Life Sciences / Regulatory Review 🧬

📌 Navigate

📋 Exec Summary

Q2 2023 was a quarter of framework construction. FDA published PCCP draft guidance in April; the FDA/Health Canada/MHRA joint guiding principles would not arrive until October. Insilico Medicine dosed first Phase II patients for an AI-designed drug (INS018_055 for IPF), crossing from proof-of-concept to clinical reality. The EU Parliament voted 499-28 to classify medical AI as high-risk by default, and the AI/ML device list was already in the mid-hundreds on a snapshot-dependent basis.

📊 What Moved

📈 Trend Arcs

Arc 1: Regulatory Framework Construction — From Principles to Structure

Velocity: Accelerating

The life sciences regulatory environment for AI entered Q2 2023 in a state of accumulated principles without binding structure. The FDA had published its 2021 action plan, the 2022 guiding principles with Health Canada and MHRA, and multiple discussion papers. None of it constituted binding guidance that a company could file against. The April 2023 PCCP draft changed that. For the first time, there was a draftable, commentable, legally significant regulatory document specifying how adaptive AI/ML devices could plan and execute changes without returning to the FDA for each iteration.

Through May and June, the comment period gathered industry input. Medical device trade associations, AI health companies, and academic institutions filed comments at a volume that reflected pent-up demand for regulatory clarity. The substantive comments clustered around three issues: (1) how to scope the "planned change" description without being so specific that it constrained algorithmic adaptation, (2) how to define the validation methodology for continuous learning without running a new clinical study for every model update, and (3) how to structure the impact assessment for changes that touched multiple device functions simultaneously.

The parallel EU AI Act trajectory reinforced the same arc: regulatory frameworks for clinical AI were being built in real time, and Q2 2023 was a quarter of framework construction rather than enforcement. Companies that engaged with the comment process were, in effect, helping write the rules they would be regulated by.

Where it stands at quarter close: The PCCP draft is in active comment period. The regulatory vocabulary for adaptive AI devices now exists in written, citable form. Companies filing new AI/ML SaMD can reference the draft framework even before it's final. The EU AI Act negotiating position is set; trilogue will determine final text through late 2023.

Arc 2: Generative AI Drug Discovery — Moving from Proof of Concept to Clinical Reality

Velocity: Accelerating

Q2 2023 was the quarter generative AI drug discovery stopped being a research narrative and became a clinical timeline question. Insilico's Phase II dosing was the headline, but it was not the only signal. Across the quarter, Exscientia, Recursion, BenevolentAI, and Absci all reported pipeline progression or partnership activity that reflected AI platforms moving candidates closer to or into clinical stages.

The Recursion/NVIDIA deal would land in July after quarter close, but the market was already pricing the same structural signal: NVIDIA made a $50M PIPE investment in Recursion — not an AI software bet but a compute-as-biology-infrastructure bet. NVIDIA's thesis was that drug discovery AI would consume GPU compute at the same scale as foundation model training, and that positioning in biology-specific compute would extend its infrastructure franchise into life sciences. The deal validated Recursion's positioning as a biology-AI company with the compute requirements to justify infrastructure-scale investment, not just software-scale investment.

The pattern across the quarter was consistent: companies that could show a specific clinical timestamp (IND filing, Phase I dosing, Phase II enrollment) were being treated differently by investors and partners than companies with platform claims but no clinical evidence. Insilico's June dosing was the quarter's clearest illustration of that distinction. The transition from "AI discovers drug targets" to "AI-discovered drug in controlled trial" was the arc's defining event.

Where it stands at quarter close: One AI-designed drug in Phase II controlled trial. Multiple others in Phase I or IND stage. The clinical milestone benchmark is set. The next quarter's question is efficacy signal — not just milestone progression.

Arc 3: FDA Digital Health Center — 510(k) Pathway Maturation for AI Devices

Velocity: Steady

The FDA's Digital Health Center of Excellence (DHCoE) had been operational since 2020, and by Q2 2023 it had accumulated enough AI/ML device review experience to show patterns. The public list was already in the mid-hundreds on a snapshot-dependent basis, and that represented a de facto jurisprudence: what submission formats worked, what predicate selection criteria the agency accepted, what clinical validation evidence was sufficient for which device types.

Radiology AI remained the dominant cleared category by count — chest X-ray, CT scan, and MRI-based AI accounted for roughly 60% of cleared devices. But the Q2 quarter showed meaningful cardiology growth (ECG analysis, arrhythmia detection) and emerging pathology and ophthalmology activity. The pattern of specialty-by-specialty maturation reflected the DHCoE's organizational approach: as examiners built familiarity with a device type, review consistency improved and timelines compressed.

The PCCP draft landed into this environment as both a clarification and a complication. Cleared AI devices were static — they shipped a version and maintained it. The PCCP framework was asking companies to think about adaptive AI that would update post-clearance. Many cleared device companies had not designed their architectures for continuous learning; they had designed for point-in-time validation. The draft guidance was therefore simultaneously relevant (for new submissions) and disruptive (for companies that had built on the static clearance model).

Where it stands at quarter close: Clearance velocity is high and accelerating modestly. Pathway maturation is real in radiology and cardiology. The PCCP framework introduces an adaptive AI pathway that most currently cleared companies are not yet equipped to use. Bifurcation between static-AI device companies and adaptive-AI SaMD companies is beginning.

🗺️ Landscape Shift

🧭 Regulatory Direction of Travel

FDA DHCoE trajectory: The pattern across Q2 was faster review for well-structured submissions and increasing examiner confidence in standard AI/ML device architectures. The 510(k) pathway is functioning for static AI devices with good predicates. The De Novo pathway is the appropriate route for novel AI device types without predicates, and the DHCoE has shown willingness to use it for genuinely novel AI-enabled diagnostics.

