Biology Atlas

Map / Assays

assay · intermediate · partly written

High-content imaging and morphological profiling

Photograph treated cells stained with several dyes and turn each cell into a long list of measurements, so compounds can be compared by how the cells look rather than by one pre-chosen readout.

Also called: Cell Painting, image-based profiling, morphological profiling, high-content screening

In his words: what imaging adds

He said: "You just may want to understand the morphologies of certain cells treated with and without a compound to understand how it affects things. I don't really understand the purpose of this, like how it differs." 1

Amit's hypothesis

How it differs from a viability readout

Your description (cells treated with and without a compound, then look at their shape) is right. The difference is that a viability screen gives one number per well, chosen in advance. Imaging stains many cell parts and measures about 1,500 features per cell, so each compound gets a fingerprint and you can ask which known drugs it resembles. The question is not fixed beforehand. 23

Reported, not yet replicated

What is physically measured, step by step more

The dyes (Bray 2016, Table 1): Hoechst 33342 stains DNA (channel 1); concanavalin A with Alexa 488, the endoplasmic reticulum (2); SYTO 14, nucleoli and cytoplasmic RNA (3); phalloidin with Alexa 568, F-actin (4, shared); wheat-germ agglutinin, Golgi and plasma membrane (4, shared); MitoTracker Deep Red, mitochondria (5). That is six dyes, five channels and eight components; the endoplasmic reticulum and RNA dyes overlap spectrally and separate imperfectly.

The chain: seed cells, add compound, stain, image (nine sites per well at 20x), correct illumination, run a machine-learning quality filter, segment each cell and measure about 1,500 features, average the cells to one profile per well (the 2013 paper concedes this does not explicitly model heterogeneity among cells), normalise to control wells, and compare profiles by correlation. 245

Reported, not yet replicated

What it cannot tell you more

Target or mechanism: in the 2013 screen, 75 of 203 active compounds had annotations, and clusters were enriched for shared mechanisms, not proven ones. The authors hedge their one mechanistic explanation (shared basic tertiary amines for one cluster): "It is possible that this chemical feature and cellular mechanism underlie the shared effect."

Biology without a visible footprint: the protocol's limitations section says some processes "may simply not yield any relevant discernible morphological phenotypes" in the chosen cell type, time and stains. Potency and exposure cannot be read from a single concentration.

In the JUMP preprint's own words compounds sharing targets "often" show similar profiles; "often" is not "reliably". 526

Reported, not yet replicated

Where it breaks and how practitioners check more

Batch and plate effects: seeding, growth, stain lots, instrument and well position change the numbers. Checks: randomise wells and spread replicates across plates, use DMSO controls, apply batch correction (Harmony and Seurat RPCA ranked in the top three of ten methods on JUMP).

Cell count masquerades as biology: many benchmark assays track cytotoxicity, so compounds that change cell number score well. A 2026 paper from the profiling field's own lab found a cell-count-only baseline scores unexpectedly well on common bioactivity benchmarks, while Cell Painting profiles beat it across a 24-assay protein-target set. Add a cell-count baseline.

Too few cells weakens profiles (aim for about 2,500 per well). Debris and bleed-through can look like phenotypes (a supervised quality classifier is used; no study of fluorescent compounds was opened). Averaging hides subpopulations. With about 1,500 features and few replicates, overfitting is a risk. A match to a known drug should be confirmed with an orthogonal assay. 2478

Reported, not yet replicated

What the field says about its own limits deep

The 2021 review lists open questions: whether reversing a phenotype gives a drug, whether profiles are reproducible across batches and sites, whether deep networks beat conventional features; polypharmacology makes profiles resemble no single mechanism; cultured cells miss whole-organism effects. It also cites two studies reporting 60- to 250-fold higher hit rates than the original assays, which is a review's summary and not an independent test.

