Panakeia Technologies

The model is not merely the product, it is the only thing that could produce this output at all, which is a stronger form of centrality than most records granted this grade.

Every other pathology vendor in this index reads morphology and reports what a trained pathologist could in principle see: a tumour boundary, a Gleason pattern, a cell count. Panakeia infers molecular status, microsatellite instability and mismatch repair deficiency, from a routine haematoxylin and eosin image. No pathologist can look at that slide and determine that status, because the signal is not one human vision was ever able to extract. The model is not accelerating human judgement, it is performing a task that previously required a separate laboratory assay.

That distinction has a commercial consequence worth stating. The company sells no platform, no viewer, no scanner and no image management system. The product is an inference from an image the laboratory already has, delivered in minutes, consuming no additional tissue and requiring no workflow change. Remove the model and there is nothing at all, not even a workflow benefit.

The research use platform extends the same capability across more than 30 cancer indications for trial screening and biomarker discovery, which is the model sold directly as a capability rather than wrapped in anything.

Graded A. Alongside Deep Bio this is one of two records in this session where the artificial intelligence is the entire company, and the technical claim here is the more ambitious of the two.

The oversight problem here is different in kind from the rest of this lane, and nothing published addresses it.

Every other pathology model in this index produces an output a pathologist can check by looking. A tumour boundary, a Gleason pattern, a cell count are all visible in the tissue, so the reviewer can compare the model's claim against what they see and disagree on the evidence. Panakeia infers molecular status from morphology, and no pathologist can verify that inference by examining the slide, because the signal is not visible to human vision. A reviewer receiving a microsatellite instability prediction has no independent means of assessing it short of ordering the very assay the product exists to displace.

That makes the confidence handling and the intended use boundary the whole safety question, and neither is described. Nothing published states whether results are returned with a confidence measure, whether low confidence cases are flagged for confirmatory molecular testing, what the recommended action is on a discordant or borderline result, or how a laboratory should treat the output in a patient where immunotherapy eligibility turns on it.

Two elements point the right way. The company positions the tools as supporting pathologists and oncologists in treatment decisions rather than determining them, and selection for the regulator's AI Airlock programme means real world use is being evaluated under supervision, which is oversight imposed externally.

The independent HER2 comparison is relevant here too, since agreement of 65 percent among models tells a reader that outputs in this class of task should be treated as probabilistic rather than definitive.

Graded C.

The scientific claim is documented to an unusual standard through the literature, while the engineering behind it is not described.

What is well established is what the models do and on what basis. Inputs are routine brightfield haematoxylin and eosin images already generated in hospital care. Outputs are molecular biomarker status, specifically microsatellite instability and mismatch repair deficiency for colorectal cancer, delivered in minutes. The multi omic framing is stated concretely as DNA, RNA, protein and metabolite changes inferred from morphology, and 2026 conference work extended this to pathway activation data. The company describes its capability as diagnostic grade explainable artificial intelligence, which is a claim about interpretability that is not elaborated anywhere located.

The peer reviewed publications are the real transparency mechanism. A validation study in npj Digital Medicine, a proof of concept in Communications Medicine and a conference abstract with a full author list are all documents a technically literate reader can obtain and assess, containing method and performance detail that no vendor web page would carry. That is a materially different disclosure route from publishing a specification, and for a scientific claim it is the better one.

What is absent from vendor material is the engineering: architecture, training corpus size and composition, evaluation methodology outside the published studies, versioning and update cadence. The explainability claim in particular is asserted rather than described, and for a model whose output cannot be checked by eye, how it explains itself is the thing a pathologist most needs to understand.

Graded B.

Institutional collaborators are named and traceable, the development corpus is not described, and the models are unambiguously the company's own.

What is traceable is the research and validation chain. The University of Leeds and Leeds Teaching Hospitals NHS Trust led the published clinical validation. The Industrial Centre for Artificial Intelligence Research in Digital Diagnostics consortium explored the platform's utility, with a named professor of pathology commenting publicly. The Friends of Cancer Research digital pathology project and the regulator's AI Airlock programme are both named. Conference abstracts carry full author lists spanning company scientists and clinicians at multiple institutions. A reader can identify every organisation that has independently touched this technology, which is more institutional traceability than any other record in this lane offers.

What is not described is the training data supply chain, which for this company is the critical one. Models inferring molecular status from morphology require images paired with sequencing or immunohistochemistry results at scale, and nothing names the source of that paired corpus, the agreements governing it, or whether it came from the same institutions that later validated the products, which would raise an independence question worth understanding.

