acceptodds
Under review as a conference paper at ICLR 2027

RNA-DFM: Masked Discrete Flow Matching for 3D RNA Inverse Design

Abstract

Designing RNA sequences that adopt desired three-dimensional structures is central to engineering functional RNA molecules, yet efficient generation with high structural fidelity remains challenging under limited structural supervision. We introduce RNA-DFM, a masked discrete flow matching model that predicts multiple nucleotide identities in parallel conditioned on the target backbone. To expand structural supervision beyond the limited coverage of experimentally determined RNA structures, we augment training with Rfam sequences paired with filtered structure predictions. We further adopt best-of-N distillation, combining structurally selected candidates with native sequences to finetune the generator on training backbones. The resulting model improves structural quality on unseen backbones under limited candidate budgets, with experiments demonstrating gains in both structural quality and sampling efficiency; Rfam augmentation further improves TM-score and native sequence recovery while reducing RMSD. Among the evaluated methods, pretrained RNA-DFM achieves the highest TM-score, the lowest RMSD, and the highest native sequence recovery of 55.9%.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA inverse design, discrete flow matching, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nSuDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper performs RNA inverse folding: it generates a nucleotide sequence from a 3D RNA backbone. The backbone is a base-neutral graph of three atoms per nucleotide, and a masked discrete flow model with a GVP encoder and a bidirectional decoder reveals masked nucleotides in parallel in K = 8 steps, independent of sequence length. Training combines 4,025 experimental structures with 5,850 Boltz-predicted structures of Rfam sequences, after which the model is fine-tuned by best-of-N distillation on structurally selected sequences paired with native sequences. It reports TM 0.445, lDDT 0.619, RMSD 9.380 A and recovery 55.9% on 88 test backbones, and sampling is 12.8x faster than order-agnostic autoregression, with +0.016 TM at N = 16.

Critical strengths
  1. The parallel decoder gives a measured efficiency gain under a controlled comparison. At K = 8 the median latency is 115 ms, against 19.14 ms per nucleotide for the order-agnostic autoregressive baseline with the same geometric network and data. At N = 16 the TM-score is 0.5196 versus 0.5033, so the 12.8x speedup comes with +0.016 TM; at an equal 1.84 s budget it produces 16 candidates against one, a +0.119 TM difference.

  2. The paper shows with a controlled substitution experiment that higher recovery does not mean a better structure. Within the same design, restoring mismatched paired bases to their native identities raises recovery from 0.5155 to 0.7608 while ViennaRNA base-pair F1 falls by 0.2743, and every intermediate dose has a paired interval below zero (Figure 2A). Recovery alone therefore does not measure structural quality.

Critical weaknesses
  1. The accuracy claims are not supported by the paper's own paired tests. Table 6 resolves only the comparisons against RIDE (lDDT +0.048, p = 0.003; recovery +0.115, p = 0.028); all 14 intervals against NA-MPNN, gRNAde and the w/o-Rfam variant cross zero, as do the four intervals for Rfam augmentation. The gap over NA-MPNN is 0.004 TM (0.445 vs 0.441, p = 0.829), yet the abstract reports the highest TM-score, lowest RMSD and highest recovery. These are point-estimate rankings, not established improvements.

  2. The sampling step count is fixed at K = 8 and never varied. Without a K = 1 result, it is unknown how much quality is lost by collapsing the eight steps into one. If K = 1 matches K = 8, the multi-step flow adds nothing to quality and the benefit comes from parallel decoding alone.

Questions for authors
  1. How much TM-score, lDDT and recovery are lost at K = 1, 2 and 4 relative to K = 8? Please report these five settings. If K = 1 is comparable to K = 8, what does the discrete flow add beyond parallel decoding?

  2. RiboDiffusion is discussed in Sections 1 and 2.1 but absent from Table 1. Please add it, or state why it cannot be trained under the same three-atom representation; the same applies to RhoDesign and RiboPO.

  3. Table 5 reports one clean-test sequence at 80%/80% local identity and coverage on both sides, and 10 at 80%/80% on the shorter sequence, while Section 3.3 mentions only the absence of exact matches. Please state this in the main text and re-report the Rfam comparison with those 10 test sequences excluded.

  4. The distillation comparison uses one fine-tuning seed and 45 backbones, with differences of +0.0105 to +0.0145 TM and standard errors of 0.005 to 0.008. Please repeat it with at least one additional seed.

  5. The abstract reports the highest TM-score, lowest RMSD and highest recovery, while Table 6 shows 14 of 16 paired intervals crossing zero. Please label these as point estimates in the abstract.

  6. Appendix A states that the encoder processes a single target conformation, while gRNAde supports conformational ensembles. Please state this scope in the main text.

4 comments
AuthorsOpenReview ↗

Response to Reviewer nSuD (1/2)

We thank the reviewer for the constructive and specific feedback. The additional experiments directly address the two central concerns regarding structural accuracy and the contribution of multi-step flow generation.

As described in our response to all reviewers, the revised benchmark weights the seven structural families equally rather than BGSU equivalence classes. This provides a more balanced assessment of cross-family generalisation and changes our headline TM-score from 0.445 to 0.414.

