FDA approves a drug credited to AI de novo design

Where things stand (16 Oct 2026)

The FDA has not yet approved any drug that meets this question's strict test. IntuitionLabs, updated 15 Oct 2026, reports that no AI-discovered or AI-designed drug has received FDA approval. The test needs a product-specific statement, from the developer and dated on or after 16 Sep 2026, saying that an AI or ML system generated the molecule, sequence or construct. AI that only scored or selected existing candidates does not count. Neither do property prediction, target identification, or bare labels like "AI-enabled discovery".

Near-term approvals look likely, but their AI credit probably fails the test:

• Zasocitinib (Takeda/Nimbus). The FDA accepted the NDA under priority review on 14 Sep 2026, with a decision due in Q1 2027 (takeda.com release). The press called it an "AI-designed pill" (Reuters, BioSpace, June 2026). Takeda's own releases never mention AI. Nimbus's R&D head said in Sep 2026 that "first AI-approved drug" is "not the term I would use" (PharmaVoice). The design relied on physics-based free-energy calculations (FEP), with ML used to predict drug-like properties; Nimbus's CEO described algorithms speeding the choice of candidates from a large pool (Japan Times). Both of those are excluded uses. I put about 3% on a qualifying statement plus acceptance.
• Intismeran (Moderna/Merck), personalized melanoma vaccine. The Phase 3 trial succeeded (19 Aug 2026), results go to ESMO on 24 Oct, and an FDA filing should follow, so approval is likely in 2027–28. Moderna's own wording is that a "deterministic machine learning algorithm" predicts and selects up to 34 neoantigens (Moderna blog; JCO 2026). That is scoring and selection of existing candidates, so it most likely fails. I give it roughly 3% acceptance under a strict reading and 20% under a lenient one.
• Lirafugratinib (Relay/Elevar). It was approved on 23 Sep 2026, inside the window. Relay credits motion-based molecular-dynamics simulation, not product-specific AI generation, and Elevar's approval release says nothing about AI.
• Zovegalisib (Relay) and aleniglipron (Structure Therapeutics). Both could be approved around 2029–30. Their developers credit generic "platform" or structure-based discovery, so the attribution is weak.
• ASC30 (Ascletis). A global Phase 3 started in Aug 2026, with an NDA planned by end-2028. It is credited to "AI-assisted structure-based drug discovery", which reads as an unexplained label.

Paths with strong attribution:

• GB-0895 (golukibart), Generate Biomedicines. Generate says it was "computationally designed using a generative optimization technology stack" built on ML models. The S-1 shows it was affinity-optimized from a reference antibody, which the question's "known scaffold" clause allows, but this creates some acceptance risk. Two replicate Phase 3 asthma trials began dosing in Jan 2026; ClinicalTrials.gov lists primary completion in Dec 2028 and Jan 2029. That points to approval around late 2030 to 2031.
• Insilico. Its Chemistry42 generative-design claims are explicit. However, the rentosertib Phase 3 runs only in China, and ISM6331 (Fast Track for mesothelioma) is still in Phase 1/2.
• Others with explicit claims, all early-stage: Recursion/Exscientia, Absci, Iambic (whose AI is described mainly as property prediction and selection), Gubra/AbbVie, and Biolojic.

Method

The reference class is AI-native molecules entering the clinic since 2020. None has been approved yet, Phase 2 success rates match the industry norm (~40%), and the typical gap from first-in-human to approval is 7–9 years.

I gave each candidate a quarterly approval curve and an acceptance probability under a strict and a lenient resolver, weighted 70/30. I added an "unknown" bucket of about 0.33 expected further AI-claimed approvals by 2031, and combined everything assuming independence (Python model, run in the scratch container).

The per-quarter chance of first resolving stays around 1–2% through 2029, then rises to about 10% in 2031 as GB-0895, Insilico and other AI-native assets mature. The end-2031 result ranges from 0.53 to 0.65 across the sensitivities I tried.

