Why AI-for-Science data needs its own breakdown
The problem with robot data is that there is no ready-made source: every example has to be produced by someone. Scientific data has the opposite problem. Public databases are free, vast and built up over decades, but they record shapes (structures, sequences, expression profiles), not the outcomes a model actually needs to learn: does this molecule bind or not, what did this perturbation change, is this compound toxic in an animal.
So the AI4S data pyramid is held open by two core layers. One is wet-lab ground truth: the most expensive, the slowest and the closest to reality, but tied to its readout method and its lab. The other is model simulation and synthetic data: cheap and unlimited, but its fidelity drops the further it gets from reality. Upstream sit the databases and the instruments that collect data. Downstream are how models mix their data and how validation results flow back. This column follows that chain through nine chapters. Biology is the main focus, with chemistry and materials covered along the way.
Nine layers, from upstream to downstream
| Position | Layer | Data | How it is produced | Typical scale | Role in the model |
|---|---|---|---|---|---|
| Upstream 1 | Public databases and literature | PDB, UniProt, ChEMBL, CELLxGENE, AlphaFold DB | Global submissions, institutional curation, government funding | Protein Data Bank (wwPDB) 261,000 entries; UniProt TrEMBL 149 million entries; CZ CELLxGENE Discover / Census about 218 million cells | The pretraining base for every foundation model; free, so not a moat |
| Upstream 2 | Instruments, reagents and capacity | Sequencers, single-cell platforms, gene synthesis, mass spectrometry, cryo-EM | Vendors sell instruments and consumables | Sequencing about $1/Gb; single-cell sequencing about 0.5 cents per cell; gene-synthesis fragments from 7 cents/bp | Sets the unit-cost curve and bottlenecks of ground-truth data |
| Core A | Wet-lab ground-truth data | Binder binding assays, perturbation atlases, phenotypic imaging, ADME, organoids and animals | Labs, automation and people | Adaptyv Bio from $149 per protein; Tahoe-100M 100 million cells; JUMP-Cell Painting 1.6 billion cell images | A tiny share of entries, but the highest weight in fine-tuning and evaluation |
| Core B | Model simulation and synthetic data | MD/DFT trajectories, distilled structures, generated sequences, virtual cells | GPU/CPU compute | OMol25: 100 million+ DFT calculations, about 6 billion CPU hours (secondary source); about 75% of AlphaFold2 samples come from self-distillation | An amplifier; fidelity decays step by step along "structure > pose > affinity > cell > efficacy" |
| Downstream 1 | Model training recipes | 27 models, including AF3, Boltz-2, ESM3, Evo 2 and scGPT, plus 1 user case study | Public databases as the base, self-distillation to amplify, a sliver of ground truth to finish | Boltz-2's roughly 1.2 million affinity binders all come from public databases | None of the open-source or academic models found publicly discloses buying commercial data |
| Downstream 2 | Validation loops and self-driving labs | Competitions, frontier-lab slow loops, cloud labs, self-driving labs | Design → synthesise → assay → feed back | Anthropic protein-binder experiment: 1,320 designs / 354 binders; about 3 weeks per round | Who owns the returning data decides who owns the flywheel: public, customer-private or platform-exclusive |
| Cross-cutting A | Benchmarks and formats | CASP, PoseBusters, ProteinGym, TDC, Virtual Cell Challenge | Prospective blind tests, in-house ground truth, retrospective curation, expert-set questions | An audit of 51 benchmark configurations found widespread leakage | Whoever judges what works defines what data is worth |
| Cross-cutting B | Vendor landscape and policy | Data-product vendors, CROs turning into data businesses, open operators, cloud data marketplaces | Subscription, per-record, revenue share, data for equity | Tempus AI data business about $316 million; only two public precedents for buying training data with cash | Policy is tightening cross-border flows and opening up non-animal data |
| Cross-cutting C | Prices, costs and mix | Price per data point, ground truth vs compute, funding | — | From 1 cent per cell to $28,000 per lab monkey; the bases vary enormously | The numerical base table for the whole column |
Price per data point and model data mix
Price per data point: from 1 cent to $28,000 (log scale, USD)
Wet-lab ground truth and compute costs on one log axis. Estimates multiply disclosed GPU/CPU hours by an assumed unit price. Company figures are self-reported, not third-party price lists. RMB converted at 7.1.
