The Signal: Bio/Health — Edition #8 — October 6, 2026

Covering September 29 – October 5, 2026

Correction: the FDA device counter was never dark. Our instrument was.

For four consecutive editions — #4 through #7 — this newsletter reported that the FDA's AI-enabled medical device counter had stopped moving. By Edition #7 we were calling it "thirteen weeks," citing a page whose most recent Date of Final Decision was June 29, 2026, and we built a wave observation on the silence.

The silence was ours.

This week the URL we had been reading returned 404. The FDA has moved the resource to its Digital Health Center of Excellence, where the page is marked content current as of 09/22/2026 — and where the dated, scrapeable table has been replaced by a "searchable resource" with no decision dates on the face of the page. That page states the FDA "has authorized over 1,600 AI-enabled medical devices for marketing in the United States as of September 2026." MedTech Dive, reading the FDA's own March 4, 2026 data pull, put the figure at more than 1,400. Roughly two hundred authorizations in six months is not a stalled regulator.

We withdraw the slowdown finding. We are not replacing it with a new count: the two numbers come from different pulls and are rounded, and without the dated table we cannot responsibly state a weekly or monthly rate. The honest position is that we do not currently have a working instrument for this measurement, and we will say so each week until we do.

This is the second consecutive edition opening with a correction, and both corrections have the same shape: not a bad judgment call on good data, but a confident claim built on a retrieval path that had quietly broken. That pattern is now the more important finding than either error.

Links: - FDA — Artificial Intelligence-Enabled Medical Devices (Digital Health Center of Excellence) - MedTech Dive — AI in medtech is booming. Track new devices here.

The Read

The week's two best items both attacked a measurement rather than producing a result. Google DeepMind published a method in Nature for watermarking AI-designed proteins without degrading them, which converts provenance for generated biology from a policy aspiration into a lab-verified technique. The Peterson Health Technology Institute found that 68% of employers and 55% of health plans now tie digital health payments to performance — the money has moved to outcome-based contracting faster than the clinical deployments it is meant to discipline. Against that, the week's largest single gesture was capital, not evidence: OpenAI's fund led a $153M round at a $1B valuation for a company whose thesis is that a meaningful share of pharma's 90% failure rate is not hard biology but bad patient matching. And the frontier-lab story we told last week reversed inside seven days: Anthropic now says it will not take its own programs into humans and will not compete with its pharma customers. Old problems, new physics — but this week the new physics was pointed at the instruments, which is where it should have been pointed at us.

Tide status

MOVED — The binding constraint on new medicines is disease understanding, not molecular design

The tide holds, and it is being tested this week from a direction it has not been tested from before: someone has put a billion-dollar valuation on the claim that part of the 90% is a matching problem.

Biossil's position is not drug repurposing, which looks for a new disease for an old molecule. It is the narrower and more falsifiable claim that some failed trials contained a real efficacy signal in a more precisely defined subgroup, diluted to nothing across a broad enrolled population. If that is true at scale, then a slice of what this tide attributes to disease → mechanism ignorance is actually drug → patient assignment error — the third failure mode the tide widened to in Edition #5, now with a company and a cap table attached rather than an argument. CEO Anthony Mouchantaf's own framing concedes the hard part is downstream: "That requires dedicated infrastructure capable of integrating AI and medicine together from discovery through development and into the clinic."

The tide's measurement-validity limb widened again too, and twice. Externally, PHTI's purchasing survey (below) shows buyers pricing engagement and outcomes rather than model performance. Internally, our own FDA device measurement broke silently for three months. And there was a genuinely good week for measurement science: two papers — one in Nature from Vadim Gladyshev's group at Harvard, one in Cell from Juan Carlos Izpisua Belmonte's team at Altos Labs — gave the epigenetic aging clock a mechanism. Eric Topol's read is that what the clocks have been measuring all along is the erosion of PRC2 low-methylated regions: the slow, chromatin-architecture layer that locks cell identity. Loss of that layer produces mesenchymal drift, now verified across 46 tissue types and tied to atherosclerosis, macular degeneration and Alzheimer's. A measurement in wide use for a decade, including commercially, has only now been told what it measures.

