Modella AI

The product is the AI, in one of its purest forms in this index. PathChat is a pathology foundation model pretrained on histology image and image text data combined with a custom trained multimodal large language model, and the company describes itself as building generative and agentic AI for biomedicine. There is no services or hardware layer. If the model is removed there is no product.

Explicitly a co-pilot. PathChat is positioned for human in the loop clinical decision making, providing assistive discussion at region, slide, and case level plus ancillary test suggestions and slide level triage, with the pathologist retaining the diagnosis.

The generative interface raises a distinct oversight question competitors do not, since a conversational model can produce fluent but wrong reasoning, and the company frames the tool as support rather than an autonomous reader, which is the appropriate posture for a language model in diagnosis.

Unusually transparent because the underlying science is published, not just described. PathChat originates in peer reviewed work from the Mahmood Lab at Mass General Brigham published in Nature, and the architecture is stated specifically: a pathology foundation model pretrained on histology image and image text datasets feeding a custom trained multimodal large language model. A buyer can read the methodology in the primary literature rather than taking a datasheet on faith, which is the strongest form of model transparency available and rare among commercial vendors.

Nothing identifies any party in the operating chain: no hosting arrangement, no sub processor list, no retention position and no statement on whether customer images contribute to model development. Three questions are specific to this record and the first applies across the whole lane.

Whole slide image files routinely embed a photograph of the slide label, which commonly carries the accession number and often the patient's name, so slides are identified records by default and any assumption that images shared for development or support are anonymous needs confirming rather than presuming.

The second follows from the product design: the assistant is described as working from a microscope, a slide viewer or a smartphone, and capture by phone changes the exposure materially, because the image transits a personal device, may persist in a camera roll or a backup, and sits outside whatever controls the laboratory applies to its own systems. That is a real accessibility advantage for a pathologist at a scope and it needs a stated position. The third follows from the acquisition.

Tissue images and the features derived from them are valuable research assets to a pharmaceutical owner, and nothing states whether customer or partner images contribute to model development, whether derived representations flow to the parent's research programmes, or what separates commercial deployment data from research use inside the group. Ask who strips embedded identifiers and when, the smartphone position, and whether anything reaches the parent.

The foundational research is strong and peer reviewed in Nature, and independent review literature discusses PathChat's diagnostic reasoning capability. But that evidence is about model capability, not about clinical deployment outcomes at health systems. The clinical grade PathChat DX is pre market, and the generally available PathChat 2a is research use only, so there is no body of real world clinical operational evidence yet. Graded on what exists in clinical practice today rather than on research promise.

Converted from Not Rated. No stewardship framework or data governance terms were located, and the prior note was right that hallucination controls and failure mode disclosure matter particularly for a generative model in pathology.

Three specific questions apply here.

Whole slide image files routinely embed a photograph of the slide label, which commonly carries the accession number and often the patient's name. Slides are therefore identified records by default across this lane, and any assumption that images shared for development or support are anonymous needs confirming rather than presuming.

The company describes its copilot as working from a microscope, a slide viewer or a smartphone. Capture by phone changes the exposure materially: the image transits a personal device, may persist in a camera roll or a backup, and sits outside whatever controls the laboratory applies to its own systems. That is a real accessibility advantage for a pathologist at a scope, and it needs a stated position.

The acquisition announced in January 2026 raises the third. Tissue images and the features derived from them are valuable research assets to a pharmaceutical owner. Nothing published states whether customer or partner images contribute to model development, whether derived representations flow to the parent's research programmes, or what separates commercial deployment data from research use inside the group.

Ask who strips embedded identifiers and when, what the position is on smartphone capture and retention, and whether any customer derived data or representation reaches the parent organisation.

Converted from Not Rated. No business associate commitment, availability statement or scope description was located, and the position is more complicated than for a conventional software vendor.

The product has been positioned for research use with a clinical version described as pre market, so for much of what the company does the health privacy rule is not the operative framework. Research use of identifiable tissue at a covered entity runs on institutional review board approval and an authorisation or waiver, and a research collaboration agreement rather than business associate terms is often the correct instrument.

The January 2026 acquisition adds a dimension that has no equivalent for an independent vendor. The counterparty is now part of a pharmaceutical group, and a pharmaceutical company receiving identifiable patient material from a provider is a familiar transaction with its own established framework, but it is not the framework that governs a software supplier. Which relationship a laboratory is entering, software vendor, research collaborator, or data provider to a pharmaceutical sponsor, determines the agreement, and the answer may differ by engagement within the same company.

A further question follows from the product itself. A generative copilot that drafts pathology reports produces text that may enter the diagnostic record. Where that text lands, who is accountable for it, and whether the drafting service is treated as part of the record system are questions for the agreement rather than the product documentation.

Establish which entity signs, what framework governs the specific engagement, and how generated report text is handled.

Converted from Not Rated. No SOC 2, ISO 27001 or equivalent attestation and no trust centre were located. The prior note attributed this to early stage status with a research use product, which was reasonable at the time and is now the wrong frame.

The company was acquired by a major pharmaceutical group in January 2026. That changes the analysis in a way this index has recorded before and should be applied rather than assumed either way. Certifications are held by legal entities, and a large parent's information security programme does not automatically extend to a recently acquired subsidiary's systems, product infrastructure or development practices. Integration takes time, often years, and during it the acquired estate frequently continues to run as it did. The opposite error is also available: assuming nothing has changed when the parent may have imposed its own controls immediately.

The right question is therefore narrow and answerable. Which entity's security programme governs the product a customer would deploy, has it been independently examined, and if the parent's certifications are cited, do they name this subsidiary and its systems in scope.

