acceptodds
Under review as a conference paper at ICLR 2027

KnitNet: Learning to Search at Inference Time for Exact RNA Inverse Design

Abstract

RNA inverse design seeks nucleotide sequences that fold into specified target structures. Existing methods rely on learned generators that can miss exact structural matches, especially beyond their training distributions, or on costly search over a vast, multimodal sequence space. Search is further complicated by globally coupled nucleotide interactions, as correcting one pairing error can create new errors beyond a fixed repair scope. To address this, we introduce KnitNet, an inference-time scaling framework that learns from verifier feedback to refine sequences for exact fold-back matching. Starting from candidates generated for a target secondary structure, KnitNet uses mismatched folds to infer nucleotide dependencies and adapt its joint repair strategy, while an online surrogate learns to predict which edits are promising. This makes search progressively more informed and compute-efficient. Crucially, KnitNet generalizes across RNA structural regimes by changing only the optimization signal to match the verifier required by each setting. Across five secondary-structure benchmarks, KnitNet increases exact solve rates by an average of 21 percentage points over continued sampling with 70% less time-to-solve. On OpenKnot57 specifically, KnitNet is the first method to achieve exact computational matches for all complex, pseudoknotted structures. Beyond 2D, KnitNet improves 3D structural self-consistency over per-target RL by 7.8% while being 39.3% faster, demonstrating generalization and effective inference-time scaling across RNA design regimes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA Inverse Design, Inference-Time Scaling, Learned Search, Counterexample-Guided Repair, Surrogate Modeling, Combinatorial Optimization, Pseudoknots

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7y9JOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes KnitNet, an RNA inverse folding design framework based on adaptive search during inference. The method constructs nucleotide dependencies through structural error feedback, dynamically adjusts the scope of joint repair, and utilizes an online proxy model to improve search efficiency. On five RNA secondary structure benchmarks, the accuracy rate improved by an average of approximately 21 percentage points, and computational structure matching was achieved for all targets on OpenKnot57.

Critical strengths
  1. Highly innovative. By combining error feedback, dependency modeling, and online learning, the method achieves adaptive joint correction during the inference phase, distinguishing it from traditional fixed-scope search methods.

  2. The experiments are comprehensive, and the results are outstanding. The study covers common secondary structures, false-knot structures, and three-dimensional designs and validates the method’s effectiveness through cross-generator experiments and ablation studies.

Critical weaknesses
  1. Lack of independent biological validation. The results rely primarily on computational structure predictors, and it has not yet been demonstrated that the designed sequences can form the expected structures in actual experiments.

  2. The advantages are not evident on some benchmarks. On traditional inverse-folding benchmarks such as Eterna100, KnitNet failed to outperform specially designed search algorithms like NEMO, indicating that its advantages are task-dependent to some extent.

Questions for authors
  1. If validation is performed using a structure predictor independent of the search process, does KnitNet still demonstrate a performance advantage?
  2. Why does KnitNet underperform NEMO on some traditional RNA reverse-folding tasks?
  3. As RNA sequence length increases further, how do the computational cost and success rate of online searches change?
5 comments

We thank the reviewer for their thoughtful and encouraging assessment. We appreciate their recognition of KnitNet as “highly innovative”, particularly its combination of error feedback, dependency modeling, and online learning to enable adaptive joint repair. We are also pleased that the reviewer finds that “the experiments are comprehensive, and the results are outstanding”, highlighting our evaluation across secondary-structure and 3D RNA design, together with the cross-generator comparisons and ablation studies.


1. Response to Weakness 1

While we agree that experimental validation is important, given ICLR’s focus on computational methodology, we believe our contribution should be assessed primarily on its methodological advances and supporting computational evidence. Our extensive evaluations across pseudoknotted, planar, and 3D RNA design settings support KnitNet’s generality, as the reviewer also recognizes among the paper’s strengths.

