acceptodds
Under review as a conference paper at ICLR 2027

sCM-RNA: Preference-Aligned Consistency Models for RNA Inverse Folding

Abstract

RNA inverse folding aims to design sequences that fold into prescribed backbone structures, supporting a wide range of applications in RNA engineering. Sequence recovery measures agreement with native sequences, but provides an incomplete assessment of structural fidelity and sequence diversity. Designing diverse sequences with high structural fidelity therefore calls for objectives that evaluate their predicted folds, together with efficient generation. We introduce **sCM-RNA**, an RNA inverse folding framework that combines continuous-time consistency modeling with preference-based alignment. The consistency model enables one-step sequence generation conditioned on target backbone structures, while the alignment stage uses preferences based on predicted structural fidelity to guide generation and explicitly promote sequence diversity. Together, these components support efficient sampling of diverse sequences whose predicted structures closely match the targets. Across six random seeds, sCM-RNA improves TM-score by **2.3%** and sequence diversity by **10.9%** relative to the strongest evaluated baseline for each metric, while achieving the highest sequence recovery. With one-step inference, sCM-RNA also achieves up to **35** faster sequence generation, supporting large-scale computational RNA design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA Inverse Design, Continuous-time Consistency Models, DPO

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_B4D4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces sCM-RNA, a one-step RNA inverse-folding framework that combines structure-conditioned flow matching, continuous-time consistency distillation, and structure-guided preference alignment. A flow-matching model is first distilled into a one-step consistency model, after which RhoFold-based structural preferences and entropy regularization are used to improve structural fidelity while preserving sequence diversity. The method is evaluated on the standard RiboDiffusion split and a recent CAMEO OOD set, with comparisons to GVP-GNN, gRNAde, RiboDiffusion, and the flow-matching teacher. The paper reports improvements in TM-score, recovery, and diversity, together with substantially lower inference latency, and further evaluates robustness using Chai-lab, multiple random seeds, pass@16, and additional ablations on the preference and diversity objectives.

Critical strengths
  1. The paper tackles an important practical bottleneck in generative RNA design by combining one-step consistency generation with structure-guided preference alignment. This gives a favorable quality–efficiency trade-off: the model improves structural fidelity and diversity while reducing generation latency by more than an order of magnitude relative to multi-step diffusion/flow baselines.

  2. The empirical validation is relatively comprehensive. Results are reported over six seeds, on both the standard in-domain split and a recent CAMEO OOD set, and the structural gains are further checked with Chai-lab and fixed-budget pass@16 evaluation. This makes the reported improvements more convincing than relying on a single folding oracle or a single in-domain benchmark.

Critical weaknesses
  1. The Stage-II distillation result is difficult to interpret. Before preference alignment, the one-step student sCM (Table 5: Recovery = 0.5382, TM-score = 0.5956) already outperforms the multi-step flow-matching teacher (Table 1: Recovery = 0.5113, TM-score = 0.5875), while diversity drops from 0.3040 to 0.2818. Since Stage II uses a simplified tangent-residual formulation rather than the original sCM recipe, and the paper neither compares against standard sCM nor varies the teacher sampling steps, it is unclear whether this gain comes from the modified distillation objective, finite-step solver error in the teacher, or simply a more concentrated output distribution.

  2. The justification for the sCM-DPO objective is also not fully convincing. While inference only queries t=0t=0, sCM-DPO samples t∼U(0,1)t \sim \mathcal{U}(0,1) without the consistency loss, and the endpoint reconstruction error scales with (1−t)2(1-t)^2 (vanishing at t=1t=1). The reported 76% ranking agreement in Appendix G.1 is measured only on filtered high-margin extreme pairs at t=0t=0, which does not validate the surrogate across the training time distribution. Moreover, the per-position probability pϕ(si∣X)p_\phi(s_i \mid X) in the token-entropy regularizer in Eq. 8 is not mathematically defined for the continuous non-autoregressive predictor.

