acceptodds
Under review as a conference paper at ICLR 2027

Geometry Is Not Enough: Ranking RNA Structure Candidates Against a Sequence Prior

Abstract

Computational RNA structure prediction produces many candidate conformations per sequence, and selecting among them remains a bottleneck, especially for long RNAs, whose candidates often share correct base-paired stems but differ in how those stems pack in three dimensions. Geometry alone can confirm that a packing is physically plausible, but not that it is the one the sequence adopts. We therefore frame evaluation as sequence–geometry alignment and introduce SIRGE, an SE(3)-invariant evaluator that conditions the geometry of every nucleotide pair on attention maps from a frozen RNA language model, used as a sequence prior. A contact-recovery objective trains these pair representations to recognise native tertiary contacts, so SIRGE judges each candidate's packing against what its sequence implies. Against six established evaluators under two distribution shifts, SIRGE ranks candidates best. On CLS-1, a benchmark of held-out, longer RNAs with broader candidate quality, it improves Kendall- alignment by 70% over the strongest baseline; on candidates from six unseen deep-learning predictors, it almost doubles the strongest baseline's top-1 hit rate. Ablations show that the prior and the contact objective are complementary, and that alignment helps most on long RNAs. An interpretability study shows that the prior carries tertiary-contact information beyond base pairing, and that SIRGE's scores track native tertiary packing where a geometry-only counterpart does not. These results suggest that for long RNAs geometric plausibility is necessary but not sufficient, and that pretrained sequence models can supply the missing reference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA 3D structure evaluation, AI for Drug Discovery, geometric deep learning, RNA language models

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_7yg7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SIRGE, an evaluator for RNA 3D structure candidates. It conditions an SE(3)-invariant geometric encoder (EquiformerV2 → nucleotide frames → IPA) on attention maps from three selected heads of a frozen RiNALMo, injected through FiLM. Training adds a tertiary-contact recovery objective to the ranking loss. SIRGE beats six baselines on CLS-1 (Kendall-τ 0.270 vs. 0.159) and on Gen-OOD (Hit@1 0.246 vs. 0.135). An interpretability analysis shows that the selected heads carry tertiary-contact signal beyond base pairing.

Critical strengths
  • The motivation is clear and grounded in biology: candidates for long RNAs often share stems but differ in helix packing, and the sequence is always available at scoring time.

  • Appendix B uses careful controls (separation binning, matching on secondary-structure distance, shuffled nulls). The Perm ablation shows that the gain depends on the prior matching the sequence.

Critical weaknesses
  • CLS-1 has only 21 RNAs, and the ± values reflect only three training seeds. The long-RNA claim rests on 11 RNAs. The “non-additive” ablation gap (0.012) is within seed noise. RNA-level paired bootstrap or significance tests are needed.

  • SIRGE-base, with no prior and no contact objective, already reaches 0.217 and beats every baseline. About half of the headline “70%” comes from the architecture. The title and abstract overstate the role of alignment, and SIRGE-base should appear in all main tables, including Gen-OOD.

  • With six candidates, random Hit@1 is about 0.167, and every baseline scores below that, with negative τ. Simple strong baselines are missing: always picking AF3, predictor pLDDT, and consensus/clustering. Most references are bound conformations from complexes.

  • There is no comparison with a non-LM prior, such as ViennaRNA/EternaFold base-pair probabilities or RiNALMo’s secondary-structure head. Without it, the LM-specific contribution is not established. Statistical potentials (rsRNASP, cgRNASP, etc.) and Liu et al. (2026) are not compared.

  • SIRGE may simply reward compactness or contact count. Contacts shared with the native should be separated from spurious ones, controlling for radius of gyration, across all long RNAs.

Questions for authors

How does SIRGE do with a BPP prior in place of the LM attention maps?

What Hit@1 do “always AF3”, pLDDT and consensus baselines reach on Gen-OOD, and how does SIRGE-base do there?

Can you provide RNA-level significance tests?

Do the conclusions hold under TM-score or lDDT?

7 comments
AuthorsOpenReview ↗

Rebuttal 1/5

We thank the reviewer for their careful and thorough review. We address each question separately

Question 1: how does SIRGE do with a BPP prior in place of the LM attention maps?

