Life Sciences / Regulatory Brief 🧬

📌 Navigate

📊 Exec Summary

This week the clinical-AI conversation moved from "what does the model score" to "who verified it, against what reference, under what monitoring," and the same question showed up in a journal framework, a preprint benchmark, a vendor access policy, and a foundation-model license.

Five things moved in regulatory pathways, life-sciences infrastructure, and AI-hybrid execution this week:

• NEJM AI proposed a phase-based evidence framework. It separates benchmark accuracy and retrospective validation from calibration, transportability, prospective clinical impact, implementation readiness, and postdeployment safety, and argues for evidence standards rather than reporting standards.
• Zero-shot LLMs beat rule-based extraction on colorectal notes. Gemini 3.5 Flash hit Macro F1 0.70 against a 0.44 rule-based baseline, and then post-hoc negation filtering degraded it to 0.58.
• Anthropic gated life-sciences model access behind verification. The Life Sciences Verification Program ties access to research credentials, security standards, and declared use cases, moves enforcement to offline monitoring with 30-day traffic retention, and excludes BAA-enabled organizations.
• Tempus licensed Recursion's TxFM RNA foundation model. A $12 million model license, alongside a data extension through November 2029 that swaps up to $84 million in discretionary fees for $42 million committed.
• Biomedical AI widened from prediction into curation and design. BELL scored 210 GBM protein-protein interactions, clearing 49.5% at HIGH confidence, while six sibling preprints pushed into RNA design, living materials, and spatial target discovery.

The pattern: Journals specify what to prove, preprints expose the reference standard, model vendors verify who is asking, and licensors price the model instead of the data. Evidence is becoming a supply chain with named checkpoints.

1️⃣ NEJM AI puts a phase gate between benchmark performance and clinical readiness

TL;DR: NEJM AI published a phase-based framework for evidence standards in clinical AI that treats technical performance and clinical readiness as different states with different proof obligations.

What happened

• On 17 September 2026, NEJM AI published "From Technical Performance to Clinical Readiness: A Phase-Based Framework for Evidence Standards in Clinical Artificial Intelligence" (DOI 10.1056/AIp2600684), ahead of print.
• The argument: clinical AI evidence standards remain inconsistent, with benchmark accuracy and retrospective validation routinely reported while calibration, transportability, prospective clinical impact, implementation readiness, and postdeployment safety are assessed far less often.
• Pharma Practice Notes read it as an evidence staircase between two worlds, and flagged the operator-relevant distinction: most existing frameworks define reporting standards (what to disclose), while this one aims at evidence standards (what to prove).
• Three adjacent signals landed in the same window. A medRxiv study across 77 global languages argued that word error rate does not track consequential transcription errors in ambient AI scribes and proposed clinical risk as the evaluation standard.
• A second medRxiv paper mapped a decade of AI-generated synthetic data in biomedicine and reported a persistent translation gap, and Bio-IT World covered AI-assisted drug repurposing with Every Cure founder David Fajgenbaum making the same point: prediction is the cheap part, patient impact is the test.

📊 Key facts (from NEJM AI)

🔗 Primary source → From Technical Performance to Clinical Readiness: A Phase-Based Framework for Evidence Standards in Clinical Artificial Intelligence

🔍 The non-obvious point

A framework in a journal is not a pathway, and the useful move is to read this as a budget document rather than a regulatory one.

• Publication is not adoption. FDA's SaMD approach and PMDA's device-AI guidance are on their own tracks, and nothing here changes a submission requirement this quarter. What it does change is the cost line you defend internally: calibration studies, cross-site transportability, and a postdeployment safety plan are now the named deliverables a reviewer will ask about.
• Reporting versus proving is the whole delta. A reporting standard is satisfied by a table in an appendix. An evidence standard is satisfied by a study you have to run. Teams that budgeted for the first and get asked for the second discover it late, after the model is frozen.
• Phase framing carries a linear-progression risk. Deployed clinical AI is iterative: models update, populations drift, workflows change. Whether the framework handles postdeployment iteration or treats deployment as an endpoint is the thing to read for.
• The ambient scribe paper is the concrete version. If word error rate does not track clinical consequence, then every scribe and transcription acceptance criterion written against WER is measuring the wrong thing, and that is a specification defect, not a model defect.

👀 What to watch

• Whether phase or evidence-tier language starts appearing in FDA SaMD and PCCP materials, or in PMDA device-AI guidance, over the next two review cycles.
• Whether any vendor publishes a transportability study structured to the framework's stages, which would be the first sign it is being used as a planning skeleton rather than a citation.