Strongest case against: most of the 2031 probability rests on GB-0895 succeeding, on time, with a resolver accepting AI optimization as generation. Developers have also been backing away from "AI-designed" labels. Case for higher: a lenient resolver could count intismeran, ASC30 or zasocitinib.

Sources:

• takeda.com, 14 Sep 2026
• pharmavoice.com, Sep 2026
• aiinlabcoat.com, 12 Jun 2026
• merck.com, 19 Aug 2026
• ascoai.org, Sep 2026
• modernatx.com blog
• generatebiomedicines.com
• Generate S-1 (sec.gov, Feb 2026)
• ClinicalTrials.gov: NCT07276724, NCT07359846, NCT06982521
• insilico.com, 10 Sep 2026
• Elevar/GlobeNewswire, 23 Sep 2026
• Ascletis/PR Newswire, 3 Aug 2026
• Iambic S-1/A, 8 Oct 2026
• intuitionlabs.ai, updated 15 Oct 2026

Question. By the last day of each quarter, has the FDA approved a therapeutic through an NDA or BLA whose developer publicly credits an AI system with its de novo design?

Definitions.

• FDA approval: an effective U.S. Food and Drug Administration marketing approval under a New Drug Application (NDA) or Biologics License Application (BLA), issued by either the Center for Drug Evaluation and Research (CDER) or the Center for Biologics Evaluation and Research (CBER). Traditional and accelerated approvals both count. An original or supplemental approval counts if the approved product contains the qualifying AI-designed therapeutic component.
• Therapeutic: a human drug or biological product approved to treat, mitigate, or prevent disease, including therapeutic proteins, antibodies, vaccines, and cell or gene therapies. Diagnostic-only products and medical devices do not count.
• Developer: the therapeutic's originator, a documented co-developer, or a successor or licensee responsible for its development; not necessarily the NDA/BLA holder. The covered-developer list does not apply.
• AI system: a computational system explicitly identified by the therapeutic's developer as artificial intelligence, machine learning, or generative AI. A proprietary system need not have publicly released code, weights, or a product name.
• De novo design: generation of a new molecular structure, biological sequence, or engineered therapeutic construct that became the approved therapeutic or an active therapeutic component of it. The developer must explicitly attribute that generation to the AI system, using “de novo” or a concrete equivalent description of generating the new structure, sequence, or construct. Human collaboration, subsequent optimization, and use of known targets, scaffolds, or sequence fragments do not disqualify the product. AI must have generated therapeutic design content, not merely scored or selected pre-existing candidates.
• Publicly credits: a product-specific statement by the therapeutic's developer in a qualifying developer-authored source under Public confirmation (A2). A general claim that the company uses AI, or an unexplained label such as “AI-enabled discovery,” is insufficient. Independent proof of the developer's attribution is not required.

Resolves YES when both an official FDA approval record and a qualifying developer attribution are publicly available by the quarter-end and identify the same therapeutic, including through documented development-code/name changes. The two statements may be published in either order. The Resolution window applies to both required publications; a new publication may reaffirm an earlier approval or attribution. Resolve on the publication date of the later qualifying document under Resolution date (A3).

Exclusions (resolve NO).

• IND clearance, permission to begin trials, successful trial results, acceptance of an NDA/BLA for review, advisory-committee recommendations, tentative approvals, emergency use authorizations, or approval outside the United States without FDA approval.
• ANDA approvals or device clearances.
• AI used only for target identification, virtual screening of existing molecules, repurposing, property prediction, trial design, manufacturing, or formulation without de novo design of an active therapeutic component.
• A novel indication for an existing molecule does not establish de novo design of that molecule.
• Approval alone without qualifying developer credit, or developer credit without an official FDA approval record.

Resolving source. FDA approval letters, official FDA approval announcements, or dated approval records in Drugs@FDA, Purple Book, or CBER's product approval records establish approval. For AI attribution, only the therapeutic developer's own-domain statements or developer-authored documents qualifying under A2 count, including company press releases and developer-authored scientific papers. Third-party reporting alone cannot establish the required developer credit. Apply A3 and the Resolution window; later withdrawal of an approval does not reverse a qualifying YES.