| Data point | Low ($) | High ($) | Basis |
|---|---|---|---|
| Adaptyv binder validation (per protein, 2 replicates) | 149 | 149 | Company website |
| Ginkgo ADME-One (per five-assay panel) | 199 | 199 | Company figure |
| Western CRO ADME (as cited by Ginkgo) | 2,000 | 5,000 | Spoken by Ginkgo management, not a third-party price list |
| Chinese CRO ADME (as cited by Ginkgo) | 1,000 | 2,500 | Spoken by Ginkgo management, not a third-party price list |
| Single-cell transcriptomics · Parse (per cell) | 0.01 | 0.01 | Company figure |
| Single-cell transcriptomics · independent benchmark (per cell) | 0.05 | 0.21 | Genentech-led benchmark |
| OMol25 DFT (per calculation, compute) | 1.2 | 3 | Estimate: 60 core-hours × $0.02–0.05 |
| Boltz-2 affinity inference (per molecule) | 0.02 | 0.02 | Estimate: 18 seconds × $4 per GPU hour |
| FEP (per compound, about 4 GPU hours) | 16 | 16 | Estimate: 4 GPU hours × $4 |
| FEP (per molecule, media figure) | 100 | 100 | Secondary source |
| Cryo-EM cloud compute (per structure) | 50 | 1,500 | Paper |
| Cryo-EM Krios instrument time (per day) | 1,000 | 4,800 | Facility price lists |
| Cynomolgus monkey · Mauritius average export price (each) | 6,425 | 6,425 | Media report |
| Cynomolgus monkey · China (each) | 10,600 | 28,200 | RMB 75,000–200,000+ at 7.1; government procurement cases plus media |
Training data composition per model: public · distilled/synthetic · private ground truth (share)
Shares are derived from entry counts or sampling ratios disclosed in each model's paper or model card. Units differ (sequences, cells, sampling weights), so use the chart for structure only; absolute amounts are not comparable across models. The public portion of Tahoe-x1 is unverified, so it is not shown.
| Model | Released | Public | Distilled / synthetic | Private truth | Basis |
|---|---|---|---|---|---|
| AlphaFold 2 | 2021 | 25% | 75% | 0% | Sampling ratio: clustered PDB 25 / self-distillation (355,993 sequences) 75; as restated in NVIDIA's model card (secondary) |
| Boltz-2 affinity | Jun 2025 | 50% | 50% | 0% | Entry counts: ChEMBL + BindingDB optimised set of about 1.2 million binders, plus 1.2 million synthetic decoys; all from public sources |
| ESM3 | 2024 | 88% | 12% | 0% | Sequence counts: 2.78B natural; 3.15B after synthetic augmentation, the 370 million difference being an estimate; model-card figures |
| scGPT | 2023 | 100% | 0% | 0% | Cell count: >33 million normal human cells from CELLxGENE (secondary) |
| Geneformer | 2023 | 100% | 0% | 0% | Cell count: Genecorpus-30M, about 30 million (later expanded version about 104 million) |
Key findings
- The base is free; the moat is not in owning data but in owning ground truth. About 75% of AlphaFold2's training samples come from self-distillation and 25% from the Protein Data Bank (wwPDB). Evo 2's OpenGenome2, with 8.84 trillion nucleotides, is open source under Apache 2.0. CZ CELLxGENE Discover / Census offers hundreds of millions of cells for free. Public databases and distilled data have solved "scale"; what remains scarce is paired ground truth with outcome labels.
- Ground-truth data now has public prices, but only in a narrow band. Adaptyv Bio at $149 per protein and Ginkgo ADME-One at $199 per panel are the only two public price lists that charge per design or per assay. Perturbation atlases, phenotypic imaging and structure determination are still project-based or donation-based. The comparison point is the RMB 500–1,000 per hour quoted for real-robot teleoperation in the robotics column; biology has no such "hourly rate" yet.
- Simulation versus ground truth is not "cheap versus expensive" but a fidelity ladder. A single DFT point costs about $1–3 (estimate), a Boltz-2 affinity inference takes about 18 seconds on a consumer GPU (media figure), and one FEP run costs about $16 (estimate). But Boltz-2's Pearson of 0.62 on the FEP+ benchmark is self-reported by its developers, and foundation models at the cell-perturbation level have not beaten simple baselines. Computation is an enrichment filter, not a substitute for experiments (see the sim-to-real fidelity ladder).
- Recipes are more open than in robotics, but the ground-truth part is the most opaque. Across 27 models and 1 user case, the entry counts of public and distilled sets are mostly findable. Private wet-lab data appears in only three forms: "used for evaluation" (Boltz-2 × Recursion), "federated fine-tuning" (OpenFold3 × pharma companies) and "competition data released as open source" (Anthropic × Adaptyv/Twist). Not one has publicly said it bought commercial training data.
