Caris Life Sciences

The models are the engine of several products, and the flagship report rests on sequencing. Caris GPSai uses neural networks on exome and transcriptome data to match a tumor to 90 cancer categories, Caris FOLFIRSTai combines about 5,000 machine learning models to predict which first-line chemotherapy sequence benefits a metastatic colorectal cancer patient, MI Clarity applies computational pathology to breast cancer recurrence risk without sequencing, and the research-use Molecular Tumor Board Report adds further signatures such as MGMTai. MI Cancer Seek, the FDA approved test at the center of the offering, is a whole exome and transcriptome sequencing assay whose approved claims rest on the sequencing rather than on the AI.

The oversight structure is described and one part is missing. Tests are ordered by physicians and reported from the company's CLIA certified laboratory, with the Pathologist in Chief and medical staff available for consultation; the report states that its therapy associations do not replace the treating physician's independent judgment; GPSai is presented as additional insight for unknown primary or ambiguous cases, and FOLFIRSTai as a consideration between two standard chemotherapy sequences. The label scope table tells the reader which results the FDA approval covers and which it does not, and the Molecular Tumor Board Report is labeled research use only for tumor boards and case consults.

What is missing is what the reader is told about each prediction's confidence and limits: whether a GPSai match below a threshold is flagged, when a FOLFIRSTai call should not be relied on, and how the research-use report's suggestion that it can help confirm a diagnosis squares with its research-use label.

The approach behind each signature is named, without version or update discipline. GPSai uses trained neural networks on exome and transcriptome data, trained on more than 250,000 Caris cases across 90 cancer categories; FOLFIRSTai is an ensemble of about 5,000 machine learning models on the molecular profile; the brain metastasis signatures state their training cohorts; MI Clarity uses computational pathology; and CODEai is the linked molecular, treatment and outcome dataset for more than 484,000 patients that the models draw on.

The approved MI Cancer Seek assay has an FDA-approved change control plan for its software, but nothing comparable is published for the signatures: no version identifiers on reports, no account of how a retrained signature is validated, and no notice to clinicians who received results from an earlier version.

A structurally short chain for the models, with the parties around the data only partly named. The signatures are Caris's own, trained and run on its infrastructure, so no outside model provider sits in the chain. The annual report names Amazon Web Services among its computing providers alongside on-premise and co-located systems, describes third party network monitoring and penetration testing, and says de-identified data flows to Precision Oncology Alliance members, biopharma partners and academic researchers, with Alliance members contributing outcome data back; Epic and Flatiron OncoEMR carry orders and results.

No subprocessor list or list of data recipients is published.

Validation is published for the approved assay and for the signatures, including against independent trial data. MI Cancer Seek's analytical and clinical performance underpinned its FDA premarket approval. FOLFIRSTai was validated on two sets independent of its training data: 467 curated real world cases and 296 cases analyzed retrospectively from the randomized phase III TRIBE2 trial, where patients treated in line with its prediction had median overall survival 17.5 months longer than those treated against it. GPSai was trained on more than 250,000 cases, validated prospectively on 3,200 cases, and reported at 95 percent accuracy in a 2025 validation across more than 97,000 cases. In August 2026 the company published a study of AI-guided therapy selection and survival in pancreatic cancer in npj Precision Oncology, and the brain metastasis signatures state their training cohorts (12,994 lung and 3,371 breast cancer cases with matched survival).

Most of this work is company-led and retrospective; the TRIBE2 validation is the strongest element because the trial data were generated outside Caris.

The clinical relationship is covered by published instruments, and the research dataset rests on a de-identification method that is not named. Caris states in its annual notice to physicians that it is a covered entity and publishes a HIPAA notice of privacy practices, a separate privacy notice and a Washington health data notice. The annual report says it contributes de-identified molecular data to the Caris Precision Oncology Alliance, whose members may contribute outcome data for joint research, and furnishes de-identified information to biopharma and academic partners, relying on various de-identification methods and acknowledging they could fail. The CODEai platform links molecular profiles with treatment and outcome data for more than 484,000 patients.

A tumor type, molecular profile, treatment sequence and survival interval together describe a person narrowly, so the method (expert determination or removal of listed identifiers) matters, and it is not stated. How outcome data is obtained and linked to profiled patients, and how long specimens and sequence are kept, are not published. Safety engineering for the signatures is not described.

Covered entity status is stated plainly and a substantive privacy document reaches the patient; no contracting instrument is published. The annual notice to physicians (September 2025) states that under HIPAA Caris is a healthcare provider and a covered entity, links its notice of privacy practices, and sets out its Stark and anti-kickback compliance policy to referring physicians. The privacy policies index adds a general privacy notice and a Washington health data notice.

No agreement for ordering institutions, business associate terms for the research platform, EHR partners or the Precision Oncology Alliance, or subcontractor provisions are published.

Controls are described with an outside check on a stated cadence, and no attestation. The annual report (Item 1C) describes a program informed by NIST CSF 2.0 with an emphasis on HIPAA, a formal IT risk assessment and external network and production application penetration tests at least annually, in-house penetration testing, independent risk assessments by outside advisors, third party network monitoring, two-factor authentication, and an incident response plan led by a Chief Information Security Officer; vendors are reviewed through questionnaires and their SOC 2 reports where available.

