Pre-seed · In active beta

Microscopy data analysis
at scale

CellWorkflow turns microscopy, from a single 3D stack to a terabyte time-lapse, into a working model of the experiment: which cells are there, what they look like, who they touch, how they change, and under which treatment. Dig into that model yourself, or let Claude, ChatGPT or your own agent run the analysis, with the numbers behind every answer.

Submit a pilot dataset → Watch an investigation
Raw microscopy t = 1/96 · Treatment A · replicate 2 · field of view 1
50 μm
ScaleTerabyte datasets, streamed to the browser DesignConditions, replicates and fields of view stay attached to the numbers EvidenceAny result opens the cells and frames it came from Your modelsSegmentation and tracking, fine-tuned on your own cells
/01 · Phenomics

Phenomics, in 3D and over time.

Phenomics is the measurement of everything a cell shows: its shape, its state, its behavior. Most of it today happens on flat images of cells in plates, thousands of wells at a time, and it works well for what it is. But a flat image leaves out the tissue around a cell, and a single snapshot cannot tell you where it was going or what happened to it next.

CellWorkflow does phenomics on the experiments you actually run: an organoid or a tissue imaged in 3D, or a live co-culture followed for hours or days. A cell's phenotype includes its shape, its neighbors and its contacts, and over time its trajectory and how it responded. From a single stack to terabytes of them, with the same rigor you would expect from a screen.

A phenotype here includes
shapemovementneighborscontactsstateresponse
/02 · The distinction

AI can read an image. CellWorkflow understands the experiment.

A microscopy volume holds all of the biology and none of the bookkeeping. Which blob is which cell, whether it is the same cell as in the last frame, whether it touched the organoid, which well it came from: none of that is written down anywhere. CellWorkflow writes it down. Segmentation, tracking, measurement and analysis turn the volume into a record of the experiment that can be queried, compared and audited, by you or by an agent.

What the pixels hold

Intensity values, arranged in space and time. Anything a biologist wants to know about them has to be worked out first, and it usually gets worked out once, by hand, in a spreadsheet nobody can find a year later.

What CellWorkflow writes down
  • Cellssegmented in 3D, in every frame
  • Identitiesthe same cell, followed through time
  • Trajectoriesspeed, displacement, persistence
  • Morphology and intensityshape, volume, markers
  • Phenotypes and statesbehavior groups, and switches between them
  • Interactionscontacts, encounters, how long they lasted
  • Neighborhoodswho is near whom, and for how long
  • Designconditions, fields of view, replicates
  • Statisticstests, effect sizes, sample sizes
  • Provenancefrom any number back to the frame
One field of view Hover or tap a cell to read its record
/03 · How it works

From the microscope file to a result you can defend.

Nothing gets exported to a spreadsheet halfway through. A figure at the end still knows which statistic it shows, which cells went into it, and which frames those cells came from.

  1. 01

    Raw microscopy

    Upload the file your microscope wrote, whether that is a 2 GB .nd2 or a 2 TB .czi. We convert it once and stream it to the browser.

    .czi · .lif · .nd2 · OME-TIFF · OME-Zarr
  2. 02

    Segmentation and tracking

    Cells and structures are found in 3D and followed through division, death and crowding. Use the current open models as they are, or fine-tune one on a few of your own annotations.

    2D · 3D · time-lapse
  3. 03

    Measurement and analysis

    Shape, intensity, speed, neighbors, contacts and how long they lasted, for every cell in every frame. Cells that behave alike are grouped into phenotypes.

    per cell · per track · per frame
  4. 04

    The record

    All of it lands in one place, together with the conditions, replicates and fields of view of the experiment, and the version of each step that produced it.

    queryable · versioned
  5. 05

    Investigation

    Ask questions in the Atlas, or let an agent call the same operations: compare conditions, split by phenotype, follow the effect over time, check confounders, pull up examples.

    Atlas · tool interface
  6. 06

    A result with its evidence

    Effect sizes with confidence intervals and sample sizes, plots, representative cells and tracks, and a link back to the frames.

    statistics · plots · provenance
app.cellworkflow.com / share / BEHAV3D Dataset
The CellWorkflow viewer showing T cells around organoids in 3D, from a BEHAV3D time-lapse

The CellWorkflow viewer on a BEHAV3D co-culture of T cells and organoids, rendered in 3D in the browser.