What sits behind the question is significant. This is a generative system that ingests tissue images and produces reports, with capture possible from a smartphone, and a model development agent operating over biomedical image data. Ask what independent security examination exists for the product estate specifically, and whether an attestation covering the acquired business is available or planned.

PathChat DX holds FDA Breakthrough Device Designation, granted January 2025. Breakthrough Device Designation is a review process commitment that grants prioritized FDA interaction, not a clearance or approval, and confers no authorization to market the device clinically, a distinction the index applies consistently to Genomate and Nucs. The generally available PathChat 2a is explicitly research use only. So there is FDA engagement but no cleared clinical product yet. Credit is due for the company stating plainly that the designation does not imply clearance, which not every Breakthrough holder does.

Converted from Not Rated. No governance framework, evaluation methodology or bias analysis was located, and two structural facts now bear on this axis more heavily than the missing documents.

The first is ownership. In January 2026 the company announced its acquisition by a major pharmaceutical company, to advance artificial intelligence driven oncology research and development. That is a legitimate and unsurprising transaction, and it changes what a laboratory is evaluating. A pathology foundation model reads tumour tissue and produces characterisations that inform which patients have which tumour features, and those features are frequently the basis on which oncology therapies are selected. The company that now owns the model also develops and sells such therapies. Nothing is alleged here. It is simply a relationship a customer should see, and should ask how it is managed: whether model development is insulated from therapeutic programmes, what governance separates the two, and who decides what the models are optimised to detect.

The second is how the models are trained. The company's whole slide foundation model is described in the published literature as fine tuned using several hundred thousand synthetic captions generated by the company's own earlier generative copilot. Training one model on the output of another from the same family means the second inherits the first's errors and blind spots as training signal rather than as noise to be corrected, and no external reference enters the loop at that step.

Independent review literature on pathology foundation models flags poor generalisability in zero shot testing, explainability, hallucination risk in generative tools and fairness across populations. Each applies here and none is addressed publicly.

The underlying science is published rather than described, which is the strongest form of model transparency available and rare among commercial vendors. The system originates in peer reviewed work from a named academic laboratory at a named institution, published in a leading journal, and the architecture is stated specifically as a pathology foundation model pretrained on histology image and image text datasets feeding a custom trained multimodal language model.

A buyer can read the methodology in the primary literature rather than taking a datasheet on faith, and can consult the reviewers' judgement rather than only the company's. Held below the top grade on the distinction this index applies to every vendor whose evidence comes from an academic origin: a published method is not a published product.

The paper describes what was built and evaluated in a research setting, and no accuracy figure, evaluation methodology or error characteristic was located for the commercial system a laboratory would deploy, which may differ in training, tuning and interface. No warranty, indemnity or remediation commitment attaches either.

The output type raises the stakes, because a generative assistant in pathology produces prose a pathologist may incorporate into a report, and a fabricated or subtly wrong statement about morphology reads exactly like a correct one. Ask for the deployed system's performance against the published evaluation, the hallucination rate, and what the assistant does when the image does not support an answer.

Converted from Not Rated. No named integration with a laboratory information system, image management system or scanner was located, and the prior note's reasoning was right that integration here is image driven rather than record driven.

What the company does describe is more interesting than an absence and cuts both ways. The copilot is presented as working from a microscope, a slide viewer or a smartphone. That is a deliberate design choice and its advantage is real: a pathologist can consult the tool at the scope without the laboratory first completing a digital pathology programme, which for most laboratories is a multi year capital project. It removes the usual barrier to adoption entirely.

The same choice is the interoperability gap. A tool reached by photographing a microscope field operates outside the laboratory information system, so nothing it produces is captured in the systems of record: no order, no audit trail linking the consultation to a case, no record that the tool was used, and no route for its output into the report except by a person retyping or pasting it. For a diagnostic aid that matters, because the question of what informed a diagnosis is one a laboratory may later need to answer.

One collaboration announced in 2025 with another vendor in this lane points toward platform integration, which would change this picture. That work was described as research rather than a shipped product.

Ask what production integrations exist with laboratory systems, how a consultation is recorded against a case, and how generated report text reaches the record.

Converted from Not Rated. No hosting, region, tenancy or residency terms were located, and the prior note identified the right question: generative model deployment raises where inference runs and whether slide data leaves the institution.

That question is sharper for this product than for a conventional pathology algorithm. Whole slide images are very large, which normally pushes analysis toward the laboratory. But a generative vision language system with a conversational interface is architecturally the opposite: it typically runs on substantial accelerated infrastructure that few laboratories operate, and the smartphone and microscope capture paths the company describes only make sense if the image is sent somewhere for processing. So the likely answer is that images do leave, and the location and terms are unstated.

Two further components need separate answers. The company describes an agent for automating model development over biomedical image data, which implies an environment where customer or partner data is used to build models rather than merely analysed. And following the January 2026 acquisition, whether infrastructure remains the subsidiary's or has moved into the parent's estate is unresolved, which matters because the parent is a pharmaceutical company with its own research systems.

Ask where inference runs for each capture path, whether images are retained after a session and for how long, what environment the model development agent operates in and on whose data, and whether any component now runs on parent infrastructure.

No product pricing is published, consistent with a research use tool and a pre market clinical version. The company is venture backed and private, so there is no listing driven financial disclosure either. Commercial terms are opaque at this stage.

Concentrated on anatomic pathology, where the generative co-pilot approach is applied across region, slide, and case level reasoning with ancillary test suggestions, making it broad within pathology rather than tied to a single tumor type. That generality is the point of a foundation model, distinguishing it from single indication pathology detectors, but coverage is confined to pathology and the clinical version is not yet available.