A 2014 review by the same community counted 118 high-content screening papers, found 60 to 80 percent used only one or two image features, and argued the Z'-factor (a screening quality score) is often not ideal for high-content screening because it needs a single univariate readout. 93

Reported, not yet replicated

Imaging plus machine learning at industrial scale: Recursion deep

Recursion runs this step as a factory: it says up to about 2 million experiments a week and a 50+ petabyte private image set (platform page, company claim). It calls proprietary data its differentiator and has released a small public slice. Its record so far, as reported: roughly 330 compounds and 17 months per programme (company claim; the release says not independently verified), one phenotype-derived drug (REC-4881) with single-arm Phase 2 data, no approved drug, and in 2025 several programmes deprioritised or paused. Cost per well, per screen or per programme is not public. 1011121314

Reported, not yet replicated

Question answered

Which known compounds or gene perturbations does this treatment make the cells resemble, and does the treated cell look more like disease or like health? The question is not fixed before the plate is made, which is the difference from a viability readout that gives one number chosen in advance. 253

Reported, not yet replicated

Readout physical

Fluorescence intensity and its spatial pattern in each imaging channel: light of one colour excites a dye, the dye re-emits at a longer wavelength, and a camera records where. Cell Painting uses six dyes in five channels (DNA, endoplasmic reticulum, nucleoli and RNA, actin, Golgi and plasma membrane, mitochondria). Software then segments each cell and measures about 1,500 features (size, shape, intensity, texture, correlation between channels). The unit of analysis is a profile (a vector of numbers), not an image. 24

Reported, not yet replicated

Not measured

It does not identify a target or mechanism: it reports similarity to annotated compounds, and a match is a hypothesis. It cannot see biology that leaves no visible trace in the chosen cell type, time and stains. Profiles differ between cell types and cannot be fully normalised away. A single concentration says nothing about potency or exposure. Averaging cells to one profile per well hides subpopulations. 25

Reported, not yet replicated

Setting

Where in cascade

Sample prep

Seed roughly 1,500 to 3,000 cells per well of a 384-well plate (typical; U2OS is a common cell line) and culture for 2 to 3 days. Add compound from a DMSO stock and incubate (24 or 48 h used; 48 h gave more distinguishable profiles in the protocol group's limited testing). Live-stain mitochondria, fix, permeabilise and add the other five dyes: 2.5 to 3 h for a batch. Wells at plate edges are excluded from imaging. 2

Reported, not yet replicated

Conditions

Perturbation type and exposure time (compounds, RNAi, overexpression or viral infection; 24, 48, 72 and 96 h were used), cell line (13 culture types were tried, including primary cells and co-cultures), and concentration all change the profile. The 2013 screen used 10 uM for 48 h. Randomising well position and spreading replicate plates across batches is advised because seeding, stain lots, instrument and well position change the numbers. Aim for about 2,500 cells per well or at least three imaged sites. 254

Reported, not yet replicated

Throughput

Imaging takes 1 to 3.5 h per 384-well plate, typically nine sites per well at 20x: 384 x 9 x 5 = 17,280 images per plate. By our arithmetic, one instrument imaging continuously would handle roughly 7 to 24 plates a day (about 2,600 to 9,200 wells), but feature extraction (about 20 h per batch of plates) and calendar time (2 weeks culture and imaging plus 1 to 2 weeks analysis) set the real pace, and a laptop is not suitable above roughly 1,000 images. Public scale: the JUMP consortium profiled 115,796 compounds across 12 laboratories. The company Recursion states up to about 2 million experiments a week (company claim, not independently verified). 2610

Our inference

Time per run

About 2 weeks for culture and imaging, plus 1 to 2 weeks of analysis (protocol abstract). Staining takes 2.5 to 3 h per batch, imaging 1 to 3.5 h per 384-well plate, feature extraction about 20 h per batch. 2

Reported, not yet replicated

Cost

Per sample

from 0.26 USD per well, stain reagents only (Reagents only at scale: about 26 cents per well. No labour, instrument time, compute or storage. Imaging, feature extraction and image storage (450,000 images, 41 million cells, 6.7 TB across 23 microscope settings in the multi-site study) are the larger hidden costs. Upper end unknown.; as of 2023 (paper accepted 2023-08-08; PDF read 2026-11-17))4

Per campaign

from 15,478 USD per 25,000-compound primary screen, internal academic rate (Academic core recharge rate sheet, high-content imaging, internal user rate, reagents supplied by the investigator's lab and so excluded; the external rate is about 1.9 times the internal one per the concept report. Compare $7,041 for a biochemical screen at the same core. Not an industry cost. This is an academic core-facility recharge rate, not a commercial or contract-lab price; no commercial price was found.; as of 2020 (sheet file dated 2020-05-27 by its file name; the date inside the sheet is unconfirmed); likely higher now)15

Engineerable variants

Change a dye

Possible but needs substantial optimisation: for example the endoplasmic reticulum dye must be bright enough to mask the RNA dye sharing its channel. 2