W1. Structural accuracy and statistical evidence

The revised evaluation provides stronger empirical support for RNA-DFM across both generation and selection regimes. Table 1 now compares nine methods on the 88 clean-test records.

Mean-of-16 results:

MethodRecoveryTM-scorelDDTRMSD
RNA-DFM0.5710.3810.61812.58
RNA-DFM w/o Rfam0.5110.3370.56613.30
RiboPO0.4520.3540.55713.93
gRNAde0.4870.3510.58013.23
NA-MPNN0.4890.3530.59312.61
RIDE0.4280.3210.55214.94
Struct2SeQ0.2770.2990.51814.77
RiboDiffusion0.4940.3090.55315.21
RhoDesign0.2830.2460.39121.31

Best-of-16 results under Boltz selection:

MethodRecoveryTM-scorelDDTRMSD
RNA-DFM0.5880.4140.64910.45
RNA-DFM w/o Rfam0.5190.3340.57112.24
RiboPO0.4630.3950.61411.28
gRNAde0.4900.3930.62410.82
NA-MPNN0.4920.3830.62512.49
RIDE0.4290.3330.57413.55
Struct2SeQ0.3100.3310.54313.14
RiboDiffusion0.5010.3280.57813.54
RhoDesign0.2890.2550.39120.25

RNA-DFM achieves the best observed performance across all eight metric–evaluation combinations. Its leading performance in both mean-of-16 and best-of-16 indicates that the observed advantage is not dependent on structural selection.

Against the strongest baseline per metric, the best-of-16 improvements are +0.087 recovery, +0.019 TM-score, +0.024 lDDT and a 0.37 Å reduction in RMSD.

Statistical evidence for augmentation. The matched Rfam ablation provides particularly strong evidence for additional structural supervision. The paired improvements are +0.079 TM-score, +0.070 recovery and +0.078 lDDT, with family-level 95% confidence intervals of [+0.056,+0.100], [+0.014,+0.127] and [+0.020,+0.136], respectively. All exclude zero, and the improvements remain directionally consistent under all four weighting schemes.

Cross-family robustness. Relative to NA-MPNN, RNA-DFM leads on five of seven structural families. Cluster 237, where RNA-DFM has a TM-score difference of −0.069, accounts for 59.1% of the records, explaining the sensitivity to aggregation. Excluding any individual family preserves positive TM-score gains (+0.022 to +0.049).

Together, these results demonstrate consistent empirical performance across evaluation regimes and family compositions, complemented by statistically supported improvements from structural augmentation.

W2 and Q1. Contribution of multi-step flow generation

We have conducted the requested controlled ablation over K = 1, 2, 4 and 8, holding the checkpoint, candidate budget (16), temperature (0.1), 5′-to-3′ order, Boltz selection, Protenix scoring and family weighting fixed.

KTM-scorelDDTRMSD (Å)Recovery
10.3480.56015.430.558
20.3850.61213.020.566
40.3970.63411.190.569
80.4140.64910.600.587

Increasing K from 1 to 8 improves TM-score by 0.066, lDDT by 0.089, recovery by 0.029, and reduces RMSD by 4.83 Å.

The pooled candidate widths are 10.0, 10.4, 10.0 and 9.9, respectively, indicating comparable candidate-pool widths across the four settings.

These results directly demonstrate the contribution of multi-step generation beyond parallel decoding alone. Progressive conditioning on previously revealed nucleotides substantially improves structural fidelity, with consistent gains across all four metrics. The results are reported in Figure 2A and Appendix F.

AuthorsOpenReview ↗

Response to Reviewer nSuD (2/2)

Q2. Additional baseline comparisons

The revised Table 1 includes RiboDiffusion, RhoDesign and RiboPO, alongside Struct2SeQ and RIDE. The expanded benchmark evaluates nine methods on the same 88 clean-test records, using the same candidate budget and structural evaluation protocol.

As shown in our response to W1, RNA-DFM achieves the best observed performance across all four metrics under both mean-of-16 and best-of-16 evaluation.

Q3. Sequence contamination and sensitivity analysis

We thank the reviewer for identifying the ambiguity in Table 5. The reported count of 10 represents local alignment hits rather than 10 distinct test sequences. All clean-test local matches involve the same PDB entry, 2OEU chain A, corresponding to a single test-manifest record (RS_03411_0_2OEU_1_A).

We have clarified the counting unit and local-identity criterion in the main text.

To directly address the concern, we repeated the Rfam augmentation comparison after excluding this affected record. Since the record forms a singleton at each level of the hierarchy, its exclusion removes one complete structural family, leaving 87 records across six families.

Metric improvement88 records (7 families)87 records (6 families)
TM-score+0.0788+0.0806
lDDT+0.0782+0.0819
RMSD reduction (Å)1.632.87

Excluding the potentially overlapping record preserves and slightly strengthens the augmentation gains. The paired differences retain their directions under all four weighting schemes.

This sensitivity analysis demonstrates that the effectiveness of Rfam augmentation does not depend on the identified local sequence overlap.

Q4. Additional fine-tuning seed for distillation

We have addressed the requested replication by evaluating two independent fine-tuning seeds for each distillation configuration. Each seed is paired with its own matched uniform-weight control, using the same initial checkpoint and candidate pool.

The reported paired differences are averaged across the two fine-tuning seeds within each record before aggregation, and the evaluation now covers all 88 clean-test records rather than 45 backbones.

