The Signal: Bio/Health — Edition #7 — September 29, 2026

The Excelsior Group · Covering September 22–28, 2026 · https://excelsiorgroup.ai/insights/signal/bio/

The Read

Two things happened this week that belong together. Anthropic announced that its AI had found a previously uncharacterised enzyme system in phage DNA with structural echoes of CRISPR, and within five days the people best qualified to judge it said, in public, that the hard part has not started — nobody knows what the thing does. That is the evidence ladder working in real time, and it is the right reaction. The second thing is our own. For three consecutive editions we refused to call Insilico's Phase 3 confirmed because ClinicalTrials.gov had not moved, and last week we escalated that refusal to tide level. The registry had in fact moved five days before we printed that. NCT07687459 now reads RECRUITING with an actual start date of September 9 2026 — the exact day Insilico said it dosed a patient — and the update was posted on September 17. Edition #6 was wrong, the escalation was built on a stale read, and we are correcting it at the top rather than burying it. The net of the two: the first generative-AI-originated medicine in a registrational trial is now registry-confirmed, and the newest AI-discovered biology is not yet anything at all. Both facts arrive the same week, and the discipline that gets them right is the same discipline.

Correction to Edition #6

Edition #6 (September 22) stated that NCT07687459 "still reads NOT_YET_RECRUITING, with an estimated start date of August 30 2026, a last update posted July 7 2026," and escalated the resulting gap to tide level as a three-week failure of confirmation. That was incorrect at the time of publication. The registry record shows a last update submitted September 14 and posted September 17 2026 — five days before Edition #6 went out — changing overall status to RECRUITING and the start date to an ACTUAL date of September 9 2026. The error was ours: a stale read, not a stale registry. Editions #4 and #5 were accurate when written; Edition #6 was not. The escalation is withdrawn and replaced by the tide entry below. Logged to bio/correction-log.md.

Tide status

One tide moves. The other holds and is illustrated rather than widened.

Biology is becoming an engineering discipline — MOVES. The top rung is now registry-confirmed. [Phase 3]

As of this morning, ClinicalTrials.gov NCT07687459 — GENESIS-IPF-3, Insilico's Phase 3 of rentosertib (INS018_055) in idiopathic pulmonary fibrosis — reads overall status RECRUITING, with a study start date of September 9 2026, type ACTUAL, last update submitted September 14 and posted September 17, enrollment 320 estimated, and 47 listed sites, of which 12 are recruiting and 35 are not yet recruiting. The 47 matches Insilico's press release exactly; the 48 we reported in Editions #4 through #6 was our own miscount. The status verified date remains June 2026, which is the one field that has not caught up.

State the consequence plainly, because we spent three editions declining to: a medicine whose target was selected by AI and whose molecule was designed by AI has a registry-confirmed actual start date in a randomised, double-blind, placebo-controlled, 52-week Phase 3 with a hard primary endpoint — annual rate of FVC decline. That is the highest rung any AI-originated asset has occupied, and it is no longer a claim resting on a company press release. Three qualifications keep it honest. Recruiting is not read out; the estimated primary completion is October 2029. Twelve of 47 sites are open, so this is a trial beginning, not a trial running. And rentosertib's own Phase 2a was a 71-patient, 12-week study — the distance between that and a 52-week registrational FVC endpoint is where the great majority of IPF programmes have died, AI-originated or not.

• ClinicalTrials.gov — NCT07687459 (GENESIS-IPF-3)
• Insilico — first patient dosed, GENESIS-IPF-3 (Sept 9 2026)

The binding constraint is disease understanding, not molecular design — holds, unwidened, and illustrated exactly [wet-lab]

No new failure mode this week, which after two consecutive widenings is worth saying out loud. What the week supplied instead is the cleanest single illustration of the tide we have had: Anthropic's agents surveyed more than 200,000 reverse transcriptases, narrowed to roughly 3,500 candidate systems, produced 20 written-up reports, and found one genuinely anomalous object — and the honest description of that object is that nobody knows what it does. Abundance at the candidate-generation stage; scarcity at the what-does-this-mechanism-actually-do stage. The tide predicts precisely this shape, and the discovery is the shape.