Reported, not yet replicated

Fewer channels or brightfield

The multi-site study tested four, five and six channels and brightfield. Its main finding was that technical quality held across microscopes while aggregated scores differed by setup (per-setup figures not extracted here). Recursion claims about 95% of genetic perturbations seen with Cell Painting are also captured from brightfield (company data, no cost figure, not validated externally). 416

Reported, not yet replicated

Live-cell version

Suggested in the protocol (live-compatible stains or fluorescent-protein tags); not tested there. 2

Reported, not yet replicated

Different perturbation or cell line

Compounds, RNAi, overexpression or viral infection, at different exposure times. Short RNAi reagents against the same gene often look unlike each other. 2

Reported, not yet replicated

Three-channel assay repurposed

A three-channel glucocorticoid-receptor translocation image screen was reused to predict compound activity in other projects (not full Cell Painting). 17

Reported, not yet replicated

Use cases

Janssen image-screen repurposing

A three-channel glucocorticoid-receptor translocation image screen was repurposed to predict compound activity in two other projects; the abstract reports hit rates up 50- to 250-fold and more diverse hit chemistry. It decided which compounds to test; cost and clinical outcome are not in the abstract. 17

Reported, not yet replicated

Broad library design

A pilot profiled over 30,000 compounds with image-based morphology and gene-expression profiles to choose library sets with high hit rates and diverse biological performance. Outcome beyond the abstract was not examined. 18

Reported, not yet replicated

SARS-CoV-2 morphological profiling

Profiling of infection across 1,425 FDA-approved compounds and clinical candidates gave 17 hits; lactoferrin was active at nanomolar concentration. The abstract calls it morphological cell profiling and does not name Cell Painting, so treat it as image-based profiling. 19

Reported, not yet replicated

Recursion REC-4881

The company says REC-4881 came from unbiased phenotypic screening with its foundation models. Phase 2 data in familial adenomatous polyposis: 13 of 15 patients (87%) had polyp-burden reductions, median 47% at 12 weeks. Single-arm and tiny; company-reported. No drug from the platform is approved, and in 2025 several programmes were deprioritised or paused. 1314

Reported, not yet replicated

JUMP-Cell Painting consortium

A shared reference dataset made so profiles from many laboratories can be compared: 1.53 billion single-cell profiles of 115,796 compounds in 12 laboratories. Its own headline is modest on biology: reproducibility reported 34% lower than single-lab data, and compounds sharing targets often (not reliably) show similar profiles. 6

Reported, not yet replicated

Before after

Status quo before

Earlier profiling assays used minimal numbers of labels, and a 2014 review of 118 high-content screening papers found that 60 to 80 percent used only one or two image features per sample, reducing each well to one number. 53

Reported, not yet replicated

What forced it

The stated aim was to light up as many cell parts as possible without leaving standard microscopes. That a viability-style readout cannot say what a compound did to the cell is our reading of why this mattered. 2

Our inference

The trick

Six dyes in five channels mark eight cell components, and software turns every cell into about 1,500 numbers, so compounds can be compared by correlation of profiles instead of by one chosen readout. 25

Reported, not yet replicated

Why it worked

Because the readout is not chosen in advance, one image set can answer questions nobody had asked when the plate was made, and a standard stain set made shared public data possible. The limits are open: the field's own 2021 review lists whether reversing a phenotype yields a drug, cross-batch reproducibility and polypharmacology as unresolved. 96

Reported, not yet replicated

Capital

JUMP consortium funding 2020-07-10

Massachusetts Life Sciences Center Bits to Bytes funding: about $750,000 per project, $6.7 million in total. 20

Reported, not yet replicated

Recursion cash and burn 2025-12-31

Cash of $753.9M at 2025-12-31; 2025 net loss $644.8M, operating cash used $371.8M, revenue $74.7M (from partnerships, not product sales). Company-reported. 12

Reported, not yet replicated

Recursion partner payments and 2026 raise 2026-11-02

Roche/Genentech $216M and Sanofi $144M received to date (upfront plus milestones); cash $561.5M at 2026-09-30; an October 2026 offering netted about $187.2M. Company-reported. 13

Reported, not yet replicated

Emerging needs

Our reading of the reports: a published all-in cost per well and per screened compound; a head-to-head hit rate against a conventional screen; evidence that a profile adds something a cell count plus a dose-response curve does not (the 2026 cell-count baseline result makes this the test); reproducibility across batches and sites; and a confirmed case in which a Cell Painting lead became a candidate drug. 896