  3. It seems like numerical inconsistencies across tables and missing aligned baselines. Purely sequence-level metrics (Recovery, MFE, Diversity) are independent of the 3D folding oracle, yet they vary for the same models across Table 1 (RhoFold), Tables 2–3 (Chai-lab), and Table 5. In addition, baseline TM-scores (gRNAde, GVP-GNN) in Table 1 (0.5032 / 0.5601) differ substantially from Appendix Table 9 (0.2932 / 0.4101), and the FLOPs reduction ratio in Figure 4(b) conflicts with Table 9. Finally, the paper omits comparisons with preference-aligned baselines such as RiboPO or applying DPO directly to the flow-matching teacher.

Questions for authors
  1. Could the authors directly compare the proposed Stage-II distillation with the original sCM training formulation under the same architecture and training budget, and test whether the one-step student still outperforms the flow-matching teacher when the teacher is evaluated with more ODE solver steps?

  2. The base sCM improves TM-score and recovery but reduces sequence diversity relative to the flow-matching teacher, whereas Stage-III alignment subsequently reduces recovery. Could the authors clarify whether the apparent pre-alignment quality gain is mainly due to a more concentrated output distribution?

  3. How well does the endpoint reconstruction error preserve the RhoFold preference ordering across the full time distribution t∈(0,1)t \in (0,1) used during sCM-DPO training? Additionally, how is pϕ(si∣X)p_\phi(s_i \mid X) in the entropy regularizer in Eq. 8 mathematically computed from the continuous predictor fϕf_\phi?

  4. Could the authors clarify why purely sequence-level metrics (Recovery, MFE, Diversity) differ across Tables 1, 2, 3, and 5 for the same models, and why baseline TM-scores and FLOPs ratios differ between the main text and Appendix Table 9?

  5. Tables 3 and 6 evaluate RhoFold-aligned models with Chai-lab, while Appendix G.2 only evaluates Chai-lab-aligned models with Chai-lab. Could the authors complete the cross-oracle evaluation by testing the Chai-lab-aligned model with RhoFold, and compare against applying DPO directly to the flow-matching teacher?

Reviewer_w5TyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces continuous-time consistency modeling towards RNA inverse folding to improve self-consistency.

Critical strengths
  1. The consistency model's sampling speed should be one of the highlights. In Fugure 4, the comparison is very straightforward and clear.
  2. The evaluation protocol is careful and reasonable. The benchmark is well-designed, with the ablation studies well-established.
Critical weaknesses
  1. The novelty of the method should be a problem. Adapting consistency modeling to RNA inverse folding did not bring too much useful insights and novelty.
  2. Although the benchmark setting is good, there are still missing baseilnes for comparison. [1,2]

[1] Tang, J., Guo, J., & Zhao, J. (2026). MORAD: Reinforcement Learning with Multi-Objective Rewards for RNA Inverse Design. arXiv preprint arXiv:2610.13305. [2] Sun, M., Cao, H., Zhang, Z., Wei, C., Wang, L., Jia, T., ... & Zhang, Y. (2025). RiboPO: Preference Optimization for Structure-and Stability-Aware RNA Design. arXiv preprint arXiv:2510.21161.

Questions for authors

See weakness.

Reviewer_4De5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The propose framework distills a flow-matching model into a one-step RNA generator, then fine-tunes it using structural preferences and a diversity objective. I found the combination makes sense, and the results are encouraging. However, my main concern is whether the proposed alignment method adds enough beyond existing approaches.

Critical strengths
  • Fast generation is useful when many candidate sequences need to be screened.

  • The paper considers both structural quality and sequence diversity, which fits the nature of inverse folding.

  • Evaluation with another folding model and secondary-structure checks gives more confidence that the gains extend beyond the training predictor.

Critical weaknesses
  • The comparisons leave the benefit of the proposed alignment unclear. RiboPO is cited but not evaluated. Applying the same preferences to the FM teacher, or simply fine-tuning on preferred sequences, would be useful controls.

  • The preference loss relies on reconstruction error as a proxy. The theory assumes that this score agrees with structural quality, but does not explain why training should produce that relationship. The empirical check provides some support, though its scope is limited.

  • The diversity objective needs a clearer implementation description, particularly how continuous outputs become nucleotide probabilities and how entropy is computed across noise samples. Stability across independently trained models also remains uncertain.

Questions for authors
  • Would the gains remain against the simpler alignment alternatives above, using the same preference data?

  • How sensitive are the results to repeating the full training procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.