A note before we start: our mail platform was suspended for a sender review between Sept 5 and Sept 10, so Edition #4 never reached your inbox. It is on the site, and the three things in it that still matter are carried forward below. Nothing was lost; it just arrived late and in a different order.

The Read

The scoreboard moved this week, and the registry didn't. Insilico dosed the first patient in what it calls the world's first Phase 3 trial of a generative-AI-designed drug — and as of this writing ClinicalTrials.gov still lists that trial as "not yet recruiting," last updated July 7. Both statements are probably true and the gap between them is the whole discipline of this newsletter. Meanwhile the accountability side of AI in medicine got its first documented casualties rather than its first hypotheticals: roughly a thousand pages of FOIA'd records show Medicare's AI prior-authorization pilot launched over its own vendor's written warning that a working product was unrealistic, with one request going eighty-three days against a seventy-two-hour target, two vendors denying more than twenty thousand requests in a single quarter, and the payment structure rewarding denials with penalties capped at five to ten percent. And ARPA-H put $62.7M behind FDA-authorized AI that would actually direct heart-failure treatment — including $15M for a supervisory model whose entire job is watching the other models for unsafe recommendations. That last line is the shape of the year: we have stopped asking whether the model is good and started building the apparatus that assumes it sometimes isn't. Old problems, new physics — and this week, for the first time, new liability. Scoreboard: still zero full FDA approvals of an AI-discovered drug.

Tide status

No tide movement this week. Both standing tides hold. One got its highest rung yet, with an asterisk.

Biology is becoming an engineering discipline — holds, top rung now occupied (company-asserted) [Phase 3]

On Sept 10 Insilico announced it had dosed the first patient in GENESIS-IPF-3: rentosertib, a TNIK inhibitor whose target was AI-identified and whose molecule was AI-designed, in idiopathic pulmonary fibrosis. 320 participants, 47 centers in China, 52 weeks, primary endpoint the annual rate of FVC decline. If it holds up, that is the first time a generative-AI-originated medicine has had a human dosed in a registrational trial — the rung this tide has been climbing toward since we opened it.

The asterisk is that we could not confirm it independently. The ClinicalTrials.gov record for NCT07687459 still reads NOT_YET_RECRUITING, with an estimated start of Aug 30 2026 and a last update posted July 7 — and lists 48 sites where the press release says 47. We flagged that stale record last week precisely so we would notice if it stayed stale. It has. This is almost certainly registry-maintenance lag at a Chinese-sited trial rather than anything sinister, and we are not calling it otherwise. But "company says dosed, registry says not yet recruiting" is exactly the situation where everyone else prints the press release. We are printing both.

What does carry peer-reviewed weight landed three days earlier in Nature Biotechnology: proteomic aging clocks applied to the rentosertib Phase 2a, n=42 IPF patients, 2,841 proteins, validated against 55,319 UK Biobank profiles. All six independent aging clocks showed reduced predicted biological age versus placebo, peaking around three to four years of reversal at week 4 in the 30 mg BID arm. The same paper restates the efficacy signal: FVC +98.4 mL in the 60 mg QD arm against −20.3 mL on placebo. That is real evidence, properly reviewed, and it is what makes the Phase 3 worth watching rather than merely worth announcing.

• Insilico — first patient dosed, GENESIS-IPF-3
• ClinicalTrials.gov — NCT07687459
• Nature Biotechnology — proteomic aging clocks in a phase 2a trial

The binding constraint is disease understanding, not molecular design — holds, and the constraint moved downstream [real-world]

Two findings this week say the drug→patient end of the pipeline is where AI is losing, not the mechanism→molecule end. An emergency-department study of LLM clinical decision support recorded no adverse events and still concluded the results "do not justify clinical deployment," because adoption decayed from 67.9% to 29.7% over four weeks — the authors' read being that "sustained clinician engagement, rather than diagnostic capability, is likely the principal barrier." Separately, STAT reported that ambient scribes genuinely save ER physicians documentation time and that those minutes do not become shorter waits or faster results for patients. Carried forward from Edition #4: Eric Topol's demonstration that two cardiovascular outcome trials enrolling more than 14,000 patients between them failed on a blood marker correlating with the actual pathology at r = 0.2, while an FDA-authorized AI imaging score that measures it directly sat available.

