The Signal: Bio/Health — Edition #6 — September 22, 2026
The Signal: Bio/Health — Edition #6 — September 22, 2026
The Excelsior Group · Covering September 15–21, 2026 · https://excelsiorgroup.ai/insights/signal/bio/
The Read
This was the week the evidence gap got a price. On Friday, Eric Topol published the most thorough public audit yet of heart-rate variability and readiness scores and came back with nothing: no peer-reviewed evidence that either metric, or any effort to move it, correlates with a health outcome. That landed ten days after Apple made a 0–10 readiness score a flagship watch feature, four days after Oura joined a CMS payment model, and days before Oura priced its IPO at $40–44 a share. In a separate register, an external validation across 39 hospitals and 211,238 patients found that Epic's End-of-Life Care Index — a mortality model running inside the dominant US electronic health record — discriminates moderately and calibrates badly, systematically overestimating death. And twelve days after Insilico announced it had dosed the first patient in what it calls the world's first Phase 3 of a generative-AI-designed drug, ClinicalTrials.gov still says that trial has not started. Three unrelated failures with one shape: in each case the thing being sold, deployed, or capitalized is a number that nobody independent has checked. Our ladder exists because that is the normal condition of this sector, not the exception.
Tide status
No tide movement this week. Both standing tides hold. One of them is now carrying an unverified claim into its third week, and we are escalating how we describe it.
Biology is becoming an engineering discipline — holds; top rung now unverified for three consecutive editions [Phase 3]
We have said since Edition #4 that the ClinicalTrials.gov record for GENESIS-IPF-3 was stale and that we would notice if it stayed stale. It has. As of this morning, NCT07687459 still reads NOT_YET_RECRUITING, with an estimated start date of August 30 2026, a last update posted July 7 2026, a status verified date of June 2026, and 48 listed sites against the 47 in Insilico's release — the same record, unchanged, twelve days after the company said a human was dosed. We set a three-week trigger. This is week three, so here is the escalation, stated plainly: the sole public evidence that a generative-AI-originated medicine has entered a registrational trial is a company press release. Registry-maintenance lag at Chinese sites remains by far the most likely explanation and we are not alleging anything worse. But the distinction our whole framework rests on is between a claim and a confirmation, and we will not print this as confirmed because it would be convenient for the thesis if it were. Worth holding alongside it: on September 3 the directors of CDER, CBER, CDRH and the Oncology Center of Excellence jointly published a statement that FDA is expanding foreign Bioresearch Monitoring inspections, prioritising scrutiny of non-IND Phase 1 and early-feasibility studies in jurisdictions where site access is constrained, and will be systematic about telling sponsors when a site cannot be inspected. A 47-site Chinese registrational trial whose public registry entry has not been touched since July is not the profile that regime was built to wave through.
- Insilico — first patient dosed, GENESIS-IPF-3
- ClinicalTrials.gov — NCT07687459
- FDA Voices — Good Clinical Practices Are Not Optional
The binding constraint is disease understanding, not molecular design — holds, widened again [real-world]
Last edition we widened this tide from the disease→mechanism end to the drug→patient end on the strength of adoption decay and null throughput results. This week it widened again, to a third failure mode: the measurement itself. A readiness score with no outcome validation and a mortality model that overestimates death are not capability failures — both systems compute exactly what they were built to compute. They fail at the step where a number is supposed to become a decision. Across three consecutive editions the pattern is consistent enough to state as a rule: in clinical AI, the constraint is almost never the model.
Waves
1. Medical forecasting maturing into a reimbursable modality — MOVED HARD. The two most widely deployed prediction products in health both failed an independent check in the same week. [real-world]
Epic's End-of-Life Care Index. Published September 11 in JAMA Network Open, surfaced through STAT's coverage on September 18 and 22: a retrospective external validation across 39 hospitals in two systems — Trinity Health (24 hospitals in 11 states; 116,749 patients, 154,063 encounters) and Kaiser Permanente Southern California (15 hospitals; 94,489 patients, 133,043 encounters). Discrimination was moderate to good — C-statistic 0.76 at Trinity, 0.81 at KPSC. Calibration was not. Both cohorts overestimated mortality, and the scaled Brier score came in at −0.01 at Trinity Health. A negative scaled Brier is the number to sit with: it means the model's probability estimates carried no net advantage over the no-information baseline at that site. Subgroup performance was, in the authors' framing, reasonably equitable across race, ethnicity and sex, and worse in the oldest patients and some diagnostic subgroups. The commentary accompanying it makes the point we would have made: the question is not only whether a model performs, but whether it is implemented in a way that supports the decision it is attached to. UCSF geriatrician James Deardorff's line to STAT is the operator's version — a model that looks good in aggregate can be bad exactly where it is used.
