BenevolentAI

The biomedical Knowledge Graph and the reasoning built on it are the company. The disclosed method is specific: a network of contextualized scientific data covering genes, proteins, diseases and compounds and the relationships between them, normalized from literature, patents, genetics, chemistry and clinical trials, with machine learning finding connections between known facts and AI based reasoning extrapolating connections not previously stated anywhere.

That is a claim no database or literature search product could make, and the company's platform enhancements have consistently been model work, including a next generation Knowledge Graph powered by natural language processing and large language models introduced to predict novel targets from scientific literature.

The oversight structure is the strongest disclosed in this category, and it is external rather than self certified. Targets predicted by the platform do not count as delivered until AstraZeneca has independently experimentally validated them and selected them for portfolio entry, a gate the partner controls and the vendor cannot grade itself against.

Company and partner scientists are described as working side by side to frame and test hypotheses, and knowledge generated in the disease programmes is fed back to improve subsequent predictions. Held at B rather than A only because no internal confidence thresholds or published prediction reliability figures were located, so what a buyer can rely on is the strength of the external gate rather than a disclosed internal standard.

Substantially more detailed than most, helped by years as a listed company that required describing platform work to investors. Disclosed specifics include the composition and normalization approach of the Knowledge Graph, the introduction of a next generation graph powered by natural language processing, the adoption of large language models to predict therapeutic targets from literature, target identification tools extended to targets best prosecuted via alternative modalities, and a distinct product for target progressibility assessment covering druggability, selectivity and competitor and patent landscape.

Held at B because no peer reviewed methods paper establishing platform predictive performance was located, and because the delisting has ended the reporting stream that produced most of this detail, so the disclosure is comprehensive but increasingly historical.

The corpus is described by source type and one major partner contribution is named, which answers more of this axis than most of the category. The knowledge graph is assembled from published literature, patents, genetics, chemistry and clinical trial data together with licensed and partner contributed datasets, so a reader knows what kinds of material sit in it, and it is not built from identifiable patient records.

The stewardship question that arises here is therefore partner data governance rather than patient data, and it is a real one: a named pharmaceutical partner's proprietary genomics, chemistry and clinical data was integrated into the graph under a collaboration. Once proprietary material is integrated into a graph and models are built over it, the question this index has learned to ask of any derived artefact applies.

A graph enriched by one partner's data is not obviously separable afterwards, and derived representations do not unlearn when a collaboration ends. Nothing published states what happens to integrated partner material at termination, whether the enriched graph is used to serve other partners, or how contributions are segregated if at all. Nothing else is enumerated either, with no hosting arrangement or sub processor list located. Ask what happens to integrated partner data at termination, whether the enriched graph serves other partners, and how contributions are segregated.

The most externally adjudicated evidence record in this category, and it earns the grade on independence rather than volume. Multiple novel targets discovered on the platform were experimentally validated by AstraZeneca and selected into its drug discovery portfolio across idiopathic pulmonary fibrosis, chronic kidney disease, systemic lupus erythematosus and heart failure, over a collaboration running from 2019 through at least 2024, which means a large pharmaceutical company repeatedly put its own laboratory resource behind the predictions and repeatedly confirmed them.

Separately the platform first identified baricitinib as a COVID-19 treatment, which subsequently received FDA approval, and no other vendor in this category can point to an approved medicine traceable to a platform prediction, though buyers should note this was repurposing of an existing molecule rather than de novo discovery. Candour is equally strong: BEN-2293 was stopped after Phase IIa with the company stating plainly it would not invest further, so the failures are on the record alongside the wins.

Not applicable in the provider sense and rated accordingly rather than penalized. The Knowledge Graph is assembled from published literature, patents, genetics, chemistry and clinical trial data together with licensed and partner contributed datasets, not from identifiable patient records.

Worth noting that AstraZeneca's proprietary genomics, chemistry and clinical data was integrated into the graph under that collaboration, so the stewardship question that arises is partner data governance rather than PHI.

Not applicable. The company is a United Kingdom based research organization contracting with pharmaceutical partners rather than a business associate of covered entities handling protected health information.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre or security page exists. Two instruments are published and they cover different things, which is worth separating.

The privacy notice is a personal data instrument governing website visitors and business contacts rather than partner research data, and under this index's reading that does not answer the question a pharmaceutical counterparty is asking. Within its own scope it is better than most: it names the specific transfer mechanisms relied on, including standard contractual clauses and international data transfer agreements, rather than gesturing at compliance, and it states that access is limited to those with a genuine business need who are trained for the processing. Naming the transfer instrument is a small thing that distinguishes a drafted policy from a generated one.

The more relevant document is a separate published privacy notice and set of terms governing the company's research partner programme, covering controller and processor roles, authorised user data, confidentiality and third party software components. That is the second vendor in this category found to publish the agreement governing its collaboration programme, and as with the first it is not a security attestation but it does let a counterparty read the confidentiality and data handling terms before making contact.

Two cautions. The company shares a name with unrelated charitable giving organisations whose privacy and security pages surface prominently in searches here; none belong to it. And it left public markets during 2025, so any filing based disclosure is historical rather than current, and its security posture should be established directly rather than inferred from the period when it reported publicly.