We thank the reviewer for the question. To answer it, we trained SIRGE with a BPP prior from the two distinct families of secondary-structure predictor that the reviewer mentioned: ViennaRNA (Lorenz et al., 2011), the classical thermodynamic method, and RiNALMo-SS, the RiNALMo language model fine-tuned for base-pair prediction (Penić et al., 2025). We apply the base pairing maps from these two methods to replace the attention maps conditioning in our SIRGE. See the updated Appendix for more details. As the table below shows, both are stronger base-pairing priors than the attention maps.

PriorBase-pair precision on CLS-1 nativesKendall-τ\tau
RiNALMo attention maps (SIRGE)0.53 / 0.50 / 0.470.270 ± 0.011
ViennaRNA maps0.650.213 ± 0.025 (−21.1%)
RiNALMo-SS maps0.560.213 ± 0.031 (−21.1%)
None (No-cond)0.230 ± 0.005 (−14.8%)

Both BPP priors return SIRGE to the level of no prior at all, and the two predictors, despite their different methodologies, give the same result. The deficit is concentrated where SIRGE's gain is concentrated, on the long RNAs (Figure 3): on the 11 RNAs of 150 nucleotides or more, SIRGE reaches a Kendall-τ\tau of 0.238, against 0.160 with the ViennaRNA prior and 0.171 with the RiNALMo-SS prior. This is the regime in which a prior without tertiary information has the most to lose: longer RNAs contain more stems and admit more ways of arranging them, so their packing is the harder part of the problem. The deficit is not a matter of predictor accuracy, since restricting to the RNAs on which each predictor recovers most of the native pairs does not close the gap. Nor is it a matter of the maps being ignored: their FiLM gates grow as large as SIRGE's on the attention maps. The maps are used, and they are not informative about packing.

Sequence changes that preserve every base pair still alter how the helices pack (Yesselman et al., 2019). Figure 5a,b supports this distinction: the stem contacts stay constant across the quality range while the tertiary contacts fall as RMSD grows, so good and bad candidates often share most of the same base pairs and differ in packing. Since the prior is shared by all candidates of an RNA and acts only through its interaction with their geometry (Section 3.1), a map that agrees equally well with every candidate's stems and holds no expectation about which distant nucleotides should touch cannot favour the correct packing. The selected attention heads carry that expectation: Appendix B shows their tertiary-contact AUC stays above chance at all separations of 24 nucleotides and more, after matching on sequence separation and secondary-structure distance.

The attention maps therefore help not through the secondary structure they encode but through the sequence-specific packing information they carry beyond it, which the contact objective trains SIRGE to recover and use as its reference. We have added both ablations to Table 4 and their construction to Appendix A.4.

AuthorsOpenReview ↗

Rebuttal 2/5

Question 2: What Hit@1 do “always AF3”, pLDDT and consensus baselines reach on Gen-OOD, and how does SIRGE-base do there?

We thank the reviewer for suggesting these simple selection baselines. We computed them on Gen-OOD and report the results below.

Selection ruleHit@1
always AlphaFold 310/84 = 11.9 %
highest pLDDT11/84 = 13.1 %
consensus14/84 = 16.7 %
random choice among the six candidates (expectation)16.7 %
SIRGE-base21.8 %
SIRGE24.6 %

The pLDDT rule selects the candidate with the highest mean pLDDT. The consensus rule selects the candidate with the lowest mean RMSD to the other five, i.e., the most central structure among the six predictions. Replacing RMSD with lDDT as the agreement measure gives 9/84 = 10.7%.

All three heuristics are at or below random selection, whereas SIRGE reaches 24.6% Hit@1. This directly addresses the concern that a simple fixed rule might suffice: no simple fixed rule can provide a reliable shortcut for identifying the best candidate across targets. Instead, generator quality is target-dependent, highlighting the need for an evaluator that makes target-specific selections.

AuthorsOpenReview ↗

Rebuttal 3/5

Question 3: Can you provide RNA-level significance tests?

We thank the reviewer for the question and report below the requested analysis.