Nevertheless, we are already pursuing experimental validation through the EternaBind challenge, where our designs are undergoing wet-lab testing for binding to target ligands, extending KnitNet toward aptamer design. We intend to incorporate these experimental results into the final version of the paper once we receive them and they are cleared for public release


2. Response to Weakness 2 and Question 2

We respectfully disagree that KnitNet advantages are task-dependent, and in fact, through our results, we demonstrate benefits across planar, pseudoknotted, and 3D inverse-design setting, as the reviewer also notes “the experiments are comprehensive, and the results are outstanding,” (Tables 1, 2, 4). The Eterna100-v2 result reflects the expected behavior of a solver explicitly engineered around that benchmark's objective, rather than an absence of advantages for KnitNet across RNA inverse-design settings.

NEMO's implementation tightly incorporates hand-crafted structural rules and domain knowledge derived from the author's Eterna experience (Portela, 2018). The NEMO paper itself demonstrates the importance of these heuristic choices, which, when ablated, drop performance by 16% and slow convergence by 15–50%. This explains why NEMO is particularly well suited to Eterna100, while also limiting its scope to only planar structures and no pseudoknots.

In contrast, while maintaining generality, KnitNet reaches 97/100 solves on Eterna100-v2 when we let it run past the 60-minute initial budget, with the last 3 targets being computationally unsolvable under current Eterna settings. This demonstrates the initial results were not a hard performance ceiling.

Across the 15 additional solves, winning lineages required a median of 60,144 oracle candidate evaluations, which is relatively short compared with the total portfolio search. The principal difficulty, therefore, was finding a productive starting sequence, and KnitNet benefitted from switching basins after apparent stagnation, a mechanism we outline in Figure 4. Inspired by NEMO's rapid turnover between unsuccessful trajectories, which also contributes to its performance, a future version of KnitNet could investigate incorporating a stagnation budget and restart rule to boost performance further.

3. Response to Question 1

To demonstrate KnitNet’s performance advantage extends beyond its search objective, we evaluated frozen Eterna100 designs using a panel of five secondary-structure predictors. KnitNet’s designs were generated and selected using ViennaRNA before panel scoring, with no reselection based on the panel results. KnitNet’s extended campaign ranks first across all three panel-averaged metrics, achieving an exact-solve rate of 32.27%, mean Jaccard of 0.7054, and mean MCC of 0.7798, compared with 31.75%, 0.6846, and 0.7616 for DesiRNA, the strongest baseline in this comparison.

Frozen output pipelinePanel exact-solve rate ↑Panel mean Jaccard ↑Panel mean MCC ↑
KnitNet extended campaign32.27%0.70540.7798
DesiRNA off-the-shelf31.75%0.68460.7616
KnitNet Generator → KnitNet search25.60%0.67000.7517
KnitNet Generator → continued KnitNet sampling22.53%0.66080.7487
Struct2Seq → KnitNet search22.80%0.66030.7422
KnitNet init → DesiRNA search25.00%0.65690.7397
Struct2Seq → continued released-model sampling19.80%0.65520.7443
KnitNet roots → NEMO25.13%0.65100.7334

Exact-solve rates are panel averages across NUPACK, MXFold2, RNAStructure, EternaFold and SPOT-RNA.

Despite known limitations in cross-oracle agreement and predictor-based evaluation (ACCORD, gRNAde), KnitNet retains its competitive advantage across all three panel-averaged metrics.


4. Response to Question 3

To answer this question we studied scaling on our internal test set of 500 RNA targets spanning 20–511 nt. For every target, we generated 96 initial sequences and then ran KnitNet for up to one hour, as well as Struct2SeQ-OA, as the strongest baseline on this dataset. We grouped targets by length and measured both exact recovery and realized folding-oracle throughput:

Target lengthKnitNet exact solves after 1 hStruct2SeQ-OA after 1 h
≤100197/200 (98.5%)176/200 (88.0%)
101–20053/61 (86.9%)24/61 (39.3%)
201–30048/112 (42.9%)6/112 (5.4%)
301–4006/71 (8.5%)2/71 (2.8%)
401–5112/56 (3.6%)1/56 (1.8%)
Overall306/500 (61.2%)209/500 (41.8%)

