Ambry Genetics

Machine learning is a feature layer inside a laboratory testing business whose value stands without it. What clinicians order from Ambry is sequencing and RNA testing, interpreted by its variant scientists and genetic counselors. Ambry describes Classifi, its classification program, as using custom validated in silico software that applies artificial intelligence and machine learning techniques, alongside RNA studies and a curated literature database, and says the tools are only as good as the experts interpreting the data.

The CARE Program, sold to health systems, runs on NCCN testing criteria and the Tyrer-Cuzick breast cancer model, a published statistical model, rather than on a model Ambry trained. Ambry's parent, Tempus AI, positions itself around artificial intelligence; this record grades what Ambry itself sells.

Ask which classification steps use a trained model and how much a model's prediction can move a variant's classification.

Expert review is asserted and the controls around the models are not described. Ambry says its variant scientists, with expertise in nonsense mediated decay, protein modeling, RNA and splicing, and more than 100 board certified genetic counselors interpret the data, and that its tools are only as good as those experts. CARE flags patients who meet testing criteria and routes them to ordering and counseling, with clinicians acting on the result.

What is not published is the mechanism: when a model's prediction may count toward a classification, what threshold it must meet, and what happens when it conflicts with other evidence.

Ask for the rules that govern how in silico predictions enter a classification and who signs off on a call that relies on them.

The approach is described in general terms and the models are not identified. Classifi is said to combine custom in silico software using artificial intelligence and machine learning with targeted short and long read RNA sequencing from Ambry's translational genomics laboratory and a curated literature database. No model, training data or version is named.

CARE's logic is identified: NCCN genetic testing criteria and the Tyrer-Cuzick breast cancer model, updated in October 2025 to offer four scores matched to breast density categories. That makes the risk assessment's method clearer than the classification models'.

Ask which in silico models Classifi uses, whether any are third party, and how model updates are versioned.

The architecture is described and no model provider is named. Classifi's in silico software is described as custom, without saying whether any component comes from an outside model or service. Tempus's notice says protected health information goes to business associates that perform work for the company, and de identified data is shared with researchers and companies, without naming them. No subprocessor list is published for Ambry or for the CARE Program.

Ask whether any part of classification or CARE sends data to an outside model or service, and for the list of processors that receive patient data.

Published research on the risk assessment and dated program figures, with nothing on the machine learning itself. Ambry reports a study in the Journal of the National Comprehensive Cancer Network (2025) in which the CARE Program's assessments matched certified genetic counselor review in 398 of 400 real world cases, across breast, ovarian, pancreatic and prostate cancer, Lynch syndrome and familial adenomatous polyposis. Its Patient for Life page reports a 54 percent diagnostic reclassification rate when the laboratory initiated reanalysis, against 4 percent when clinicians requested it, and more than 100 new gene and disease relationships identified a year.

No evaluation of Classifi's in silico models, such as accuracy against later evidence or effect on uncertain results, is published.

Ask for the measured contribution of the machine learning tools to classifications and the rate at which model supported calls are later revised.

The data use position is stated broadly and in the company's favor. Tempus's notice, which covers Ambry, says de identified data may be used and shared with Tempus researchers and developers and with third parties including academic researchers, hospitals, laboratories and life science, insurance and pharmaceutical companies, and that Tempus may conduct research and product development with it. A sale of protected health information requires authorization. No retention period is stated for samples, sequence data or CARE Program histories, and no de identification method is named.

A grant to use and share patient data, however specific, describes what the company may do rather than how data is protected. Safety engineering for the classification models is not published.

Ask how long samples, sequence data and CARE histories are kept, whether they train Tempus or Ambry models, and which de identification method applies.

A substantive privacy document reaches the patient, issued by the parent. Ambry's notice of privacy practices link now resolves to Tempus's notice, last modified 14 May 2026, which covers Tempus AI and its affiliates including Ambry Genetics Corporation. It describes uses and disclosures of protected health information, disclosure to business associates that perform work for Tempus, research use and patient rights including correction.

The notice does not state covered entity status in so many words, and no business associate agreement is published for the CARE Program, which runs inside health systems' workflows and collects patient histories on their behalf.

