acceptodds
Under review as a conference paper at ICLR 2027

RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning

Abstract

RNA design aims to identify sequences that fold into specified secondary structures. Existing methods formulate the task as target-specific search or conditional generation. However, natural RNA evolution proceeds through sequence variation and selection, with compensatory substitutions, whereas these methods do not explicitly model this process. To address this limitation, we propose a two-stage framework comprising RNA Inverse-Folding Flow (RNA-IFlow) and RNA-IFlow-RL. RNA-IFlow uses structure-conditioned Dirichlet Flow Matching to model coordinated variation across the sequence, while RNA-IFlow-RL maps the learned flow to a pairing-preserving finite policy and refines it with thermodynamic feedback. Our framework achieves leading performance on multiple benchmarks, reaching 85.19% Pass@1 on Rfam-27. Further analyses reveal thermodynamic gains, policy dynamics, and robustness across settings. Our work couples coordinated variation with thermodynamic selection, offering a novel paradigm for RNA design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA design, flow matching, reinforcement learning, thermodynamic optimization, policy refinement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_5Eh8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a two-stage framework, RNA-IFlow and RNA-IFlow-RL, that combines structural conditional Dirichlet Flow Matching with reinforcement learning for RNA inverse folding design. The method achieves thermodynamic feedback-based sequence optimization through finite strategy mapping and structural constraints.

Critical strengths
  1. The method is somewhat innovative. By transforming continuous flow matching into a finite reinforcement learning strategy and introducing a pair-preservation constraint, it provides a new technical framework for RNA sequence optimization.

  2. The method demonstrates good performance and efficiency. It outperforms existing methods on multiple benchmarks, and its number of parameters and inference time are significantly lower than those of RNA-Design-LM.

Critical weaknesses
  1. The effectiveness of Flow-to-Policy Mapping has not yet been fully demonstrated. Table 10 in the appendix shows that finite policies not optimized via RL perform significantly worse than the original Flow model, indicating that the mapping itself fails to effectively preserve generative performance.

  2. The improvement over the strongest baseline is limited. For example, the Pass@8 score for Eterna100-v2 increased only from 63.67% to 65.00%, and further evidence is needed to demonstrate the statistical significance of this improvement.

Questions for authors
  1. Why does Flow-to-Policy Mapping significantly reduce initial generation performance?
  2. Can authors provide a comparison with other reinforcement learning methods using the same backbone network and training budget to demonstrate the independent contribution of limited policy design?

I will raise my score if authors can anwser my questions clearly.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5Eh8

Thank you for recognizing the framework and the efficiency results, and for asking about the mechanism and the controls.

W1 and Q1 — Mapping rationale and initial performance loss. FPM converts the continuous Dirichlet predictor into a finite policy whose transition probabilities RL can evaluate explicitly. Conditional-mean inputs reuse the predictor, and normalized keep-or-resample kernels define actions over structural units. Shorter inference is a secondary benefit.

The conversion changes initialization, inputs, and dynamics together, so the exact conditional mean does not guarantee that the native sampling distribution is preserved. A nonlinear predictor evaluated at a mean need not equal the average of its predictions over random inputs. P@8 falls from 52.00% for native flow to 24.33% for the finite policy, then reaches 65.00% after RL. This shows effective post-training alongside an initial conversion cost.

W2 — Significance and practical value. On Eterna100-v2, P@8 rises numerically from 63.67% to 65.00%. The practical contribution also includes the smaller model and faster measured generation: 87.55M versus 357.92M parameters, and median resident-model K=8 generation of 0.73s versus 4.41s. These support a favorable quality–cost trade-off under the reported protocol.

Q2 — Matched RL comparison. This control has not been completed. RNA-Design-LM differs in backbone, capacity, and generation procedure, and Tables 11–12 vary our own hyperparameters. Following your suggestion, the follow-up comparison will start from the same supervised checkpoint and fix targets, success criteria, rollout and optimization budgets, checkpoint selection, and evaluation.

Thanks for your response. But I did not see the "follow-up comparison"?

