Researchers have recently made stunning breakthroughs in the practice of protein design with the help of machine learning technologies. A recent work demonstrated these advances at the International Conference on Machine Learning (ICML) 2025 in Vancouver, Canada. Vsevolod Viliuga, team lead on the challenge-winning team, leads development of GAFL, an innovative technique we’re calling Geometric Algebra Flow Matching. It’s a new innovative approach that vastly increases the speed and precision of de novo protein design.
The work illustrates just how critical a role flexibility plays in functional proteins. These proteins frequently have elaborate arrangements that allow them to be malleable, creating flexibility to alter their conformation. This built-in, structural wiggle-room is essential to how proteins work. The nature of the team’s approach has great potential to change the way in which scientists develop and design these proteins.
The Team Behind the Breakthrough
Vsevolod Viliuga, the study’s lead author, worked closely with a diverse team of researchers. This fellowship team features one of its leaders, Frauke Gräter, and co-author Leif Seute from the Heidelberg Institute for Theoretical Studies (HITS). Jan Stühmer, who is a group leader at HITS as well, played an important role in the studies. Their breadth of expertise and unconventional research directions have pushed this work to the leading edge of computational biology.
Viliuga underscored the broader significance of their findings. They observed that GAFL accelerates the modeling process and produces higher quality designs of protein structures. In the midst of this challenge, the team’s efforts underscore an encouraging trend. They are using machine learning to design proteins that mimic their natural counterparts.
Advancements in Protein Structure Design
The GAFL method described here is a significant step toward high-throughput design of protein structures. In addition to being three times faster than current models, it has very high designability which is critical for producing functional proteins. This combination of speed and accuracy is poised to have an outsized effect across the biotechnology and biomedical fields.
Underlying GAFL’s method is the field of flexibility-conditioned protein structure design. Through the use of flow matching approaches, the team of researchers have robustly engineered protein structures to pursue dynamic behavior that is vital to their functions. This molecular flexibility enables diverse and improved interactions with other biomolecules, creating vast opportunities for development as drugs and therapeutic interventions.
Read more about their collaboration and IGRA research, and view the full the research team’s poster presentation online at icml.cc/virtual/2025/poster/46289. It shares their discoveries with an international audience of scientists and researchers.
Implications for Future Research
This study isn’t only a theoretical improvement. It opens doors to real-world applications in just about every major scientific discipline. GAFL is able to model proteins that exhibit structural flexibility. This critical capability opens the door for researchers to develop innovative biotechnological solutions that were once beyond reach.
The world of machine learning moves quickly. Viliuga and his colleagues are working on techniques that will ultimately establish new models and procedures for designing complex proteins. This would imply new and exciting discoveries in diseases understood at the molecular level. Beyond that, it might allow us to create highly tailored treatments that take advantage of what proteins do best.
