Drug discovery still fails roughly 9 times out of 10, and a single approved drug costs upward of $2.6 billion and over a decade to reach a pharmacy shelf. Recursion Pharmaceuticals just used an "AI Operating System" to compress part of that timeline into months — and an AI-selected molecule cut a hereditary disease's polyp burden by a median 43% in a mid-stage trial.
For most of the last century, finding a new drug meant testing one hypothesis at a time: pick a target, synthesize a molecule, run years of lab work, and hope. That process is why most biotech R&D budgets get burned on candidates that never reach a patient. There's a different way to run the experiment — one that treats biology itself as a dataset large enough to train AI models on, rather than a single hypothesis you test and discard.
Recursion (Nasdaq: RXRX), a 600-plus person biotech based in Salt Lake City, has spent over a decade building exactly that: a phenomics-driven "AI Operating System" that has already driven more than $500 million in upfront and milestone payments from pharma partners, and put one of its own AI-selected drugs into a Phase 1b/2 trial that reduced colorectal polyp burden by a median 43% in just 13 weeks.
Inside the AI Operating System Reshaping Drug Discovery
Recursion's platform — internally called Recursion OS — runs on a different raw material than most AI systems: phenomics, the systematic imaging of how millions of living cells respond to genetic or chemical perturbations. Lab robots dose cell cultures with thousands of compounds and gene edits, then photograph the results at massive scale, building what the company describes as one of the largest proprietary maps of cellular biology in existence.
That imagery feeds Phenom, Recursion's family of AI foundation models, which learn to recognize patterns invisible to a human eye under a microscope — which compounds nudge a diseased cell back toward a healthy state, and which genetic targets are worth pursuing at all. In January 2026 the company demonstrated LOWE, a natural-language AI workflow engine that lets bench scientists — not just data scientists — query the platform and launch new computational experiments directly.
If you work in biotech R&D, venture investing, or pharma business development, this matters because it reframes target discovery and early compound screening as a computation problem you can run in parallel, rather than a wet-lab bottleneck you run one hypothesis at a time.
How the Platform Actually Moves a Molecule Toward the Clinic
The full technical architecture spans wet-lab automation, foundation-model training, and clinical translation — Sityos AI will cover the deeper mechanics in future deep dives. At a high level, three things had to come together for Recursion's results to happen:
- Phenomic screening at scale — automated wet labs generate the raw imaging data that trains the AI directly, rather than relying on published literature alone.
- A dedicated supercomputer — Recursion built BioHive-2 with NVIDIA, now described as the fastest supercomputer owned and operated by any pharmaceutical company, to train and run Phenom at the scale phenomics data demands.
- Pharma-scale validation — every AI-flagged target or molecule still goes through the same regulatory gauntlet as any other drug candidate; Recursion's edge is in how many fewer dead ends it walks down first.
The 2024 all-stock merger with UK-based Exscientia — itself a pioneer in AI-designed small molecules — added automated chemical synthesis capability directly into that pipeline, so an AI-flagged target can move into an AI-designed molecule without leaving the platform.
The Results: Over $500M in Deals and a Trial That's Already Reading Out
Recursion's pharma partnerships are the clearest proof the platform produces results partners are willing to pay for ahead of approval. Roche and Genentech — partners since a 2021 deal covering up to 40 potential medicines in neuroscience and GI oncology — have paid Recursion more than $213 million to date, including a $30 million payment in November 2025 for a whole-genome phenotypic map of microglial immune cells. Sanofi has accepted all five of the AI-discovery program packages proposed under its collaboration, paying $134 million in milestones so far. Across all partners, Recursion has now earned more than $500 million in upfront and milestone payments, with over $100 million more expected by the end of 2026.
Recursion's own pipeline shows the same pattern translating to patients. REC-4881, an AI-selected MEK inhibitor repositioned for familial adenomatous polyposis (FAP) — a rare inherited disorder that fills the colon with precancerous polyps — produced a median 43% reduction in polyp burden at the Week 13 assessment of its Phase 1b/2 TUPELO study. Recursion is now expanding trial eligibility down to age 18 and plans to meet the FDA in the first half of 2026 to discuss a registration pathway.
The company also enters this phase under new leadership: Najat Khan, Ph.D. took over as CEO and President on January 1, 2026, succeeding co-founder Chris Gibson, who moved to Chair of the Board before stepping back to a strategic advisor role this year. Recursion ended its most recent reported quarter with $754 million in cash and equivalents — a runway into early 2028 without needing to raise additional financing, a buffer most single-asset biotechs simply don't have.