So what: The tide is unchanged in direction and better instrumented. Two practical consequences. First, if you are evaluating an AI-discovery platform, ask separately what it claims about mechanism selection and what it claims about patient selection — Biossil is a bet that the second is underpriced, and it is the first bet of its size on that specific proposition. Second, treat any biomarker you are paying for as unvalidated until someone has shown what it is physically measuring; the aging clocks were accurate and unexplained for ten years, which is the exact profile of a measurement that can be accurate for the wrong reason.

Links: - The Globe and Mail — Biossil raises $153-million in OpenAI-led funding - BetaKit — Biossil hits unicorn status with new OpenAI-led financing - Ground Truths — Loss of Cell Identity Drives Human Aging

HOLDS, no movement — Biology is becoming an engineering discipline

Nothing moved the top clinical rung this week. NCT07687459 (GENESIS-IPF-3, rentosertib) stands where Edition #7's correction left it: RECRUITING, start date 2026-09-09 marked ACTUAL, 320 estimated enrollment, 47 sites. No new registry activity, and we did not re-read the registry this run — see the Coverage Log for why, and treat the Edition #7 record as the current one rather than a fresh confirmation.

"No tide movement this week" is the default here and remains a feature.

Waves

Wave 1 — The regulatory regime for adaptive/learning systems takes shape [policy]

MOVED, in two directions, neither of them the one we were watching.

First, the FDA restructured how it publishes AI device authorizations: hub page moved, dated table replaced by a search interface, content current as of September 22. Whatever the intent, the practical effect is that the public, longitudinal, machine-readable record of AI device authorizations — the single best outside measurement of how fast this regime is actually clearing products — is harder to read than it was a month ago. We lost three months to that. Anyone else building a tracker on the old table lost the same three months.

Second, and sharper: a model retirement is a regulatory event. Yujan Shrestha, MD, writing in the week's Where Tech Meets Bio, makes the argument cleanly. Claude Sonnet 4 shipped in May 2025 and was retired in June 2026 — about 13 months, shorter than many De Novo reviews. Anthropic gives at least 60 days' deprecation notice; the FDA's goal for a Traditional 510(k) is 90 days. One of his team's API keys logged 10,236 failed requests in a single day after a retirement. Swapping the model string in a cleared device can be a design change requiring verification and validation. His remedies are concrete: write model replacement into a Predetermined Change Control Plan agreed with the FDA in a Pre-Submission; hold a sequestered, labeled test set that can requalify a replacement model in days rather than months; pin dated model snapshots instead of aliases that move on their own.

Roadmap implication: This is the first well-posed statement of a structural mismatch that will hit every generative-AI device on the market: frontier model lifecycles are now shorter than the regulatory clearance cycles of the products built on them. If you are building or buying a regulated product with a frontier model inside it, the PCCP is not paperwork — it is the only mechanism that lets your supplier's roadmap change without your clearance lapsing. Ask your vendor two questions: what is pinned, and what is the requalification path. A vendor who cannot answer is carrying an unpriced risk on your behalf.

Links: - Where Tech Meets Bio #86 - FDA — Predetermined Change Control Plan guidance (via the AI-Enabled Medical Devices hub)

Wave 2 — Clinical AI matures into a reimbursable modality [real-world]

MOVED. Payment discipline has arrived ahead of clinical deployment.

PHTI's 2026 State of Digital Health Purchasing, published September 29, surveyed 321 digital health decision-makers at health plans, employers and health systems between July 1 and 22. The findings, in order of how much they should change your plan:

• 68% of employers and 55% of health plans now use performance-based contracts for digital health.
• 85% of those contracts tie at least a quarter of fees to performance; 31% put more than half of fees at risk.
• 47% of purchasers say fewer than a quarter of eligible members ever enroll.
• 64% name strong member engagement as the prerequisite for increasing digital health spend — engagement, not clinical effect size.
• On AI specifically, health systems have gone broad on administration and stayed narrow on clinical judgment: 71% deploy AI for clinical documentation; only 18% deploy AI for clinical decision support enterprise-wide.

Set that beside what this wave already carries. Edition #5 logged ED decision-support adoption decaying from 67.9% to 29.7% over four weeks. Edition #6 logged Epic's End-of-Life Care Index externally validated across 39 hospitals and 211,238 patients with a respectable C-statistic and poor calibration in both systems. Now the buyers have independently arrived at the same conclusion by a different route: they are not paying for model performance, because model performance has not predicted value. They are paying for enrollment.