Wearable readiness scores. Eric Topol's September 19 review is the most complete public treatment of heart-rate variability and readiness scores we have seen, and its conclusion is a null. The outcome-association literature for HRV — the mortality and morbidity links everyone cites — was built on ECG-derived HRV. Consumer devices measure pulse-rate variability by optical plethysmography, which is a different quantity: one study in over 900 adults found PRV non-uniformly underestimated across cardiovascular, endocrine, neurological and respiratory disease and called it an invalid surrogate; a systematic review of 43 studies concluded PPG-derived HRV should not be treated as universally interchangeable with the ECG-derived kind across devices, populations and contexts. A UK Biobank analysis of more than 46,000 participants using genetically predicted HRV failed to reproduce the HRV–mortality link, which argues the association is not causal. On readiness scores specifically: a review of 14 composite recovery and readiness scores found HRV contributes 86% of the composite, resting heart rate 79%, activity and sleep duration 71% each — and only resting heart rate is consistently accurate across devices. The algorithms are proprietary, unstandardised, not comparable between vendors, and revisable by the vendor without notice. Topol found no study linking any readiness score to a health outcome. The best available is a company-authored, uncontrolled observation that a 10-point Recovery increase tracked roughly half a stroke per round in 389 professional golfers.
Roadmap implication: the reimbursable unit is still a workflow with a billing code attached — but this week supplies the diligence question that goes with it. For any predictive product, ask for the calibration curve, not the AUC. Discrimination tells you the model ranks patients correctly; calibration tells you whether the number it prints means what the interface says it means. Epic's index ranks well and prints a number that is wrong, and it is inside most American hospitals. For consumer health, the question is narrower and harder: name one randomised outcome. Two adjacent data points from the same week, both about uptake rather than accuracy: Talkspace's municipal contracts are delivering well under projection — one Seattle deal worth over $2M assumed 2,200–4,400 registrations and treated 442 young people — and Omada, at $310M trailing-twelve-month revenue and 1.1 million members, used its first investor day to float expansion into chronic kidney disease, heart failure and sleep apnea, categories where PHTI's September 9 assessment already found virtual management does not slow progression.
- JAMA Network Open — Accuracy and Equity of the End-of-Life Care Index in Predicting 1-Year Mortality
- STAT — Geriatrician explains why AI for older adults deserves careful scrutiny
- Ground Truths — The Big Holes in Wearable Heart Rate Variability and Readiness Scores
2. The regulatory regime for adaptive and learning systems — MOVED. FDA took the word "animal" out of its nonclinical testing rules and wrote "computer models" in. [policy]
On September 21 FDA issued a direct final rule (Federal Register 2026-19350, with companion proposed rule 2026-19349) revising its nonclinical testing terminology. The substance is definitional and therefore easy to under-read: "animal tests" and "animal studies" become "nonclinical tests" and "nonclinical studies" across the affected parts, with related terms including "preclinical" and "in vitro" updated to match, aligning the regulations with the Food and Drug Omnibus Reform Act of 2022. The rule names computer models alongside human cells, organs-on-chips and other advanced technologies as methods that can supply nonclinical evidence. Acting Commissioner Kyle Diamantas framed it as supporting animal studies where they remain appropriate and validated alternatives where they can carry the evidentiary load. Because it is a direct final rule, significant adverse comment withdraws it in favour of the ordinary notice-and-comment track.
Here is the calibrated version. This does not validate any computational model, does not create a pathway, and does not tell a sponsor that an in-silico package will be accepted in place of a tox study. What it does is remove a definitional obstacle: the regulation no longer presupposes that nonclinical evidence is animal evidence. That is a necessary condition for in-silico evidence to have standing, and it is now met. It is not a sufficient one, and the gap between the two is where the next three years of argument will happen.
