Recursion

The models are what the company is built around. The Recursion OS pairs automated wet labs, running up to 2 million experiments a week into a proprietary dataset of more than 50 petabytes, with machine learning foundation models and purpose built language models, in a "lab-in-the-loop" where model predictions are tested in the lab and every result feeds back into training.

The platform page says it spans target identification to clinical trial enrollment, and the internal pipeline and pharma partnerships both come out of it. The labs, compute and data exist to train and test the models; without them the company would be a conventional biotech with a pipeline it no longer had a way to extend.

The oversight structure is described and the decision points are not. Predictions from the Recursion OS are tested in the company's automated labs, a wet and dry "lab-in-the-loop" run by what it calls bilingual teams of scientists and technologists, so model output meets physical evidence before a program moves on. Human judgment over the portfolio is visible in practice: the company has publicly deprioritized and paused clinical programs.

What is not published is where people decide: which model outputs are committed to synthesis or to a program without further review, the thresholds for advancing a target or molecule, and who signs off.

Models are named, their training data characterized, and versions published. The scientific materials page lists the platform's methods by name and venue: the Phenom masked autoencoder models for cell microscopy (CVPR 2024), their scaled successor ("ViTally Consistent", ICML 2025), MolE, a molecular graph foundation model (Nature Communications, 2024), TxPert for perturbation prediction using multiple knowledge graphs (Nature Biotechnology, 2026), Boltz-2 for binding affinity (preprint, 2025, developed with outside collaborators) and Boltz-ABFE, alongside active learning over synthon space for molecular design. Training data is characterized: the RxRx datasets, including public releases, and a proprietary corpus of more than 50 petabytes from up to 2 million weekly experiments. Model generations are named (Phenom-1 and its successor, Boltz-1 and Boltz-2), and the OpenPhenom model is released publicly.

What a partner cannot see is which versions run in the internal platform at a given time; the company also describes purpose built language models without naming them.

A largely in house chain, with its data suppliers named. Recursion builds its own models on its own compute (BioHive-2, built with NVIDIA technology) and generates most of its data in its own labs; licensed data comes from named data partners holding human genomic and clinical data. Some methods are developed with outside collaborators and released openly (Boltz-2).

The chain is not enumerated beyond that: no subprocessor list, no statement of outside cloud providers, and nothing on whether partner data enters shared models or stays in partner specific ones. The company is also silent on what happens to representations derived from the licensed human data when a license ends.

Pipeline progress that can be checked, and peer reviewed method papers. The pipeline lists seven programs across oncology and rare disease, with clinical studies named on the company's site (TUPELO in familial adenomatous polyposis, DAHLIA, ENLYGHT, ZINNIA, ELUCIDATE and EXCELERIZE, where the first patient was dosed with the MALT1 inhibitor REC-3565), interim Phase 1 dose escalation data for REC-617 presented in December 2024, and a Phase 1b/2 update on REC-4881 in 2025. The scientific materials page lists peer reviewed methods behind the platform: TxPert on predicting transcriptomic perturbation effects (Nature Biotechnology, 2026), the Phenom microscopy models (CVPR 2024; ICML 2025), MolE, a molecular foundation model (Nature Communications, 2024), perturbative maps (PLOS Computational Biology, 2024), and a DCAF1 ligand found with a drug target interaction model (Journal of Chemical Information and Modeling, 2023).

Not all of that evidence is favorable. The company has also deprioritized clinical programs, no medicine from the platform is approved, and its claim of faster, cheaper progress from hit to IND enabling studies against industry averages is stated without published method.

The platform uses human derived data, and nothing published answers the questions that raises. Recursion's home page features a post on "leveraging the power of patient data" in drug discovery, and its data corpus includes material licensed from companies holding human genomic and clinical data. No published statement covers consent provenance for that licensed data, the de-identification standard applied, permitted secondary uses, retention, or what happens to trained model representations if a license ends.

The website privacy notice (last updated 24 February 2025) covers website visitors and, separately, the health data of clinical trial participants and patient program enrollees, and defines de-identified data, but it does not describe the discovery platform's datasets. Partners' proprietary data is handled under private contracts, and no categorical commitment on its use in training is published. Safety engineering is not addressed.

A privacy document covers patient data in some settings, and no position is stated for the platform. Recursion's privacy notice (last updated 24 February 2025) covers website visitors and, in separate sections, the health data of clinical trial participants and patient program enrollees, whose data it shares with contract research organizations and investigators. As a drug sponsor Recursion is not usually a business associate, and its discovery partners are pharmaceutical companies rather than covered entities.

The platform itself uses licensed human genomic and clinical data, and nothing published says whether any of it is protected health information, how it was de-identified, or under what agreements it arrives.