Label weightingBest-of-16 TM-score gain95% CI
N = 2+0.0147[+0.0013,+0.0281]
N = 4+0.0112[−0.0028,+0.0252]
Argmax+0.0047[−0.0112,+0.0205]

N = 2 weighting improves 60% of the test records relative to the matched control. Its confidence intervals exclude zero from candidate budgets of four upwards, demonstrating benefits beyond a single selection budget.

The observed ordering across label-weighting strategies supports the motivation for using softer order-statistic weighting rather than restricting supervision to only the highest-scoring candidate.

Q5. Abstract and interpretation of accuracy claims

The abstract has been revised according to the updated family-weighted evaluation. It reports RNA-DFM's observed performance among the nine evaluated methods, while distinguishing these benchmark results from statistically supported improvements established through controlled ablations. In particular, the augmentation effect is supported by our matched experiments and additional statistical analysis.

Q6. Single-conformation input

RNA-DFM is designed for RNA inverse folding conditioned on a specified target backbone conformation. Our benchmark evaluates both RNA-DFM and gRNAde under the same single-conformation setting, ensuring a consistent comparison. Ensemble-conditioned design represents a distinct problem setting.

Summary

The revision addresses the reviewer's principal concerns through an expanded nine-method benchmark, a controlled sampling-step ablation, a contamination-exclusion sensitivity analysis, and two-seed distillation experiments over the full clean-test benchmark.

In particular, the substantial quality improvement from K = 1 to K = 8 demonstrates the contribution of multi-step flow generation, while the contamination sensitivity analysis and replicated distillation experiments strengthen the evidence for our training contributions.

Thank you for the additional experiments. The K ablation addresses my main concern about multi-step generation. The added baselines, overlap clarification, and second fine-tuning seed also strengthen the evaluation.

Regarding the paired tests, non-significant results do not establish equivalence. The new evidence supports Rfam augmentation, while rankings against competing methods should remain descriptive. Please report both weighting schemes and keep this distinction clear.

Most of my concerns are now addressed. I am updating my recommendation from weak rejection to weak acceptance.

Thank you for your constructive feedback and for reconsidering our work! We appreciate your suggestions regarding statistical interpretation and the reporting of both weighting schemes.

Reviewer_Fg4GOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a flow matching model for RNA 3D inverse design. It uses a GVP network to encode the target backbone and generates RNA sequences in parallel by progressively revealing masked nucleotides. It further adopt Best-of-N distillation to fine-tune the generator using structurally selected candidates. Experiments show competitive sequence recovery and structural quality.

Critical strengths

The paper introduces masked discrete flow matching to RNA 3D inverse design, enabling parallel nucleotide generation and substantially faster sequence sampling than autoregressive methods.

Critical weaknesses
  • The proposed method mainly combines existing techniques, with limited methodological novelty. Masked discrete flow matching and its use in structure-conditioned sequence generation have already been studied in [1–3], and discrete diffusion has also been explored for RNA 3D inverse design [4]. Some closely related studies are also not adequately discussed.

  • Augmenting RNA inverse design datasets with predicted structures is not new either. Similar approaches have been used in [5-6]. In this paper, Rfam augmentation only improves TM-score from 0.439 to 0.445, and the difference is not statistically significant.

  • The auxiliary rate regularization uses softplus(logits) to construct predicted rates, while the actual sampler uses transition rates based on softmax(logits). These parameterizations are not equivalent and may even have different optimal probability distributions. Why should this auxiliary loss improve the actual discrete flow matching process? The ablation in Appendix C.3 also shows no statistically significant improvement.

  • Multiple nucleotides are sampled independently within each generation step, although RNA base pairing and long-range interactions involve strong dependencies. Moreover, once a nucleotide is revealed, it cannot be revised. Could this random and irreversible parallel sampling scheme hurt structural consistency?

  • The paper also lacks sufficient analysis of different unmasking strategies.

  • Best-of-N distillation essentially performs supervised fine-tuning on structurally selected sequences, closely following existing Best-of-N distillation ideas [7]. Its advantage over methods that optimize structural rewards or preferences more directly [8-9] is not established. The improvement is relatively small: at N=8, Protenix TM-score increases by only 0.0145 over random-candidate fine-tuning, and the experiments use just one fine-tuning seed.

  • The main experiments do not provide strong evidence that RNA-DFM achieves better structural quality. For example, its TM-score is only 0.004 higher than NA-MPNN (0.445 vs. 0.441, p=0.829), while its lDDT is lower. Most structural performance differences are not statistically significant.

  • The reported 12.8× speedup only covers sequence sampling, excluding the expensive structure prediction and candidate selection steps, so it does not show an equivalent improvement in end-to-end RNA design efficiency.

[1] Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design, ICML

[2] All-atom inverse protein folding through discrete flow matchin, ICML

[3] Co-design protein sequence and structure in discrete space via generative flow.

[4] A Hyperbolic Discrete Diffusion 3D RNA Inverse Folding Model for Functional RNA Design.

[5] RiboDiffusion: Tertiary Structure-based RNA Inverse Folding with Generative Diffusion Model.

[6] Deep generative design of RNA aptamers using structural predictions, Nature Computational Science.