All statistics below take the RNA as the unit of analysis: each RNA contributes its Top-3 ordering accuracy, averaged over the three seeds, confidence intervals (CI) are 95% percentile bootstrap intervals over RNAs (20,000 resamples), and tests are two-sided Wilcoxon signed-rank tests on the paired per-RNA differences. On CLS-1 SIRGE's Top-3 accuracy has a 95% interval of [0.270, 0.349], above the mean of every baseline (0.212 to 0.265), and its lift over each baseline excludes zero:

ComparisonMean Improvement (↑)95% CIRNAs where SIRGE improvesp-value
SIRGE vs. HomRank+0.043[+0.024, +0.063]16/21< 0.001
SIRGE vs. lociPARSE+0.055[+0.023, +0.088]16/210.004
SIRGE vs. ARES+0.052[+0.022, +0.085]16/210.003
SIRGE vs. RNA3DCNN+0.093[+0.058, +0.132]18/21< 0.001
SIRGE vs. EquiRNA+0.076[+0.042, +0.112]17/21< 0.001
SIRGE vs. PAMNet+0.097[+0.057, +0.141]18/21< 0.001

On Gen-OOD SIRGE's Top-3 interval is [0.246, 0.393], above the mean of every baseline (0.179 to 0.206), and its lift over each baseline excludes zero:

ComparisonMean Improvement (↑)95% CIRNAs where SIRGE improves (ties)p-value
SIRGE vs. HomRank+0.127[+0.044, +0.210]33/84 (38)0.012
SIRGE vs. lociPARSE+0.111[+0.016, +0.206]32/84 (38)0.056
SIRGE vs. ARES+0.139[+0.044, +0.234]35/84 (31)0.007
SIRGE vs. RNA3DCNN+0.139[+0.044, +0.234]36/84 (35)0.016
SIRGE vs. EquiRNA+0.144[+0.026, +0.255]25/51 (17)0.029
SIRGE vs. PAMNet+0.139[+0.048, +0.230]37/84 (29)0.004

EquiRNA's published neighbourhood size (26) cannot build a graph for RNAs of 26 nt or fewer; its row is paired on the 51 Gen-OOD RNAs it scores (28–135 nt).

Reviewer_AaA2OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether a pretrained RNA sequence prior improves selection among candidate three-dimensional RNA structures. SIRGE combines an equivariant atom encoder, nucleotide-local frames, and invariant point attention. Three attention maps from frozen RiNALMo modulate pairwise geometric features through FiLM. Candidate-level listwise ranking is supplemented by per-nucleotide lDDT ranking and tertiary-contact recovery during training.

The authors report improvements over six learned evaluators on longer held-out RNAs and on predictions from six unseen candidate generators. Matched ablations and controlled contact probes support the value of contextual sequence conditioning. The contribution is a useful integration and empirical study; its strongest defensible claim is improved ranking within the tested comparisons, rather than a general necessity of language-model priors for RNA evaluation.

Critical strengths
  1. The matched ablations support the value of sequence conditioning. On CLS-1, SIRGE achieves Kendall ranking alignment of 0.270, compared with 0.230 when sequence conditioning is removed and 0.217 when both the sequence prior and contact objective are removed (Table 4). These comparisons retain the underlying architecture, providing evidence that the proposed additions contribute to improved ranking. This is consequential because it helps isolate the contribution beyond simply building a stronger geometric model.

  2. The sequence-prior analysis includes thoughtful controls. The attention heads are selected on validation RNAs and then examined for structural information on held-out RNAs. Appendix B controls for sequence separation and secondary-structure relationships, and Figure 8 reports RNA-level bootstrap confidence intervals. These checks strengthen the evidence that the prior contains useful information about contacts between distant parts of an RNA, supporting the motivation for combining sequence information with candidate geometry.

Critical weaknesses

The practical Gen-OOD selection benefit is not established against fixed-generator strategies. Section 4.2 makes Hit@1 its main metric for choosing the best candidate from six predictors, but Table 3 compares only learned evaluators. An independent calculation of all six fixed-generator strategies, using the released RMSD labels and the repository's 84-target manifest, gives DeepFoldRNA the highest observed Hit@1: 26/84 (30.95%), compared with SIRGE's reported 24.6%. The baseline calculation was checked against the released metric implementation; SIRGE's predictions were not independently reproduced. The authors should report all six fixed strategies under their exact protocol and include the RMSD of selected candidates. These comparisons would clarify the benefit of adaptive selection over a fixed-predictor choice. The strongest fixed strategy was identified retrospectively, so its advantage on future targets is not established. No statistical significance claim is made. This concern is specific to the practical Gen-OOD Hit@1 claim and does not invalidate the reported improvements over learned evaluators or the CLS-1 results.