2️⃣ Zero-shot LLMs beat rule-based note extraction, then negation filters break them

TL;DR: On 2,704 colorectal discharge notes, zero-shot LLMs outperformed rule-based and clinical NER baselines, and deterministic negation filtering made both LLMs worse.

What happened

• A medRxiv preprint benchmarked dictionary rules, clinical NER, zero-shot LLMs, and hybrid variants for colorectal cancer symptom extraction from MIMIC-IV v2.2 discharge notes accessed under a PhysioNet data-use agreement.
• Scope: 2,704 CRC discharge notes, 46 target symptoms drawn from the Memorial Symptom Assessment Scale and EORTC QLQ-CR29.
• Reference standard: 200 adjudicated notes, two raters, pooled kappa 0.71 and macro kappa 0.49.
• Gemini 3.5 Flash led at Macro F1 0.70 (Micro F1 0.86, Macro Precision 0.74). Claude Haiku reached Macro F1 0.63 (Macro Recall 0.71). Rule-based scored 0.44, clinical NER 0.38.
• Adding fixed-window post-hoc negation filtering dropped the hybrids to 0.58 (Gemini) and 0.54 (Claude).
• A companion preprint posted 16 September used the same MIMIC-IV corpus to derive LLM-based symptom clusters and relate them to patient outcomes, on the argument that routine documentation is continuous where patient-reported outcome surveys are episodic.

📊 Key facts (from the medRxiv preprint)

"Post-hoc negation filtering paradoxically degraded LLM performance"

🔗 Primary source → Leveraging Large Language Models for Colorectal Cancer Symptom Extraction from MIMIC-IV Clinical Notes

Companion study: Large Language Model-derived Symptom Clusters and Patient Outcomes in Colorectal Cancer from MIMIC-IV Clinical Notes

🔍 The non-obvious point

The headline is the 0.70, but the number that should change how you write an acceptance specification is the 0.49 macro kappa on the reference standard.

• The ceiling is human. Two trained raters agreed at macro kappa 0.49 across 46 symptoms. A model scored against that reference cannot be cleanly separated from reference noise, which means any validation protocol for clinical-note extraction needs inter-rater agreement stated in the spec, not just the model metric. Specs that name only a model threshold are unfalsifiable.
• Deterministic guardrails are a change that needs validating. The instinct to wrap an LLM in a rule-based negation filter is exactly what failed here: fixed-window syntactic scope overrode correct predictions. Post-processing is not a free safety layer. It is a modification with its own performance profile, and under a change-control regime it is a change.
• The absences are the same absences as item 1. No prospective deployment, no monitoring or escalation design, no institution-specific calibration, and no evidence of generalizing past the MIMIC-IV note distribution. That is a technical-performance result, cleanly, and the framework in item 1 is the map of what it does not yet cover.

👀 What to watch

• Whether either preprint adds external validation on a non-MIMIC note corpus, which is the transportability step that turns this from an NLP result into a deployable workflow.
• Whether the companion symptom-cluster work holds when clusters are tied to outcomes prospectively, since that is the point at which an extraction pipeline starts making a clinical claim.

3️⃣ Anthropic gates life-sciences model access behind verification and monitoring

TL;DR: Anthropic opened a Life Sciences Verification Program that ties model access to verified credentials and declared use cases, and replaces real-time blocking with offline monitoring backed by 30-day traffic retention.

What happened

• On 18 September 2026, Anthropic launched the Life Sciences Verification Program (LSVP) in beta. Endpoints News reported it as the reopening of biology use for Mythos after Anthropic had previously restricted life-sciences work on the model.
• Qualification runs through a verification process covering research credentials, security standards, and ethical research oversight, plus a declaration of intended use cases.
• Two grant tiers: Standard Use, a team-wide grant covering most life-science work with annual renewal, and High-risk Use, a project-specific add-on with additional vetting and six-month renewal.
• Enforcement shifts from real-time blocking to offline monitoring against declared use cases, and LSVP traffic carries a 30-day retention requirement so flagged activity can be reviewed.
• Scope limits: the beta covers teams and institutions on Anthropic surfaces across Mythos 5.1, Opus 5, and Sonnet 5 under Enterprise Frontier Safeguards. It does not cover individual plans or third-party platforms, and BAA-enabled organizations are not supported.
• Anthropic onboarded dozens of early-access organizations and projected hundreds of organizations in the first week of the beta.
• Separately, Covington Digital Health reported that three California AI employment bills passed the legislature and sit with Governor Newsom, who has until 30 September to sign or veto.