On components the position is implicitly in house. No third party model, framework or pretrained foundation model is acknowledged, and the company describes proprietary insights into cancer biology alongside its artificial intelligence, suggesting domain knowledge embedded in the approach rather than a general model applied to pathology.

Graded C.

The best evidence base of any vendor built in this session, and its distinguishing feature is that the company keeps submitting to evaluation it does not control.

The anchor is a blinded multi site clinical validation led by the University of Leeds and Leeds Teaching Hospitals NHS Trust, published in npj Digital Medicine in January 2026 and described as the largest real world blinded multi site validation of artificial intelligence enabled molecular profiling from routine diagnostic images. Academic led, blinded, multi site and real world is the strongest available study design short of a randomised trial, and publication in a peer reviewed digital medicine journal puts it in front of reviewers with no stake in the outcome. Earlier blinded multi site results were presented at the ASCO gastrointestinal symposium in January 2025 with a named author list spanning company scientists and clinicians from multiple institutions, and proof of concept across more than 30 cancer indications appeared in Communications Medicine in 2024.

The unusual part is the comparative work. The breast product was included in an independent study benchmarking ten artificial intelligence models against each other and against three pathologists on HER2 expression, and the published result is not flattering to anyone: models agreed with each other 65 percent of the time, models and pathologists agreed 65 percent, and the pathologists agreed with each other 70 percent, with low expression cases hard for both. A vendor that enters a head to head comparison capable of producing that finding, and that also joins a multi stakeholder agreement and reliability project, is behaving in a way this index almost never encounters.

Selection for the United Kingdom regulator's AI Airlock programme adds regulator supervised real world evaluation.

Graded A.

Nothing published describes the training corpus, and on this record that omission is larger than on any other in the pathology lane because of what the models must have been built from.

A model that infers microsatellite instability from a haematoxylin and eosin image can only be trained on slides paired with the molecular ground truth for the same patients. That means image plus sequencing or immunohistochemistry result plus, for validation, outcome. Assembling that linkage at the scale needed for more than 30 cancer indications means access to substantial archives of patient tissue and their molecular results, gathered originally for clinical care or for research programmes with their own consent frameworks.

Nothing published names the source institutions, states the consent or research authorisation basis, describes the de identification standard, or explains what data access agreements govern the holdings. For a company whose entire technical proposition rests on that paired corpus, the absence of any account of its provenance is the substantive gap.

The academic partnerships mitigate by implication. Validation led by a university and National Health Service trust, and participation in a research consortium and a multi stakeholder agreement project, all operate under governance frameworks, and a company embedded in that environment is unlikely to have assembled data carelessly. Implication is not disclosure, and the studies are not the same thing as the development corpus.

On customer data nothing states whether images processed clinically or for pharmaceutical customers are retained or inform continued development.

Graded D on the absence, with the note that a published statement of data provenance would move this several grades and would cost the company nothing it has not already told its ethics committees.

No published position was located, and the data flows on this record run in two directions that both matter.

Outward, the clinical products process patient slide images in National Health Service laboratories, so patient data is handled under United Kingdom data protection law and National Health Service information governance, and nothing published describes the arrangement, the retention position or where processing occurs.

Inward is the more substantial question and it is specific to this business. The company's research platform sells biomarker profiling across more than 30 cancer types to pharmaceutical customers, and a model that infers molecular status from morphology can only have been built by pairing large numbers of slide images with the molecular results for the same patients. That paired corpus is the company's core asset and it necessarily came from patients whose tissue and sequencing data were collected for clinical care or research elsewhere. Nothing published describes the source institutions, the consent or research authorisation basis, the de identification standard applied, or any data access agreements.

The academic collaborations partly answer this by implication rather than disclosure, since research led by a university and National Health Service trust would proceed under research ethics approval, and a research consortium partnership operates under its own governance. That is inference about specific studies, not a published position covering the company's data holdings.

No business associate style agreement, protected data handling summary or retention statement was located, and nothing addresses cross border transfer for research customers outside the United Kingdom.

Graded D.

No published security posture was located. No trust centre, no service organisation control report, no information security management certification, no penetration testing statement, no vulnerability disclosure policy, no subprocessor list and no incident notification commitment were found.