AuthorsOpenReview ↗

Response to Reviewer 5Eh8

Thank you for checking. The follow-up comparison has not been completed. The controlled comparison with other reinforcement-learning formulations on the same backbone and budget remains planned work: it requires adapting each formulation to the continuous Dirichlet predictor and re-running post-training from the same supervised checkpoint, which we could not complete within the discussion period.

Table 4 lists what the current results contain. Tables 11 and 12 vary the hyperparameters of our own method; none of them isolates the policy design from the RL algorithm, and the revised paper will state this limitation explicitly.

Table 4. P@8 on Eterna100-v2 (Table 10 of the submission).

SettingP@8 (%)
Native flow52.00
Finite policy, no RL24.33
Finite policy + RL (RNA-IFlow-RL)65.00
Reviewer_np34OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors are trying to solve the problem of RNA design by introducing finite-policy reinforcement learning. Specifically, RNA-IFlow and RNA-IFlow-RL were introduced. Compared with the existing pipelines, this two-stage framework is trying to decode the evolutionary inform behind the natural RNAs. However, the overall quality of this submission is not that good. Specifically, several pieces of information are missed / incomplete (see below).

Critical strengths

The authors are trying to solve the problem of RNA design by introducing finite-policy reinforcement learning. Specifically, RNA-IFlow and RNA-IFlow-RL were introduced. Compared with the existing pipelines, this two-stage framework is trying to decode the evolutionary inform behind the natural RNAs.

Critical weaknesses
  1. [Major] The authors should, at least, introduce the full form of each metric at the first appearance. Specifically: What is P@1? What is P@8? Why are these metrics selected to evaluate the performances of each pipeline? According to the initial submission I have received, it is quite hard for me to grasp the flow of the paper. For the inverse folding task, several key metrics (i.e., overall recovery rate, sequence similarity) are missed. Additionally, compared with the secondary structure, the corresponding 3d structure of each generated sequence is definitely worthy of being paid more attentions. For instance, have the authors ever evaluated the structural difference / similarity (e.g., global RMSD) among the ground truth RNAs and the generated sequences from your framework?
  2. [Major] Have the authors trained / fine-tuned the framework on (real) experimental data? If the answer is yes, please include some more technical details during the rebuttal period. If the answer is no, why don't you do this? As the authors mentioned (i.e., lines 295-301), several public datasets are involved. Similarly, why are they selected? Have you ever tested how different the shortlisted datasets are? Based on the best of understandings, most of the training data (i.e., 10 million sequence-structure pairs) the authors used should be in silico data. Then have the authors assessed the quality of each entry, except for the filtering based on the sequence length? If the answer is yes, please at least disclose some statistical results afterwards. Otherwise, I may question the overall quality of the submission.
  3. [Major] Regarding the case study (i.e., Figure 4), why did the authors pick this reference sequence? Is it a 'real' RNA whose 3D structure has been solved? If the answer is yes, please at least include the PDB accession inform (e.g., PDB-ID). If the answer is no, what is the significance of this 'design'?
  4. [Major] This submission needs to be polished. It cannot be accepted at this present.
  5. [Major] It's a good practice to include detailed subheading for each section in Appendix. However, the authors should at least refer the corresponding table(s) (e.g., Tables 13 - 20) while drafting the text, instead of only pasting the table itself.
Questions for authors

Please refer back to "Critical Weaknesses".

2 comments
AuthorsOpenReview ↗

Response to Reviewer np34

Thank you for the suggestions on evaluation, data quality, and presentation.

W1 — Metrics and structural evaluation. The revision will define each metric at first use. P@K is the fraction of targets for which at least one of the first K candidates has the target as its unique MFE secondary structure under the stated evaluator; MFE@8 allows tied minima. Target probability is the target's Boltzmann probability, normalized ensemble defect is the expected fraction of incorrectly paired nucleotides, and Pair-F1 measures base-pair agreement. Together they capture folding success and ensemble quality.