Graded C because nothing independently examined was located. Ask for an attestation, and ask whether the research partner terms still reflect current practice after the restructuring.

Correctly structured, with the platform not presented as a regulated device and regulatory activity at the asset level. Disclosed programmes have moved through real regulatory steps: BEN-2293 completed a Phase IIa study in atopic dermatitis before being discontinued, BEN-8744 for ulcerative colitis progressed from clinical trial application to a Phase I trial, BEN-28010 was declared a clinical candidate, and BEN-34712 for ALS was described as in IND enabling work.

Buyers should treat the published pipeline as historical unless a newer table is issued, since the company is now private and the 2024 restructuring prioritized five assets and shifted strategy toward partnering earlier stage programmes.

This company addresses the bias question directly, in its own words, and it is the only vendor located in this category that engages the representativeness dimension rather than only the technical one.

On the domain relevant question for a knowledge graph, it states the problem itself: biomedical literature is heavily biased toward a small number of diseases typically of interest to the developed world, and it says it applies data normalisation strategies to down-weight that skew when making predictions. A system that extrapolates novel connections from published literature inherits publication bias directly, so a vendor naming that and describing a mitigation is doing more than the category norm.

It then goes further, onto ground almost nobody in this field touches. It states that biobanks and other data sources provided by government and commercial organisations are heavily biased toward patients of European descent with other ethnicities often underrepresented, says it has developed tools to quantify diversity in data, and established a named internal initiative to raise awareness of the issue. That is an explicit acknowledgment that its inputs do not represent the global population, paired with a stated attempt to measure the gap rather than assert it away.

Held at B rather than A on three counts. The disclosure sits in authored commentary rather than in product documentation, a model card or a methods paper, so it is not where a buyer would look. No results from the diversity tooling are published, so the size of the skew and the effect of the normalisation are both unquantified. And the company underwent a strategic overhaul and left public markets during 2025, so whether that initiative and its tooling survived the restructuring is unestablished and should be asked directly rather than assumed.

Related caution retained: improving extraction quality from the literature does not change what is missing from the literature.

Two passes located no peer reviewed methods paper establishing platform predictive performance, no published error characteristic and no warranty, indemnity or remediation commitment. The substantive finding on this record is about where the existing disclosure came from and what happens to it now, and it generalises beyond this company.

Considerable detail is public, covering the composition and normalisation of the knowledge graph, a next generation graph built with language processing, the use of large language models to predict targets from literature, and a distinct product assessing target progressibility across druggability, selectivity and the competitive and patent landscape.

Almost all of it was produced by the reporting obligations of a public listing, which required describing platform work to investors on a schedule and to a standard. The delisting has ended that stream. So the disclosure is comprehensive and increasingly historical, describing a platform as it stood when someone was obliged to describe it, with nothing compelling an update and no way for a reader to tell how much has changed since.

Disclosure that exists as a byproduct of an external obligation decays when the obligation lapses, and a buyer reading it should date it rather than treat it as current. Ask what the platform does now, for predictive performance with a denominator, and for the prospective success rate of platform originated programmes against the industry base rate.

Not applicable. This is a target identification and discovery platform with no provider workflow surface and no EHR touchpoint. Integration effort is directed at normalizing heterogeneous scientific and partner datasets into the Knowledge Graph.

Historically not applicable, since the platform was operated in house and delivered through collaboration with scientists working side by side rather than as software deployed to a customer, and no tenancy, hosting or residency terms were located. The December 2024 strategy statement is forward looking and worth tracking: the company said it intends to transform its existing technology into more flexible standalone product offerings enabling faster integration and customised applications. If that materializes the axis becomes assessable, but nothing shipped was identified in this review.

This grade records a change of state rather than an act of concealment, and the trajectory is the point. As a listed company BenevolentAI disclosed more than almost anyone in this category, publishing a three pillar monetisation model covering end to end discovery partnerships, the clinical and preclinical pipeline and knowledge exploration products, along with detailed platform progress, target delivery counts and adverse pipeline decisions.

That reporting stream ended with the 13 March 2025 delisting from Euronext Amsterdam, executed as a merger into Osaka Holdings after which the original entity ceased to exist and the acquiring Luxembourg vehicle took the BenevolentAI name. The company is now private, no pricing or current deal economics were located, and a buyer evaluating it today is working from a disclosure record that stops in early 2025.

Disease agnostic by design and demonstrated across an unusually wide indication range for this category, with disclosed work in idiopathic pulmonary fibrosis, chronic kidney disease, systemic lupus erythematosus, heart failure, atopic dermatitis, ulcerative colitis, ALS and COVID-19 repurposing. Target identification is also described as extending to targets best prosecuted via alternative modalities rather than being locked to small molecules.

Held at B rather than A because the 2023 pipeline review narrowed the portfolio to five prioritized assets and the 2024 restructuring reduced scope further, so demonstrated breadth is historical while current capacity is smaller than the record implies. Buyers should ask what the platform is resourced to cover now rather than what it has covered.