[7] BOND: Aligning LLMs with Best-of-N Distillation, ICLR.

[8] RIDER: 3D RNA Inverse Design with Reinforcement Learning-Guided Diffusion, ICLR.

[9] RiboPO: Preference Optimization for Structure- and Stability-Aware RNA Design.

Questions for authors

Please see Critical Weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer Fg4G (1/3)

We thank the reviewer for the constructive and technically detailed feedback. The revision strengthens the evidence for RNA-DFM through additional analyses of structural augmentation, unmasking strategies, selection-aware distillation and evaluation robustness.

To assess generalisation across structural families, the revised benchmark weights seven structural families equally rather than BGSU equivalence classes. Cluster 237 (52 of 88 records) consequently receives 14.3% instead of 59.1% of the weight. Our headline TM-score changes from 0.445 to 0.414. Appendix Table 7 reports the alternative weighting schemes.

W1. Methodological novelty and related work

RNA-DFM builds upon discrete flow matching [1–4] and introduces task-specific adaptations for RNA 3D inverse design. All four works identified by the reviewer were cited in the original submission. The revised Related Work section clarifies the distinction between these foundations and our RNA-specific contributions:

  • Categorical corruption operates directly on the four-nucleotide alphabet.
  • Masking is aligned with the inverse design objective rather than serving solely as an auxiliary perturbation.
  • Reveal order provides an explicit mechanism for controlling access to base-pair context.
  • Multiple supervision signals from each backbone motivate large-scale predicted-structure augmentation.

Beyond the generative framework, our contributions include decontaminated structural augmentation using 5,850 records, selection-aware best-of-NN distillation with order-statistic weights and a matched control, and evaluation using distinct selection and judging predictors.

W2. Predicted-structure augmentation

Our matched ablation establishes the value of additional structural supervision for RNA inverse design. With identical architecture, training schedules and seeds, the only difference is the inclusion of Rfam augmentation:

MethodRecoveryTM-scorelDDTRMSD (Å)
RNA-DFM0.5880.4140.64910.45
w/o Rfam0.5190.3340.57112.24

Augmentation improves all four metrics, including a substantial TM-score gain of +0.079. Family-level wild-cluster-bootstrap-tt 95% confidence intervals exclude zero for TM-score [+0.056, +0.100], recovery [+0.014, +0.127] and lDDT [+0.020, +0.136]. The improvements remain directionally consistent under record, unique-sequence, equivalence-class and family weighting.

Our contribution extends previous augmentation approaches [5,6] through scale, explicit decontamination and controlled evaluation. We removed 530 predicted structures with TM-score ≥0.5\geq0.5 to any evaluation backbone, and our sequence audit found no exact matches with test sequences. Reference [6] is now explicitly cited in the augmentation section.

W3. Softplus rates and the softmax sampler

The non-equivalence identified by the reviewer is explicitly analysed in Section 3 and Appendix B. The auxiliary rates

r~a=softplus(ℓa)+10−6\widetilde{r}_a=\mathrm{softplus}(\ell_a)+10^{-6}

need not sum to the prescribed hazard. Consequently, the equivalence between the rate loss and weighted cross-entropy does not apply. The gradient is

sigmoid(ℓa)(1−ra⋆/r~a)\mathrm{sigmoid}(\ell_a)(1-r_a^\star/\widetilde{r}_a)

The unconstrained optimum over nonnegative rates is c(t)πac(t)\pi_a. However, the softplus floor and shared logits restrict this solution, meaning that the combined objective need not attain the categorical optimum. Setting r~a=c(t)qa\widetilde r_a=c(t)q_a would instead recover weighted cross-entropy.

Our two-seed ablation further characterises the empirical effect:

MetricCECE + Rate
TM-score0.35150.3757
Recovery0.48880.5413
lDDT0.56670.5598
RMSD (Å)12.5913.39

The auxiliary objective improves TM-score and recovery, with trade-offs in lDDT and RMSD. We explicitly report these mixed effects and distinguish this training configuration from the principal contributions of the paper.

AuthorsOpenReview ↗

Response to Reviewer Fg4G (2/3)

W4. Independent parallel reveals

RNA-DFM balances parallel generation efficiency with conditional dependence through multi-step unmasking. Figure 2A quantifies how structural quality varies with the number of unmasking steps K. Unlike single-step generation (K = 1), multi-step generation conditions subsequent predictions on previously revealed nucleotides.

We additionally investigate whether explicitly incorporating base-pair relationships into the reveal order improves structural consistency. At K = 8, the geometrically pair-ordered strategy achieves a TM-score of 0.407, compared with 0.414 for the simple 5′-to-3′ strategy.

This suggests that enforcing geometric pairing order alone provides limited additional benefit in our setting. We acknowledge that within-step independence and irreversible commitments remain limitations. Re-masking or iterative refinement represents a potential extension beyond the current framework.

W5. Unmasking strategies

Appendix D.1 provides a controlled comparison of four unmasking strategies, holding the checkpoint, candidate budget, temperature and number of steps fixed:

OrderRecoveryTM-scorelDDTRMSD (Å)
5′-to-3′0.5880.4140.64910.45
Random0.5700.4140.61811.29
Pair-ordered0.5750.4070.64111.32
Confidence0.5760.4010.63611.20

The 5′-to-3′ strategy matches random unmasking on TM-score while improving recovery, lDDT and RMSD. The narrow TM-score range (0.401–0.414) indicates limited sensitivity to the four tested strategies, supporting the choice of 5′-to-3′ as the default.