/04 · Model Hub

Start from the best open models. Make them yours.

Your cells, stains and microscope are not what the open models were trained on, and the record is only as good as the cells it finds. Start from a current open model, correct it on a few of your own cells, and keep the result as your lab's model: versioned, shared with your team, and better with every experiment.

Illustration of one field of view, before and after fine-tuning.
  1. 01

    Start from an open model

    Run Cellpose-SAM, microSAM or StarDist on your images as they are.

  2. 02

    Correct what it gets wrong

    Fix merged, missed and split cells in the viewer, in 2D or 3D. It shows you the cells it is least sure about first, so every correction counts.

  3. 03

    Fine-tune on our GPUs

    Training runs on GPUs we manage, with nothing to install.

    T4 L4 A100 H100
  4. 04

    Keep your lab's model

    It is versioned and shared with your team, ready for the next experiment. Every result records the version that produced it and the annotations that trained it.

Why a few corrections go a long way: in the Cellpose 2.0 study, 500 to 1,000 annotated cells brought a pretrained model close to one trained on 200,000, and a human in the loop cut that to 100 to 200. Pachitariu and Stringer, Nature Methods, 2022

The models you start from

We package and test new releases as they come out, so they reach your lab without an install.

Trained by you

Pixel and object classifiers

Paint a few strokes on your own images, in 2D or 3D, and train a classifier for structures no open model knows.

Built on APOC by Robert Haase

2D · 3DTrained on your data
Trained by you

Behavior classifiers

Label a few dozen tracks, say engaging or scanning, and every cell in every experiment gets a label, checked on held-out replicates.

Built into CellWorkflow

Tracks over timeTrained on your data
Trained by you

Your own weights

Already trained a model in your lab? Bring its weights and run it like any other, on the same GPUs.

From your lab

2D · 3DYour weights
/05 · Investigate

Investigations, not summaries.

Type the question you would ask a postdoc. The agent works out what to check first, calls CellWorkflow for each step, reads the result, and decides what to do next. It cannot make a number up, because every number it gets back comes from an operation you can rerun yourself.

Illustrative session tcell-organoid-07 · 2 conditions · 8 replicates · 16 fields of view Agent interface · early access
CellWorkflow

measures, runs the tests, keeps the record.

The model

plans, reasons, writes it up.

Works with

Claude, ChatGPT and Codex, your company's agents, or a model running on your own hardware.

The operations are exposed through the Model Context Protocol, so any assistant that speaks it can use them, and the platform gets better as the models do. The session above runs on a simulated experiment we built to show the flow; the numbers are not from a real study.

/06 · Evidence

Every conclusion leads back to evidence.

When the agent tells you that engaging T cells spend more time on the organoid after treatment, you should not have to take its word for it. Click the number and you get the test, the eight replicate values behind it, the 547 tracks, and the frames those tracks were measured in. Rerun the operation and you get the same result, because the versions are recorded with it.

A conclusion, opened

Phenotype C1, Engaging, shows increased contact with the organoid after Treatment A: , , over in , with behind it.

Same inputs, same result.Operations are versioned. Run the same one on the same data next year and you get the same number.
Share the evidence, not just the figure.Send a result to a colleague or a reviewer with its statistics, tracks and the steps behind it, and they can check it without redoing the analysis.
Back to the frame.Any track opens in the viewer at the frame you are arguing about, with its segmentation drawn over the raw image.
/07 · What you can measure

The phenome, measurement by measurement.

All of this is measured per cell and per frame. Together it is the phenome of the experiment: stored with its design, and available to you in the interface and to agents through the same operations.

01

Segmentation

Cells, nuclei and structures in 2D and 3D. Current open models out of the box, click-to-segment for the awkward cases, fine-tuning on your own annotations.

02

Tracking

Identity kept through division, death and crowded co-cultures, with a lineage tree and a quality score for every track.

03

Morphology and intensity

Volume, surface, shape and texture per object, and marker intensity per channel and per frame.

04

Spatial analysis

Neighbors, distance to structures, niches, and whether two cell types are found together more often than chance would put them.

05

Temporal analysis

Speed, displacement, persistence, state changes, and how the population shifts over the time-lapse.

