Recursion Completes Supercomputer for AI Drug Discovery

Artificial intelligence (AI)-based drug developer Recursion said it has completed the fastest supercomputer to be wholly owned and operated by any pharmaceutical company worldwide, using technology developed by collaboration partner Nvidia.

The supercomputer, BioHive-2, ranks No. 35 on the TOP500 list of the world’s most powerful supercomputers across all industries as of this month.

BioHive-2 operates four times faster than Recursion’s original supercomputer BioHive-1 in benchmark performance tests, according to the company.

BioHive-2 consists of an NVIDIA DGX SuperPOD AI supercomputer, powered by 63 DGX H100 systems with a total of 504 NVIDIA H100 Tensor Core GPUs interconnected by NVIDIA Quantum-2 InfiniBand networking.

“Recursion has spent the last decade generating and aggregating one of the largest biological and chemical datasets in the world, purpose-built for training new AI models. With BioHive-2 now online, we have significantly more computational horsepower to accelerate our use of our ever-growing dataset, extending our ability to train larger and more generalizable foundation models and AI agents to industrialize our drug discovery efforts,” Ben Mabey, CTO, said in a statement.

Recursion has also developed new foundation models that include Phenom-1, a deep-learning model designed to extract biologically meaningful features from images of cells. The company says it has demonstrated a need for training larger models since the performance of its model has increased as the size of the training data and the number of model parameters have grown.

The experimentation and training to produce Phenom-1 required several months of computational time using BioHive-1. By developing BioHive-2, Recursion said, it can now execute multiple AI projects of similar or greater size in parallel in shorter timeframes, enabling teams from both Recursion and its Valence subsidiary to pursue advanced AI drug discovery and unlock additional value from Recursion’s data.

Valence is a pioneer in low-data small molecule drug design which Recursion bought for $47.5 million last year.

Investors signaled approval of Recursion’s supercomputing advance with a buying surge that sent the company’s stock rising 10% Monday, from $8.60 to $9.48. Nvidia was all but flat, rising 0.58% from $898.78 to $903.99.

A smaller model similar to Phenom-1, called Phenom-Beta, was released for external use earlier this year through NVIDIA BioNeMo, a generative AI cloud-based platform designed to enable faster discovery and design of drugs. Phenom-Beta is designed to flexibly process cellular microscopy images into general-purpose embeddings at any scale, from small projects to billions of images.

“Meaningful representations”

“Phenom-Beta can turn a series of image inputs into meaningful representations that are foundational to analyzing and understanding the underlying biology,” Mabey wrote 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.

The Phenom series is among several different models Recursion has developed to accelerate drug discovery using biological, chemical, and real-world patient data. Phenom-Beta is the first third-party model to be made available on BioNeMo.

Last year Recursion’s Nvidia-based platform enabled the AI drug developer to develop a large protein-ligand interaction prediction dataset. Recursion predicted the protein target or targets for approximately 36 billion chemical compounds on Enamine’s REAL Space chemical library, which consists of 48 billion make-on-demand molecules.

REAL Space offers the largest supply of commercially available compounds, according to Kyiv, Ukraine-based Enamine and Salt Lake City, UT-based Recursion. Last December, the companies launched a collaboration of undisclosed value to generate compound screening libraries with insights from Recursion’s protein-ligand interaction predictions spanning across Enamine’s massive library.

An NVIDIA DGX SuperPOD based system called Eos ranks No. 10 among supercomputers worldwide. Eos is based on the NVIDIA DGX H100 with Xeon Platinum 8480C processors, NVIDIA H100 accelerators, and Infiniband NDR400 and it achieves 121.4 Pflop/s.

“Accelerated computing, combined with the power of generative AI, is propelling the pharmaceutical industry into a new, advanced era of drug discovery,” said Rory Kelleher, global head of business development for life sciences at Nvidia. “BioHive-2, powered by NVIDIA DGX AI supercomputing, is poised to accelerate the development of additional industry-leading foundation models across biology, chemistry, and healthcare.”

No Byline Policy

Editorial Guidelines

Corrections Policy

Source