Read together: capability is not the binding constraint anywhere on this tide. Selection isn't solved, adoption isn't solved, and throughput isn't solved. Those are the expensive parts.

Waves

1. The AI-drug-discovery show-me window — MOVED. The first dosing, the first bankruptcy, and a price finally printed. [Phase 3]

Three things happened to this category in seven days and they point in different directions, which is what a real inflection looks like.

The dosing is covered in the tide above. The bankruptcy is new and matters: BioXcel Therapeutics, an AI-driven neuroscience biotech with a marketed product (Igalmi, for agitation), filed Chapter 11, with Teva as stalking-horse bidder at $57.5M upfront plus $67.5M in milestones. That is the first time a company that built its identity on AI drug discovery, got a drug approved, and commercialized it has still gone to the wall. Worth sitting with: approval was never the hard part of the business model.

The price came from an unexpected direction. Merck KGaA is paying PostEra "mid-double-digit millions" for two fertility programs — oral small-molecule agonists of the FSH receptor and the LH/choriogonadotropin receptor, designed on PostEra's platform, aimed at replacing the injectables used in IVF. A number. An actual, printed number, in a category where the standard disclosure is silence — and set against Owkin signing Servier as its fourth top-20 pharma for K Pro and MOSAIC, again with terms undisclosed (see Ripples). Two data points don't make a market, but "mid-double-digit millions for two named programs" is now the only public anchor anyone has.

And the negative finding is as loud as the positives: no pure-play AI drug-discovery venture round closed in this window at all. Both dedicated AI-DD funding trackers show nothing. Every health-AI dollar this week went to clinical operations, revenue cycle, and monitoring — not to discovery.

Roadmap implication: the category is bifurcating in front of us. Assets with clinical data are advancing and can now point to a dosed Phase 3; platforms without them are being repriced, and one has been liquidated. Underwrite the asset, not the platform — and note that the money agrees, because it stopped funding platforms this week entirely.

• Where Tech Meets Bio #84

2. The regulatory regime for adaptive/learning systems — MOVED. Institutions, not guidance. [policy]

The FDA did something more consequential than another authorization: on Sept 11 it published a final order (91 FR 57785, docket FDA-2026-N-9907) creating an entire new device class — Cardiovascular Machine Learning-Based Notification Software, Class II with special controls, built on the Viz.ai De Novo and applied retroactively to Aug 3 2023. Read the special controls, because they are the policy: clinical performance testing on independent real-world data from at least three geographically diverse sites with demographic and hardware-type representation; a human-factors assessment of how the output could be misinterpreted; and labeling that warns against over-reliance. The regulator has written generalization and automation bias into a device class definition.

Alongside it, FDA granted De Novo authorization to "Queen of Hearts" (Powerful Medical), an EKG model that flags STEMI and also occlusion MI that lacks the classic ST-elevation pattern — a higher evidence bar than a 510(k), for a genuinely harder clinical claim.

The institutions arrived the same week. CMS named Stephanie Carlton its first Chief Clinical AI Officer, with a four-pillar strategy — public trust, data sharing, market-access pathways, reimbursement — and the ACCESS model as the vehicle, with an evidence checkpoint set at December 2028. FDA named Jared Seehafer its first Deputy Commissioner for Technology and AI. Worth knowing, per NYT reporting relayed by STAT: Seehafer previously proposed centralizing AI oversight in the commissioner's office or at HHS, CDRH director Michelle Tarver pushed back, HHS officials wanted then-Commissioner Makary to fire her, and he refused. Where AI authority sits inside FDA is an open fight, and it now has a named participant on each side.

Britain went further and more coherently: a commission delivered a 119-page report with 44 recommendations proposing staged authorization — deployment "within a tightly controlled scope, based on MHRA review of initial evidence," expanding as prespecified performance and safety thresholds are met — plus mandatory provider reporting on how tools perform in live clinical settings and post-deployment monitoring as models adapt. MHRA is expected to accept and phase it in.