W6. Best-of-N distillation versus reward optimisation

Our additional experiments reveal an important distinction between improving average candidate quality and improving the best candidate available for selection. This distinction motivates our selection-aware distillation objective.

We applied RIDER's RL procedure to its released weights and evaluated 53 targets in 15 clusters using Boltz-2, with USalign (-mol RNA) normalised by the native structure:

StatisticPre-RLPost-RLGain95% CIp
Min-of-160.2830.429+0.147[+0.082,+0.186]0.0006
Mean-of-160.4460.516+0.071[+0.036,+0.099]0.0066
Best-of-160.5880.599+0.011[−0.014,+0.033]0.36

RL significantly improves the minimum and mean candidate quality, whereas improvement in best-of-16 is not statistically resolved.

The difference between the mean and best-score gains is:

Δmean − Δbest = 0.0706 − 0.0113 = 0.0593.

This is equivalent to the reduction in the gap between the best and mean scores before and after RL. Thus, 84% of the mean gain reflects the narrowing mean-to-best gap rather than improvement in the best candidate itself.

The ordering Δmin > Δmean > Δbest holds on 38 of 53 targets, with a paired Wilcoxon test giving p = 1.1 × 10⁻⁸ for Δmin − Δbest.

These findings suggest that reward optimisation in this setting primarily improves the lower tail of the candidate distribution. However, practical design pipelines typically generate multiple candidates and retain the best one, making the upper-tail statistic particularly relevant.

Following BOND [7], our method incorporates this selection objective into supervised fine-tuning by weighting each candidate according to its probability of winning a best-of-N draw. Uniform weighting of the same candidate pool provides a matched control.

Our weighting improves best-of-16 TM-score by +0.0147 relative to uniform weighting, with a 95% CI of [+0.0013,+0.0281] and improvement on 60% of records. The confidence intervals exclude zero from four candidates upwards, demonstrating benefits across multiple candidate budgets.

AuthorsOpenReview ↗

Response to Reviewer Fg4G (3/3)

W7. Statistical evidence for structural quality

The revised family-weighted evaluation provides a more informative assessment of performance across structural families. Relative to NA-MPNN, RNA-DFM achieves improvements of +0.032 in TM-score, +0.026 in lDDT and +0.097 in recovery.

The per-family decomposition further reveals the consistency and heterogeneity of these improvements:

Structural familyRecordsTM-score difference
12319+0.038
1993+0.098
23752−0.069
3129+0.083
4251−0.031
5043+0.036
5891+0.073

RNA-DFM outperforms NA-MPNN on five of seven structural families. The relatively challenging cluster 237 accounts for 59.1% of the records, explaining why the comparison is sensitive to the aggregation scheme.

Importantly, the observed family-weighted improvement is not dependent on a single structural family. Excluding any one family preserves a positive TM-score difference, with leave-one-family-out gains ranging from +0.022 to +0.049.

These results demonstrate the directional robustness of the observed improvements to changes in structural family composition. Together with the statistically supported augmentation gains reported in W2, they provide complementary empirical evidence for the effectiveness of RNA-DFM, without implying statistically significant superiority over individual baselines.

W8. Generation speed and end-to-end efficiency

RNA-DFM achieves a 12.8× generation-time speedup through parallel nucleotide generation. This provides a practical latency advantage: below 3.3 seconds per target, the autoregressive sampler cannot complete a single design under the evaluated setting, whereas RNA-DFM can.

Our matched wall-clock analysis also shows comparable recovery between the two generation paradigms, with a difference of 0.004 over their shared operating range.

We agree that end-to-end efficiency must additionally account for structural prediction. Appendix F reports that folding 16 candidates with Boltz takes 204–429 seconds per target, depending on sequence length, while candidate generation accounts for only 0.2–5% of pipeline runtime for seven of the nine evaluated methods.

This motivates addressing efficiency at two complementary levels:

  • Parallel generation reduces sequence generation latency.
  • Best-of-N distillation targets candidate quality under a limited candidate budget, potentially reducing the number of expensive structural predictions required.

Figure 2B characterises structural quality as a function of candidate budget N. Since folding dominates the pipeline, improving best-of-N performance is directly relevant to reducing end-to-end computational cost.

We have moved the end-to-end runtime analysis alongside the generation speedup claim to clarify their respective practical benefits.

Summary

In response to the reviewer, we have clarified the RNA-specific methodological contributions, strengthened the augmentation analysis, expanded the discussion of unmasking strategies, provided an empirical comparison between reward optimisation and selection-aware distillation, and clarified the distinction between generation latency and end-to-end efficiency.

We believe these additional results provide stronger empirical support for RNA-DFM and a more precise understanding of its methodological contributions, practical advantages and current limitations.

Reviewer_KcAcOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

RNA-DFM is an 3d-based RNA inverse design model that features a geometric encoder to capture 3d information from RNA backbones and a discrete flow matching decoder that parallelly generates sequences. Compared to existing baselines, they outperform them on the proposed metrics.