[1] Liu, Xiaocheng, et al. "Quality assessment of RNA 3D structure models using deep learning and intermediate 2D maps." Communications Biology 9.1 (2026): 293.

Questions for authors
  1. Can you confirm the fixed DeepFoldRNA result against the exact evaluation inputs used for Table 3, and report all six fixed-generator rules? Please include paired target-level comparisons and the structural quality of selected candidates alongside Hit@1.

  2. Could the authors provide paired target-level confidence intervals for the main performance differences? Appendix D specifies that the reported standard deviations measure variation across training seeds. Uncertainty estimates across RNA targets would provide complementary evidence about the robustness of the average gains. This request concerns the main performance comparisons; Figure 8 already includes RNA-bootstrap intervals for the contact analysis.

  3. RNArank [1] is cited but not compared. Can you provide an appropriate comparison, or explain a specific obstacle?

  4. Can you report the Gen-OOD results for SIRGE-base, No-cond, No-ct, and Perm? Table 4 reports ablations only on CLS-1, while Appendix D describes evaluating the selected checkpoints on both datasets. The requested results would clarify whether the same components account for the Gen-OOD advantage.

8 comments
AuthorsOpenReview ↗

Rebuttal 1/4

We thank the reviewer for their careful and detailed review. We address each question in a separate comment below.

Question 1: Fixed-predictor evaluation strategy We thank the reviewer for a precise and reproducible calculation. We confirm the 26/84 of always-DeepFoldRNA and report below all the six fixed-generator rules, each with the RMSD of the candidates it selects, as requested.

Fixed ruleHit@1Median selected RMSD (Å)Mean selected RMSD (Å)
Always DeepFoldRNA26/84 = 31.0%9.6412.12
Always DRfold19/84 = 22.6%10.6714.57
Always RhoFold13/84 = 15.5%13.2716.86
Always AlphaFold 310/84 = 11.9%16.3320.52
Always RoseTTAFoldNA8/84 = 9.5%13.8617.35
Always trRosettaRNA8/84 = 9.5%12.7315.64
Fixed rule chosen without the labels (expectation over the six)16.7%12.7516.18

The six rules span 9.5% to 31.0%, so a fixed rule is only as good as the choice of generator. Consistently choosing the correct generator is therefore the selection problem itself, not a baseline for it. Always-DeepFoldRNA is a baseline only if DeepFoldRNA can be named the best of the six before the labels are seen. The review notes that the rule "was identified retrospectively", and we agree: it is the best of six rules, selected on the test labels. A fixed rule chosen without the labels gives 16.7% Hit@1 and 16.18 Å mean selected RMSD in expectation, which is what picking a candidate at random gives.

The review also notes that the advantage of always-DeepFoldRNA "on future targets is not established". This can be tested. Dataset 2 of the same benchmark release (Martinović et al., 2024) applies the same six generators to the twelve CASP15 RNA targets, all released after the benchmark's training cutoff.

Fixed ruleHit@1 on Gen-OOD (84 targets)Hit@1 on CASP15 (12 targets)
Always DeepFoldRNA26/84 = 31.0%2/12 = 16.7%
Always DRfold19/84 = 22.6%1/12 = 8.3%
Always RhoFold13/84 = 15.5%3/12 = 25.0%
Always AlphaFold 310/84 = 11.9%4/12 = 33.3%
Always RoseTTAFoldNA8/84 = 9.5%1/12 = 8.3%
Always trRosettaRNA8/84 = 9.5%1/12 = 8.3%

The best rule on Gen-OOD, always-DeepFoldRNA, falls to chance on CASP15. Conversely, the best rule on CASP15, always-AlphaFold 3, is below chance on Gen-OOD. No fixed rule is above chance on both sets. To serve as a baseline, a fixed rule must be chosen before the labels of a new set are seen, yet the ranking of the six rules does not carry over from one set to the next. The benchmark's authors reach the same conclusion: no single tool outperforms the others, because different tools excel on different datasets and metrics (Martinović et al., 2024).