The absence is partly explained by the company's centre of gravity, which is scientific rather than commercial. Panakeia publishes in peer reviewed journals, presents at oncology conferences and participates in regulatory and multi stakeholder programmes, and its public communication is organised around evidence rather than around enterprise procurement. That is a legitimate emphasis for a company at this stage and it does not answer a hospital security review.

Two credentials exist by implication and neither is an information security posture. UKCA marking as a medical device entails conformity assessment covering software lifecycle, and cybersecurity forms part of that assessment for software devices, so material exists in the technical file. The company also markets regulatory expertise in developing software as a medical device as a service to research customers, which indicates a quality management system behind it, though no certification is named.

The research business raises the sharper version of the question. Pharmaceutical customers sending trial patient images for biomarker screening will require documented security assurance as a condition of engagement, so the artefacts almost certainly exist. As with several records in this session, the assurance is being provided privately to individual counterparties rather than published.

Graded D on published evidence rather than on any judgement about the underlying engineering.

A clear, honestly bounded United Kingdom position with a genuinely forward looking regulatory engagement, limited by geography.

Two clinical products, for colorectal and breast cancer, are UKCA marked and clinically deployed in the United Kingdom. The wider platform covering more than 30 cancer indications is research use only, and the company states that split consistently in the same breath rather than letting registered status colour the whole catalogue. That discipline is worth crediting, because a company with two registered products and thirty research ones has an obvious temptation to blur them and does not.

The distinguishing element is selection for the Medicines and Healthcare products Regulatory Agency AI Airlock programme, a regulatory sandbox supporting real world evaluation of artificial intelligence medical devices. That is a regulator working through how software of this kind should be assessed, with this product as a live case, and participation means the company is engaging with the hardest regulatory question its technology poses rather than waiting to be asked. Very little in this index sits inside a regulatory sandbox.

What holds this below the top grade is reach. No European CE marking under the in vitro diagnostic regulation and no United States clearance were located, so clinical availability is one country while the research platform is global. For a technology whose value proposition is displacing molecular assays, the markets where those assays are most expensive are precisely the ones where clinical registration is absent.

Graded B.

The best position on this axis in the entire session, and it was earned by submitting to comparisons the company could not control.

The substantive act is participation in an independent multi vendor study of HER2 expression assessment, benchmarking ten artificial intelligence models against each other and against three pathologists. The published finding is unflattering across the board: models agreed with one another 65 percent of the time, models and pathologists agreed 65 percent, the pathologists agreed with each other 70 percent, and low HER2 expression was difficult for both. A vendor entering a study capable of producing that result, and letting it be published, is doing the opposite of what commercial incentive suggests. It also gives a buyer the single most useful piece of information available in this category, which is that outputs of this kind vary between tools and between humans and should be treated accordingly.

Participation in the Friends of Cancer Research digital pathology agreement and reliability project is the same behaviour institutionalised, a multi stakeholder effort examining exactly the comparability question vendors usually avoid. The blinded multi site validation led by an external university and hospital trust adds independent testing across institutions, which is how population and laboratory generalisation is actually demonstrated.

What is still missing is conventional disclosure. No model card, no training population composition, and no performance breakdown by patient ancestry, age or institution were located, which matters because morphology to molecular inference could plausibly vary with tissue processing and with population genetics.

Graded B.

Nothing published addresses liability, and the regulatory position plus one specific programme do more work here than a contract clause would.

UKCA marking establishes an accountable manufacturer under a device regulation regime with post market surveillance obligations, adverse event reporting duties and recall powers, so a hospital using the colorectal or breast product within its registered indication has a regulatory relationship rather than only a commercial one. Selection for the regulator's AI Airlock programme adds supervised real world evaluation, which means performance in practice is being observed by a party with authority rather than only by the vendor.

The exposure is nonetheless sharper than for most pathology software, for the reason set out on the autonomy axis. A false negative microsatellite instability result may mean a patient is not offered immunotherapy they would benefit from, and may mean Lynch syndrome goes unidentified, which has consequences for the patient's relatives as well as the patient. Those harms are delayed, diffuse and unlikely to be traced back to the prediction that caused them, and unlike a morphology call there is no way for the reviewing pathologist to have caught the error by looking.

Nothing published states whether confirmatory molecular testing is recommended in any circumstance, what the intended use text says about reliance, or where responsibility sits between vendor, laboratory and treating oncologist. No indemnity, limitation or performance warranty was located.

The published validation performance is the substitute a buyer actually has, and it is a stronger one than most records offer.

Graded C.

The workflow claim is strong and the integration specifics are undocumented.