Critical strengths
  • I like the parallel decoding strategy because of the discrete flow. I think order doesn't really matter much which order you fill the sequence, and the authors' choice of decoder reflects that.
  • I like the data augmentation strategy, but I have comments and improvements on this which I share below.
  • Good choice of baselines, covers most of the popular and well-used community models.
  • The best-of-N approach to RNA design is great, since it tries to maximise diversity while also boosting the refoldability-TMscore of the designs.
Critical weaknesses
  • I'd include Struct2Seq as a baseline given that it's a powerful model for RNA inverse design [1]. PDB data is very limited now and you're unable to get a massive boost on real-world RNA design just by using it. I understand the authors are trying to do 3D inverse folding but ultimately, if the aim is to design sequence that are viable in wet-lab experiments, you'd wnt to try out all methods that do the job. Struct2Seq scores highly during experimental validation [1]

  • Boltz is not a great RNA structure predictor on its own. I'd consider using RNAPro instead as a structure prediction model [2]. I wouldnt trust the outputs of Boltz if you're using it to create an augmented training set. The confidence head (that gives plddt) of Boltz might be very uncalibrated, so it's not a good measure of how realistic an RNA structure is. I totally get the appeal of using Protenix and Boltz though ... they can be used in single-sequence mode in a de novo design pipeline, but it defeats the purpose of having reliable 3d structure.

  • From a novelty perspective, it feels like the RNA-DFM framework is very close to gRNAde but with a different decoder, using discrete flows as opposed to an autoregressive transformer. The encoder is also GVP-based, so I'm unsure if there's any new insight being provided.

  • I'd refrain from adding an "up-arrow" for the Recovery metric. It's a bit myopic. In reality, recovery is not really a good metric for RNA inverse design since it's highly problem dependent. In reality, sometimes experimentalists know some parts of the RNA they're trying to design, or they know nothing about the design target. In such cases, recovery could be very important or not that important. Recovery and novelty are usually at odds, and is also problem dependent. I'd consider specific design targets, perhaps those offered in Eterna-100.

  • Again, regarding metrics, I'd caveat the lDDT column since Boltz pLDDTs are probably not calibrated well, given how little RNA data there is on the PDB.

[1] https://www.biorxiv.org/content/10.64898/2026.05.21.726960v1

[2] https://www.biorxiv.org/content/10.64898/2025.12.30.696949v1

Questions for authors
  1. While I understand the intention of augmenting the training set with synthetic RNA, can the authors please explain why the majority of trainign data is synthetic data (59% from RFAM), especially knowing that there is no good RNA 3d structure predictor currently.

  2. Not a question and more of a comment, but I think it'd be great to caveat the metrics based on the comments I made under "Weaknesses".

3 comments
AuthorsOpenReview ↗

Response to Reviewer KcAc (1/3)

We thank the reviewer for the constructive feedback, particularly the suggestions regarding experimentally validated inverse design methods and the reliability of RNA structure predictors. These comments motivated additional experiments that strengthen the empirical foundations of RNA-DFM.

As described in our response to all reviewers, the revised benchmark assigns equal weight to seven structural families, providing a balanced assessment of generalisation across diverse RNA structures.

W1. Inclusion of Struct2Seq

Struct2Seq is now included in the revised benchmark:

MethodRecoveryTM-scorelDDTRMSD (Å)
Struct2Seq0.3100.3310.54313.14
RNA-DFM0.5880.4140.64910.45

Struct2Seq conditions on secondary structure through dot-bracket representations rather than three-dimensional coordinates. We therefore evaluate it using its supported input representation and explicitly clarify this distinction in the manuscript.

We recognise the importance of its experimental validation in RNA design. Our computational comparison complements this evidence by evaluating the ability of different methods to recover specified three-dimensional structures.

W2. Structure predictor reliability and RNAPro validation

Following the reviewer's suggestion, we conducted additional experiments with RNAPro to evaluate the robustness of our conclusions to the choice of structure predictor.

First, we folded the native sequences of all 88 clean test records using three predictors in single sequence mode:

PredictorFamily average TM-scoreRecord average TM-score
Boltz0.6810.809
Protenix0.7030.826
RNAPro0.7550.873

RNAPro indeed achieves higher accuracy on native sequences. To reproduce a practical de novo design setting, RNAPro was executed without MSA or template inputs, with RibonanzaNet2 enabled. We verified this configuration from the inference logs.

Independent evaluation with RNAPro

Our primary evaluation separates candidate selection from structural assessment: Boltz selects candidates, while Protenix independently evaluates their structures.

To further assess robustness, we conducted an additional evaluation using RNAPro and compared the resulting rankings with our main benchmark:

MethodProtenixRNAProRank: Boltz / Protenix / RNAPro
RNA-DFM0.4140.37061 / 1 / 1
gRNAde0.3930.33962 / 3 / 2
NA-MPNN0.3830.33893 / 4 / 3
RiboPO0.3950.32564 / 2 / 4
w/o Rfam0.3340.30167 / 5 / 5
RiboDiffusion0.3280.30006 / 8 / 6
RIDE0.3330.29805 / 6 / 7
Struct2Seq0.3310.28308 / 7 / 8
RhoDesign0.2550.20989 / 9 / 9

RNA-DFM ranks first under all three structure predictors. The Spearman correlation between the Protenix and RNAPro rankings is 0.90, demonstrating substantial agreement despite differences in their absolute structural scores.

