Reproduce “ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Angxiao Yue, Zichong Wang, Hongteng Xu
Renmin University of China researchers developed ReQFlow, a rectified quaternion flow model for efficient and high-quality protein backbone generation. The model achieves 97.2% designability with an scRMSD of 1.071 and provides up to a 63x speedup in inference time while robustly generating long-chain proteins.
Key insights
- Using unit quaternions with exponential format SLERP provides superior numerical stability across the full range of rotation angles and avoids renormalization needs, leading to more robust geometric transformations.
- Flow rectification effectively re-wires generative paths to be more direct and efficient, preserving marginal distributions and reducing transport costs, thereby accelerating inference significantly.
- An exponential step size scheduler for rotation updates during inference is crucial for maintaining high designability and stability even with a drastically reduced number of sampling steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.