The Signal: Bio/Health — Edition #10 — October 20, 2026
The Signal: Bio/Health — Edition #10 — October 20, 2026
The Excelsior Group Covering October 13 – 19, 2026
The Read
This week the sector's credibility question stopped being about models and became about who is allowed to check the work. STAT put the first sustained, named-source scrutiny on Isomorphic Labs — in talks at $40 billion-plus, no disclosed clinical candidate, a public pledge of human trials by year's end now about ten weeks from expiring, and independent researchers documenting where AlphaFold's predictions actually fail. On the same day, the comment window closed on the FDA's generative-AI device docket, and the hospital lobby's answer amounted to a demand for verifiability: reversibility, output traceability, multiple third-party evaluators, and the developer rather than the hospital carrying post-market responsibility. Meanwhile a Nature audit of the transparency rule we already have found that all 35 certified EHR developers passed certification while only 57% explicitly named the risks they monitor. One pattern runs through all three: this sector produces attestations at scale, and the market keeps pricing them as evidence.
Last week's standing rule was that retrospective review is audit, not oversight. This week's companion: certification is attestation, not verification.
Tide status
No tide movement this week. Both standing tides were reinforced; neither moved.
"Biology is becoming an engineering discipline" — reinforced from the tooling side. The npj Drug Discovery review naming "TechBio 3.0" describes exactly the engineering transition we hold: generative design with toxicity and PK constraints inside the objective function rather than bolted on afterwards, and automated experiments feeding structured data back into the models. But the same review states plainly that as of mid-2026 no AI-designed small molecule has FDA marketing approval. A tide confirmed by its own proponents' admission of what has not happened yet is a tide holding still, not moving. This tide last moved on 2026-09-29, on a registry-confirmed Phase 3 start.
"The binding constraint on new medicines is disease understanding, not molecular design" — reinforced, and from a useful new direction. Schrödinger's CSO makes the argument structurally: machine-learning accuracy degrades the further you move from the training distribution, and genuine discovery happens precisely there — novel scaffolds, new resistance mutations, targets with few known ligands. The closed loop closes where data already exists. That is our thesis restated in an engineer's vocabulary, not a change to it.
Waves
The AI-drug-discovery show-me window — MOVES, and the scrutiny has finally arrived
[deal] · Last week we logged Isomorphic Labs' reported $40 billion-plus funding talks and wrote that the wave was moving on capital rather than evidence. This week the evidence question got asked in public by someone with a byline. STAT's Allison DeAngelis reported that despite partnerships with Eli Lilly and Novartis, Isomorphic has disclosed little about its drug pipeline, and that it is unclear whether the company will meet CEO Demis Hassabis' pledge to begin human trials by year's end — a pledge already pushed back once. The piece also cites independent researchers who have identified concrete limitations in AlphaFold's predictive abilities, including missing protein binding sites and producing infeasible answers. The most quotable defence came from GV's Krishna Yeshwant: "They're really trying for breakthroughs, and that means that it has to be OK to try stuff that doesn't work."
Roadmap implication: Yeshwant's defence is correct and it is also the whole problem. A research organisation permitted to fail is a good thing; a $40 billion private mark that cannot be checked against a disclosed pipeline, a registry entry, or a published failure is a different thing wearing the same clothes. The asymmetry to price: Isomorphic's upside is narrated continuously and its downside is unobservable. Two dates now matter more than any model release — December 31, when the year-end trial pledge either lands or lapses, and ESMO Madrid, October 23–27, the nearest hard catalyst on our tracker. A missed year-end trial start, unaccompanied by a disclosed reason, should be read as information, not noise.
- Does Isomorphic Labs have what it takes to revolutionize drugmaking?