Table 33 reports sequence recovery for the case study. Many sequences can fold to the same target, so similarity to a single reference is neither necessary nor sufficient for inverse-folding success. The task is sequence design for a specified secondary structure, and we did not perform three-dimensional filtering or RMSD evaluation. The current conclusions concern computational secondary-structure folding only.

W2 — Training data and quality. Section 4.1 already lists the data sources; the revision will make their roles explicit. RNAErnie's published pretraining uses biological sequences from RNAcentral. Design-specific supervised training uses the 10 million solver-generated sequence–structure pairs released with RNA-Design-LM. RL uses computational thermodynamic rewards on the released EternaWeb targets. None of the design stages use experimentally measured folding labels.

Our construction checks cover format, legality, length, and duplicate pairs. Retained supervised lengths are 6–500 nt. In the 1–64, 65–128, 129–256, and 257–510 nt bins, the supervised fractions are 7.97%, 13.22%, 27.10%, and 51.71%, and the RL fractions are 29.82%, 34.59%, 18.35%, and 17.24%. Exact target-structure overlap with the evaluation sets is reported separately: nine RNAsolo target structures occur in supervised training, and Tables 7–8 report results with and without them.

W3 — Figure 4 reference. Anemone is a 214-nt Eterna target (puzzle 8935338). The case illustrates computational inverse folding for a specified secondary structure. Its reference is the player-designed Sample Solution (V2/Vienna2), not an experimentally resolved RNA with a cited PDB structure, and the caption will say so.

W4 and W5 — Presentation and appendix references. The revision will introduce notation before the equations, correct the formulas identified by the reviewers, and separate the variation–selection motivation from measured findings. Tables 13–20 and the other appendix tables will be cited at the relevant points in the main text, each with a brief statement of its question, fixed quantities, and finding.

Reviewer_np34OpenReview ↗

Re: Authors' Responses

Dear Authors:

Thank you very much for your clarification. At this present, I do not see an updated PDF, including your planned revisions listed above. Therefore, I will keep my initial rating.

Reviewer np34

Reviewer_dLnsOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for RNA secondary structure inverse folding that combines conditioned Flow Matching with reinforcement learning. The first stage uses structure-conditioned Dirichlet Flow Matching to learn an RNA sequence distribution, with structure-aware decoding to enforce valid target base pairs. The second stage converts the pretrained model into a finite-step discrete policy and further optimizes it using paired-site updates and PPO with thermodynamic rewards.

Critical strengths

The paper presents an RNA design framework combining flow matching, structural constraints and RL, enabling direct sequence generation at inference time without additional thermodynamic search.

Critical weaknesses
  1. The central idea overlaps considerably with existing work. Prior studies [1,2] already use RL to learn RNA design policies that generalize to unseen target structures, yet neither is discussed in the paper. Moreover, [3,4] have studied policy-gradient optimization of flow matching models, but are only briefly mentioned in the appendix. The paper does not sufficiently discuss its connections to and differences from these works, raising concerns about its methodological novelty.
  2. The paper emphasizes modeling RNA evolution through a variation–selection process, but the actual method does not model evolutionary trajectories or use sequence covariation information. The so-called coordinated variation mainly relies on bidirectional contextual predictions and valid base-pair constraints. In particular, SPD simply multiplies the predicted probabilities at two positions and restricts the results to six valid base pairs, while [5] already uses joint actions for paired positions. I am not convinced that the method really captures compensatory substitutions beyond ordinary structure-conditioned generation.
  3. I have concerns about the justification for flow-to-policy mapping. Although Eq. (6) uses the exact conditional mean of the Dirichlet distribution, this does not imply that the resulting discrete policy approximates the original flow matching dynamics. The new policy no longer follows the learned velocity field, but instead uses a manually defined keep-or-resample transition. More importantly, Table 10 shows that switching from the original flow-based generator to the finite policy without RL reduces P@8 from 52% to 24.33%. This raises questions about whether FPM actually preserves the generative capabilities of the pretrained model.
  4. The “Thermodynamic trajectory refinement” essentially uses standard group-relative advantages and PPO, assigning the same terminal advantage to every action along a trajectory. It does not introduce a new credit assignment mechanism. Moreover, Table 12 shows that using 2 and 8 policy steps achieves P@8 scores of 65.11% and 65.78%, respectively, with only a small difference. This also raises questions about the actual necessity of the proposed multi-step variation and refinement process.
  5. The PPO formulation may face issues with high-dimensional policy updates. Eq 12 computes the importance ratio from the product of action probabilities across all structural units. For long RNA sequences, even small changes at individual positions can make the joint ratio move far outside the PPO clipping range, potentially leading to frequent clipping.
  6. The paper lacks a matched comparison using the same backbone with standard masked-token pretraining, followed by the same finite policy and RL procedure. This makes it difficult to isolate the contribution of flow matching pretraining. Similarly, alternative FPM constructions and approaches that directly optimize the original flow matching process are not sufficiently examined.
  7. The improvements over RNA-Design-LM SL+RL are quite small. For example, P@8 on Eterna100-v2 only increases from 63.67% to 65%, and the reported 95% confidence interval includes zero. On RNAsolo, the method achieves lower P@8, and at a sampling budget of 1,024 on Eterna100-v2, its P@K also falls below RNA-Design-LM SL+RL.

