acceptodds
Under review as a conference paper at ICLR 2027

Base-and-Sugar Dual-Frame Flow Matching for RNA Co-Design

Abstract

Current frame-based RNA generative models represent each nucleotide with a single rigid coordinate frame. However, our analysis of seven nucleotide frame constructions across static RNA chains and multi-state relation groups reveals a clear trade-off: base-centered frames better preserve base-mediated relations, whereas sugar-centered frames better support sugar-phosphate backbone reconstruction. Motivated by this complementarity, we introduce DuetRNA, a joint sequence-structure generation model for RNA co-design that performs flow matching over coupled base-and-sugar frames. Empirically, we evaluate DuetRNA under both the inverse-folded (IF) backbone-design protocol and generated-sequence (GS) co-design protocol. DuetRNA improves scTM-validity over the strongest matched baselines by 7.67 and 4.33 % under IF and GS, respectively. Controlled ablations further support the necessity of the dual-frame modeling in DuetRNA.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative AI, RNA 3D structure modeling, RNA design, flow matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eQvAOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce DuetRNA, a method based on dual-frame RNA structure modeling for unconditional generation of RNA biomolecules. A range of experiments suggests the usefulness of modeling dual rather than single RNA frames to improve generative modeling and structural robustness of deep learning methods developed for RNA. Empirically, dual frames seem to encode additional priors that help model RNA structures.

Critical strengths
  1. The authors identify a key issue in current RNA generative modeling work: a single choice of geometric nucleotide frame does not capture all of RNA's structural diversity for distribution learning.
  2. DuetRNA, the authors' new method based on this idea, features dual RNA nucleotide frame tracks to address this concern, supporting training and sampling in a straightforward manner.
  3. Empirically, DuetRNA's dual frames appear to improve unconditional generative modeling of single-chain RNA, suggesting future work may benefit from such a parametrization of RNA.
Critical weaknesses
  1. The authors do not empirically compare their method to a state-of-the-art baseline, RiboDiff [1]. This is necessary to ensure that DuetRNA is truly offering new capabilities beyond what existing methods offer for RNA generation.
  2. Dual frames do not seem to guarantee improved RNA structural modeling (in all cases). For example, if I'm understanding it correctly, Table 4 on Page 9 of the main text suggests that a sugar-GS-only frame parametrization offers considerably better structural accuracy for C1′–N9 and C1′–N1 bond lengths.
  3. Most importantly, the paper's writing style is hard to follow, as it appears to be at least partially LLM-generated. For instance, phrases such as A complementary dual-frame reconstruction analysis..., Inter-residue O3′–P P99 deviation..., and The present evidence is limited to... read like opaque LLM text. Additionally, Figure 1 in the paper is rather unexciting for an ML conference paper, since it simply reviews different ways of representing RNA nucleotide structures without any connection to ML itself. This makes the paper, to me, read more like an incremental (RNA biology-centric) improvement for a single RNA generation task instead of a compelling and expressive method for many RNA modeling contexts.

References:

[1] Li, X., Uehara, M., Su, X., Scalia, G., & Ji, S. (2026, April). A Joint Diffusion Model with Pre-Trained Priors for RNA Sequence-Structure Co-Design. In International Conference on Learning Representations (Vol. 2026, pp. 105615-105639).

Questions for authors
  1. Would the authors please rewrite the LLM-like phrases flagged in Weakness (3)?
  2. Are the authors planning on releasing open-source code for this? If so, how easy is the implementation to follow, since existing RNA generative models tend to produce very verbose and unintuitive code implementations?
Reviewer_kmG3OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces DuetRNA, an RNA sequence–structure generation model that assigns each nucleotide separate base-centered and sugar-centered rigid frames. A coupled SE(3) flow predicts these frames, with a relative-pose loss, a sequence head, and torsion-conditioned heavy-atom reconstruction. The motivation comes from an analysis of seven frame definitions on static and multi-state RNA structures. Evaluation covers unconditional generation of 40–150 nt RNAs using inverse-folded sequences (IF) and directly generated sequences (GS). The headline IF self-consistency rate improves from RNA-FrameFlow's reported 41.00% to 48.67%, while the GS difference from RiboGen is smaller and not statistically significant. Ablations, training-seed results, and direct geometry analyses examine both the benefit and limitations of the representation.

Critical strengths

The representation has a clear geometric motivation. Base-centered frames show lower canonical-pair pose drift in the multi-state analysis, while the sugar frame provides a useful coordinate system for the backbone (Section 4.2; Appendix D.4). This analysis helps explain why separating the two could be useful for RNA generation.

The controlled generative ablations are convincing for the implemented model. With data, training schedule, sampler, evaluator, and approximate parameter count held fixed, IF validity drops from 48.67% to 34.67% without the base frame and to 30.00% without the relative-pose loss (Table 3a). The independent training runs in Table 4a also exceed the published RNA-FrameFlow IF result.

Critical weaknesses

The reconstruction audit does not separate the choice of frame from the choice of decoder. Appendix D.5 states that the single-frame controls use a sugar-rooted torsion decoder without frame-specific template refitting. A base-centered frame may therefore reconstruct poorly because the decoder is mismatched to it. The generative ablation tests Sugar-GS-only, but does not include a base-centered single-frame model with a compatible decoder. I find the improvement over the tested sugar-only model credible; the broader claim that one frame cannot adequately represent both parts of a nucleotide needs this additional control.

