The FDA now accepts computer models as evidence in medical device review. It has not said which computer models count.

On 26 October, a University of Leeds spinout called adsilico raised £10M to sell that ambiguity back to device makers as a product. Virtual Chimera, its platform, uses generative AI to build synthetic hearts and blood vessels from anonymised clinical scans, then runs physics-based simulations of an implant inside each one.

PXN Ventures led the round. Northern Gritstone put in £5M, Parkwalk returned, and total equity now stands at £13.5M after a £3.5M seed in 2024. The team grows from 14 people to 34.

🎯
The cheque is small. The regulatory position it buys is not.

FDA device guidance covers physics-based models. It excludes standalone machine learning on its face.

The customer is a device maker carrying a $522M development bill and a 13.7% approval rate.

The round is small; the evidence question is not

adsilico has already finished projects with heart valve and neurovascular device manufacturers. Its cardiac work will be presented at TCT, the structural heart conference where device makers shop for the next generation of implants. That detail matters more than the headline number. A simulation vendor with completed manufacturer projects has crossed from research demo to procurement line item.

“Patients miss out on the chance of a life-saving device, or face uncertainty about whether it will work for them, because people like them weren’t represented when it was designed and tested,” said Sheena Macpherson, chief executive and co-founder.

The company was founded in 2024 by Macpherson, Dr Nishant Ravikumar and Dr Zeike Taylor. It sits at Nexus, the University of Leeds innovation hub, which explains the investor base: PXN Ventures, Northern Gritstone and Parkwalk all fund UK university spinouts. Northern capital backing a northern lab, with the first cheques written close to home.

What changed between 2023 and now is not the modelling. It is admissibility. As we wrote in October, Rivercell raised €22M to build the lab before the model, and AstraZeneca paid $1.5B for an option on insitro’s AI. Both were bets on data and biology. This one is a bet on the paperwork a regulator will accept.

What the FDA did, and what it left outside

Two documents define adsilico’s market, and they do not say the same thing.

The first is the November 2023 guidance on computational modeling and simulation in device submissions. It sets out a risk-informed framework for proving a model is credible for a stated context of use, built on ASME V&V 40. Its scope is narrow: physics-based, mechanistic, or other first-principles models. Standalone machine learning and AI models are explicitly out of scope.

The second is the direct final rule the agency issued on 21 September 2026. It replaces “animal tests” and “animal studies” with “nonclinical tests” and “nonclinical studies”, and recognises cell-based assays, organs-on-chips and computer models as acceptable methods. The rule applies to drugs and biological products. It does not change evidentiary standards, and it does not cover devices.

DocumentScopeStandalone AIDevices
FDA CM&S guidance, Nov 2023 Physics-based and mechanistic models ✗ Excluded ✔ Yes
FDA NAMs final rule, 21 Sep 2026 Nonclinical testing terminology ◐ Named, not defined ✗ Drugs and biologics only

Scope comparison from FDA guidance and direct final rule, 2023–2026

Read together, the position is precise. The physics step inside Virtual Chimera sits inside the 2023 framework. The generative model that creates the anatomy sits outside both documents. adsilico sells the pair as one product, with clinical oversight and documented uncertainty at each stage.

That packaging is the actual asset. A regulator does not approve a model. A regulator accepts an evidence package for a specific question, such as whether a valve size will fit a 72-year-old woman with a calcified annulus. The company’s method — provenance, assumptions, validation, uncertainty, clinical review — is written for that conversation. Its own site calls this support for a regulatory conversation, not approval.

A few hundred patients is the expensive part

Heart valves, stents and flow diverters reach the market after trials that enrol a few hundred people. That sample is too small to represent everyone who will receive the implant.

A JAMA Internal Medicine review of 195 high-risk implant trials published between January 2016 and May 2022 found women made up a median 33% of participants. The share fell to 29% in cardiovascular trials and 27% in stent trials, and rose to 44% in heart valve trials. The UK’s independent review on equity in medical devices found evidence that some groups receive sub-optimal care because of bias in device design and use.

The cost of getting this wrong late is measurable. Average development cost for a complex medical device is around $522M including failure, according to 2022 research for the US Department of Health and Human Services. An estimated 13.7% of complex devices that enter non-clinical testing go on to approval.

Those two numbers explain the sales pitch. If a manufacturer can test twenty device sizes across thousands of synthetic anatomies before recruiting anyone, it can kill a bad design earlier and recruit a narrower, better-targeted cohort. Simulation does not replace the trial. It changes what the trial has to prove.

Macpherson’s framing is about representation, and the commercial logic runs through it. A device sized on a homogeneous cohort will underperform in the patients who were never modelled. Paying for a virtual population is cheaper than paying for a post-market surprise.

Who else is selling the same evidence

adsilico is not alone, and the category is not new.