Industry analysts tracking AI-originated drug candidates broadly report Phase 1 success rates as high as 80-90%, versus a historical industry average near 52% [DATO A VERIFICAR — secondary industry analysis, methodology not independently confirmed against a primary source]. Biotechs without an AI-derived pipeline are increasingly competing for licensing dollars against companies that can show that kind of de-risking before a human ever takes the drug.
The Technology Stack Behind the Platform
None of this is sold as off-the-shelf software — Recursion's stack is built and operated internally, with one notable external dependency that's reshaping pharma infrastructure decisions industry-wide.
| Component | Role in the Platform | Availability |
|---|---|---|
| Recursion OS | End-to-end orchestration linking wet-lab automation, imaging, and AI models | Proprietary, internal use only |
| Phenom | Foundation models trained on phenomics imaging data for target and compound discovery | Proprietary; partner access via collaboration deals |
| LOWE | Natural-language workflow engine for scientists to query data and launch experiments | Internal tool, demoed publicly January 2026 |
| BioHive-2 | NVIDIA-built supercomputer for training and running Phenom at scale | Owned and operated by Recursion |
| NVIDIA BioNeMo | Cloud platform Recursion uses to host and distribute its own foundation models | NVIDIA platform; Recursion is the first hosting partner |
NVIDIA's relationship with Recursion goes beyond a customer contract — NVIDIA invested $50 million directly in the company and is using Recursion as a flagship case for biology-specific supercomputing. [REQUIERE VERIFICACIÓN — whether BioNeMo hosting terms are exclusive to Recursion or available to other biotechs on comparable terms].
Who Should Be Watching This Case Study
This is essential reading for pharma business-development teams evaluating AI-platform partners, biotech investors trying to separate genuine phenomics infrastructure from AI-branding, and engineering leaders at any data-heavy R&D organization curious how foundation models perform when trained on proprietary, single-company datasets instead of public internet text. It's less useful if you're looking for a tool you can sign up for directly — Recursion's platform isn't sold as a SaaS product, and the only way in today is a partnership, an equity stake, or a job application.
More AI-in-Pharma Case Studies on Sityos AI
This case study is part of an ongoing series tracking how AI is actually changing drug discovery economics — not just the press-release version. Related breakdowns worth reading next:
- Insilico Medicine: how a generative-chemistry pipeline took a drug from target to Phase IIa in 30 months for roughly $6 million.
- JPMorgan's COiN platform: how contract-intelligence AI replaced 360,000 manual legal hours a year in an entirely different regulated industry.
- New case study every week: we publish a fresh, fully-sourced AI automation or enterprise AI deployment breakdown on sityos.com.
Follow the Sityos AI case study series for the next breakdown
Every case study on sityos.com is sourced directly from company filings, press releases, and primary research — not recycled press copy. New breakdowns publish weekly.
Frequently Asked Questions
Can other biotechs use Recursion's AI Operating System?
Not directly. Recursion OS, Phenom, and LOWE are internal tools. Outside companies access the platform's output only through a formal partnership or collaboration deal, such as the ones Recursion holds with Roche/Genentech and Sanofi.
How is Recursion different from Insilico Medicine's AI drug discovery platform?
Both use AI to shorten drug discovery, but the core data is different. Insilico's PandaOmics mines multi-omics datasets to find disease targets and then generates candidate molecules with Chemistry42. Recursion's Phenom models are trained primarily on phenomics — large-scale cellular imaging — and the company built its own supercomputer, BioHive-2, specifically to run that imaging-based approach at scale.
How soon could an AI-discovered Recursion drug actually reach patients?
REC-4881, the furthest-along internally discovered candidate, is still in Phase 1b/2 trials with an FDA registration-pathway discussion planned for the first half of 2026 — realistically still several years from a possible approval. The faster near-term payoff shows up in partnership deals, where Recursion gets paid for de-risked targets long before any drug reaches the market.
Pharma R&D budgets are under more margin pressure every year, and the companies proving they can de-risk a target before spending nine figures on a Phase 3 trial are the ones picking up partnership dollars first. Recursion's $500 million-plus in deals — and a 43% trial result to back it up — is a preview of what that competition will look like through the rest of the decade.