scDEFT · single-cell Drug EFfect Transducer
Know which patients will respond to a drug, and why, before the trial.
scDEFT reads a patient's pre-treatment cells, predicts how they will change under a drug, and names the gene programs that separate responders from non-responders.
What scDEFT tells a development team
Who will respond
Response is predicted from pre-treatment cells alone, so patients can be stratified before dosing: to enrich a trial, or to find the subgroup where a drug that missed its endpoint actually works.
How each patient's cells will change
scDEFT predicts the drug-induced shift in every cell population it tracks, per patient, rather than one average effect for the whole cohort.
Why non-responders fail
Each prediction is traced back to named gene programs in specific cell types, with a control that separates a true change in cell behavior from a shift in which cells are present.
What a different drug would do
A drug can be applied to a patient's cells it was never given to. In IBD, applying a JAK inhibitor to patients who received anti-TNF predicted the real JAK-inhibitor effect better than their own drug did: an early proof of concept for asking what an untried drug would do.
Solid tumors
Tested across six solid tumor trials
We rebuilt the engine on a second, harder disease area: six published oncology trials covering five tumor types and fourteen treatment arms, from checkpoint inhibitors to chemotherapy to endocrine therapy, including seven combination regimens. Every patient had a biopsy before and after treatment and a clinical response label.
Predicting how tumor cells will change
Share of achievable signal recovered on held-out patients, from the average-patient baseline (0%) to the data's reproducibility ceiling (100%).
On new tumor types and drug classes the model had never seen, performance matched our IBD result. Grouping cells into transcriptional neighborhoods again delivered roughly a fourfold gain. Tumor results are scored per cell type; fold-to-fold variation is about 10 points.
The trials
- Breast, ER+letrozole; ribociclib + letrozole27 pts
- ColorectalPD-1 inhibitors with or without CAPOX chemotherapy22 pts
- Melanomanivolumab; nivolumab + ipilimumab; pembrolizumab20 pts
- Ovarian, high-grade serouscarboplatin + paclitaxel11 pts
- Colorectal, dMMRtoripalimab; toripalimab + celecoxib8 pts
- Breast, triple-negativepaclitaxel; paclitaxel + atezolizumab5 pts
All six are published, publicly available datasets.
Growing the tumor atlas
We screen every published single-cell study of drug treatment in solid tumors against strict criteria: human tumor biopsies, single-cell RNA sequencing, at least three patients sampled both before and after treatment, clinical responder and non-responder labels, and drug-based therapy. Few studies clear them all.
Qualifying trials queued next
- Head and neck: neoadjuvant anti-PD-1, with or without anti-CTLA-4
- Gastric: nivolumab with disulfiram
- Rectal: total neoadjuvant chemoradiotherapy
- Melanoma: BRAF and MEK inhibitors
- Triple-negative breast: pembrolizumab
- Additional melanoma and ovarian cohorts
As of September 2026. Counts change as new studies are published and screened.
First tested in inflammatory bowel disease
We first tested the engine where matched before-and-after patient data already exist: three published longitudinal cohorts in inflammatory bowel disease, treated with anti-TNF antibodies or a JAK inhibitor. The question was whether scDEFT could predict each patient's response from biopsies taken before treatment.
Predicting response before treatment
AUROC, cross-validated across patients. The dashed line is chance.
scDEFT was the only method whose 95% confidence interval (0.56–0.84) excluded chance.
Predicting how cells will change
Share of achievable signal recovered, from a shared-drug baseline (0%) to the data's reproducibility ceiling (100%).
The attributed gene programs independently recovered the main responder and non-responder cell states reported by the original study authors.
How predictions are tested
Every prediction goes to the bench.
Clinical atlases show scDEFT can predict. To show it's right about why, we test its predictions in patient-derived tumor organoids using a four-arm design. Sonoporation, a brief ultrasound pulse with microbubbles that temporarily opens cell membranes, is the only thing that differs between arms, and it changes only how much drug gets into the cells.
| Standard delivery | With sonoporation | |
|---|---|---|
| Vehicle | Baseline | Sonoporation control |
| Drug | What the clinic sees | Delivery removed as a limit |
- scDEFT predicts which patients' cells should respond, and whether delivery is the limit.
- The four arms test it. The difference between the two drug arms, net of the vehicle arms, is the causal effect of delivery.
- Results retrain the model, so each experiment makes the next prediction sharper.
Starting with solid tumors
In two clinical atlases, the engine's predictions closely track real patient outcomes. Our focus is solid tumors, starting with the antibodies and small molecules scDEFT has been trained on, with antibody-drug conjugates as the natural next step. For each, a single question decides the next program: did the drug fail because the biology was wrong, or because it never reached the inside of the tumor cell at a working dose?
Clinical data can't settle it. Exposure is inferred from plasma levels, animal studies, and biopsy markers, but delivery inside tumor cells is never experimentally controlled, so its effect can't be separated from the drug's. So we generate the data that can.
Solid tumors are where we start, not where we stop. scDEFT is disease- and drug-agnostic, and the same four-arm test applies wherever tissue stands between a drug and its target, including inflammatory disease, and rare diseases and neuroscience indications with a physical delivery barrier, such as Duchenne muscular dystrophy, cystic fibrosis, and Charcot-Marie-Tooth disease.
Evidence · clinical atlases
Learning from what happened
scDEFT learned from patients who were treated and observed. It predicts well, but like any model trained on observational data, it cannot separate a drug that failed to arrive from a drug that arrived and didn't work.
Application · solid tumors
Controlling what goes in
In patient-derived tumor organoids, we control intracellular delivery with benchtop sonoporation, holding the drug fixed. Training scDEFT on these experiments lets it attribute each failure to delivery or to mechanism.
How a program runs through scDEFT
Predict
scDEFT screens your molecule in silico across patient cell states and flags where delivery or mechanism is the likely limit.
Validate
Selected predictions are tested in patient-derived organoids, with and without controlled intracellular delivery.
Attribute
Each outcome is scored for delivery versus mechanism and fed back to retrain the model.
Decide
You get a go, redesign, or stop recommendation, with the gene programs and patient subgroups behind it.
Have a program that missed its endpoint?
We're taking on a small number of solid tumor programs for early partnership.