Helix

Artificial intelligence is a feature layer on a sequencing and data business whose value stands without it. What Helix sells is Exome+ sequencing in its own CLIA certified, CAP accredited laboratory, clinical genetic tests and population screening programs for health systems, and GenoSphere, a linked dataset of exomes with years of EHR and claims history.

The learned components sit in the research tools: Cohort Builder's AI Research Assistant, which the company says is trained on its clinico-genomic data and workflows and builds and refines cohorts through conversation, and a GenoSphere connector that lets a researcher's own AI assistant query aggregate statistics. The clinical tests describe embedded decision support without describing a model.

The research tool's oversight is described, and the clinical return of results is not. In Cohort Builder a researcher sets or reviews criteria in a visible drag-and-drop interface, checks the cohort's demographics before analysis, and can use the AI assistant to build and refine criteria by conversation; the company limits variant-level outputs to feasibility use and tells researchers to validate variant data independently before using it for trial eligibility. That is a described human check on the AI's output.

What is missing is the clinical side: who confirms a screening finding before it reaches a participant, what genetic counseling accompanies it, and how a later reclassification reaches someone told years earlier that they were clear. The providers page says patients get counseling support and results with embedded decision support, without describing the process.

The assay and data model are described; the AI is described only in general terms. Exome+ is a proprietary sequencing assay extending beyond a standard exome, and GenoSphere links it to OMOP-harmonized EHR data and claims, refreshed quarterly. Cohort Builder's AI Research Assistant is said to be trained on Helix's clinico-genomic data and platform workflows, to build and refine cohorts through conversation and to keep context across a session; the connector exposes aggregate queries to a researcher's own AI assistant.

No model or supplier is named for the assistant, no evaluation of how accurately it translates a request into criteria is published, and no version or update practice is described.

The architecture is described and the model provider behind the AI assistant is not named. The parties that are named sit around the data: Databricks, whose platform hosts Helix Research Workspaces with individual-level access to more than 550,000 records; Google Cloud, which processes the support chatbot's data; Anthropic's Claude, into which the aggregate-only GenoSphere connector was launched for researchers' own use; and life sciences partners with data access agreements. Cohort Builder's assistant, which works on Helix's own data, is described without saying whose model it runs on.

No subprocessor list is published.

Named deployments with dated figures, and published research short of independent validation of the AI tools. Population genomics programs run at about 20 health systems, including Cone Health, Sanford Health, WellSpan, St. Luke's, HealthPartners, Nebraska Medicine and Memorial Hermann, with more than 530,000 exomes sequenced and 1.5 million lives in enterprise programs. The providers page reports that network data resolves the status of 45 percent of variant of uncertain significance carriers, with 96 percent of reclassified variants confirmed benign and a 9 percent overall increase in pathogenic calls, linked to the company's research. The home page's estimated life years saved is a modeled figure built on a published cost-effectiveness framework, labeled as such.

No evaluation of Cohort Builder's AI assistant or the query connector is published.

Specific controls on two access routes, and a wide one opened in September 2026. The GenoSphere connector returns only de-identified aggregate statistics, suppresses small counts automatically and works only through provisioned partner accounts, which blocks the usual attack of narrowing a cohort to one person. Helix Research Workspaces, launched on Databricks on 24 September 2026, gives researchers individual-level access to more than 550,000 linked records (exome, 13 or more years of EHR history, medical and pharmacy claims) inside project-based workspaces with access controls. The privacy notice says de-identified data may be used for any lawful purpose, states CLIA retention of test records (2 years) and proficiency records (7 years), and says customer support chatbot data is processed by Google Cloud, used to improve the chatbot and deleted after one year. The HIPAA notice (effective 21 May 2026) permits research use under legal safeguards.

No de-identification method is named for a dataset where an exome plus a decade of clinical history narrows a person sharply, and nothing states what leaves a Research Workspace. Safety engineering for the AI assistant is not published.

Covered entity status is evident and a substantive privacy document reaches the patient; no contracting instrument is published. Helix publishes a HIPAA notice of privacy practices (effective 21 May 2026) for its germline testing and related screening services, covering uses and disclosures, research use under legal safeguards, patient rights and the requirement that business associates protect PHI. A separate privacy notice covers the website, accounts, research participation and the support chatbot. Research participants enroll through their health system under an IRB-approved protocol with consent.

No agreement for health system partners, business associate terms for Databricks or other processors, or statement of which Helix entity contracts is published.

