acceptodds
Under review as a conference paper at ICLR 2027

RNA Design for Target Structure by Language Model

Abstract

RNA design, the task of finding a sequence that folds into a target secondary structure, has broad biological and biomedical impact but remains computationally challenging due to the exponentially large sequence space and exponentially many competing folds. Traditional approaches treat it as an optimization problem, relying on per-instance heuristics or local search. We instead reframe RNA design as conditional sequence generation and introduce a reusable neural approximator, instantiated as an autoregressive language model (LM), that maps target structures directly to sequences. We first train our model in a supervised setting on random-induced structure-sequence pairs, and then use reinforcement learning (RL) to optimize end-to-end metrics. We also propose methods to select a small subset for RL that greatly improves RL efficiency and quality. Across four datasets, our approach outperforms state-of-the-art systems on key metrics such as Boltzmann probability while being faster, establishing conditional LM generation as a scalable, reusable alternative to per-instance optimization for RNA design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA design, language model, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ePTVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper frames RNA inverse folding as conditional sequence generation. A pretrained LM is adapted to this setting, trained by SL on structure-sequence pairs generated by SAMFEO, then fine-tuned with GRPO on a filtered subset. The authors report that their model, taking the best of 10⁴ samples per target, achieves the highest p(y*|x) and the best NED.

Critical strengths
  • The method is clearly specified and cleanly give a solution a well-defined subtask of RNA design.
  • decontamination of the RL set is reasonable and principled, the threshold is calibrated against cross-family minimum structural distances in RNAStrAlign.
Critical weaknesses
  • The method is a direct combination of mature components, with little methodological innovation specific to this problem.
  • The baselines are from 2023 or earlier.
Questions for authors
  • The margins in Table 1 are small in absolute terms (e.g. 0.581 vs. 0.588 in p(y*|x) on Eterna100), and no seeds, variance, or significance tests are reported. Could the main results be reported over multiple runs with variance?
  • Why is there no comparison with post-2023 methods such as DesiRNA or Montparnasse?
  • Sec. 4.1 gives four reasons for not using existing RNA LMs, but at least three do not apply to nucleotide models such as Evo 2, which is autoregressive, operates at single-nucleotide resolution, and has public checkpoints. Given that the paper’s central claim is that pretraining transfers, could the authors explain the choice of a text-pretrained backbone over a nucleotide-pretrained one?
  • Table 1 reports four test sets, but Fig. 5 and App. Figs. S2–S4 cover only three, with Eterna100-v2 appearing only in Table 1. Could the authors explain the omission?
Reviewer_NHcnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops a reusable, structure-conditioned RNA inverse-folding model from Qwen2.5-0.5B. It combines supervision on ten million SAMFEO-generated pairs, pairing-constrained decoding, and group-relative policy-gradient refinement using ViennaRNA target probability. AoN/NSD filtering selects a smaller RL target set. Best-of-10,000 evaluation includes standard benchmarks, long-target analysis, and independent folding engines. The evidence supports a useful integrated training recipe; formulation-level priority and the specific contribution of the selection rule need narrower positioning.

Critical strengths

The benefit is larger on difficult long targets than overall arithmetic means suggest. Table S3 reports geometric-mean target probability of 0.0153 for SAMFEO versus 0.163 for SL+RL on 19 MFE-solvable Eterna100 targets longer than 200 nt (10.7x). Figures 4-5 additionally support pretrained initialization and RL refinement within the evaluated recipe (pp.7-8,18).

Appendix S3 includes a substantive freeze-then-evaluate check: the same ViennaRNA-selected designs improve target probability and reduce normalized ensemble defect versus SAMFEO under both independent engines on all three evaluated benchmarks. This is evidence beyond reranking with a new engine. MFE results are mixed, so universal metric superiority is not established (pp.19-20, Tables S7-S9).