Waves

1. The AI-drug-discovery show-me window — MOVED. The category answered last week's exit question with an S-1. [deal]

Edition #6 ended on a blunt question for platform companies: if the category cannot raise, what is the exit? We suggested the realistic answer might be acquisition by a frontier lab. The week answered differently. Iambic Therapeutics filed to go public on Nasdaq under "IAM" — the S-1 carries a September 21 2026 filing date and the trade coverage landed September 23–24 — with J.P. Morgan, Jefferies, BofA Securities and Citigroup leading. Share count and price are not set.

What is being sold matters more than the fact of the filing. Proceeds go mainly to IAM1363, an oral, brain-penetrant HER2 inhibitor already in Phase 1/1b in HER2-altered solid tumours (NCT06253871), which the company says is the only known HER2 TKI binding the inactive DFG-out conformation, with a registrational trial targeted as early as 2027. Two further programmes — IAM217 (KIF18A) and IAM-C1 (CDK2/4) — are guided into the clinic. The platform behind them is two named models, Enchant (a multimodal transformer predicting preclinical and clinical endpoints) and NeuralPLexer (protein–ligand structure prediction fusing physics with ML). The company has raised roughly $461.8 million since 2019 with NVIDIA and the Qatar Investment Authority among its backers, and has partnerships with AbbVie, Takeda (headline value above $1.7 billion), Revolution Medicines and Lundbeck. It lands into a genuinely reopened window: Aktis Oncology ($318M, January), Eikon Therapeutics ($381M, February), Generate:Biomedicines ($400M) and Kardigan ($400M upsized, June) have all priced this year.

Now the self-correction. For two editions we have used the Cure AI drug-discovery venture funding tracker as a measurement, reporting a lengthening gap since Accipiter Bio's $10.5M seed extension on July 31. That entry is still the most recent, which would make it nine weeks — but the tracker itself has not been updated since September 4. A tracker that is five weeks stale cannot measure a nine-week drought; at this point the instrument is the more likely explanation than the phenomenon, and we are retiring it as a headline measurement rather than repeating it a third time. What we can say without it: the largest health-AI round logged in Fierce's tracker inside this window was Blair Health's CAD $4.24 million on September 22, led by BDC's Thrive Venture Fund. Against last week's $600M, $250M and $150M, that is a quiet week by any instrument.

Roadmap implication: the exit question has an answer and it is better than the one we guessed — the public market, at a stage where the lead asset is a Phase 1 molecule with a named binding mode rather than a platform narrative. That is the diligence template for the next filings in the queue: what molecule, what rung, what mechanism claim can a chemist check. A platform without a clinical asset is not what cleared this bar.

• SEC — Iambic Therapeutics Form S-1
• Bioxconomy — AI drug hunter Iambic files for Nasdaq IPO
• Cure — AI Drug Discovery Venture Funding Tracker 2026

2. Generative design crossing into wet-lab function — MOVED. An AI found a new system; the field said finding it was the easy part. [wet-lab]

On September 23 Anthropic disclosed the first result from the life-sciences group and Bay Area wet lab it stood up in spring 2026. Given one high-level instruction to look for interesting reverse transcriptases, roughly 950 Claude agents spent 21 hours and 210 million tokens, gathered more than 200,000 RTs, narrowed to about 3,500 candidate systems and wrote up the 20 most compelling. One agent, reading raw sequence, flagged a tandem repeat array sitting beside an unusual reverse transcriptase gene in a jumbo bacteriophage. Anthropic named it ART — array-associated reverse transcriptases — and reports that first experiments show the array is expressed as a set of distinct short RNAs, which is the property that makes CRISPR programmable. Feng Zhang reviewed the preprint and said it merits further investigation. All laboratory work was done by human scientists at BSL-1 and BSL-2, with no human pathogens.

