Clinical data infrastructure · Frontier AI

We build the data substrate for biological intelligence.

Biology's next generation of models is bottlenecked not by architecture but by data. OmicsBank deploys agentic AI inside hospital systems to convert fragmented clinical records into standardized, linked, longitudinal patient state — and makes that state available to the teams building on top of it.

See the data

12.3MPatients

30M+Imaging studies

6M+Pathology slides

500K+Biospecimens

300K+Sequencing files

17Therapeutic areas

Thesis

Foundation models for biology are training on the wrong substrate.

Protein structure prediction worked because the PDB existed — decades of curated, standardized, machine-readable structures. Biology's next frontier is not the molecule but the patient: disease trajectory, treatment response, multi-system interaction over time.

There is no PDB for that. What exists instead is fragmented across hospital silos — EMR in one system, imaging in PACS, sequencing on a lab server, pathology on glass. Roughly 80% of it is unstructured free text. None of it is linked at the patient level, and almost none of it is queryable.

OmicsBank exists to build that missing substrate, inside the institutions that generate the data, under their governance and on their terms.

The network

What's in it.

Modality	Scale	Detail
EMR / EHR	12.3M patients	Longitudinal, OMOP CDM-mapped, full unstructured notes included
Imaging	30M+ studies	Radiology DICOM and EDF
Pathology	6M+ slides	Digitized IHC and H&E with annotations
Biobank	500K+ specimens	FFPE, fresh-frozen, blood, PBMC
Omics	300K+ files	Panel, WES, WGS, RNA-seq with rich metadata

Cardiology 2.12M · Neurology 1.68M · Internal medicine 1.66M · Endocrinology 880K · Gastroenterology 840K · Urology 705K · Oncology 668K · Gynaecology 588K · Respiratory 565K — across 17 therapeutic areas.

Why disease-enriched cohorts beat population biobanks

Population biobanks were designed for epidemiology: broad, healthy-leaning cohorts sampled to represent a population. Disease AI needs the opposite — high event rates, deep phenotyping, and longitudinal follow-up on patients who are actually sick.

Dimension	UK Biobank	NIH All of Us	OmicsBank
Cohort size	500K	750K+	12.3M
Disease event rate	5–8%	~10–15%	100%
Neuro cases	~4,000	~8,000	1.67M pool
Cancer cases	~25,000	~30,000	668K · 70%+ panel sequenced
MI / stroke / diabetes	~12,000 MI	~20,000 MI	2.1M cardio pool
Single-cell RNA	—	—	Prospective
Proteomics	Olink subset	—	3 platforms · full IP
Unstructured notes	Limited	Partial EHR	Full, bundled
IP ownership	UKB retains	NIH retains	Client can own derived IP
Access timeline	6–18 months	Months	1–3 months
South Asian ancestry	~3%	~5%	Majority

Generation platforms

10x GenomicsParse BiosciencesIlluminaOlinkSeerThermo Fisher

Partners and backers

Built with the infrastructure and capital behind frontier compute.

90Hospital units

10Diagnostics labs

20,000+Beds

3Countries

For pharma & biotech

AI-driven discovery is now clinically validated. The constraint is the input data.

The field no longer has to argue that computational target discovery works — it has been demonstrated in humans. What limits the next wave of programs is the depth, linkage, and disease-enrichment of the data those engines run on. That is the layer OmicsBank builds.

668K oncology patients with 70%+ panel sequencing. 2.1M cardiometabolic. 1.67M neurology. A 100% disease event rate — arithmetic a population biobank cannot reach at any cohort size.

Precedent: what the field has published

Insilico Medicine's programs are the clearest published demonstration of the discovery-to-clinic loop — and of exactly where cohort depth becomes the binding constraint.

Nature Medicine · 2025

Rentosertib — first clinical proof-of-concept for an AI-discovered target

Insilico's TNIK inhibitor, with both target and molecule generated computationally, reported phase 2a results in idiopathic pulmonary fibrosis. Across 71 patients at 21 sites, the 60 mg once-daily arm showed a mean FVC change of +98.4 mL against −20.3 mL on placebo, with exploratory biomarker analysis supporting the proposed mechanism.

The data dependency

Novel-target programs live or die on phenotyped patient cohorts deep enough to validate mechanism and stratify responders — before a trial, not after it.

Frontiers in Aging Neuroscience · 2022

ALS target discovery from a few hundred cases

Using PandaOmics across CNS samples (237 cases, 91 controls) and iPSC-derived motor neurons from Answer ALS, Insilico and collaborators at Johns Hopkins, Mass General, Mayo Clinic and others nominated 17 high-confidence and 11 novel targets; eight previously unreported genes rescued neurodegeneration when suppressed in a Drosophila model.

