Recursion

The directions that can currently be confirmed are: a Phase 2 efficacy signal in a disease with no approved therapy (REC-4881 in FAP), a molecule the platform designed rather than repurposed that has not yet dosed a patient (REC-7735), $350 million in partner payments received with one target option exercised out of up to forty, an agentic layer reported through internal efficiency metrics, and a burn rate that is filed quarterly with the SEC. Recursion is the only company on this site that is legally required to say, on a schedule, what all of this costs.

Scope note: this page tracks the public record only: SEC filings, company releases, the peer-reviewed literature and third-party reporting. No inside information. It records how far public material goes, not how well the company is doing.

Overview

1. The first clinical signal, and what the platform contributed to it

What can be confirmed: on December 8, 2025 Recursion filed Phase 1b/2 results from the TUPELO trial of REC-4881, an allosteric MEK1/2 inhibitor, in familial adenomatous polyposis. At 4 mg once daily: "75% of evaluable patients showing reductions in total polyp burden and a 43% median reduction after 12 weeks of treatment (n=12)." Twelve weeks after stopping therapy, "82% of evaluable patients (9 of 11) maintained a durable reduction in total polyp burden, with a 53% median reduction observed from baseline." Four of ten patients achieved a ≥1-point improvement in Spigelman stage. Grade 3 treatment-related adverse events occurred in 15.8% of safety-evaluable patients, with no Grade ≥4 reported. FAP has no approved medical therapy.

The provenance is the part worth reading slowly, and Recursion states it plainly. An unbiased phenotypic screen in "the earliest version of the Recursion OS" identified MEK1/2 inhibition as a mechanism that reverses APC loss-of-function signatures. "Guided by this AI-driven insight, Recursion in-licensed REC-4881 from Takeda and redirected REC-4881 — originally evaluated clinically in solid tumors — as a mechanistically aligned therapeutic candidate for FAP."

So the platform's contribution here is a mechanism and an indication, not a molecule. The compound was made by somebody else and had already been in patients. That is a smaller claim than "AI discovered a drug" and a more interesting one: the hard part of FAP was not chemistry, it was knowing that a MEK inhibitor was the thing to point at it. REC-4881 is the first MEK1/2 inhibitor ever studied clinically in this disease.

Boundary: eleven patients and no control group. The durability result rests on 9 of 11 evaluable patients in a single-arm trial. The comparison is a retrospective natural-history analysis (n=55) showing 87% of untreated patients with annualized polyp-burden increase. This is genuinely useful context, and not a randomised control. Confidence: high that the data are as reported (filed with the SEC, named principal investigator); medium for efficacy, at this size and without randomisation.

Open leads: the FDA's view on a registrational path, expected in 2H26; updated data at the CGA-IGC Presidential Plenary on November 2, 2026; whether the effect holds in the expanded population now enrolling from age 18.

2. The molecule the platform designed: REC-7735

What can be confirmed: REC-7735 is a PI3Kα H1047R inhibitor "precision designed to show >100-fold selectivity for the H1047R mutant over wild type." Recursion reports that "the differentiated development candidate was delivered in 10 months and 242 compounds from first novel hit." The IND has cleared and the Phase 1/2 ZINNIA study is to start in the second half of 2026.

Why this matters more than the efficiency headlines elsewhere on this site. "242 compounds from first novel hit" is an unusual disclosure: it is a number a medicinal chemist can hold against their own experience and judge. Most AI-design claims are expressed as time saved against an unstated baseline. This one names the quantity that a discovery programme actually spends.

Boundary: no human data. This is a designed molecule with a selectivity claim and a cleared IND, which is a different thing from the FAP programme: there the platform picked the target and licensed the compound; here it made the compound and has not yet dosed anyone. Confidence: medium (company-reported design metrics, regulatory clearance is checkable, no clinical data).

3. Partner validation: $350 million received, one option exercised

What can be confirmed: as of the Q2 2026 filing, Recursion has received "$216 million in upfront and milestones payments from the Roche and Genentech collaboration" and "$134 million in upfront and milestone payments from the Sanofi collaboration." The Roche/Genentech deal covers "up to 40 potential small molecule discovery programs, each carrying the potential for more than $300 million in development, commercialization, and net sales milestones."

The Q2 event is the first exercise of a Validated Target Option. Genentech advanced "a previously unexplored neuroscience target into a small molecule early discovery program." The underlying asset is described concretely: "the first whole-genome CRISPR knockout map generated from a subset of over 1 trillion internally manufactured iPSC-derived neuronal cells," with candidate targets passing successive pathway, functional and disease validation stages jointly designed with Genentech.