- California Prop 38 explained, Isomorphic AI: Readout Newsletter
The regulatory regime for adaptive/learning systems — MOVES, as the comment window closes
[policy] · The comment period on FDA docket FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, closed on October 19. The discussion paper, issued August 18, proposed a two-axis risk framework. The American Hospital Association filed on the closing day, and its letter is the most legible statement yet of what the buying side actually wants. AHA argues that "activity level" — whether a device merely informs or actually acts — does not by itself predict risk, and asks FDA to add reversibility, downstream safeguards, output traceability, and likelihood of error to the heuristic. It asks for labelling guidance developed jointly with the FTC that discloses the level and type of clinician oversight. It wants multiple third-party evaluators rather than one, and it wants developers, not hospitals, to carry primary responsibility for post-market performance monitoring.
Roadmap implication: read this as a procurement document, not a policy comment. Every item on AHA's list is a feature a vendor will have to ship: traceable outputs, reversible actions, disclosed oversight tier, a post-market monitoring plan the developer owns. For anyone building clinical AI, the commercially relevant deadline is no longer FDA's rulemaking calendar — it is the next hospital RFP, which will start asking these questions well before the agency mandates them. Note also what AHA asked FDA to leave alone: clinical decision support and administrative AI, which it wants explicitly out of scope. That is the boundary the next two years of lobbying will be fought over.
- AHA Provides Feedback to FDA on Regulatory Considerations for Generative AI-Enabled Medical Devices
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices
Clinical-AI autonomy vs. mandated human oversight — MOVES: the disclosure rule we already have is not being followed
[policy] · A brief communication published in npj Digital Medicine on October 14 audited whether EHR developers' publicly posted Risk Management Summaries actually meet the ASTP/ONC HTI-1 Final Rule requirements for AI-enabled predictive decision support interventions. Across the 35 developers that disclosed at least one such tool in their certified products: 91% outlined a risk mitigation strategy, 80% reported a formal governance structure, 77% referenced all eight required attributes at least once, 71% described their overall risk analysis process — and only 57% explicitly identified the individual risks they monitor. Every one of the 35 had nonetheless been certified as meeting the requirement. The authors call for standardised reporting templates and more rigorous certification review, and note the caveat that this may reflect early-stage adoption.
Roadmap implication: this is the cleanest available test of a question the whole governance debate assumes away — does a transparency mandate produce transparency? On this evidence, partially, and unevenly, with certification passing everyone regardless. The practical consequence for any health system: an AI tool's certified status tells you it filed, not that it told you what it monitors, and in 43% of cases it did not. It also sharpens the AHA letter above, which asks FDA for model cards and opposes ONC removing them — the hospitals are defending a disclosure regime whose own compliance data is this soft. Our operating rule, extended: retrospective review is audit, not oversight; and certification is attestation, not verification. Ask vendors for the risk list itself, not the certificate.
- Electronic health record developer adherence to federal AI risk transparency requirements
- Electronic Health Record Developer Adherence to Federal AI Risk Transparency Requirements (preprint)
Abundance of molecules vs. scarcity of validated mechanisms — MOVES, with a name attached
[in-silico] · Two statements landed in the same week and they frame the frontier between them. Insilico Medicine's Alex Zhavoronkov posted that the field approaches a "pharmaceutical singularity" and that "prompt to drug is now possible in low-moderate novelty space," with the real moat being operational efficiency and delivery. Against that, a review by Krish Ramadurai and Abhirup Banerjee in npj Drug Discovery names the current generation "TechBio 3.0" — the closed loop — while stating that no AI-designed small molecule has FDA marketing approval as of mid-2026, and listing applicability-domain checking among the unresolved problems: a model will return a precise-looking score for a molecule it effectively knows nothing about. Schrödinger's CSO Robert Abel supplies the mechanism from the execution side, and one striking operational number: roughly 80% of the computational modelling jobs on Schrödinger's own HPC are now launched by Bunsen, its agentic co-scientist — which does not predict molecular properties itself, but chooses among validated physics methods and runs them after a scientist approves the work plan.
