Predicting the Future of Protein Science
Boston University Professors Dima Kozakov and Sandor Vajda Win Top Honors at Global CASP-16 Competition in Category of Multiprotein Complexes
A BU-led interdisciplinary research team, bringing together expertise in biomedical engineering, applied mathematics, and theoretical physics, has won top honors at the internationally renowned CASP-16 (Critical Assessment of Structure Prediction) competition. The group earned top ranking in predicting multiprotein complexes, a critical benchmark of accuracy and innovation in computational modeling. Multiprotein complexes involve two or more interacting proteins that perform specific cellular functions. Accurately modeling these complexes is essential to understand cellular processes, improve understanding of disease progression, and design targeted therapeutic strategies.
The team was led by Dima Kozakov, Research Professor of Biomedical Engineering at BU, and Sandor Vajda, Professor of Biomedical Engineering, Director of the Biomolecular Engineering Research Center, and Hariri Institute Faculty Affiliate at BU, with support from Dr. Taras Patsahan, Deputy Director of Research and Head of the Soft Matter Theory Department at the Institute for Condensed Matter Physics of the National Academy of Sciences of Ukraine.
Held every two years, CASP is the field’s premier global competition, tasking top international teams with predicting the three-dimensional structures of proteins and protein complexes whose experimental structures have not yet been released. The competition provides a rigorous, blinded test of emerging computational methods.
According to the official evaluator’s report, the BU team’s work (identifier G274) made significant advances in the accurate prediction of protein multimers—one of the most complex challenges in structural biology. Evaluators noted that their models outperformed all other entries by a wide margin, achieving the highest accuracy in the category.
The high prediction accuracy achieved by the team reflects a novel protocol jointly developed by the Kozakov laboratory and the Sandor laboratory over the past three years. Central to this approach is integrating the physics of protein interactions and the geometry of the conformational space into the machine learning models.
Traditional machine learning approaches often rely on biased sampling shaped by training data, leading to inefficiency and reduced accuracy when predicting novel interactions outside the training set. In contrast, the Kozakov/Vajda method employs systematic sampling of high-interest regions, significantly improving efficiency and accuracy in exploring the vast conformational landscape.
“While the approach demonstrated strong performance in specific CASP tasks, its core innovation—combining machine learning with physics-based sampling—is broadly applicable,” says Kozakov. “It is particularly advantageous in scenarios with limited training data, where physics-informed guidance can compensate for data sparsity and improve predictive robustness across a range of biomolecular modeling challenges.”
This success not only highlights the team’s technical innovation in algorithm design and modeling but also has broad implications for biomedical research. “Accurate prediction of protein complexes accelerates our understanding of diseases, supports drug development, and enhances synthetic biology efforts,” says Vajda. “By improving the tools available to interpret the molecular machinery of life, our collaboration offers potential to shape the future of precision medicine and therapeutic design.”
Their achievement also reflects the critical role academic research plays alongside major industry leaders like DeepMind, whose AlphaFold programs have revolutionized single-protein structure prediction.
Learn more about this work on the Kozakov lab and Vajda lab websites.
About this event: Held every two years, CASP is the field’s premier global competition, tasking top international teams with predicting the three-dimensional structures of proteins and protein complexes whose experimental structures have not yet been released. The competition provides a rigorous, blinded test of emerging computational methods.
The most recent event was conducted over the summer months of 2024. Nearly 100 research groups from around the world submitted more than 80,000 models on 100+ modeling entities, yielding 300 targets in five prediction categories. Results were presented at the CASP-16 conference in Punta Cana, Dominican Republic, December 1-4, 2024, and evaluated by independent assessors.
In 2018, CASP joined with a similar contest called CAPRI (Critical Assessment of Predicted Interactions), which focuses on predicting how molecules interact. Both events, challenge researchers to submit models for a set of macromolecules and macromolecular complexes (proteins, RNA, ligands) for which the experimental structures are not yet public. Independent assessors compare the models to experimental structures. The Google-owned company DeepMind achieved excellent results at CASP-14 in 2020 with their machine learning based protein structure prediction program AlphaFold-2. The leaders of the AlphaFoild-2 project, Dennis Hassabis and John Jumper, shared the 2024 Nobel prize in chemistry. More recently, DeepMind and its subsidiary, Isomorphic Labs, released the program AlphaFold-3 that further increases accuracy.