Recursion

The models and the data flywheel are the company rather than a capability inside it. The Recursion OS learns from cellular imaging and multiomic data generated by the company's own high throughput experimentation, and the entire operating structure exists to feed it, including BioHive-2, described at completion as the fastest supercomputer wholly owned and operated by any pharmaceutical company, built with NVIDIA technology.

The Exscientia combination added generative chemistry design and automated synthesis, extending model driven decision making from target biology through molecule design. Removing the models does not leave a lesser product, it leaves no product.

The oversight mechanism is structural and reasonably strong: model hypotheses are tested against wet lab experimentation at industrial scale, so predictions meet physical evidence continuously rather than accumulating unchecked. Human strategic judgment is also demonstrably in the loop, as the May 2025 decision to deprioritize three clinical stage programs shows leadership overriding platform momentum.

Held at B because no disclosure was located on how candidate selection is gated between model output and committed spend, meaning the specific human decision points are inferred from company behaviour rather than documented.

Named systems and disclosed infrastructure put this above the opaque end of the category: the Recursion OS, the Phenom family of foundation models, BioHive-2 and a stated data corpus of more than 60 petabytes generated in house or licensed from partners including Helix and Tempus. Held at B rather than A because no peer reviewed methods paper establishing core platform performance was located in this review, so the headline capability claims rest on company communications.

Widely circulated specific figures for model accuracy and patent counts trace to secondary commentary rather than to primary company or peer reviewed sources and should not be relied upon in diligence.

The systems and the data suppliers are both named, which answers more of this axis than most of the category manages. A named operating system, a named foundation model family and named compute infrastructure are disclosed, alongside a stated corpus of more than sixty petabytes generated in house or licensed from partners, with the licensors identified by name. Naming who supplied your data is the corpus half of this axis answered properly, and it is rarer than naming a model.

Held below the top grade because of what those particular suppliers hold. The licensed data comes from companies holding human genomic and clinical data, so human derived material flows into the platform, and no published disclosure was located covering consent provenance for it, the de identification standard applied, permitted secondary use, or what happens to derived model representations if a licensing relationship ends.

That last question is the one this index has learned to ask of any derived artefact: a model trained on licensed human data is not obviously returnable, and the licence ending does not unlearn it. The distinction matters for comparison inside this category. A platform training on human data carries stewardship obligations that a platform training on molecular structures does not, and the disclosure here does not yet match the obligation. Ask for consent provenance, the de identification standard, and the position on derived representations at licence end.

Genuinely mixed, and the mix is the useful signal. On the positive side, pharma partners have paid: more than 450 million dollars in upfront and realized milestone payments received against a disclosed potential exceeding 20 billion dollars, which is third party validation of delivered work rather than a vendor assertion, and multiple programs reached the clinic.

On the negative side, the company promised roughly ten clinical readouts within eighteen months at the time of the Exscientia combination, then in May 2025 deprioritized three clinical stage programs, paused a fourth and wound down a preclinical program, narrowing to six active projects. That attrition is the honest counterweight to platform scale claims, and it is exactly the evidence a buyer evaluating industrialized discovery should weigh. No approved medicine has yet emerged from the platform.

Unlike pure chemistry platforms this axis genuinely applies, which is why it is graded rather than marked not applicable. The disclosed data corpus includes material licensed from Helix and Tempus, both of which hold human genomic and clinical data, so human derived data flows into the platform.

No published disclosure was located covering consent provenance for that licensed data, de identification standard, permitted secondary use, or what happens to derived model representations if a licensing relationship ends. The distinction matters for comparison within this category: a platform training on human data carries stewardship obligations that a platform training on molecular structures does not, and the disclosure here does not yet match the obligation.

Not applicable in the direct sense. Counterparties are pharmaceutical partners and data licensors rather than covered entities transferring protected health information under a business associate agreement, so the governing instruments are commercial data licences. Note that this places the consent and permitted use questions raised on the stewardship axis inside private contracts rather than inside a disclosed compliance framework.

No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre was found. As a Nasdaq listed registrant, however, the company files a required annual cybersecurity disclosure, and it is the most substantive in this category.

Three things lift it above the norm. The programme is anchored to a named external standard, with the company stating that its information security approach is informed by the National Institute of Standards and Technology Cybersecurity Framework. Almost no vendor in this lane names any framework at all, and naming one gives a reviewer something to assess the programme against. Second, governance is specific rather than gestural: the audit committee holds express charter oversight of cybersecurity processes, controls and procedures, is responsible for monitoring and reviewing mitigation efforts, and receives regular reports on strategy, risk and mitigation as well as notification of potentially reportable incidents. A dedicated internal information security team reviews threat intelligence and engages third party consultants.

Third, and rarest, the company discloses that it has experienced cybersecurity incidents in the past while stating that none has been determined material. Most registrants say nothing on that point. Publishing the unflattering half is a pattern this index credits wherever it appears.

