acceptodds
Under review as a conference paper at ICLR 2027

ReSTFlow: Stable Few-Step Rectified Flow for RNA Backbone Generation

Abstract

RNA molecules play central roles in biological regulation and therapeutic development, with many of their functions closely tied to three-dimensional structure. Recent diffusion and flow-based generative models provide effective frameworks for unconditional RNA backbone generation. However, existing approaches often require many sampling steps to maintain steric validity, limiting their efficiency in large-scale candidate generation. Here, we introduce ReSTFlow, a rectified flow framework that integrates steric regularization and trajectory correction for few-step, high-quality RNA backbone generation. The first stage trains STFlow, a sterically regularized flow model that represents each nucleotide as a local rigid frame and learns continuous transport from a noise distribution to RNA backbone conformations under clash-aware supervision. In the second stage, the learned flow is rectified using trajectories constructed from STFlow-generated samples, resulting in straighter transport paths that can be accurately integrated with substantially fewer sampling steps. At inference, trajectory correction stabilizes coarse translational updates during sampling. With 20 sampling steps, ReSTFlow achieves 75.0% steric validity, and 69.8% of its sterically valid backbones are designable, demonstrating strong performance in the few-step setting. These results establish ReSTFlow as an effective framework for high-quality RNA backbone generation under limited sampling budgets.

Primary area:
generative models
Keywords:
RNA backbone generation, rectified flow, flow matching, generative modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kx9COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce ReSTFlow, a rectified flow method for RNA 3D backbone generation. The authors propose steric clash loss terms to improve rectified flow fidelity for RNA structures. Empirical results suggest that ReSTFlow achieves higher sample validity while reducing the computational costs required to generate samples.

Critical strengths
  1. ReSTFlow demonstrates that few-step generation of RNA 3D backbones is possible and beneficial in some cases.
  2. With ReSTFlow, the authors propose a steric clash loss for RNA structures that may be useful in future work.
  3. The paper is well-written and easy to follow.
Critical weaknesses
  1. The novelty of ReSTFlow is somewhat limited. For instance, rectified flows for images and even protein 3D structures [1] have already been studied, so applying them to RNA structures is not a surprising leap forward in terms of methodological development.
  2. Frame-based representations of RNA, though useful, are more complex to work with than all-atom representations of 3D point clouds, such as those adopted by AlphaFold 3 [2]. It would be nice for the authors to discuss further why frame-based representations of RNA should still be considered.
  3. Looking at Tables 1 and 2, it seems that rectified flows, as ReSTFlow has implemented them, notably degrade the average diversity of generated RNA backbones in exchange for increased designability and steric validity. It would be helpful for the authors to study this effect further so readers can better understand whether or not this is an inherent trade-off that rectified flows present for this low-data RNA task.

References:

[1] Yue, A., Wang, Z., & Xu, H. (2025). Reqflow: Rectified quaternion flow for efficient and high-quality protein backbone generation. ICML 2025.

[2] Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., ... & Jumper, J. M. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630(8016), 493-500.

Questions for authors
  • For Figure 1(b), would the authors please plot something like "Designability + steric validity + diversity + novelty" against log10 inference time, so that readers can more easily see how composite RNA sample quality is affected by the number of inference timesteps used?
Reviewer_mYRaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Given a target length, this paper generates RNA backbones from noise without sequence or template conditioning, with the goal of keeping quality at 10-20 sampling steps. It trains STFlow, an SE(3) frame-and-torsion flow model with a differentiable van der Waals clash penalty; rectifies the flow on 7,850 clash-free noise-backbone pairs generated by its own corrected sampler; and applies a history-conditioned correction to the translational velocity at inference. At 20 steps it reports 75.0% steric validity and 69.8% designability among valid backbones, against 51.0% and 60.8% for RNA-FrameFlow, an overall designable yield of 52.4% versus 31.0%, and 1.03 s versus 1.77 s per 100-nt sample.

Critical strengths
  1. Each component targets a different failure mode of few-step sampling, and the ablation separates them. Table 2 removes the clash loss (designability 69.8 to 63.8, validity 75.0 to 69.0), the rectification stage (to 59.3 and 51.2) and trajectory correction (to 68.0 and 71.2) one at a time. The mechanism can therefore be checked component by component.

  2. The few-step result is reported as joint designable yield, which is the quantity that matters for candidate generation. At 20 steps ReSTFlow reaches 52.4% (69.8% designability x 75.0% validity), against 31.0% for 20-step RNA-FrameFlow and 32.4% for RNA-FrameFlow at 100 steps, at 1.03 s versus 1.77 s per 100-nt sample. The claim therefore rests on the joint yield.

