A drug is approved on AI-simulated evidence, without an efficacy trial

Where things stand (mid-October 2026)

I found no approval, pending application, regulatory pathway or formal proposal, at either FDA or EMA, under which a new active substance could be approved on computer-simulated efficacy evidence in place of any human efficacy trial. All the 2026 policy activity points to fewer or smaller human trials, not none:

• FDA's "plausible mechanism" draft guidance (Feb 23, 2026) still requires one adequate and well-controlled clinical investigation plus confirmatory evidence. After licensure, new gene-editing "variants" can be added to an existing license on the strength of mechanism. Those would be supplements, not new-active-substance approvals, and the evidence is mechanistic reasoning rather than simulation. (Arnold & Porter, Feb 2026: https://www.arnoldporter.com/en/perspectives/advisories/2026/02/fda-advances-a-plausible-mechanism-framework-for-rare-disease-drug-development-and-shifts-to)
• The revised Substantial Evidence draft guidance (June 22, 2026, part of "Operation TrialBlazer") makes one pivotal trial the default. It does not make zero trials acceptable. The statute itself defines substantial evidence as "adequate and well-controlled investigations, including clinical investigations." (https://www.arnoldporter.com/en/perspectives/advisories/2026/06/fda-issues-revised-draft-guidance-on-demonstrating-substantial-evidence-of-effectiveness)
• ARPA-H's SURPASS program (Sept 30, 2026) is about simulation-augmented trials "with fewer participants." It is a five-year research program and is not a substitute for trials. (BioSpace: https://www.biospace.com/drug-development/us-government-launches-ai-driven-programs-to-overhaul-clinical-trials)
• Other FDA computational initiatives target animal testing or safety, both excluded here. These are the Sept 21, 2026 rule on non-animal methods and the first in-silico drug development tool accepted in June 2026 (for predicting liver injury). (PharmTech, Sept 2026 roundup: https://www.pharmtech.com/view/regulatory-roundup-september-2026)
• FDA leadership is in flux. Makary is gone and Heidi Overton has been nominated as commissioner; the permanent heads of the drugs and biologics centres are focused on rare-disease flexibility using human data. (BioSpace, Sept 2026)
• EMA has become more cautious, not less. In March 2026 it recommended that tecovirimat no longer be used for mpox (FirstWord, Mar 27, 2026). Tecovirimat had been authorised on animal data, and later human trials failed. The CHMP's September 2026 new-medicine opinions are all conventionally trial-supported. EU regulatory sandboxes under the pharma reform only become fully applicable around 2028.

Base rate and reference class

There have been zero qualifying approvals across decades of FDA new-molecular-entity and biologics approvals and EU new-active-substance authorisations (roughly 90 a year combined). The nearest analogue, the FDA Animal Rule, is excluded. It was proposed in 1999 and finalised in 2002, and its first new-entity approval (raxibacumab) came only in December 2012, about a decade from rule to first use. No computational equivalent has even been proposed.

The EU has authorised products with no human efficacy trials: obiltoxaximab in 2020 and tecovirimat in 2022, both on animal efficacy evidence plus human pharmacokinetic bridging. That evidence fails this question's computational-substitution test, both under the "solely animal" exclusion and the "dose selection only" exclusion.

Paths to YES

1. A new regulatory pathway. A deregulatory, pro-AI US administration could create an "in-silico rule," or the EU could accept modelling under exceptional circumstances. This would need a policy shift and then a product never tested for efficacy in humans, all approved and documented before end-2031. I put the shift at about 5% and the follow-through at about 20–30% given it.
2. A borderline case. For example, an EU medical countermeasure whose assessment report frames modelling as the human efficacy evidence, or a gene-editing platform variant described in modelling terms. Most such cases fail the explicit exclusions, so I give this path well under 0.5%.

Paths to NO

These dominate. The statute requires clinical investigations. Developers can almost always run at least an exploratory efficacy study. US countermeasures go through the excluded Animal Rule. The resolution also requires official records explicitly stating that no human efficacy trial has been run in any indication and that simulations replaced such trials.

Forecast construction

I computed quarterly hazards in Python: 0.01% for Q4 2026 (no candidate under review), rising to 0.03–0.04% per quarter in 2027, 0.05–0.06% in 2028, 0.07% in 2029, 0.08% in 2030 and 0.09% in 2031. The rise reflects possible policy change and the lag before an approval could follow. Compounded, these give 0.15% by end-2027, 0.65% by end-2029 and 1.3% by end-2031.

