Bristol Myers Squibb is significantly boosting its artificial intelligence capabilities by acquiring an NVIDIA DGX SuperPOD built on the chipmaker’s cutting-edge Vera Rubin architecture. This strategic investment underscores BMS’s commitment to revolutionizing drug discovery and development through advanced computing power. The pharmaceutical giant is set to become the first in the life sciences sector to deploy a DGX SuperPOD featuring the Vera Rubin architecture, which NVIDIA unveiled earlier this year as the successor to its current generation of AI computing systems.
The new high-performance computing cluster will comprise eight DGX Vera Rubin NVL72 systems. Each of these rack-scale systems integrates NVIDIA’s Vera central processing units and Rubin graphics processing units, delivering unparalleled computational throughput for complex AI workloads. BMS plans to leverage this formidable infrastructure to train proprietary AI models and execute sophisticated predictions across its extensive research programs. This will encompass critical work involving compounds, proteins, and a vast array of other scientific data crucial for therapeutic innovation.
While financial terms of the acquisition remain undisclosed, this move represents a substantial expansion of BMS’s existing NVIDIA infrastructure. Company executives noted that their current DGX SuperPOD, which has been operational for approximately three years, is now two to three generations behind the Vera Rubin architecture. The plan is to seamlessly integrate the new Vera Rubin system with their existing infrastructure, creating a unified computing environment accessible from BMS research sites worldwide. The comprehensive SuperPOD software stack will enable intelligent scheduling of training, prediction, and development workloads across the entire infrastructure. This expansion is expected to provide a broader spectrum of scientists with direct and democratized access to these powerful computing resources.
Greg Meyers, BMS’s Chief Digital and Technology Officer, highlighted the escalating computational demands driven by the deployment of increasingly sophisticated AI models across the organization’s research endeavors. Erin Davis, Vice President of Research Business Insights and Technology at BMS, echoed this sentiment, stating that the current infrastructure is operating at full capacity. She attributed this surge in demand to large-scale predictions involving complex molecules and the development of proprietary internal foundation models. Importantly, Davis emphasized that the new system will not be confined to a select group of computational researchers. BMS intends to make it broadly available across the research organization, eliminating the waiting periods and access limitations that have characterized the use of its current infrastructure.
The integration of AI is profoundly transforming BMS’s approach to drug discovery. The company revealed that AI now informs the design of every small-molecule program and the majority of its large-molecule programs. This technology is strategically applied across multiple stages of the R&D pipeline, including target identification, lead optimization, large-molecule predictions, and the development of internal AI models. AI-enabled target identification, for instance, has already demonstrated the potential to significantly reduce manual research efforts by several weeks. The computationally intensive nature of large-molecule prediction workloads is a primary driver for the increased demand for advanced graphics processing capacity.
Robert Plenge, BMS’s Chief Research Officer, expressed optimism about the new system’s potential to empower scientists to evaluate a far greater number of potential drug candidates in the early stages of development. “Maybe before we could do 10 and now we can do dozens,” Plenge remarked, illustrating the dramatic leap in scale. This enhanced computational screening allows researchers to meticulously assess potential compounds before narrowing down to a smaller, more promising cohort for synthesis and rigorous laboratory testing.
BMS employs a novel methodology it terms “Predict First.” This approach utilizes model-generated predictions to effectively filter out molecules that do not meet predefined essential properties before candidates are even considered for synthesis. Payal Sheth, Senior Vice President of Therapeutic Discovery Sciences at BMS, explained how researchers leverage these predictions to pinpoint molecules exhibiting the precise combination of desired attributes. “We use predictions as a way to prioritize synthesis of molecules with multi-parameter optimization,” Sheth stated. “This ensures precious laboratory experiments are aligned with progressing molecules that have the highest probability of success.” By precisely narrowing the scope of compounds subjected to laboratory analysis, researchers can concentrate their experimental efforts on molecules that demonstrably align with a program’s predicted requirements, optimizing resource allocation and accelerating progress.
Furthermore, BMS has harnessed AI to significantly expand its library of CELMoD compounds. These are specifically engineered to selectively degrade cancer-causing proteins and are currently under investigation for their efficacy in treating blood cancers and other diseases. The company’s modeling work has enabled researchers to explore a wider array of protein targets and potential compounds before committing to experimental validation of specific candidates. Additionally, BMS is employing AI tools to dramatically reduce the time required to produce medicines for clinical trials. Plenge indicated that this process has already seen a reduction of 20% to 30% and anticipates further improvements of up to 50% in the coming years. He cited an experimental sickle cell disease treatment in early clinical development as a prime example of AI-supported research, suggesting that its discovery might not have been feasible without the company’s advanced AI tools. It is important to note that these time-saving figures pertain to the period of identifying and producing candidates for clinical testing, not their subsequent performance in trials.
The Vera Rubin system will also grant researchers access to NVIDIA’s BioNeMo Agent Toolkit, a suite of specialized tools designed for biological and drug-discovery applications. BioNeMo offers capabilities for protein-structure prediction, molecular generation, molecular docking, sequence analysis, and genomics. Critically, it can seamlessly integrate multiple computational tools within a single research workflow, fostering greater efficiency and collaboration. BMS executives reiterated that human researchers will continue to play a vital role in reviewing model outputs and making the ultimate decisions on which compounds or programs advance.
BMS is simultaneously introducing new tools designed to lower the barrier to entry for initiating complex computing tasks. The company anticipates that researchers will soon be able to submit certain prediction requests using intuitive natural-language instructions, further democratizing access to advanced computational resources. The entire environment will be meticulously managed through NVIDIA Mission Control, which provides comprehensive capabilities for cluster provisioning, infrastructure monitoring, and intelligent workload management, according to BMS.
This unified infrastructure is poised to enhance collaboration by enabling data and model outputs generated at one research site to be seamlessly utilized by teams at other locations. For example, datasets originating from a program in Lawrenceville, New Jersey, can be integrated into models used by researchers in San Diego. Sheth underscored the strategic intent behind this shared environment: to effectively capture and leverage knowledge from experiments and research programs across the entire organization. “The compute infrastructure is what connects all of our scientists together and ensures that our learnings are institutionalized,” Sheth stated.
The two SuperPODs will operate within a common data environment, facilitating access to shared datasets and model outputs for teams across different sites. This environment will encompass a rich tapestry of information derived from experiments, clinical readouts, and research partnerships. BMS plans to strategically allocate this enhanced computing capacity across various critical areas, including small- and large-molecule design, clinical research, and the development of digital-twin applications. While specific details regarding the planned digital-twin initiatives and the exact capacity allocation for each area were not provided, the scope of this investment signals a broad embrace of AI across the R&D spectrum.
Meyers emphasized the significant efficiency gains offered by the Vera Rubin system, noting its superior computing capacity relative to its electricity consumption. Both BMS and NVIDIA project that this eight-system cluster will deliver up to 10 times the performance per megawatt compared to the infrastructure it is replacing. “When you host these things, you have to pay an electric bill,” Meyers observed. “Think of it as 10 times more compute capacity per watt spent… Electricity is not getting cheaper.” This focus on energy efficiency underscores the economic and environmental considerations driving advanced computing adoption. BMS has not yet disclosed a specific deployment date or the location where the new system will be hosted.
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/23905.html