A described security program with outside help, and no attestation. Recursion's annual report on Form 10-K carries the cybersecurity section listed companies must file: it says the program is informed by the National Institute of Standards and Technology Cybersecurity Framework, that the audit committee oversees cybersecurity risk and receives regular reports and notice of potentially reportable incidents, that an internal information security team reviews threat intelligence, and that the company engages third party consultants. It also discloses that the company has experienced cybersecurity incidents, none determined to be material.

A framework the program is "informed by" is not an audit of how it runs, and no SOC 2 report, ISO/IEC 27001 certificate or trust center is published. That matters for partners handing over unpublished targets and assay data.

No device claim is made and the platform is scoped as internal research and development. The Recursion OS discovers and designs drug candidates; it makes no diagnostic or treatment recommendation to clinicians or patients, so no clearance applies to it. The regulatory standing that exists sits with the drug candidates, several of which are in clinical studies under the company's own trials, and none is approved; that is the drugs' position, not the platform's.

The platform page says the OS extends to clinical trial enrollment; where model output shapes trial design or submissions, FDA expects the sponsor to establish the model's credibility for that use, and the company states no position on it.

Published evaluation material on the method's main confounder, with no governance framework. Recursion has released large public datasets and benchmarks built so outsiders can test its approach, including RxRx3 and the RxRx3-core benchmark for drug target interactions in high content microscopy (2025), and papers on building and benchmarking perturbative maps (PLOS Computational Biology, 2024) and on obtaining cellular embeddings across experimental batches. Batch effects, where the same perturbation looks different across plates, sites and runs for reasons that are not biological, are the failure mechanism most likely to mislead a phenomics platform, and the company publishes material that lets others measure how well it is handled.

What is not addressed is who the models generalize to. Cell lines carry their donors' genetic backgrounds, and the company also uses licensed patient data; no analysis of cohort composition, ancestry or representativeness is published, and there is no stated governance framework, model card or intended use statement. Safety engineering is not addressed.

A mechanism checks outputs before they are acted on, and nothing stands behind them. The platform's design is that model predictions are tested in Recursion's own automated labs before a program advances, so a wrong prediction is meant to fail at the bench rather than in a partner's hands. When programs fail later, the company reports it: its pipeline has been narrowed publicly as programs were deprioritized.

For a partner, that is the whole of what is published. No error rate for target or compound predictions, no success rate of platform originated programs against industry base rates, and no warranty, remediation or fee consequence when a platform derived hypothesis fails is stated; partnership terms are negotiated privately, and upfront payments are not described as at risk.

No clinical record integration exists: the Recursion OS is the company's own internal discovery platform, not software a provider or partner installs, and no clinical record system is involved. For a pharma or biotech partner the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments. None is named as connected.

The platform runs on the company's own experimental data and on licensed datasets, including human genomic and clinical data from outside data partners, and its outputs feed internal and partnered programs. The platform page mentions "translational clinical tools" and clinical trial enrollment, but names no record system, research data platform or interface standard, and nothing describes how a partner's data or systems exchange data with it. That leaves no integration evidence at all, which for a partnership business is an absence rather than a defect in the platform.

An internal platform, with location implied rather than committed. The Recursion OS is not deployed to customers: partners' data and questions come to Recursion's own infrastructure, which includes its BioHive-2 supercomputer and automated labs, with the company headquartered at 41 South Rio Grande Street, Salt Lake City. The company does not state where partner data, licensed patient data or models are stored, whether any of it runs on outside cloud providers, or how one partner's data is separated from another's and from Recursion's own programs.

Recursion also releases some models publicly (its OpenPhenom microscopy model is distributed through a cloud model catalog), which anyone can run in their own environment; that is a research artifact, not access to the discovery platform.

The commercial structure is discoverable even though no price is published. Recursion sells no software license; it works with pharmaceutical partners through discovery collaborations and also develops its own pipeline, and the partners page describes pharmaceutical, technology and data partnerships. As a listed company it discloses the aggregate economics of those deals, such as total potential milestones and payments received, but those are what particular partners negotiated for particular scopes, not a price a new partner can apply to its own program.

The disclosures show the shape of the terms (upfront payments, research funding, milestones and royalties) and nothing about how a new partner's own deal would be priced.

Coverage is named, with validation behind part of it. The internal pipeline concentrates on oncology (solid tumors, lymphoma, B-cell malignancies, small cell lung cancer) and rare disease (familial adenomatous polyposis, hypophosphatasia), with clinical studies running in both. The platform is claimed to work across biology and chemistry, and partnerships extend it further, including neuroscience with Genentech, where the first target has been unlocked. Validation is strongest where the clinical programs are; neuroscience and other partnered areas are earlier.