Roadmap implication: note the careful qualifier in Zhavoronkov's own claim — low-to-moderate novelty. Both camps are describing the same boundary from opposite sides: the loop closes where data exists, and first-in-class targets and new modalities are exactly where it does not. The week's deal flow confirms where the smart money is going (see ripples): not to model launches, but to the apparatus that manufactures trustworthy data — Lilly/BigHat, AstraZeneca/Carterra, PacBio/TychoBio. The investable question for the next 18 months is not who has the best model. It is who owns the data factory and the execution layer, and whether anyone attaches a confidence estimate to a prediction before spending money synthesising it.
- Weekly Tech+Bio #87: TechBio 3.0 Is Closing the Loop. Is "Prompt to Drug" Next?
- Interview with Robert Abel, Chief Science Officer at Schrödinger
Ripples
Agilent buys the layer between the slide and the model
[deal] · Agilent agreed on October 13 to acquire Proscia, a vendor-agnostic digital pathology software platform, for $240 million at closing plus up to $30 million in performance milestones. Proscia connects scanners, lab information systems, image analysis applications and AI tools across more than 4,000 labs, hospitals and clinics and 16 of the top 20 pharma companies, has trailing-twelve-month revenue above $18 million, and holds FDA 510(k) clearance for primary diagnosis. Agilent had distributed Proscia's software since 2022 and closed the $950 million Biocare acquisition in June.
So what: an instruments-and-reagents company paying roughly 13x trailing revenue — about 15x with milestones — for a software layer is a statement about where it thinks value is migrating. Agilent already owned the stains and the scanners; what it did not own was the operating layer between the slide and the model. Watch for the same logic to play out in every other modality where an incumbent sells the consumable and a startup owns the data pipe.
ByteDance spins out an AI pharma company at a $1.5bn opening mark
[deal] · ByteDance spun out Anew Labs with a $290 million angel round at a reported ~$1.5 billion valuation (RMB 10 billion). The company has four preclinical assets, led by an oral IL-17A/F inhibitor.
So what: a consumer-internet platform company spinning a drug-discovery subsidiary out at a $1.5bn mark on four preclinical assets is the same pricing phenomenon as Isomorphic, in a different jurisdiction and without the AlphaFold pedigree. It is also a reminder that the compute-rich incumbents entering this field are no longer only American. Pair this with China's proposed restrictions on foreign acquisition of biopharma platform technologies — announced the same week, and explicitly exempting licensing deals, which is where the $120 billion of Chinese drug licensing this year actually sits.
Lilly ships a model trained on a startup's wet-lab output
[wet-lab] · Eli Lilly's TuneLab released NanoLab, a nanobody developability model trained on 2,500+ VHH antibodies characterised by BigHat Biosciences across six one-week design rounds. In the same week, AstraZeneca and Carterra announced they will co-build hardware and software to produce standardised binding data for lab-in-the-loop biologics discovery, and PacBio and TychoBio will sequence 10,000+ samples to train models predicting how antisense therapies behave across the transcriptome.
So what: three deals, no model launches — all three are about manufacturing data a model can trust. Note what BigHat got out of it: its wet-lab characterisation work is now embedded in a major pharma's public tooling, which is a different and more durable commercial position than being a discovery vendor. For anyone evaluating platform companies, "whose data trains the incumbent's model" is becoming a more useful question than "whose model is best."
Partial reprogramming reports its first human eyes
[Phase 1] · Life Biosciences reported early Phase 1 data for its three-factor epigenetic reprogramming therapy in glaucoma: no treatment-related side effects to date, and visual-field gains in two of three patients at eight weeks. It is a conference presentation in three people.
So what: logged at the correct rung and no higher. Three patients, eight weeks, a conference podium, and an endpoint — visual field — that is subject to learning effects and high test-retest variability. This is the first human safety signal for a Nobel-adjacent idea that has carried a decade of mouse data, and it deserves to be on the tracker; it does not yet deserve a conclusion. The number that would change our view: a controlled trial, more than thirty patients, with structural as well as functional endpoints.
