Recursion's BioHive-2, Powered by NVIDIA GPUs, Joins World's Top 35 Supercomputers

NVIDIA has expanded its partnership with clinical-stage biotech startup Recursion, utilising its AI supercomputer, BioHive-2, to expedite pharmaceutical research and development. 

BioHive-2 integrates 504 NVIDIA H100 Tensor Core GPUs within an NVIDIA Quantum-2 InfiniBand network, achieving a staggering two exaflops of AI performance. This system is nearly five times faster than the initial BioHive-1, facilitating more efficient processing of complex biological data. The resulting NVIDIA DGX SuperPOD is nearly five times faster than Recursion’s first-generation system, BioHive-1.

Positioned at Recursion's headquarters in Salt Lake City, BioHive-2 now ranks 35th on the Top 500 list of the world's fastest supercomputers, marking a significant leap from its predecessor.

“Just as with large language models, we see AI models in the biology domain improve performance substantially as we scale our training with more data and compute horsepower, which ultimately leads to greater impacts on patients’ lives,” said Recursion’s CTO, Ben Mabey.

The power of BioHive-2 allows Recursion's scientists to conduct AI-driven experiments, greatly reducing the dependency on traditional wet-lab research. By incorporating AI into workflows, the team can achieve approximately 80% of the outcomes with only 40% of the hands-on lab work.

Under the Hood

Founded in 2013 by Christopher Gibson, Recursion integrates experimental biology, bioinformatics, and artificial intelligence on a combined lab-to-cloud platform. This approach enables the identification of disease treatments that can be modelled using cellular data.

Its collaboration with global biopharma leaders, including Bayer AG, Roche, and Genentech, is supported by a vast database exceeding 50 petabytes of biological, chemical, and patient data. This repository not only enhances the training of its AI models but also speeds up the discovery and optimisation of new drugs. 

One of its notable achievements includes the Phenom family of foundation models, which transforms microscopic cellular images into data models, streamlining the understanding of underlying biological processes.

Phenom models, for example, assist in discovering and improving treatments for diseases and cancers, with previous models accurately predicting COVID-19 drug candidates.

The partnership initiated in July between NVIDIA and Recursion quickly demonstrated its value. 

In less than a month, using BioHive-1 and NVIDIA DGX Cloud, they analysed a vast chemical library to predict protein targets for billions of chemical compounds. Its innovations include LOWE, an AI workflow engine with a natural-language interface designed to make scientific tools more accessible, and a billion-parameter AI model built to predict molecular properties crucial in healthcare.

The shared vision between NVIDIA and Recursion emphasises AI's transformative potential in simulating and understanding complex biological structures.