KnitNet outperforms the strongest baseline across all evaluated length groups, although exact recovery and folding throughput decline with length: throughput falls from approximately 37 evaluations/s for targets ≤100 nt to 2 evaluations/s for targets of 401–511 nt, reducing the candidates explored within one hour. Additional compute still helps: extending search from one to two hours increased exact solves from 8/127 (6.3%) to 15/127 (11.8%) among targets of 301–511 nt. On ten held-out targets of 759–956 nt, one-hour search achieved no exact solves but improved mean best pair Jaccard from 0.668 to 0.764 (+14.4%), versus 0.718 for matched continued sampling. These results show that KnitNet improves approximate recovery even on much longer RNAs, while exact recovery remains challenging. Alongside higher oracle costs and a larger search space, weaker guidance may contribute: Chen et al. (2025) report declining accuracy with length for nearly all tested secondary-structure predictors, with severe degradation above 500 nt (https://doi.org/10.1088/1674-1056/adea9c).

We thank the reviewer again for their encouraging assessment and constructive suggestions. We hope the additional cross-predictor evaluation, length-scaling analyses, and clarification of KnitNet’s performance on Eterna100-v2, address their remaining questions. If these additions resolve their concerns, we would be grateful if they would consider raising their score.

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

KnitNet combines a structure-conditioned diffusion generator with inference-time search for RNA inverse design. Its search constructs joint repair scopes from folding errors, ranks proposed edits using an online surrogate, and updates repair priorities based on subsequent feedback.

The method achieves good computational results on pseudoknotted targets and improves several existing generators, but the evidence for its methodological novelty, self evaluation, individual components remains insufficient.

Critical strengths
  1. Results on pseudoknotted benchmarks are good, including 57/57 computational matches on OpenKnot57 in individual runs.
  2. Comparisons against continued sampling use identical initial candidate pools and matched post-initialization wall-clock budgets.
  3. The appendices provide detailed implementation descriptions, data-overlap checks, and additional baseline comparisons.
Critical weaknesses
  1. The conceptual novelty appears limited. The contribution is primarily integration of prior components.
  2. Evaluation relies heavily on the predictors used during optimization. Without independent-predictor validation or experimental evidence, it is difficult to assess whether improvements reflect robust structural design or adaptation to particular computational oracles.
  3. The component ablations provide limited support for the central mechanisms: removing surrogate ranking reduces coverage from 306/500 to 297/500, while fixing the repair problem yields 301/500.
  4. The advantage over strong baselines narrows substantially when search guidance is aligned. On PseudoBase++, RNSS-adapted DesiRNA reaches 334.7/339 versus KnitNet’s 335.7/339. KnitNet also trails specialized methods on Eterna100-v2.
Questions for authors
  1. Can the strongest RNSS-adapted baselines be included in the main table, with statistical comparisons against KnitNet?
  2. How well do optimized sequences perform under an independent folding predictor that was not used for labeling, search, or candidate selection?
4 comments

We thank the reviewer for their careful assessment and constructive questions. We address each concern below.

1. Response to Weakness 1

Our conceptual novelty comes from the fundamental paradigm shift we introduce in RNA inverse design: we are the first to combine learned biological priors with an online, adaptive repair method we develop, thereby achieving a highly generalizable algorithm. The result is a reusable, memory-informed repair formulation that seamlessly adapts to any RNA design setting without depending on specific outputs from a limited number of oracles, which is a unique capability of our method. KnitNet encapsulates and extends beyond all the capabilities of prior works, which would otherwise be limited to learned generation (Struct2SeQ-OA), costly search (DesiRNA), or could only design planar targets (SentRNA, RIFT-VAE). By doing so, we also demonstrate empirical gains of having such a generalizable method by being the first to solve all targets in OpenKnot57, as well as reach 97/100 solves on Eterna100-v2 (extended campaign described in Section 4 of this rebuttal), the best computational results reported.

Furthermore, our algorithm specifically tackles a central conceptual limitation of prior approaches, that of fixing one mismatch creating another conflict outside the original repair region. Our method addresses this by converting foldback information into direct guidance for how we should repair the sequence, and our dynamic error graph adapts where and what to mutate. This solves the issue of prior methods (RNAblueprint, DRAG) having a static graph that cannot capture the evolving interdependencies between nucleotides or not being able to design pseudoknots, respectively. Another reason we designed the graph to retain search history was to learn from previous rounds which mutations are favorable in new repair rounds, thereby saving on repeated and expensive oracle calls and making our method much more computationally efficient, as shown by the 2.1-fold speedup in time-to-solve relative to the strongest baseline on OpenKnot57 (Table 9).