Place it on the ladder carefully, because the headlines did not. The computational half is a search result. The wet-lab half is expression in laboratory strains plus biochemical and structural characterisation — real, and enough for [wet-lab], which is two rungs below anything preclinical and six below anything a patient sees. And the central fact, which Anthropic states itself, is that the function is unknown. Lucas Harrington — co-founder of Preventive and Mammoth Biosciences, PhD from Jennifer Doudna's lab — put it the way we would have: "Finding a weird cluster of genes and repeats is often the easy part," and the real discoveries come from working out what the system actually does. STAT's Brittany Trang added the commercial read, that this is hard to separate from marketing Claude's scientific capabilities ahead of a long-awaited IPO, and asked the question that ought to be asked of every result of this kind: who corrects the feel-good headlines if the lab work shows it is nothing. Dario Amodei has also noted that a Stanford group previously found a system similar in some ways.

None of that makes the work uninteresting. The number that generalises is 20 reports out of 200,000 candidates — the scarce input in this kind of discovery has always been expert attention, and what Anthropic industrialised is not the biology but the triage. The daily Signal carried this on September 23 as a second data point in its AI crosses into original biology wave, alongside the Arc/Stanford Evo phage-genome work from August; our contribution here is the rung and the pushback, not the news.

Roadmap implication: for any AI-discovery claim crossing your desk in the next quarter, the question is not whether the model found something novel. Novelty is now cheap and will get cheaper. The question is who is funded to determine function, on what timeline, and what happens to the valuation if the answer is "nothing." That is a wet-lab throughput and attention problem, and it is the bottleneck the tide above describes.

• Anthropic — Claude discovers a novel enzyme system
• TechCrunch — Anthropic says its biology lab has already found something big
• STAT Morning Rounds — Breaking down Anthropic's big biology 'discovery'

3. Frontier AI labs moving into the clinical and discovery layer — MOVED. The labs stopped being vendors this quarter. [deal]

Three facts now sit together. Anthropic has its own wet lab and its own scientists running BSL-1/2 experiments. Anthropic signed a collaboration with Novo Nordisk on September 16 to apply Claude across drug discovery and R&D. And the enterprise layer keeps thickening underneath: Bristol Myers Squibb put Claude in front of more than 30,000 employees earlier this year with a stated goal of halving the time from target selection to lead molecule; Merck has committed up to $1 billion to Google Cloud for Gemini Enterprise across R&D, manufacturing and commercial; Amgen deployed ChatGPT Enterprise. Set that beside the corporate-development lead Anthropic was hiring for in September to run acquihires and tuck-ins across drug discovery, which we logged in Edition #6.

The useful frame, which Andrii Buvailo made this week and we will borrow with attribution, is that these deals carry two value propositions needing two different measurements. The operational one — regulatory and medical writing, knowledge retrieval, software engineering, planning — has real if mixed evidence from outside pharma and is measurable in cycle time. The scientific-reasoning one is much harder to measure, and a new Cell benchmark assessing frontier models and smaller specialist models across 17 aging-biology tasks found that model size alone does not decide performance on biological data. That finding is the whole argument for the grounded specialist layer — Schrödinger, OWKIN, Causaly, BenchSci, Insilico, Lantern — built on physics simulation or curated knowledge graphs rather than scale.

Roadmap implication: when a pharma signs a frontier-lab enterprise deal, ask which of the two propositions is being bought and what the success metric is. If the answer is "scientific reasoning" with no named endpoint, that is a procurement without a checkpoint. And note the structural point for anyone holding techbio equity: a frontier lab with its own bench, its own corp-dev function and direct pharma contracts is no longer only a supplier to this category.

