BioHive is powered by NVIDIA AI to accelerate the work of scientists in the medical field. It has risen more than 100 places in the world's top 500 supercomputers.
图注：BioHive-2 与 Recursion 公司的员工。 从左至右：Paige Despain、John Durkin、Joshua Fryer、Jesse Dean、Ganesh Jagannathan、Chris Gibson、Lindsay Ellinger、Michael Secora、Alex Timofeyev 和 Ben Mabey。
BioHive-2 debuted at Recursion's Salt Lake City headquarters and is billed as the pharmaceutical industry's largest supercomputing system. In the latest edition of the world's TOP500 list of supercomputers, BioHive-2 ranked 35th, up more than 100 places from its predecessor.
This advancement represents the company's recent efforts to leverage NVIDIA technology to accelerate drug discovery.
Ben Mabey, CTO at Recursion, said, "We're seeing that, just like large language models, scaling up training with more data and computing power can dramatically improve the performance of AI models in biology, which ultimately has a positive impact on patients' lives." Ben Mabey has been exploring the use of machine learning in the medical field for more than a decade.
BioHive-2 搭载了 504 个 NVIDIA GPU,并通过 NVIDIA Quantum-2 InfiniBand 网络互联,可提供 2 exaflops 的 AI 性能。 NVIDIA DGX SuperPOD 因此比 Recursion 的初代系统 BioHive-1 快近 5 倍。
High performance solves complexity challenges
Performance is key to rapid progress because "biology is extremely complex," says Mabey.
Finding a new drug candidate can take scientists years and millions of experiments in wet labs.
This work is crucial. Recursion's scientists run more than 2 million of these experiments every week. But going forward, they will use AI models on BioHive-2 to direct their platform to the most promising areas of biology to run experiments.
"With AI, we now get 80 percent of the value of 40 percent of wet lab work, and that percentage will increase even further in the future," he says. ”
Biological data drives medical AI advancements
Recursion is working with biopharmaceutical companies such as Bayer AG, Roche and Genentech. Recursion has amassed more than 50 petabytes (petabytes) of biological, chemical, and patient databases, and has developed powerful AI models that accelerate drug discovery.
Mabey joined Recursion more than seven years ago, in part because of the company's commitment to building such datasets. "It's one of the largest biological datasets in the world, built with AI training in mind, and it includes both biological and chemical data," he said. ”
Create an AI atmosphere
By processing this data on BioHive-1, Recursion has developed a series of foundational models called Phenom. These models translate a series of microscope-observed images of cells into meaningful representations that can be used to understand the biology underlying them.
One of these models, Phenom-Beta, is now available as a cloud API and is the first third-party model on NVIDIA BioNeMo, a generative AI platform for drug discovery.
After months of research and iteration, BioHive-1 used more than 3.5 billion images of cells to train Phenom-1. Recursion's scaled system enables more robust models to be trained on larger datasets in less time.
The company also leverages NVIDIA DGX Cloud, hosted by Oracle Cloud Infrastructure, to provide additional supercomputing resources for its work.
Caption: Just like training a large language model to generate missing words in sentences, the Phenom model is trained to generate masked pixels in the image of the cell.
The Phenom-1 model serves Recursion and its partners in a variety of ways, including finding and optimizing molecules to treat a wide range of diseases and cancers. Early models have helped Recursion predict drug candidates for COVID-19, with 9 out of 10 successes.
Recursion announced its partnership with NVIDIA last July. Less than 30 days later, the combination of BioHive-1 and DGX Cloud successfully screened and analyzed a vast library of chemistries predicting protein targets for approximately 36 billion compounds.
In January, Recursion showcased LOWE, an AI workflow engine with a natural language interface that can help scientists more easily use the company's tools. In April, the company also mapped a self-developed billion-parameter AI model that aims to provide a new way to predict the properties of key molecules in the medical field.
Recursion uses NVIDIA software to optimize the system.
"We love CUDA and NVIDIA AI Enterprise, and we're looking at whether NVIDIA NIM will make it easier for us to release our models internally and to our partners," he said. ”
A shared vision for healthcare
In a fireside chat with Recursion's chairman, NVIDIA founder and CEO Jensen Huang described the vision of "simulated biology," and these achievements take that vision one step further.
"You can now recognize and learn the language of almost any structured object, and you can translate it into any structured object," Huang said...... This is the generative AI revolution. ”
"We have similar views," Mabey said. ”
He added: "We're in the early stages of a very interesting era, and just as computers have accelerated chip design, AI can also speed up drug design. Biology is much more complex, so it takes years to see results. But when one looks back, one can see that this was a real turning point in the medical field. ”
Learn about NVIDIA's AI platform for Health & Life Sciences:
And subscribe to NVIDIA Healthcare News: