The LOWE Down on Recursion’s New LLM Orchestration Work Engine from JP Morgan
By Allison Proffitt
January 9, 2024 | About halfway through his talk yesterday at the JP Morgan Healthcare conference, Chris Gibson, CEO of Recursion, took a risk. “I’m going to do something pretty new for JP Morgan,” he quipped, “which is that I’m going to exit PowerPoint.” The comment drew laughs from the audience, even as Gibson launched a live demo of Recursion’s new large language model orchestration work engine, LOWE.
Gibson began his presentation outlining many of the Recursion talking points: the company’s mission to create the right datasets to map and navigate biology, creating a T-shaped funnel that will allow fast and early failure, and industrializing drug discovery through the Recursion OS.
Data volume is the Recursion strategy. The robotic wet lab does up to 2.2 million phenomics experiments a week, and each station is monitored by a camera, Gibson reported. Another roboticized system sequences up to 100,000 exomes per week. The Recursion vivarium has cameras in cages of “hundreds” of animals, extracting data on animal behavior and study progression—identifying signals of toxicity months before it is evident on histology slides, he added. Recursion’s cell images maps span trillions of relationships. And the company built its own DMPK data generation platform.
“This is what it takes to win in the field of TechBio,” Gibson said. “You have to build data or access data that you can deeply, deeply trust. And you have to build it in a way that allows you to aggregate that data over time.”
TechBios generally begin with a point solution, Gibson observed. It was Recursion’s genesis too, he noted, referring to an August 2016 paper on Cell Painting. But changing the drug discovery industry will require many integrated point solutions as modules across many diverse steps. This means continuously adding and updated modules to the Recursion OS, many of which, Gibson said, were originally only available at the command line to highly trained data scientists.
But when Gibson left the safety of the slide deck and entered the wilds of conference internet access, he was presenting what he believes will be the future of TechBios: a tool to leverage large language models not to understand the world, but as orchestrators of workflows.
LOWE Live
To start his demo of LOWE, Gibson asked for a list of targets known to be involved in non-small cell lung cancer. LOWE isn’t only scanning published literature; it’s delving into Recursion’s proprietary data cache. “The differentiator for us is that we have 50 petabytes of proprietary biological data at our fingertips, and so we’re in a position to do much more than just look at public data,” Gibson said.
He narrowed his search to targets with similar phenotypes as known cancer-causing genes; picking one of those genes, he used Recursion’s MatchMaker to find the top 100 active compounds for his gene of choice. LOWE returned the molecules with the highest probability of interacting with that gene, filtered by solubility, and placed orders for those molecules from chemistry providers.
“This will be the part that’s a little bit not real,” Gibson quipped. “We’re going to pretend like we ordered those compounds and that it’s been six to eight weeks, and they’ve arrived, and our team has unpacked them.”
With compounds in hand, LOWE can connect to the experimental design tool and design and run phenomics experiments in the integrated wet lab. With the resulting hits, LOWE applies generative modeling to identify novel molecules based on the hits with improved ADME properties, similar chemical properties, and similar 3D structures.
“We can see that this generative AI algorithm has generated new molecules that may or may not exist in anybody’s catalog, that may or may not be orderable that we could then custom-synthesize and take forward,” he said.
Gibson firmly believes that this is a vision of the future.
“I think this represents how drug discovery will be done at every company in the industry five to ten years from now,” he said. “It means that any scientist, regardless of their training, can interact with a wide variety of tools… Each new version of each tool is now instantly accessible at the fingertips of our scientists, and in many ways, it’s not a graphical user interface, it’s a discovery user interface.”
Between Now and Then
Back in the familiarity of PowerPoint, Gibson recounted Recursion’s other 2023 wins including partnerships with Genentech and Bayer, multiple Phase 2 trials, and the creation of Phenom-1, the largest phenomics-based foundational model. As part of Recursion’s ongoing partnership and investment with NVIDIA, Phenom-Beta is now available on BioNeMo.
The company also closed two acquisitions in May 2023. LOWE was developed at Valence Labs. Cyclica built two modules now in the Recursion OS: MatchMaker, an AI-enabled deep learning engine that predicts the polypharmacology of small molecules as the foundation for small molecule drug discovery, and POEM (Pareto Optimal Embedding Model), a similarity-based property prediction model.
In the near-term future, Gibson predicted more INDs from partnerships, more Phase 2 trial starts and data readouts, and additional partnerships. But the most ambitious goal: Recursion OS moves toward autonomous discovery.
“The future looks really, really bright,” he said in closing, “not only for Recursion, but I think for the TechBio industry and the biopharma industry writ large.”