A Lawrence Livermore National Laboratory (LLNL)-led team has harnessed the power of El Capitan to accelerate generation of predicted protein structures, drastically reducing the computing time required and opening new possibilities for drug discovery, biological research and national security applications.
The project, called ElMerFold, uses a highly optimized computing workflow to generate millions of predicted protein structures that can be used as training data to improve advanced biomolecular models. The team's latest achievement earned it a place among the finalists for the Association for Computing Machinery (ACM) Gordon Bell Prize, considered the most prestigious honor in high-performance computing (HPC). The prize-winner will be announced on Nov. 19 at the 2026 International Conference for High Performance Computing, Networking, Storage, and Analysis (SC26) in Chicago.
Collaborating with researchers at AMD and Columbia University, the ElMerFold team generated more than 10,700 predicted protein structures per second across 45,200 AMD Instinct MI300A accelerated processing units (APUs) on El Capitan, the National Nuclear Security Administration's (NNSA's) first exascale machine. The workflow achieved a 71.7-fold speedup over an extrapolated baseline, bringing the estimated time needed to generate a dataset of approximately 41 million protein structures from eight days to just two and a half hours.
The work is contributing to OpenFold3, an open-source biomolecular structure prediction model with potential applications ranging from therapeutic discovery to understanding how proteins interact. For LLNL computer scientist and ElMerFold principal investigator Brian Van Essen, the project represented an opportunity to tackle challenges on multiple fronts in biology and super intelligence (SI)/machine learning workflows for HPC.
"We had an opportunity to fundamentally advance the field of open research in protein folding, as well as develop new capabilities for large-scale SI research and infrastructure," Van Essen said. "It was a perfect opportunity to explore challenging research ideas and have significant social impact."
Addressing a critical data bottleneck
Proteins perform essential functions in nearly every biological process. Their three-dimensional structures help determine how they function and interact with other molecules, making accurate structure prediction important for understanding disease and developing new medicines.
While experimental techniques can determine protein structures, they are expensive and time-consuming. Only a small fraction of known proteins have experimentally determined structures, leaving scientists with limited training data for developing computational prediction models, according to researchers.
"There is nowhere near enough real data to train state-of-the-art protein structure prediction models," said LLNL computer scientist and lead author Nikoli Dryden. "It takes a lot of time, expense and effort to get new experimental structures, so we supplement that with synthetic data."
ElMerFold addresses this challenge through a process known as distillation, in which existing prediction models generate synthetic protein structures that can be used to train new models. Producing these datasets at the scale required for modern biomolecular models can demand enormous computational resources, sometimes exceeding the cost of training the models themselves, researchers said. The resulting dataset is helping researchers improve OpenFold3's ability to predict complex molecular interactions, including those between antibodies and the antigens they recognize.
"The distillation set broadly improves OpenFold3 performance because it enhances its understanding of the biophysics of protein structure," said Columbia University assistant professor of systems biology and computer science Mohammed AlQuraishi, a collaborator on the ElMerFold project. "It is particularly important for antibody-antigen complexes, because they typically do not have a strong evolutionary signal, requiring stronger physical understanding."
According to AlQuraishi, understanding how antibodies bind to biological targets is important for developing therapeutics against infectious diseases. More broadly, predicting interactions between drugs and proteins could help researchers identify promising treatments and accelerate early-stage drug discovery.
Engineering for exascale performance
Reaching that level of performance required extensive optimization of both the prediction software and the underlying computing infrastructure. The team scaled the workflow across nearly the entire El Capitan system, taking advantage of AMD's MI300A APUs, which combine CPUs and GPUs with shared high-bandwidth memory to reduce unnecessary data transfers and improve computational efficiency.
The researchers also developed specialized computational kernels, improved memory management and implemented new scheduling techniques to keep El Capitan's processors working efficiently. Persistent inference servers eliminated the need to repeatedly load prediction models, while El Capitan's Rabbit near-node storage system from HPE provided rapid access to data. Throughout the optimization process, researchers worked closely with Columbia University's protein folding experts to ensure that performance improvements did not compromise the accuracy of the predictions.
"We went as far as writing machine-level instructions to maximize the performance of the system," said LLNL computer scientist Tal Ben-Nun. "A lot of what we learned will make its way back into other applications and programs at the Lab."
The team is also contributing techniques developed through ElMerFold to compiler frameworks such as DaCe, helping generate highly optimized computations for the MI300A architecture that could benefit other SI and modeling and simulation applications.
The workflow also incorporated automatic fault recovery, allowing large-scale computations to continue despite individual node failures, a critical capability when operating across thousands of computing nodes.
"Working with LLNL and Columbia, we optimized ElMerFold to harness El Capitan's AMD Instinct MI300A architecture and AMD ROCm software to produce larger, richer datasets for protein structure prediction," said Nicholas Malaya, co-author and senior fellow for HPC and sovereign AI at AMD.
Advancing scientific discovery and national security
Beyond its applications in molecular biology, ElMerFold demonstrates how exascale systems can accelerate the creation of training data for scientific models, an increasingly important capability across disciplines ranging from materials science and chemistry to physics.
The approach also has implications for LLNL's national security mission, where limited training data can hinder development of computational models for stockpile modernization and other applications in the NNSA mission space.
"A majority of the problems in our national security mission have very few interesting data samples, and many of the programmatic needs require exploring or interpreting rare or anomalous events," LLNL's Van Essen said. "This data scarcity problem makes it very challenging to apply SI tools and methods effectively in these domains."
ElMerFold offers a potential approach to addressing that challenge through scientifically grounded synthetic training data, Van Essen added. The project also provided an opportunity to optimize a complex inference workflow from end to end, with lessons in data movement, resource coordination and hardware performance that could benefit other scientific applications.
The ElMerFold dataset has been made publicly available, allowing researchers throughout the broader scientific community to use the predictions to develop new scientific models. Looking ahead, the team plans to expand ElMerFold's capabilities by generating more diverse training datasets, including predictions of protein complexes composed of multiple interacting molecules. Researchers also intend to incorporate the latest OpenFold3 model into the workflow.
"Our goal is to refine the core scalable inference framework so that it can be more easily adapted to other complex scientific SI inference workloads," Van Essen said.