ReQFlow

ReQFlow is a generative model for de novo protein backbone generation developed by Angxiao Yue, Zichong Wang, and Hongteng Xu at Renmin University of China, and presented at ICML 2025. It generates a local translation and a 3D rotation for every residue in a chain, starting from random noise, producing backbone geometry that a sequence designer can then populate.

Backbone generation is the step that largely fixes the fold and therefore the designability of a de novo protein, and diffusion and flow-matching models have become the standard tools for it. Their two persistent weaknesses are quality and cost: generated backbones often fail the designability test, especially for long chains, and sampling requires hundreds of network evaluations. ReQFlow addresses both by changing how rotations are represented and how the flow is trained. Each residue rotation is encoded as a unit quaternion and its flow on SO(3) is constructed by spherical linear interpolation in an exponential format, which keeps the rotation update numerically stable — matrix-based interpolation introduces measurable numerical error in the same setting.

The second half of the method is rectification. A base quaternion flow model (QFlow) is trained first; it then generates its own noise-backbone pairs, and the model is retrained on those pairs so that sampling trajectories become closer to straight lines. Straighter trajectories can be integrated in far fewer steps, which is where the speed comes from. The result sits alongside RFdiffusion, Genie2, and FoldFlow2 as an unconditional backbone generator, and is distinguished by holding its quality as the step budget shrinks.

#Key Features

• Quaternion rotation flow: Per-residue rotations are unit quaternions interpolated by SLERP rather than rotation matrices, giving numerically stable updates and rotation steps roughly 20% faster than frame-based alternatives.
• Flow rectification: Retraining on self-generated noise-backbone pairs straightens the generative path, which is what lets sampling steps drop by an order of magnitude without a quality collapse.
• Stable at low step counts: Designability holds from 500 steps down to 20, where several competing generators degrade sharply once the budget is halved.
• Structural data filtering: Rectification training keeps only samples passing scRMSD and TM-score thresholds and discards backbones with excessive loop content or an outlier radius of gyration.
• Length range: Models trained on PDB generate chains from 60 to 512 residues; a separate SCOPe-trained model covers the 60-128 residue range.

#Technical Details

The denoising network follows the SE(3) frame formulation used by FrameFlow, with the rotation component replaced by the quaternion flow. Two training corpora are used: 23,366 PDB backbones of length 60-512, and 3,673 SCOPe backbones of length 60-128. To build the rectified model, the trained QFlow generates 7,653 backbones on the PDB setting and 3,167 on SCOPe, and ReQFlow is trained on those noise-backbone pairs after filtering.

Evaluation follows the standard designability protocol: sequences are designed for each generated backbone with ProteinMPNN, refolded with ESMFold, and a backbone counts as designable when the self-consistency RMSD is at or below 2 A. Generating 50 backbones each at lengths 100, 150, 200, 250, and 300, ReQFlow reaches a designable fraction of 0.972 against 0.908 for Genie2 and 0.904 for RFdiffusion, with the lowest scRMSD at 1.071 +/- 0.482 A and a notably tighter spread than the baselines. Diversity and novelty, measured by pairwise TM-score and maximum Foldseek TM-score to the PDB, remain competitive. At 50 sampling steps the designable fraction is still 0.912, taking 1.78 s where RFdiffusion and Genie2 need 37x and 63x longer for a 300-residue backbone.

#Applications

ReQFlow serves the first stage of a de novo protein design pipeline: generate candidate backbones, design sequences onto them, and validate the fold computationally before ordering genes. Its speed profile suits campaigns that need large candidate pools — enzyme scaffold libraries, structural building blocks for nanomaterials, or the initial diversity sweep before a conditional design step. Because the quality advantage grows for longer chains, it is particularly relevant to groups generating backbones beyond the 200-residue range, where step-efficient samplers usually lose designability first.

#Impact

The model's practical contribution is showing that rectified flow, already established for image generation, transfers cleanly to SE(3) protein geometry, and that the choice of rotation parameterization is not cosmetic — quaternion SLERP yields both a numerical stability and a wall-clock benefit over matrix interpolation. The release includes code on GitHub and trained checkpoints on Hugging Face, though the repository ships no license file, which leaves reuse terms undefined for anyone building on it commercially. As an unconditional generator, ReQFlow produces backbones only: sequence design and experimental validation remain separate steps, and the reported designability is a computational self-consistency measure rather than wet-lab evidence that the designs fold and express.