Pulling the other way: HTI-5 is due to be finalized "in the next few weeks" and as proposed would eliminate 34 of 60 EHR certification criteria, including recently added transparency requirements for AI tools bundled with EHRs. Carried forward from Edition #4, still live: FDA's TEMPO pilot (four participants, unchanged this week — SonderMind, Limbic, Cadence Solutions, Dexcom) lets generative-AI devices reach Medicare patients under enforcement discretion before authorization; and Vara holds a world-first CE mark to report normal mammograms with no radiologist read, conditioned on a drift monitor that reverts the site when performance moves.

Roadmap implication: three governments converged this week on the same instrument — authorize narrowly, then require evidence from live deployment — and the American version is simultaneously being hollowed out at the EHR-certification layer. For anyone building here, the special controls in that Federal Register order are now the de facto spec sheet: multi-site independent real-world validation, demographic and hardware coverage, a documented automation-bias analysis. Build to that and the rest is paperwork. Standing item: comment on docket FDA-2026-N-7874 before Oct 19 — five weeks.

• Federal Register — Classification of Cardiovascular ML-Based Notification Software
• Fierce Healthcare — federal officials outline AI ambitions
• STAT — UK unveils recommendations on AI regulation in medicine

3. Clinical-AI autonomy vs. mandated human oversight — MOVED HARD. The first documented harm, and the first AI hired to supervise AI. [real-world]

ARPA-H's ADVOCATE program committed $62.7M over four years ($33.7M guaranteed in year one) to build partially autonomous, FDA-authorized AI that directs heart-failure treatment — assessing symptom severity, prescribing drugs, ordering labs. The awards: Kaiser Permanente up to $16.3M (implementation, ~2,500 patients across 21 medical centers, pilot then randomized trial), Duke up to $15.5M (multi-site validation across five health systems), Stanford up to $15M, Tempus AI up to $9.5M (patient app), UpDoc up to $9.2M, Atman Health up to $7.7M. FDA authorization packages expected within two years; a 39-month prototype-to-clinic path. Target population 6.7M Americans; ARPA-H projects $28B in annual savings. Program manager Haider Warraich: "If we can do this with heart failure, then we will have essentially created a template that can work across conditions."

The single most interesting line item is Stanford's $15M, which funds a supervisory AI that watches the clinical agents for unsafe recommendations. The federal government is paying for an oversight model because it does not believe human oversight will scale to agentic clinical care. That is a concession with enormous consequences, and it is being made in a grant announcement rather than a guidance document.

Then the other side of the ledger. The EFF published roughly 1,000 pages obtained through FOIA litigation on CMS's WISeR model — AI-assisted prior authorization in Original Medicare for skin substitutes and epidural pain injections, live across six states since January. The records show: a 72-hour turnaround target and one request that went 83 days; two vendors alone denying more than 20,000 requests in the first quarter of 2026; one vendor, Virtix, denying more than it approved; vendors paid on denied requests, with quality-score penalties capping their downside at 5–10% of payments; and Innovaccer internally flagging that a working product by the January launch date was unrealistic, with functionality gaps persisting into April. Provider complaints document cancelled kyphoplasty and delayed pain management.

Roadmap implication: this is the first instance of AI in American healthcare producing documented patient harm through an incentive structure rather than a model error — the model did what it was paid to do. Every conversation about clinical-AI safety has been about accuracy; this one is about who pays whom for which answer. If you are building anything that sits between a patient and a payment decision, the diligence question is no longer "how accurate is it" but "what is it paid to conclude, and what is the penalty for being wrong." Expect that question from regulators within a year.

