AI-driven biomarker discovery

01Overview and value chain#

Markers EC: FDA PMA review for multi-cancer tests + CMS MolDX coverage + EU IVDR 2017/746 | OECD: Biotechnology and health | Regulator: FDA (USA), EMA (EU), NMPA (China)

AI-driven biomarker discovery is what happens when a biomarker stops being a molecule someone hypothesized and becomes a pattern a model found. The input is population-scale molecular measurement — methylation arrays across hundreds of thousands of plasma samples, whole-exome and transcriptome profiles across comparable cohorts — and the method is supervised learning against outcomes: cancer present or absent, tissue of origin, treatment response. The output only matters if it survives the two gates every diagnostic must pass, analytical and clinical validation, and the field’s recent history is the story of those gates becoming real: Galleri, the multi-cancer early-detection test built on methylation-pattern classification, reported full NHS-Galleri trial results at the 2026 ASCO Annual Meeting showing a four-fold higher cancer detection rate and a substantial reduction in stage IV diagnoses, and has an FDA PMA application submitted with an advisory committee anticipated; Caris Life Sciences, whose AI models profile molecular features across whole-exome and transcriptome data, carries an FDA approval for MI Cancer Seek since November 2024, MolDX approval for its Caris ChromoSeq assay, Q2 2026 revenue of $263.7 million and full-year guidance of $1.03–1.04 billion. The economics are the biomarker business at its largest: infrastructure sized for up to one million tests per year on one side, four laboratories totalling over 275,000 square feet on the other.

Key directions of AI-driven biomarker discovery:

1. Multi-cancer early detection (MCED Tests): methylation-pattern classifiers over cell-free DNA that report a cancer signal and predicted tissue of origin — Galleri’s NHS-Galleri readout showed a four-fold higher detection rate versus standard screening, with results presented from Pathfinder 2 across more than 35,000 participants.
2. AI molecular profiling for therapy choice (Profiling Platforms): whole-exome and transcriptome profiling combined with trained models that predict immunotherapy response and metastatic risk — Caris’s AI models for NSCLC and for brain-metastases risk in breast and lung cancer.
3. Regulatory-grade AI assays (FDA and MolDX Paths): the shift of AI-discovered markers into cleared and covered tests — FDA approval for MI Cancer Seek in November 2024, MolDX approval for Caris ChromoSeq, and an FDA PMA application for Galleri with an advisory committee anticipated.
4. Population-scale screening infrastructure (Million-Test Capacity): laboratory footprints built before demand lands — CLIA-certified capacity for up to one million tests per year in Research Triangle Park, over 275,000 square feet across four Caris laboratories.

Sectoral value chain#

[cohort + biobank] ──> [population-scale measurement] ──> [model training]
                                                                  │
                                                      (validation against outcomes)
                                                                  │
                                                                  ▼
[covered clinical test] <── [regulatory review, coverage] <── [locked assay]

Value chain levels#

Cross-cutting technologies of the sector:

• Cell-free DNA methylation sequencing (cfDNA Methylome): the measurement substrate of multi-cancer detection — fragmentomic and methylation signals over plasma.
• Whole-exome and transcriptome profiling (XOME Profiling): the tissue-side substrate that therapy-selection models train on.
• Clinical laboratory information systems (LIS at Scale): the chain-of-custody and reporting spine a million-test laboratory runs on.

02US#

The US hosts both the multi-cancer frontier and the profiling-at-scale business, and its regulatory institutions — FDA review, Medicare coverage through MolDX — decide which AI-discovered markers become products.

FDA PMA pathway, MolDX coverage, CLIA laboratory operation#

• GRAIL (Palo Alto / Research Triangle Park): reported full NHS-Galleri trial results at the 2026 ASCO Annual Meeting — a substantial reduction in stage IV cancer diagnoses and a four-fold higher cancer detection rate — presented Pathfinder 2 results from more than 35,000 participants, and has submitted an FDA PMA application for Galleri, with an advisory committee anticipated.
• GRAIL’s operating base: a CLIA-certified laboratory in Research Triangle Park, North Carolina, CLIA and CAP accreditations, infrastructure capable of meeting demand of up to one million tests per year, $110 million of equity financing completed with Samsung C&T and Samsung Electronics in June 2026, and Q2 2026 revenue of $44.7 million.
• Caris Life Sciences (Phoenix): FDA approval for MI Cancer Seek in November 2024, MolDX approval for Caris ChromoSeq as of August 2026, Q2 2026 revenue of $263.7 million with full-year 2026 guidance of $1.03–1.04 billion, and over 275,000 square feet of laboratory space across four laboratories.

03CN#

China’s role in AI-driven biomarker discovery is a fast-growing domestic testing market and an AI-models research base; the sourced ledger evidence for its vendors is not yet ledger-grade, so the article carries market structure rather than vendor claims.