Ask in what capacity Ambry holds CARE Program data, as a covered entity or as the health system's business associate, and for the agreement that applies.

A recognized certification is named without the artifact. Ambry's information and security statement, dated 27 July 2026, says it holds SOC 2 Type 2 certification, independently audited by an AICPA organization, with testing every year. The report itself, its date and its scope, including whether it covers the CARE Program application, are not published, and no penetration testing statement or vulnerability disclosure program is named.

Ask for the current SOC 2 Type 2 report and confirmation that its scope includes the CARE Program and the classification systems.

No device claim is made. Ambry's tests are run in its own licensed laboratory, with state licenses and New York exceptions listed in its legal pages, and the classification models are part of the laboratory's own interpretation rather than software delivered to customers. The CARE Program applies NCCN testing criteria and the Tyrer-Cuzick model to identify patients for testing, a guideline application whose recommendations a clinician acts on.

No FDA clearance is claimed for any test or for CARE, and no regulatory position for CARE's risk scores is published.

Ask how Ambry classifies the CARE Program's risk scores for regulatory purposes and whether any test you would order is FDA authorized.

Validation is claimed and no evaluation of the models is published. Ambry calls its in silico software custom validated and describes an interdisciplinary team behind classification, but publishes no method or results for how the machine learning tools perform, no change control for them, and nothing by ancestry, which matters because uncertain results fall unevenly across ancestry groups where reference data are thin.

The CARE Program has a published accuracy study against genetic counselor review, which evaluates the guideline engine rather than a learned model, and reports no results by patient group.

Ask for the validation of Classifi's models, any results by ancestry, and how changes to the models are approved.

A published commitment attaches to the result, and the patient has a correction route. Ambry's Patient for Life program says that when new discoveries help explain a test result, updated reports are shared proactively with the ordering clinician, at no additional cost and with no time limit, and that clinicians may request reanalysis without limits on number or timing. As the laboratory reporting the result, Ambry is also covered by Tempus's notice, which lets a patient ask for incorrect or incomplete information to be corrected.

No error rate is published for model supported classifications, and no warranty or indemnity attaches to a result.

Ask how quickly updated reports follow a reclassification and how often a reclassified variant had been called with model evidence.

Integration is claimed with no system named. Ambry's October 2025 release says the CARE Program embeds Tyrer-Cuzick scores with breast density directly into the electronic health record, and earlier releases describe CARE as integrating with health records. The CARE page describes a web based application reached from any browser, supporting assessment, ordering, results and counseling referral, without naming a record system, an app marketplace listing or a standard.

Ask which record systems CARE runs inside today, at which health systems, and whether risk scores and test results arrive as discrete data.

A single hosted service with location implied. Ambry runs testing in its own laboratory in California, and the CARE Program is a web application it hosts. Nothing published says where sequence data or CARE histories are stored, which cloud carries CARE, or whether data now moves into Tempus systems after the acquisition.

Ask where CARE and test data are hosted, in which region, and whether any of it is processed in Tempus infrastructure.

No price is published and the posture is discoverable. Ambry's billing page says nine in ten patients pay nothing for testing, that it is in network with most US health plans, that cash pricing may be available for the uninsured or those choosing not to use insurance, and that financial assistance is assessed on location, household size and income. A cost estimator is offered online. The CARE page adds that patients who do pay average less than 100 dollars out of pocket.

No cash price, panel price, institutional rate or CARE Program fee for health systems is published.

Ask for the cash price of the panels you would order and what a health system pays to run the CARE Program.

Coverage is named with validation behind part of it. Ambry tests across hereditary cancer, cardiology, neurology and rare disease, with exome and RNA testing for diagnostic cases. The CARE Program is scoped to hereditary cancer risk assessment in health systems and community hospitals, and its 2025 validation covered breast, ovarian, pancreatic and prostate cancer, Lynch syndrome and familial adenomatous polyposis. HCA Healthcare's Sarah Cannon Cancer Network is cited as a user.

The machine learning in Classifi carries no stated scope or validation by gene or specialty.

Ask which genes and specialties Classifi's models are validated for, and which cancer types CARE assesses beyond those in the study.