No SOC 2, HITRUST or ISO/IEC 27001 attestation of Caris's own is named, and the laboratory's CLIA and CAP credentials cover testing quality. The CODEai research platform is reachable by outside organizations, which makes its access control and export limits part of the security question.

An approval in FDA's database with its indications, and the unapproved parts named as such. MI Cancer Seek was approved through premarket approval P240010 on 5 November 2024 as a next generation sequencing test on FFPE tumor tissue with companion diagnostic indications; supplements include an FDA-approved predetermined change control plan for minor software modifications (S001, February 2025) and later manufacturing and quality control changes.

The MI Cancer Seek page publishes a label scope table: variants in 228 genes, MSI, TMB, one copy number alteration in breast cancer and 8 companion diagnostic indications are within the approved label, while the remaining genes, HRD, RNA results and the AI signatures GPSai and FOLFIRSTai are not, with the instruction not to imply that all services are FDA approved. The sample report states that it mixes FDA approved tests and laboratory developed tests, and the Molecular Tumor Board Report is labeled research use only.

Validation is published and evaluation across patient groups is not. The signatures are validated on large cohorts (GPSai on more than 97,000 cases, FOLFIRSTai on real world and phase III trial data), and the annual report acknowledges the risk that its AI may need retraining or be limited by regulation. Nothing published reports signature performance by ancestry, sex, age, insurance status or treating site, and no governance framework or monitoring process for the signatures after release is described. The FDA-reviewed change control plan for MI Cancer Seek covers the approved assay's software, not the AI signatures.

The question matters most for signatures trained on observed outcomes. FOLFIRSTai and the brain metastasis signatures learn from treatment and survival as they happened, so where a group was treated less aggressively or diagnosed later, that history can enter the model as if it were biology.

Performance is published with its denominators, and the approved and unapproved parts of the report are separated in writing; nothing makes the company answerable for a prediction. The GPSai overview reports 95 percent accuracy in a 2025 validation across more than 97,000 cases, after prospective validation on 3,200 cases; FOLFIRSTai reports that patients treated in line with its prediction lived 17.5 months longer in validation on 467 real world cases and 296 cases from the phase III TRIBE2 trial. The MI Cancer Seek page lists which results are within the FDA approved label (variants in 228 genes, MSI, TMB, one copy number alteration, 8 companion diagnostic indications) and which are not, naming GPSai and FOLFIRSTai among the latter, and the sample report states that treatment decisions rest on the physician's independent judgment.

No guarantee, indemnity or correction route attaches to a signature result.

Named bidirectional integrations with major record systems, in production at scale. Caris joined Epic's ORA network as the first whole exome and whole transcriptome partner, integrating test ordering and results with structured genomic data inside Epic, and its profiling is available inside Flatiron's OncoEMR for ordering, order tracking and result delivery. The EHR integrations page reports more than 315 organizations and 3,330 clinical locations integrated, with an average 2.2 day turnaround improvement, and quotes UPMC Hillman Cancer Center on ordering and viewing results inside Epic. The company says it can integrate with any CMS-certified record system.

It is not stated which result elements arrive as discrete data in each system, or whether updated interpretations reach the record.

The company runs its own infrastructure plus a named cloud, with no data location committed. Testing happens in Caris laboratories on shipped specimens, and the annual report describes on-premise, co-located and cloud computing infrastructure, naming Amazon Web Services as a provider whose interruption could affect results. The clinical pipeline, the CODEai research platform offered to outside biopharma and academic users, and the data shared with Precision Oncology Alliance members are different uses of the same material.

No region, residency commitment, separation between clinical and research environments, retention schedule for sequence and linked outcomes, or subprocessor list is published.

No price is published and the posture is discoverable. Caris bills insurance, Medicare and state Medicaid programs directly, says cost varies with the profile ordered and the number of biomarkers and technologies used, performs medically necessary tests regardless of network status, appeals denials on the patient's behalf, requires an Advance Beneficiary Notice where Medicare coverage criteria are not met, and offers a financial assistance program. GPSai and FOLFIRSTai are included at no additional cost, and the Molecular Tumor Board Report is available on request. The annual report says average selling prices are set by payer arrangements and do not differ materially by cancer type, and that MI Profile's average rose with the MI Cancer Seek launch.

No list price, self-pay rate or biopharma rate is published.

Coverage is named product by product and signature by signature, with validation behind each. MI Cancer Seek profiles solid tumors in adult and pediatric patients with its FDA-labeled indications; FOLFIRSTai applies only to metastatic colorectal cancer patients eligible for first-line FOLFOX or FOLFIRI and is validated in that population; GPSai applies to cancers of unknown primary and ambiguous cases across 90 categories; MI Clarity is limited to breast cancer recurrence risk; the Molecular Tumor Board Report's tumor-specific modules name their cancers (colorectal, breast, pancreatic, glioblastoma). Blood-based profiling (Caris Assure) and early detection (Caris Detect) are offered separately.