Can a software model replace a mouse in an FDA filing?
For most of the past decade the honest answer was no. Regulators wanted human-relevant evidence, and the tools could produce predictions without proving they tracked human biology closely enough to trust.
That answer is being rewritten. On September 14, GATC Health became the first AI-native member of the Critical Path Institute's New Approach Methodologies (NAMs) Developer Coalition. The coalition's job is certification.
The Bottleneck Moves to Standards
The Critical Path Institute (C-Path) launched the NAMs Developer Coalition in May 2026, ahead of the MPS World Summit in Amsterdam. Its members build organ chips, complex in-vitro systems, and in silico (computer-modelled) platforms. The goal is one shared qualification framework, so a buyer can tell which tool answers which question.
"The launch of NAMs-DC marks a turning point for the integration of human-relevant science into regulatory drug development," said Klaus Romero, chief executive officer of C-Path. The work that follows is unglamorous: agreed validation, agreed contexts of use, agreed documentation.
The framing matters for investors. A model that predicts toxicity is a software asset. A model that has cleared a qualification pathway becomes a regulated input, and regulated inputs carry different economics because they shorten the path to an investigational new drug application.
The FDA and the EMA have committed to New Approach Methodologies as a priority; standardised evidence packages are the missing piece.
Platforms that can document validation will separate from platforms that can only demo predictions.
What the FDA Actually Accepted
The FDA made ISTAND (Innovative Science and Technology Approaches for New Drugs) a permanent qualification pathway in July 2025. In June 2026 its drug centre accepted the first Letter of Intent for an in-silico drug development tool: an AI-driven digital liver model that predicts drug-induced liver injury (DILI).
The model compares the chemical structures of new candidates against historical reference drugs with known DILI risk. Its output is designed to sit inside a weight-of-evidence review, alongside existing assays rather than in place of them.
First computational tool in ISTAND
An AI digital liver model became the first purely in-silico tool to enter FDA qualification. · FDA, 2026
The distinction from earlier milestones is worth keeping straight. The FDA accepted an organ-on-a-chip liver model under the ISTAND pilot in 2024. A chip is a physical system. The 2026 submission is computational, and that raises a different set of questions about reproducibility and data provenance.
A draft FDA guidance published in 2026 now spells out validation expectations for non-animal methods, in-vitro and in silico alike. That document is the template every NAMs developer will be measured against.
The Economics of Failing Later
Drug development fails expensively, and it usually fails late. Preclinical work alone can consume $15 million to $100 million per program, according to GATC's own estimate. A model that flags a toxic compound before the first animal is dosed attacks that bill directly.
Preclinical cost per drug program
Preclinical development alone can consume the budget of a small company. · GEN, 2026
GATC's platform is called Operon. The company says it simulates biology end to end and delivers three to five optimised compounds ready for preclinical testing in four to six months. Traditional high-throughput screening runs 36 to 48 months for the same stage.
AI-assisted discovery cycle
Operon compresses years of hit-to-lead work into a single discovery sprint. · GATC Health, 2026
| Stage | Traditional discovery | In-silico-first |
|---|---|---|
| Discovery time | 36–48 months | ✔ 4–6 months |
| Output | Screen hits | ✔ 3–5 optimised compounds |
| Animal use | Required for most filings | ◐ Reduced, not eliminated |
| Regulatory status | Established | ◐ Qualification in progress |
GATC Health platform data, 2026; FDA ISTAND program status
Validation claims are where the caution belongs. Its risk report, Derisq, was assessed at 91% specificity for off-target risks and 86% sensitivity by the University of California, Irvine. Those figures describe a prediction tool, not a cleared regulatory standard.
Whether AI can meaningfully predict how drug candidates may behave in humans is becoming increasingly established. The more important question now is which methods can withstand independent scrutiny, and demonstrate that their outputs can inform real development decisions.— Jayson Uffens, chief technology officer and chairman, GATC Health
The funding record is thin but real. PitchBook lists a $26 million later-stage round in September 2025 and about $15.4 million raised before it. As we wrote in September, Faro AI's $37.3 million Series B attacked the same clinical-development bottleneck from the trial side.
Where the Model Still Fails
Biology resists compression. A digital liver that predicts DILI from chemical structure works on a defined question with a large historical dataset behind it. Move to a new organ, a new mechanism, or a rare disease with few reference compounds, and the same model loses its footing.
The C-Path coalition's own framing concedes the gap. Pharmaceutical companies have used NAMs in discovery for years, while regulatory applications have lagged and remain unstandardised. Qualification is slow because it demands reproducibility across laboratories, documented data provenance, and a context of use narrow enough to audit.
Why did the FDA keep animal testing as the default?
Investors should read qualification timelines the way they read clinical timelines. A promising model with no qualified context of use is a preclinical bet. The ISTAND pathway and the NAMs coalition exist to convert one into the other, and both take years.
Will regulators default to in-silico evidence by 2030?
Probability: 60% — two major regulators are committed and the first qualification pathway is live, though each tool still needs years of evidence.
✅ Arguments for
The EMA has signalled the same direction, creating a competitive standard across two markets.
The first in-silico tool is already inside qualification, which sets a template for the next ones.
Confirmation criteria: a second qualified in-silico tool and a public qualification framework from C-Path.
❌ Arguments against
Sponsors keep animal arms in trials to protect their filings.
A single high-profile prediction failure would slow adoption across the field.
Disconfirmation criteria: no new ISTAND acceptances within two years, and no sponsor filing that leans on in-silico safety data.
Development scenarios
🟢 Optimistic scenario (25%)
Implications: Qualified platforms re-rate as infrastructure; preclinical cost curves bend for the first time in decades.
🟡 Base-case scenario (55%)
Implications: In-silico tools work as de-risking filters upstream while remaining optional in the formal dossier.
🔴 Pessimistic scenario (20%)
Implications: Regulators tighten evidence rules; smaller platforms with thin validation budgets stall.
A second in-silico tool accepted into ISTAND qualification.
C-Path publishing the shared NAMs qualification framework.
The EMA opening a parallel pathway to ISTAND.
A large sponsor filing an IND that cites in-silico DILI data.