AI for science asks what models can do for a laboratory. Science for AI asks the reverse: what does a laboratory have to record for a model to learn from it? This is a working guide to that question.
3 of 96 wells reach the paper.
An illustrative protein activity screen, not real data. Color shows activity relative to the reference protein. Select any well to read its record.
The best-known success of AI in science rested on a dataset before it rested on a model. The Protein Data Bank opened in 1971 with seven structures. Over the next five decades structural biologists deposited their results in that one open archive, in a shared format, and journals came to require it.1 By the time AlphaFold 2 was trained, the archive held about 170,000 structures, and the field had been running a blind test since 1994 in which predictions are scored against structures nobody has yet seen.2
1 Berman et al., J. Comput. Aided Mol. Des. (2014).
2 DeepMind (2020); Jumper et al., Nature (2021).
Structures in the Protein Data Bank
Total entries at the end of each year
1971 Opens with 7 structures
1994 First CASP blind test
2020 AlphaFold 2 trained on about 170,000
2025 247,274 entries
The record that structure prediction learned from took fifty years to build. The archive passed 250,000 entries in 2026. Data: RCSB PDB.3
| Year | Entries |
|---|---|
| 1976 | 13 |
| 1977 | 36 |
| 1978 | 42 |
| 1979 | 53 |
| 1980 | 69 |
| 1981 | 85 |
| 1982 | 117 |
| 1983 | 153 |
| 1984 | 175 |
| 1985 | 195 |
| 1986 | 213 |
| 1987 | 238 |
| 1988 | 291 |
| 1989 | 365 |
| 1990 | 507 |
| 1991 | 694 |
| 1992 | 886 |
| 1993 | 1,582 |
| 1994 | 2,871 |
| 1995 | 3,812 |
| 1996 | 4,984 |
| 1997 | 6,548 |
| 1998 | 8,603 |
| 1999 | 10,959 |
| 2000 | 13,583 |
| 2001 | 16,393 |
| 2002 | 19,387 |
| 2003 | 23,531 |
| 2004 | 28,680 |
| 2005 | 34,014 |
| 2006 | 40,419 |
| 2007 | 47,553 |
| 2008 | 54,456 |
| 2009 | 61,744 |
| 2010 | 69,486 |
| 2011 | 77,411 |
| 2012 | 86,162 |
| 2013 | 95,500 |
| 2014 | 105,052 |
| 2015 | 114,295 |
| 2016 | 125,100 |
| 2017 | 136,161 |
| 2018 | 147,330 |
| 2019 | 158,809 |
| 2020 | 172,815 |
| 2021 | 185,401 |
| 2022 | 199,691 |
| 2023 | 214,191 |
| 2024 | 229,660 |
| 2025 | 247,274 |
Most experimental science has neither the archive nor the test. Results sit in supplementary spreadsheets, instrument exports and notebooks, in formats that change from lab to lab. Experiments that did not work are rarely written down in a form anyone else can use. The reasons an experiment was run at all, which is most of what a scientist knows, are not recorded anywhere.
Automated laboratories are about to multiply the volume. In July 2026 the US National Science Foundation committed $380 million to 20 programmable cloud laboratories, with shared standards for protocols, metadata, data and AI as part of the brief.4 One of them, the DREAM Cloud Lab at Northwestern, plans to build and measure more than 300,000 proteins and release up to 30 million data points.5 Whether records like these are something a model can learn from is decided by what gets captured at the bench, and it is decided now.
4 NSF announcement, 22 July 2026.
5 Northwestern Engineering, July 2026.
A reader who is confused can write to the first author. A model cannot.
Whatever a model needs has to be in the record already. These eight questions are a way to check a dataset before release. Each one names the part of the record that answers it.
Names drift; identifiers do not. Give the full sequence, structure or composition, not a nickname. Say which numbering convention you are using. Put a unit on every quantity.
Capture sequences or structure identifiers, units, ontology terms, the numbering frame.
A scan of genomics journals found gene names silently converted to dates by spreadsheet software in about a fifth of papers with supplementary Excel gene lists.6
The same sample gives a different number at another temperature, in another buffer, on another instrument. Conditions are what let a model explain variance instead of memorizing it.
Capture protocol version, instrument and settings, reagent lots, temperature, timing, operator or agent.
Automated experimentation helps here. It lets a lab capture metadata at every step of a workflow, which makes the experiment easier to reproduce and the data easier to learn from.7
A value without its spread teaches false confidence. Report replicates one by one, not only their mean. Keep each QC verdict together with the rule it was judged against. State the floor and ceiling of the assay.
Capture replicate-level values, QC pass or fail with the criterion, detection limits, calibration.
The literature is a record of successes, so a model trained on it has never seen failure. A dataset that keeps only the hits teaches a model where the peaks are and nothing about the ground between them.
Capture every sample that was run, including nulls, inactive variants and failed runs, flagged and not deleted.
One group transcribed the failed syntheses from its own notebooks and trained on them with the successes. The model’s recommended reactions worked 89% of the time; the chemists’ own judgment, 78%.8
Data from two plates, two weeks or two labs can be merged only if something ties them together. Shared controls and a reference sample carried through every run let a model learn the science and not the batch.
