Recursion at JPM: Exscientia Merger and the Coming Virtual Cell

By Allison Proffitt 

February 4, 2025 | Chris Gibson, CEO of Recursion, gave the company update last month at the 2025 J.P. Morgan Healthcare Conference.  

Gibson gave updates on several Recursion pipeline programs for indications including precision oncology, rare disease, and other indications. He explored six candidates and credited the successes thus far to the Recursion Operating System. The Recursion OS includes both the wet lab operations of Recursion—“generating massive quantities of real data across hundreds of millions of perturbations, generating rich, high dimensional ‘omics data, phenomics, transcriptomics, proteomics, and now increasingly different kinds of chemical data, automated chemical synthesis, etc.”—and the machine learning models the company is building, including the latest foundation model  for chemistry the company published with Novo Nordisk.  

These models, “are taking that [wet lab] data and learning to understand and identify patterns in those data so that we can increasingly predict all of the experiments we have not run yet,” Gibson said. “We began with a point solution in phenomics and today we are generating or aggregating with our partners data from patients to cells, cells to organoids, organoids to animals, and animals to people in the clinic.”  

Gathering data and training machine learning and AI models across the various layers is the secret to industrializing drug discovery, Gibson believes. And it’s also leading the drug discovery space to a tipping point.  

While in years past, Gibson has shown the Recursion flywheel where real world research feeds the model and hypotheses are sent back to be tested in the lab, we are quickly approaching a dramatic inversion of that diagram, he said.  

“As these models become really, really good, as they become highly predictive, you essentially can build what we’ll call here a virtual cell,” Gibson said. “All of the sudden you have this transposition where your wet lab becomes the validation tool for the simulated output of the virtual cell, as opposed to… your wet lab as a data generation engine to build, one day, a virtual cell.”   

Virtual Cell Building Blocks 

Building a virtual cell that can simulate biology in a robust way requires data and algorithms at three different layers, Gibson said: strong pathway models, strong protein models, and strong atomistic models.  

Recursion is already “far ahead” in pathway models, Gibson believes. “We’ve already knocked out every gene in the genome of dozens of different cell types. We’ve put millions of chemical perturbations on top of these cells. We have rich omics data from phenomics to transcriptomics. We have time course data increasingly with bright field imaging. All of these datasets are giving us a very, very robust understanding of the gene networks that are at play.”   

For the protein model layer, Gibson said Recursion is dedicated to partnering. AlphaFold kicked off a revolution, Gibson said, led by open-source companies. “We believe this layer will be commoditized, and that if you partner with the right group—which we are doing—you will have access to the state-of-the-art frontier protein folding, protein-protein interactions, protein-ligand interaction models.” Recursion already has a partner in this space, Gibson hinted, but didn’t name names. “Give us a few months,” he said. “We think they are on the very forefront of protein folding.”   

Finally for the atomistic layer, Gibson is leaning on Recursion’s August 2024 merger with Exscientia. “Exscientia had built an incredible chemistry platform that was allowing them to really generate best-in-class programs quickly across really challenging multi-parameter optimization situations. And Recursion had built sort of a biology-first platform that was allowing us to find first-in-class biology. What we hope to do, if we bring the companies together in the best way, is find programs that are going to be simultaneously both first-in-class and best-in-class.”  

Gibson predicted many synergies. The combination of Exscientia’s chemistry platform with Recursion’s biology platform will hopefully “dramatically reduce the amount that we’re outsourcing,” Gibson said, “Sorry to the CROs!” He is also excited about the expertise the merger brings together: “a team who’s taken programs from start to clinic from those two companies! That’s a rare set of people in the tech bio field.”  

Through the Exscientia merger, Gibson said, Recursion gained a quantum mechanics/ molecular mechanics (QM/MD) team and dataset that will position the company for success in the atomistic layer. “We can’t share everything today, but in the coming months, I think we will be able to tell the world that we are the leaders in this layer as well.” 

Even with predicted dominance in the three layers needed for a virtual cell, Gibson acknowledged challenges. While many foundation models are good at modeling specific areas, building multi-modal models with different kinds of data able to tackle different kinds of questions is more challenging. “I think that’s likely to be the biggest challenge that we face: building the single model,” he said.  

When asked which cell type he planned to start with, Gibson said Recursion has the most data in the human umbilical vein endothelial cells, dating back to Gibson’s own vascular biology lab days, though various Recursion partnerships have yielded large datasets in GI oncology, neuroscience, and more.  

LOWE Update 

It wasn’t until Q&A that Gibson mentioned LOWE, the large language model/ orchestration work engine that dominated his presentation at last year’s J.P. Morgan Healthcare Conference. “LOWE was rolled into our collaboration with Bayer,” Gibson said. “The team at Bayer is now using LOWE with our scientists to collaborate on all of the programs that are advancing through the pipeline together.”  

This is indicative of all of the partnerships Recursion wants to pursue in the future. “I think we’ll find ways to learn and to teach with our large pharma partners, and we won’t do a partnership if it’s not like that. We want learning and teaching to go both ways.”