Caris Life Sciences

The AI Insights signatures and the CodeAI platform are genuine model outputs, and MI Clarity is AI native, applying computational pathology to stratify recurrence risk with no genomic sequencing required. Held back from A because the flagship revenue product, MI Cancer Seek, is a sequencing assay whose FDA approved companion diagnostic function rests on the molecular profiling itself; the AI signatures are layered on top of it.

The delivery vehicle implies human deliberation, which is a genuine strength, and the design of the review is not described.

Two controls sit around the output. The report is issued by an accredited clinical laboratory under a regulated quality system and signed out by a qualified director, which is externally imposed and inspected. And the artificial intelligence signatures are delivered inside a molecular tumour board report, which is the right destination: a molecular tumour board is a multidisciplinary meeting where oncologists, pathologists and scientists weigh a case together, so a signature arrives into a structured deliberation rather than as an answer to a single clinician under time pressure. That is a better oversight context than most predictive tools in this index enjoy.

What is not published is what the deliberation is given to work with. Whether the report shows a signature's confidence, the cohort it was derived from, its performance characteristics, the features driving a particular prediction, or the conditions under which it should not be relied upon. A tumour board can only scrutinise what it can see, and a bare risk category is difficult to argue with.

The interaction between components matters too. A single report carries an approved companion diagnostic result and one or more predictive signatures of different provenance, and nothing states how the report distinguishes them for the reader.

Ask what a tumour board sees alongside each signature, how disagreement is recorded, and how the report separates regulator approved results from company validated predictions.

Signatures are named and their training cohorts are quantified, which is better than most. What each one is remains undescribed, and one distinction on this record is easy for a reader to lose.

What is public: named signatures for chemotherapy efficacy prediction, tissue of origin, and brain metastasis risk, with training cohort sizes stated for the metastasis work at close to thirteen thousand lung and over three thousand breast cancer cases with matched survival outcomes. Quantifying a training cohort is a real disclosure and few vendors in this index do it.

What is absent: the method behind any signature, the features used, validation design, whether validation was internal or external, performance figures, versioning, and how a signature change is communicated to clinicians who received earlier reports.

The distinction that matters most is regulatory rather than technical. The company's sequencing assay holds full premarket approval, which is a high bar and correctly noted elsewhere on this record. The artificial intelligence signatures layered on top are a different thing: one is described by the company as clinically validated, which is a claim the company makes rather than a status a regulator conferred. A clinician reading a single report containing both an approved companion diagnostic result and a predictive signature may reasonably assume the same authority stands behind both. It does not, and the report should make that clear.

Ask which components are within the approved assay's scope and which are not, and for the validation evidence behind each signature.

Good instruments cover the clinical relationship properly, with an explicit covered entity position, a published notice of privacy practices and a separate website statement, and one claim deserves closer reading than it usually receives.

The company describes its clinico genomic platform as fully de identified and compliant, at a scale of hundreds of thousands of patients with over a million data points per patient integrating molecular profiles with treatment information and survival outcomes. De identification is a defined legal standard and the claim may well be correct.

It is also being made about a dataset unusually rich in exactly the attributes that make de identification hard: a tumour type, a molecular profile, a treatment sequence and a survival interval together describe a person very narrowly, and a million data points per patient is a great deal of surface. The method used is not stated, and for a dataset of this shape the distinction between removing enumerated identifiers and a statistical determination is material rather than technical.

A second question follows from the outcomes themselves and is easy to miss. Survival and treatment response are not generated by the laboratory; they accrue after the test, elsewhere, which means an acquisition pathway exists that brings longitudinal outcome data back and links it to a profiled patient. Nothing published describes it, and it is the step where a de identified profile meets a real person's subsequent history. Ask which method applies and who certified it, and how outcome data is obtained and linked.

Clinical and analytical validation of MI Cancer Seek was published, and the AI signatures disclose their training cohorts with specificity, including 12,994 NSCLC and 3,371 breast cancer cases with matched survival outcomes for the brain metastasis signatures. Publishing cohort sizes for individual model signatures is materially more transparent than aggregate accuracy claims.

Good instruments and one claim that deserves closer reading than it usually receives.

The instruments are real: an explicit covered entity position, a published notice of privacy practices, and a separate website privacy statement. Those cover the clinical relationship properly.

