A U.S. Department of Veterans Affairs initiative has created the nation’s largest biorepository of genetic data from over 800,000 veterans, enabling research that could reshape understanding of disease across diverse populations. Ravi Madduri, a senior scientist at Argonne National Laboratory with joint appointments in human genetics and data science, leads efforts to analyze this dataset using supercomputing resources.

The Million Veteran Program, launched in 2011, collects health and genetic information from veterans who voluntarily participate. Madduri’s team sought to identify genetic variants linked to medical conditions by analyzing the full dataset, which includes participants from over 40 ancestral backgrounds — a level of diversity rarely seen in biomedical research.

Initial attempts to run genome-wide association studies on existing systems stalled when data size exceeded memory capacity. The original code, designed for central processing units, spent most of its processing time multiplying large numerical grids — a task well-suited to graphics processing units. Madduri’s team, working with Argonne’s Alex Rodriguez, rewrote portions of the analysis pipeline to leverage GPU acceleration, transforming a projected 481,000 CPU-core-hour computation into a tractable job.

The optimized version, named SAIGE-GPU, is built on a widely used genetics framework and described in a paper published in Bioinformatics. By shifting computation to GPU-friendly operations, the team reduced processing time from weeks to hours, enabling analysis of 2,068 traits across nearly 636,000 veteran participants. The results, published in Science, identified 13,672 genomic regions associated with various conditions, including 1,600 previously undetected links in non-European ancestry groups.

Madduri emphasized that these findings represent a discovery phase, not clinical application. Each identified region requires years of follow-up research before potential therapeutic use. The computational advance allows researchers to study larger, more diverse populations with greater statistical rigor than previously possible.

Data sharing remains constrained by privacy and institutional boundaries. Unlike some scientific fields, medical data cannot be freely exchanged. To address this, Madduri helped develop APPFL, a federated learning framework that enables collaborative model training without moving raw data. Instead, each participating institution processes data locally and shares only model updates, preserving confidentiality while improving collective analysis.

Collaboration across institutions requires both technical infrastructure and mutual trust. The upcoming SC26 conference in Chicago will bring together researchers, computational resource providers, and data stewards to discuss frameworks like federated learning and automated data governance. Madduri plans to present work on scaling these approaches through the DOE’s Genesis Mission, which aims to streamline cross-institutional research workflows.

Madduri has attended the Supercomputing Conference regularly since 2001, observing shifts from early data transfer tools to current AI-driven computing paradigms. He notes that while hardware capabilities have grown, the core challenge remains enabling diverse institutions to work together effectively.

The veteran genomic dataset continues to offer opportunities to study conditions that may have been overlooked due to historical underrepresentation. By combining computational innovation with careful data governance, researchers aim to turn this unique resource into actionable insights for public health, particularly for populations historically excluded from genetic studies.