Our inference

ML role

Does

Classical features come from CellProfiler (hand-designed measures), and a machine-learning filter finds bad images. Models then predict activity and mechanism from profiles: structures, Cell Painting and L1000 each predicted activity well in 6 to 10% of 270 assays and 21% combined; mechanism-of-action classification reached 89 to 98% on 10 classes (a review's figure, with plates shared between train and test as the hidden risk). Recursion trained a 300M-parameter vision transformer (Phenom-1) on 95M images, using 128 A100 GPUs for about 20K GPU hours (third-party blog). Batch correction: in a benchmark of ten methods on JUMP, Harmony and Seurat RPCA ranked in the top three. 22122237

Reported, not yet replicated

Cannot

It cannot name the target: similarity is not binding. It can learn the dataset instead of the biology: a cell-count-only baseline scores unexpectedly well on common bioactivity benchmarks (many well-predicted assays track cytotoxicity), although Cell Painting beat that baseline across a 24-assay protein-target set. Leaky splits where train and test share plates inflate scores. 822

Reported, not yet replicated

Data availability

JUMP data on the Cell Painting Gallery (AWS) public

Images and profiles from the JUMP consortium: 1.53 billion single-cell profiles for 115,796 compounds across 12 laboratories (preprint abstract); the Gallery held 656 TB as of January 2024. 24256

Caveats. The licence is reported differently: CC0 on the AWS registry page, CC BY 4.0 in the Gallery paper. The Gallery is large, so use a subset. The compound data are mostly one cell line (U2OS); reproducibility was reported 34% lower than in single-lab data. Batch and plate effects must be handled.

Contested

RxRx3-core (Recursion) public

A public slice of Recursion's imaging data: 222,601 images, 18 GB, under 1% of its data, released 2025-05-28. 11

Caveats. A company-released subset of data the company calls its differentiator; licence not checked here.

Reported, not yet replicated

Run the assay yourself in-house

Stain reagents cost about 26 cents per well at scale; the larger costs are imaging time, feature extraction and terabytes of images. 4

Caveats. No labour, instrument or compute price found.

Price. from 0.26 USD per well, reagents only (Reagent price at scale from the multi-site Cell Painting study.; as of 2023)4

Reported, not yet replicated

Cell Painting staining kit (Thermo Fisher Image-iT) purchase

A kit of 10 plates of Cell Painting stains: a reagent line only. 26

Caveats. Reagents only: no imaging, instrument time, labour or analysis; volume per well not confirmed, so no per-well figure is recorded. No CRO price for Cell Painting was found.

Price. 1,195.56 to 1,458 USD per 10-plate kit (Thermo Fisher catalogue page: $1,195.56 online offer to 2026-12-31, $1,458.00 list; read 2026-11-17.; as of 2026-11-17)26

Reported, not yet replicated

Teach this

In one sentence

Imaging stains many parts of the cell and turns each cell into about 1,500 numbers, so you compare what compounds do to the whole cell instead of reading one chosen number. 2

Reported, not yet replicated

The five-minute version

A viability assay asks one question you chose in advance. Cell Painting asks none: six dyes light up eight cell parts, a camera photographs each well, and software measures about 1,500 features per cell. Averaged per well, that is a profile; two compounds with similar profiles probably do something similar, so a new compound can be matched to known drugs. The costs are mostly hidden: reagents are cheap (about 26 cents a well) but imaging, analysis and terabytes of images are not. The traps are batch effects and cell number: a 2026 paper found cell count alone predicts many benchmarks. A match is a hypothesis, never a target. 248

Reported, not yet replicated

What a listener will ask

Why not just use a viability assay?

It gives one pre-chosen number per well and cannot say what a compound did to the cell; imaging keeps the whole appearance so you can compare compounds by look-alike.

Does it tell you the target?

No. It tells you which annotated compounds the profile resembles; a target needs confirmation by an orthogonal assay.

Could a plain cell count do as well?

On many benchmarks, surprisingly, yes (a recent result from the field's own lab); Cell Painting beat the count baseline on a protein-target benchmark, so test against the baseline.

The common wrong idea

Imaging is just a prettier way of asking whether cells are alive.

It does not ask that: it records what the cell looks like and lets you compare compounds by resemblance, and its main trap is that plate, batch and cell count can look like biology.

