Described as the biggest system within the pharmaceutical trade, BioHive-2 on the Salt Lake Metropolis headquarters of Recursion debuts at present at No. 35, up greater than 100 spots from its predecessor on the most recent TOP500 checklist of the world’s quickest supercomputers.
The advance represents the corporate’s most up-to-date effort to speed up drug discovery with NVIDIA applied sciences.
“Simply as with giant language fashions, we see AI fashions within the biology area enhance efficiency considerably as we scale our coaching with extra knowledge and compute horsepower, which finally results in better impacts on sufferers’ lives,” mentioned Recursion’s CTO, Ben Mabey, who’s been making use of machine studying to healthcare for greater than a decade.
BioHive-2 packs 504 NVIDIA H100 Tensor Core GPUs linked on an NVIDIA Quantum-2 InfiniBand community to ship 2 exaflops of AI efficiency. The ensuing NVIDIA DGX SuperPOD is sort of 5x sooner than Recursion’s first-generation system, BioHive-1.
Efficiency Powers By Complexity
That efficiency is vital to speedy progress as a result of “biology is insanely advanced,” Mabey mentioned.
Discovering a brand new drug candidate can take scientists years performing thousands and thousands of wet-lab experiments.
That work is important; Recursion’s scientists run greater than 2 million such experiments every week. However going ahead, they’ll use AI fashions on BioHive-2 to direct their platform to essentially the most promising biology areas to run their experiments.
“With AI within the loop at present, we are able to get 80% of the worth with 40% of the moist lab work, and that ratio will enhance going ahead,” he mentioned.
Organic Knowledge Propels Healthcare AI
Recursion is collaborating with biopharma corporations similar to Bayer AG, Roche and Genentech. Over time, it additionally amassed a greater than 50-petabyte database of organic, chemical and affected person knowledge, serving to it construct highly effective AI fashions which can be accelerating drug discovery.
“We consider it’s one of many largest organic datasets on Earth — it was constructed with AI coaching in thoughts, deliberately spanning biology and chemistry,” mentioned Mabey, who joined the corporate greater than seven years in the past partly on account of its dedication to constructing such a dataset.
Creating an AI Phenomenon
Processing that knowledge on BioHive-1, Recursion developed a household of basis fashions referred to as Phenom. They flip a sequence of microscopic mobile pictures into significant representations for understanding the underlying biology.
A member of that household, Phenom-Beta, is now out there as a cloud API and the primary third-party mannequin on NVIDIA BioNeMo, a generative AI platform for drug discovery.
Over a number of months of analysis and iteration, BioHive-1 skilled Phenom-1 utilizing greater than 3.5 billion mobile pictures. Recursion’s expanded system allows coaching much more highly effective fashions with bigger datasets in much less time.
The corporate additionally used NVIDIA DGX Cloud, hosted by Oracle Cloud Infrastructure, to offer extra supercomputing assets to energy their work.
The Phenom-1 mannequin serves Recursion and its companions in a number of methods, together with discovering and optimizing molecules to deal with quite a lot of ailments and cancers. Earlier fashions helped Recursion predict drug candidates for COVID-19 9 out of 10 occasions.
The corporate introduced its collaboration with NVIDIA in July. Lower than 30 days later, the mixture of BioHive-1 and DGX Cloud screened and analyzed an enormous chemical library to foretell protein targets for about 36 billion chemical compounds.
In January, the corporate demonstrated LOWE, an AI workflow engine with a natural-language interface to assist make its instruments extra accessible to scientists. And in April it described a billion-parameter AI mannequin it constructed to offer a brand new technique to predict the properties of key molecules of curiosity in healthcare.
Recursion makes use of NVIDIA software program to optimize its techniques.
“We love CUDA and NVIDIA AI Enterprise, and we’re seeking to see if NVIDIA NIM might help us distribute our fashions extra simply, each internally and to companions,” he mentioned.
A Shared Imaginative and prescient for Healthcare
The efforts are a part of a broad imaginative and prescient that Jensen Huang, NVIDIA founder and CEO, described in a fireplace chat with Recursion’s chairman as shifting towards simulating biology.
“Now you can acknowledge and be taught the language of just about something with construction, and you’ll translate it to something with construction … That is the generative AI revolution,” Huang mentioned.
“We share an analogous view,” mentioned Mabey.
“We’re within the early phases of a really fascinating time the place simply as computer systems accelerated chip design, AI can velocity up drug design. Biology is way more advanced, so it’s going to take years to play out, however wanting again, individuals will see this was an actual turning level in healthcare,” he added.
Study NVIDIA’s AI platform for healthcare and life sciences and subscribe to NVIDIA healthcare information.
Pictured at high: BioHive-2 with just a few members of the Recursion crew (from left) Paige Despain, John Durkin, Joshua Fryer, Jesse Dean, Ganesh Jagannathan, Chris Gibson, Lindsay Ellinger, Michael Secora, Alex Timofeyev, and Ben Mabey.