Against that, the negative finding we have now carried for three editions has gotten large enough to be the story: FDA's AI-Enabled Medical Device List page was last refreshed 09/04/2026 and the most recent Date of Final Decision in the table is still June 29, 2026. That is roughly twelve weeks without a new AI-enabled device authorization appearing on the list. We are not yet calling it a slowdown in authorizations rather than a lag in list maintenance — those are different claims and we cannot separate them from outside — but twelve weeks is long enough that it now belongs in anyone's regulatory timeline assumptions. Minor institutional note in the same week: W. Alex Smith joined FDA's Digital Health Center of Excellence from Hogan Lovells, where he was senior director of regulatory compliance.
Roadmap implication: the standing action item is unchanged and the clock is shorter — comment on docket FDA-2026-N-7874 before October 19 2026, now under four weeks. Add a second: if you have an in-silico safety story, the September 21 rule is the first time the regulatory text is neutral about the substrate. Early comment on the direct final rule is cheap and the window is open.
- FDA — Updates Regulations to Advance Innovative Alternatives to Animal Testing
- FDA — Artificial Intelligence-Enabled Medical Devices
3. Clinical-AI autonomy vs. mandated human oversight — MOVED. The autonomy question moved from the device counter to the payment counter. [policy]
On September 17 Counsel Health announced it is joining the CMS ACCESS model's Early Cardio-Kidney-Metabolic track, with Oura as technology and marketing partner. Counsel Health is an AI-native primary care company that delivers care through a messaging interface backed by board-certified physicians; Oura will promote the program to its members and, with consent, contribute biometric data. Target conditions are hypertension, obesity, hyperlipidemia and prediabetes; coverage spans nearly all US states; no out-of-pocket cost to eligible beneficiaries; start in early 2027.
Read it against what we logged last edition. CMS named its first Chief Clinical AI Officer with ACCESS as the vehicle and a December 2028 evidence checkpoint; FDA's TEMPO pilot offers enforcement discretion to participants in exchange for real-world data. ACCESS is where the federal government is buying clinical AI before the evidence exists, deliberately, in order to generate the evidence. What is new this week is the product shape now entering it: an AI-first primary care service whose data feed is a consumer wearable — one whose core derived metric, per the same week's review, has no validated link to any health outcome. The interesting part is that this is not obviously wrong. If you believe evidence has to be generated somewhere, a time-limited federal model with a stated checkpoint is a better place to run the experiment than the open consumer market, which is where these scores live today with no checkpoint at all.
Roadmap implication: for anything entering ACCESS or TEMPO, the diligence question is what happens at the checkpoint. December 2028 is the date CMS has put on the record. Ask what evidence the participant is committing to produce by then, who adjudicates it, and what the downside is if it does not arrive — and pair it with the question we added last edition for anything sitting between a patient and a payment decision: not how accurate is it, but what is it paid to conclude.
- Businesswire — Counsel Health Joins ACCESS Model to Bring AI-Native Chronic Care to Medicare Beneficiaries
- Fierce Healthcare — Oura, Counsel Health join CMS ACCESS model
4. The AI-drug-discovery show-me window — MOVED. The capital drought is now the measurement. [deal]
Last edition we logged, as a negative finding, that no pure-play AI drug-discovery venture round closed in the window. It has now happened twice in a row, and going to the dedicated tracker rather than the week makes it worse: the most recent entry on the AI drug-discovery venture funding tracker is Accipiter Bio's $10.5M seed extension on July 31 2026. That is roughly seven and a half weeks without a logged round in the category, while health-AI capital in the same period went to an employer health plan ($600M), a clinical-answers engine ($250M), prior authorization ($150M), fall detection, cardiac monitoring and revenue-cycle agents. Capital is not asking discovery platforms to prove anything right now; it has stopped asking them at all.
Two data points on the other side, pointing in opposite directions. Isomorphic Labs put its head of drug discovery, Chris Butler, on the record with Semafor on September 15: the company is on track, moving smoothly through preclinical, partnered with Novartis, Eli Lilly and Johnson & Johnson, sitting on the $2.1B it raised in May 2026 led by Thrive Capital, with its models kept entirely in-house. He declined to give clinical timelines. Set that beside the contrast we drew last week and it sharpens rather than softens: the best-capitalised venture in AI drug design is still preclinical with no named molecule, and its head of discovery will not say when that changes. Meanwhile the cheapest possible version of the same activity produced a candidate — see the Sheo ripple below — using two consumer chatbots and a CRO.