• Where Tech Meets Bio #85 — Big Pharma Is Betting On LLMs and Agents. What Would Count As Success?
• Endpoints — Novo signs Anthropic deal to test Claude Science in R&D

4. The regulatory regime for adaptive and learning systems — minor update. The in-silico qualification path took one more step; the device counter stayed dark. [policy]

Two small movements and one large absence. Absentia Labs says FDA accepted its AI model for predicting drug-induced cardiotoxicity into the ISTAND pathway, which qualifies novel drug-development tools that do not fit existing categories. This would be the company's second organ model in the programme after its Digital Liver, which was the first AI drug-development tool accepted into an FDA qualification programme. Flag the sourcing honestly: the cardiotoxicity acceptance reaches us as a statement from the company's CEO reported in trade press, with the retrievable primary release covering the earlier liver model. Treat it as company-asserted until FDA's own ISTAND listing reflects it — the same standard we just spent three editions applying to Insilico, and the same standard we owe here.

That matters because it is the practical follow-through on the September 21 direct final rule that replaced "animal tests" with "nonclinical tests" across FDA's regulations and named computer models among the acceptable methods. The rule removed a definitional obstacle; ISTAND is where a specific model actually earns standing. The standing action item is unchanged and the clock is now three weeks: comment on docket FDA-2026-N-7874 by October 19 2026.

The absence: FDA's AI-Enabled Medical Device List page still carries a modification date of 09/04/2026 and the most recent Date of Final Decision in the table is still June 29 2026. That is thirteen weeks. We have carried this as a negative finding for four editions and declined each time to say whether it is an authorisation slowdown or a list-maintenance lag, because from outside we cannot separate them. At thirteen weeks the distinction stops mattering for planning purposes: if you are modelling a clearance date for an AI-enabled device, you have had no public evidence of the queue moving for a quarter.

Nothing in FDA's press announcements during the window touched AI, digital health or software — the week produced an MCT8 deficiency approval, a PMTA framework notice and a consent decree.

• FDA — Artificial Intelligence-Enabled Medical Devices
• FDA — Updates Regulations to Advance Innovative Alternatives to Animal Testing
• BioSpace — Absentia Labs' Digital Liver Model Becomes First AI Drug Development Tool Accepted Into FDA Qualification Program

Ripples

Novo Nordisk will pay Orbis up to $1.4B for AI-designed oral macrocycles — and we missed it last week [deal]

Announced September 17, so it belongs to Edition #6's window and was not in it. Copenhagen-based Orbis Medicines signed a multi-target discovery and licence deal with Novo Nordisk worth up to $1.4 billion in upfront and milestone payments plus tiered royalties, with Novo also taking an undisclosed equity stake. The work uses Orbis's nGen platform, which pairs generative AI with automated chemistry to produce orally bioavailable macrocycles against cardiometabolic targets — the structural class that normally has to be injected.

So what: this is the second disclosed price in AI-designed chemistry in a month, after PostEra/Merck KGaA's "mid-double-digit millions" for two fertility programmes, and it is roughly an order of magnitude larger. Read the two together and the pricing signal is clearer than any venture round: pharma will pay platform-scale money when the AI is pointed at a formulation-class problem — making an injectable modality oral — rather than at target discovery. That is a narrower and more defensible claim than "AI finds better drugs," and it is where the money actually went. Recording the miss as well as the deal: our sweep last week caught the funding drought and not the largest AI-chemistry licence of the quarter, which is a lesson about weighting trade press over trackers.

• Businesswire — Orbis Medicines Enters Multi-Target Drug Discovery Partnership with Novo Nordisk Worth up to USD 1.4 Billion
• BioPharma Dive — Novo partners with Orbis, deepening investment in oral peptide drugs

Mayo Clinic and Thermo Fisher are building a million-biospecimen pre-symptomatic atlas [wet-lab]

The two launched Precure, LLC, a Mayo-majority joint venture generating multi-omics data from one million biospecimens linked to longitudinal clinical records, using Thermo Fisher's Olink Explore HT proteomics and Orbitrap mass spectrometry. The stated aim is detecting molecular disease signals years before diagnosis. No deal value disclosed.

