Freenome

Machine learning classifiers on multiomic signal are what convert a blood draw into a cancer result, and the company is expanding into fragment level deep learning and a methylation foundation model. Held back from A for the same reason as GRAIL: this is a wet lab assay plus a classifier, and the assay chemistry is as load bearing as the model. The vendor's own reporting of assay upgrades improving limit of detection and signal to noise makes that explicit.

Bounded by the clinical pathway it sits in, and the boundaries are clear.

The test is ordered by a clinician for average risk colorectal cancer screening. The validation cohort was built with a pre colonoscopy arm specifically to mirror that intended use population. A positive result does not diagnose anything; it routes the patient to diagnostic colonoscopy, which remains the definitive procedure.

What strengthens this beyond the vendor's own framing is that an independent body has now defined the role. The American Cancer Society's updated colorectal screening guideline includes blood based testing as a category of screening option, which means the place of this test in the pathway is set by the guideline rather than asserted by the manufacturer.

The honest limitation belongs here too, because it shapes what the oversight model has to catch. Reported sensitivity for advanced precancerous lesions sits between roughly 14 and 22 percent depending on assay version and analysis, against far higher sensitivity for established cancer. Colonoscopy's value is partly that it removes precancerous lesions before they progress, and a blood test at this level of precancer sensitivity does not substitute for that. The company reports the precancer figures openly rather than leading only with the cancer number.

Method and evidence are well described. The reported numbers are not consistent across venues, and that is what holds this to a B.

What is disclosed is substantial: a multiomics approach combining genomic, epigenomic and proteomic signal with machine learning classifiers, an expansion into fragment level deep learning and a methylation foundation model, a pivotal validation published in JAMA, and successive assay versions with performance reported for each.

The problem is which number a reader should use. The first version was reported at roughly 81 percent sensitivity for colorectal cancer and 14 percent for advanced precancerous lesions at 90 percent specificity, census adjusted, as published. The next generation version has been described in company materials at 85 percent and 22 percent. A securities filing in July 2026 reports the updated test at 80.4 percent and 18.2 percent on a weighted basis, with 41.9 percent for advanced lesions carrying high grade dysplasia.

Those figures may all be defensible descriptions of different analyses across different cohorts and weightings. Nothing published makes clear which analysis each figure comes from, and the marketing figure is the highest of the set while the securities figure is the lowest.

Check what is in the numerator before carrying any percentage from this category into a comparison. Establish the cohort, the weighting and the specificity threshold for each figure quoted.

The data rights are disclosed clearly, in the wrong document, and they are broad. Commercial agreements state that the company will have access to real world patient and multimodal molecular data generated through commercialisation, and that this data will feed machine learning models to improve its platform across multiple cancer indications.

That is a substantial secondary use grant: a person screened for one cancer contributes multiomic data that trains classifiers for a pipeline spanning many further indications, and the sample they gave for a single screening decision becomes an input to products they will never see. The disclosure is real and it is creditable that it was made at all.

The problem is where it appears, since it is stated in commercial announcements written for investors and partners rather than in a patient facing notice explaining what happens to a blood sample and the data derived from it. The people whose material is being described are the only audience not addressed.

Three searches, one targeting the company's own legal pages directly, did not surface a privacy notice, a notice of privacy practices, or any statement of specimen or data retention, and the same approach applied to the closest comparable company in this index returned its equivalent documents immediately, so this is a difference in what is published rather than in how hard it was looked for. Establish how long the specimen is kept, how long sequencing data is kept, what consent covers model training, and whether a patient can withdraw.

Strong on design and venue. The pivotal PREEMPT CRC validation was published in JAMA in 2025, and blood based testing was subsequently included in the American Cancer Society's updated colorectal cancer screening guideline, which is independent recognition by a guideline body rather than vendor assertion. That combination is rare in this index and it is what earns the A.

The reported performance needs care, because at least three different figure sets are in circulation for what is broadly the same product line.

The first version was reported at roughly 81 percent sensitivity for colorectal cancer and 14 percent for advanced precancerous lesions at 90 percent specificity, census adjusted, as published. Company materials describe a next generation version at 85 percent and 22 percent from a 966 sample head to head validation. A securities filing in July 2026 reports the updated test at 80.4 percent and 18.2 percent on a weighted basis, with 41.9 percent for advanced lesions carrying high grade dysplasia, and cites modelling projecting 7.7 percent more life years gained and 9.5 percent fewer colorectal cancer cases and deaths per 100,000 screened against the first generation test.

