Invitae

Machine learning is a feature layer inside a laboratory testing service whose value stands without it. What a clinician orders from Invitae is a sequencing test, interpreted by the laboratory's scientists under the standard variant classification framework and returned with genetic counseling.

The learned component is the Invitae Evidence Modeling platform, named on the variant classification pages: gene specific models, trained on curated evidence and kept only when they pass accuracy thresholds, whose calibrated predictions enter classification as one line of evidence among several. Former Invitae scientists report that this evidence moved 42 percent of the variants it touched out of uncertain significance, which shows how much the model changes the result, and the scientists still make the call.

Ask which genes and evidence categories use model predictions today, and how often a classification on your patients rests on them.

The oversight structure is described, and what happens after a model supported call proves wrong is not. Model predictions are one line of evidence inside the standard classification framework, weighed by the laboratory's scientists rather than issued as a classification. Each model is gene specific, trained on input curated to a single evidence category, and used only if it clears an accuracy threshold, with calibrated scores mapped to defined evidence strengths. That is a stated rule for when the system may contribute.

The missing part is the response to error: how a model is withdrawn when it drifts, and how variants classified with its evidence are reviewed and reissued.

Ask how model supported classifications are monitored after release and how clinicians are told when one changes.

The approach is named without version and update discipline. The preprint describes the platform's four design features: inputs curated to correspond to single evidence categories, gene specific model training and validation, accuracy threshold filtering, and calibrated pathogenicity scores mapped to classification categories. It names one component, a modified version of the SpliceAI splicing model, with code released publicly.

No inventory of the models in use, no version identifier on reports and no description of how retrained models are introduced is published.

Ask which models contribute to your patients' reports, how a model update is validated, and whether reports record the model version used.

The chain is structurally short. The classification models are built in house on Invitae's curated evidence, including a modified version of the openly published SpliceAI model, and run inside the laboratory's own interpretation process, so no outside model provider receives patient data for this step.

What is not published is the rest of the chain after the acquisition: which Labcorp systems and outside service providers handle sequence data and reports, and whether any processor sits between the laboratory and the record integrations. No subprocessor list is published.

Ask for the outside parties that process Invitae sequence data and reports today, and the terms that govern them.

Published research on the interpretation models, short of independent validation. A medRxiv preprint posted in February 2025 by fifteen authors, all former Invitae employees and most now at Labcorp, describes the Evidence Modeling platform applied to more than 800,000 variants across more than a million people. It reports that 42 percent of those variants would have been variants of uncertain significance without the model evidence, and that definitive classifications agreed with later ClinVar submissions more than 99 percent of the time, measured prospectively.

The work is not peer reviewed, and the concordance figure measures agreement with other laboratories' submissions rather than clinical outcomes. Invitae also reports more than five million patients tested, which is scale rather than evidence of benefit.

Ask for the peer reviewed version of the study, the false classification rate for model supported calls, and how many reclassifications have followed them.

Patient controls over de identified data are stated, and the questions machine learning raises are not answered. Invitae's privacy pages say its policies give patients rights beyond legal obligations over how their de identified data is shared, and that patients set preferences in the Invitae portal for certain research and commercial activities. Nothing published says whether patient results train the Evidence Modeling platform, how long samples and sequence data are kept, which de identification method applies, or whether preferences set before the acquisition carry over to Labcorp.

Safety engineering for the interpretation models, meaning how a model error would be detected and how affected reports are handled, is not published.

Ask whether patient data trains the classification models, how portal preferences govern that use, and the retention schedule for samples and sequence data.

Invitae publishes a HIPAA Notice of Privacy Practices, effective 22 January 2021, covering how it uses and shares protected health information and patients' rights over it, and a privacy policy from Invitae Corporation, effective 25 January 2024, which says the two documents together govern patient information. As the testing laboratory, Invitae acts as a covered entity toward patients rather than as a business associate of the ordering health system.

Both documents predate Labcorp's acquisition, the site's legal footer is now Labcorp's, and neither document says which company holds patient records today or how they move between Invitae and Labcorp Genetics.

