Recursion Pharmaceuticals

Recursion's case rests on one claim: that cellular images are a training signal nobody else has at volume. The automated labs run up to 2 million weekly experiments, and every plate of images feeds models that try to map the cellular disruptions driving disease. Ten-plus years of that produces the 50+ petabyte proprietary dataset spanning phenomics, transcriptomics, proteomics, ADME and de-identified patient data — the asset Recursion treats as the moat, and the reason the company describes itself as an AI-native data factory rather than a software vendor. The second asset is the loop itself. Bilingual wet/dry teams run physical automation and machine learning as one continuous system, and generative de novo chemistry narrows millions of molecules to a synthetically accessible short list before anything reaches the bench. Recursion says it has advanced candidates after synthesizing only around 300 compounds, roughly 90% fewer than the industry average — a concrete, checkable-efficiency number rather than a general speed claim. BioHive-2, co-developed with NVIDIA, carries the training and simulation load. The third asset is clinical. REC-3565, a selective MALT1 inhibitor, has dosed its first patient in Phase 1 for relapsed or refractory B-cell lymphomas. REC-1245, an RBM39 degrader, showed well-tolerated safety and predictable pharmacokinetics in solid tumors with no DLTs, supporting dose escalation per the Q1 2026 update. REC-4881, a MEK1/2 inhibitor, is in Phase 2 for familial adenomatous polyposis with a data update in play, and a hypophosphatasia program extends the rare-disease side. Genentech optioning the first neuroscience target into joint early discovery in August 2026 is the strongest external validation yet that a partner looked at the platform output and paid to move it forward. Where to be careful. The pipeline is concentrated in oncology and rare disease, so a program outside those areas will not find much read-through. The headline metrics — 50+ petabytes, 2 million weekly experiments, the ~90% reduction in compounds synthesized — are company-reported and not externally audited. Clinical traction is early: a first Phase 1 patient dosed and a clean safety readout are safety signals, not efficacy proof; the efficacy question is still open. And the relationship model is a multi-year joint discovery program, which means you are committing to a partnership and a shared roadmap, not evaluating a tool in a week. Fit: large pharma or funded biotech with a hard oncology or rare-disease target and appetite for a multi-year collaboration, especially teams that would rather own proprietary wet-lab data than rent predictions. Not fit: academic groups wanting open or low-cost access, programs outside oncology and rare disease, and teams that need a fast bounded proof of concept. Against software-only prediction players, Recursion's differentiator is that it generates its own experimental data and has carried candidates into patients.