Read the ratio, not the headline. A pharmaceutical company with access to the data chose to spend its own discovery resources on one target the platform produced. That is a real, informed vote, and it is one, out of up to forty, several years in. Both halves belong on the record.

Boundary: commercial validation of target discovery. No molecule, no candidate, no trial follows from this yet. Confidence: high (SEC-filed, named partner, dated, with cumulative payments stated).

4. The agentic layer, and the numbers it reports

What can be confirmed: Recursion reports three internal agents. A Target Discovery Agent lets scientists "mine and extract insights from proprietary maps in hours rather than weeks." Drug Design Agents cut "structural analysis time… from 4 hours to 30 minutes," with "agent-generated hypotheses now driving design cycles in active programs." A Clinical Strategy Orchestration Agent contributed to "a 1.3 to 1.6x increase in enrollment rates versus historical benchmarks."

These are internal efficiency metrics against internal baselines, which is what makes them hard to weigh. Four hours to thirty minutes is a real operational change for whoever was doing that analysis; it says nothing about whether the resulting molecules are better. The enrollment figure is the most interesting of the three, because trial enrollment is a rate-limiting step nobody solves with chemistry, and "historical benchmarks" is not defined.

Boundary: speed of internal work. Confidence: low-to-medium (company-reported, no stated baseline, no external check).

5. The clock

What can be confirmed, from the quarter ended June 30, 2026: total revenue $7.7 million, down from $19.2 million a year earlier; net loss $131.0 million; cash, cash equivalents and restricted cash $556.8 million, down from $753.9 million at the end of 2025; cash runway expected "to extend into early 2028"; full-year cash operating expense guidance cut by $15 million to under $375 million; general and administrative costs down "primarily driven by a decrease in salaries of $4.9 million as a result of headcount reductions in the year." A "strategic reprioritization of our clinical portfolio" took place in the second quarter of 2025. Najat Khan became CEO on January 1, 2026, with co-founder Chris Gibson moving to chair.

Why a balance sheet belongs on a page about science. Every other organization tracked on this site can describe its progress without pricing it. Recursion cannot: it burns roughly $100 million a quarter in operating cash, files the number, and has told the market when the money runs out. That makes it the only place in this field where the cost of the thesis (industrialised phenomics, proprietary data, an owned clinical pipeline) is visible at all. It also means the decisive test is dated: something has to work before early 2028, or the terms change.

Boundary: finance. Confidence: high (SEC filings).

What could not be found

• No randomised controlled data for REC-4881. The contextual comparison is a retrospective natural-history cohort, and the company says so.
• No peer-reviewed publication of the TUPELO results. The next disclosure is a conference presentation in November 2026.
• No detail on the 2025 clinical reprioritization: how many programmes were discontinued, or which.
• No baseline behind the agent metrics. "Historical benchmarks" for enrollment is not defined anywhere public.
• No platform-native molecule with human efficacy data. REC-7735 is the design claim and has not dosed a patient; REC-4881 is the efficacy signal and was licensed from Takeda.
• No split of the $216 million and $134 million between upfront payments and earned milestones.

Method and scope

Built from SEC-filed exhibits, namely the Q2 2026 results (2026-08-05), the full-year 2025 results (2026-02-25) and the TUPELO data release (2025-12-08), read at the SEC rather than from press coverage, plus the company's own descriptions of its platform. Every figure quoted above was read in a filed document.

Confidence is assigned claim by claim, not to a whole direction. The labels:

• High: a regulatory filing, a public artifact, or execution by an outside institution;
• Medium: an official release or company-reported figure: inspectable in detail, unverified outside;
• Low: a single-source claim, or a metric reported against an undisclosed baseline.

Two operating rules: revisions are dated by version and are not treated as new projects; evidence that a project exists is not evidence that the technology has succeeded.

One remark specific to this company. The temptation with Recursion is to read the FAP data as "AI discovered a drug," and the company's own filing does not support that reading: it says the platform found the mechanism and the molecule was in-licensed from Takeda. Held to what it actually claims, the result is still notable: a phenotypic screen pointed at a disease nobody had pointed a MEK inhibitor at, and eleven patients later there is a durable signal. Whether that generalises is what REC-7735 and the next thirty-nine Genentech options are for.

Primary sources

SEC filings

• Q2 2026 results and Genentech option exercise, Exhibit 99.1 (2026-08-05)
• TUPELO Phase 1b/2 results at 25 weeks, Exhibit 99.1 (2025-12-08)
• Full-year 2025 results and business update, Exhibit 99.1 (2026-02-25)
• Recursion filing index, CIK 0001601830

Company

• Recursion platform
• RxRx3 dataset