One repository, every lab’s cells.
Free for labs that share, a membership for teams that use the data, and a place where cell data converts.
Vision
Every lab that images cells should know where its cells sit among everyone else’s, and every dataset made from cells should carry that answer with it. Datum Line is one shared repository of raw cell images from every lab, on one map, and the product is a membership to it. Labs join free by sharing the control images they already take, and keep a share of what their data earns. Teams that use the data pay: screening teams first, for a check on every batch they run; AI teams the most, for aligned images from every lab by API. The arrow after that is left open on purpose.
The habit comes first, the way Strava’s segments place a run among everyone who ran the same stretch of road. Here the road is the cell line, and the field is every lab that ran it. Everything below serves one rule: uploading must be effortless, and the answer must be worth looking at.
01Why now
Cell data has a price. One vendor pays its experts more than $1.5M a day, about $85 an hour on average, at a $10B valuation. [10] Roche and Genentech paid Recursion a second $30M milestone for an accepted whole-genome map of microglial immune cells; $213M from the partnership to date. [11] A $1.8B virtual-cell data initiative, part new money and part existing funding; Meta, Google DeepMind and Isomorphic put in $300M together and get a year of exclusive access first. [12]
Cell data does not convert. EU-OPENSCREEN ran one Cell Painting protocol at 4 sites, with cells, serum and dye lots supplied centrally. Within a site, 84 to 94% of compounds replicate; across sites, 54 to 68% pair, and the authors call the profiles “not directly comparable between imaging sites” [13]. On the public JUMP data, a classifier names the lab from DMSO wells alone at 99.9% against 9.1% chance. Whoever makes cell data comparable sets its exchange rate.
02The path
| Stage | What it is | Who pays | What has to be true | Status |
|---|---|---|---|---|
| Map | every lab’s DMSO wells placed on one map; the box, the code and the public reference | nobody: open, and never crippled | labs upload the controls they already image | today: 11 labs, 129 batches, 1,871 plates of public data |
| Members | the repository, and a check on every batch: where the cells sit, which lab they look like, what moved | teams that screen all year, $5–15k per cell line per quarter; companies that buy cell data, $10–50k per dataset | the map holds for the line | next: the first paying member, a screening team; prices proposed on board A15 |
| API | aligned raw images from every lab, each traced to the lab that made it | AI labs and model builders, by use, and the most; each lab gets a share | labs choose to share; the map makes labs agree | then |
| … | what the map makes possible next: journals asking for raw control images with every paper, alerts when a lot drifts across labs, a graded benchmark, models trained on every lab’s controls | open |
03The habit loop
| Step | In Datum Line | Design rule |
|---|---|---|
| Trigger | the microscope finishes a plate; a new line or lot arrives | the trigger is the instrument, not an email |
| Action | nothing: the export folder syncs | any click is a leak; the reference starves on friction |
| Reward | where you landed, which features moved, who else moved | real uncertainty: nobody knows their position until they look |
| Investment | your history becomes a trajectory; your baseline tightens | leaving costs your own past, which is honest switching cost |
04Six magic moments
Each moment turns an image the lab already has into something it could not see before, and each is ordered by when a new member meets it.
| # | Moment | What the lab sees | When |
|---|---|---|---|
| 1 | Your past, diagnosed | old plates dragged in on day one appear as a path across the field map: "you left the envelope in March" | first ten minutes |
| 2 | Show the cells, not the score | when a plate moves, its nuclei next to their nearest neighbours from labs in distribution: same size, more texture | every flag |
| 3 | The fix | what the in-distribution labs do differently: seeding density, media or serum lot | after a flag |
| 4 | A forecast | "at your current drift, this line leaves the envelope in about five weeks" | after three readings |
| 5 | Someone else's plate saves your screen | a lot you use is drifting at other labs, before your own plates show it | once labs share lots |
| 6 | The check as currency | a verifiable link for a methods section, a CRO handoff, a grant report | when results leave the lab |
Moments 1 to 4 work for a single lab on day one, against the public reference. Moments 5 and 6 exist only once the network does, and no competitor can copy them without the labs.
05Membership
Membership is access to the repository for a period, not ownership of the data. A lab that shares joins free; everyone else pays by how they use it. The check on every batch is what screening teams pay for first; aligned images from every lab are what AI teams pay the most for. A biotech’s first step costs nothing: a backup of its raw images, off-site and immutable, the copy cyber insurers ask for, encrypted with its own key; its control wells join the map when it chooses.