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

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

[3] McAllister et al., Flow Matching Policy Gradients, ICLR2026.

[4] Su et al., Discrete Flow Matching Policy Optimization, 2026.

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

Questions for authors

Please refer to Critical Weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer dLns

Thank you for the thoughtful questions on the method's contribution, the conversion step, and the supporting controls.

W1 — Relationship to prior work. The main text will discuss Eastman et al., CoCoRNA, FPO, and DoMinO. As outlined in the General Response and our reply to Reviewer FakN, RNA-IFlow differs in how the policy is built: it converts a continuous Dirichlet predictor into a finite policy whose transitions jointly update legal target pairs or unpaired sites.

W2 — Variation–selection and paired updates. Variation–selection is the design motivation (see General Response), and we will remove wording that could suggest phylogenetic or natural evolutionary information. SPD normalizes the product of two contextual nucleotide distributions over six legal pair states, so every update from a legal initialization stays pair-compatible. Pair legality is a constraint on the policy, not a guarantee that a design folds to the target.

W3 — Why introduce FPM. FPM gives RL explicit action probabilities while reusing the pretrained Dirichlet predictor. Native flow integrates continuous simplex states, whereas FPM uses normalized keep-or-resample transitions. The conversion changes initialization, predictor inputs, and dynamics together, and a nonlinear predictor evaluated at a mean need not equal the average of its predictions over random inputs. P@8 falls from 52.00% for native flow to 24.33% for the finite policy, and post-training then raises it to 65.00%.

W4 — TTR and horizon length. TTR applies a group-relative terminal advantage to each transition through a PPO surrogate, using established GRPO/PPO components. Table 12 reports P@8 of 65.11% at H=2 and 65.78% at H=8, each from three 64-update continuation runs from a common warm-start.

W5 — Joint ratios and clipping. We agree that small unit-level changes can accumulate in the joint log ratio. In the final 350-update segment, trajectories are refreshed before each optimizer step, and the logged ratio diagnostics are computed before the update. The revision will add a length-stratified post-update ratio and clipping diagnostic.

W6 — Matched controls. The appropriate control for the pretraining objective is a same-backbone comparison of masked-token and flow-matching pretraining under the same finite-policy/RL procedure. Comparisons with other RL formulations require faithful adaptation to a continuous Dirichlet predictor. Both are follow-up work (see General Response).

W7 — Success rates and efficiency. On Eterna100-v2 at K=8, the P@8 advantage is small (+1.33 points). Our main case is the quality–cost trade-off: 87.55M versus 357.92M parameters and median resident-model K=8 generation of 0.73s versus 4.41s. The revised discussion will also cover RNAsolo-764 and larger sampling budgets.


