An AI-discovered drug meets a Phase 3 endpoint in a hard disease

Where things stand (as of 17 Oct 2026)

Nothing has resolved this yet. No drug credited to AI has ever met a Phase 3 primary endpoint in Alzheimer's disease (AD), ALS, pancreatic cancer (PDAC) or glioblastoma (GBM). AI-native pipelines are still early. An ASCO 2026 analysis, summarised by IntuitionLabs (updated Oct 2026), counted 117 AI-enabled assets in human trials, of which only 8 had completed Phase 2. The furthest-along ones are in other diseases: Insilico's rentosertib (IPF, Phase 3 since Sep 2026) and Generate's GB-0895 (asthma). In ALS, Verge's AI-derived VRG50635 failed and the company pivoted (Fierce, Dec 2025). In PDAC, the RAS-inhibitor wave is producing Phase 3 wins, for example daraxonrasib in RASolute 302, approved Aug 2026. But that success predates the opening date, and its discovery paper describes "structure-guided" design with no AI credit. So the question turns on a few drugs whose AI credit is peripheral, plus things that are not yet visible.

Paths to YES (rough cumulative contribution by end-2031)

• BXCL501 / dexmedetomidine for agitation in Alzheimer's dementia (~3.5%).
  • The AI credit is solid. BioXcel's 10-K says its EvolverAI/ML platform identified its pipeline compounds and their new indications.
  • BioXcel filed for Chapter 11 on 27 Aug 2026. Teva is the stalking-horse bidder, and its release mentions only the at-home schizophrenia/bipolar use.
  • The second Phase 3 needed for the AD indication (TRANQUILITY In-Care) was never started.
  • YES needs four things: a new owner funds that trial (~20–25%), it is reported by 2031, it succeeds (~70%), and a resolver accepts "agitation associated with AD dementia" as treating AD (~40–45%). The hazard falls mostly in 2030–31.
• AR1001 / mirodenafil, POLARIS-AD (~0.7%).
  • This is the only near-term event: a 1,535-patient trial with topline at CTAD on 16 Nov 2026. The primary endpoint is CDR-SB, analysed in the overall and monotherapy populations.
  • Its Phase 2 missed all primary and secondary efficacy endpoints, so I put success at ~7%.
  • AriBio's AI credit is weak. It says only that its "AI-powered" ARIDD platform is "being used to develop AR1001". In 2021, ARIDD meant "Advanced, Rapid, and Integrated Drug Development", and the drug was licensed from SK Chemicals in 2011.
  • The primary observation period also ended before the opening date (primary completion 29 Jun 2026, last safety visit 27 Jul 2026). I assumed about a 50% chance a resolver treats database lock as the "data cutoff".
• Setidegrasib, Astellas (~1.4%).
  • A Phase 3 in first-line KRAS G12D PDAC is running (NCT07409272, 614 patients, OS primary endpoint, primary completion Aug 2029). Success chance is about 55%.
  • Astellas's only AI language is a vague line on its targeted-protein-degradation webpage: the drug's rapid progress drew on "our artificial intelligence and robotics capabilities". The discovery paper (Commun Chem 2025) describes structure-based design with no AI. This likely fails the "describe the discovery contribution" test unless Astellas later makes a more explicit claim.
• BPM31510, BPGbio (~0.8%). It has AI provenance (Bayesian "NAi" platform) and promising single-arm Phase 2 data in GBM (median OS 19.3 months, topline due fall 2026). But a randomized Phase 3 would still need funding, and GBM Phase 3 success rates are very low.
• GFH375 (GenFleet, ~0.5%) and atebimetinib (Immuneering, ~0.25%). Both are in PDAC Phase 3s with good odds of success, but neither has qualifying AI credit.
  • GenFleet's 2026 interim report credits AI drug design only to newer projects such as GFH946.
  • Immuneering's 10-K explicitly contrasts its platform with "AI approaches".
• Other or unforeseen drugs (~2.2%). These include new explicit AI claims for other successful PDAC RAS drugs, or AI-native drugs that reach Phase 3 by 2029.

Hazard shape

I summed the per-path yearly probabilities into quarterly hazards, then computed the cumulative series in Python (1 − ∏(1−h)).

Quarter or year	Hazard	Main driver
Q4 2026	~0.44%	AR1001 topline
2027–28	~0.1–0.25% per quarter	GFH375, atebimetinib
2029–31	rising to ~0.65–0.7% per quarter	setidegrasib, a possible BXCL501 trial, BPM31510

The cumulative probability reaches about 8.2% by end-2031.

Strongest case against my estimate

Higher: PDAC now has many likely Phase 3 wins, and AI branding is spreading, so some developer may retroactively and explicitly credit AI design, or a lenient resolver may accept vague credits like Astellas's or AriBio's. Teva might also see the AD-agitation market as worth one more trial.

Lower: the criteria are strict ("describe the discovery contribution"). The BXCL501 path needs a chain of uncertain steps plus a favourable reading of "for Alzheimer's disease". Taken together, these points roughly balance, and I kept about 8%.