AuthorsOpenReview ↗

Rebuttal 2/4

Question 2: Could the authors provide paired target-level confidence intervals for the main performance differences? Appendix D specifies that the reported standard deviations measure variation across training seeds. Uncertainty estimates across RNA targets would provide complementary evidence about the robustness of the average gains. This request concerns the main performance comparisons; Figure 8 already includes RNA-bootstrap intervals for the contact analysis.

We thank the reviewer for the question and report the requested analysis.

The table below reports the paired difference between SIRGE and each learned evaluator on the main metric of each benchmark, Kendall-τ\tau alignment on CLS-1 (Table 1) and Hit@1 on Gen-OOD (Table 3). Each RNA contributes its metric averaged over the three seeds, the difference is taken per RNA, and the 95% confidence interval is a percentile bootstrap over RNAs (20,000 resamples).

SIRGE minusCLS-1 Kendall-τ\tauGen-OOD Hit@1
ARES+0.136 [+0.077, +0.195]+0.135 [+0.036, +0.230]
PAMNet+0.207 [+0.136, +0.285]+0.151 [+0.052, +0.246]
RNA3DCNN+0.190 [+0.131, +0.248]+0.151 [+0.052, +0.250]
EquiRNA+0.201 [+0.133, +0.267]+0.131 [+0.020, +0.242]†
lociPARSE+0.126 [+0.078, +0.172]+0.111 [+0.012, +0.210]
HomRank+0.110 [+0.053, +0.169]+0.155 [+0.067, +0.238]

†^\dagger EquiRNA can score only 51/84 Gen-OOD RNAs because its 26-nearest-nucleotide neighbourhood is undefined for RNAs of length ≤26.

All twelve intervals exclude zero, so SIRGE’s lead over every learned evaluator on both benchmarks is robust to the choice of RNA targets. We will add this table to the appendix.

AuthorsOpenReview ↗

Rebuttal 3/4

Question 3: RNArank is cited but not compared. Can you provide an appropriate comparison, or explain a specific obstacle?

We thank the reviewer for the question and report the requested experiment.

We retrained RNArank on our training set and evaluated it with the metrics of Tables 1–3. SIRGE outperforms RNArank on every metric on both datasets. We have added these rows to Tables 1–3.

CLS-1 Kendall-τ\tauGlobalD1 (2–5 Å)D2 (5–10 Å)D3 (10–15 Å)
RNArank0.171 ± 0.0100.381 ± 0.0400.260 ± 0.0210.051 ± 0.016
SIRGE0.270 ± 0.0110.453 ± 0.0170.347 ± 0.0100.209 ± 0.007
CLS-1Top-3Hit@10Hit@50Hit@100
RNArank0.272 ± 0.0070.365 ± 0.0270.635 ± 0.0730.794 ± 0.027
SIRGE0.308 ± 0.0130.571 ± 0.0670.738 ± 0.0340.810 ± 0.027
Gen-OODKendall-τ\tauHit@1Top-3
RNArank−0.017 ± 0.0040.115 ± 0.0180.222 ± 0.018
SIRGE0.110 ± 0.0540.246 ± 0.0180.317 ± 0.041
Reviewer_N6q7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, SIRGE, a structural quality evaluation model integrating RNA sequence information and three-dimensional geometric features, is proposed. The author uses the pre-trained attention maps of RiNALMo as a sequence prior, modulates the geometric representation of nucleotide pairs by FiLM, and ranks the candidate RNA structures by combining IPA and three-level contact supervision.

Experiments show that SIRGE outperforms the six baseline methods on CLS-1 and Gen-OOD datasets, especially on long RNA and candidate structures generated by unknown structure predictors.

Critical strengths

It is pointed out that it is difficult to distinguish RNA candidate structures with similar secondary structures but different tertiary stacking methods only by geometric information. The introduction of sequence prior has reasonable biological motivation.

SIRGE is superior to six baseline methods in both test sets, and ablation experiments show that sequence prior and three-level contact supervision have complementary effects.

Critical weaknesses

RiNALMo attention heads are selected according to the ability of base pairing recognition of secondary structure. Although the author proves that these features contain certain tertiary contact information, they lack direct comparison with pure secondary structure priors, so it is difficult to determine whether the performance improvement really comes from sequence-specific tertiary structure information.

