Computational Science Accelerates Research Innovation at Bristol Myers Squibb

BMS knew it needed to change how it provided compute for research, as neither traditional cloud nor traditional on-prem approaches were fully serving scientific needs. “To guarantee GPU availability and maximize computational resources, we decided to establish a centralized and modernized infrastructure for drug discovery. To pull this off, we needed a proven platform that could scale with us as we progressed along this journey. That is why we chose the NVIDIA DGX SuperPOD architecture, which included the capability to expand to cloud as needed, a colocation strategy for hosting, and an acceleration partner to clear all the early hurdles,” said Mayo.

NVIDIA DGX™ partner Equinix expertly managed the DGX SuperPOD infrastructure, colocation data center, and interconnectivity, while Mark III, an NVIDIA solutions partner, delivered crucial AI and operational expertise.

“Mark III’s expertise was the catalyst behind our rapid evolution from isolated, single-node computing to seamless, high-performance multi-node training and ensured we not only optimized our systems but continuously advanced through routine audits and precise workload sizing,” said Brian Wong, Director of Research Computing at BMS. “The Equinix Private AI with DGX solution provided a ready-to-run AI platform and seamless hybrid AI development environment, without the operational headaches commonly associated with on-premises equipment,” added Wong.

As a result, scientists gained computational resources to scale up and out without additional headcount or expenses for system administration. Wong added, “Leveraging DGX SuperPOD with Equinix’s seamless integrations with public cloud providers ensured cost-effective data movement while achieving 55% overall cost savings compared to the prior model. The cluster has demonstrated peak utilization rates of up to 90% during certain periods. Our scientists can now easily adjust resources to meet workload demands, increasing nodes for large language model (LLM) training when needed and reallocating them to deep learning tasks as required.”

After becoming operational in March 2024, the DGX SuperPOD provided a centralized AI platform for BMS research teams, offering a wealth of information and services. One of the projects was the development of a large-scale foundational model for oncology utilizing millions of CT scans from clinical trials. Manual analysis of these scans would have been impractical and the consolidated analysis virtually impossible.

To address this, the team used NVIDIA MONAI to process and transform the data, training it on the DGX SuperPOD to automatically segment lesions, tumors, and organs in standard CT scans. They enhanced the model’s robustness by training it on both internal and public data, using self-supervised learning and mask image modeling. This involved masking 80%–90% of thousands of CT scans and having the model reconstruct the missing image data.

The results were transformative. The model not only matched human accuracy in lesion detection, but surpassed it by quantifying complex features like texture and heterogeneity, aspects previously reliant on subjective interpretation, in near real time. The impact on the pace of research was profound, with improvements in both segmentation accuracy and processing speed. In addition, the model can now be fine-tuned for downstream tasks like organ segmentation and brain metastases detection, paving the way for further advances in drug discovery.

In immunotherapy oncology, predicting who will respond to a treatment is a major challenge. BMS researchers tackled this using large language models and transformers to analyze complex clinical trial data. The innovation was to treat diverse data—including genomics, lifestyle, and treatment details—as grammatical sentences, similar to natural language processing. By embedding this data alongside outcomes like survival and adverse events, BMS built a model that predicts patient outcomes with unprecedented accuracy.

Success ultimately hinged on training models with extensive data, combining internal clinical data, public datasets, and scientific literature. Leveraging the DGX SuperPOD’s computational power, BMS models outperformed standard baseline transformers, transforming their data into actionable insights that could revolutionize oncology treatment strategies.

With dedicated compute resources, extensive data, deep technical expertise, and urgent scientific questions,  researchers can now easily launch high-potential experiments. “By combining curiosity and courage with technology skills and capacity, we have assisted our research community in making innovative discoveries, all in the service of our patients,” adds Mayo. Looking forward, the successful implementation of the DGX SuperPOD has positioned BMS to leverage AI for continued innovation and advancements in drug discovery across all therapeutic areas and scientific functions.