The one citation to hand over: 2

Still open here

Linked from

Sources

  1. Amit's dictation of 2026-11-17: what the atlas must contain. Kept in the atlas repository; not published.
  2. Bray et al.. Cell Painting protocol (Bray et al., Nature Protocols 2016). Nature Protocols (2016). paper; opened 2026-11-17 all records citing this
  3. Singh, Carpenter, Genovesio. Review of high-content screening statistics (Singh, Carpenter, Genovesio; PMC4230961). (2014). review; opened 2026-11-17 all records citing this
  4. Tromans-Coia et al.. Assessing the performance of the Cell Painting assay across different imaging systems. Cytometry A (2023). paper; opened 2026-11-17 all records citing this
  5. Gustafsdottir et al.. First Cell Painting paper (Gustafsdottir et al., PLoS ONE 8:e80999). PLoS ONE (2013). paper; opened 2026-11-17 all records citing this
  6. JUMP Cell Painting dataset preprint (bioRxiv 2023), abstract. bioRxiv (2023). preprint, not peer reviewed; opened 2026-11-17 all records citing this
  7. Arevalo et al.. Batch correction benchmark on JUMP (Arevalo et al., Nature Communications 2024), abstract. Nature Communications (via RePEc) (2024). paper; opened 2026-11-17 all records citing this
  8. Seal et al.. Counting cells can accurately predict small-molecule bioactivity benchmarks (Seal et al., Nature Communications 2026). Nature Communications (2026). paper; opened 2026-11-17 all records citing this
  9. Chandrasekaran et al.. Image-based profiling for drug discovery review (PMC7754181). (2021). review; opened 2026-11-17 all records citing this
  10. Recursion platform page. Recursion (2026). organisation-page, not peer reviewed; opened 2026-11-17 all records citing this
  11. Recursion news: accelerating AI drug discovery with open-source datasets (2025-05-28). Recursion (2025). news, not peer reviewed; opened 2026-11-17 all records citing this
  12. Recursion fourth quarter and full year 2025 results (press release, 2026-02-25). BioSpace (company press release) (2026). news, not peer reviewed; opened 2026-11-17 all records citing this
  13. Recursion third quarter 2026 results (press release via Finviz, 2026-11-02). Finviz (company press release) (2026). news, not peer reviewed; opened 2026-11-17 all records citing this
  14. Recursion reworks pipeline after the Exscientia merger (Fierce Biotech, 2025-05-05). Fierce Biotech (2025). news, not peer reviewed; opened 2026-11-17 all records citing this
  15. MSU Assay Development and Drug Repurposing Core recharge rates (file dated 2020-05-27). Michigan State University (2020). report, not peer reviewed; opened 2026-11-17 all records citing this
  16. Recursion news: phenomic foundation models and the brightfield comeback (2024-09-11). Recursion (2024). news, not peer reviewed; opened 2026-11-17 all records citing this
  17. Simm et al.. Repurposing high-throughput image assays (Simm et al., Cell Chemical Biology 25:611). Cell Chemical Biology (2018). paper; opened 2026-11-17 all records citing this
  18. Wawer et al.. Toward performance-diverse small-molecule libraries (Wawer et al., PNAS 111:10911). PNAS (2014). paper; opened 2026-11-17 all records citing this
  19. Mirabelli et al.. Morphological profiling of SARS-CoV-2 infection (Mirabelli et al. 2021), Europe PMC listing. PNAS (via Europe PMC listing) (2021). paper; opened 2026-11-17 all records citing this
  20. Massachusetts Life Sciences Center Bits to Bytes funding (2020-07-10). Massachusetts Life Sciences Center (2020). organisation-page, not peer reviewed; opened 2026-11-17 all records citing this
  21. Moshkov et al.. Predicting compound activity from phenotypic profiles and chemical structures (Moshkov et al. 2023), Europe PMC listing. Europe PMC listing (2023). paper; opened 2026-11-17 all records citing this
  22. Tang et al.. Morphological profiling review (Tang et al., arXiv 2312.07899). arXiv (2023). review, not peer reviewed; opened 2026-11-17 all records citing this
  23. MLOps community post on Recursion Phenom-1. MLOps community. web, not peer reviewed; opened 2026-11-17 all records citing this
  24. Thermo Fisher Image-iT Cell Painting Kit (I65500). Thermo Fisher Scientific (2026). vendor-page, not peer reviewed; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (4)

This record has been changed 4 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Throughput class
high
Emerging
yes
Status
partly written
Last updated
2026-11-17
Record
content/assays/high-content-imaging.json