Capture controls on every plate, anchor samples across rounds, calibration to standard units, batch identifiers.
Scientific data is full of near-duplicates: homologous proteins, one compound in two assays, a time series cut at random. Ship the split with the data, and say what it is meant to hold out.
Capture train and test assignment, the grouping rule (by family, scaffold, date or lab), known overlaps with public sets.
A survey of machine-learning-based science found data leakage affecting at least 294 papers across 17 fields.9
A dataset with no license, or with terms that are silent on model training, is not ready however clean it is. Neither is one whose consent or embargo status nobody can state.
Capture license, consent and embargo status, a persistent identifier, how to cite.
A clear, accessible usage license is one of the FAIR principles.10
Every design choice encodes knowledge: what was expected, what it rested on, what was ruled out. None of it survives in a results table. It is the part of the record nobody keeps, and the subject of the next section.
Capture the question, the choice made, the alternatives rejected, a prediction written down before the result.
Laboratory software answers one question well: what happened. Plate, well, sample, condition, value. It does not hold the answer to the question a scientist would ask first: why does this well exist?
The first layer can be captured by instruments, automatically. The second has to be written down by whoever made the decision, a scientist or an agent, before the result is known. Kept together, they turn an archive into something a model can be scored against: a prediction, an outcome and a verdict for every choice.
Captured by instruments.
Eleven fields, and every one describes the well. None of them explains it.
Written down by the person or agent who decided.
Before the plate ran
After
Our working hypothesis is that the second layer is what separates a model that predicts assay results from one that can plan an experiment. It is a hypothesis, not a finding. The nearest evidence comes from surgery, where one group stopped labeling more video frames and instead harvested the reasoning in surgeons’ narrated teaching videos; models trained on it outperformed larger general-purpose models on questions about intent and rationale.11
11 Perez, Rau et al., arXiv (2026).
It also pays for itself before any model is trained. A lab that writes its predictions down first can count how often it was right, see which assumptions a whole round of designs was resting on, and tell a refuted idea from an experiment that never tested it.
Guidance on data usually stops once a dataset can be found and is well described. That is necessary, and a model needs more. This ladder is our working synthesis. The lower rungs map onto standards that already exist; the top rung is where we think the open problem is.
Files exist. What they mean lives in filenames and in someone’s memory.
Usable by the person who made it.
Persistent identifiers, standard formats, descriptive metadata and a license: the FAIR principles.10
Usable by another scientist who goes looking.
A datasheet explains how the data was collected and what it should not be used for.12 Machine-readable metadata such as Croissant lets ML tools load it without hand-written parsing.13
Usable by a training pipeline.
Conditions, uncertainty, negatives, comparability, splits and rights are all in the release, in the spirit of criteria such as NIH Bridge2AI’s seven dimensions of AI-readiness.14
Usable by a model that predicts outcomes.
The record also carries the decisions: why each experiment was run, what was predicted, and how the result changed the next one.
Usable by a model that plans experiments.
Sixteen checks, two for each question. Mark the ones your dataset already meets. Your marks stay in this browser; nothing is sent anywhere.
None of this starts from zero. These are the frameworks and community resources we lean on, and where each one helps.
| Framework | What it gives you | Reference |
|---|---|---|
| FAIR Guiding Principles | Findable, accessible, interoperable, reusable. The baseline for any shared dataset. | 2016 10 |
| Datasheets for Datasets | A structured account of why and how a dataset was made, and where it should not be used. | 2021 12 |
| Croissant | A metadata format from MLCommons that lets ML frameworks load a dataset directly. | 2024 13 |
| Bridge2AI AI-readiness criteria | Seven dimensions for biomedical data: FAIRness, provenance, characterization, pre-model explainability, ethics, sustainability, computability. | 2024 14 |
| Data readiness for scientific AI at scale | Readiness levels and processing stages for datasets large enough to train foundation models. | 2025 15 |
| An actionable framework for AI-ready data | Concrete steps for data publishers and repositories, from the Open Data Institute. | 2026 16 |
| Resource | Why it works | Reference |
|---|---|---|
| Protein Data Bank | One open archive and one format for macromolecular structures, with deposition required by journals. | 1971 1 |
| Open Reaction Database | An open schema for chemical reactions that keeps conditions and outcomes together, from benchtop work to automated screens. | 2021 17 |
Science4AI is maintained by Guo Ye, a data scientist at the DREAM Cloud Lab in Northwestern University’s Center for Synthetic Biology, one of the 20 nodes in the NSF Programmable Cloud Laboratories Test Bed. It grows out of the day-to-day work of making an automated protein-engineering lab’s output usable for training AI.
This is a working document and it will change. If something here is wrong, missing or out of date, open an issue or propose an edit on GitHub. A related project, Autonomous Lab Bench, proposes a cross-laboratory benchmark for scientific agents.
Names show where the maintainer and collaborating researchers work. They do not imply that any institution endorses this site. The views here are the maintainer’s own and do not represent NSF, Northwestern University or any other institution.