These are plausibly different analyses of different cohorts under different weightings. What is not published is which analysis produced which number, and the highest figures appear in commercial materials while the lowest appear in a securities filing.

This index grades evidence design and venue and does not re verify underlying results. It does expect a reader to be told which analysis a percentage came from, and that is the gap here.

The data rights are disclosed clearly, in the wrong document, and they are broad.

The commercial agreements state that Freenome will have access to real world patient and multimodal molecular data generated through commercialisation, and that this data will feed machine learning models to improve its platform across multiple cancer indications. That is a substantial secondary use grant: a person screened for colorectal cancer contributes multiomic data that trains classifiers for a pipeline spanning lung and more than ten further indications.

The disclosure is real and it is creditable that it was made. The problem is where it appears. It is stated in commercial announcements written for investors and partners, not in a patient facing notice explaining what happens to a blood sample and the data derived from it.

Three searches, one targeting the company's own legal pages directly, did not surface a privacy notice, a notice of privacy practices, or any statement of specimen or data retention. The same search approach applied to the closest comparable company in this index returned its HIPAA notice and its provider data protection addendum immediately, so this is a difference in what is published rather than in how hard it was looked for.

Establish before contracting: how long the specimen is kept, how long sequencing data is kept, what consent covers model training, and whether a patient can withdraw.

Nothing published was located that sets out the posture.

The status is not ambiguous. A clinical laboratory that performs testing on referral and bills for it is a covered entity under HIPAA in its own right, not merely a business associate, and covered entities with a direct treatment relationship are required to produce and distribute a notice of privacy practices. Freenome operates the laboratory that runs the test.

Across three differently phrased searches, including one aimed at the company's own legal pages, no notice of privacy practices, business associate agreement, role statement or review cadence was found. None of the three routes to a higher grade in this index is taken.

The comparison matters for fairness in both directions. The same searches run against the nearest peer in this category returned a full notice of privacy practices written for patients and a published provider facing data protection addendum containing the contractual commitments. That establishes both that such material is findable when it exists and that publishing it is achievable in this category.

Absence of a retrieved document is not proof that none exists, and a buyer should simply ask. But a patient handing over a blood sample for genomic analysis cannot currently read what governs it.

The distinction that governs this grade is the same one that applies across laboratory based diagnostics, and it is easy to get wrong.

A CLIA certified laboratory operating at high complexity is subject to a demanding quality regime, and that regime is the right assurance that a test result is analytically valid and that the laboratory is competent to produce it. It is not an information security attestation. CLIA says nothing about how the sequencing data generated is protected once it exists.

No SOC 2 report, HITRUST certification, ISO 27001 or trust centre was surfaced across three searches, including one aimed at the company's legal and security pages.

The holding at stake is significant and growing. This is multiomic sequencing data on a screening population, with an explicit commercial intention to accumulate real world molecular data to train models across more than ten cancer indications. A dataset assembled for that purpose is a durable asset rather than a transient processing artefact, and an independent security attestation is the ordinary way a buyer verifies it is protected.

The company completed its Nasdaq listing on 20 July 2026 and trades as FRNM, so securities disclosure obligations now apply where previously they did not. Periodic filings may surface a fuller picture than a private company publishes, and this axis is worth re reading once the first annual report is filed. The grade reflects what is published today.

Positioning is stated precisely and without overclaim. The Premarket Approval application for SimpleScreen CRC was submitted in August 2025 and remains under FDA review, with the company anticipating a decision in 2026 and stating an intent to file a supplemental PMA for the next generation assay. The test is consistently described as under review rather than approved, and the lung product is identified as a laboratory developed test.

One real representativeness step, and the subgroup question left open in a disease where it matters more than most.

The published pivotal results were census adjusted to reflect the composition of the US population, which is a genuine generalisability correction rather than a claim, and it was done in a peer reviewed venue. Very little in this index goes that far.

What is not published is performance broken down by group, and colorectal cancer is a poor place for that gap. Incidence and mortality differ substantially across racial groups in the United States, early onset disease is rising, and a screening test intended to reach people who currently go unscreened will be deployed disproportionately into exactly those populations. Whether sensitivity, specificity and advanced lesion detection hold across subgroups is the question that determines whether this test narrows a screening gap or widens it.

Credit where it is due on candour: the low sensitivity for advanced precancerous lesions is reported plainly rather than buried, and it is the least flattering number the company has. A vendor that publishes its weakest metric alongside its strongest is behaving correctly even when the metric is poor.