📊 Key facts (from Anthropic)

"To qualify for these grants, each applicant goes through a verification process that includes a review of their research credentials, security standards, and ethical research oversight."

🔗 Primary source → Introducing the Life Sciences Verification Program

🔍 The non-obvious point

The 30-day retention requirement and the BAA exclusion are the same fact viewed from two directions, and together they draw a line straight through a life-sciences company.

• Monitoring needs the traffic, and PHI workflows cannot hand it over. Offline enforcement only works if the prompts and completions persist long enough to review. That is incompatible with a business-associate posture, so the research side of an organization is in scope while the clinical-facing, PHI-touching side is out. Anyone planning a single model stack across discovery and clinical operations now needs two.
• Your application text becomes a compliance artifact. Enforcement runs against declared use cases, which means drift between what you declared and what you actually run is an enforcement event. That is functionally an intended-use statement with a document-control obligation, and it belongs under the same change-control discipline as a device labeling claim.
• The six-month High-risk renewal is a periodic review cycle in disguise. Treat renewal as a scheduled evidence package, not a form. Projects that cannot restate scope, safeguards, and oversight every six months will lose access mid-program.
• What is absent matters for planning. No pricing or commercial terms, no quantitative biology-safety or utility evaluation of the refined safeguards, and no date for individual-plan or third-party-platform availability. Teams standardized on third-party model surfaces cannot build a schedule against this yet.

👀 What to watch

• 30 September: Governor Newsom's sign-or-veto deadline on the three California AI employment bills, which would set hiring-side AI obligations for any employer operating in California.
• Whether Anthropic publishes a supported path for BAA-enabled organizations, since that is the single gate between LSVP and clinical-facing deployment.
• Whether the High-risk Use tier accumulates published criteria, which would turn a private vetting process into something a regulated team can design against.

4️⃣ Tempus licenses Recursion's TxFM RNA foundation model for $12 million

TL;DR: Tempus and Recursion extended their data license through November 2029, converted up to $84 million in discretionary fees into $42 million committed, and signed a separate $12 million license for Recursion's TxFM RNA foundation model covering oncology diagnostic, clinical, and internal R&D use.

What happened

• On 15 September 2026, Tempus and Recursion announced an extension of their existing multi-year data license agreement and a new RNA Foundation Model License Agreement.
• The data license now runs through November 2029, with amended economics replacing up to $84 million in discretionary fees with $42 million in committed payments.
• Tempus licenses TxFM, Recursion's RNA foundation model, for oncology diagnostic, clinical, and internal R&D use, for a $12 million fee.
• Recursion disclosed the agreements in an 8-K dated 15 September 2026.
• Structurally, this is a clinical-data platform paying for access to a drug-discovery foundation model, rather than the usual direction of a model developer paying for data.

📊 Key facts (from the Tempus and Recursion announcement)

🔗 Primary source → Tempus and Recursion Extend Existing Data License Agreement and Enter New License Agreement for Recursion's RNA Foundation Model

🔍 The non-obvious point

The $12 million headline is the smaller number and the larger signal. Confidence here is medium: this is a single-source commercial disclosure, not a pattern yet.

• Both sides traded ceiling for floor. Up to $84 million discretionary became $42 million committed. Recursion buys revenue predictability at the cost of upside. Tempus caps exposure at the cost of optionality. That swap is what a data relationship looks like once it stops being speculative, and it is the shape to expect in renewals across the category.
• There is now a public unit price for an RNA foundation model. $12 million for oncology diagnostic, clinical, and internal R&D rights is the first comparable most AI-biology platforms have had. It is also roughly a quarter of the committed data spend, which says the data still prices higher than the model it trained.
• The scope is the regulatory question. A model built for discovery, licensed into diagnostic and clinical use, moves into an evidence regime it was not designed for. If TxFM-derived features ever reach a reportable result, analytical and clinical validation sits with Tempus as the diagnostics operator, not with Recursion as the licensor. The license transfers the model, not the evidence burden.

👀 What to watch

• Whether Tempus describes TxFM-derived features inside a regulated assay or confines them to research use, which is the fork between a licensing story and a submission story.
• Whether a second diagnostics company licenses a discovery-side foundation model at comparable terms, which is what would turn a $12 million data point into a market price.

5️⃣ Biomedical AI moves into curation and design, and BELL scores the evidence

TL;DR: BELL automates evidence retrieval, scoring, and explanation for biological interactions, clearing 49.5% of 210 glioblastoma protein-protein interactions at HIGH confidence, while six sibling preprints push biomedical AI from prediction into design.