• EFF — New records reveal problems with Medicare's AI prior-authorization experiment
• STAT — Medicare's WISeR AI prior-authorization pilot rushed launch, delayed care
• STAT — ARPA-H funds health tech companies to build AI cardiologists

4. Clinical AI maturing into a reimbursable modality — MOVED. The evidence caught up with the category, and it isn't flattering. [real-world]

Ambient documentation is the most-deployed clinical AI in America and this was the week its evidence base got tested in public. Mass General Brigham ended a multi-year head-to-head pilot of Microsoft DAX/Dragon Copilot against Abridge and chose Microsoft, declining to say why — set that against Abridge CEO Shiv Rao's 2024 line to STAT: "We've done several head-to-heads now and we haven't lost a single one." Separately, STAT reported that the minutes scribes save ER physicians do not convert into shorter waits or faster results, and that nobody is running a study on whether scribes improve patient outcomes at all. PHTI published that virtual CKD management solutions "show no consistent evidence of slowing disease progression" versus usual care, with spending reductions "immaterial when spread across the full attributed CKD population."

The vendors' response is instructive. Abridge moved into revenue cycle with pre-bill claim review — comparing coded diagnoses and DRGs against clinical documentation before submission, first customer Reid Health — and Rao's framing is the most honest sentence in health AI this year: "Health systems aren't reimbursed for the care that they deliver. They're reimbursed for the care they document." Suki, meanwhile, launched a research arm explicitly to "establish a gold-standard framework for measuring the performance and impact of ambient clinical intelligence." When the category leader pivots toward billing and the challenger starts funding its own evidence base in the same week, the clinical-benefit story has stopped selling on its own.

The money agrees. Every financing this week landed in operations rather than diagnosis: Forus $150M Series C at a $3B valuation (Bain Capital Ventures) for agentic prior-auth cutting weeks to under 48 hours; Verily took an undisclosed NVIDIA investment extending its March $300M round, funding Forecast 1.0, a multimodal genomic-plus-EHR model for chronic-disease risk; Inspiren $70M Series C at over $500M for fall detection in senior living; Implicity $40M; Epsilon Health $27.6M to run an "AI-native radiology practice" — selling the read, not the software; GenHealth $16.5M for revenue-cycle agents.

Roadmap implication: unchanged from Edition #4 and now better evidenced — the reimbursable unit is a workflow with an existing billing code attached. What is new is the downside case: three independent negative findings in one week against the two most-deployed categories. Ask any clinical-AI company for its adoption curve at week twelve, not week one, and for one outcome measured on a patient rather than on a clinician's calendar.

• STAT+ AI Prognosis — AI can't fix health care (Sep 9)
• Fierce Healthcare — Abridge expands into revenue cycle

Ripples

1. DeepMind precomputed every possible single-letter change in the human genome [in-silico]

AlphaGenome Atlas (Sept 8) publishes molecular-effect predictions for all ~9 billion possible single-nucleotide variants, roughly 27,000 predictions per variant across gene expression, transcription and splicing in hundreds of human and mouse cell types — about one petabyte, described as more than thirty times the size of the AlphaFold Database. It ships with a composite AlphaGenome Variant Impact (AVI) score merging AlphaGenome and AlphaMissense. On retrospectively solved rare-disease cases, AVI placed the known causal variant in a patient's top 50 candidates 29.5% of the time against 12.5% for CADD. Free web portal for academics; commercial access via Google Cloud "coming soon."

So what: the honest reading is that this is a spectacular measurement artifact with a modest decision improvement attached — 29.5% is a large relative gain over CADD and still means the causal variant is missing from the shortlist seven times in ten. The accompanying paper is a bioRxiv preprint, not peer-reviewed, DeepMind's own blog does not state the benchmark comparison (it came via press), and DeepMind flags enhancers as a weak spot. Treat as [in-silico] and watch the diagnostic-yield studies. The strategically interesting phrase is "commercial access coming soon": the rare-disease interpretation market is about to have a free academic tier and a priced enterprise tier from the same vendor.

• DeepMind — AlphaGenome Atlas

2. The first AI-drug-discovery company to get a drug approved has gone bankrupt [deal]

BioXcel Therapeutics filed Chapter 11. Teva is stalking-horse bidder at $57.5M upfront plus $67.5M in milestones for a company whose lead product, Igalmi, was approved and on the market.

So what: every bull case in this category implicitly ends at approval. BioXcel got there and it wasn't enough — commercial execution, payer access and capital structure killed it, none of which an AI platform improves. When you underwrite an AI-discovery company, model the launch, not the IND.