RNAPro therefore provides independent supporting evidence for the performance of RNA-DFM.

Validation of augmentation without Boltz in evaluation

We further investigated whether the observed augmentation benefit depends on using Boltz for candidate selection, given that the training augmentation also uses Boltz predictions.

We folded all 3,417 unique candidate pairs from the augmented and nonaugmented models using RNAPro. RNAPro then selected the best candidate for each record, and Protenix independently evaluated the selected designs. Thus, Boltz was entirely removed from the evaluation process.

Evaluation protocolRfam TM-score gain95% CI
Boltz selects, Protenix scores+0.0788[+0.0564,+0.0998]
RNAPro selects, Protenix scores+0.0639[+0.0176,+0.1089]

The augmentation benefit remains statistically supported without Boltz participating in either selection or evaluation. Its direction is also consistent under all four aggregation schemes.

Importantly, RNAPro and Boltz select the same candidate in only 70 of 352 decisions (19.9%). Despite substantial differences in candidate selection, the augmentation benefit persists.

These results demonstrate that the benefits of Rfam augmentation persist even when Boltz is entirely removed from the evaluation process.

AuthorsOpenReview ↗

Response to Reviewer KcAc (2/3)

W3. Methodological contributions beyond gRNAde

RNA-DFM shares the geometric inductive bias of GVP based encoders with gRNAde, while introducing a fundamentally different generation process through masked discrete flow matching.

Retaining an established geometric encoder allows us to focus on the contribution of the generation mechanism. The revised experiments establish several practical properties of this formulation.

1. Flexible generation schedules. The number of sampling steps can be selected at inference time using the same trained model, independently of sequence length. Increasing the number of steps from one to eight improves TM-score from 0.348 to 0.414, demonstrating the contribution of progressive conditional generation beyond single step parallel decoding.

RNA-DFM also supports alternative reveal orders without retraining, including random, confidence based, geometric pairing based and 5′-to-3′ generation.

2. Efficiency under limited candidate budgets. Structural prediction dominates the computational cost of practical design pipelines. Folding 16 candidates requires 204–429 seconds, compared with 2.55 seconds for generating them.

RNA-DFM reaches a TM-score of 0.395 with only two candidates, matching the score achieved by the strongest baseline using 16 candidates. This corresponds to an eightfold reduction in the number of structural predictions at comparable observed structural quality.

3. Selection aware distillation. Our best-of-N distillation incorporates structural selection into the generator's training objective. It improves best-of-16 TM-score by 0.015 over a matched control trained on identical candidate pools with uniform label weights, isolating the contribution of the selection weighting.

4. Learning under limited structural supervision. Masking enables multiple supervision patterns from each RNA backbone, complementing our large scale structural augmentation in a setting where experimental structures remain scarce.

The revised Related Work section clarifies these contributions in relation to gRNAde, CoFlow, ADFLIP and RIdiffusion. Our contribution is therefore not a new geometric encoder, but an RNA inverse design framework that combines flexible generation, structural augmentation and selection aware optimisation.

W4. Interpretation of sequence recovery

We agree that native sequence recovery should not be interpreted as a universal measure of RNA design success. Its relevance depends on the design objective, the available sequence constraints and the desired degree of sequence novelty.

Our manuscript includes a controlled perturbation experiment that directly demonstrates this distinction. Restoring mismatched paired nucleotides to their native identities increases recovery from 0.516 to 0.761, while ViennaRNA base pair F1 decreases by 0.274. This shows that increasing agreement with the native sequence does not necessarily improve structural consistency.

We also find that selecting candidates according to recovery provides a weaker signal for structural quality than selection using predicted structures.

Nevertheless, recovery remains a useful complementary metric because it measures native sequence agreement independently of any structure predictor. We clarify that its interpretation is descriptive rather than a universal optimisation objective.

Regarding Eterna-100, it addresses a distinct design task. Eterna-100 specifies secondary structures in dot-bracket notation and evaluates designs using secondary structure folding algorithms such as ViennaRNA or EternaFold. In contrast, RNA-DFM conditions on a specified three-dimensional backbone and evaluates structural agreement using TM-score, lDDT and RMSD.

Applying RNA-DFM to Eterna-100 would require an additional stage to construct three-dimensional backbones from secondary structure targets. The resulting performance would therefore reflect both backbone construction and inverse folding, rather than isolating the capability of RNA-DFM. We consequently focus on benchmarks with experimentally determined three-dimensional target structures, consistent with the intended task.

W5. Distinction between lDDT and pLDDT

The lDDT reported in our manuscript is not Boltz pLDDT or a predictor confidence score.

It is a geometric lDDT computed using C4′ atom distances between the Protenix predicted structure of the designed sequence and the experimental target structure, with a 15 Å inclusion radius and distance tolerances of 0.5, 1, 2 and 4 Å.

Consequently, the calibration of Boltz pLDDT does not directly affect the reported lDDT metric. Boltz confidence is used only as one filtering criterion in the structural augmentation pipeline.

AuthorsOpenReview ↗

Response to Reviewer KcAc (3/3)

Q1. Why does predicted structural data account for 59% of the training dataset?

