Recursion Pharmaceuticals

Recursion Pharmaceuticals is a drug discovery company in Salt Lake City, Utah, built on photographs of cells. Its laboratories treat human cells with a genetic change or a chemical compound, photograph them, and have a neural network turn each picture into a list of numbers. Across thousands of treatments, those with similar numbers are taken to be biologically related. The method is called phenomics, the study of visible traits, and the company's name for its platform is the Recursion Operating System.^[1]

Recursion went public in April 2021 and in November 2024 merged with Exscientia, another drug discovery company that traded on Nasdaq.^[1][2] Chris Gibson, a founder, was chief executive officer until the end of 2025.^[3]

Chris Gibson began the work as a student in the laboratory of Dean Li at the University of Utah, which studied a blood-vessel disease of the brain, cerebral cavernous malformation. A hypothesis in the laboratory about a protein, RhoA, did not hold up. The team then screened thousands of drugs on diseased and healthy cells and used machine learning to find small differences between the images, and Gibson says that the algorithm picked hits more often than the people did. He finished a doctorate in bioengineering in 2013 and left medical school to start a company with Li and Blake Borgeson.^[4] Recursion was launched in Salt Lake City on November 5, 2013.^[1]

The company's initial public offering was priced on April 16, 2021, at $18.00 a share for 24,242,424 shares, with expected gross proceeds of $436.4 million, and its stock trades on Nasdaq under the symbol RXRX.^[1] On November 20, 2024, Recursion completed its merger with Exscientia, which became a wholly owned subsidiary.^[2] In May 2025 it reduced its programs from 11 to seven, and a month later a restructuring removed about 20 percent of its staff, around 160 jobs.^[3] The board named Najat Khan, the company's chief research and development officer and chief commercial officer, to succeed Gibson as chief executive officer on January 1, 2026. Gibson became chair.^[3]

A Recursion experiment follows a fixed routine. Cells are grown and given a genetic perturbation, a compound or both. One to two days later they are fixed and stained, then imaged.^[5] For years the staining was Cell Painting, a protocol developed at the Broad Institute that labels eight cellular components with dyes. According to an interview with a Recursion vice president, a project in 2021 suggested that plain brightfield images, which need no dyes and can be taken of living cells, carry enough signal, and by the summer of 2024 brightfield had become the main modality.^[5] The same interview put the archive at more than 19 petabytes of image and perturbation pairs.^[5]

The numbers come from a masked autoencoder. During training the model sees a cell image with most of it hidden, 75 percent in the version described in the interview, and learns to fill in the rest; the list of numbers for a picture is what its encoder produces.^[5] A joint embedding predictive architecture asks for less. It predicts a representation of the hidden part, where the autoencoder has to rebuild the pixels. In a paper from the company at CVPR 2024, masked autoencoders recalled known biological relationships, curated from public databases, with a relative improvement of up to 11.5 percent over weakly supervised classifiers.^[6] The software that explores the relationships is called the Recursion Map.^[1]

In June 2025 Recursion and researchers at the Massachusetts Institute of Technology released Boltz-2, an open model that predicts the structure of a protein and a ligand together and estimates how tightly they bind. The release says it approaches the accuracy of physics-based free energy calculations while running up to 1,000 times faster, and the code and weights are available under the MIT license.^[7] It was trained on the company's BioHive-2 supercomputer, which NVIDIA describes as a DGX SuperPOD of 63 DGX H100 systems.^[8]

In November 2025 the company reported data for REC-617, an oral inhibitor of the enzyme CDK7, in a Phase I/II trial for advanced solid tumors. Of 29 patients treated across six dose levels by September 29, one had a confirmed partial response and five had stable disease. Two stopped because of a treatment-related adverse event. The maximum tolerated dose was 10 milligrams once daily.^[3]

Since August 19, 2026, the front page of Molecular Reality Corporation has said "everybody's racing to build the brain, we're building the eye." The brain column in the company's chart has five names, and Recursion is not among them.^[9] Recursion describes the first layer of its platform as one that generates its own biological and chemical data.^[1]

• Brain lab
• Self-driving laboratory
• Isomorphic Labs
• Lila Sciences