Controls are asserted with nothing independent behind them. The company describes one designed control, the GenoSphere connector's aggregate-only results with automatic small-count suppression and provisioned accounts, and project-based access controls and monitoring in the new Research Workspaces. No SOC 2, HITRUST or ISO/IEC 27001 attestation, penetration testing statement, vulnerability disclosure program or trust center is published, and the laboratory's CLIA and CAP accreditation covers testing quality rather than information security.

What needs protecting is large and outward facing: more than 550,000 linked exome and clinical records, workspaces opened to outside researchers with individual level access, data delivered into customers' own environments, and an AI query route.

The position is stated for the laboratory tests and the research tools, with no FDA status claimed. Helix's clinical tests are validated in its CLIA certified, CAP accredited laboratory and offered as laboratory developed tests, with no FDA clearance claimed; the research tools are scoped away from clinical use, with Cohort Builder's variant outputs described as for informational feasibility use by qualified researchers, not for clinical diagnosis or patient management.

The clinical tests carry embedded decision support for medication adjustment, referral and cascade screening, which sits inside the laboratory's report rather than under device review. Population screening returns findings to people without symptoms, so the counseling and confirmatory pathway falls to the health system.

The research program is governed, and the AI tools are not covered by anything published. The Helix Research Network runs under an IRB-approved protocol with participant consent, return of individual results and reporting back to participants, which governs how data is collected and used for research. Nothing comparable is published for the Cohort Builder assistant or the query connector: no evaluation of how accurately the assistant turns a request into cohort criteria, no review process for changes to it, and no monitoring.

On bias, the company describes GenoSphere as drawn from a diverse population and says Cohort Builder shows ancestry representation to users, but publishes no composition figures and no performance by ancestry for the variant interpretation behind its clinical tests or its reclassification work. Its own FAQ notes that ancestry predictions become more accurate as more diverse participants join.

Mechanisms let a user check the output, and nothing stands behind it. Cohort Builder shows the inclusion and exclusion criteria in a drag-and-drop view and the resulting cohort's demographics, so a researcher can see what the AI assistant built before relying on it, and the company states that variant counts are for informational feasibility use by qualified researchers, not for clinical diagnosis, and must be independently validated before use in trial eligibility. On the clinical side, family variant testing is offered at no additional cost within 90 days of a relative's report.

No error rate is published for the assistant's cohort definitions, and no guarantee, indemnity or remediation attaches to either the research tools or the clinical results.

A named record system and a named data standard, without a named live site. The providers page says clinicians can order Helix tests inside their EHR and that Helix integrates with Epic so results appear where they practice, with embedded clinical decision support and pharmacogenomic results returned in minutes from stored sequence. On the research side, GenoSphere's clinical data is harmonized to the OMOP common data model and refreshed quarterly, so analyses carry across to other datasets built the same way.

The company names no health system as live on the Epic integration, and how results are represented in the chart (discrete coded findings or documents) and how a later reclassification reaches the record are not described.

Several access routes are described, and no data location is stated for any of them. Researchers can work on GenoSphere in place through managed access, receive data into their own trusted research environment, query aggregates through the connector, or, since September 2026, work on individual-level records in Helix Research Workspaces built on Databricks. The privacy notice says chatbot data is processed on Google Cloud.

No hosting provider for the core dataset, cloud behind the Databricks workspaces, region, residency commitment, separation between the roughly 20 contributing health systems, or retention schedule for sequence and linked records is published.

No price is published and the posture is discoverable. Clinical tests are billed to Medicare, Medicaid and commercial insurance, with a cash-pay price offered on request, a financial assistance program and monthly payment plans; family variant testing is free within 90 days of a relative's report. Health plans can buy Helix Genomic Advantage, described as one sequence at one cost with reusable results. Population genomics programs for health systems and multi-year data access agreements for life sciences (Alnylam and AstraZeneca among those announced) are contracted separately.

No cash-pay figure, program fee or data access rate is published.

Coverage is named across three businesses, with validation behind part of it. Helix sells clinical genetic tests (hereditary cancer, cardiovascular, pharmacogenomics, carrier screening, rare disease exome) validated in its laboratory, population screening programs run by about 20 health systems, and research data and tools for life sciences, where the AI sits. The AI tools are scoped to research feasibility and explicitly not to clinical diagnosis or patient management.

The clinical tests carry laboratory validation and the screening programs named deployments; the AI assistant's scope is stated clearly but no evaluation stands behind it.