A sequencing company becomes part of a cancer vaccine's manufacturing chain
[Phase 3] · Tempus expanded its agreement with Moderna and Merck to run sample logistics and the sequencing used to design each patient's personalised cancer vaccine, V940 (intismeran autogene), which is in Phase 3 in resected high-risk melanoma.
So what: this is what individualised therapy looks like when it stops being a science story and becomes an operations story. A personalised vaccine is not manufactured once — it is manufactured per patient, and the sequencing step is inside the production line, not upstream of it. The economics of this class of therapy will be decided by turnaround time and cost per patient at that step, not by the immunology. Watch for the first disclosed vein-to-vein time.
An independent evaluator argues administrative AI is making billing more expensive
[policy] · The Peterson Health Technology Institute's report Driving Adoption of High-Value AI in Healthcare carries a "last updated October 15, 2026" date on its page. Its core argument: under current incentives, administrative AI tends to raise the cost of billing rather than reduce waste, while high-value clinical AI is held back by payment, purchasing and policy barriers. PHTI proposes outcome-based payment plus a four-domain adoption roadmap.
So what: flagged with a caveat — we could not establish whether October 15 is first publication or an update, and no press coverage pins it, so treat the date as soft. The finding itself is the sharpest version of a thesis we have been building for three editions: the revenue-cycle AI boom is an arms race between payers and providers in which both sides spend more and neither side's patients benefit. Administrative AI is the easiest clinical-AI budget to approve and the hardest to show a return on. If PHTI is right, the correction comes when a CFO runs the numbers on their own billing-cost line.
Pipeline watch
Isomorphic Labs — UPDATED, no phase change, and the clock is now visible. Still preclinical, still no disclosed candidate, still no clinicaltrials.gov entry. New this week: the year-end human-trials pledge is now the subject of on-record press scrutiny, and independent researchers' documented AlphaFold limitations are in print alongside it. The valuation moved last week; the evidence has still not.
BigHat Biosciences — UPDATED. BHB810 (CDH17-directed VHH-Fc ADC) remains Phase 1 in gastric/GEJ cancer. Platform-side, BigHat's VHH characterisation data now trains Eli Lilly's publicly released NanoLab developability model — a commercial position worth tracking separately from the clinical asset.
intismeran autogene (V940, Merck/Moderna) — UPDATED, no phase change. Phase 3 in resected high-risk melanoma. Tempus is now inside the manufacturing chain for per-patient vaccine design.
rentosertib (ISM001-055, Insilico Medicine) — no change, context added. Phase 3 in IPF, registry-confirmed (NCT07687459). The npj Drug Discovery review independently restates the Phase IIa result — mean +98.4 mL forced vital capacity at the top dose in 71 patients — and calls it the field's most advanced proof point. Still the most advanced AI-originated asset anywhere.
Iambic Therapeutics — UPDATED. IAM1363 (HER2) and portfolio remain Phase 1/1b. Filed for a Nasdaq IPO; cited in the npj review as a working example of weekly design-make-test cycles. A public filing will, for the first time in this cohort, force disclosure of what an AI-native discovery platform actually spends and produces.
Unchanged this week: Orbis/Novo, PostEra/Merck KGaA, Generate Biomedicines, Superluminal, insitro, Recursion, Xaira, EvolutionaryScale/Chai, Anthropic/Adaptyv, Sheo.
Scoreboard
Zero full FDA approvals of an AI-discovered drug — now confirmed not by our own count but by a peer-reviewed review of the field, which states that no AI-designed small molecule has FDA marketing approval as of mid-2026. Also still zero: AI-originated assets with disclosed Phase 2 efficacy effect sizes; protein-design pharma partnerships with disclosed economics. One AI-originated asset has a registry-confirmed Phase 3 start (rentosertib, NCT07687459). New this week: 43% of certified EHR developers do not explicitly identify the AI risks they monitor, and all of them are certified anyway.
The Signal: Bio/Health is published weekly by The Excelsior Group. Archive and past editions: https://excelsiorgroup.ai/insights/signal/bio/