CLS-1 not only changed the RNA length and candidate quality, but also significantly changed the proportion of structure generators. In addition, some reference structures in Gen-OOD may come from molecular complexes, while candidate structures are predicted by single strand. These factors affect the interpretation of the generalization conclusion.

Questions for authors

Q1. Can the predicted secondary structure or base pairing probability map be used instead of RiNALMo attention maps to verify whether the tertiary structure prior provides additional benefits?

Is the accuracy of Q2. RiNALMo prior related to the performance improvement of SIRGE? When prior is not accurate, will it reduce the sorting performance?

Q3. Can the evaluation results of MD and FARFAR2 candidate structures in CLS-1 be reported separately, so as to exclude the influence of generator distribution change?

Q4. How much RNA in Gen-ood comes from molecular complexes? Is there any difference between the evaluation performance of bound and unbound RNA?

Q5. Can you provide confidence intervals or significance tests based on independent RNA samples to verify the stability of experimental results?

6 comments
AuthorsOpenReview ↗

Rebuttal 1/5

We thank the reviewer for their careful and detailed review. We address each question in a separate comment below.

Weakness 1 and Question 1: replacing the RiNALMo maps by a pure secondary-structure prior

We thank the reviewer for proposing this ablation. We ran it with two distinct families of secondary-structure predictor: ViennaRNA (Lorenz et al., 2011), the classical thermodynamic method, and RiNALMo-SS, the RiNALMo language model fine-tuned for base-pair prediction (Penić et al., 2025). We apply the base pairing maps from these two methods to replace the attention maps conditioning in our SIRGE. See the updated Appendix for more details. As the table below shows, both are stronger base-pairing priors than the attention maps.

PriorBase-pair precision on CLS-1 nativesKendall-τ\tau
RiNALMo attention maps (SIRGE)0.53 / 0.50 / 0.470.270 ± 0.011
ViennaRNA maps0.650.213 ± 0.025 (−21.1%)
RiNALMo-SS maps0.560.213 ± 0.031 (−21.1%)
None (No-cond)0.230 ± 0.005 (−14.8%)

Both secondary-structure priors return SIRGE to the level of no prior at all, and the two predictors, despite their different methodologies, give the same result. The deficit is concentrated where SIRGE's gain is concentrated, on the long RNAs (Figure 3): on the 11 RNAs of 150 nucleotides or more, SIRGE reaches a Kendall-τ\tau of 0.238, against 0.160 with the ViennaRNA prior and 0.171 with the RiNALMo-SS prior. This is the regime in which a prior without tertiary information has the most to lose: longer RNAs contain more stems and admit more ways of arranging them, so their packing is the harder part of the problem. The deficit is not a matter of predictor accuracy, since restricting to the RNAs on which each predictor recovers most of the native pairs does not close the gap. Nor is it a matter of the maps being ignored: their FiLM gates grow as large as SIRGE's on the attention maps. The maps are used, and they are not informative about packing.

Sequence changes that preserve every base pair still alter how the helices pack (Yesselman et al., 2019). Figure 5a,b supports this distinction: the stem contacts stay constant across the quality range while the tertiary contacts fall as RMSD grows, so good and bad candidates often share most of the same base pairs and differ in packing. Since the prior is shared by all candidates of an RNA and acts only through its interaction with their geometry (Section 3.1), a map that agrees equally well with every candidate's stems and holds no expectation about which distant nucleotides should touch cannot favour the correct packing. The selected attention heads carry that expectation: Appendix B shows their tertiary-contact AUC stays above chance at all separations of 24 nucleotides and more, after matching on sequence separation and secondary-structure distance.

The attention maps therefore help not through the secondary structure they encode but through the sequence-specific packing information they carry beyond it, which the contact objective trains SIRGE to recover and use as its reference. We have added both ablations to Table 4 and their construction to Appendix A.4.

AuthorsOpenReview ↗

Rebuttal 2/5

Question 2: Is the accuracy of RiNALMo prior related to the performance improvement of SIRGE? When prior is not accurate, will it reduce the sorting performance?

We thank the reviewer for this question and answer in two steps: which information in the RiNALMo prior the gain depends on, and what happens to the ranking when the prior lacks it.

