QuantHealth

Simulation is the entire offering. The Large Real-World Drug Model is described as a clinical trial foundation model trained on real world data from more than 350 million patients and large volumes of therapeutic datapoints, used to predict individual patient response and aggregate those into a simulated trial result. There is no consulting or data brokerage layer being dressed as AI; the model output is the deliverable.

Structurally low risk on this axis because nothing the model produces touches a patient. Output informs protocol design decisions made by clinical development and operations teams: inclusion and exclusion criteria, comparator arms, endpoints, indication selection.

The real exposure is different from clinical AI and worth naming plainly: a simulation that wrongly predicts a subgroup will not respond could exclude that population from a trial, which shapes who the eventual therapy is studied in and labeled for. Human sponsors retain every decision, but the influence on trial design is substantive rather than advisory.

Scale is disclosed specifically while method is not. The company names the model, its training base of more than 350 million patients, and its versioning, and reports headline accuracy figures including 88 percent for Phase 2 and 83.2 percent for Phase 3 predictions. What is missing is how accuracy is defined and validated, whether predictions were made prospectively before readout or fitted retrospectively, and what counts as a correct prediction of a trial outcome. Vendor materials state predictions are made without sponsor data at 85 percent accuracy on primary endpoints, but no peer reviewed methodology paper was located to substantiate the claim.

One supplier is named openly, which deserves credit because naming a data supplier at all is more than most companies training on real world data will do: a healthcare data partner that aggregates records from provider organisations into research datasets and describes them as de identified, with the partnership presented as improving dataset structure and closing demographic gaps. What is not addressed is everything that determines whether the disclosure is meaningful.

The training corpus is described variously across different releases in terms of hundreds of millions of patients, a smaller count of individuals, millions of mechanistic variables and tens of billions of drug patient interactions, and only one supplier is named against all of it, so the remaining provenance is unaccounted for and a reader cannot tell whether the named partner supplied most of the corpus or a fraction of it.

No de identification standard is stated either, and the choice matters at this scale: whether safe harbour or expert determination was applied changes what residual identifiability remains, and across hundreds of millions of longitudinal records the re identification literature is not reassuring. Nothing describes the consent basis under which the underlying patients' records entered a commercial training corpus, which is the question this axis exists to ask. The sponsor side is separately unaddressed, since a customer running a simulation may supply its own trial data and no handling, retention or segregation terms were located.

Volume of reported activity is substantial but the evidence is almost entirely vendor generated. The company reports simulating more than 350 trials across 23 therapeutic areas with up to 90 percent predictive accuracy, having grown from a reported 100 simulations at 85 percent accuracy, and cites a case where simulation identified a likely non responder subgroup for a top ten pharmaceutical company with 31.4 million dollars in claimed savings.

None of these figures were independently verified in this review, the pharmaceutical partner is unnamed, and no peer reviewed publication was located. Strategic investment from a major pharmaceutical venture arm and an Accenture backed funding round are meaningful third party signals of credibility, but they are commercial rather than scientific validation. Contrast with Unlearn.AI, whose comparable claims sit in peer reviewed literature and a regulatory qualification.

Converted from Not Rated. Provenance is better disclosed than the prior review found, and the substance behind the disclosure is still thin.

What is now evidenced: at least one data source is named openly, a healthcare data partner that aggregates records from provider organisations into research datasets and describes them as de identified. The partnership is presented as improving dataset structure and closing demographic gaps. Naming a data supplier at all is more than most companies training on real world data will do, and it deserves credit.

What is not addressed is everything that determines whether that disclosure is meaningful. The training corpus is described variously as 350 million patients, 100 million individuals, five million mechanistic variables and 30 billion drug patient interactions across different releases, and only one supplier is named against it, so the remaining provenance is unaccounted for. No de identification standard is stated: whether safe harbour or expert determination was applied changes what residual identifiability remains, and at a scale of hundreds of millions of longitudinal records the re identification literature is not reassuring.

Nothing describes the consent basis under which the underlying patients' records entered a commercial training corpus, which is the question this axis exists to ask.

The sponsor side is separately unaddressed. A customer running a simulation may supply its own trial data, and no handling, retention or segregation terms for that were located.

Ask which de identification standard was applied, and what happens to sponsor supplied data.

Converted from Not Rated. The prior note was right that customers are trial sponsors rather than provider organisations, and the position is more nuanced than a simple absence.

