Суперкомпьютерийн хүчин чадлыг нэмэгдүүлэх нь генийн өгөгдлийн шинжилгээнд хангалтгүй байгааг эрдэмтэд тогтоов

When a million veterans shared their genetic data by enrolling in the Department of Veterans Affairs’ “Million Veteran Program,” they expected nothing back. This did not surprise Ravi Madduri, a senior scientist with the Data Science and Learning Division at the U.S. Department of Energy’s (DOE) Argonne National Laboratory. He is abundantly clear about the terms under which these veterans donated their data: “Even after giving everything to the country, they found it in their hearts to give more without expecting anything in return.” That gift, which constituted the nation’s largest biorepository and most ancestrally varied cohorts of veteran data in the world, created a debt that Madduri, who also holds a joint appointment in the Department of Human Genetics at the University of Chicago, is working to repay. “What I feel when I look at that data is that I have an obligation to deliver to the veterans on their intentions in doing this,” he says. Those intentions, as he reads it, are that others, veteran or not, can benefit from whatever the data can reveal. “We cannot dilly dally,” Madduri says. After arriving at Argonne as a graduate student in 2001, Madduri has spent the last two decades at the intersection of supercomputing and biomedicine, much of the time building the computing that turns the veterans’ data into results impacting areas from post-traumatic stress disorder to heart disease to cancer. He explains that machines have largely held up their end, but the obstacles between precious data and delivered results are not always solved with faster machines. Understanding the challenges is the path toward better science. What the Math Was Unable to Solve Any two human genomes differ at millions of specific positions. A genome-wide association study asks a seemingly simple question across an enormous volume of data. For example, does a given genetic variant travel together with a particular medical condition? On its own, the question is simple. Multiplied across every variant worth testing and every condition recorded in a patient’s chart, it becomes an enormous math problem, which is what supercomputers exist to solve. The harder question is whether the code is written to let them. Madduri’s team has run this problem on the Summit system at DOE’s Oak Ridge Leadership Computing Facility and on the Aurora system at the Argonne Leadership Computing Facility. On Summit, the team found that the job would not finish on the machine’s CPUs. As the cohort grew, the matrix the analysis needed for each group no longer fit in memory, and the job stalled. “We decided to open the box and look at the code,” Madduri said. They found that the issue was not merely a speed problem, but a coding one. The analysis had been written for ordinary processors because that is what the genetics community was running on. Underneath the algorithms, though, the code spent nearly all of its time doing one thing: multiplying very large grids of numbers. This is exactly the operation that graphics processors exist to perform, and what makes them the engine of modern artificial intelligence training. As Madduri notes, the same computational architecture accelerating AI discovery “is also foundational for population genomics.” So, the team, led on this effort by Argonne’s Alex Rodriguez, rewrote that part of the code for GPUs, and an analysis that would have taken roughly 481,000 CPU-core hours in the original version became tractable. This shift did more than speed up a slow job. It made a virtually impossible job possible. The workhorse they accelerated is a widely used genetics package called SAIGE, which made the rewritten version SAIGE-GPU, described in a separate paper in Bioinformatics. “We are not good at coming up with names. Sorry,” Madduri says sheepishly. Before SAIGE-GPU, nearly every genome-wide association study started by picking a disease then looking for the variants linked to it. Madduri is blunt about the cost of that approach. “One of the challenges I found in biomedical research is that it is very incremental,” he explains. “Every year, they get new datasets, they analyze the data, they find a new set of genetic variants, and this happens across all of the different diseases.” The reason, he adds, is limited access to large-scale computing. His team chose not to be incremental and decided to solve the computational challenge once and for all. “One of the challenges we set for ourselves was to just take all the data that we have available, all the genetic data for all the veterans that have been sequenced as part of the Million Veteran Program,” Madduri says. The team, the Million Veteran Program consortium, published the results in Science, with Madduri’s group among the authors, and published the computational work separately in Bioinformatics. The study reported on 2,068 traits across 635,969 veterans, and the results flagged 13,672 regions of the genome linked to one or more conditions. About 1,600 of these regions surfaced only because nearly a third of the participants were not of European ancestry. According to the Science paper, most genetic studies to this point have been drawn overwhelmingly from people of European descent. These 1,600 associations were previously “invisible” in the populations studied. That ancestral range also shaped how the analysis had to be run. The groups could not simply be pooled. That is, if a condition is more common in one ancestral group than another, the analysis has no way to tell a genuine disease marker from a marker of simply belonging to that group. Each group had to be run separately, and the results combined afterward. The 1,600 associations exist in no single group’s results. They appear only when the separate analyses are combined. Madduri is careful about what the whole exercise means and what it does not. “This is all preclinical,” he