For the roughly 80% of clinical trials that run late, the delay is rarely about science. It is about people: the patients who could not be enrolled, the data points that came in noisy, the control arms that took an extra year to fill. Each delayed trial is reported to cost sponsors around $15 million [Prism News, retrieved 2026]. Mantis Biotech, a San Francisco startup founded in 2025, is betting that some of those missing patients can be modeled instead of recruited.
Mantis is building what it calls digital twins of humans, computational models of anatomy, physiology, and behavior assembled from disparate real-world datasets and stitched together with synthetic data [TechCrunch, Mar 2026]. The company's stated ambition reaches beyond drug development into sports performance and defense applications [Medium, Mar 2026], but the near-term wedge is biomedical: giving researchers and device makers a way to stress-test hypotheses, simulate cohorts, and pressure-check study designs before a single human subject is enrolled. In April, Mantis disclosed $7.4 million in seed funding led by Y Combinator, with participation from Decibel VC, Liquid 2, StoryHouse Ventures, Pioneer Fund, and Spot VC [The AI Insider, Apr 2026].
The standard of care today
It is worth being precise about what Mantis is trying to displace, because the patient population here is broad: anyone enrolled in, or excluded from, a clinical trial. Today, the standard of care for trial design is a combination of historical control data, registry pulls, electronic health record mining, and prospective recruitment through academic medical centers and contract research organizations. Synthetic control arms, where a model-generated cohort substitutes for some or all of a placebo group, are an emerging but not yet mainstream tool. The FDA has accepted external control arms in narrow circumstances, particularly in rare disease and oncology, and the EMA has issued reflection papers on the use of real-world evidence, but neither agency has issued a general clearance for fully synthetic patients in pivotal trials. The bar for regulatory acceptance remains case-by-case and is anchored in transparency about the underlying data and model assumptions.
That is the world Mantis is walking into, and it is the world its closest named competitor, Unlearn.AI, has been working in for several years, primarily in neurology indications such as Alzheimer's and multiple sclerosis. Mantis is earlier, smaller, and broader in stated scope.
The bet
The company's pitch, as described in press coverage, is that the bottleneck in modern medicine is not algorithms but data availability, and that high-fidelity synthetic data drawn from many modalities can fill gaps that no single hospital system or registry can [TechCrunch, Mar 2026]. Crunchbase characterizes the product as a unified biomedical device testing and regulatory platform [Crunchbase, retrieved 2026], which would suggest an early commercial focus on medical device makers, a customer set that often faces shorter and more design-driven regulatory pathways than drug sponsors do. That framing is reported rather than confirmed by the company, and it is the part of the thesis worth watching as Mantis moves from seed deck to first paying contracts.
Why it could be big
The tailwinds are real. Sponsors are under pressure to compress trial timelines, regulators are increasingly open to model-informed drug development, and the device industry has been quietly absorbing simulation tools for years, particularly in cardiovascular and orthopedic implants. If Mantis can produce digital twins that hold up under regulatory scrutiny for even a narrow class of devices, the addressable market expands quickly, because every avoided month of trial delay maps directly to the $15M-per-trial figure cited in industry reporting [Prism News, retrieved 2026].
The investor syndicate suggests the thesis has resonance with technically literate funds. Decibel VC has a track record in deep infrastructure, and Y Combinator's willingness to lead and follow at the seed stage gives Mantis network access into both the developer and biotech operator pools.
| Metric | Value |
|---|---|
| Seed funding raised | 7.4 $M |
| Reported cost per delayed trial | 15 $M |
| Clinical trials delayed by data issues | 80 % |
| Current headcount | 3 people |
The team and traction
Mantis is led by founder and CEO Georgia Witchel, who is 24 [Medium, Mar 2026] and came to entrepreneurship from a background in professional ice climbing and high-risk expedition work [Startup Strides, 2026]. The company currently has three employees [Y Combinator, retrieved 2026]. That is a small team for a problem this large, and Witchel has spoken publicly about the ethical questions raised by predictive modeling of human biology [Medium, Mar 2026], a posture that matters in a category where overclaiming has historically drawn regulator attention.
The honest counterfactual
The most credible concern is regulatory acceptance. Synthetic data in clinical contexts faces a high evidentiary bar, and Unlearn.AI, the named competitor, has spent years building published validation studies and FDA engagement around a focused set of indications. Bears will note that breadth (medicine, sports, defense) is hard to defend at seed stage with three employees, and that the device-and-regulatory-platform framing reported by Crunchbase has not yet been substantiated with named customers in public reporting. The bull answer is that Mantis is explicitly positioning around data generation rather than a single therapeutic claim, which gives it room to sell into device makers, contract research organizations, and simulation-curious pharma teams while validation work matures. The TechCrunch coverage also suggests the company is leaning into multi-modal data assembly as the technical moat [TechCrunch, Mar 2026], which is a different bet than competing on a single disease model.
What to watch
Over the next twelve months, the meaningful milestones for Mantis are concrete: a first publicly named customer, ideally a device maker; a peer-reviewed validation paper or a documented regulatory interaction (an FDA pre-submission meeting would qualify); and the first hires beyond the founding three, particularly in regulatory affairs and clinical biostatistics. A Series A in late 2026 or 2027 would not be surprising given the seed size and investor mix, but the more important signal will be whether the digital twins Mantis builds can survive contact with a regulator's statistical reviewer.
The disease state Mantis is ultimately trying to serve is every disease state, which is both the appeal and the difficulty of the bet. For now, the patient population that benefits most directly is the one that never makes it into a trial at all.
Pulse Raman covers biotech, digital health, and clinical AI for Startuply.