Recursion combines biological experiments and AI for drug discovery

Recursion is a drug-discovery and development company that combines automated biological experiments, machine learning and chemistry. Its central proposition is a recurring loop: generate experimental data, use models to identify promising directions, test those directions and feed the results back into research. The company should be evaluated as a research and development partner, rather than assumed to be a self-service molecular chatbot.

01 / ProductThe platform links computational predictions with physical experiments

The Recursion OS description places biological data generation, computational models, molecular design and clinical development within a connected research system. Its important feature is the return path from experiment to model. Predictions can influence what scientists test, while the resulting measurements can expose weaknesses in the computational picture and inform the next iteration.

This is materially different from generating a plausible molecule in isolation. A proposed molecule still needs an experimental context, a synthesis route and evidence about the properties that matter for a program. A biological association also needs interpretation: a cell response may suggest a useful direction without establishing that the same mechanism will produce a desired effect in people.

Recursion’s pipeline page organizes programs by development stage and therapeutic area. It is a view of ongoing development, not a catalog of medicines available for routine use. The stage distinction matters because discovery evidence, early clinical observations and an approved treatment answer different questions and carry different levels of uncertainty.

The completed Exscientia combination announcement states that Exscientia became a wholly owned subsidiary in November 2024. Readers encountering older Exscientia material should therefore interpret it within the combined company. Historical capabilities or pipeline labels still need to be checked against Recursion’s current offer before planning a collaboration.

02 / AudienceResearch partners and open-science users have different entry points

Recursion is most relevant to organizations with scientific questions that can benefit from a substantial discovery collaboration. Such a team needs to contribute disease knowledge, define meaningful experiments and assess the evidence returned. It should have a clear view of where its internal capabilities end and what it wants a partner to add.

The partner page describes collaborations with pharmaceutical companies, including Roche and Genentech, Bayer, Merck KGaA and Sanofi. These relationships illustrate a program-based commercial model. Their disclosed scope or potential milestone value should not be treated as a menu available to every prospective customer, nor as a guaranteed outcome for a new program.

An academic machine-learning team may have a different need. The RxRx site provides a route to public datasets and tools, including links to Boltz-2 and OpenPhenom. Access to a public resource is useful for research exploration, but it does not grant access to Recursion’s whole proprietary data environment, laboratory capacity or partnership services.

Abridge provides a useful boundary comparison: its clinical documentation workflow is a different use of healthcare AI from discovering medicines. NVIDIA represents a different layer: computational infrastructure and software used by many research builders. Neither is a simple substitute for a drug-discovery partnership; the reader should compare the missing capability in its own research process.

03 / WorkflowA proposed collaboration tests a biological hypothesis through several gates

Imagine a proposed project investigating whether a cellular response points to a tractable disease mechanism. Begin by stating the hypothesis and the observations that would weaken it. A model-generated ranking is only useful if the team knows what evidence would make a candidate worth further work. Otherwise, a long list of associations can become activity without a clear scientific decision.

Choose an experimental system appropriate to the question and record its limitations. In this proposed evaluation, scientists would specify controls, relevant perturbations and the measurements used to judge a response. This is not a description of Recursion’s private protocol. It explains the information a partner needs to understand before interpreting any computationally selected experiment.

Examine whether the model is learning the intended biological distinction or an incidental difference in how the data were produced. For example, a candidate that separates two experimental batches may appear useful until evaluated under a changed acquisition condition. Keep model exploration and confirmation data distinct so the program does not confuse a discovered pattern with an independently supported result.

Move a selected hypothesis into an explicit confirmation stage. Ask which alternative explanations remain and what additional measurement could distinguish them. A negative result can still improve the program if it rules out a tempting but unsupported path. The learning loop becomes valuable when it changes subsequent decisions, rather than simply generating more data for a growing archive.

If the work progresses into chemistry, separate the desired molecular properties from the system’s numerical scores. Review tradeoffs across the program’s objectives and check the feasibility of making and testing proposed compounds. A model may favor a property that is easy to predict while the harder, decisive property remains poorly measured. The research plan needs to make that imbalance visible.

At each decision gate, preserve the version of the hypothesis, input data and resulting experiment. This gives both parties a common basis for deciding whether to continue, redirect or stop. The proposed deliverable is a traceable research decision backed by measurements. It is not a promise that an AI-selected candidate will advance successfully through clinical development.

04 / PricingPartnership economics cannot be reduced to an API tariff

The reviewed official pages do not publish a universal subscription or per-prediction fee for Recursion OS. The commercial route centers on partnerships, while public datasets and tools form a separate access path. A prospective collaborator should define the scientific scope and rights under discussion before trying to estimate the cost of the relationship.

The partner page includes arrangements with upfront payments, possible milestones and royalties. Those are transaction-specific structures, and contingent future amounts are not earned revenue simply because a maximum is announced. For a new project, the practical questions concern deliverables, responsibility for later experiments and ownership or licensing of resulting assets, all of which require the actual agreement.

05 / DistinctionsRepeated measurement gives the AI proposition a concrete mechanism

The meaningful distinction is the connection between model output and a system that can generate new experimental evidence. If a team only reorders existing knowledge, its learning is constrained by the data already available. A laboratory-linked process can instead choose an experiment to reduce uncertainty, then revise the next computational choice based on what happened.

This does not establish that every experiment is informative or every model is reliable. The mechanism is an opportunity to learn, not a guarantee. A useful partnership discussion should therefore include examples of how unexpected observations changed a program and how the team distinguishes a productive failure from a measurement problem.

Recursion’s company history also shows the evolution of its public research activity, acquisitions and computational resources. That history helps explain why older materials may use different product or program language. For current decisions, the relevant evidence is the present platform scope and the particular project, rather than an accumulated list of historical announcements.

06 / QuestionsAsk what is open, what is contracted and what remains experimental

Is a public model equivalent to access to the platform?

No such equivalence is established by the reviewed sources. A public resource can support an independent experiment under its own terms, while a partnership may involve proprietary data, laboratory work and shared program responsibilities. Inspect the license and documentation for the particular resource instead of treating the company’s open-science page as a blanket commercial license.

How should a pipeline milestone be interpreted?

Read it at the level of the specific program and study. Advancing a candidate is evidence of development activity, while the clinical questions depend on study design, endpoints, population and results. This blueprint does not compare treatment effectiveness or extrapolate a company’s platform claims into patient benefit.

How can a partner assess the learning loop?

Request a project-level account of how data become a decision, how that decision becomes an experiment and how results feed back into future work. Focus on the point where uncertainty is actually reduced. Large compute or data totals may describe capacity, but they do not reveal whether the next proposed experiment is the right one for the scientific question.

07 / DecisionChoose a defined scientific collaboration or a bounded research resource

Recursion belongs on a shortlist when an organization needs a partner capable of connecting computational discovery with experimental work and development. The useful starting point is a sharply defined hypothesis and the evidence needed to advance it. For independent researchers, the public tools offer a more bounded route, with a different scope and no assumption of access to the complete platform.