Why Advanced Computing Matters for Genomics, Even When Supercomputers Get Faster

When one million veterans enrolled in the U.S. Department of Veterans Affairs’ Million Veteran Program and shared their genetic data, they did not expect anything in return. Ravi Madduri, a senior scientist in the Data Science and Learning Division at the U.S. Department of Energy’s Argonne National Laboratory, views that contribution as an obligation to deliver useful knowledge to the participants and to others who may benefit from the research.

Madduri, who also holds a joint appointment in the Department of Human Genetics at the University of Chicago, has spent more than two decades working where supercomputing and biomedicine meet. Much of that work has involved developing the computing methods needed to turn the veterans’ data into findings related to post-traumatic stress disorder, heart disease, cancer, and other conditions.

Faster hardware has helped, but it has not resolved every obstacle between data and scientific results. In many cases, the challenge lies in software design, data access, and coordination between institutions.

What the Math Could Not Solve

Any two human genomes differ at millions of specific positions. A genome-wide association study examines whether a genetic variant is associated with a particular medical condition. The question is straightforward in isolation, but testing millions of variants against many conditions recorded in patient charts creates a massive computational problem.

Madduri’s team ran this type of analysis on the Summit system at the U.S. Department of Energy’s Oak Ridge Leadership Computing Facility and on the Aurora system at the Argonne Leadership Computing Facility. On Summit, the team found that the analysis would not finish using the system’s CPUs. As the cohort grew, the matrices required for each group no longer fit in memory, causing the job to stall.

The team examined the code and found that the problem was not simply insufficient processing speed. The software had been written for conventional processors because that was the hardware most commonly used by the genetics community. However, much of the computation involved multiplying very large grids of numbers, an operation that graphics processing units are designed to handle efficiently and that also supports modern artificial intelligence training.

Led in this effort by Argonne’s Alex Rodriguez, the team rewrote that part of the code for GPUs. An analysis that would have required approximately 481,000 CPU-core hours in its original form became tractable. The change did more than accelerate a slow task: it made an otherwise impractical analysis possible.

The team optimized SAIGE, a widely used genetics package. The GPU-enabled version was named SAIGE-GPU and was described in a separate paper in Bioinformatics.

Before SAIGE-GPU, many genome-wide association studies began by selecting a disease and then searching for genetic variants associated with it. Madduri describes this as an incremental process. Researchers received new datasets, analyzed them, identified another set of variants, and repeated the process for different diseases. Limited access to large-scale computing was one reason for this approach.

Madduri’s team instead sought to analyze all available genetic data from veterans sequenced through the Million Veteran Program. The Million Veteran Program consortium published the results in Science, while the computational work was published separately in Bioinformatics.

The study examined 2,068 traits across 635,969 veterans and identified 13,672 regions of the genome associated with one or more conditions. Approximately 1,600 of those regions appeared only because nearly one-third of the participants were not of European ancestry. Most earlier genetic studies had drawn predominantly from people of European descent, leaving these associations effectively invisible in the populations previously studied.

The diversity of the cohort also affected the analysis method. Researchers could not simply combine all participants into one group. If a condition is more common in one ancestral group than another, pooling the groups can make it difficult to distinguish a true disease marker from a marker associated with group membership. Each ancestral group therefore had to be analyzed separately, with the results combined afterward. The 1,600 associations did not appear in any single group’s results. They emerged only from the combined analysis.

Madduri emphasizes that these findings are preclinical. They represent a discovery phase and provide a statistical map of where researchers should look. Turning one of the 13,672 regions into a treatment requires years of additional research.

The computational improvements have also enabled larger and more rigorous follow-up analyses. Tasks that took 14 days when the Science paper was written now take less than a couple of hours, allowing researchers to repeat the analysis with larger cohorts and greater statistical power.

When Data Cannot Travel

The veterans’ genomic data have an advantage that many medical datasets do not: one agency collected the data under a common agreement, making it possible to bring them to Summit. Medical data are generally more difficult to share between institutions because of privacy, security, and authorization requirements.

One alternative is federated learning. Instead of moving data to a central location, each institution trains a model using the data it already holds. The institution sends back model updates rather than the underlying records. Those updates are combined into a new model, which is distributed back to the participating sites, and the process is repeated.

Madduri describes collaboration as the first requirement for this approach. Institutions must agree to work together, while the technical infrastructure must allow them to train models using their own authorized computing resources.

Madduri was a co-developer of APPFL, the Argonne Privacy-Preserving Federated Learning framework. APPFL helps move models rather than data. The supporting infrastructure also addresses access to different computing environments. Globus Compute provides a common interface across facility schedulers, allowing APPFL to submit training jobs without requiring separate code for every scheduler.

Authorization remains a central issue. Computing must be tied to a project allocation, and each collaborator’s job must run under that institution’s credentials on a machine the institution is authorized to use. A site with unique data can improve the resulting model without transferring those data to the other participants.

Many Different Research Worlds

Madduri has attended the SC Conference regularly since SC01 in Denver, whose theme was “Beyond Boundaries.” At that conference, he demonstrated his master’s thesis, which involved using the Internet to transfer large amounts of data reliably between locations. That reliable file transfer work later evolved into Globus.

Since SC01, Madduri has attended the conference approximately 20 times. He follows developments in parallel processing, AI training at scale, parallelization, and resource-utilization monitoring. His group also attends the Computational Approaches for Cancer Workshop each year. More recently, the group has become interested in AI agents.

Madduri’s research experience spans several fields. Before focusing on biomedicine, he worked with high-energy physicists, earthquake engineers, and cosmologists. He describes himself as part of a small group that participates in both human genetics and supercomputing conferences.

What Is Still Missing

The Million Veteran Program has produced a detailed map of genetic associations, but it has not yet produced a cure. One remaining challenge is enabling separate institutions to compute together as a coordinated system. Much of that coordination is still performed manually.

For earlier experiments, Madduri’s team had to communicate separately with the National Energy Research Scientific Computing Center, Argonne, and Oak Ridge. This process involves substantial human intervention. He hopes the American Science Cloud, being developed under the DOE’s Genesis Mission, will eventually automate that coordination through application programming interfaces.

For now, Madduri continues working to turn the Million Veteran Program’s data into findings that can support research on cancer, post-traumatic stress disorder, heart disease, and other conditions. Progress depends on both improved algorithms and the infrastructure required to run them across institutions.

Madduri plans to return to Chicago for SC26 in November 2026.