The claim concerns the research dataset. The company describes its clinico genomic platform as fully de identified and compliant with the health privacy rule, and states its scale as more than 484,000 patients with over a million data points per patient, integrating molecular profiles with treatment information and survival outcomes. De identification is a defined legal standard and the claim may well be correct. It is also being made about a dataset that is unusually rich in exactly the attributes that make de identification hard: a tumour type, a molecular profile, a treatment sequence and a survival interval together describe a person very narrowly, and a million data points per patient is a great deal of surface. The method used, expert determination or removal of the enumerated identifiers, is not stated, and for a dataset of this shape the distinction is material.

A second question follows from the outcomes themselves. Survival and treatment response are not generated by the laboratory. They accrue after the test, elsewhere, which means there is an acquisition pathway bringing longitudinal outcome data back and linking it to a profiled patient. Nothing published describes it.

Ask which de identification method applies and who certified it, and how outcome data is obtained and linked.

An explicit published role statement, the required instrument alongside it, and a compliance disclosure most laboratories keep private.

The company states directly, in an annual notice addressed to referring physicians, that under the health privacy rule it is a healthcare provider and a covered entity. That removes the ambiguity this index usually has to resolve by inference. Obligations attach to it directly and are enforceable by the regulator rather than flowing through a customer contract. It publishes its notice of privacy practices in full, and maintains a separate website privacy statement so that the two audiences, patients and site visitors, are not served by one document doing both jobs badly.

The same annual notice does something else worth recording, because it belongs to a compliance domain this index has not previously touched and which matters more for laboratories than for software vendors. It sets out the company's position on the physician self referral law and the anti kickback statute, stating that it is its policy to comply with both and explaining the restrictions on financial relationships between referring physicians and laboratories. Diagnostic laboratories have historically been a principal enforcement target in that area, because the referral relationship creates obvious inducement risk. A laboratory that puts its position in writing to the physicians who refer to it, annually, is managing that exposure openly.

A referring institution should still confirm the terms governing data returned into its own systems, and any separate arrangement covering research participation, which the privacy notice governs only in part.

No SOC 2, HITRUST, ISO 27001 or equivalent information security attestation and no trust centre were located.

The laboratory accreditation distinction applies as it does throughout this lane. The company operates accredited clinical laboratories, and that accreditation is demanding, real, and concerned with analytical validity, proficiency and specimen handling rather than information security. The estate at issue extends well beyond the laboratory bench: a whole exome and whole transcriptome pipeline, a research platform served to biopharmaceutical customers over the web, and a linked clinico genomic database.

A route exists and is recorded rather than relied upon. The company listed on a United States exchange in 2025, so its annual report must describe processes for assessing and managing cybersecurity risk, board oversight and accountable management. That item was not retrieved in this pass and nothing here rests on it.

One feature of this business raises the stakes above the ordinary. The company sells access to a web based research platform containing linked molecular, treatment and survival data on hundreds of thousands of patients. That platform is by design reachable by external organisations, which makes access control, authentication, query logging and export restriction load bearing controls rather than background hygiene. A de identified dataset that can be queried flexibly by outside parties invites a different threat model from a static file transfer.

Ask what attestation exists, how research platform access is granted and monitored, and what controls limit export and cohort narrowing.

A distinction buyers can easily miss, and the company documents it rather than blurring it: MI Cancer Seek holds FDA approval with companion diagnostic indications, while the Caris AI Insights signatures delivered in the Molecular Tumor Board Report are research use only. The company's own materials caution against inferring that all available services are FDA approved and publish a summary of what falls in and out of the approved labeling. Clear scoping of what is approved and what is not is precisely what this axis rewards.

No governance framework, subgroup performance analysis or bias evaluation was located, and two distinct problems apply to the signatures on this record.

The first is shared with all clinical genomics. Interpretation and molecular signatures rest on reference data and cohorts drawn disproportionately from patients of European ancestry, so performance for other populations is an open question rather than a demonstrated equivalence. Nothing published reports signature performance by ancestry.

