pairmixer

Triangle Multiplication is All You Need for Biomolecular Structure Representations

By Jeffrey Ouyang-Zhang, Pranav Murugan, Daniel J. Diaz, Gianluca Scarpellini, Richard Strong Bowen, Nate Gruver, Adam Klivans, Philipp Krähenbühl, Aleksandra Faust, and Maruan Al-Shedivat

Genesis Research & UT Austin

Please see our GitHub Repository for instructions on how to use this model.

Model Description

Pairmixer is an efficient alternative to AlphaFold3's Pairformer architecture for biomolecular structure prediction. By eliminating triangle attention while preserving geometric reasoning through triangle multiplication, Pairmixer achieves:

• 4× faster inference compared to Pairformer
• 34% lower training cost
• Competitive accuracy on standard benchmarks

Citation

If you use Pairmixer in your research, please cite:

@inproceedings{pairmixer2025,
  title={Triangle Multiplication is All You Need for Biomolecular Structure Representations},
  author={Ouyang-Zhang, Jeffrey and Murugan, Pranav and Diaz, Daniel J. and Scarpellini, Gianluca and Bowen, Richard Strong and Gruver, Nate and Klivans, Adam and Kr{\"a}henb{\"u}hl, Philipp and Faust, Aleksandra and Al-Shedivat, Maruan},
  year={2025},
}

Additional Resources

• Paper (ArXiv)
• GitHub Repository

License

This model is released under the MIT License.