3. Owkin signs a fourth pharma with no number; PostEra prints one [deal]

Owkin licensed K Pro and its MOSAIC multimodal patient data to Servier for oncology — the fourth top-20 pharma after AstraZeneca (May), Sanofi (June) and Boehringer Ingelheim (September), and the fourth consecutive deal announced with terms undisclosed. In the same week Merck KGaA disclosed it is paying PostEra "mid-double-digit millions" for two AI-designed fertility programs (oral FSH-receptor and LH/choriogonadotropin-receptor agonists, targeting the injectables used in IVF), with PostEra keeping its PMOS program in-house and trials planned early next year.

So what: we flagged Owkin's three silences last week as a price signal. A fourth silence, in the same seven days that a comparable transaction printed a mid-double-digit-million figure for two named programs, sharpens it considerably. "Mid-double-digit millions" is now the public anchor for what an AI-designed program with a pharma partner is worth. Ask any AI-platform company you're looking at to price itself against that number.

4. A retired medicinal chemist designed a drug candidate with off-the-shelf ChatGPT [preclinical]

At ACS Fall 2026, Guibai Liang — a veteran big-pharma chemist, now co-founder of Sheo Pharmaceuticals — presented a novel renin inhibitor for hypertension designed using consumer ChatGPT and Gemini, no proprietary model. The team started from Takeda's abandoned TAK-272, interrogated the models on structure-activity questions (the role of the piperidine hydrogen-bond donor, whether the morpholine oxygen was necessary, metabolic stability), designed a new chemical class, had a CRO synthesize ~200 compounds across several iterations, and landed on SHEO-054, which reportedly outperformed aliskiren — the only FDA-approved renin inhibitor — in a monkey hypertension model. Liang: "it's impossible for human chemists to remember all those data to come up with a clear picture."

So what: heavy caveats — the structure was not disclosed, the data is self-reported conference material, and there is no paper. We are not moving anything on the ladder for this. But the enabling condition is worth more than the result: a dense, published, abandoned literature around a dead target. That describes a very large share of pharma's discontinued programs, and it is an asset class nobody is pricing. The interesting question this raises is not whether frontier models design molecules — it is whether the moat in AI drug discovery is the model at all, given a retired chemist with a $20 subscription and a CRO budget got this far.

5. The complete connectome of a fruit fly brain, reconstructed by AI [wet-lab]

Google Research and HHMI Janelia published in Cell the complete AI-reconstructed connectome of the male fruit fly brain and nerve cord — 166,000 neurons and 125 million synapses, the largest ever produced, and the first enabling neuron-by-neuron comparison between sexes.

So what: file this next to AlphaGenome Atlas. Both are the same move — AI used to produce a complete measurement of a system rather than a prediction about one — and both are the kind of substrate the virtual-cell field keeps saying it lacks. The pattern to watch across the next year is capability shifting from generating candidates to finishing atlases, because atlases are what downstream models are starved of.

6. Anthropic is hiring someone to buy biotech companies [deal]

A job listing for a biotech and life-sciences corporate development lead, paying up to $600K, to run "acquihires, tuck-in acquisitions, minority investments, and strategic partnerships" with AI-native companies across drug discovery, clinical development, regulatory and medical writing, lab automation, and healthcare data infrastructure.

So what: following the Coefficient Bio acquisition and the John Jumper hire, this is the clearest public signal that a frontier lab intends to buy its way into the life-sciences stack rather than partner into it. For anyone holding an AI-native life-sciences asset, the buyer universe just added a bidder with a very different cost of capital than pharma — and a very different view of what a model company is worth.

Pipeline watch

Scoreboard: zero full FDA approvals of an AI-discovered drug — unchanged. New sub-line as of this week: one AI-designed drug with a patient dosed in a registrational trial (company-asserted, not yet registry-confirmed).

The Signal — Bio/Health is published weekly by The Excelsior Group. Archive, including Edition #4: https://excelsiorgroup.ai/insights/signal/bio/