2. Response to Weakness 2 and Question 2

We agree that evaluation beyond the optimization oracle is important. We therefore evaluated our frozen Eterna100 designs using five secondary-structure predictors that were not used for training-label generation, search, or candidate selection. KnitNet’s designs were generated and selected using ViennaRNA before panel evaluation, without subsequent retraining or reselection.

KnitNet’s extended campaign (see Section 5 of our rebuttal) ranks first across all three panel-averaged metrics. Despite known limitations in cross-oracle agreement and predictor-based evaluation (ACCORD, gRNAde), KnitNet retains its competitive advantage across all three panel-averaged metrics.

Frozen output pipelinePanel exact-solve rate ↑Panel mean Jaccard ↑Panel mean MCC ↑
KnitNet extended campaign32.27%0.70540.7798
DesiRNA off-the-shelf31.75%0.68460.7616
KnitNet Generator → KnitNet search25.60%0.67000.7517
KnitNet Generator → continued KnitNet sampling22.53%0.66080.7487
Struct2SeQ → KnitNet search22.80%0.66030.7422
KnitNet init → DesiRNA search25.00%0.65690.7397
Struct2SeQ → continued released-model sampling19.80%0.65520.7443
KnitNet roots → NEMO25.13%0.65100.7334

We are also pursuing experimental validation through the EternaBind challenge, where our designs are undergoing wet-lab testing for ligand binding. We intend to incorporate these results into the final paper once received and cleared for public release.

3. Response to Weakness 3

To address the reviewer’s concern, we conducted follow-up component ablations on OpenKnot57 and Eterna100-v2, using a matched budget of 15,000 logical oracle evaluations to assess both final coverage and efficiency throughout search. These experiments strengthen the evidence that adaptive repair and learned ranking make complementary contributions to KnitNet’s performance.

MethodEterna100-v2 Exact (/100) / AUCFULL ΔAUC 95% CIOpenKnot57 Exact (/57) / AUCFULL ΔAUC 95% CI
FULL–KnitNet85 / 0.768—57 / 0.963—
NO-SURROGATE82 / 0.743+0.025 [0.006, 0.045]56 / 0.948+0.015 [0.004, 0.029]
FIXED-PROBLEM72 / 0.701+0.067 [0.032, 0.105]55 / 0.936+0.027 [0.009, 0.050]
RANDOM-SCOPE77 / 0.714+0.054 [0.020, 0.089]54 / 0.928+0.035 [0.013, 0.061]
LINEAR-SURROGATE81/ / 0.736+0.032 [0.009, 0.057]56 / 0.951+0.012 [0.002, 0.026]
FIXED + NO-SURROGATE72 / 0.688+0.080 [0.044, 0.121]55 / 0.930+0.033 [0.011, 0.057]

The ablations show why all components matter: adapting the repair problem unlocks additional solutions (13 additional Eterna100-v2 targets), while surrogate ranking improves oracle efficiency on both benchmarks, with paired AUC gains of 0.025 and 0.015. The forward audit supports this interpretation: concordance of 0.631/0.624 and positive top-k enrichment show that the surrogate identifies promising edits before evaluating them. These follow-up experiments use 64 ExtraTrees instead of 16, with the rest of the algorithm unchanged. This hyperparameter adjustment improved Eterna100-v2 coverage from 81 to 85/100, while retaining 57/57 on OpenKnot57. The table reports component contributions within this updated configuration.


4. Response to Weakness 4 (Eterna100-v2 performance)

With additional test-time compute on the remaining unsolved Eterna100-v2 targets, going beyond the initial 60 minute/target budget, KnitNet reaches 97/100 solves overall, while only requiring a median of 60,000 additional oracle calls. This follow-up result highlights KnitNet can unlock state-of-the-art performance across varied benchmarks, underlining the advantage of our adaptive algorithm.

