BostonGene

Higher than most in this lane because the models carry more of the product: the company positions an AI foundation model of tumor and immune biology, with Kassandra performing cell deconvolution to reconstruct the tumor microenvironment from sequencing data, an inference task rather than a measurement.

Held back from A because Tumor Portrait remains a DNA and RNA exome assay run in the company's own regulated laboratory, so the wet lab is still load bearing, consistent with how GRAIL, Freenome, Caris, and GeneDx are graded.

A real external control exists and it does not reach the parts a buyer is asking about.

The control is the laboratory itself. A clinical laboratory report of this kind is issued under a regulated quality system and signed out by a qualified laboratory director, and that requirement is imposed and inspected by someone other than the company. Any result reaching a physician has passed through it. That is stronger oversight than most vendors in this index can point to, and it is worth stating before the gaps.

The gaps concern the components this record is actually indexed for. The report incorporates a computational reconstruction of the tumour microenvironment and a trial ranking the company describes as fully automated. Nothing published states how either is reviewed: whether a person examines the deconvolution output against the underlying data, whether the trial ranking is reviewed before it reaches the ordering physician or is generated and delivered without human sight, what a reviewer is shown about why a trial ranked where it did, or what happens when a clinician disagrees.

The distinction matters because the two outputs carry different consequences. An erroneous biomarker call is checkable against the sequencing data. A trial ranking that quietly omits a suitable study produces no signal at all, and the physician has no way to know what was not shown.

Ask who signs out the computational components, whether the trial ranking is reviewed or automatic, and whether a physician can see the full candidate set rather than the ranked subset.

Detailed where it matters most and thin on a second product, and the asymmetry is worth seeing.

The deconvolution algorithm is disclosed properly. It is named, its method class is stated as a decision tree machine learning approach to cellular deconvolution from bulk expression data, its training construction is described as artificial transcriptomes built from a large set of expression profiles, its output resolution is quantified at up to fifty one distinguishable cell types, and the whole is published in a peer reviewed journal a reader can consult. That is transparency in the form this axis asks for: a customer can read how it works rather than being told that it works.

The trial matching capability is disclosed very differently. It is described as a fully automated scoring algorithm relying on more than eighty dynamic parameters. A parameter count is not a description of any of them, and it is the kind of figure that reads as sophistication while remaining unfalsifiable, because nobody outside can check a count of things none of which are named. The same company that published its deconvolution method in detail has published nothing comparable here.

Also absent across both: model versioning and update practice, and any statement of how a change to either algorithm is validated and communicated, which matters because both operate inside a regulated laboratory whose reports are clinical documents.

Ask what the eighty parameters are and how they are weighted, and how model versions are controlled and disclosed when a report is issued.

Nothing identifies any party in the chain: no hosting arrangement, no sub processor list, and no retention position covering specimens, raw sequence or derived data was located in two passes. What the company receives is broad by design and permanent by nature.

A tumour specimen retrieved from the treating pathology department, a normal sample from the patient providing germline sequence, and accompanying clinical documentation including progress notes and pathology reports, from which whole exome and transcriptome sequence is produced.

Germline sequence is durably identifying and it concerns the patient's relatives as well as the patient, which is a property no consent form fully resolves, because the people it discloses about were never parties to it. Nothing states whether residual specimen or sequence is retained after reporting, whether patient derived data contributes to model development or reference building, who inside the company can reach an identified record, or what a patient can request in the way of deletion.

The model development question is not hypothetical: the deconvolution algorithm is reported as trained on a large set of expression profiles, and the company runs research collaborations with cancer centres and pharmaceutical partners. One clarification belongs on the record because it is easy to misread. The laboratory accreditations do not answer any of this. They govern how the test is performed, not what becomes of the data afterwards. Ask for the retention schedule across specimen, sequence and derived data, and the deletion path.

Clinical and analytical validation of the combined RNA and DNA exome assay was published in Communications Medicine, part of the Nature portfolio, across more than 2,200 tumors, reporting high reproducibility and clinical actionability in 98 percent of cases. Publishing a validation study of the assay that underpins the AI, in a peer reviewed venue with a stated cohort size, is the standard this axis asks for. Corroborated by nine abstracts accepted at ASCO 2026 developed with MD Anderson, Weill Cornell, and the Parker Institute.

The material held is among the most sensitive in this index and almost nothing is published about its stewardship.