Critical weaknesses
  1. The efficiency claim is not separated from the integrator. Every reported budget varies the number of Euler updates (L = 10 to 100, Section 3.4), and no comparison at matched NFE against a higher-order or adaptive solver is reported. It is therefore not shown how much of the gain comes from the straighter flow and how much from the particular first-order update.

  2. Two supporting metrics change with validity. Diversity is clusters divided by the number of valid backbones, so the ratio falls as validity rises even when the absolute number of clusters does not: at 20 steps the counts implied by Table 1 are about 168 for ReSTFlow and 155 for RNA-FrameFlow, while the reported ratios are 0.373 and 0.507. Novelty is computed only on the valid subset, whose composition changes with method. The reported checkpoint was chosen by steric quality and diversity (C.3), and (K, alpha, beta) by a sweep scored with the same 600-sample protocol (D.2); no held-out set is defined.

Questions for authors
  1. Table 1 reports diversity as clusters divided by the number of valid backbones, and novelty only on that subset. Please give the absolute cluster counts and the novelty statistics on all 600 samples; at 20 steps the ratios 0.373 and 0.507 correspond to about 168 and 155 clusters given 75.0% and 51.0% validity.

  2. Section 3.4 uses L Euler updates and the tables vary L from 10 to 100, but no comparison at matched NFE against a higher-order or adaptive solver is reported. On the same STFlow checkpoint, does midpoint or Heun integration at 20 NFE close part of the reported gap?

  3. C.4 notes that RNA-FrameFlow keeps its original training data. At 20 steps STFlow and RNA-FrameFlow are close (59.3/51.2 vs 60.8/51.0), but at 100 steps STFlow is higher (63.9/56.8 vs 57.6/56.2); the abstract's comparison to the 100-step result uses the released weights. Please report RNA-FrameFlow retrained on the same 7,053-chain snapshot.

  4. C.3 selects checkpoints by steric quality and diversity, and D.2 selects (K, alpha, beta) with the same 600-sample protocol used in Table 1. Was any held-out length set or independent seed set used? If not, please report the main metrics on held-out lengths.

  5. How many STFlow candidates were generated to retain the 7,850 clash-free pairs, and what is the acceptance rate per length? This determines how much of stage 2 is distillation of the model's own clash-free tail.

  6. Section 4.1 states a 4 A cutoff "or better" while C.1 states "below 4 A", and steric validity checks only inter-nucleotide clashes. Please make both scopes explicit in the main text.

  7. The 100-step joint yield is not stated: 31.0% corresponds to the 20-step model (60.8% x 51.0%), and the 100-step value is 32.4% (57.6% x 56.2%). Please state it explicitly, and add one sentence on why RiboFlow, RNAFlow and Morehead et al. are not comparable baselines.

Reviewer_c2kzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ReSTFlow, a two-stage rectified flow framework for few-step RNA backbone generation. The first stage trains STFlow, a flow model on nucleotide frames that represents rotations as unit quaternions and adds a clash loss. The second stage rectifies STFlow on clash-free noise–backbone pairs generated by STFlow. At inference, a history-conditioned trajectory correction stabilizes the translational updates.

Critical strengths
  1. Relevant task. Few-step RNA backbone generation with steric validity is a practical and underexplored problem. Improving steric validity under a limited sampling budget is useful for large-scale candidate generation.

  2. Detailed evaluation and reporting. The paper reports steric validity, designability, diversity, novelty, local geometry, and length-resolved results, and documents its implementation and evaluation settings in detail.

Critical weaknesses
  1. Limited novelty. The core pipeline closely follows ReQFlow [1]. STFlow uses the same quaternion SLERP interpolation and exponential schedule for rotations as QFlow (the first stage of ReQFlow), and ReSTFlow uses the same single rectification round on filtered, model-generated noise–backbone pairs. The RNA-specific parts, the clash loss and the trajectory correction (inspired by PMI [2]), are relatively incremental. As a result, the paper's own contribution is limited.

  2. Weak baseline comparison. RNA-FrameFlow is the only external baseline, and it is evaluated with its released checkpoint trained on different data. As the authors note, the performance gap cannot be attributed solely to the proposed method.

[1]. ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation, ICML 2025

[2]. Free Lunch for Stabilizing Rectified Flow Inversion, ICLR 2026

Questions for authors

See Weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.