Milir Labs is building analysis software for organoids and other 3D cell models. Drug-screen plates and images go in. Measurements a lab can trust, compare and reproduce come out.
Simulated dose-response
A simulation. Each of the 380 cells has its own sensitivity, so the curve emerges from the cells, the way it does in a real screen. Drag the dose, or drag the organoid to turn it.
Drug screens in organoids produce dense, beautiful data. Too often it ends up in spreadsheets and one-off scripts, where a single edge-well artifact can quietly change the answer. We are building the analysis layer that makes the answer trustworthy.
We are starting with one workflow and doing it properly: imaging and viability readouts from oncology organoid drug screens. Scroll through it.
Reader exports, plate maps and image stacks, from any instrument. Before anything is analyzed, we check layouts, units and concentration series, so a mislabeled column never becomes a result.
Reads CSV and XLSX exports, OME-TIFF stacks, and 96- and 384-well layouts.
Segmentation finds each organoid and records its count, diameter, area and estimated volume. A drug's effect is no longer confused with how large the organoids happened to grow.
Built on open segmentation approaches in the Cellpose and StarDist family, benchmarked on organoid morphologies.
Control separation, replicate agreement and spatial artifacts are tested first. A problem is flagged with its reason, and every excluded well is recorded rather than quietly dropped.
Z′ factor, SSMD, replicate CV, and row, column and edge-effect diagnostics.
Dose-response fits arrive with confidence intervals and a plain fit-quality label. Each result stores its inputs, software versions and settings, so an old analysis can be reproduced exactly.
Four-parameter logistic fits with bootstrap intervals. IC50, Hill slope, Emax, AUC, and growth-rate-corrected GR50.
1 well excluded: B02, no organoids detected. The exclusion is recorded in the analysis.
There is no shortage of AI for drug discovery. The unglamorous layer underneath it, where plates become numbers, is where results are won or lost. We think that layer deserves the most care.
A narrow scope lets us get measurement, quality control and statistics right. It is also the part of the work labs feel every single week.
We lead with transparent, published statistics and use machine learning only where it clearly earns its place, such as finding organoids in an image. You will never get a black-box response score you cannot question.
Every metric ships with an interval and a fit-quality label. When the data cannot support a number, we say so. That is the standard a reviewer or a sponsor should expect.
We do not sell assays, reagents or screening services. Our only incentive is that your analysis is correct, portable, and works with whatever instruments and protocols you already use.
Analyses are exportable and versioned. Any future use of data to train prediction models happens only with explicit, written permission from the lab that produced it.
| Capability | Spreadsheets and scripts | General plate or image software | Wet-lab CRO platforms | Milir |
|---|---|---|---|---|
| Organoid-aware measurement | Manual | Generic | Inside their service | Built in |
| Quality control before fitting | Rare | Partial | Internal, not always visible | Always, with reasons |
| Growth-corrected metrics (GR50) | Rare | Rare | Varies | Standard output |
| Reproducible, versioned analyses | Depends on one person | Partial | Within their system | Every run |
| Works with your own assays | Yes | Often vendor-tied | Their assays only | Vendor-neutral |
| Compare across runs and labs | Painful | Limited | Their data only | Designed for it |
A general view of each category, not of any specific company. Products within a category differ.
The FDA Modernization Act 2.0, signed in 2022, removed the federal requirement that drugs be tested in animals before human trials, allowing cell-based methods in its place. Regulators and pharma have kept moving toward human-relevant models since.
Patient-derived organoids are now used to test drug sensitivity in oncology research, on plates of up to 384 wells. The data is growing faster than the tooling built to read it.
Open segmentation models and well-established statistics mean a small, focused team can build a rigorous, vendor-neutral analysis layer without inventing new science.
Starting with assay analysis gives us a way in. Standardized, opt-in data gathered over time is what would later make response prediction credible.
We are early, and we would rather say so. This is the order we are working in.
Defining the data model and benchmarking open segmentation and curve-fitting methods against public datasets.
Running real plates from partner labs through the full workflow and fixing what breaks.
Reference tracking, batch-effect correction, multi-run comparison and combination analysis.
Models trained on opt-in, standardized data, released as research tools with stated limits.
We are deliberately small. If you care about getting organoid data right, we would like to hear from you.