| Who joins | What membership gives | Price, proposed | Role |
|---|---|---|---|
| labs and core facilities | the repository and the map; a check on every batch; the leaderboard | free: sharing controls earns it | the map |
| biotech companies | their raw images backed up off-site and immutable, private to them; their controls join the map when they choose | free | the hook |
| teams that run cell screens all year | a check on every batch they run, against every lab on the same line | $5–15k per cell line per quarter | first revenue |
| AI labs and model builders | aligned raw images from every lab, by API | the most: priced by use, not yet set | main revenue |
| companies that buy cell data | each dataset they buy, placed on the map | $10–50k per dataset | data revenue |
| benchmark partners | a test set with faults planted on purpose | $50–200k per year | the standard |
| What members get | Why it is worth paying for |
|---|---|
| The check | the comparison a lab cannot make alone: on the public data, 6 of 129 batches sat off the map, in 3 of 11 labs (results/e9) |
| Aligned data | it converts: on the map, agreement between labs on the same drugs goes from 0.75 to 0.78 (board A14) |
| Early access | a window before data goes public, which consortia already sell: a year, at the virtual-cell initiative [12] |
| The benchmark | AI labs and model builders graded on planted faults, with a cell-count baseline on every task (board A16) |
What never gets sold. A lab’s treated data, its results, its name without consent, a ranking of instruments, or a verdict for hire. Membership buys access and the check; it never buys a better verdict. Prices and shares are proposals (boards A15 and A22) until the first partner labs and the first members set them.
06Trust, without the token
The open-science half of the product, and the reason a lab would share.
| Promise | How | Status |
|---|---|---|
| Control wells only | nothing else is shared: no compound, no target, no result; the frame is applied on the lab’s side | today |
| A private backup | raw images a company backs up with us are encrypted with its own key; Datum Line cannot read them | proposed |
| Every lab leaves a fingerprint | 99.9% of DMSO wells name their own lab on plates the model never saw, so a dataset claimed from one lab and made in another is caught | today |
| Signed and timestamped | every check signed, its fingerprint written to a public log: anyone can verify it, nobody can backdate it; a blockchain is one such log, not a requirement | proposed |
| The lab keeps its data | sharing is a choice and can be withdrawn; a share of every membership goes to the lab that made the data | proposed |
| Known members | members verified at sign-up; unusual access flagged; every access in a log the members can read | proposed |
| Open code and data | the code is MIT on GitHub, the public data is JUMP’s under CC0, and every number comes from a script anyone can rerun | today |
| Low risk by design | images of cell lines: no human genetic data and no pathogen data; the map holds DMSO wells only | today |
For a decentralized-science audience this is the point: open data, verifiable records and contributor upside, without a token to speculate on and without betting on anyone’s results.
07The game
Every batch a lab sends is ranked against every batch the field has seen, on its DMSO wells alone: how typical the cells are, how alike the plates are, how steady the lab is from batch to batch, and cells per well against the field. Nothing in the score touches a treated well or a result, so the only way to climb is better culture and imaging practice. On the public data, lab 5 holds the field’s longest streak, 33 batches (results/e14). A lab sees its rank, its personal best, its streak and the lab to beat.
| Do | Don’t | Why |
|---|---|---|
| rank practice, never results | rank a lab’s results or its drugs | the only way up is better culture and imaging |
| labs are numbers until they choose a name | publish a lab’s name or rank without consent | shame makes labs hide their worst plates, which teach the field most |
| credit for uploading flagged plates | penalise a batch for sitting off the map | a failure uploaded is worth more than a success |
| explain every flag with cells and a likely fix | show a bare red dot | a verdict without a reason reads as judgement |
| let the instrument trigger uploads | send nudges to protect a streak | the reward is the answer, not the visit |
08Metrics
North star: the share of a lab’s plates that upload with no human action. If it is high, the loop turns; if people have to remember, nothing else matters.
| Metric | What it tells you |
|---|---|
| automatic upload share | the habit exists without effort |
| flagged plates uploaded, not deleted | labs trust the product with bad news |
| labs per line, lines per lab | the map is widening |
| labs sharing to the repository | labs trust the repository with their data |
| paying members, and how many renew | membership is worth paying for, again |
| share of each membership paid to the lab | contributors are earning, so more will share |
| checks cited outside the lab | the check has become currency |
09What could go wrong
| Risk | Why it is real | Answer |
|---|---|---|
| cold start | no map without labs, no labs without a map | seed from JUMP’s public controls; U2OS first, any other line from its first batch |
| labs will not share | data is a lab’s edge, and pharma data is intellectual property | a lab that keeps its images private still gets the check, as a paying member, the way screening teams do; sharing is free, optional and pays the lab |
| AI teams train on public data for free | JUMP and other public sets are open to anyone | the open map stays free; members pay for what public data lacks: every lab aligned, new batches as labs share them, the lab behind every image |
| public data must stay free | NIH-funded labs must share the data behind their findings | charge members for the check, alignment and compute, never for access to data a lab owes the public |
| buyers want designed data at scale | the deals above buy factories and consortia, not leftovers | place factory and consortium data on the map too; the benchmark grades their models |
| a platform builds its own map | phenomics companies hold the most images | no one can check their own drift against themselves; the reference has to be neutral and open |
References
- Chandrasekaran SN, et al. JUMP Cell Painting dataset. bioRxiv, 2023.
- TechCrunch. Mercor quintuples valuation to $10B with $350M Series C. 27 October 2025.
- Recursion. Recursion reports third quarter 2025 financial results and provides business update. Press release, 5 November 2025.
- Jungco K. Google, Meta join $1.8B push to build AI “virtual cells”: here’s what we know. eWeek, 9 October 2026.
- Wolff C, …, Schmied C. Morphological profiling data resource enables prediction of chemical compound properties. iScience 28(5):112445, 2025.