Sources

• ClinicalTrials.gov: NCT05531526, NCT07409272, NCT07562152
• AriBio release, 3 Aug 2026 (EIN Presswire); Hankyung, 1 Oct 2026; ChosunBiz, 5 Jun 2026; Korea Biomedical Review, Sep 2021
• Greeley et al., AR1001 Phase 2 (PMC12501329)
• BioXcel–Teva releases, 28 Aug 2026 (GlobeNewswire); BioXcel Q1 2026 release; BioXcel FY2020 10-K (SEC); Bondoro case summary, Sep 2026
• Immuneering FY2025 10-K (SEC)
• Astellas targeted-protein-degradation page and Phase 3 release, 15 Apr 2026; Yoshinari et al., Commun Chem 2025
• GenFleet interim report (HKEX, 16 Sep 2026)
• BPGbio release, 28 May 2026; Targeted Oncology, 17 Jul 2026; Drug Discovery Trends, Jan 2024
• Fierce Biotech on Verge, Dec 2025
• IntuitionLabs, "AI-Discovered Drugs in Clinical Trials 2026"

Question. By the last day of each quarter, has a Phase 3 trial reported that a drug credited to AI-driven discovery met its primary endpoint for Alzheimer's disease, ALS, pancreatic cancer, or glioblastoma?

Definitions.

• Drug: a therapeutic small molecule or biological product, including a repurposed drug and a drug tested in combination with other treatments. Devices, diagnostic tools, and non-drug interventions do not count. There is no restriction to covered developers.
• Credited to AI-driven discovery: the drug's developer or discovery partner publicly identifies the drug and explicitly credits an artificial-intelligence or machine-learning system with discovering or designing the candidate, selecting it through screening, or identifying a therapeutic target, mechanism, or new indication that led to its development for the qualifying disease. The statement must describe the discovery contribution, not merely call the company or drug “AI-powered.” De novo molecular design is not required. AI used only for trial recruitment, patient stratification, endpoint measurement, analysis of clinical results, manufacturing, or administrative work does not qualify.
• Phase 3 trial: a human interventional trial prospectively designated Phase 3/III in a public trial registry or protocol. A combined Phase 2/3 or 2b/3 trial counts if the reported result is the prespecified confirmatory Phase 3 analysis. Phase 2 studies described only as “pivotal” do not count.
• Qualifying disease: Alzheimer's disease (including biomarker-defined early Alzheimer's disease), amyotrophic lateral sclerosis (ALS), pancreatic cancer, or glioblastoma, including molecularly defined subtypes. Treatment of any one suffices. A basket trial counts only if the disease cohort has its own prespecified primary efficacy analysis.
• Met its primary endpoint: the report explicitly states that the prespecified primary efficacy analysis succeeded under the trial's statistical success rule, in the intended beneficial direction. Superiority and noninferiority both count. Clinical or surrogate efficacy endpoints count; safety, tolerability, pharmacokinetic, and feasibility endpoints alone do not. Where success requires all co-primary endpoints, all must be met; where the prespecified rule permits success on any one primary endpoint, one suffices with the required multiplicity adjustment. A prespecified interim analysis crossing the efficacy boundary counts.

Resolves YES if a qualifying results publication reports primary-endpoint success for at least one qualifying drug trial and qualifying public credit to AI-driven discovery is available by the horizon. FDA approval, a successful second trial, and proof that AI was indispensable are not required. Combination trials count when the AI-credited drug is part of the experimental regimen evaluated in the successful primary comparison, not merely background treatment common to both groups.

The results publication must fall within the Resolution window, and the primary analysis's observation period or data cutoff must end after the opening date, under the Observation period convention. Earlier AI-discovery documentation may establish the drug's provenance. If the discovery credit is published after the results, the resolution date is the later qualifying publication date under A3.

Exclusions (resolve NO). No qualifying report by the horizon; trial launches, enrollment announcements, forecasts of success, favorable trends without primary-endpoint success, secondary or exploratory endpoint successes, post-hoc subgroup successes, and retrospective changes to the primary endpoint after unblinding. A Phase 1 or Phase 2 result does not qualify even if the drug's AI provenance is established. A report that merely republishes an analysis with a pre-opening data cutoff does not qualify.

Resolving source. Trial results must appear in a peer-reviewed full research article reporting the trial or a trial sponsor's official top-line results release on its website or as its attributable release through a newswire. Conference abstracts, preprints alone, media reports, and investor commentary alone do not qualify. Trial phase, endpoints, and success rules may be established from ClinicalTrials.gov, an entry in a WHO-recognized primary trial registry, or the publicly available trial protocol/statistical analysis plan. AI-discovery credit must be in an official drug-developer or discovery-partner statement or an authored research publication; executive interviews alone do not count. These are narrower resolving-source requirements than Public confirmation (A2). Apply Resolution date (A3) and the Resolution window; once YES, always YES.