Method and evidence are well described, and the reported numbers are not consistent across venues, which is what holds this to a middling grade rather than a high one. What is disclosed is substantial: a multiomics approach combining genomic, epigenomic and proteomic signal with machine learning classifiers, an expansion into fragment level deep learning and a methylation foundation model, a pivotal validation published in a major journal, and successive assay versions with performance reported for each.

The problem is which number a reader should use. Sensitivity for the primary indication and for advanced precancerous lesions is reported at materially different levels across the published validation, company materials and a securities filing, on different bases and at different weightings. Those figures may all be defensible descriptions of different analyses across different cohorts, and nothing published makes clear which analysis each figure comes from.

One observation belongs on the record because it is factual rather than an imputation of motive: the marketing figure is the highest of the set and the figure in the securities filing, where the legal exposure for overstatement is greatest, is the lowest. That pattern is worth noticing wherever it appears.

Check what is in the numerator before carrying any percentage from this category into a comparison, and establish the cohort, the weighting and the specificity threshold behind each figure quoted to you.

Prescription in, report out, with nothing published about how either travels.

No native electronic health record integration, structured result delivery, discrete result coding or laboratory interface is described. For a screening test intended for repeat use in an average risk population, that matters in the same way it does for any recurring screen: a result that lands outside the record is a result a practice has to track manually alongside colonoscopy intervals and other screening due dates.

What exists is distribution reach rather than integration depth, and the two should not be confused. The commercial partner is described as leveraging existing infrastructure to streamline access to nearly 400 health systems. That describes how many organisations can be sold to, not how the test is ordered inside their systems or how the result reaches the chart.

The pathway consequence is worth stating. A positive result must reliably produce a diagnostic colonoscopy, and the follow through rate is the number that determines whether earlier detection becomes earlier treatment. Interoperability is what makes that follow through reliable rather than dependent on manual chase.

A laboratory service, so the architecture is total transfer by design. The blood sample and everything derived from it leave the ordering practice and come to rest with the laboratory. There is no configuration in which the data stays with the buyer.

No hosting provider, region, residency statement or subprocessor list is published.

One wrinkle is specific to this record and a buyer should understand it. The commercial arrangements make a milestone payment conditional on successful technology transfer to the commercial partner, which means the assay, and plausibly elements of the data pipeline around it, are contractually intended to move to a different corporate entity. A residency answer given today by the developer may not describe where the work runs once the partner is operating it at scale.

Ask who will physically run the laboratory process at the point of purchase, in which facilities, and whether the answer changes on approval. The counterparty question raised on the commercial axis has a data location counterpart, and it is not addressed anywhere public.

No published test pricing, which is unsurprising for a product still under review, and unusually detailed disclosure of the commercial structure behind it.

Start with the fact that matters most to a buyer: the counterparty will probably not be Freenome. The exclusive license covering US commercial rights was signed with Exact Sciences in August 2025, and Abbott acquired Exact Sciences in March 2026, so on approval the colorectal test is commercialised by Abbott. There is one counterparty here, not two, and later company releases refer to it as Abbott formerly Exact Sciences. Freenome retains a narrower product of its own, SimpleScreen CRC paired with its SimpleScreen Lung laboratory developed test, offered solely to patients eligible for both screenings. A purchaser therefore needs to know which product configuration it is buying before it knows who it is contracting with.

The deal economics are public in a way almost nothing in this index is: 75 million dollars upfront, up to 700 million in milestones, and a royalty range stated as 0 to 10 percent depending on profitability.

One disclosure goes further than any pricing page could, and it is the reason for the B. The milestone structure publicly encodes a performance bar. The 100 million dollar payment tied to approval of the next generation test was stated as contingent on benchmarks including at least 19 percent sensitivity for advanced precancerous lesions and at least 83 percent overall colorectal cancer sensitivity, with a reduced payment where performance falls below.

In July 2026 the company reported that the updated test met all primary and secondary endpoints, at 18.2 percent and 80.4 percent, and stated in the same release that the milestone payment will be set at 70 million dollars.

Read those two things together, because they describe two different bars. The study cleared the endpoints the company set for itself. The same figures sit just below the thresholds the contract set, and the milestone is now stated at the lower figure. Only the first of those appeared in the headline. A buyer rarely gets to see a contractually defined performance threshold, a reported result against it, and the commercial consequence, all in public.

The company completed its Nasdaq listing on 20 July 2026 under the ticker FRNM. That will produce periodic filings in time, but none has yet been filed as a public operating company, so this grade does not yet reflect any listing advantage.

Clearly bounded: blood based screening in average risk colorectal cancer populations, with a lung indication in development. The company states explicitly that no single technology detects every cancer, which is an unusually candid scope limitation for this category.