- The bottleneck in validation loops is latency and ownership, not equipment. One round of a binder loop takes about 3 weeks. Most cloud labs selling "machine time" have not scaled: Strateos moved to on-premises deployment, and about 80% of Emerald Cloud Lab's capacity is tied to a single university. The "platform-exclusive data" story of Lila Sciences and Periodic Labs has not yet produced a peer-reviewed closed-loop result. Whoever owns the returning data owns the flywheel.
- Benchmarks leak widely, and "question-setter = grader = service provider" is a structural problem. An audit of 51 benchmark configurations found train–test leakage in most. The ground truth for BenchBB, Proteinbase and the Arc Virtual Cell Challenge all comes from a single lab. A neutral, multi-lab, pre-registered validation benchmark still has no operator, which is the clearest gap outside the pricing chapter.
- Capital and policy point the same way: money goes to "model plus in-house wet-lab loop", while policy tightens cross-border data. The largest private rounds of the past 24 months include Chai Discovery's $400 million Series C (a single secondary source, unverified) and about $550 million raised in total by Lila Sciences. Xaira Therapeutics' $1 billion in Apr 2024 falls outside the window and serves only as a baseline. Since Apr 2025 the NIH has barred countries of concern from accessing its controlled-access databases. The BIOSECURE Act became law through the FY2026 NDAA, and human genetic resources leaving China require review. China's role looks more like a "validation factory" than a "data exporter".
Compared with the robot-data column
| Dimension | Robotics / embodied AI | AI for Science (bio) |
|---|---|---|
| Most expensive layer | Real-robot teleoperation (RMB 500–1,000/hour) | Wet-lab ground truth (from $149 per protein; animal studies over $10,000 per animal) |
| Cheapest layer | Human video (on the order of 100,000 hours already free and open source) | Public databases (free, hundreds of millions of entries) |
| Failure mode of synthetic data | Sim-to-real gap, world-model hallucination | Fidelity ladder: structures are usable, affinity and cell-level data are not trustworthy |
| Recipe transparency | Low (only two models give per-source hours and weights) | Medium (public and distilled data findable; ground truth opaque) |
| Precedents for buying with cash | Yes (Motus2 credits five commercial data vendors) | Almost none (only two deals, both clinical real-world data) |
| Returning data | Deployment rollouts plus human corrections, tied to the embodiment | Validation results, tied to the readout method and the lab |
| Policy | Training-ground subsidies, dataset standards | Cross-border data restrictions, opening of non-animal data |
Both fields point to the same thing: the cheap layers are being flattened by open source, and value is migrating to the two ends, producing ground truth and judging validation.
The robotics column is at AI Data: Robotics.
Chapters and entity cards
| # | Chapter | What it covers |
|---|---|---|
| 01 | Public databases and literature | The free, vast but biased top of the AI4S data chain: the training base for structure, sequence and single-cell models. |
| 02 | Instruments, reagents and sequencing/synthesis capacity | Sequencing, single-cell, gene synthesis, mass spectrometry and automation: the collection equipment and the cost-per-data-point curve. |
| 03 | Wet-lab ground-truth data production | The most expensive layer, the closest to truth, the most tied to its readout: who makes it, who pays, who owns the data. |
| 04 | Model simulation and synthetic data | Training data produced by computation rather than experiment: physics simulation, distillation and self-labelling, generative synthesis, and their fidelity and cost. |
| 05 | Model training data recipes | Which public databases, distilled data and private ground truth more than 30 biology and chemistry foundation models used, and in what shares. |
| 06 | Validation loops and self-driving lab feedback | Prediction, wet-lab validation, and results flowing back as training and reward data: who runs them, how fast, who owns the data. |
| 07 | Benchmarks, evaluation and data formats | Where benchmark ground truth comes from, who sets and who grades, evidence from leakage audits, and data-format standards. |
| 08 | Data vendor landscape and policy | Who sells biological data, capital flows over the past 24 months, and how policy is rewriting data supply. |
| 09 | Prices, costs and mix | Unit prices for wet-lab ground truth, simulation compute, equipment and data licences, plus model data-mix ratios and funding proxies. |
The column also includes 243 entity cards, browsable by type:
This section is personal research. Every number is labelled in the original with its source and basis; self-reported company numbers are marked as company figures, and independently unverified ones are flagged. Judgements hold as of the date written and are not investment or R&D advice.
Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Oct 15, 2026.