Strongest case against my estimate

AI capability and the political push (ARPA-H, a new FDA deputy commissioner for technology and AI, the EU sandboxes) could move faster than past regulatory change, and a resolver might count a borderline modelling-heavy authorisation. I have allowed for this, but it does not justify more than about 1–2%, because no candidate, proposal or legal basis exists today, and the Animal Rule history suggests even a new pathway would take years to produce an approval.

Question. By the last day of each quarter, has the FDA, or the European Commission following EMA assessment, approved a drug containing a new active substance for which no human efficacy trial has been run in any indication or population, accepting computational or AI-simulated evidence of efficacy instead, excluding FDA Animal Rule approvals?

Definitions.

• Approval: an FDA marketing approval through an NDA or BLA, or a European Commission centralised marketing authorisation following EMA assessment, for a human medicinal product. Accelerated approval, conditional marketing authorisation, and authorisation under exceptional circumstances count. A positive EMA/CHMP opinion alone, emergency-use authorisation, clinical-trial permission, expanded access, or compassionate use does not.
• New active substance: at least one ingredient recognised in the approving regulator’s assessment as a new active substance, new molecular entity, or novel biological active ingredient at its initial approval in that jurisdiction. Small molecules and biological medicinal products, including vaccines and cell or gene therapies, are eligible. A new formulation, dose, route, indication, combination of previously approved ingredients, generic, or biosimilar alone is insufficient.
• Human efficacy trial: any prospective interventional study administering the new substance, alone or in combination, to living human participants and assessing its therapeutic, preventive, or diagnostic efficacy. This includes uncontrolled and single-patient trials, challenge studies, and studies with exploratory efficacy endpoints or surrogate endpoints used to infer benefit, regardless of phase, country, sponsor, result, or publication status. Safety, tolerability, pharmacokinetic, bioavailability, or mechanistic pharmacodynamic studies without efficacy endpoints do not count. Ex vivo studies of human tissue are not human trials.
• No trial has been run: as of the qualifying approval, no participant anywhere has received the substance in a human efficacy trial in any indication or population, including an ongoing, discontinued, unsuccessful, or unpublished trial. Registration or planning without dosing does not count as running a trial. Trials of other substances do not disqualify the product.
• Computational or AI-simulated evidence of efficacy instead: the regulatory assessment explicitly accepts model-generated predictions of human efficacy as a substitute for human efficacy-trial evidence for the new substance. Mechanistic simulations, quantitative systems pharmacology, virtual patients, and AI-generated efficacy predictions can qualify; use of AI is not mandatory. Supporting laboratory, animal, safety, pharmacokinetic, or observational evidence is permitted, but simulations must replace, rather than merely supplement or analyse, human efficacy-trial evidence.

Resolves YES when official regulatory records establish all of the following:

1. A qualifying approval has been granted.
2. The product contains a qualifying new active substance.
3. The assessment affirmatively establishes that no human efficacy trial of that substance has been run in any indication or population, rather than merely reporting that no pivotal trial or no trial in the approved population was required.
4. The regulator explicitly relies on computational or AI-simulated efficacy evidence in place of human efficacy trials.
5. The qualifying approval is not under the FDA Animal Rule, 21 CFR Part 314 Subpart I or Part 601 Subpart H.

Exclusions (resolve NO). No qualifying confirmation by the quarter-end; silence about trials; waiving only Phase 3 or confirmatory trials; extrapolation to children or another indication after human efficacy trials of the same substance elsewhere; synthetic control arms paired with actual treated trial participants; model-assisted analysis of actual efficacy trials; AI-assisted drug design alone; simulations used only for dose selection or safety; and replacement of animal testing alone. An EU approval supported solely by animal efficacy evidence also fails the computational-substitution requirement. A post-approval human efficacy-trial obligation does not disqualify an approval if that trial had not begun dosing at approval.

Resolving source. Official FDA approval letters and agency-authored reviews in Drugs@FDA (https://www.accessdata.fda.gov/scripts/cder/daf/) or CBER product approval records; EMA European public assessment reports, particularly the initial assessment report, at https://www.ema.europa.eu/en/medicines; and European Commission authorisation decisions in the Union Register of medicinal products. These regulatory sources are the exclusive resolving sources, narrower than Public confirmation (A2). Sponsor claims, trial-register searches, and press reports alone cannot establish the absence of trials or regulatory acceptance. Apply the Resolution window and Resolution date (A3): use the publication date of the last official document needed to establish all requirements, not an earlier approval or assessment date. Later trials or withdrawal do not undo a qualifying YES.