Unlearn builds digital twins of trial participants for drug development and raised a $50M Series C in February 2024. Virtonomy sells virtual patients to device companies and raised more than €5M in a Series A in December. Emmi AI raised €15M in 2025 for industrial simulation out of Linz and was acquired by Mistral AI in May 2026. Zenithon AI raised €8.85M on 30 September 2026 for physics world models.

Of that group, adsilico is the one selling into a regulated clinical pathway where the buyer has a submission deadline. That is a narrower market with a slower sales cycle and a harder validation bar. It is also stickier. A simulation accepted in one submission becomes the reference for the next.

Market sizing depends on who is counting. Grand View Research puts in-silico clinical trials at $4.1B in 2026, reaching $5.6B by 2030 at a 7.7% compound annual growth rate. Roots Analysis puts the same market at $4.35B in 2026 and $12.6B by 2040. The gap between those paths is the difference between steady substitution and a step change in how devices are cleared. Investors are being asked to underwrite the second path while the first is what the current evidence supports.

From projects to a repeatable contract

The open question is commercial, not technical. Project work has to convert into repeat engagements, and each new anatomy family and device class needs its own validation dossier. That is expensive. It is also the moat. A competitor can copy the model architecture; copying the validation record takes years and access to clinical imaging.

Device makers buy slowly. The budget sits with regulatory affairs, the sign-off runs through quality, and the first engagement is usually a narrow question about sizing or risk stratification. adsilico’s expansion into flow diverters for brain aneurysms is a test of whether the method travels. Those cases have the smallest trial populations and the least tolerance for a design error, which makes them the best proof point and the hardest sale.

What has to be true by 2028

The bear case is straightforward. Simulation stays an internal design tool, never a named input in a clearance decision. Procurement remains project-by-project, the market grows at the low end of forecasts, and adsilico stays a well-regarded consultancy with software attached.

The bull case needs three things. A device maker cites a simulation study as material evidence in a public submission. A regulator accepts it without a bespoke review cycle. The cost of building a validated virtual population falls faster than the cost of running a small trial.

None of those is guaranteed. All three are plausible enough that £10M looks cheap against the option value.

Will a simulation study carry a device submission by 2028?

🔮
At least one implantable device will be cleared with a simulation study named as material evidence in its FDA submission by the end of 2028.

Probability: 65% — the 2023 framework already defines how to prove credibility, so the remaining work is sponsorship and a test case, not new regulation.

✅ Arguments for

ASME V&V 40 gives reviewers a shared vocabulary for credibility, so the argument is procedural rather than novel.

The 13.7% approval rate and $522M average cost create a large prize for any tool that kills bad designs earlier.

UK and US regulators have both signalled concern about device bias, which makes representative anatomy a regulatory interest, not a vendor claim.

Confirmation criteria: a public FDA summary or clearance document that cites a named simulation platform for a clinical endpoint.

❌ Arguments against

The 2023 device guidance excludes standalone machine learning, and the 2026 rule covers drugs and biologics only.

A regulator can accept a physics model and still treat the generative anatomy step as an unvalidated input.

Device makers carry liability for the implant. A simulation that is wrong in a narrow population is worse than no simulation.

Falsification criteria: three years pass with no public submission citing simulation as material evidence.

Scenarios

🟢 Optimistic scenario (25%)

A large device maker uses a simulation dossier in a PMA supplement, the agency accepts it, and the in-silico market compounds toward the Roots Analysis path.

Consequences: validated platforms become infrastructure, and the earliest accepted vendor captures multi-year contracts.

🟡 Base scenario (55%)

Simulation is accepted as supporting evidence, not primary evidence. Vendors grow as specialist suppliers, and the market tracks the 7.7% path.

Consequences: adsilico's £13.5M supports a durable niche business, with slower upside than the round implies.

🔴 Pessimistic scenario (20%)

Regulators keep simulation out of clearance decisions, device makers treat it as a design aid, and generative anatomy is challenged on validation grounds.

Consequences: the category consolidates into consulting, and standalone simulation vendors lose pricing power.
Leeds spinout adsilico lands £10M Series A to fix medtech’s trial diversity gap
The round terms, total equity, trial-diversity data and the exact scope limits of the FDA device guidance and 2026 rule.
The clearest public account of what the 2023 device guidance does and does not cover.
Northern Gritstone invests £5 million into University of Leeds spinout adsilico
Primary announcement of the Series A, the platform mechanics, headcount plan and investor quotes.
Investor-side source for the round structure and the validation claims behind Virtual Chimera.
FDA updates regulations to advance innovative alternatives to animal testing
The 21 September 2026 direct final rule: terminology, scope, and the explicit statement that evidentiary standards did not change.
Primary text for the claim that the 2026 rule covers drugs and biologics, not devices.