- 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
-
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.
-
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
-
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.
-
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
- If validation is performed using a structure predictor independent of the search process, does KnitNet still demonstrate a performance advantage?
- Why does KnitNet underperform NEMO on some traditional RNA reverse-folding tasks?
- As RNA sequence length increases further, how do the computational cost and success rate of online searches change?
5 comments
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 pipeline | Panel exact-solve rate ↑ | Panel mean Jaccard ↑ | Panel mean MCC ↑ |
|---|---|---|---|
| KnitNet extended campaign | 32.27% | 0.7054 | 0.7798 |
| DesiRNA off-the-shelf | 31.75% | 0.6846 | 0.7616 |
| KnitNet Generator → KnitNet search | 25.60% | 0.6700 | 0.7517 |
| KnitNet Generator → continued KnitNet sampling | 22.53% | 0.6608 | 0.7487 |
| Struct2Seq → KnitNet search | 22.80% | 0.6603 | 0.7422 |
| KnitNet init → DesiRNA search | 25.00% | 0.6569 | 0.7397 |
| Struct2Seq → continued released-model sampling | 19.80% | 0.6552 | 0.7443 |
| KnitNet roots → NEMO | 25.13% | 0.6510 | 0.7334 |
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 length | KnitNet exact solves after 1 h | Struct2SeQ-OA after 1 h |
|---|---|---|
| ≤100 | 197/200 (98.5%) | 176/200 (88.0%) |
| 101–200 | 53/61 (86.9%) | 24/61 (39.3%) |
| 201–300 | 48/112 (42.9%) | 6/112 (5.4%) |
| 301–400 | 6/71 (8.5%) | 2/71 (2.8%) |
| 401–511 | 2/56 (3.6%) | 1/56 (1.8%) |
| Overall | 306/500 (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.
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.