Deciding without the key variable
Exposure is inferred from blood, animal models, or bulk tissue, but delivery inside the patient's tumor cells is rarely measured and never experimentally controlled.
Did your drug fail, or did it never get in?
When a program misses, plasma levels, animal studies, and biopsy markers can suggest why, but they rarely settle whether the drug hit the wrong biology or never reached its target inside the patient's tumor cells. We test it directly, so you fix the right thing before your next trial.
Our team built AI and organoid science at
- Novartis
- IBM Research
- Harvard Medical School
- Mayo Clinic
Selected for
- C10 Labs × LabCentral AI BioHub, Spring 2026 cohort
- NVIDIA Inception Program
The Physicality Gap
Two opposite problems. One identical trial result.
A molecule that works in a dish meets tissue, stroma, and membranes in a patient. Development teams track exposure carefully: plasma pharmacokinetics, tissue levels in animals, target expression and pharmacodynamic markers in biopsies. But these measure delivery indirectly, and they show what happened alongside response, not what caused it. When a drug fails, they rarely say for certain whether the biology was wrong or the drug never reached the inside of the tumor cells at a working dose.
The fixes are opposite. Guess wrong, and you shelve a drug that works or fund another trial that fails for the same reason.
A call you can't defend
Boards and partners want to know why a program missed. "The biomarker didn't hold up" isn't an answer anyone can act on.
Good drugs die for the wrong reason
A drug that never got in and a drug that failed once inside need opposite fixes. Confusing them wastes the next trial.
Why it matters
Every unexplained failure is a decision made in the dark.
Clinical trials are the most expensive experiments in medicine, and most of them fail. When a drug fails for lack of efficacy, the next move depends entirely on why. Guessing is costly in both directions.
0of drug candidates that enter clinical trials fail. Lack of efficacy accounts for 40–50% of those failures.1
$0Mmedian cost of a single pivotal oncology trial measuring overall survival.2
$0Bestimated capitalized cost per approved drug, most of it the cost of the failures along the way.3
Rescue the asset
The biology works. Re-engineer the delivery (linker, payload, formulation, dose) or enroll the patients the drug can reach, instead of writing off years of work.
Stop with confidence
No delivery fix will help. Stop before the next trial and move the budget to programs that can win.
Pick the right patients
Knowing whose tumors a drug actually reaches turns a broad trial into an enriched one: smaller, faster, and more likely to succeed.
Price the asset correctly
A failed program is mispriced until someone knows why it failed. Causal evidence changes what a shelved asset, an in-licensing deal, or a board decision is worth.
1. Sun et al., Acta Pharmaceutica Sinica B, 2022 (analysis of 2010–2017 clinical data). 2. Hsiue, Moore & Alexander, Clinical Trials, 2020 (pivotal trials for FDA-approved cancer drugs, 2015–2017). 3. DiMasi, Grabowski & Hansen, Journal of Health Economics, 2016.
Try it
Two failed programs. Raise delivery and see which one was worth saving.
Clinical data can't change how much drug gets into a cell. Our experiments can. We use sonoporation, a brief pulse of ultrasound with microbubbles that temporarily opens cell membranes, to push more drug inside tumor cells while the drug itself stays exactly the same. If response rises with delivery, the drug was delivery-limited. If it doesn't, the mechanism is the problem.
Patient-derived tumor organoid (illustrative)
Cells reached by drug0%
Tumor cells killed0%
Both programs look like failures
Without sonoporation, neither kills many cells, just as in the clinic. Turn up sonoporation to find out why.
Illustrative simulation. Sonoporation is used only as an in vitro laboratory tool.
The experiment behind it
Four arms. One variable. A causal answer.
What makes this different: most AI platforms check predictions against more observational data. We test scDEFT's predictions with a physical intervention in patient-derived tissue, and the results become training data no public dataset contains.
Every program runs through the same four-arm design in patient-derived tumor organoids. The only thing sonoporation changes is how much drug gets into the cells, not the drug and not the cells. That makes it an instrumental variable: a controlled lever that separates the effect of delivery from the effect of the drug's mechanism, something observational data can't do.
Standard delivery
With sonoporation
Vehicle only
Drug
BaselineNo drug, no sonoporation
Sonoporation controlConfirms sonoporation alone doesn't affect the cells
What the clinic seesDrug at normal delivery; much of it never gets in
Delivery removed as a limitSame drug, pushed inside the cells
In the sonoporation arms, ultrasound (left) makes microbubbles at the tumor’s surface oscillate, briefly opening cell membranes so the drug can get in.
How we read it
- Compare the two drug arms. If the drug works once sonoporation gets it inside, the program was held back by delivery. The fix is delivery or patient selection, not a new target.