The federal health privacy rule reaches identifiable health information held by covered entities and their business associates. Here the customers are pharmaceutical sponsors, which are not covered entities, and the training corpus is described by its supplier as de identified. Properly de identified data sits outside the rule entirely. On that reading the axis does not apply, in the way it does not apply to a vendor holding no health data at all.

The distinction from that cleaner case is why this stays at C rather than moving up. For a vendor whose data is not health data, the non application is structural and verifiable. Here health data is the entire product, and the rule's non application depends wholly on the quality of a de identification step that is asserted rather than described. Which standard was applied, by whom, and with what expert determination behind it are unstated, and at this scale that is the load bearing question rather than a technicality.

Two further gaps. Where a sponsor supplies its own clinical trial data for a simulation, that data may well be identifiable and the contracting position for it is not published. And no statement addresses obligations flowing from the upstream provider organisations whose records reached the corpus through a data partner.

Ask which de identification standard applies, and what governs sponsor supplied identifiable data.

Converted from Not Rated, and confirmed by a second differently phrased search using the company's own model and product names.

No SOC 2 report of either type, no ISO 27001, no HITRUST, no trust centre and no penetration testing statement was retrieved. A listing on a major cloud marketplace exists, which requires passing that provider's listing process and is not a security attestation.

The customer base makes the omission conspicuous rather than routine. The company reports working with eight of the twenty largest pharmaceutical companies and four of the five largest. Those organisations run some of the most demanding vendor security assessments in any industry, because a sponsor's unpublished trial design, endpoint strategy and asset prioritisation are among the most commercially sensitive documents it holds. A simulation platform sees all of it, often years before anything is public.

So the sensitivity here is commercial rather than clinical, which is worth stating precisely. The corpus itself is described as de identified, meaning a breach would not primarily expose patients. It would expose which molecules a sponsor is prioritising, which trials it expects to fail, and which populations it plans to exclude. That is material non public information about listed companies.

Absence of a retrieved document is not proof none exists, and agreements with pharmaceutical partners of that size will have required security review. None of it is public.

Ask for the attestation and its period, and for the segregation terms between sponsor engagements.

Converted from Not Rated. The prior note's contrast is the right one and it has not moved.

No device pathway applies and none is claimed. Trial simulation informs a sponsor's design decisions rather than diagnosing or treating anyone, so no clearance is required and its absence is not a deduction.

What makes this a C rather than higher is that a regulatory route does exist for methodology of this kind and has not been taken. A company operating in adjacent territory obtained a qualification opinion from the European Medicines Agency for its methodology, which is the mechanism by which a regulator examines and endorses a modelling approach used in trial design. That precedent establishes both that the route is available and that a peer judged it worth pursuing. No equivalent qualification, advice procedure or regulatory engagement was located here.

The distinction matters because of what the product is used for. Simulation output influences protocol design, endpoint selection and population definition on trials that regulators will later assess. A methodology that has been qualified carries weight in that conversation; one that has not is a vendor tool the sponsor must defend itself.

One forward looking statement needs checking rather than carrying. In July 2025 the company said third party validations were under way with pharmaceutical partners and through external peer review. A year on, no published result was located. Establish whether those have completed and what they found.

Converted from Not Rated. The prior note raised the concern as a possibility. The company's own published case study confirms it is the marketed use.

No governance framework, subgroup performance analysis or bias disclosure was located.

The mechanism is straightforward. A model trained on historical real world data inherits the demographic composition of that data, including whoever was under treated, under diagnosed or absent from the record entirely. Applying it to predict who will respond to an investigational therapy, and then designing a trial around that prediction, risks narrowing the enrolled population along the same lines the source data was already skewed by.

What moves this from hypothetical to concrete is how the capability is sold. A published account of work for a large pharmaceutical customer describes the simulation identifying a likely non responder subgroup that could otherwise have diluted and weakened the trial's results, with savings quantified. Excluding predicted non responders is presented as the value delivered. That is a legitimate design technique with a long history in enrichment strategies, and it is also precisely the operation that decides who is studied and therefore who the eventual label describes.

One genuine counterweight deserves recording. A named data partnership is framed explicitly around addressing disparities in clinical trials and closing demographic information gaps, which is the right instinct applied at the input layer.