The data dependency

That study drew on a few hundred cases. Our neurology pool is 1.67M patients, with linked imaging, labs, and full clinical narrative.

J. Chem. Inf. Model. · 2024

Target engines are multimodal by design

PandaOmics ranks targets by combining omics evidence with biomedical text, and Insilico's suite extends to trial-outcome forecasting. The published methodology is explicit that target and biomarker hypotheses are generated from multimodal omics plus text, not from any single assay.

The data dependency

Multimodal engines need multimodal inputs, linked per patient. Ours arrive pre-normalized to OMOP CDM with pathology, imaging, pharmacy, and omics on the same identifier.

Rentosertib, PandaOmics, Chemistry42 and inClinico are programs and platforms of Insilico Medicine. Cited here as published precedent for the class of work this substrate supports; OmicsBank was not a party to these studies.

Program 01

Target discovery and validation

Multi-omics linked to longitudinal clinical outcomes, so a candidate target can be interrogated against real disease trajectory rather than cell-line proxies.

Clinical + genomics + proteomics + metabolomics, patient-linked
Differential signal across responder and non-responder strata
Comorbidity and polypharmacy context from full EMR history

Program 02

Patient stratification and biomarker discovery

Define and validate a stratification hypothesis on a cohort large enough to hold out a test set, with unstructured clinical narrative included at no additional cost.

Molecular subtyping against observed outcomes
Companion-diagnostic feasibility and prevalence estimation
Domain-specific biomarker panels and biobank construction

Program 03

Prospective data generation

When the modality you need was never part of routine care, we identify eligible patients from longitudinal EMR, consent them under IRB/EC-approved workflows, and generate it.

Single-cell and spatial transcriptomics — 10x Genomics, Parse
Targeted and unbiased proteomics — Olink, Seer, mass spectrometry
On-demand IHC and H&E digitization from banked or fresh tissue

Program 04

Translational and indication expansion

Test whether a mechanism holds in an adjacent indication before committing a program to it, using the same normalized substrate across therapeutic areas.

Cross-indication cohort construction across 17 therapeutic areas
Real-world response and progression signals
South Asian ancestry depth for populations underrepresented elsewhere

DeliverablesOMOP CDM tables · FASTQ/BAM · expression and protein matrices · digital slides · QC reports

Timeline1–3 months to delivered cohort

Book a technical call

For frontier AI

Corpora, environments, and evals for biomedical agents.

Language models learned reasoning from static corpora. Agents learn from environments — states, actions, transitions, rewards. Biology has had no such environment: no setting where a model can propose an intervention, observe a trajectory, and be scored against ground truth. We are building it from real patient data.

Environments can be delivered inside a federated boundary, so the agent trains where the data lives and no patient state crosses an institutional or national line.

State

Patient state at t

Linked EMR, labs, imaging, pathology, pharmacy, and omics — normalized to OMOP CDM as one machine-readable representation.

Action

Interventions taken

Prescriptions, procedures, dose changes, referrals, diagnostics ordered — coded via RxNORM and procedure ontologies.

Transition

Observed state at t+1

Longitudinal follow-up across a 12.3M-patient network with a 100% disease event rate.

Reward

Scoreable outcomes

Clinical outcomes, biomarker response, progression, readmission, survival.

Coverage

Enough trajectories

17 therapeutic areas. 2.1M cardiology, 1.67M neurology, 668K oncology patients.

Program 01

Pre-training and post-training corpora

High-volume EMR and imaging at scale, delivered model-ready with metadata intact.

12.3M patient records including full unstructured narrative
30M imaging studies, 6M digitized pathology slides
Parquet, JSON, or flat files; co-located compute where data cannot move

Program 02

RL environments

Trajectory environments constructed from observed care, for agents that must act rather than answer.

Clinical reasoning over diagnostic and management trajectories
Treatment-sequencing policies in oncology, CKD, and immunology
Counterfactuals drawn from a disease-enriched cohort, not a synthetic prior

Program 03

Evaluation harnesses

Held-out real-world trajectories as an eval set for medical agents — outcomes, not multiple choice.

Scored against observed patient outcomes
Provenance and consent lineage intact through evaluation
Stratified by therapeutic area, ancestry, and care setting

Program 04

Digital twin substrate

Patient-state transition models trained on multimodal longitudinal data rather than single-modality claims or synthetic proxies.

Multi-system state representation per patient
Trajectory modelling across lines of therapy
Grounded in real disease progression at population scale

SafeguardsDe-identified, consent-tracked, IRB/EC-approved; trajectories never reconstructed to a re-identifiable level

Book a technical call

For RWE & CROs

Evidence that survives a regulatory read.

Real-world evidence fails on provenance more often than on statistics. Every record in the network carries consent state, transformation lineage, and source attribution from ingestion through export — because governance is enforced at the data layer, not reconstructed at submission time.

