Announcing an artificial intelligence (AI) drug discovery partnership with Nvidia earlier this month, Recursion co-founder and CEO Chris Gibson, PhD, spoke as much about dollars as he did about drugs or data.
AI, Gibson asserted, will succeed where government policies, politicians, and payers haven’t—namely in finally reining in the sky-high and climbing costs of new treatments. But it may take a little longer than the 5-to-10-year timeframe he called his “pie-in-the-case” ideal scenario before the savings emerge, he cautioned.
“If it doesn’t happen in 5–10 years, it’s going to happen in 10–15,” Gibson said recently, addressing reporters from GEN Edge and other news outlets during a wide-ranging Q-and-A session at an evening reception held in San Francisco during the recent 42nd Annual J.P. Morgan Healthcare conference.
AI will cut drug discovery costs, Gibson explained, by dramatically reducing the expense associated with failed drug candidates, and doing so far earlier in the development process than is now the case.
“It makes failure cheap and easy at the early stage,” Gibson said. “So, we can explore more broadly, and every scientist can do 5 or 10 new programs a day. It doesn’t mean that they’re all going to translate. In fact, 99% of them aren’t. But the failure’s happening in the first six hours or the first six weeks of the first six months, as opposed to happening six years down the road in Phase II or Phase III.”
“I’m sure we’ll have Phase II or III failures. That’s inevitable. The question is, how can we reduce that failure rate over time? It’s tools like this that are going to help scientists explore more at scale,” Gibson added. “That’s why I’m so convinced—I think it’s inevitable—that in 10–15 years, ML and AI will have dramatically reduced prices.”
By how much, he was asked?
“I don’t think I could make a prediction,” Gibson acknowledged.
Phenomics foundation
At the glitzy, invitation-only event, highlighting the Nvidia-Recursion partnership, Recursion announced it will be the first hosting partner of Nvidia’s to release a potential series of AI foundation models for external use, to be hosted on Nvidia’s BioNeMo™ generative AI cloud-based platform designed to enable faster discovery and design of drugs. The series is called “Phenom,” a play on the words “phenomenal” and “phenomics,” the latter being the systematic study of a cell’s phenotype in response to many different chemical or genetic perturbations.
“Recursion has been so far ahead of the rest of the field in our use of images of human biology, and in particular images of human cells to build a new kind of omics. Rather than looking at genomics, we’re looking at what we call phenomics,” Gibson said. “We think this will become as big and exciting as genomics. It’s just a new layer of biology, and it’ll be in there with proteomics, metabolomics, etc.”
“I think we’re far enough ahead that we felt confident it made sense for us to share some of what we’ve learned with the broader industry, and we wanted to do that with the very best technology partner,” Gibson explained. “We think NVIDIA is really well placed as the company where we wanted to be partnered.”
One phenomics foundation model released by Recursion to researchers “subject to commercial limitation” is Phenom-Beta, which is designed to flexibly processes cellular microscopy images into general-purpose embeddings at any scale, from small projects to billions of images.
“Phenom-Beta can turn a series of image inputs into meaningful representations that are foundational to analyzing and understanding the underlying biology,” Recursion chief technology officer Ben Mabey wrote January 9 on the company’s blog.
Phenom-Beta was trained using the RxRx3 dataset, a publicly available dataset Recursion released last year that contains approximately 2.2 million images of HUVEC cells across ~17,000 genetic knockouts and 1,674 known chemical entities. “We were excited to discover that as we increased the size of the training data and the number of parameters, the model’s performance increased,” Mabey observed.
Mabey added that Recursion’s most advanced foundation model, Phenom-1, “is currently in production for our internal teams and close partners.”
Recursion also demonstrated its new software designed to perform complex drug discovery tasks using a natural language interface called LOWE, short for Large Language Model-Orchestrated Workflow Engine.
Those steps and tools range from finding significant relationships within Recursion’s Maps of Biology and Chemistry, to using Recursion’s MatchMaker digital chemistry tool (acquired when the company bought Cyclica in May 2023 for $40 million) to identify drug-target interactions; to generating novel compounds and scheduling them for synthesis and experimentation.
Using MatchMaker, Recursion successfully screened Enamine REAL Space, reportedly the world’s largest searchable chemical library, to predict the protein target or targets for the approximately 36 billion chemical compounds contained within.
Much of the initial testing and infrastructure development for the screening project was completed using Recursion’s in-house supercomputer BioHive-1 and an NVIDIA DGX SuperPOD, ranked No. 9 among the top 125 most powerful supercomputers in the world across any industry by TOP500 as of November 2023. The final analysis was made possible by NVIDIA’s DGX Cloud, an advanced AI-training-as-a-service solution.
$50M investment
Recursion gained access to DGX Cloud after Nvidia invested $50 million in the Salt Lake City, UT-based AI drug developer, a partnership the companies announced last July. That money is a proverbial drop in the bucket compared to the $1 billion that, according to Gibson, Recursion has spent developing its Recursion Operating System (OS). Recursion OS is designed for generating, analyzing and deriving insight from massive biological and chemical datasets to industrialize drug discovery.
Through a private placement, Nvidia acquired more than 7.7 million shares of Recursion’s Class A common stock at $6.49 per share on July 11, 2023—a 4% discount from that day’s closing share price of $6.78. Through the collaboration, Recursion agreed at the time to accelerate development of its AI foundation models for biology and chemistry by using Nvidia’s cloud services.