So what: file this with the AlphaGenome atlas and the fly connectome as the same species of artifact — measurement infrastructure, not inference. The distinguishing feature here is the linkage: a million specimens tied to longitudinal outcomes is a dataset built for supervised learning on disease trajectory, which is precisely the disease→mechanism gap our second tide says is binding. If the tide is ever going to move, it moves because of assets like this rather than because of a better model. It is also a reminder that the scarce resource in this decade is annotated human biology, and that the institutions holding it — Mayo here, Tempus with its 100,000-then-million whole-genome programme — are positioning as infrastructure rather than as customers.

• Where Tech Meets Bio #85

Generate's GB-0895 data existed for a month while we recorded it as unretrievable [Phase 1]

We carried "GB-0895 ERS 2026 COPD data unretrieved" as an open item for three consecutive editions with confidence downgraded. It was retrievable the whole time, in trade press we did not search. The Phase 1 COPD data were accidentally released early on around August 25 and reported August 27, ahead of their scheduled September 7 presentation at the European Respiratory Society Congress. The study: placebo-controlled, ascending-dose, 40 patients. A single dose produced rapid and sustained reductions across four COPD biomarkers; in patients with elevated blood eosinophils the company describes broad and durable anti-inflammatory activity lasting at least six months; half-life near 100 days. Six serious treatment-emergent adverse events occurred in five patients, none Grade 3 or above and none deemed drug-related. A companion asthma poster supported twice-yearly dosing at 300mg. The stock fell from an all-time high of $20.38 on August 25 to $16.57 by that Thursday.

So what: two separate readings. On the asset — a ~100-day half-life supporting twice-yearly dosing is the commercially interesting number, more than the biomarker movement, because dosing interval is what differentiates an anti-TSLP entrant against an established competitor. On our process — a three-edition "unretrieved" flag caused by searching only the company's own media centre is exactly the failure mode this newsletter exists to catch in other people, and the honest response is to say so and change the retrieval rule. The item is now closed.

• BioSpace — Generate Bio drops as data accidentally leaked ahead of European conference

Ashish Jha: AI is closing the expertise gap that kept biological weapons rare [policy]

In STAT on September 28, the former White House Covid response coordinator argued that biological weapons stayed rare largely because building them required expertise that is hard to acquire and harder to combine, and that AI erodes exactly that barrier by making someone trained in one field competent across several. It ran alongside a piece marking 25 years since the 2001 anthrax letters, whose author — Daniel Jernigan, who worked the CDC investigation — notes that detection technology is substantially better now while the response infrastructure has been weakened by cuts to civilian and military biodefence programmes.

So what: this is the governance counterpart to the ART discovery in the same week, and it is why Anthropic's BSL-1/2-and-no-human-pathogens framing is in its announcement rather than being an afterthought. The operative asymmetry for anyone building in this space: existing biosecurity controls screen synthesis orders against known hazards, which is a lookup problem, while the value of a generative or search-based system is producing sequences that are not in the lookup table. Better detection paired with weaker response is not a stable equilibrium, and a company that industrialises novel-sequence discovery is going to be asked about it on a shorter timeline than it expects.

The week's health-AI funding was a rounding error [deal]

The largest round logged in Fierce Healthcare's tracker inside the window was Blair Health's CAD $4.24 million on September 22, led by the Business Development Bank of Canada's Thrive Venture Fund, for clinical infrastructure that lets generalist physicians deliver specialty care through AI-encoded protocols.

So what: a single quiet week is noise, not signal, and we are logging it as such rather than building a thesis on it — a discipline we should have applied to the funding tracker two editions ago. What is worth carrying forward is the shape rather than the volume: even the small round this week is care-delivery infrastructure sitting on existing payment rails, which is the same destination as last week's $600M, $250M and $150M. Discovery still is not where the venture money goes; the difference this week is that the category found a different door, and it was the Nasdaq.