5. Response to Weakness 4 and Question 1 (KnitNet outperforms RNSS-adapted baselines)

We will revise the main table to include both DesiRNA–RNSS (our RNSS-aligned adaptation of the published DesiRNA) variants and paired statistical comparisons against KnitNet.

We want to clarify that the apparent proximity on PseudoBase++ is primarily a ceiling effect: 80.7% of targets were already solved at initialization. KnitNet nevertheless ranks first in exact recovery, solve AUC, and capped time to solve (TTS). On less staturated benchmarks, like OpenKnot57 and our 500-target test set, KnitNet outperformed RNSS-adapted DesiRNA and demonstrated a statistically signifcant performance gap.

Moreover, across the three benchmarks KnitNet achieves a 1.07-3.14x speedup and requirs 2.35-11.6x fewer oracle calls for solving targets. KnitNet especially solves pseudoknots faster than DesiRNA because it overcomes a methodological limitation, which was also reported by the authors: their pseudoknot-capable implementation was approximately 100 times slower than when run on planar targets, and also limited to a specific oracle's reward output. In contrast, KnitNet can be adapted to any oracle and be deployed across any design setting. Our higher throughput is experimentally consequential: within a fixed compute budget, it enables the generation of larger and more diverse candidate libraries for downstream synthesis and wet-lab screening– a central requirement of practical RNA design campaigns.

PanelMethodExactSolve AUCCapped TTS (min)TTS ratioΔ exact (95% CI)Δ AUC (95% CI)Δ TTS, min (95% CI)
OpenKnot57KnitNet56.67/570.974240.7731.00×———
OpenKnot57DesiRNA–RNSS, KnitNet roots55.00/570.930602.0822.69×+1.67 [0.33, 3.67]+0.0436 [0.0072, 0.0914]−1.309 [−2.742, −0.216]
OpenKnot57DesiRNA–RNSS, native initialization54.33/570.919212.4243.14×+2.33 [0.33, 5.00]+0.0550 [0.0151, 0.1055]−1.651 [−3.164, −0.453]
PseudoBase++-339KnitNet335.67/3390.988210.3541.00×———
PseudoBase++-339DesiRNA–RNSS, KnitNet roots335.00/3390.985340.4401.24×+0.67 [−0.67, 2.33]+0.0029 [−0.0013, 0.0081]−0.086 [−0.242, 0.038]
PseudoBase++-339DesiRNA–RNSS, native initialization334.67/3390.980460.5861.66×+1.00 [−0.67, 3.67]+0.0077 [0.0013, 0.0159]−0.232 [−0.476, −0.039]
500-target test setKnitNet275/5000.5257214.2281.00×———
500-target test setDesiRNA–RNSS, KnitNet roots260/5000.4931615.2051.07×+15 [4, 27]; p=0.0167+0.03256 [0.01347, 0.05225]−0.977 [−1.567, −0.404]
500-target test setDesiRNA–RNSS, native initialization255/5000.4830815.5071.09×+20 [7, 34]; p=0.0055+0.04264 [0.02057, 0.06469]−1.279 [−1.941, −0.617]
500-target test setOff-the-shelf DesiRNA189/5000.3630219.1091.34×+86 [70, 103]; p=5.75×10⁻²⁵+0.16270 [0.13330, 0.19358]−4.881 [−5.807, −3.999]

We thank the reviewer again for their constructive feedback, which helped us clarify KnitNet’s contribution and strengthen the supporting evidence. We hope the additional analyses and comparisons address their concerns. If these additions resolve their reservations, we would be grateful if they would consider raising their score.

Dear Reviewer zefs,

We would be grateful if you could let us know, while the discussion period is still open, whether our rebuttal and additional analyses have satisfactorily addressed your concerns.

Our response clarifies KnitNet’s methodological novelty (Section 1) and provides further evidence of its strong performance through independent-predictor evaluation (Section 2), expanded component ablations (Section 3), and comparisons with strong baselines (Sections 4–5).

Thank you again for your time and thoughtful feedback.

