NVIDIA, Genentech Launch Multi-Year Generative AI Research Collaboration
By Allison Proffitt
November 21, 2023 | Genentech and NVIDIA announced today a multi-year strategic AI research collaboration to accelerate drug discovery and development.
“Our collaboration will begin with optimizing Genentech’s proprietary machine learning models on NVIDIA DGX Cloud, our AI supercomputer that offers training-as-a-service, including our software platforms NVIDIA BioNeMo for generative AI applications in drug discovery,” said Kimberly Powell, NVIDIA’s vice president of healthcare, in an early briefing last week.
NVIDIA DGX Cloud provides dedicated instances of AI supercomputing and software hosted by NVIDIA cloud service provider partners. NVIDIA BioNeMo enables biotech companies to customize models at scale and integrate BioNeMo cloud application programming interfaces directly into computational drug discovery workflows. Genentech will use these computing resources to speed discovery with their proprietary data biological data.
“Generative AI has arrived at just the right time,” Powell said. “Today more than ever in history, we can generate [data]—and Genentech does! They can generate terabytes of digital biology in days, petabytes in months. And we can use generative AI methods and combine it with powerful simulation to computationally, in silico, drive this drug discovery process.”
“The scale of the search space is enormous!” agreed John Marioni, SVP and Head of Computational Sciences, Genentech Research and Early Development. “There are thousands of genes, millions of variants, all part of the processes that are underlying how cells are active,” he said. “Identifying which one of these different processes to target is, itself, hard enough, but then you add the complexity of inter-individual variability onto this, and the fact that there are tens of thousands of diseases, and you can imagine that the target identification is really challenging. Once you’ve got that target, discovering an antibody that can target it, discovering a small molecule that’s going to have the effect you want, that’s—again—a trillion parameter space.”
Genentech’s approach is a “lab-in-a-loop” framework: experimental data feeds computational models that uncover patterns and make new, experimentally testable predictions. The goal is for scientists to quickly assess these predictions in the lab and feed results back into the models, allowing for iterative development of better therapies. Genentech has petabytes of proprietary data, not to mention public data, to feed these computational models, Marioni said. The challenge is accelerating the iterative compute processes.
“It is the ability to do this quickly that is going to be transformative,” he said. “The tightness of that loop: going from the experimental data into the AI model predictions, and then synthesizing in the lab, testing, closing that loop, making it as tight as possible is absolutely essential. And so much of that is around accelerated scientific computing. If we’re able to do that well, it’s going to get us to solutions so much faster.”
And NVIDIA is certainly fast. Earlier this month, NVIDIA posted the results of several benchmarking tests. In MLPerf HPC, a benchmark for AI-assisted simulations on supercomputers, NVIDIA GPUs trained OpenFold, a model that predicts the 3D structure of a protein from its sequence of amino acids, in just 7.5 minutes. A version of the OpenFold model and the software NVIDIA used to train it will be available soon in NVIDIA BioNeMo, a generative AI platform for drug discovery, the company said.
“This is the kind of acceleration where—especially when you’re looking at lab-in-a-loop and the pace at which you can generate data and being able to retrain models at speeds like that—that loop is going to go from something that would otherwise be several months into hours or days,” Powell said.
On NVIDIA’s side, NVIDIA AI experts will gain insights into AI-related challenges in drug discovery and development, and plans to use these insights to improve its BioNeMo platform and others to further accommodate the requirements of models used by the biotech industry.
“Our teams will be continuously exchanging expertise on the advancement of science, the state of the art on methods emerging in accelerated computing, and AI and simulation across the entire drug discovery process,” Powell added.
Marioni declined to say which indications Genentech was applying this accelerated process to first. He expects the strategic collaboration to have “impact right across all the areas within our organization.” Both declined to share specific milestones or any financial details.
When asked if they foresee this collaboration as applicable to work beyond drug discovery—into clinical trials and manufacturing, for instance—they both emphasized the current focus is certainly research and early development. But, Marioni added, the strategic collaboration is structured in such a way that other parts of Genentech and Roche could propose projects.
And Powell predicted that the ability to input desirable properties into a generative AI model will lead to thinking more broadly for all parts of the pharma value chain.
“If you know what’s necessary to be successful for manufacturing, can you—upstream of that—inject those properties into the generative and predictive capabilities of the model? That is absolutely what we can do. I do think you’re going to see this interesting and tight connection,” she said.