Ask how excluded subgroups are characterised, and whether exclusion recommendations are audited for demographic pattern.

Scale and identity are disclosed specifically while the thing that would make the claims assessable is not. The company names the model, states a training base in the hundreds of millions of patients, publishes its versioning, and reports headline accuracy figures separately for two trial phases, which is more specific than a single platform number and acknowledges that the two are different prediction problems.

What is missing is the one question that decides whether any of it means anything for trial prediction: whether the predictions were made prospectively before readout or fitted retrospectively to trials whose outcomes were already known. Those produce very different numbers from the same model, and a retrospective figure is close to uninformative because the outcome was available during development.

Nothing states which was done, nor how accuracy is defined, what counts as a correct prediction of a trial outcome, or how validation was performed, and no peer reviewed methodology paper was located to substantiate a further claim about predictions made without sponsor data. No warranty, indemnity or remediation commitment attaches.

The stakes make the distinction practical rather than academic, since a sponsor using these predictions may decline to run a trial, which is a decision no subsequent readout ever corrects. Ask whether the reported accuracy is prospective, how many predictions were locked before readout, the definition of a correct prediction, and the false negative rate on trials that succeeded.

Converted from Not Rated. This is a genuine category non application rather than a gap, and the company does not pretend otherwise.

The platform operates on aggregated real world datasets inside sponsor development workflows. It does not sit in a care setting, does not receive orders, does not return results to a clinician and has no reason to connect to a provider's record system. Grading it down for lacking clinical integration would penalise a business model rather than describe it, which this index has ruled against where an axis mismatches a vendor's position.

Declining to overclaim is the creditable part. Vendors in adjacent positions frequently assert integration breadth as a trust signal whether or not it is relevant, and this one does not.

The interoperability that does exist runs upstream and through a partner rather than directly. A named data collaborator aggregates records from many provider organisations into structured research datasets, so clinical record data reaches the model as a curated product rather than through any integration this vendor builds or maintains. That is a supply relationship, and it means integration quality and breadth are properties of the partner.

On the delivery side, availability through a major cloud marketplace gives sponsors a procurement and provisioning path, which is a distribution channel rather than a technical integration.

What is not described, and would matter to a sponsor, is how simulation outputs move into the systems where protocols are actually authored and managed.

Ask how outputs reach trial design and operations systems, and in what format.

Converted from Not Rated. No hosting, region, tenancy or residency terms were located.

One delivery route is documented. The foundation model is offered through a major cloud provider's marketplace, which gives a sponsor a procurement path and indicates the platform runs on that provider's infrastructure. As with other records in this pass, naming a cloud provider is not a residency answer: it says nothing about region, whether processing stays within a given jurisdiction, or how customer environments are separated.

The marketplace listing does raise a question worth resolving directly, because the answer changes the exposure profile materially. Software offered through a cloud marketplace can be delivered from the vendor's own environment or deployed into the customer's account. For a pharmaceutical sponsor, a deployment inside its own tenancy would keep unpublished trial designs and asset priorities within its perimeter, while a hosted service would not. Nothing published distinguishes them.

Jurisdiction is the second open item. The company is headquartered in Israel with a New York presence, and its customers include the largest pharmaceutical companies in Europe and the United States, several of which operate under data transfer constraints. No transfer mechanism, regional hosting option or subprocessor list was located.

Separation between sponsor engagements is the third. A platform serving eight of the twenty largest pharmaceutical companies is holding competing sponsors' confidential design work simultaneously.

Ask which deployment model applies, in which region, and how sponsor tenancies are isolated.

No pricing is published. The company quantifies value in customer savings terms, citing 31.4 million dollars delivered to one unnamed pharmaceutical company, which is a return claim rather than a cost disclosure and cannot be independently checked. The engagement model appears to be enterprise partnership with pharmaceutical sponsors rather than licensed software, and terms are not disclosed.

Broad by design and broader than comparable vendors. Deployment spans a reported 23 therapeutic areas including oncology, immunology, cardiometabolic disease, and gastroenterology, and capability covers trial design optimization, indication selection, drug repurposing, asset evaluation, and probability of technical success modeling across development phases.

That breadth contrasts sharply with Unlearn.AI, whose coverage is bounded by disease specific twin generators and continuous endpoints. Breadth here reflects a general purpose model rather than per disease qualification, which is a strength commercially and a weakness evidentially.