Roadmap implication: The near-term revenue question for clinical AI is no longer "is the model good" but "will anyone use it, and can you prove it monthly." A company selling into this market needs instrumented engagement data as a first-class product feature, not a reporting afterthought — 85% of contracts now require it to get paid. And the 71%-versus-18% split is the clearest available statement of where health systems currently believe AI is safe: near the chart, not near the decision.

Links: - PHTI — 2026 State of Digital Health Purchasing - PHTI announcement — Majority of employers and health plans use performance-based contracting

Wave 3 — Frontier AI labs and the clinical layer [deal]

REVERSES. We had the direction wrong last week, and the correction arrived in seven days.

Edition #7 ran this wave under the heading "the frontier labs stopped being vendors," on three facts: Anthropic operating its own wet lab, the Novo Nordisk collaboration, and a thickening enterprise layer. The inference was that owning a lab was a step toward owning assets.

This week Anthropic said the opposite. Reported by Endpoints' Andrew Dunn and relayed in Where Tech Meets Bio: Anthropic does not plan to start human testing or Phase 1 trials for its pipeline programs, with Eric Kauderer-Abrams saying it will not compete with its pharma customers. The logic is obvious once stated — if you are selling the reasoning layer into every large pharma's R&D organisation, owning clinical assets that compete with those customers is a conflict you cannot paper over.

And the supplier position kept thickening in exactly that direction. At Roche's Pharma Day on September 28, Anthropic, OpenAI and NVIDIA appear as named suppliers on Roche's own autonomy roadmap — not as collaborators with a stake in the molecules.

Roadmap implication: Restate the wave: frontier labs are consolidating as the supplier layer for pharma R&D, not entering drug development. That is a less dramatic thesis and a more investable one — the labs are choosing recurring revenue across the whole industry over equity in a handful of assets, which is the same choice AWS made and the opposite of the one Google made in Isomorphic. Two caveats we are keeping on the record: the claim reaches us as a reported quote rather than a company statement, and the report as quoted does not say what Anthropic's "pipeline programs" are or what happens to them if they never enter humans — partnering or licensing is the natural route, and nobody has said so. Treat this as company-asserted via trade press until Anthropic says it in its own voice.

Links: - Endpoints News — Q&A: Anthropic's life sciences team on its enzyme research, biology lab and a preclinical pipeline - Endpoints News — Unpacking Anthropic's 100-day sprint into biopharma - Where Tech Meets Bio #86

Wave 4 — The AI-drug-discovery show-me window [deal]

MOVED on capital, not on evidence — and the gap between the two widened this week.

Six financings and a licence inside seven days, the largest cluster this newsletter has logged: Biossil $153M at a $1B valuation (OpenAI Startup Fund leading); Enveda $311M Series E at $2B, led by Catalio, taking it past $845M raised; Anew Labs $290M at a reported $1.5B valuation in its first external round, with ByteDance reportedly holding 56%; Basecamp Research $140M including NVIDIA and S32 to launch a therapeutics pipeline; BigHat $75M Series C behind an AI-designed ADC already in Phase 1; and Earendil Labs signing a research collaboration with Genentech valued above $1.5B on bispecific cancer antibodies. Precision Neuroscience's $250M and Varda's $251M sit alongside as adjacent hardware and manufacturing bets.

Against that, the week's disclosed evidence was Xaira's first programs — real numbers, honestly reported, and still preclinical (below).

One more datum for the scoreboard, pointing the other way: Richard Bonneau has left his role as Genentech's global head of AI for drug discovery, per Endpoints' Andrew Dunn — the latest in a run of senior AI departures there that also includes KyungHyun Cho, Stephen Ra and Nathan Frey. Two sources cited tension with Aviv Regev; Genentech denies it. We are not running this as a verdict on anything. It is one of the few observable signals about whether large-pharma AI organisations are actually working, and four senior exits from the same group is worth a row in the ledger.