says. “This is the discovery phase.” The analysis produces a statistical map of where to look. Turning any one of those 13,672 regions into a treatment takes years of downstream work by researchers who pick up and follow the data, which is why Madduri says the team never stopped. “Things that took 14 days when we wrote the Science paper now take less than a couple of hours,” he says. While this means getting an answer sooner, the real win is being able to ask the same question of a bigger population “with a larger cohort size and more statistical rigor than the paper that we published a couple of years ago.” When Data Cannot Travel The veterans’ genomes have an unusual advantage over other medical data. One agency gathered the data under one agreement, which is what made bringing them to Summit possible. In medicine, Madduri says, data “cannot be shared with one another as freely as you would probably share cosmology data.” The alternative is to move the model, he says: “Federated learning allows you to train models without moving data around.” This means each institution trains on the copy it already holds, behind its own walls, and only sends back the adjustments the model learned. Those adjustments are combined into a single updated model, which goes out to everyone, and the cycle repeats. Asked how several institutions are persuaded to compute together without pooling anything, he corrects the premise. “It is not the institution,” he says. “The first layer of that puzzle is people wanting to work together, to collaborate.” Then, the part that makes it work: “It is a people federation.” Madduri was a co-developer of APPFL, the Argonne Privacy-Preserving Federated Learning framework, which provides a means to help “move” models rather than data. The machinery underneath is closer to access control than to machine learning. “You do not have to write code for each scheduler,” he says. Globus Compute presents one interface across the facilities’ schedulers, and APPFL submits its training jobs through it. Even with these technical challenges, the harder problem is permission. “All the computing that happens is authorized computing that is tied to a project allocation,” Madduri says. Federated training means that each collaborator’s job has to run under their own institutional credentials on a machine they are personally entitled to use. A site with data nobody else has makes that shared model better for everyone. “If one site has data that are unique to it and other sites do not have that, then the model becomes more robust,” Madduri adds. Many Different Worlds Madduri has been a regular attendee at the SC Conference since SC01 in Denver, which was themed “Beyond Boundaries,” and focused on moving data across them. “I was a grad student, and I had a demo in the DOE booth demonstrating my master’s thesis, which at that time was to use the Internet to move large amounts of data reliably from one point to the other,” he says. That capability was called reliable file transfer, and, in his words, it “has now transformed into what is called ‘Globus’.” The first tool he helped build moved research data around. The latest one lets the data stay in place. Since SC01, Madduri has been to the conference roughly 20 times and is specific about what he takes home. “I learn parallel processing, how people are scaling AI training, how people are dealing with parallelization, how people are dealing with resource utilization monitoring,” he notes. His group attends the Computational Approaches for Cancer Workshop every year. Lately, the interest has moved to AI agents. “The agentic approaches are the next frontier for us,” he declares. Madduri is also aware that he walks in unusual research domains. Before biomedicine, he worked with high-energy physicists, earthquake engineers, and cosmologists. “When I go to SC, I live in many different worlds,” he says. “I am part of a very small group of people that go to a [combined] human genetics conference and a supercomputing conference.” What Is Still Missing For Madduri, his debt is not settled. More than a decade after the first veterans enrolled, the Million Veteran Program data have produced a map but not yet a cure. What stands between those two things is the work of getting separate institutions to compute as one. To date, that still is being done largely by hand. “For the earlier experiments, it was us talking to NERSC, us talking to Argonne, us talking to Oak Ridge,” Madduri says. “It is a lot of human in the loop.” He is hoping the American Science Cloud being built under the DOE’s Genesis Mission will take that over: “It would be automated, run using APIs.” For now, Madduri remains focused on turning data into triumph, using a very special cohort of data to reveal the next breakthrough against cancer or spare someone a lifetime of anxiety. Until then, he marks progress one good idea, one technical advancement, and one SC Conference at a time. “My connection to Supercomputing, to SC, is long and very fruitful,” Madduri says. “I will be back in November, in Chicago, for SC26.” Join the Conversation at SC26 SC26 brings together the people building the computing infrastructure the work described herein depends on: systems, data platforms, networks, and storage that let research run at scale and across institutions. Attendees engage with technical papers, workshops, tutorials, posters, and Birds of a Feather (BoF) sessions covering AI, quantum computing, and HPC alongside practitioners from national laboratories, industry, and academia working on similar problems. Madduri will also be part of the SC26 Regional Tech Hub, a multimedia forum making its conference debut in Chicago, where he will contribute to the High-throughput Computing Infrastructures and Systems area. Join us in Chicago, November 15–20, 2026, for a week of discovery, collaboration, and community with a spotlight on regional technology leaders. Register today! Supplied by SC26