Bioactivity Assay Data Structuring
Pull the activity data scattered across patents, literature, and internal lab records out together with its test conditions, and organize it into a comparable, trainable structured dataset. The hard part isn't copying down the numbers—it's making sure every number carries its context, otherwise the values simply aren't comparable to one another.
Where This Kind of Task Gets Hard
Three easily underestimated pitfalls1. Units and dimensions aren't consistent
Activity concentrations are often expressed in a variety of units—nM, µM, µg·mL⁻¹, and more—and even a single paper may mix them internally. Converting from a mass concentration to a molar concentration requires the molecular weight, and the molecular weight often has to be found in another table or another paper—break the chain and the value is left floating.
If you pull out only the numbers without keeping the original units and the conversion basis, this batch of data is essentially unverifiable downstream.
2. Metric types get conflated
IC50, EC50, Ki, and Kd are not the same quantity; their measurement principles and applicable scenarios all differ. Normalizing them into a single "activity value" looks tidy but actually makes the model learn a target that mixes several physical meanings, producing results that are hard to interpret.
3. Test conditions determine comparability
The same IC50 is not comparable between two experiments that differ in cell line, temperature, incubation time, or substrate concentration. When conditions are missing, the value is still "true", but it has no comparative meaning.
So structuring must carry the conditions along: value + unit + metric type + test subject + test conditions. Miss any one of these, and the data point is discounted downstream.
One more, more insidious problem
Activity data inside charts and figures. Points obtained from curve fitting need human verification, and image resolution itself sets an upper bound on precision. This kind of data can be extracted—but it must be treated differently from the unambiguous values in body-text tables.
What We Cover
Grouped by data form and field- Test substances and their identifiers (names, development codes, abbreviations)
- Activity metric types: IC50 / EC50 / Ki / Kd / MIC, etc.
- Original values and original units, including range values and inequalities (>, <)
- Converted values and the basis for conversion
- Test subjects: targets, cell lines, strains, animal models
- Assay method type and key condition parameters
- Mapping between concluding statements and values
- Source location: page and table numbers in literature, patents, or lab records
Table forms, body-text narrative forms, and chart forms can all be handled, but the three follow different processing paths: tables are structured by column mapping; narrative text requires first recognizing the "substance—metric—value—condition" sentence structure; and charts go through a manual verification process.
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; can link compound SMILES / InChI identifiers
- 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 scenariosStructured extraction of test substances, activity metrics, value units, test subjects, and methods, covering patent and literature data in pharma R&D.
PK/PDTable extractionUnit normalizationFrequently Asked Questions
Questions we're often asked before working togetherUnits aren't consistent within the same batch of data—how do you handle that?
Original values and original units are preserved as-is; converted values go into a separate field, with the basis for conversion recorded (for example, the molecular weight used and its source). Original values are never overwritten, so downstream consumers can independently verify whether the conversion is correct.
Are IC50, EC50, Ki, and Kd annotated separately?
Yes. The metric type is a separate field and is not mixed in with the value. These quantities differ in meaning and are not interchangeable; mixing them would cause downstream models to learn the wrong target, so we do not normalize them into "a single activity value".
What if an activity value is missing its test conditions?
It is recorded faithfully as incomplete conditions, with no imputation. When conditions such as cell line, temperature, or incubation time are missing, the value loses comparability—flagging that is more valuable than filling in a plausible-looking value.
Can data in charts and figures be extracted?
Yes, but precision is limited by image resolution. Results of this kind carry confidence flags and go through human review, and are not treated by the same standard as the unambiguous values in body-text tables.
What formats do you output?
Delivered as JSON / CSV / Excel, with fields corresponding one-to-one to the co-designed Schema; when compounds are involved, SMILES / InChI identifiers are linked so it can be merged with structure data.
Continue Reading
Related pagesTo match activity data to compounds, you first have to turn structure diagrams into machine-readable structures.
Browse all the solutions we currently offer by industry domain and specific task.
Send us a representative paper or table, and we can run a small pilot annotation batch so you can see directly whether the field design and boundary handling match your expectations.