Domestic testing demand, sequencing-industry base, model research#

• Sequencing capacity as the substrate: China’s clinical sequencing industry is among the world’s largest by volume, which makes it the natural home market for population-scale biomarker programs built on domestic cohorts.
• Regulatory separation: NMPA routes register diagnostics as devices while AI-based clinical decision software follows its own review track — the same split the FDA and IVDR frameworks manage elsewhere.
• The open question: which Chinese AI-biomarker programs will publish trial-grade evidence and regulatory approvals at the standard this page’s table requires — the dossier base does not yet carry ledger facts that would let a vendor be tabled.

04EU#

Europe contributes the population-scale proof: the NHS-Galleri program is the field’s largest real-health-system test of whether multi-cancer detection changes outcomes, and the IVDR is the framework European deployment must clear.

NHS-Galleri at population scale, IVDR certification, health-system evaluation#

• The NHS-Galleri program: full trial results reported at the 2026 ASCO Annual Meeting showed a substantial reduction in stage IV cancer diagnoses and a four-fold higher cancer detection rate — the largest evidence event yet for multi-cancer early detection, run inside a national health system rather than a registry.
• Pathfinder 2 as the interventional complement: results from more than 35,000 participants presented alongside the NHS readout.
• IVDR as the European gate: any European deployment of an AI-discovered biomarker assay certifies under IVDR 2017/746 — the evidentiary bar the corpus’s own cfDNA precedent passed, and the one multi-cancer tests have not yet faced in Europe.

05Leading companies and research institutes#

06Tech stack and innovations#

The stack is a measurement-to-model pipeline whose innovations are less about new mathematics than about owning the data supply and locking the model into an assay a regulator can inspect.

1. Methylation-pattern classification (Signal over Noise):
   • cell-free DNA methylation carries tissue-of-origin information; classifiers read it across hundreds of genomic regions per sample.
   • case: Galleri’s NHS-Galleri readout — a four-fold higher cancer detection rate against standard screening, with a reduction in stage IV diagnoses.
2. Multimodal molecular profiling with trained models (XOME + AI):
   • whole-exome and transcriptome profiles feed models trained on outcomes to predict therapy response and metastatic risk.
   • case: Caris’s AI models for NSCLC and brain-metastases risk in breast and lung cancer, and its Caris Detect and Caris Assure launches.
3. Regulatory-grade assay locking (Model to Product):
   • a discovered classifier becomes a product only when locked to an assay, validated analytically and clinically, and carried through PMA, MolDX or IVDR review.
   • case: FDA approval for MI Cancer Seek (November 2024), MolDX approval for Caris ChromoSeq, and Galleri’s submitted FDA PMA application.

07Value chains and production pipelines#

Industrial pipeline of an AI-discovered biomarker assay (CLIA/CAP regulated-laboratory regime)#

┌───────────────────────────┐      ┌───────────────────────────┐
│ 1. Cohort and biobank     │ ───> │ 2. Population-scale       │
│                           │      │ measurement               │
└───────────────────────────┘      └───────────────────────────┘
                                                 │
                                                 ▼
┌───────────────────────────┐      ┌───────────────────────────┐
│ 4. Assay lock and         │ <─── │ 3. Model discovery and    │
│    validation             │      │ training                  │
└───────────────────────────┘      └───────────────────────────┘
              │
              ▼
┌───────────────────────────┐      ┌───────────────────────────┐
│ 5. Regulatory and         │ ───> │ 6. Clinical operation at  │
│    coverage               │      │ scale                     │
└───────────────────────────┘      └───────────────────────────┘

Stage 1: Cohort and biobank

Consented samples with outcome truth are assembled through clinical partnerships — the asset that decides what a model can ever learn, and the reason programs court national health systems.

Stage 2: Population-scale measurement

Plasma and tissue run through methylation, exome and transcriptome measurement at cohort scale — the step whose cost curve sets which biomarker classes are discoverable at all.

Stage 3: Model discovery and training

Supervised models search molecular profiles for outcome-predictive patterns; candidate classifiers emerge with their discrimination measured against held-out cohorts.

Stage 4: Assay lock and validation

The classifier is frozen into a locked assay and proven reproducible — CLIA and CAP accreditation govern the laboratory, and the assay becomes a fixed analytical object a regulator can review.

Stage 5: Regulatory and coverage

Evidence runs through FDA PMA review, MolDX-style Medicare coverage and IVDR certification by route — GRAIL’s PMA application and Caris’s MolDX approvals are the working examples.

Stage 6: Clinical operation at scale

The covered test runs as laboratory medicine: capacity for up to one million tests per year on one side, over 275,000 square feet of laboratories on the other, reporting to clinicians.