Mapped to OMOP CDM and CDISC-SDTM at rest, so the substrate a study runs on is the same one a regulator expects to see.

Program 01

External and synthetic control arms

Construct comparator arms from real patients matched on baseline covariates, for trial designs where a randomized control is infeasible or ethically fraught.

Rare disease, oncology, and progression studies
Longitudinal depth sufficient to model trajectory, not just endpoint
Multimodal matching — not claims data alone

Program 02

Feasibility and site selection

Answer whether a protocol is recruitable before it is written, with counts by inclusion and exclusion criteria across the live network.

Cohort counts against draft I/E criteria
Site-level prevalence and throughput estimates
Protocol amendment impact modelling

Program 03

Protocol-to-patient matching

Structured records make matching a query rather than a chart review. Sites move from weeks of manual screening to minutes.

Automated screening against active protocols
Genomic and biomarker eligibility included in the match
Deployed at the site, inside the firewall

Program 04

Post-market evidence and safety

Long-run outcome and safety signal detection on a population followed continuously, with pharmacy and lab data linked to the clinical record.

Comparative effectiveness in routine care
Adverse event and signal detection across pharmacy linkage
Label expansion and HEOR evidence packages

StandardsOMOP CDM · CDISC-SDTM · SNOMED CT · RxNORM · ICD-10

AuditConsent state and lineage traceable per record

ComplianceHIPAA · GDPR · DPDP · ISO 27001

Book a technical call

For hospitals

Your data, structured, inside your firewall.

Every hospital writes a book of clinical knowledge each year, and almost none of it gets read. Deploying OmicsBank gives an institution the AI-ready foundation every downstream application depends on — before any question of external licensing arises. Live in under 90 days, on-premise, air-gapped if required.

Clinical records that are not structured today cannot be retroactively standardized in five years. Every month without a data lake is longitudinal patient history permanently lost to future care and research.

Program 01

Unified patient view

One record per patient across EMR, PACS, LIMS, pharmacy, and genomics — replacing a view that today exists nowhere in the building.

AI co-pilots for diagnosis and tumour boards, trained on your own patterns
Lost-patient recall across the full longitudinal record
Genomic, clinical, and imaging data unified per patient, not per department

Program 02

Operations and throughput

The same structured foundation powers the operational layer, not just the clinical one.

Bed optimization and capacity forecasting
Billing and coding co-pilots
Demonstrable governance for regional data protection audits

Program 03

Research and trial site readiness

Retrospective studies without months of chart abstraction — and protocol matching that makes you the preferred site in your region.

Real-world evidence generation from your own population
Protocol-to-patient matching in minutes
Investigator-initiated research on a queryable substrate

Program 04

Governed licensing revenue

License de-identified cohorts to pharma, biotech, and AI labs — governed by you, on your terms, with revenue share.

Consent and lineage enforced before any export
Institution approves every request
Projected 5–7x first-year return from licensing, trial revenue, and operational savings

Book a technical call

Platform

The stack

Six layers, deployed on-premise. Order matters — each layer is a precondition for the next.

Ingestion

Small language models tuned for clinical text ingest EMR/EHR, PACS/DICOM, LIMS, pharmacy, genomics, and whole-slide pathology. Handles 3K to 30M patient records with no change in architecture. Typical time to production: under 90 days.

HL7 / FHIRDICOMFASTQ / BAM / VCFWSI

Agentic structuring

Patent-pending agentic extraction converts unstructured clinical narrative — discharge summaries, radiology reports, operative notes, scanned correspondence — into structured, coded, machine-readable records. Eleven patents filed. Runs entirely inside the firewall.

11 patents filedOn-prem inferenceNo external LLM calls

Normalization

Every record harmonized to a single schema. This is what makes cross-site federation possible: two hospitals on different EMR vendors in different jurisdictions resolve to the same representation.

OMOP CDMSNOMED CTICD-10RxNORMCDISC-SDTM

Governance

Automated PII removal, consent-state tracking, and full lineage audit execute before any record leaves the secure vault. Compliance is enforced at the data layer, not bolted on at export.

HIPAAGDPRDPDPISO 27001IRB / EC workflows

Federation

Models train across sites without patient-level data crossing an institutional or national boundary. The query moves; the data doesn't.

NVIDIA FLARERhino Health FCP compatible

Delivery

Datasets ship model-ready with rich metadata. Co-located GPU environments available for workloads where data cannot move at all.

ParquetJSONFlat filesCo-located compute

The substrate is the bottleneck. We're building the substrate.

Every month a health system runs without structured data is longitudinal patient history that cannot be reconstructed later. Every month the field trains biomedical models on synthetic proxies is a month spent optimizing against the wrong distribution.

Book a technical call