Roadmap implication: the scoreboard is unchanged at zero full FDA approvals of an AI-discovered drug. But the diligence question has moved from the one we set in Edition #5 — model the launch, not the IND — to a blunter one for platform companies specifically: if the category cannot raise, what is the exit? The buyer universe added a frontier lab this quarter. For an AI-native discovery asset, that may now be the realistic path rather than the IPO, and it prices very differently.
- Semafor — AI biotech Isomorphic Labs is not slowing down, exec says
- Cure — AI Drug Discovery Venture Funding Tracker 2026
Ripples
A retired chemist designed a renin inhibitor with off-the-shelf ChatGPT and Gemini [preclinical]
At ACS Fall 2026 in Chicago, Guibai Liang — a veteran medicinal chemist who co-founded Sheo Pharmaceuticals after leaving big pharma — described designing a novel renin inhibitor for hypertension using ChatGPT and Gemini with no proprietary model in the loop. Renin is one of the most picked-over targets in the industry; every major pharma ran a program and aliskiren is the only one FDA ever approved. Sheo started from Takeda's TAK-272, put mechanistic structure-activity questions to the chatbots, used their synthesis of decades of published SAR to design a new chemical class, had a CRO make roughly 200 compounds across several iterations, and landed on SHEO-054, which the team reports outperformed aliskiren in a monkey model of hypertension.
So what: treat the result as unverified — conference talk, self-reported, structure undisclosed, no paper, and one primate model is a long way from a development candidate. Treat the method as real. The enabling condition Liang himself names is a dense, published, abandoned literature, which describes a large share of pharma's dead targets. If that generalises even partly, a meaningful slice of the value AI platforms are being funded to create is available to a two-person team with API access and a CRO budget. That is a margin question for every platform company in the category, and it lands in the same week the category could not raise.
Oura priced its IPO in the same week its core metric was audited [deal]
Oura set an indicative range of $40–44 per share for an offering of 50,000,000 shares — 13,500,000 from the company and 36,500,000 from existing stockholders — resolving the pending S-1 we flagged last edition. The amended filing states the company intends to apply approximately $526.4 million of net proceeds to anticipated tax withholding and remittance obligations.
So what: the sequencing is the point. Within one week Oura joined a CMS payment model, priced an IPO, and had the metric family at the centre of its product characterised by the field's most credible reviewer as having no validated link to a health outcome. None of these contradict each other — consumer demand for readiness scores is real and is what is being priced — but a public-market story built on health claims now has a citable, high-profile null sitting against it. For anyone holding or underwriting: the risk is not that the devices are inaccurate. Resting heart rate is accurate. The risk is that the composite score is a product feature dressed as a clinical measure, and that distinction is now on the record.
Talkspace's city contracts delivered a fraction of the projected uptake [real-world]
Reporting from the Seattle Times and Proof News found that Talkspace's municipal deals to provide mental-health support to young people are running well below expectation. In Seattle, a contract worth over $2 million assumed 2,200 to 4,400 individuals would register for therapy services; 442 young people received treatment.
So what: this is the adoption-decay finding from Edition #5 in procurement form. There, emergency-department clinician use of LLM decision support fell from 67.9% to 29.7% over four weeks. Here, a municipal buyer paid for capacity against a registration forecast and got roughly a tenth to a fifth of it. Same failure, different side of the transaction: the model works, the distribution does not. If you are selling into public buyers, expect utilisation floors and per-engagement pricing to start appearing in contracts, and price accordingly.
OpenEvidence raised $250M at a $15B valuation [deal]
Confirmed by CEO Daniel Nadler to STAT after surfacing in Forbes coverage of the company's new agreement with Memorial Sloan Kettering: a quiet $250 million round led by Byers Capital and Andreessen Horowitz, valuing the clinical-answers company at $15 billion.