Roadmap implication: The show-me window is being financed rather than closed, which is the normal shape of a category between its first registrational asset and its first approval — and it means private marks in AI discovery are now set by platform narrative plus a Nasdaq comparable (Iambic, filed September 21), not by clinical data. If you are pricing into this, the asymmetry is that the first genuine Phase 2 failure of a flagship AI-originated asset will reprice the whole set, and the first approval will not reprice it nearly as much, because approval is already in the price.

Links: - Where Tech Meets Bio #86 - Fierce Biotech — Genentech inks $1.5B bispecific pact with fast-rising Earendil - BusinessWire — BigHat Biosciences announces $75 million Series C financing

Ripples

1. DeepMind watermarked AI-designed proteins without breaking them, and published the method in Nature [wet-lab]

Nature published "Function-preserving watermarking of AI-generated proteins" on September 30, the paper behind DeepMind's SynthID Bio. For sequences, the paper reports 100% true-positive detection at a 0.1% false-positive rate on its SARS-CoV-2 RBD binder set, with no significant population-level difference in SPR-measured binding affinity and comparable nanomolar to sub-nanomolar binders against SARS-CoV-2 RBD, VEGF-A and PD-L1. The structure method, which fine-tunes AlphaFold 3, exceeds 99.8% true-positive detection at the same false-positive rate with negligible effect on LDDT and template-modelling scores. The cost is stated rather than hidden: non-distortionary watermarking at temperature 0.5 reduces the pass rate by 41.7%. Code and in vitro data are being open-sourced, aimed at DNA synthesis screening. External reviewers included James Diggans of Twist Bioscience and Sarah Carter of Science Policy Consulting; in vitro validation was run by Adaptyv Bio, and the bacteriophage genome work with the Hie lab at Stanford and the Arc Institute.

The daily Signal ran this on September 30 with the numbers; what the weekly adds is the rung and one observation the daily did not make. Adaptyv Bio is the same third-party wet lab that validated Anthropic's minibinder designs, which this newsletter already carries on its pipeline tracker. Two frontier labs' most load-bearing biology claims are now being independently confirmed by the same vendor. That is good for comparability and bad for independence, and nobody is tracking it.

So what: Provenance for generated biology is now a published, function-preserving technique rather than a governance aspiration — which is what makes AI protein design defensible to a regulator instead of merely promising. If you design or order synthetic sequences, ask your synthesis provider when it will screen for this and treat the answer as a diligence item. And if you are relying on third-party wet-lab validation as your independence story, find out how many of your competitors use the same lab.

Links: - Nature — Function-preserving watermarking of AI-generated proteins - Google DeepMind — SynthID Bio: Watermarking methods for synthetic biology

2. Xaira finally showed its pipeline, with hit rates attached [in-silico]

After five consecutive editions in which this newsletter logged "nothing in window" for Xaira, the company published its X-Design platform and first programs on October 2. X-Design designs antibodies de novo, optimising simultaneously for binding, humanness, selectivity, cross-reactivity and developability; the current iteration is called Vega.

The numbers, as disclosed: XA-1, an oncology program requiring precision-constrained epitope binding with surrogate-species cross-reactivity, reached a progressable binder in three weeks and a lead in seven, at 17 nM affinity against human target and 83 nM against cynomolgus, from a 10% confirmed binder rate (18 of 182). XA-4, a GPCR target with no well-understood epitope that had resisted both conventional screening and llama immunisation, produced a 7.6% hit rate (7 of 92 screened), 183 nM IC50 antagonism and 32 nM EC50 binding. Validation targets CXCR4 and APJ returned 35% and 29% hit rates. Both XA-1 and XA-4 are in lead optimisation.

So what: This is the most useful disclosure of the week because it is specific enough to be wrong. A seven-week design-to-lead on a hard GPCR is a genuine platform result, and the honest rung is [in-silico] moving into wet-lab confirmation, not clinical anything — there is no IND, no trial, no candidate nomination. Use these hit rates as the benchmark the next antibody-design platform has to beat: 7.6% on a target that defeated llama immunisation is now the public number to argue with.