What is stated repeatedly and matters commercially is that the products require no additional laboratory testing and no changes to existing workflow. They consume routine brightfield images the laboratory already produces for diagnosis, so adoption does not require new tissue, new stains, new instruments or a new process. For a National Health Service pathology department that is the difference between a pilot and a project, and it is the strongest interoperability property a diagnostic can have: fitting into what already happens rather than asking for something new.

The research side implies further integration capability, with the company offering to add other stains or omics data to enrich research scope and to validate customers' own predictive biomarkers, which requires ingesting external data in varied forms.

What is absent is every technical specific. No image management system or laboratory information system is named, no interoperability standard is cited, no application programming interface is published, and nothing describes how a result reaches the pathologist or the report. That last point is the material gap for a clinical product: a molecular status prediction is only useful if it lands in the signed case and the oncology conversation, and nothing describes whether it arrives as a structured result in the laboratory system or as a separate output somewhere else.

No scanner compatibility statement was located either, which matters when the input is a whole slide image and scanner variation is a known source of model performance drift.

Graded C.

Nothing published. No hosting arrangement, no region, no residency option, no subprocessor list, no retention position, no export terms and no availability commitment were located.

The architecture is not even inferable here, which is unusual. The product consumes whole slide images and returns a molecular status prediction, and that could run on premise inside a hospital, in a vendor operated cloud receiving uploaded images, or embedded inside a partner image management system. Those are materially different arrangements and nothing indicates which applies.

The United Kingdom clinical deployment makes the question concrete rather than theoretical. National Health Service organisations operate under specific information governance expectations for patient data, including where processing occurs and who holds it, and any laboratory running these products has satisfied itself on that point. The answer exists for those customers and is published for none, so the next hospital repeats the work.

The research business adds a second and separate path. Pharmaceutical customers using the platform for trial screening across more than 30 indications are sending patient images from their own trial sites, potentially across borders, and nothing describes where that processing happens or under what terms.

Availability deserves brief mention. A tool positioned on returning in minutes what otherwise takes weeks becomes part of a diagnostic pathway once adopted, and no uptime or continuity position is published.

Graded D on the absence rather than on any evidence of a problem.

Nothing is published. No price, no unit of charge, no licensing structure, no implementation fee, no contract term and no minimum was located in vendor or third party material.

The two market structure makes the unit question interesting and unanswered. On the clinical side a National Health Service pathology laboratory running the colorectal product would most naturally be charged per case or per test, and the commercial argument is directly comparable to the assay it displaces, since microsatellite instability testing by immunohistochemistry or sequencing has a known cost and turnaround. That comparison is the whole business case and the company publishes neither side of it. On the research side, pharmaceutical customers using the platform for trial screening across more than 30 indications are buying something closer to a service or a data capability, priced entirely differently.

The displaced cost framing is what makes the silence conspicuous rather than ordinary. A product whose value proposition is delivering in minutes what currently takes weeks and a separate laboratory test invites an explicit cost comparison, and a published per case figure set against the cost of the assay would be the most persuasive thing this company could show a laboratory director.

One structural signal exists without a number. Clinical deployment is in the United Kingdom through National Health Service pathways, where procurement is institutional and prices are typically negotiated rather than listed, which explains the absence without removing it.

Graded D.

Two clinical indications and a research platform spanning more than 30, with the gap between them stated honestly rather than blurred.

Clinically the coverage is narrow and deep: colorectal cancer for microsatellite instability and mismatch repair deficiency, and breast cancer, both UKCA marked and deployed in United Kingdom practice. Within colorectal the indication is well chosen, because microsatellite instability status determines immunotherapy eligibility and identifies Lynch syndrome, which carries implications for the patient's relatives, so a single inference does two clinically distinct jobs.

The research platform is where breadth sits, profiling biomarkers across more than 30 cancer types for patient screening and clinical trials, sold to pharmaceutical customers for biomarker discovery, validation of predictive signatures and prospective patient selection. The company also offers to add other stains or omics data to enrich research scope and to independently validate customers' own predictive biomarkers through its clinical partner network.

The honest part is that the company does not let the research breadth colour the clinical claim. Material consistently states which products are registered and which are research use only, in the same sentence, rather than implying that 30 indications are clinically available.

Geographically the clinical footprint is the United Kingdom alone, with no European or United States clinical registration located, so the research platform is global while clinical deployment is one country.

Graded B: exceptional depth in two indications, genuine breadth confined to research, coverage limited by geography.