Folding self-consistency leaves substantial uncertainty about the yield of usable designs. The largest gain is under IF, which uses an external inverse-folding model to supply eight candidate sequences. For direct GS co-design, the Boltz-1 result is 38.50% versus 34.17%, with a confidence interval crossing zero (Table 2). The generated coordinates also have 26.55 clashes per 100 heavy atoms, glycosidic-bond RMSDs of 0.417/0.484 Å, and a 38.6% phosphodiester-linkage outlier rate (Table 4b). Only 55.2% of the 600 generated structures in the continuity analysis are break-free; Appendix E.8 even shows a fragmented structure passing the scTM criterion. The paper is candid about these issues, but does not report how many scTM-valid samples also pass basic chemical and continuity checks. That intersection is needed to assess the practical design improvement.

Questions for authors

Could the authors clarify how templates and torsional coordinates are adapted to each frame in Appendix D.5? How would a base-centered single-frame model perform with a compatible decoder? This would help distinguish a limitation of the representation from a limitation of the reconstruction setup.

What fraction of IF- and GS-scTM-valid samples also passes the chain-continuity and local bond-geometry checks? With explicit acceptance criteria, does DuetRNA still outperform the matched sugar-only control on this joint measure?

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

In this paper, DuetRNA, a method of SE(3) Flow Matching based on double Frame, is proposed for the joint generation of RNA sequences and three-dimensional structures. It is found that the base center Frame is more suitable for describing the relationship between base pairing and stacking, while the ribose center Frame is more suitable for skeleton reconstruction. Therefore, the model constructs two coupled Frame for each nucleotide, and introduces relative pose constraints to predict sequence and twist angle at the same time. Experiments show that this method is improved by 7.67 percentage points compared with RNA-FrameFlow under IF protocol and 4.33 percentage points compared with RiboGen under GS protocol, but the latter is not statistically significant. Ablation experiments verify the effectiveness of double Frame and its coupling constraints.

Critical strengths

Through large-scale Frame analysis, the author reveals the complementarity of geometric representation of base and ribose, which provides a sufficient basis for double Frame design.

After removing the constraint of Base Frame or relative pose, the IF efficiency decreased by 14.00 and 18.67 percentage points respectively, indicating that both core components have made practical contributions.

Critical weaknesses

SE(3) Flow Matching, IPA and torsion angle modeling have all been well studied, and NuFold and others have also explored multi-local Frame and flexible nucleotide representation. The paper lacks a fair comparison with the stronger single Frame+Torsion parameterization method, which can not fully prove the necessity of generating two Frames independently.

Only 55.2% of the generated structures meet the standard of no chain breakage adopted in this paper, and there are still a lot of atomic collisions and glycosidic bond errors. In addition, the improvement of relative RiboGen under GS protocol is not statistically significant, so the evidence of improving the actual Co-design capability is still limited.

Questions for authors

Q1. Can it be compared with the single Frame model with matching parameters and richer representation of torsion or base-orientation to prove the advantages of double frames?

Q2. How many generated samples satisfy scTM validity, skeleton continuity and reasonable atomic geometric constraints? Is the performance advantage still valid after being limited to physically valid samples?

Q3. Why use relative posture loss instead of directly restricting the bond length and bond angle of glycosidic bonds? Have these alternatives been compared?

Q4. Does double Frame directly improve the geometric quality of base pairing and stacking in the generated structure? Can you provide corresponding quantitative indicators?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Final Discussion Status

Dear Area Chairs,

Thank you for overseeing the review of our submission. As the author–reviewer discussion period concludes, we briefly summarize the main concerns raised across the reviews, how we addressed them in the revised manuscript and follow-up responses, and the resulting discussion status.

DuetRNA couples base and sugar frames in a flow-matching model, using RNA-specific geometric priors to jointly generate sequences and all-atom structures. We are pleased that all three reviewers recognized the geometric motivation for this dual-frame representation and its empirical benefits. We addressed their concerns with the additional experiments, analyses, and revisions summarized below.

  • Frame representation and coupling (kmG3 and exCU). We repeated reconstruction with compatible decoders, trained a Base-only control, and tested the suggested bond-length/angle supervision. These experiments support the benefits of the dual-frame representation and relative-pose coupling (Tables 3–5; Appendix D.5).

  • Geometry and physical validity (all three reviewers). Joint validity and FR3D analyses confirm DuetRNA's gains over single-frame controls. We clarified glycosidic bond errors and specified a closed-form post-processing step. Training on curated data further improves backbone continuity and GS validity without architectural changes (Table 3; Appendix E.6).

  • External baselines (eQvA and exCU). We documented that RiboDiff's training/inference code and trained checkpoints were not publicly available. Our follow-up using RNA-FrameFlow outputs from Table 2 confirms DuetRNA's higher joint validity, including with bond-geometry filtering. We also report clash comparisons on a shared atom set.

  • Presentation (eQvA). We moved the model overview to Figure 1, reorganized Section 4, and clarified terminology and prose. We also explained training and sampling variability (Table 7) and the trade-offs between all-atom and frame-based generation.

As a result, following three rounds of discussion, Reviewer eQvA (Confidence: 4) raised their rating from 1 → 2 → 3 (Weak Acceptance) and confirmed they are inclined to accept the paper. Reviewer kmG3 (Confidence: 3) remains at 3 (Weak Acceptance). The revised manuscript addresses both technical questions with additional experiments and analyses. Reviewer exCU (Confidence: 3; Rating: 2) explicitly confirmed in their latest comment that Q1, Q3, and Q4 are resolved, leaving only the external-baseline joint-validity comparison, which we have now provided in our follow-up. At the time of writing, exCU has not yet responded to our follow-up, so we respectfully ask the Area Chairs to assess this evidence alongside the revised manuscript and the reviewers' updates.

Thank you for your time and careful consideration of our submission.

Best,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.