Links: - Xaira — De Novo Design of Progressable Antibodies to Therapeutic Targets

3. OpenAI's fund bought the thesis that failed drugs failed the wrong patients [deal]

Biossil, out of Toronto, emerged from stealth in April with $70M and a contrarian plan for an AI company: instead of designing new molecules, acquire or license ones that failed their own clinical trials. On September 29 it raised a further $153M (C$216M) at a $1B valuation, led by the OpenAI Startup Fund, with Founders Fund, Quiet Capital, Modern Capital, Golden Ventures, Panache Ventures, Duke University's endowment and the Abu Dhabi Investment Council participating. Programs span sickle cell disease, idiopathic pulmonary fibrosis, glioblastoma, breast cancer and Alzheimer's. The round also fits OpenAI's own life-sciences push — GPT-Rosalind, a reasoning model for biology and translational medicine released in April through its trusted-access program with Amgen, Moderna, the Allen Institute and Thermo Fisher, on top of the earlier Novo Nordisk partnership.

So what: The rung here is [deal] and nothing else — no asset in this portfolio has been shown to work in a redefined subgroup, and retrospective subgroup rescue is one of the most reliable ways to fool yourself in clinical statistics. The reason to take it seriously anyway is that it is a cheap, falsifiable bet on a specific decomposition of the 90% failure rate, and it will generate real readouts faster than de novo discovery can. Watch for the first prospective trial in a model-defined subgroup, which is the only result that will settle it. If you hold AI-discovery positions, note that this thesis competes with yours for the same capital while claiming a shorter path to data.

Links: - The Globe and Mail — Biossil raises $153-million in OpenAI-led funding to bring resuscitated drugs to market - BetaKit — Biossil hits unicorn status with new OpenAI-led financing

4. Roche put a six-level ladder under "autonomous lab" and a number under its AI claims [real-world]

At Pharma Day on September 28 — one day inside Edition #7's window, and missed by that sweep — Roche laid out a progression from fully manual through automated instruments, connected workflows and AI-assisted workflows to domain autonomy and lab-wide autonomy, with a parallel ladder for the human role: Operator → Assistant → Approver → Auditor. NVIDIA, OpenAI and Anthropic appear as named technology partners.

The number that matters: 40% of pipeline decisions from Q4 2025 through Q2 2026 had a tracked AI or computational contribution — of which 41% were rated fundamental or critical, 48% supportive and 10% informative. Roche also says TargetNexus, its target-assessment agent, is on track to inform 80% of research portfolio decisions by Q4. One program is claimed to have gone from lead identification to clinical candidate in 18 months.

So what: This is an investor-strategy presentation, not a filing or a publication, and the 80% figure is a target rather than a result. Even so it is the most quantified public account we have of AI inside a large pharma's decision process, and the Operator → Auditor framing is the honest version of what AI adoption does to a job: it moves people from execution to accountability, which is a harder transition than the vendor pitch admits and requires different people. The usable artifact is the ladder itself. Most organisations claiming an "autonomous lab" are at connected workflows; ask which rung, and ask what the human is accountable for at that rung.

Links: - FourWeekMBA — Roche Maps the Autonomous Lab in Six Levels (Pharma Day 2026) - Roche Holding AG — Pharma Day 2026 transcript

5. Two cancer-vaccine readouts went opposite ways, and the confound is the story [Phase 3] [Phase 2]

Merck and Moderna's intismeran autogene met its Phase 3 endpoints in resected high-risk melanoma in August. Days later, BioNTech and Genentech scrapped a mid-stage colorectal cancer study after their personalised vaccine failed. STAT's Allison DeAngelis reported the field's reaction on October 5. The obvious explanation is tumour immunology — melanoma is immunologically "hot," colorectal notoriously "cold." Researchers are cautioning against that reading, for a reason that belongs in this newsletter: intismeran was given with Keytruda; BioNTech's vaccine was given alone. Jason Luke of Strand Therapeutics put the open question plainly: why does Moderna have good data in melanoma and BioNTech not in colon cancer.

So what: This is the evidence ladder doing its job on a comparison that will be made carelessly everywhere else. Two trials at adjacent rungs in different tumours with different backbones do not constitute a verdict on personalised cancer vaccines, and anyone telling you the category is validated or dead this month is reading press releases. For our tracker, intismeran remains the nearest hard catalyst: still zero effect sizes disclosed since the August 19 win, with ESMO Madrid, October 23–27, the likely venue. The BioNTech failure is now the comparator that will frame however those numbers land.

