Bristol Myers Squibb Scales AI Across R&D to Accelerate Next-Generation Medicines

BMS selected NVIDIA DGX SuperPOD to create a centralized AI computing environment for drug discovery and development, with the flexibility to connect to public cloud resources as needed. Equinix managed the infrastructure, colocation data center, and connectivity, while Mark III Systems delivered crucial AI and operational expertise. Together, the partners helped BMS progress from isolated, single-node computing to high-performance multi-node training.

The resulting environment gave BMS a ready-to-run hybrid AI platform without requiring additional system-administration headcount. Researchers could shift resources among large language model training, deep learning, medical imaging, and other workloads as their needs changed. The platform supports the full AI lifecycle—from model development and training to scalable, production-grade inference.

The deployment enabled BMS to translate AI into measurable scientific impact across the R&D pipeline. AI-enabled target identification saves scientists weeks of manual work, while AI-assisted expansion of BMS’s CELMoD compound library—molecules engineered to selectively degrade cancer-causing proteins, with applications in blood cancer treatment and beyond—has opened the door to new targets and new potential medicines across a wider range of diseases. BMS also applies AI during lead optimization through its Predict First methodology, using design predictions to inform experimental gating and prioritize which molecules should be tested in the wet lab.

BMS is using its centralized AI platform to develop and apply foundation models in oncology. One project involved building a large-scale model using CT and MRI scans from clinical trials—datasets that would have been impractical to analyze manually at scale.

The team used NVIDIA MONAI to process and transform the data, then trained the model on DGX SuperPOD to automatically segment lesions, tumors, and organs in standard CT scans. Using internal and public data with self-supervised learning and mask image modeling, the team masked 80%–90% of thousands of CT scans and trained the model to reconstruct the missing image data, learning rich representations from large volumes of internal datasets.

The results were transformative. The resulting model matched human accuracy in detecting lesions and tumors and went beyond it by quantifying muscle mass, fat and bone density. The work accelerated image analysis while improving segmentation accuracy and processing efficiency. BMS also fine-tuned its internally trained foundation model on public datasets for downstream tasks including organ segmentation and brain metastases detection, paving the way for further advances in oncology research and clinical decision-making.

“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.”

Brian Wong
Director of Research Computing, BMS

In immuno-oncology, predicting which patients will respond to treatment remains a major challenge. BMS researchers applied large language models and transformers to analyze complex clinical-trial data.

The team represented information—including genomics, lifestyle, and treatment details—as structured sequences similar to sentences in natural language. It embedded this information alongside outcomes such as survival and adverse events, enabling models to learn patterns across different kinds of clinical data.

Training combined internal clinical data, public datasets, and scientific literature. Using DGX SuperPOD, BMS models outperformed standard baseline transformers, turning complex clinical information into insights that can support oncology research and clinical decision-making.

Scientific results have driven AI adoption across the research organization. BMS has put large-scale predictions for large molecules into production and is developing proprietary foundation models. As these workloads grew, demand outgrew the capacity of the original environment.

To meet that demand, BMS is deploying a second NVIDIA DGX SuperPOD powered by eight NVIDIA DGX Vera Rubin NVL72 systems. The new systems deliver up to 10x the performance per megawatt of the infrastructure they replace.

BMS plans to combine its existing and expanded resources into a unified environment with a common data plane accessible from every BMS site. The expansion will extend advanced AI beyond specialist teams, support faster experimental cycles and allow scientists to explore larger chemical spaces while spending less time securing and managing computing resources.