Critical weaknesses

Section 8 makes a broad first-study claim for autoregressive language models in RNA inverse folding. RWKV-IF (bioRxiv, 2025; doi:10.1101/2025.06.13.659654) already describes structure-to-sequence conditional language modeling and synthetic training from random sequences. This warrants a precise comparison and narrower contribution statement. The abstract alone does not establish the exact autoregressive setup or duplication of this complete training recipe; the criticism should therefore be understood as a positioning question, rather than a claim that the methods are identical.

AoN/NSD filtering has a plausible motivation and an existing full-set/subset comparison, but Figure S6 changes both target selection and training-set size and evaluates on the parent RL pool. It supports a useful subset, without isolating the particular selection principle. Its legend gives 2.8K selected plus 10.3K remaining structures, inconsistent with the main pipeline's 2.7K selected from 12.2K. The reported 1.74x runtime advantage also applies to the disclosed H100-versus-five-CPU-worker configuration, not general resource efficiency (pp.5-9,17).

Questions for authors
  1. How should the first-study claim be scoped relative to RWKV-IF, and which specific contribution beyond structure-conditioned language modeling is supported by this paper’s evidence?

  2. What conclusion about the AoN/NSD selection principle can the existing full-set/subset comparison establish, and which experiment populations explain the Figure S6 counts relative to the main pipeline?

Reviewer_Nm83OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a language model-based method for RNA secondary structure inverse design. The method uses Qwen2.5-0.5B as the backbone, performs supervised fine-tuning on structure–sequence pairs generated by SAMFEO, and combines constrained decoding with reinforcement learning to directly optimize the Boltzmann probability of target structures.

Critical strengths

The paper combines language models, constrained decoding, and reinforcement learning into a reusable framework for RNA secondary structure inverse design.

Critical weaknesses
  • The central idea overlaps considerably with prior work. For example, [1-3] already learn reusable RNA design policies that can generate sequences for unseen target structures. Although [2-3] use RL policies rather than pretrained language models, their underlying ideas are closely related to conditional generation. More directly, Struct2SeQ [4] uses a structure-conditioned autoregressive Transformer decoder with reinforcement learning for RNA inverse design. Therefore, reframing RNA inverse design as conditional sequence generation is not a new idea, and the paper does not sufficiently discuss its differences from these existing works.

  • The authors adopt Qwen2.5-0.5B pretrained on natural language, but reinitialize its input embeddings and LM head. Although it outperforms the same architecture trained from scratch, this does not necessarily establish the specific value of natural-language pretraining for RNA design. Moreover, the necessity of a 0.5B-parameter model for a task involving only four nucleotides and sequences up to 400 nt is not well justified.

  • Constrained decoding samples from a masked and renormalized probability distribution, whereas the GRPO loss in Section 4.3 is written using the original model probabilities. If the gradients are computed using the unmasked probabilities rather than the actual sampling policy, the resulting policy gradient would be inconsistent. In particular, positions with only one valid action have a probability of 1 under constrained decoding and should not contribute a corresponding policy gradient.

  • The proposed GRPO is essentially a group-normalized REINFORCE objective that assigns the same sequence-level advantage to all nucleotide generation actions. It does not specifically address long-range dependencies or credit assignment in RNA design.

  • The experiment lack comparisons with simpler alternatives such as standard REINFORCE or reward-weighted supervised fine-tuning.

  • The RL data selection relies on two empirical thresholds without sufficient justification.

  • NSD is a coefficient of variation of rewards and does not necessarily reflect the learning potential of a target structure.

  • The improvements over SAMFEO in average Boltzmann probability are quite small. For example, the score on Eterna100 only increases from 0.581 to 0.588, while the number of MFE solutions actually decreases from 77 to 75.

[1] Eastman et al., Solving the RNA Design Problem with Reinforcement Learning, PLOS Computational Biology, 2018.

[2] Runge et al., Learning to Design RNA, ICLR2019.

[3] Hu et al., CocoRNA: Collective RNA Design with Cooperative Multi-agent Reinforcement Learning, ICML2026.

[4] He et al., Struct2SeQ: RNA Inverse Folding with Deep Q-Learning, 2026.

Questions for authors

Please refer to the Critical Weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.