The second is specific to signatures trained on observed outcomes, and it is the more interesting one. These models are trained on real cohorts with matched survival, one reported at close to thirteen thousand lung cancer cases and another at over three thousand breast cancer cases. A model trained that way learns the outcomes of the treatment decisions that were actually made, at the institutions that made them. It therefore encodes practice patterns alongside tumour biology. Where a group of patients was systematically treated less aggressively, diagnosed later, or had less access to a therapy, their worse outcomes enter the training data as signal, and a model that predicts poor response for such patients may be reproducing a care disparity while appearing to describe a biological one. Because the prediction then informs treatment selection, the effect can compound rather than wash out.

That is measurable. Ask for signature performance stratified by ancestry, insurance status and treating site type, and for evidence that predicted poor response reflects biology rather than the historical treatment those patients received.

Training cohorts are quantified, which few vendors in this index do, with sizes stated for the metastasis risk work across many thousands of lung and breast cancer cases with matched survival outcomes, and the signatures are named individually rather than presented as one platform. Held at C on a distinction that is regulatory rather than technical and that a clinician is very likely to lose.

The company's sequencing assay holds full premarket approval, which is a high bar and is correctly credited elsewhere on this record. The signatures layered on top are a different thing: one is described by the company as clinically validated, which is a claim the company makes rather than a status a regulator conferred.

A clinician reading a single report containing both an approved companion diagnostic result and a predictive signature may reasonably assume the same authority stands behind both, because the two arrive on the same page under the same letterhead with the same apparent finality. It does not, and the report should make that clear.

This index has recorded certification non transfer in several forms, and this is the sharpest, because the two categories are not merely adjacent in a company's portfolio, they are adjacent on one document a clinician acts on. What is otherwise absent is the method behind any signature, the features used, validation design and whether it was internal or external, performance figures, and versioning. Ask which components fall within the approved assay's scope, and for the validation evidence behind each signature separately.

No named record system integration, interoperability standard or interface documentation was located, and the workflow described is document based.

The primary output is a molecular tumour board report, a rich document combining sequencing results, biomarker findings and predictive signatures. Delivered as a document it is readable and complete, and it is well suited to the multidisciplinary meeting it is designed for. It is poorly suited to everything that happens afterwards. A report that arrives as a document sits in the chart as an attachment rather than as discrete results, so the individual biomarker findings are not available to decision support, cohort identification, trial matching or any later query. The next clinician must open and read it.

That has a specific consequence in oncology. Molecular findings retain relevance for years, and a patient's actionable alteration may become actionable again when a new therapy or trial appears. If the finding is not stored as structured data, nothing can surface it later, and the patient depends on someone remembering.

The research platform raises the inbound version of the same question. Linking treatment and survival outcomes to profiled patients requires clinical data flowing in from institutions, and nothing published describes how that is collected, at what cadence or through what interface.

Ask whether results can be delivered as structured discrete data, what standards are supported, and how outcome data reaches the platform.

No hosting location, region, tenancy, retention or subprocessor information was located.

The physical testing side is straightforward, since profiling is performed in the company's own accredited laboratories on shipped specimens. The computational estate is where the questions sit, and it has two distinct parts that should be assessed separately rather than together.

The clinical pipeline produces a patient's report and is subject to the laboratory's regulated quality system. The research platform is a web application served to external biopharmaceutical and academic customers, containing linked molecular, treatment and survival data on hundreds of thousands of patients. Those have different users, different access models and different risk profiles, and how they are separated technically determines what an external customer's query can reach.

Retention is the second question and it is a long one in this domain. Sequencing data is expensive to generate and remains informative indefinitely, and the research value of the dataset depends on keeping it. How long raw sequence, derived profiles and linked outcome records persist, and whether a patient can obtain deletion from the research dataset having consented to a clinical test, is not addressed publicly.

Residual specimens are a third, governed separately from data.

Ask where each environment runs, how the clinical and research estates are separated, the retention schedule across specimen, sequence and linked outcomes, and the subprocessor register.

Structure is partially disclosed rather than priced. The Molecular Tumor Board Report containing the AI signatures is stated as available at no additional cost when ordering MI Cancer Seek, which is a real commercial disclosure. Test pricing itself is not published and runs through diagnostic reimbursement channels; biopharma data partnerships are separately negotiated. As a public company (NASDAQ: CAI) aggregate revenue is disclosed in filings, but test level pricing is not.

Precisely bounded: solid tumor molecular profiling for adult and pediatric patients aged one and older with previously diagnosed malignancy, with signature level specificity by cancer type including colon, breast, ovarian, pancreatic, and lung.