Also as part of the partnership, Recursion also committed to working with Nvidia to expand its BioHive-1 supercomputer over four times, by adding more than 500 NVIDIA H100 GPUs to the more than 300 NVIDIA A100s already in place, Recursion disclosed November 9 in its most recent Form 10-K quarterly report for the third quarter of 2023. The expansion is set to be completed in the first half of 2024.
“We project that upon completion and benchmarking, BioHive-1 will be in the top 50 most powerful supercomputers in the world across any industry (according to the TOP500 list) and will be the most powerful supercomputer owned and operated by any biopharma company,” Recursion predicted.
A significant source of the drug development cost savings projected by Gibson as a result of AI is envisioned to come from changing how clinical trials are conducted.
An often-quoted 2020 study published in BMJ Open reported an estimated median cost of $48 million for the 225 pivotal clinical trials conducted for the 101 new molecular entities approved by the FDA between 2015-2017—$87.2 million in today’s dollars. The 225 individual clinical trials had a median estimate of $19 million ($34.5 million today) per trial and $41,413 ($75,256) per patient.
“We’re going to design a trial that’s ten times less expensive. It’s because we’re going to pick a target based on much broader systems biology data that on average, across hundreds of targets, I think will have eventually a higher probability of success,” Gibson said. “At the end of the day, if you were to go out to our first 10 readouts and then our next 10 and then our next 100 and do this across all the companies in the space, we think we will bring down the price.”
“I don’t think we’re going to simulate clinical trials for 20 years, but we’re going to have synthetic control arms,” Gibson said.
Nine disclosed programs
Recursion is eager to cut clinical trial costs as it develops its pipeline, which includes nine disclosed programs. Five are in clinical phases, including its lead oncology program targeting AXIN1 or APC mutant cancers (REC-4881), to be assessed in the Phase II LILAC trial set to start by early Q1 2024; and all four of its rare and other disease programs:
- Cerebral cavernous malformation (REC-994), now in the Phase II SYCAMORE trial (NCT05085561), set to release proof-of-concept data in the second half of 2024.
- Neurofibromatosis type 2 (REC-2282), now in the Phase II/III POPLAR trial (NCT05130866), for which Phase II safety, tolerability, pharmacokinetics, and preliminary efficacy data is also expected to be shared in 2H 2024.
- Familial adenomatous polyposis (REC-4881), under study in the Phase II TUPELO trial (NCT05552755), for which Phase II safety, tolerability, pharmacokinetics, and preliminary efficacy data is expected in the first half of 2025, and
- Clostridioides difficile infection (REC-3694), for which Recursion expects to launch a Phase II proof-of-concept trial this year. The company completed a Phase I study in September 2023, reporting that REC-3964 had been well tolerated in healthy volunteers with no serious adverse events.
Recursion is also in preclinical phases for programs for RBM39 HR-proficient ovarian cancer (targeting RBM39; IND-enabling studies), and a cancer immunotherapy (target Delta), as well as two discovery phase programs, cancer immunotherapy (target Alpha) and MYC-driven oncology.
Streamlining trials by using patient data to identify novel cancer targets is a keystone of a Recursion collaboration also announced this month. Recursion agreed to pay Tempus $160 million cash or equity over five years, in return for preferred access to more than 20 petabytes of what the companies called one of the world’s largest proprietary, de-identified, patient-centric oncology datasets, spanning DNA, RNA, health records, and more. The data is intended to support the discovery of potential biomarker-enriched therapeutics at scale through the training of causal AI models.
“Our dream is to go back to the same patient, six or 12 months later, whose genetics helped us find a potential treatment, and try to enroll them in a study,” Gibson said. “That molecule is more likely to do well if you are treating the patient who also helped you find the medicine. This is the future: it’s creating virtuous cycles of learning and iteration.”
“We will undercut them”
What’s to stop biopharmas from simply pocketing the money they would save from lower clinical trial expenses thanks to AI, Gibson was asked.
“If they do that, we’re going to find a drug and undercut them on price,” he replied. “If we’re successful enough to be around in 10 years and a drug company is doing that, I will have built with our team a machine that will allow us to discover a medicine with a much lower failure rate, maybe against the same disease they’re going after, and we will undercut them on price.”
It’s a response, Gibson said, that has been successfully applied in retail: “That’s exactly what Amazon did: They made it easy, convenient, faster predictive. And what did they do? They brought down prices.”
Gibson acknowledged the challenging capital markets climate for Recursion and other biopharmas but said that didn’t and wouldn’t stop his company from pursuing acquisitions like Cyclica and Valence, a pioneer in low-data small molecule drug design which Recursion bought for $47.5 million last year.
“You’ll see us continue to take and make bold bets because I think ultimately whether or not Recursion is successful, we’re running an experiment that is ultimately going to give us or others behind us a lot of data about how one could try and take a swing at discovering and developing medicines better,” Gibson said.
“I think it’s going to be us. But even if it’s not, it’s the company behind us. It will absolutely be different in the next ten years,” Gibson added. “Other CEO’s and heads of R&D’s from many of the large pharma companies, they are starting to believe and they’re hiring incredible people. If you’re not already on board with building in this [AI] space in biopharma, in eight years you will look like the equivalent of, let’s say, what a large traditional auto company looks like in the space of transportation.”