Held at B rather than A because a filing describes a programme rather than evidencing an examined one. There is no independent attestation, no scope statement and no trust centre through which a customer could request evidence. That gap matters more here than for peers handling molecular data only, since the platform ingests licensed human genomic and clinical data alongside partners' proprietary target information.

Correctly structured: the discovery platform is not a regulated device, and regulatory standing sits at the asset level where multiple programs have reached clinical stage across oncology and rare disease under active INDs. No approval has been obtained. Buyers should read the 2025 pipeline reprioritization as a reminder that clinical stage status is not durable, since three clinical stage programs were deprioritized in a single quarter.

No formal governance framework or bias statement exists, but a second look finds substantial published material on one half of the domain relevant question, and the split between the two halves is itself informative.

What the company does publish, openly and at scale. It released a foundation model for microscopy trained by self supervised learning on more than three million images drawn from publicly accessible datasets, distributed through a major cloud model catalogue. It released an accompanying challenge dataset of over 200,000 wells covering hundreds of genetic knockouts and more than a thousand compounds across eight concentrations, with computed embeddings and an associated drug target interaction benchmark, described in a workshop preprint. Most pointedly, an earlier release was built specifically to let the community evaluate experimental batch correction methods. Batch effect is the confounder that most threatens this entire approach, because the same perturbation can look different across plates, sites and runs for reasons that are not biological, and the company published the dataset that lets outsiders test how well anyone handles it, including itself. Independent researchers have since used that material to publish generalisation failures and limits. Releasing the benchmark on which your own paradigm can be shown to be fragile is the strongest form of this behaviour.

What remains unaddressed is the biological representativeness question, and it is not the same thing. Cellular models carry the genetic background of their donors, so a morphology model trained predominantly on cell lines from narrow ancestral backgrounds may generalise unevenly to patient populations that differ from them. With human genomic data also licensed into the corpus, cohort composition compounds it. Nothing published addresses either.

So the company publishes openly on the confounder that threatens its method's technical validity, and says nothing about the one that determines who its outputs generalise to. Held at B on that, and on the absence of any formal framework, model card or intended use statement.

Two passes located no peer reviewed methods paper establishing core platform performance, no published error characteristic, no evaluation methodology and no warranty, indemnity or remediation commitment, so the headline capability claims rest on company communications alone. One finding here is a diligence warning rather than a grading point and it belongs on the record.

Widely circulated specific figures for model accuracy and patent counts trace to secondary commentary rather than to primary company material or peer reviewed sources, which means a buyer who encounters those numbers in a deck or an article is looking at something with no locatable origin.

This index has recorded the reverse pattern, where a vendor cites generous third party characterisations of itself, and this is its cousin: figures acquire authority through repetition rather than through a source, and a diligence process that accepts them is accepting a claim nobody made.

Named systems and disclosed infrastructure sit on the other side of the ledger and are graded on the supply chain axis rather than here, because naming what you run is not the same as measuring what it does. Ask for the primary source behind any performance figure you have been shown, for a peer reviewed methods paper on core platform performance, and for the prospective success rate of platform originated programmes against the industry base rate.

Not applicable. This is a preclinical discovery platform with no provider workflow surface and no EHR touchpoint. Data integration effort is directed at internal experimental pipelines and licensed external datasets rather than at clinical systems.

Not applicable in the software sense. The Recursion OS is an internal operating platform rather than a product deployed into a customer environment, and the commercial relationship is discovery partnership rather than licensing, so questions of tenancy, hosting and residency do not arise for a counterparty. A partner's data protection question is therefore contractual, covering what the partner contributes and what the platform retains, rather than architectural.

One distribution surface does exist and should not be confused with the commercial platform. The company's openly released microscopy foundation model is published through a major cloud provider's model catalogue, so an outside organisation can run that model in its own cloud environment. That is a free public research artefact rather than access to the discovery platform, and using it involves no relationship with the company at all. A buyer evaluating a partnership is not evaluating the same thing.

Unusually specific, and public listing is the reason. As a Nasdaq registrant (RXRX) the company discloses partnership economics with real numbers rather than adjectives: approximately 200 million dollars in expected milestone payments over 24 months at the time of the Exscientia combination, more than 20 billion dollars in total potential milestones before royalties, more than 450 million dollars already received, and approximately 850 million dollars in combined cash. Pipeline changes including deprioritizations are announced through quarterly results rather than allowed to disappear quietly, which is the harder half of transparency.

Deliberately narrowed rather than broad. Following the May 2025 prioritization the internal pipeline runs six active development programs, four in oncology and two in rare disease, with the company describing these as its consistent focus. The underlying platform is claimed to work across biology generally and the partnership portfolio extends further, but the demonstrated internal application is concentrated. Buyers outside oncology and rare disease should treat applicability as a partnership discussion rather than a proven setting.