06

Phenotype discovery

Behavior groups with the measurements that define them and example cells for each, so you can rename or merge them when they do not match the biology.

07

Interactions

Contacts and encounters between cell types or with structures such as organoids, how long they lasted, and what happened afterward.

08

Statistics

Comparisons across conditions with one value per replicate, effect sizes, confidence intervals and confounder checks.

09

Charts and provenance

Plots and tables that know which cells they came from, ready for a paper, a lab meeting or an agent's report.

/08 · Applications

Questions labs bring to CellWorkflow.

The experiments differ. The question mostly does not: what did the cells do, did the treatment change it, and would it hold up in review.

01 · Immuno-oncology

Which T cell phenotypes engage the organoid, and does the treatment change that?

Contact time, encounters and killing per phenotype, compared across cell products and doses. This is the experiment CellWorkflow grew out of, with the BEHAV3D authors.

Co-culture · Cell therapy
02 · Organoid and developmental biology

How do cells move, divide and change state as the tissue forms?

Lineage, migration and shape over days of imaging, with identity kept through divisions and without correcting tracks by hand.

Lineage · Morphogenesis
03 · Live-cell pharmacology

Does the compound change behavior before it changes viability?

Movement, shape and interaction phenotypes over time, per condition and dose, compared at the replicate level.

Dynamics · Dose response
Field note

"The bottleneck was never the imaging. It was running the analysis: weeks of manual work per experiment, stitched across desktop software, notebooks and scripts. A platform that keeps the whole chain connected turns that into part of the experimental loop, not a quarterly project."

SA
Scientific advisors Co-developers of BEHAV3D · Princess Máxima Center / imAIgene Lab
/09 · Scientific team

Built by the people who built BEHAV3D.

CellWorkflow exists because the original BEHAV3D authors wanted the pipeline they pioneered to be usable by every imaging lab, not just the ones that can dedicate a senior analyst to it for two to four weeks per dataset.

Developed at
imAIgene Lab imAIgene Lab The lab where CellWorkflow was developed, and where the BEHAV3D behavioral phenotyping framework continues to be extended.
Part of
Institute for Biomedical Research Sols-Morreale IIBM Sols-Morreale The Institute for Biomedical Research Sols-Morreale, which hosts the lab. It is run jointly by the Spanish National Research Council and the Autonomous University of Madrid.
Spanish National Research Council (CSIC) CSIC The Spanish National Research Council, the public research organization the institute and the lab belong to.
Scientific advisor

Dr. Anne C. Rios

Group leader at the Princess Máxima Center for Pediatric Oncology, senior author of the BEHAV3D platform papers, and pioneer of 3D large-scale single-cell imaging of patient-derived organoids and immune co-cultures. Closely involved in shaping the platform's behavioral readouts.

Nat Biotechnol2023 Nat Protoc2024
Scientific advisor

Dr. Maria Alieva

Co-first author of Uncovering the mode of action of engineered T cells in patient cancer organoids (Nature Biotechnology, 2023), the study that introduced BEHAV3D, and first author of the BEHAV3D protocol (Nature Protocols, 2024). Established the behavioral phenotyping framework CellWorkflow productionizes.

Nat Biotechnol2023 Nat Protoc2024
Where we are

CellWorkflow is a pre-seed venture. The platform was developed at the imAIgene Lab and is now being built into a product together with the labs that use it, so the groups who come in early still shape what ships next.

CellWorkflow is built in collaboration with, but is independent from, the Princess Máxima Center. Scientific advisors shape the analytical design of the platform and do not endorse its commercial claims.

/10 · Direction

One experiment today. A lab's memory of its biology next.

Once a lab has a few experiments phenotyped this way, the interesting questions stop being about one dataset. We want a lab to be able to ask across everything it has imaged: whether a phenotype has shown up before, which earlier experiments responded the same way, and whether an effect is consistent between them.

"Have we observed this phenotype before?" "Which earlier experiments produced a similar response?" "Is this effect consistent across them?"

Send us one experiment.

Upload a dataset from something you imaged recently, any format, any size. We run it end to end and go through the first analysis with you on a 30-minute call, with a scientist rather than a sales rep.

Submit a pilot dataset → Talk to a scientist