Ask which legal entity is the covered entity for your patients' results today, and for the current notice under Labcorp ownership.

Controls are asserted with nothing independent behind them. Invitae's privacy policy carries a security measures section, and no SOC 2 report, HITRUST certification or ISO/IEC 27001 certificate is published by Invitae or by Labcorp for genetic testing, and no trust center, penetration testing statement or vulnerability disclosure program is named. Laboratory accreditation, where held, covers testing quality rather than information security.

The estate holds sequence data and results for more than five million patients and moved between companies in 2024.

Ask Labcorp for its current security attestation covering Invitae systems, with scope and date.

No device claim is made, and the models sit inside a laboratory's own interpretation process. The Evidence Modeling platform produces evidence that Invitae's scientists use to classify variants in tests the laboratory reports; it is not sold or delivered as software to anyone else. No FDA clearance is claimed for any Invitae test or for the platform, and no regulatory position for the tests is set out on the pages that describe them.

Ask for the regulatory status of the tests you would order under Labcorp ownership and whether the classification models fall within the laboratory's validated test procedures.

A testing method is published and the subgroup results are not set out. The 2025 preprint describes how models are built and kept: curated inputs per evidence category, gene specific training and validation, accuracy threshold filtering and calibration. It also says that the platform has the potential to reduce the gap in uncertain results across race, ethnicity and ancestry groups, a known problem because reference data under represent many ancestries.

No results by ancestry group are presented on Invitae's pages, the preprint is not peer reviewed, and no governance process for changing or retiring models under Labcorp is described.

Ask for the rate of uncertain results by ancestry group with and without model evidence, and who approves changes to the models.

Mechanisms let a person check the output, and nothing published stands behind it. Model evidence is weighed by the laboratory's scientists before a classification is reported, and the report reaches a clinician with genetic counseling included, so an implausible result can be questioned before it is acted on. Invitae's Notice of Privacy Practices sets out patient rights over their health information.

No error rate for model supported classifications, no commitment to reissue reports when a classification changes, and no warranty or remediation attached to a result are set out on the pages describing the platform.

Ask whether Invitae commits to notify clinicians and reissue reports when a variant is reclassified, at no charge, and for the error rate of model supported calls.

A named record system with a stated count of deployments, without named sites. Labcorp's July 2025 article on genetic testing integrations names Epic's Aura network for ordering and resulting Invitae tests and says Invitae and Labcorp teams have completed more than 400 customer record integrations, with results arriving securely in the record.

No health system is named as live, no other record system is named, and whether variant level results arrive as discrete data or as a report document is not described.

Ask for reference sites on your record system and whether results, including later reclassifications, arrive as discrete data.

A single laboratory service with location implied. Invitae tests and interprets in its own laboratory operation, now part of Labcorp, and returns results through its portal and record integrations. Nothing published says where sequence data and classification models are hosted, whether processing moved into Labcorp's systems after the acquisition, or where data rests.

Ask where sequence data and reports are stored today, whether they sit in Invitae or Labcorp systems, and in which region.

No price is published, and how the testing is sold can be established. Invitae's provider billing questions describe billing per clinical area, payment options, a financial assistance program and cost estimates on request, and explain that patients and providers may receive bills from Invitae, from Labcorp Genetics or from both since the acquisition. Genetic counseling is included at no additional cost. The answers sit in an interactive page, and no patient pay figure, institutional rate or panel price is set out.

Ask for the patient pay price and the institutional bill rate for the panels you would order, and which entity, Invitae or Labcorp Genetics, will bill your patients.

Coverage is named, with validation behind the model component. Invitae sells germline testing in six areas: oncology, women's health, pediatrics and rare disease, cardiology, neurology and genetic risk assessment. The Evidence Modeling platform is built gene by gene, with each model validated before use and discarded if it misses the accuracy threshold, so its coverage is the set of genes where a model passed, and the preprint reports concordance across the variants it was applied to.

Which genes and specialties carry validated models today is not listed on Invitae's pages.

Ask for the list of genes where model evidence is used, and the validation result for the genes on the panels you would order.