So what: this is the largest private valuation in the clinical-decision-support layer and it was raised without an announcement, which tells you something about how competitive the round was. Note what OpenEvidence sells: not a prediction, not a diagnosis, but retrieval and synthesis over the literature with citations attached — the one clinical-AI product category whose output a physician can independently check in seconds. In a week defined by unverifiable numbers, the highest-priced asset is the one that shows its sources.
Angle Health's $600M is the week's largest health-AI round, and it is an insurer [deal]
Angle Health raised $600 million at a $2.7 billion valuation, led by Vitruvian Partners, to expand its AI-native health plans for small businesses. The company was founded by former Palantir engineers.
So what: the pattern we called out last edition holds and hardened. The three largest health-AI raises this week — $600M into insurance administration, $250M into clinical retrieval, and the $150M prior-authorization round still echoing from last week — all sit on existing payment rails. Zero went to discovery or diagnosis. The money is not betting on new biology; it is betting on taking cost out of the paperwork between a patient and a payment. That is a real and probably correct trade. It is also the reason the discovery scoreboard has not moved.
The largest connectome yet: 166,000 neurons and 125 million synapses [in-silico]
Google Research and HHMI Janelia published in Cell the complete AI-reconstructed wiring diagram of the male fruit fly's brain and nerve cord — 166,000 neurons and 125 million synapses, the largest connectome assembled to date. Paired with the existing female map, it allows neuron-by-neuron comparison between sexes and establishes the reconstruction methods for zebrafish and mouse brains.
So what: file this with the AlphaGenome Atlas from last edition as the same species of artifact — a complete map rather than a prediction. These are the inputs that make the tide real, and they are also the reason to keep the ladder strict: a complete wiring diagram of a fly is a genuine scientific achievement and sits at the bottom rung of anything clinical. The honest framing is that the measurement layer of biology is being industrialised far faster than the inference layer that would turn measurements into medicines.
Pipeline watch
- Insilico Medicine · rentosertib (ISM001-055), IPF · Phase 3 · ESCALATED. No change in the registry. NCT07687459 still NOT_YET_RECRUITING, estimated start August 30 2026, last update posted July 7 2026, status verified June 2026, 48 sites vs 47 in the release — twelve days after the announced first dosing and three consecutive editions after we first flagged it. Company-asserted, not registry-confirmed, and we now say so at tide level.
- Isomorphic Labs · first candidates · preclinical · updated disclosure. Head of drug discovery Chris Butler on the record September 15: on track, moving through preclinical, partnered with Novartis, Lilly and J&J, models kept in-house, clinical timelines declined. Still no named clinical candidate.
- Sheo Pharmaceuticals · SHEO-054, renin inhibitor, hypertension · preclinical · ADDED. AI role is off-the-shelf LLM-assisted SAR reasoning, not a proprietary platform — added precisely because it is the cheap-path counterexample. Self-reported at ACS Fall 2026; one monkey model; no structure disclosed and no paper. Verify against any publication or IND filing.
- Generate Biomedicines · GB-0895 (golukibart) · Phase 1 · no change, confidence still downgraded. The Sep 8 ERS asset remains unretrieved and no company release or peer-reviewed document has appeared confirming the ERS 2026 COPD numbers. The only primary GB-0895 data in public is still the ERS 2025 asthma Phase 1 abstract. Third consecutive edition carrying this unresolved.
- Merck / Moderna · intismeran autogene · Phase 3 · no change. Still zero effect sizes disclosed since the August 19 Phase 3 win. ESMO Madrid, October 23–27, remains the next hard catalyst on this tracker — one month out.
- PostEra / Merck KGaA · oral FSH-R and LH/CG-R agonists · preclinical · no change. No registry entry yet; trials guided to early next year.
- Superluminal, insitro, Recursion, Xaira, EvolutionaryScale/Chai, Anthropic/Adaptyv · no change. Nothing in window.
Scoreboard: zero full FDA approvals of an AI-discovered drug. Unchanged since we opened this tracker. Sub-line unchanged and now explicitly caveated: one AI-designed drug with a patient reported dosed in a registrational trial, company-asserted, still not registry-confirmed after three weeks.
The Signal: Bio/Health is published weekly by The Excelsior Group. Archive and back issues: https://excelsiorgroup.ai/insights/signal/bio/