Links: - STAT — Divergent trial results on cancer vaccines have the field grappling with hard questions

6. The FDA opened a pilot to cut first-in-human timelines, and the AI content is zero [policy]

This newsletter has never run the FDA's Expedited IND Pilot — "Operation TrialBlazer" — announced September 15. It is logged as a standing gap in Edition #7 and we are closing it now because the application window closes October 30, 2026.

The pilot pairs sponsors with Qualified Research Institutions whose scientific review lets the FDA assess IND components on a rolling basis rather than waiting for a complete submission, and encourages IRB review and site activation to run in parallel. The FDA expects to select 8–10 sponsor-QRI pairs for the first cohort. The agency's own framing of the problem is the part worth quoting back to anyone who thinks US trial infrastructure is fine: first-in-human trials take up to two years in the United States and are completed much faster in China and Australia.

There is no mention of artificial intelligence, computational tools or modelling anywhere in the announcement.

So what: That absence is the finding. This newsletter's tide says the rate-limiting stages are disease → mechanism and drug → patient, and the loudest thing the regulator did about the drug → patient bottleneck this quarter contains no AI at all — it is process redesign and parallelisation. If you are holding a preclinical AI-designed asset heading for an IND, the highest-expected-value regulatory action available to you in October is an application to this pilot, not a model improvement. Eight to ten slots, 24 days left.

Links: - FDA — Launches Expedited IND Pilot, Begins Accepting Applications - BioSpace — FDA opens Operation TrialBlazer for applications, hoping to reshore early trials

Pipeline watch

The scoreboard for whether AI drug discovery is actually working. Four entries moved this week; ten did not.

ADDED — BHB810 (BigHat Biosciences) · CDH17-directed VHH-Fc antibody-drug conjugate · gastric and advanced GI tumours · [Phase 1] First patient dosed September 1, 2026 in a first-in-human dose-escalation trial, principal investigator Alexander Spira at NEXT Oncology-Virginia; preclinical work across roughly 30 tumour models showed complete or near-complete clearance. The company describes it as "the first program we've taken from AI-generated design into human studies." A $75M Series C followed on September 23. No NCT number in the release — to verify. This belongs on the tracker and should have been added in Edition #6; the dosing announcement predates our window by five weeks and we missed it. Second candidate BHB299 guided to the clinic in 2027.

UPDATED — Xaira Therapeutics · XA-1 (oncology), XA-4 (GPCR) · [in-silico] → lead optimisation Five editions of "no disclosed clinical programs" ends. Hit rates, affinities and timelines now public (see Ripple 2). Still preclinical; no IND, no candidate nomination. Verify when an IND lands.

UPDATED — Anthropic / Adaptyv Bio · platform, no clinical asset Anthropic reported as saying it will not run human testing or Phase 1 trials for its pipeline programs and will not compete with its pharma customers — company-asserted via trade press, not yet in Anthropic's own voice. Separately, Adaptyv Bio is now also the in vitro validation partner on DeepMind's SynthID Bio work, making it the common validator behind two frontier labs' flagship biology claims.

UPDATED — intismeran autogene (Merck / Moderna) · [Phase 3] Still zero effect sizes disclosed since the August 19 Phase 3 win. ESMO Madrid, October 23–27, remains the hard catalyst. BioNTech/Genentech's colorectal failure is the new comparator (see Ripple 5).

No change this week: rentosertib (Insilico, Phase 3, registry-confirmed per Edition #7); Iambic Therapeutics (Phase 1/1b, S-1 filed September 21); Orbis Medicines / Novo Nordisk (preclinical); PostEra / Merck KGaA (preclinical); GB-0895 golukibart (Generate Biomedicines, Phase 1 COPD); Superluminal Medicines (preclinical, Phase 1 guided to end-2026); insitro (preclinical, MASH candidate clinic-bound); Recursion (Phase 1, REC-4881 TUPELO data due November); Isomorphic Labs (preclinical); EvolutionaryScale / Chai Discovery (platform); SHEO-054 (Sheo Pharmaceuticals, preclinical).

The Signal — Bio/Health Weekly · The Excelsior Group · Archive and all editions: https://excelsiorgroup.ai/insights/signal/bio/