What the company receives is broad by design. A tumour specimen retrieved from the treating pathology department, a normal sample from the patient providing germline sequence, and accompanying clinical documentation including progress notes and pathology reports. From that it produces whole exome DNA and RNA sequence. Germline sequence is durably identifying and it concerns the patient's relatives as well as the patient, which is a property no consent form fully resolves.

Nothing located states retention periods for specimens, raw sequence or derived data; whether residual specimen or sequence is retained after reporting; whether patient derived data contributes to model development or reference building; who inside the company can access an identified record; or what a patient can request in the way of deletion.

The model development question is not hypothetical. The deconvolution algorithm is reported as trained on a large set of bulk expression profiles, and the company runs research collaborations with major cancer centres and pharmaceutical partners. Whether clinical testing data flows into those activities, and under what basis, is the single question a patient or an ordering institution would most want answered.

The laboratory accreditations do not answer any of this. They govern how the test is performed, not what becomes of the data afterwards.

Ask for the retention schedule across specimen, raw sequence and derived data, the position on secondary research use and model training, and the deletion path.

A scoping determination with an unusually clear answer, and one pathway that needs separate handling.

This company operates a high complexity clinical laboratory that receives specimens on a physician's order and returns a clinical report. A laboratory of that kind is a health care provider transmitting health information electronically, which makes it a covered entity in its own right rather than a business associate of anyone. That is a stronger position than most records in this index, because the obligations attach directly rather than flowing through a contract, and they are enforceable against the company by the regulator.

The operating footprint supports it. The company publishes its accreditation and licensing documents openly, including its national certificate of accreditation, its New York State licence, and separate clinical laboratory licences and permits for several other states that operate their own regimes. Multi state licensure of that kind is not obtainable without a functioning compliance programme.

The pathway that needs separate treatment is the biopharma side. The company runs research collaborations and partnerships in which tumour and immune profiling informs trial design and biomarker discovery. Patient derived molecular data moving to a pharmaceutical partner is a different transaction from returning a result to the ordering physician, and it requires either de identification or a patient authorisation. Nothing located states which applies, what patients are told, or whether participation is separable from receiving the test.

No notice of privacy practices was located either, which a covered entity is required to provide. Ask for it, and ask what governs molecular data used in partner research.

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

The distinction this index has had to draw repeatedly applies squarely here. The company publishes a full and genuine set of accreditation documents: a national certificate of accreditation, a New York State licence, and clinical laboratory licences and permits for several further states. Those are demanding to obtain and hold, and they are real. They examine laboratory quality, analytical validity, proficiency testing, personnel qualification and specimen handling. They examine no part of information security, and a reader scanning a page of credentials will not see the difference.

That matters more here than for a laboratory alone, because the product is not only the assay. It is a computational platform holding tumour and germline sequence data, clinical notes and pathology reports, delivering results through a customer portal, and feeding partner research programmes. None of that estate is covered by a laboratory accreditation.

One practice deserves explicit credit and is uncommon. The company's own certification page states that credentials are reviewed regularly, that status can change with renewal cycles or issuing body updates, and that a reader should contact the company to confirm current standing for formal compliance purposes. Most vendors publish a static list of marks with no such caveat, and this index has separately found lists containing standards the vendor could not plausibly hold. Telling a reader the list may age is the opposite failure mode and a good sign.

Ask whether an information security attestation exists, what it covers, and how the portal and sequence data estate are protected.

Regulatory position is stated plainly: the Tumor Portrait platform is approved under CLIA, CAP, and the New York State Department of Health, the standard framework for a laboratory developed test, and New York State approval is a meaningfully higher bar than CLIA alone. No FDA clearance or approval is claimed, and none is implied. Held back from A because no FDA pathway or timeline is disclosed for products where one might eventually apply.

Better than most in this category, and the reason is that the method is published and one of its limits is stated openly.

The cell deconvolution algorithm is described in a leading cancer biology journal, and the clinical and analytical validation of the assay beneath it appears in a Nature portfolio journal across more than two thousand tumours. That is peer reviewed disclosure of the method rather than a claim about it.

More unusually, the company's chief medical officer has publicly acknowledged a specific limitation: the deconvolution approach does not assess spatial heterogeneity or temporal dynamics from a single biopsy. He then described the mitigations, multi regional and longitudinal sampling and a separate spatial platform. Naming what your method cannot see, in a trade publication rather than a footnote, is the behaviour this axis exists to reward.

