---
title: "Breakthrough in Supercomputer Workflow Accelerates Protein Structure Prediction"
url: https://newscentral.site/breakthrough-in-supercomputer-workflow-accelerates-protein-structure/
language: en
publisher: "News Central Site"
section: "Science"
published: 2026-11-17T22:10:01.000Z
updated: 2026-11-18T00:44:22.869Z
id: ee077602-2195-4fc0-a12a-e5e91a8a3049
source: "Phys.org https://phys.org/news/2026-11-optimized-supercomputer-workflow-protein.html"
attribution: "Link to https://newscentral.site/breakthrough-in-supercomputer-workflow-accelerates-protein-structure/ and name News Central Site when you quote or summarize this story."
---

# Breakthrough in Supercomputer Workflow Accelerates Protein Structure Prediction

Researchers at Lawrence Livermore National Laboratory have developed an optimized computing workflow that enables the generation of predicted protein structures at an unprecedented rate.

A breakthrough in supercomputer workflow has enabled researchers to accelerate the generation of predicted protein structures at an unprecedented rate. This achievement, led by Lawrence Livermore National Laboratory (LLNL), promises to revolutionize various fields including drug discovery and biological research.

The optimized computing workflow is part of a project called ElMerFold, which leverages advanced biomolecular models for improved accuracy. By generating millions of predicted protein structures as training data, the team aims to enhance these models' capabilities.

Researchers from LLNL collaborated with experts at AMD and Columbia University to push the limits of high-performance computing on El Capitan, a machine capable of exascale operations. The result was astonishing: over 10,700 predicted protein structures per second were generated across thousands of accelerated processing units.

The team's achievement has earned them a spot among finalists for the prestigious ACM Gordon Bell Prize, with the winner to be announced next month at the SC26 conference in Chicago.

The project's scope extended beyond the realm of protein folding, as it aimed to develop new capabilities for large-scale superintelligence research and infrastructure.

According to Brian Van Essen, this collaboration offered a unique chance to tackle pressing challenges in biology and machine learning workflows for high-performance computing.

Proteins play a crucial role in nearly every biological process, and understanding their three-dimensional structures is essential for determining how they function and interact with other molecules.

The experimental determination of protein structures is often expensive and time-consuming, limiting the amount of available data for developing computational prediction models. Researchers acknowledge that this shortage of real data hinders the training of state-of-the-art protein structure prediction models.

Nikoli Dryden notes that acquiring new experimental structures requires significant resources in terms of time, expense, and effort. As a result, researchers must supplement these limited data with synthetic information to aid their predictions.

Existing prediction models are generating synthetic protein structures to train new models in a process called distillation.

This approach is crucial for producing large data sets required by modern biomolecular models, which can be computationally expensive and sometimes costlier than the models themselves.

The resulting data set is being used to enhance OpenFold3's ability to predict complex molecular interactions, particularly those between antibodies and antigens they recognize.

According to Mohammed AlQuraishi, a collaborator on the ElMerFold project, this distillation set improves OpenFold3 performance by refining its understanding of protein structure biophysics.

Understanding how antibodies bind to biological targets is essential for developing therapeutics against infectious diseases, and predicting drug-protein interactions could accelerate early-stage drug discovery.

To achieve the desired performance level, the research team had to thoroughly optimize both the prediction software and the underlying computing infrastructure. This involved leveraging the capabilities of El Capitan's processors by scaling the workflow across nearly its entire system.

The researchers exploited the unique features of AMD's MI300A APUs, which integrate CPUs and GPUs with shared high-bandwidth memory, thereby reducing unnecessary data transfers and enhancing computational efficiency. By doing so, they were able to streamline the computing process and minimize delays.

Specialized components were developed to fine-tune El Capitan's processors, including optimized computational kernels, advanced memory management strategies, and innovative scheduling techniques. This allowed the system to operate at peak performance levels without compromising its processing capacity.

To further enhance efficiency, the team implemented persistent inference servers, eliminating the need for repeated model loading, and utilized El Capitan's Rabbit near-node storage system from HPE for rapid data access. Throughout this optimization process, the researchers collaborated closely with Columbia University experts in protein folding to ensure that performance gains did not come at the expense of prediction accuracy.

The team also made significant contributions to compiler frameworks such as DaCe, enabling the generation of highly optimized computations tailored to the MI300A architecture. This expertise could potentially benefit other applications and programs within the lab, as expressed by LLNL computer scientist Tal Ben-Nun: "We went as far as writing machine-level instructions to maximize the performance of the system."

The optimization of ElMerFold has led to significant advancements in protein-structure prediction, thanks to its ability to harness the power of El Capitan's AMD Instinct MI300A architecture and AMD ROCm software.

This achievement is particularly noteworthy given the importance of exascale systems in accelerating the creation of training data for scientific models. As a result, researchers can now generate larger and more detailed datasets, opening up new avenues of inquiry across various disciplines such as materials science, chemistry, and physics.

The implications of ElMerFold extend beyond molecular biology to LLNL's national security mission, where limited training data poses a significant challenge. This scarcity of data hinders the development of computational models for stockpile modernization and other applications in the NNSA mission space.

To address this challenge, researchers are exploring the use of scientifically grounded synthetic training data, which ElMerFold has shown to be an effective approach. By doing so, scientists can better interpret and understand rare or anomalous events, a crucial aspect of their work in national security missions.

The ElMerFold data set has been made accessible to researchers worldwide, enabling them to create new scientific models using the predictions. This move is expected to foster further development of protein-structure prediction methods.

To enhance its capabilities, the team aims to generate a wider range of training data sets, including those for predicting complex proteins composed of multiple interacting molecules. The inclusion of these diverse datasets will provide researchers with more robust tools for their work.

The researchers plan to integrate the latest OpenFold3 model into the ElMerFold workflow, ultimately refining the core scalable inference framework to make it adaptable to various scientific workloads.

---
Source: [Phys.org](https://phys.org/news/2026-11-optimized-supercomputer-workflow-protein.html)  
Published by News Central Site: https://newscentral.site/breakthrough-in-supercomputer-workflow-accelerates-protein-structure/