- Check against the vehicle arms. They subtract any effect of sonoporation itself, so what's left is the drug.
- If nothing changes, delivery wasn't the problem. The mechanism is, and no delivery redesign will rescue it.
Sonoporation is used only as an in vitro laboratory tool, with low-energy settings that open membranes briefly without damaging cells.
Lab in the loop
Every experiment makes scDEFT sharper.
scDEFT's predictions aren't taken on faith. Each one is tested in the four-arm experiment, and every result, confirmed or not, goes back into training. Over time the model learns what no public dataset contains: how drug response changes when delivery is controlled.
The engine
scDEFT predicts who will respond, and why.
Our AI reads a patient's cells before treatment, predicts how they'll change under a drug, and traces the prediction back to named gene programs. It was trained and tested on matched before-and-after biopsies from nine clinical trials and cohorts, in five solid tumor types and in inflammatory bowel disease, not on cell lines. For patients held out from training, its predictions closely track what actually happened.
Predicting response before treatment (AUROC, IBD cohorts)
Best published signature
0.49
Pseudo-bulk classifier
0.50
Cell composition
0.57
scDEFT
0.70
0.50 is chance. scDEFT was the only method whose confidence interval excluded it.
0patient cells across two clinical atlases: six solid tumor trials and three IBD cohorts
0of achievable signal recovered predicting how tumor cells change, on patients the model never saw
0patients with matched pre- and post-treatment biopsies, across 17 treatment arms
Working with us
From an unexplained result to a defensible decision.
Bring us a program
A molecule in development, or one that missed its endpoint. We scope the question with your team in one working session.
We predict, then test
scDEFT screens the drug across patient cell states. We confirm key predictions in patient-derived organoids, with and without controlled delivery.
You get a decision
Delivery or mechanism, which patients respond, and what to change: advance, redesign delivery, re-select patients, or stop.
Why we're different
Most AI platforms find the next molecule. We explain the one you have.
Typical AI drug discovery platform
Varies the drug. Learns from correlation.
- Works in discovery: new targets and molecules
- Trained on large observational or cell-line datasets
- Can predict a failure, not diagnose it
- Predictions checked against held-out data
Causal Data AI
Varies the drug and its delivery. Learns cause.
- Works in translation: molecules already in development
- Trained on patient outcomes and controlled-delivery experiments
- Tells you why: delivery or mechanism
- Predictions put to a four-arm wet-lab test
Where teams start
Built for programs where getting in is the problem.
Antibodies and small molecules
Checkpoint inhibitors, chemotherapy, and targeted and endocrine therapies: the drug classes scDEFT has already been trained on in solid tumors.
Programs that missed an endpoint
Separate a salvageable drug from a failed hypothesis, and find the subgroup where it works.
Next: ADCs and complex modalities
Antibody-drug conjugates, bispecifics, and RNA therapeutics, where getting the payload into the right cell is often the whole problem.
Beyond oncology
Solid tumors first. Built for any disease where delivery decides the outcome.
Nothing in scDEFT or the four-arm design is specific to cancer. Any tissue that stands between a drug and its target, whether a tumor's stroma, an inflamed gut lining, or scar tissue, raises the same question and gets the same test.
Solid tumors
Antibodies and small molecules, with antibody-drug conjugates next. scDEFT tested across six trials in five tumor types, with seven more qualifying trials queued.
Inflammation and immunology
scDEFT was first tested in inflammatory bowel disease, where it predicted anti-TNF response before treatment better than published signatures.
Fibrosis
Dense scar tissue is a physical barrier to drugs, making delivery a leading suspect whenever a promising antifibrotic falls short.
The Causal MoA Atlas
A map of how drugs act once delivery is controlled.
Every four-arm experiment adds measurements that exist nowhere else: how a drug changes a patient's cells when it reaches them, and when it doesn't. Across drugs, tissues, and diseases, those results build into an atlas that makes every new prediction better than the last. It grows with each program we run.
Illustrative. The atlas is being built program by program.
The team
Drug development, AI, and organoid science at one table.
Rhodemann LiFounder, CEO. Three-time healthcare founder; exit to ConMed.
Murthy Devarakonda, PhDCo-Founder, Chief AI and Data Officer. Former Novartis Head of AI for Target ID; creator of scDEFT.
Farideh Moeinvaziri, PhDCo-Founder, Senior Scientist. Former Harvard Medical School postdoctoral fellow.
Have a program you can't explain?
If you're deciding whether to advance, redesign, or stop a program, a causal answer is the right next step. We're taking on a small number of solid tumor programs this year.