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

The paper proposes a two-stage framework combining flow matching and reinforcement learning for RNA design. The first stage models coordinated sequence variation through structure-conditioned Dirichlet flow matching, while the second stage converts the learned flow into a finite RL policy for thermodynamic selection with structural constraints on both stages. Experiments on multiple benchmarks demonstrate the SOTA performance of the model.

Critical strengths

-- The task is very important, there has been a significant amount of work done in this domain with not so much significant progress yet.

-- Motivation for the selected approach, while not perfectly sound, is reasonable.

-- Structure aware terminal decoding is a nice approach to partially mitigate the lack of conditioning during sampling in discrete flow matching approaches.

Critical weaknesses

-- Notations are hard to follow, e.g. eq. 6. The dependence of x on t is not properly introduced.

-- Due to loss normalization by sequence length in eq. 11, each predicted nucleotide has smaller weight if it comes from longer sequence. As a result, training overfocuses on short sequence predictions. Moreover, the expression in eq. 20 for CE term differs from eq. 11 in normalization (it actually has a correct one).

-- Eq. 10 approaches advantage expression as GRPO, however the citation of it is absent.

-- The model is trained using SAMFEO sequence designs, yet SAMFEO is not present as a baseline. Some partial results are present in the Appendix, they should be more comprehensive and in the main text.

-- The leakages between training and test data are not discussed in sufficent depth: no splits based on structure similarities.

-- The ablation studies for several design choices are absent: accounting for dependencies between paired nucleotides both RNA-IFlow and RNA-IFlow-RL; RL regularization terms, necessity of CE term; RL reward coefficients.

Questions for authors

-- please firstly consult weaknesses above

-- some tables in the Appendix have problems with formatting (e.g. tables 20,27,28,30,33). Please consider reformatting them.

-- it's hard to assess the novelty of the policy part, as it's written not in a complex way and you don't cite previous works. As a results, we don't know if it's something new or repetition of the GPRO-like case.

4 comments
AuthorsOpenReview ↗

Response to Reviewer FakN

Thank you for the detailed technical comments.

W1 — Notation. Before Eq. (6), the revision will define the sequence state xtx_t, the nucleotide xt,jx_{t,j}, the simplex input zt,j=μαt(xt,j)z_{t,j}=\mu_{\alpha_t}(x_{t,j}), and the level αt\alpha_t. It will distinguish discrete policy steps from continuous flow time and use separate notation for clean supervised nucleotides.

W2 — Loss normalization. Eq. (11) describes the executed reduction incorrectly. Training used the sum of valid-token losses divided by the number of valid tokens in each batch, not an average of per-sequence means. We will correct Eq. (11) to match this batch reduction and Eq. (20). The implementation is unchanged, so no retraining is needed. Whether the model is robust on long sequences remains a separate empirical question.

W3 and Q3 — Policy contribution and GRPO. Eq. (10) uses a group-relative advantage and the surrogate uses PPO clipping. The revision will cite GRPO (Shao et al., 2024) and PPO (Schulman et al., 2017).

Our policy construction differs from the closest flow-based RL methods. FPO uses a proxy ratio based on the flow-matching loss, and DoMinO computes probabilities of a native discrete-flow sampler. RNA-IFlow instead builds a pairing-constrained discrete policy from a continuous Dirichlet predictor. Conditional-mean inputs reuse the predictor (Eq. 6), a normalized keep-or-resample kernel defines the transitions (Eq. 7), and joint actions restrict target pairs to six legal states (Eq. 8). Starting from a legal sequence, every transition preserves target-pair legality, and the transition probabilities supply the RL ratio (Eq. 12).

We regard this RNA-specific construction, motivated by variation–selection, as the contribution. GRPO/PPO and paired actions have precedents. FPO and DoMinO could also be adapted to RNA constraints; matched comparisons remain follow-up work.