Best regards,
The authors

Reviewer_mC98OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

In this comprehensive study, the authors are trying to explore the potential of 'search' in the field of RNA inverse design. A promising framework called KnitNet has been proposed. Besides the pipeline itself, various types of experiments, across 2D and 3D structural alignment on different public benchmarks (e.g, OpenKnot57, Order-Agnostic Struct2SeQ), have been conducted. As a potential publication in the field of (RNA) biology, several visualizations focusing on the bio-dirversities of the generated sequences have also been posted (i.e., Appendix C), which is pretty amazing, compared with most of the existing preprints / submissions. In general, it's a good work which deserves to be accepted, after addressing several minor concerns listed below.

Critical strengths

Please refer back to "Paper Summary".

Critical weaknesses
  1. [Minor] Performances under different MMseqs2 identities: Would the authors please consider adding some follow-up experiments to test the performances of KnitNet under different MMseqs2 identities, except for 80% (i.e., line 994)?
  2. [Minor] RNAs shorter than 12 nt: I notice that such shorter RNAs were not included in the existing benchmarks (i.e., Table 6). Thus, the authors are highly recommended to test KnitNet on such shorter entries and I'm really looking forward to see follow-up results (if any) during the rebuttal period.
Questions for authors

Please refer back to "Critical Weaknesses".

5 comments

We thank the reviewer for their thoughtful and encouraging assessment of our work. In particular, we appreciate the reviewer's recognition of KnitNet's promise and the breadth of our experiments, spanning 2D pseudoknotted, 2D thermodynamic, and 3D RNA inverse design. These evaluations demonstrate how our method can improve design across different structural representations and benchmarks.

“several visualizations focusing on the bio-dirversities of the generated sequences” … “which is pretty amazing”

We also appreciate the reviewer's positive assessment of our biological analysis (e.g., Appendix C). We devoted particular attention to characterizing the diversity and structural properties of the designed RNAs. These computational analyses provide encouraging evidence of biological plausibility and motivate future wet-lab evaluation.

We address the two requested extensions below.

1. Response to Question 1

We thank the reviewer for this suggestion. We chose the 80% identity cutoff following established practice in RNA benchmarking: SPOT-RNA uses an 80% cutoff for sequence nonredundancy filtering, and RNAFlow uses an 80% cutoff for its sequence-similarity split (Singh et al., 2019; Nori et al., 2024). Nevertheless, we agree that evaluating stricter thresholds provides useful additional evidence.

Starting from the same 500-target ablation panel used in the paper, we constructed additional nested evaluation subsets excluding references with detected training matches at ≥70% or ≥60% identity, while retaining the original ≥80% coverage requirement for both sequences. We kept the trained checkpoint and evaluation budget fixed, reusing the original scored generator candidates per target.

Evaluation setExact recoveryExact rateMean best Jaccard
Full ablation panel139/50027.8%0.8102
T7099/33030.0%0.7949
T6099/32930.1%0.7944

The observed exact-recovery rate does not decrease on the stricter subsets, while mean best Jaccard decreases modestly. These results support the generator's ability to design targets whose reference sequences have no detected close training match.

Beyond this internal sensitivity analysis, we audited the available reference sequences of our external 2D benchmarks against the same training corpus, retaining the bilateral 80% coverage requirement. The completed checks show no detected matches at or above 80% identity for the references summarized below, and most have no detected match at or above 60%.

BenchmarkHighest detected identityMean best detected identity*No detected hit at ≥60%
OpenKnot5779.61%78.38%55/57
PseudoBase++79.81%78.32%314/339
Eterna100-v275.00%71.38%96/100
Rfam-Learn-Test79.95%78.83%17/100
Rfam-Taneda79.94%78.78%4/29

Together, the observed exact-recovery rates on stricter internal subsets and the limited detected sequence overlap in the completed external checks provide additional support for generalization beyond close sequence matches.

2. Response to Question 2

We thank the reviewer for this constructive suggestion, which motivated us to extend the evaluated length range. Our PDB and literature search identified 3 unique and designable RNA structures below 12 nucleotides. We therefore constructed a panel of 25 distinct targets of 8–11 nucleotides: the three supported experimental motifs (an 8-nt UUCG hairpin and PDB entries 1IDV and 2MXJ) and 22 synthetic structures obtained by folding biologically-feasible sequences with ViennaRNA. KnitNet received only the target structures, not their source sequences.