PCCP adoption signals: The April draft was the first time companies could plan a PCCP-structured filing. Industry adoption of PCCP in new submissions won't show up in cleared device data until 2024 at the earliest. The comment period volume — multiple substantive comments from major device trade associations — suggests the industry is engaging seriously rather than waiting to see what the final rule requires.

Evidence standards in motion: The PCCP draft introduced the concept of post-market performance monitoring as a substitute for pre-market clinical validation of each adaptive change. This is a significant shift in how FDA thinks about evidence: instead of a one-time clinical study that validates the device forever, the agency is moving toward ongoing performance monitoring as the validation mechanism for adaptive systems. This is the right approach for learning algorithms but requires companies to build monitoring infrastructure they have not historically needed.

EU AI Act implications: The Parliament's 499-28 vote on high-risk classification means EU-market clinical AI companies face conformity assessment requirements under both MDR/IVDR and the AI Act. The overlap between these two regulatory frameworks — what's required by each, where they share requirements, where they conflict — is not yet resolved. Companies planning EU submissions in 2024-2025 needed to begin planning for dual-compliance architecture in Q2 2023.

International alignment assessment: The US/Canada/MHRA PCCP joint principles would not land until October, after Q2 closed. The EU's trajectory is independent and stricter (high-risk default vs the US's more case-by-case approach). The practical consequence is a two-tier international landscape: US/Canada/UK moving toward aligned adaptive AI frameworks; EU moving toward a stricter high-risk compliance regime. Companies building global regulatory strategy need different postures for these markets.

💰 Funding & Deal Pattern

Q2 2023 life sciences x AI capital flows were defined by compute-scale bets and clinical-stage validation premiums.

🔍 The Counter-Narrative

• The consensus: 250+ FDA AI/ML clearances means the agency has become broadly comfortable with AI devices. The reality: FDA became comfortable with a specific, narrow archetype: static, prediction-only, radiology-image-based, with clear predicates and well-bounded indications. For anything adaptive, novel-indication, or using generative outputs, the regulatory bar remained high and uncertain. Builders extrapolating "FDA approves AI quickly" to novel device types were reading a misleading signal.

• The consensus: Insilico Phase II dosing proves AI drug discovery works. The reality: It proved AI can design a drug candidate safe enough to enter a controlled trial — a timeline milestone, not an efficacy benchmark. Phase II readouts were not yet available. The honest version: AI drug discovery produces clinical-stage candidates faster. Whether those candidates perform better in trials is a question the data hadn't answered yet.

📐 Builder's Benchmark

Clearance timelines by pathway (estimated medians, Q2 2023):

Clinical validation evidence being accepted:

• Retrospective multi-reader studies: still the dominant validation design for diagnostic AI. Accepted for most 510(k) submissions when the device is a decision support tool.
• Prospective performance testing: increasingly expected for De Novo and any device making autonomous recommendations (not just decision support).
• Real-world performance data post-clearance: the PCCP framework's innovation — explicitly enabling post-market performance data to substitute for pre-market clinical validation of adaptive changes.

PCCP adoption rate: Zero cleared devices had used a PCCP structure at Q2 2023 close. The draft guidance was two months old. First PCCP-structured clearances expected in 2024 at the earliest.

Reimbursement signal:

AI-enabled diagnostic reimbursement remained the largest unresolved commercial constraint. FDA clearance established safety and effectiveness; it did not establish reimbursement. CMS coverage of AI-enabled devices was ad hoc — some CAD/AI codes existed for mammography and CT pulmonary embolism; most AI diagnostic tools were reimbursed only if billed under the base procedure code with no AI-specific add-on. The AMA's CPT process for AI-specific codes was running slowly. The practical consequence: cleared AI devices still largely monetized through licensing or utilization fees to health systems rather than direct reimbursement, limiting the market to health systems with capital budgets, not fee-for-service economics.

👀 What to Watch

1. PCCP comment period close (Q3 2023): FDA will summarize and respond to public comments on the PCCP draft. The agency's response to industry's scope and validation methodology concerns will determine how actionable the final guidance is for companies with continuous learning architectures. Watch for the FDA's position on post-market performance monitoring as a validation substitute — this is the highest-stakes clarification.

2. EU AI Act trilogue progress (September–December 2023): Parliament, Council, and Commission are in active negotiation. Key outstanding issues: the definition of "high-risk" within the medical device carve-out, the overlap and deconfliction between AI Act and MDR/IVDR requirements, and the implementation timeline. Watch: whether the final text creates a single conformity assessment process or two parallel ones.

3. Recursion quarterly results (August 2023): Post-NVIDIA deal, Recursion's reporting will show whether the compute-as-biology-infrastructure thesis is translating into pipeline acceleration or primarily into narrative. Watch: clinical pipeline progression beyond INS018_055 (for Insilico), new IND filings, and whether NVIDIA-delivered compute is showing up in experimental throughput metrics.

4. CMS AI diagnostics reimbursement updates (Q3–Q4 2023): The AMA CPT Category III code process for AI-enabled diagnostics is moving. Watch for new proposed codes for cardiology AI and pathology AI — these would be the first step toward routine reimbursement outside the legacy CAD code structure. First reimbursable AI codes outside radiology would be a material commercial signal.

5. DHCoE published Q2-Q3 clearance data (late Q3): The FDA periodically updates its AI/ML medical device list. The Q3 update will show whether specialty diversification — particularly pathology and ophthalmology — accelerated in the second half of the year, and whether any De Novo pathways opened new device categories.

📎 Sources

Key references for this quarter. Links provided where available; historical entries may reference publications by title and date.