W4 — SAMFEO. The existing SAMFEO results (Appendix A.9, Tables 26–27) will move into the main comparison. Over 100 Eterna100-v2 targets with three runs each, SAMFEO reaches 66.67% uMFE success under a 60-second search cap per run. Its mean search time is 43.30s, and 172 of 300 runs reach the cap while still returning candidates. RNA-IFlow-RL reaches 65.00% P@8 with eight samples generated in a median of 0.73s under its timing protocol. Capped search success and fixed-eight-candidate success are different evaluations, so the table will separate the two budgets and report timeouts, returned counts, and scoring scope.

W5 — Training/test separation. Exact full-target dot-bracket matching finds no overlap between the supervised corpus and Eterna100, Eterna100-v2, or Rfam-27, and nine overlaps with RNAsolo-764. The RL targets have no exact overlap with either Eterna collection. The RNAsolo-755 analysis excludes the nine supervised overlaps.

These checks cover exact structure identity only, not approximate structural similarity, sequence homology, or family membership, and the revision will state this scope. We also plan a structure-based ID/OOD evaluation with similarity criteria and split rules fixed before evaluation.

W6 — Missing ablations. The KL term anchors the structured resampling distributions to a reference distribution, and the CE term retains clean-nucleotide supervision. Table 18 tests coefficient sensitivity only, and removal ablations for KL, CE, pairing, and reward components remain follow-up work. The pairing controls will distinguish terminal decoding from paired actions inside the finite policy.

Q2 — Table formatting. Following your suggestion, Tables 20, 27, 28, 30, and 33 will receive consistent widths and explicit units, and their captions will state the comparison and the fixed quantities.

I thank the authors for the provided replies.

  • The utility of the method is questionable if it loses in performance to the model its training data came from.
  • Additionally, the compute budget is unclear, if SAMFEO runs on CPU, the setups are not directly comparable.
  • Finally, why compare a pass@3 setup against pass@8? Please provide pass@1 for both methods alongside the rest of the metrics.
  • It will be also useful if the authors upload a revised version of the manuscript with other points addressed.
AuthorsOpenReview ↗

Response to Reviewer FakN

Thank you for the follow-up. We answer your three points in turn. All numbers are on Eterna100-v2 over the same 300 target–seed groups (100 targets × 3 seeds).

1. Success rate and utility. SAMFEO generated the supervised corpus and is a per-target search method. As Table 1 shows, it reaches a higher success rate than RNA-IFlow-RL (P@8: 66.67% vs. 65.00%; P@1: 66.33% vs. 54.00%) at a much larger per-target cost. The value of RNA-IFlow-RL lies in this trade-off between success and cost.

Table 1. Same 300 groups. Each SAMFEO group is a single native search, not a three-run union.

MethodReturnedP@1 (%)P@8 / any (%)Time per targetDevice
RNA-IFlow-RL854.0065.000.73 s (median)GPU
RNA-Design-LM SL+RL851.3363.674.41 sGPU
SAMFEO (60 s cap)9.9266.3366.6743.30 s (mean)CPU

2. Pass@3 versus pass@8. The SAMFEO value is not a three-run union. Both methods are averaged over the same 300 groups: SAMFEO returns 9.92 candidates per group and RNA-IFlow-RL returns 8. For SAMFEO, P@1 is the first sequence in the stored return order (66.33%; 66.67% if the highest-probability sequence is used). In 199 of its 200 successful groups every returned sequence is a uMFE hit, so its P@1 and group-level success nearly coincide.

3. Compute budget. We agree that CPU search and GPU generation are not directly comparable. Table 2 truncates the stored SAMFEO searches at shorter caps: the retained set at time T is the last frontier checkpoint produced by T, rescored with the same scorer. SAMFEO needs about 5 s to match the P@1 of RNA-IFlow-RL and 20–30 s to match its P@8.

Table 2. Success versus time cap. SAMFEO times include process start-up (about 0.4 s) on a shared CPU node; the RNA-IFlow-RL time is resident-model GPU generation. The comparison gives an order of magnitude, not an equal-hardware benchmark.