• Fierce Healthcare Fundraising Tracker '26

Owl Posting asks whether organoids are useful for Alzheimer's research [preclinical]

Published September 21, one day outside the window, and noted rather than summarised because it is the best available treatment of a question this newsletter will need shortly: whether the model systems being proposed as the validation layer for AI-generated hypotheses can actually carry the weight. That question is about to become commercial — Polyphron raised a $20 million seed this month to pair an automated human micro-tissue "Tissue Foundry" with a simulation layer explicitly for testing what AI-generated therapeutic hypotheses do in human biology, with a former DeepMind Gemini researcher as co-founder and chief AI scientist.

So what: if the bottleneck is function rather than candidates — which is what both tides now say — then the companies selling validation throughput are positioned on the correct side of the constraint, and they are a category with almost no capital in it relative to the design side. Worth watching whether that asymmetry persists.

• Owl Posting — Are organoids useful for Alzheimer's research?

Pipeline watch

• Insilico Medicine · rentosertib (ISM001-055), IPF · Phase 3 · CONFIRMED — escalation withdrawn. NCT07687459 now RECRUITING, start date September 9 2026 ACTUAL, last update posted September 17, 320 estimated enrollment, 47 sites (12 recruiting, 35 not yet), status verified date still June 2026. Highest rung occupied by any AI-originated asset, and now registry-confirmed. Next hard checkpoint is enrollment completion; estimated primary completion October 2029.
• Iambic Therapeutics · IAM1363 (HER2), IAM217 (KIF18A), IAM-C1 (CDK2/4) · Phase 1/1b · ADDED. S-1 filed September 21 for a Nasdaq listing under "IAM." IAM1363 in Phase 1/1b (NCT06253871) in HER2-altered solid tumours; registrational trial targeted as early as 2027. Platform: Enchant and NeuralPLexer. $461.8M raised since 2019; NVIDIA and QIA among investors. Verify the registrational start against the registry when guided.
• Orbis Medicines / Novo Nordisk · undisclosed oral macrocycles, cardiometabolic · preclinical · ADDED. Up to $1.4B in upfront and milestones plus tiered royalties and an undisclosed equity stake, announced September 17. nGen platform: generative AI plus automated chemistry. No named asset or registry entry; verify when one appears.
• Generate Biomedicines · GB-0895 (golukibart) · Phase 1 · RESOLVED, three-edition flag closed. Phase 1 COPD data (40 patients, placebo-controlled ascending dose; four biomarkers reduced; ≥6-month durability in the elevated-eosinophil group; ~100-day half-life; 6 serious TEAEs in 5 patients, none Grade 3+ or drug-related) released accidentally around August 25, reported August 27, ERS presentation September 7. Confidence restored. Phase 3 in severe asthma proceeding.
• 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 — now under four weeks out, and the single hardest catalyst on this tracker.
• Isomorphic Labs · first candidates · preclinical · no change. Nothing new since Chris Butler's September 15 comments. Still no named clinical candidate, still no timeline.
• Anthropic · ART enzyme system + Adaptyv minibinder platform · platform, no clinical asset · UPDATED. ART disclosed September 23: wet-lab expression and structural characterisation done, function unknown, preprint reviewed by Feng Zhang. Anthropic now operates its own BSL-1/2 lab with human scientists. Verify function against any peer-reviewed publication; this stays at [wet-lab] until one exists.
• Sheo Pharmaceuticals · SHEO-054 · preclinical · no change. Still conference-talk-only; no paper, no structure.
• PostEra / Merck KGaA, Superluminal, insitro, Recursion, Xaira, EvolutionaryScale / Chai · no change. Nothing in window.

Scoreboard: zero full FDA approvals of an AI-discovered drug. Unchanged since this tracker opened. The sub-line changes for the first time: one AI-designed drug is registry-confirmed as recruiting in a Phase 3 registrational trial, actual start September 9 2026 — no longer company-asserted, and no longer carrying our caveat.

The Signal: Bio/Health is published weekly by The Excelsior Group. Archive and back issues: https://excelsiorgroup.ai/insights/signal/bio/