What happened

• On 18 September 2026, bioRxiv posted BELL (Biomodel Evidence and LLM-based Logic), a biocuration support framework for interaction-level verification (DOI 10.64898/2026.09.18.721350).
• Five-step pipeline: entity grounding, database ranking, evidence retrieval across seven biological databases, heuristic four-dimension scoring, and LLM-generated explanation with a recommended curator action.
• Applied to 210 protein-protein interactions from a curated glioblastoma multiforme mechanistic model: 49.5% reached HIGH confidence and 72.4% received a positive curator recommendation.
• BELL is integrated into the KALIMBA curation platform.
• Six preprints in the same window widen the surface: StressNET for mechanical stress inference in tissues, miRstring for mature miRNA decoding and artificial small RNA design across species, an LLM network-medicine framework for combinatorial drug repurposing in ER+ breast cancer built from 595,122 PubMed abstracts, time-series foundation models for cognitive workload from eye-tracking, data-driven design of engineered living hydrogels, and autonomous AI-driven nanoscale spatial mapping in the 5xFAD Alzheimer's model.

📊 Key facts (from the bioRxiv preprint)

"We present BELL (Biomodel Evidence and LLM-based Logic), a biocuration support framework that automates evidence retrieval, scoring, and explanation for interaction-level verification."

🔗 Primary source → BELL: Biomodel Evidence and LLM-based Logic

🔍 The non-obvious point

For regulated builders, BELL's contribution is not accuracy. It is the shape of the artifact it produces.

• A scored, explained, database-linked recommendation is reviewable. A raw model answer is not. Four-dimension scoring plus a written rationale plus a named source set is the structure of an auditable evidence record. Any team building AI into a quality system or a submission dossier should copy the output format before it copies the method, because the format is what survives an inspection.
• The 49.5% cuts both ways. Half of a curated GBM model's interactions did not clear the evidence bar. Read that as a measurement of how thinly evidenced the published mechanistic record is, not as a model failure. If your target rationale rests on curated interaction databases, half of that support may not hold up to programmatic retrieval.
• The stated gaps repeat the week's theme. No comparison against expert-curator time, no inter-rater benchmark beyond the recommendation outputs, no prospective deployment, no evidence of generalizing beyond the GBM set. Without a curator-time delta there is no business case, only a demo.
• The design-side preprints are the ones that eventually create a pathway question. miRstring designs artificial small RNAs and the hydrogel work designs engineered living materials. Designed sequences and designed materials are products, not predictions, and they land in a different regulatory conversation than a classifier does.

👀 What to watch

• Whether curation frameworks of this type start reporting curator-time deltas and inter-rater comparisons, the two numbers that would justify integrating one into a regulated evidence workflow.
• Whether the RNA-design and living-materials preprints attract follow-on work with characterization data, which is the step that converts a design model into a regulated product candidate.

📊 The pattern

Every item this week is a statement about who checks the work and against what. NEJM AI names the checkpoints between a benchmark and a clinic. The MIMIC-IV preprints show the reference standard itself agreeing at macro kappa 0.49, and show a deterministic guardrail making things worse. Anthropic checks the requester before the request and keeps 30 days of traffic to check afterward. BELL scores the published evidence and finds half of it thin. Tempus pays for a model and inherits the validation burden that comes with using it clinically.

The trend to name: verified evidence supply. The models are no longer the scarce input. Reference standards, provenance, declared scope, and audit-ready artifacts are. Teams that can produce those on demand will ship into regulated settings. Teams holding a benchmark score will keep discovering, late, that they built the wrong evidence.

👀 Watchlist

• 30 September, California AI employment bills. Governor Newsom's sign-or-veto deadline on three bills regulating AI use in employment, which reaches any employer operating in California regardless of sector.
• BAA-enabled access under LSVP. Until Anthropic supports business-associate arrangements, the program covers research and excludes PHI-touching clinical operations, forcing dual model stacks inside a single company.
• Six-month High-risk Use renewal. The first renewal cohort will reveal whether vetting criteria are published and designable-against, or stay discretionary.
• External validation outside MIMIC-IV. The colorectal extraction and symptom-cluster work becomes deployable only when it transports to a non-MIMIC note corpus at a named institution.
• TxFM inside a regulated assay. Whether Tempus places foundation-model-derived features into a reportable oncology result, and what analytical and clinical validation it publishes if it does.
• Phase-based language in regulator materials. Watch FDA SaMD and PCCP documents and PMDA device-AI guidance for evidence-tier vocabulary, the first sign the framework is influencing submissions rather than citations.

📎 Sources

Sources of truth

Click to verify or go deeper.

Commentary we read