospectively when the model's behavior becomes abnormal.
So the real difficulty is not "recognition" but "judgment"—is that substituent in the diagram a non-specific structure or a drawing error? Which compounds does the Markush variable definition actually cover? Which interpretation should the blurred bonds in a scan take? These require a chemistry background to decide, and cannot be solved by pushing model metrics up another two points.
We wrote a column piece on how datasets for this kind of task should be structured and why edge samples are the hardest to obtain: the PEN dataset.
What We Do Is Not Human-Outsourcing-Style Annotation
The difference isn't the price—it's the deliverables- Billed per item for box selection and label tagging
- Delivers only an annotation result
- Errors are found through after-the-fact sampling
- Cannot answer "why was this labeled this way"
- The Schema is unilaterally specified by the client, and boundaries are guessed
- Co-design the Schema with your algorithm team first, defining fields, types, granularity, and boundaries
- Delivers the dataset + annotation guidelines + QC report + provenance log
- Three-tier QC: automated checks → team review → domain expert arbitration
- Every data point can be traced back to the source location, annotator, and modification time
- Edge cases are written back into the guidelines, so the next batch doesn't repeat the same mistakes
Delivery Pipeline
Four steps in one line · fully loggedDeliverables
What you get is not just an annotation result- Structured dataset: JSON / CSV / Excel, with fields corresponding one-to-one to the Schema; structures output as SMILES and InChI
- Annotation guidelines document: field definitions, edge cases, acceptance criteria
- QC report: sampling ratio, error categories, rework records
- Provenance log: each data point maps to its source location, annotator, and modification time
Related Cases
Real delivery scenariosRecognition and structure-drawing correction of full compound names, Markush formulas, non-specific structures, and complex substituents, covering patent and literature data in chemistry and chemicals.
OCSRStructure drawingMarkushFrequently Asked Questions
Questions we're often asked before working togetherAutomated recognition tools already exist—why is human work still needed?
Public benchmarks show that automated tools perform well on tabular patents, but structure recognition accuracy drops markedly for non-tabular patents and papers; the models also output structures that are syntactically valid but chemically impossible in valence. Such results do not raise errors—they quietly enter the training set. Tools are good for producing a first draft, but the judgment must be made by people with a chemistry background.
Can you handle Markush formulas and non-specific structures?
Yes. Variable definitions in Markush formulas, the scope of substituents, and non-specific structures (polymers, mixtures, undefined crystal forms, etc.) are all within scope. We record the value boundaries of variables and uncertain items as agreed for the project, rather than forcing them into a single structure.
Can you process scans and low-quality patent pages?
Yes. Blurred bonds, structures spanning pages, and captions separated from the body text are judged by experts using context, and the basis for the judgment and confidence status of each entry is marked in the provenance log.
What formats do you output?
Structures are output as SMILES / InChI, while the structure diagrams and entity–relation records are also retained; the dataset as a whole is delivered as JSON / CSV / Excel, with fields corresponding one-to-one to the co-designed Schema.
How is data confidentiality ensured?
An NDA is signed and the authorized personnel list is locked before the project starts; data is processed in the client's designated environment or an isolated domain—it never leaves that domain, is never reused, and is never used to train our own models; raw and intermediate data are deleted on the agreed schedule, and operation logs can be exported for review. See Six Data Compliance Clauses for details.
Continue Reading
Related pagesStructure formulas and formulations in materials patents often appear on the same page, so the two workflows can be run together.
Browse all the solutions we currently offer by industry domain and specific task.
Send us your data sources, target Schema, and acceptance criteria, and we'll start by providing an executable annotation plan and a boundary checklist.