The proportion results from the limited availability of experimental structures and the construction of the filtered augmentation dataset, rather than optimisation of a mixing ratio.

Following dataset splitting and filtering, we retain 4,025 experimental structure records. To expand structural supervision, we construct the augmentation dataset using the following procedure:

Filtering stageRecords retained
Available single chain predictions21,291
Confidence filtering13,341
Selection for sequence coverage6,380
Structural decontamination5,850

The final training dataset therefore comprises 4,025 experimental records and 5,850 augmented records, with the latter accounting for 59.2%.

Importantly, the augmented sequences are genuine Rfam sequences, rather than synthetically generated sequences. Only their associated three-dimensional conditioning structures are predicted.

We specifically evaluate whether these filtered predictions provide useful additional supervision through a matched ablation:

MethodRecoveryTM-scorelDDTRMSD (Å)
RNA-DFM w/o Rfam0.5190.3340.57112.24
RNA-DFM0.5880.4140.64910.45

The improvements in recovery, TM-score and lDDT are statistically supported, with confidence intervals excluding zero. All improvements remain directionally consistent across the four aggregation schemes.

Moreover, as demonstrated in W2, the augmentation gain remains statistically supported when RNAPro replaces Boltz for candidate selection and Protenix independently evaluates the selected structures.

These results suggest that predicted structures do not need to match experimental accuracy to provide useful additional supervision. Rather than assuming that every predicted structure is accurate, we directly evaluate the effectiveness of filtered Rfam augmentation through controlled ablations and independent structural assessment.

Q2. Interpretation of evaluation metrics

We have clarified the interpretation of the reported metrics and evaluation protocol:

  • Recovery measures native sequence agreement and is not a universal indicator of design success.
  • lDDT measures geometric agreement with the experimental target rather than the confidence of a structure predictor.
  • TM-score and RMSD assess structural agreement between predicted designs and experimental targets.
  • Independent structure predictors are used for candidate selection and subsequent evaluation.

The additional RNAPro experiments further assess the dependence of our conclusions on the choice of structure predictor.

Summary

In response to the reviewer, we have expanded the benchmark to include Struct2Seq, conducted additional validation using RNAPro, verified the augmentation benefits without Boltz participating in evaluation, clarified the methodological contributions beyond gRNAde, and refined the interpretation of sequence recovery and structural metrics.

Importantly, RNA-DFM ranks first under all three structure predictors, and Rfam augmentation retains a statistically supported TM-score improvement of +0.0639 when Boltz is entirely removed from the evaluation process. These additional experiments strengthen the evidence for both the effectiveness of RNA-DFM and the robustness of its evaluation.

We thank the reviewer for the valuable suggestions that have substantially strengthened the experimental analysis and methodological presentation of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We sincerely thank all three reviewers for their constructive and detailed feedback. Their comments have helped us strengthen the evaluation protocol, expand baseline comparisons, clarify our methodological contributions, and conduct additional controlled experiments.

Before addressing the individual reviews, we would like to clarify an important change in the evaluation protocol that affects the reported benchmark results.

Revised aggregation protocol

The submitted version aggregated the 88 clean test records over BGSU nonredundant equivalence classes. The revised evaluation assigns equal weight to the seven structural families (1/7 each).

The test set (88 records, 53 unique sequences and 15 sequence clusters), candidate budget (16), and selection criterion (highest Boltz TM-score), independent structural judge (Protenix) remain unchanged.

Motivation

Both aggregation schemes are valid but measure different aspects of model performance. However, given the substantial imbalance in structural family sizes, equal weighting better reflects the ability of a model to generalise across diverse RNA structures, rather than primarily reflecting its performance on the most represented family.

Specifically, the purine riboswitch family (structural cluster 237) contains 52 of the 88 records and 13 of the 22 equivalence classes, receiving 59.1% of the total weight under the original aggregation. Equal weighting reduces its contribution to 14.3%, ensuring that each structural family contributes equally to the overall benchmark. Appendix Table 7 reports the family weights under alternative aggregation schemes.

Impact on RNA-DFM performance

Metric (best of 16)SubmittedRevised
TM-score ↑0.4450.414
lDDT ↑0.6190.649
RMSD (Å) ↓9.38010.45
Recovery ↑55.9%58.8%

The revised aggregation changes the relative contributions of structurally distinct targets, resulting in lower absolute TM-score and higher RMSD, alongside improvements in lDDT and recovery.

Impact on comparative performance

Reviewers Fg4G and nSuD both referred to the TM-score difference between RNA-DFM and NA-MPNN of 0.004 (0.445 versus 0.441). Under equal weighting of structural families, this difference is 0.031 (0.414 versus 0.383).

RNA-DFM achieves the highest recovery, TM-score and lDDT, and the lowest RMSD among the nine evaluated methods, both when averaging over 16 candidates and when selecting the best candidate.

Furthermore, the matched Rfam augmentation experiment demonstrates statistically supported improvements, with consistent gains across all four aggregation schemes.

We report the revised evaluation throughout the manuscript and the individual responses to ensure consistency and transparency. We believe the expanded benchmark and additional analyses provide a more comprehensive assessment of RNA-DFM across diverse RNA structural families.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.