What holds it at B is the training corpus, which the company also describes and which raises the question nothing answers. The model is reported as trained on artificial transcriptomes constructed from a large set of bulk expression profiles. Synthetic training relocates the representativeness question rather than removing it: an artificial mixture is only as representative as the reference cell profiles used to build it, and those profiles come from real cohorts with their own composition. If the references under represent particular ancestries, disease states or treatment histories, the model's reconstruction will be least reliable for exactly those patients, and nothing published reports performance by subgroup.

Ask what the reference profiles are drawn from, and for deconvolution accuracy broken down by patient population rather than by cell type.

One algorithm is published to an academic standard and the other is described by a parameter count, and the contrast inside a single company is what makes this record useful. The deconvolution algorithm is disclosed properly: it is named, its method class is stated, its training construction is described as artificial transcriptomes built from a large set of expression profiles, its output resolution is quantified at a stated number of distinguishable cell types, and the whole is published in a peer reviewed journal a reader can consult.

That is transparency in the form this axis asks for, because a customer can read how it works rather than being told that it works. The trial matching capability is described as a fully automated scoring algorithm relying on more than eighty dynamic parameters.

A parameter count is not a description of any of them, and it belongs with the other claim shapes this index records as reading like sophistication while remaining unfalsifiable, since nobody outside can check a count of things none of which are named. The company that published its deconvolution method in detail has published nothing comparable here, which establishes that the disclosure is possible and was not made.

Also absent across both: model versioning and update practice, and any statement of how a change to either algorithm is validated and communicated, which matters because both operate inside a regulated laboratory whose reports are clinical documents. Ask what the parameters are and how they are weighted, and how versions are controlled when a report is issued.

The workflow is documented clearly and it is largely manual, which is the finding.

The company describes the process plainly: the ordering team completes a requisition form and submits it with supporting documents, the company retrieves the tumour specimen from the pathology department, the patient provides a normal sample at home or a local laboratory, and results are delivered through a customer portal accessed by an emailed activation link. That is an honest and useful description, and the specimen logistics are genuinely well handled, including retrieval from the pathology department rather than placing that burden on the ordering clinician.

What is absent is the electronic path. No result delivery into the ordering physician's record system was evidenced, no standards based interface, and no named integration with any record platform. For a diagnostic company this is the axis's central question, because a genomic report that lives only in a separate portal sits outside the record where oncology decisions are documented and where the next clinician to see the patient will look. It also means the report is not available to any downstream system that might surface it later.

The consequence compounds for the trial matching capability. A ranked trial list delivered to a portal is seen once, by whoever opens it. Eligibility changes as a patient progresses, and a static document in a separate system cannot follow that.

Ask whether results can be delivered as structured data into the record, what standards are supported, and whether any named integration exists in production.

Nothing was located on hosting, region, tenancy, retention or subprocessors for the computational platform or the results portal.

One structural feature makes the question more pressing than it would be for a domestic laboratory. The company operates a joint venture in Japan with named corporate partners, and runs research collaborations with academic centres and pharmaceutical companies internationally. Genomic sequence moving across borders engages transfer rules that vary considerably, and several jurisdictions treat genetic data as a special category with its own conditions, independent of any general privacy analysis. Japan in particular operates its own personal information regime with specific requirements for transfers abroad.

Nothing published states where sequence data is held, whether analysis for a Japanese patient occurs in Japan or elsewhere, how the joint venture's data estate relates to the parent's, or whether a customer can elect where processing happens.

The laboratory side is comparatively clear, since physical specimens are processed in the company's own accredited facility. The computational side is where the data actually lives and persists, and it is undescribed.

Ask where sequence and derived data are stored and processed, what separates the joint venture's estate from the parent's, what transfer mechanism supports any cross border flow, and for the subprocessor register covering the analysis pipeline and the portal.

No public pricing. Contact the vendor. Two commercial surfaces with different structures: clinical testing ordered by providers, and separately negotiated biopharma research and development partnerships such as the announced Takeda collaboration. No rate card published for either.

Clearly bounded: oncology and immune mediated disease, spanning clinical care through the provider facing Tumor Portrait test and drug development through biopharma partnerships. The company states both surfaces explicitly rather than implying general applicability.