Method (time)P@1 (%)Success (%)
SAMFEO, 1 s37.3342.33
SAMFEO, 2 s46.0051.00
SAMFEO, 5 s54.6758.33
SAMFEO, 20 s62.3363.67
SAMFEO, 30 s64.3365.67
SAMFEO, 60 s66.3366.67
RNA-IFlow-RL, 0.73 s54.0065.00

The revised table will list hardware, wall-clock time, returned candidates and timeouts for each method, together with the one-time training cost of RNA-IFlow-RL (supervised training: 84.1 A100 GPU-hours). Internal oracle counts were not recorded for the native searches.

Long targets. RNA-IFlow-RL is stronger than SAMFEO on targets longer than 256 nt (Table 3): it solves 12 of 25 such targets versus 9 for SAMFEO and 0 for FastDesign, and four targets are solved only by RNA-IFlow-RL, against one only by SAMFEO. On these targets 72 of 75 SAMFEO searches reach the 60 s cap, and SAMFEO's success stays flat as the budget grows (0.04, 0.32, 0.33 and 0.35 at 1, 5, 10 and 60 s), below the 0.48 of RNA-IFlow-RL. Taken together with Table 2, the advantage of RNA-IFlow-RL is largest where search is most expensive. On targets up to 256 nt SAMFEO is stronger (60 vs. 53 of 75 targets).

Table 3. By target length. A target is solved if it succeeds in at least 2 of its 3 groups; group success is the mean over target–seed groups.

≤ 256 nt (75 targets)> 256 nt (25 targets)
Solved: RNA-IFlow-RL5312
Solved: SAMFEO609
Solved: FastDesign500
Solved only by RNA-IFlow-RL04
Solved only by SAMFEO71
Group success: RNA-IFlow-RL0.710.48
Group success: SAMFEO0.770.35
SAMFEO searches at the 60 s cap100/22572/75

Revised manuscript. We are not uploading a revised PDF during the discussion period. The changes are specified in the General Response and in our individual replies.

I thank the authors for their additional clarifications, but I tend to maintain my initial evaluation scores.

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

This paper tackles the problem of RNA inverse folding by proposing a flow based framework consisting of RNA-IFlow and RNA-IFlow-RL. The framework is capable of modelling sequence variation over the complete RNA sequence and can enforce consistent compensatory substitutions during variation. GRPO style RL is also used for further improving the flow model. Experiments show that RNA-IFlow-RL obtains competitive performance on 3 benchmarks, and have faster inference speeds compared to other baseline models.

Critical strengths
  1. The models achieve better performance while achieving significantly faster inference speed compared to other baseline models.
  2. The evolution inspired paired updates gives a principled way to preserve RNA structural constraints during sequence refinement.
Critical weaknesses
  1. There is a large overlap between Eterna100 and Eterna100-v2 benchmark (81/100 structures), this should be highlighted in the table, or the table should represent this as an aggregated column containing the model performance over the union of the 2 benchmarks.
  2. The RNAsolo benchmark is hidden in the appendix, while showing that the RNA-Design-LM paper, which is very similar to this paper in methodology (SL/RL pipeline, 0.5/0.25/0.25 thermaldynamic reward etc), outperforms RNA-IFlow on every metric. This should be included in the main results table. Also, Eterna100 and Eterna100-v2 are all synthetic benchmarks, and rfam-27 is relatively small compared to RNAsolo-764. It seems like RNA-IFlow loses its advantage on non-synthetic target structures, this weakens the broader applicability of the paper's method.
  3. The main difference between this paper and the RNA-Design-LM paper is that this paper uses a bidirectional backbone model, flow based training, and a structure constrained refinement policy. The lack of ablation studies makes it unclear where did the improvement of the RNA-IFlow-RL model come from. There is one experiment showing that RL did improve RNA-IFlow, but there is no controlled experiments for the individual components mentioned above. There should be extra ablation showing where exactly the gains of the model come from.
Questions for authors
  1. Given the performance on RNAsolo-764, does RNA-IFlow-RL generalize less well on non-synthetic structures as compared to RNA-Design-LM?
  2. Could the authors provide controlled ablations to show where exactly the improvements over RNA-Design-LM come from?
1 comment
AuthorsOpenReview ↗

Response to Reviewer CANy

Thank you for noting the efficiency results and for pressing on evaluation breadth and ablations.