We generated 96 candidates per target (seed 42). Under each folding evaluator, a target was considered solved if at least one candidate exactly reproduced the input base-pair set.

MethodViennaRNARibonanzaNet-SS
KnitNet (96 candidates)23/2525/25
KnitNet + search when needed25/2525/25

All targets were solved directly under RibonanzaNet-SS. Vienna-guided search resolved the remaining two ViennaRNA. These results demonstrate KnitNet's ability to design the tested 8–11-nt RNAs under both evaluators.

We sincerely thank the reviewer again for their encouraging assessment and thoughtful feedback. We hope that the clarifications and additional results address their questions, and would welcome any remaining concerns or suggestions for further analysis. If these additions satisfactorily address their concerns, we would be grateful if the reviewer would consider improving their assessment.

Reviewer_mC98OpenReview ↗

Re: Authors' Responses

Dear Authors,

I would love to say thank you for the detailed responses towards my concerns.

I hereby confirm you all that I WILL read each and every word of yours while considering my final recommendation.

Please feel free to post some more supporting information (if applicable) during the rebuttal period.

Reviewer mC98

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for recognizing KnitNet’s contributions as follows: reviewer 7y9J described the method as “Highly innovative” and noted that “The experiments are comprehensive, and the results are outstanding.” Reviewer mC98 called it “a good work which deserves to be accepted,” while reviewer zefs highlighted the strong pseudoknot results and matched-budget comparisons that clearly distinguish the performance boost obtained through KnitNet. We appreciate these assessments and have uploaded a revised manuscript (changes highlighted in violet) addressing the additional questions:

  • Independent-oracle evaluation (reviewers zefs and 7y9J). Reviewers zefs and 7y9J asked whether KnitNet’s advantage extends beyond the search predictor used for optimization. We show that KnitNet maintains its competitive advantage when evaluated with five independent oracles, exceeding the results of the strongest baseline by 3.0% in Jaccard and 2.4% in MCC.

  • Component ablations (reviewer zefs). Reviewer zefs questioned the contributions of individual components. Additional ablations on OpenKnot57 and Eterna100-v2 show that adapting the repair scope and error priorities enables 13 additional Eterna100-v2 solves (85 versus 72). Removing either adaptive repair or learned ranking significantly reduces solve AUC on both benchmarks, showing that both components improve the use of a fixed oracle-evaluation budget. All paired 95% confidence intervals for these AUC differences exclude zero.

  • Aligned baselines and statistical comparisons (reviewer zefs). Reviewer zefs requested stronger aligned baselines and statistical comparisons. With identical initialization and structural-margin guidance, KnitNet gains 1.67 OpenKnot57 solves over the strongest competing search method (95% CI: [0.33, 3.67]) and reduces capped mean time to solve from 2.08 to 0.77 minutes, maintaining our lead even when adapting existing methods to our optimization objective.

  • Sequence-length scaling (reviewer 7y9J). Reviewer 7y9J asked how performance scales with length. We show that KnitNet outperforms the strongest baseline in every tested length group from 20–511 nt, solving 306/500 versus 209/500 targets within one hour.

  • Extended planar-benchmark performance (reviewers zefs and 7y9J). Reviewers zefs and 7y9J also questioned planar-benchmark performance. An extended campaign combining search and target-specific refinement reaches 97/100 cumulative exact solves on Eterna100-v2, a leading result in computational solutions on this benchmark, showing that the initial 60-minute results were not a performance ceiling.

Together, these results establish KnitNet as the leading RNA design framework across pseudoknotted, planar 2D, and 3D structures, combining strong exact-solve performance with broad generalization. Our novel integration of inference-time learning and adaptive repair delivers substantial gains over strong baselines and converts additional computation effectively.

We thank you once again for your constructive feedback and insightful questions, which enabled us to substantiate these advances and further demonstrate the strength of KnitNet’s contribution.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.