W1 — Benchmark overlap. As you note, Eterna100 and Eterna100-v2 share 81 of 100 target structures, and we will state this explicitly in the note to Table 1.

W2 and Q1 — RNAsolo and generalization. Our results are mixed across natural-family collections, and we do not claim better generalization to non-synthetic structures. However, on Rfam-27, which contains 27 targets from natural RNA families, P@1/P@8 is 85.19%/87.65% for RNA-IFlow-RL versus 81.48%/86.42% for RNA-Design-LM SL+RL. To be sure, RNA-IFlow-RL is indeed slightly behind RNA-Design-LM SL+RL on RNAsolo-764 (73.12% versus 75.31%), though it also has fewer parameters (87.55M versus 357.92M). Following your suggestion, we will move the RNAsolo results into the main table.

On the other hand, we have not tested family-disjoint generalization. Evaluating by RNA family and by synthetic versus non-synthetic source is the natural next step for identifying where the method applies. We will take up this suggestion to improve our experiments, and we appreciate your kind reminder.

W3 and Q2 — Component ablations. Indeed, the requested controlled ablations are missing. Table 10 shows that RL raises the finite policy's P@8 from 24.33% to 65.00%, but it does not separate the contributions of the backbone, the pretraining objective, and the structural policy. The appropriate control uses our own bidirectional backbone, with initialization, data, and training exposure fixed, to compare masked-token and flow-matching pretraining, each followed by the same finite-policy/RL procedure. We plan it as follow-up work. RNA-Design-LM serves as a performance baseline in our work, and its different architecture makes it unsuitable for isolating these components.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful reading. Several comments identify real errors and missing controls. We respond to each below and summarize our position here.

Contribution. RNA-IFlow is motivated by variation under structural constraints followed by thermodynamic selection. Its central contribution is a finite RNA policy built from a pretrained continuous Dirichlet predictor. Conditional-mean inputs reuse the predictor, and normalized keep-or-resample transitions act on unpaired sites and legal target pairs, giving RL explicitly evaluable transition probabilities. This differs from FPO, which uses a proxy ratio based on the flow-matching loss, and from DoMinO, which computes probabilities of a native discrete-flow sampler. To our knowledge, combining flow-based variation with RL-based selection through such a pairing-constrained finite policy has not been done for RNA inverse folding. The contribution lies in this construction; the optimizer uses established GRPO/PPO components. Variation–selection is the motivation for this design, not a claim that the model reproduces natural evolution or recovers phylogenetic information.

Evidence. On the three Table 1 benchmarks, RNA-IFlow-RL has the highest P@1 and P@8 point estimates, with 87.55M parameters versus 357.92M for RNA-Design-LM. Under the reported protocol, resident-model K=8 generation takes a median of 0.73s versus 4.41s. However, the success-rate advantage is not uniform: it is small on the Eterna benchmarks, and RNAsolo-764 and larger sampling budgets favor RNA-Design-LM SL+RL. Converting the continuous predictor into the finite policy lowers P@8 at initialization (52.00% for native flow, 24.33% for the finite policy), and RL then raises it to 65.00%. Our conclusions concern computational secondary-structure folding, and training–test separation was checked by exact target-structure identity.

Revision. We will correct the loss normalization, clarify notation, define the metrics at first use, cite GRPO/PPO and discuss prior work, identify the benchmark overlap (Eterna100 and Eterna100-v2 share 81 of 100 targets), cite the appendix tables with their questions and findings, improve table formatting, and move RNAsolo and the SAMFEO comparison into the main text. Further experiments, including same-backbone comparisons of pretraining objectives, removal ablations, alternative RL formulations, and family- or structure-disjoint evaluation, are discussed in our individual responses, and we will report their results in the revised version where they are completed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.