1. The gain depends on the kind of information the prior carries, not on its base- pair accuracy.

The two secondary-structure priors of Question 1 recover native base pairs better than the selected RiNALMo heads, yet neither improves ranking. Likewise, the Perm ablation (Table 4) preserves every value of the RiNALMo maps but destroys their correspondence to the RNA sequence, reducing performance to 0.232. Thus, providing base-pair information alone is insufficient. What distinguishes the effective RiNALMo prior is that, beyond recovering base pairs (Figure 9b), it also carries sequence-specific tertiary-contact information related to helix packing (Appendix B). The performance gain therefore depends on this additional tertiary signal, rather than simply on the presence or accuracy of secondary-structure information.

2. An uninformative prior costs the gain and nothing more.

As the paper claims, the prior contributes information that geometry alone cannot provide, and removing this information reduces performance by about 14% (Table 4). Importantly, the degradation is bounded: Perm represents an extreme case in which the prior retains the same numerical values but no longer corresponds to the RNA sequence, yet its performance is essentially identical to No-cond (0.232 vs. 0.230; Table 4). Thus, an uninformative prior does not actively mislead SIRGE; it simply removes the benefit of sequence conditioning. Even in this regime, SIRGE still outperforms all baselines on CLS-1.

This comparison also clarifies what the prior contributes. Geometry alone can distinguish and organize candidate folds, as shown by the SIRGE-base embedding in Figure 4, but this organization is not necessarily oriented toward the native structure. Candidates with similar RMSD can still differ substantially in fold and plausibility (Figure 5a,b), so geometric plausibility alone does not determine which candidate is closer to the native. The sequence prior provides a shared, RNA-specific reference and acts only through its interaction with candidate geometry (Section 3.1), allowing SIRGE to orient geometric variation toward native-like packing. When this sequence correspondence is destroyed, the model falls back to the geometry-only ordering rather than inverting it. Moreover, if the maps helped merely as additional input or through generic RNA statistics, Perm should still outperform No-cond; the fact that it does not indicates that the gain specifically comes from sequence-specific correspondence with the candidate geometry.

In short, an informative prior is useful and improves ranking. An uninformative one returns SIRGE to its geometry-only performance, which already outperforms every baseline.

AuthorsOpenReview ↗

Rebuttal 3/5

Question 3: Can the evaluation results of MD and FARFAR2 candidate structures in CLS-1 be reported separately, so as to exclude the influence of generator distribution change?

We thank the reviewer for the question. Both generators are present in training, molecular dynamics with 94% of the training candidates and FARFAR2 with 6% (Appendix C), and every CLS-1 RNA has candidates from a single generator: molecular dynamics for the 15 PDB RNAs, FARFAR2 for the 6 RNA-Puzzles targets.

The table below reports it for every model on each group, mean ± std over three seeds.

ModelMD RNAs (15)FARFAR2 RNAs (6)
ARES0.156 ± 0.0220.075 ± 0.006
PAMNet0.062 ± 0.0160.065 ± 0.003
RNA3DCNN0.093 ± 0.0200.047 ± 0.019
EquiRNA0.086 ± 0.0250.025 ± 0.007
lociPARSE0.162 ± 0.0130.097 ± 0.009
HomRank0.191 ± 0.0210.079 ± 0.008
SIRGE0.323 ± 0.0110.143 ± 0.014

SIRGE leads every baseline within each generator.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

We thank all the reviewers for their careful and constructive reviews.

Strengths identified by the reviewers

  • Clear, biologically grounded motivation: geometry alone cannot separate candidates that share their secondary structure but differ in tertiary packing, and the sequence is always available at scoring time (N6q7, 7yg7).
  • Consistent gains over the baselines on both benchmarks (N6q7).
  • Matched ablations that isolate the contribution of the sequence prior and the contact objective beyond a stronger geometric model (N6q7, AaA2, 7yg7).
  • Careful controls in the sequence-prior analysis of Appendix B (AaA2, 7yg7).
  • Clear presentation (N6q7, AaA2, 7yg7).

How we addressed the questions and concerns

  • Does the gain come from secondary structure? (N6q7, 7yg7) We trained SIRGE with base-pairing priors from ViennaRNA and RiNALMo-SS in place of the attention maps. Both recover native base pairs more accurately than the attention maps, yet both return SIRGE to the level of no prior. The attention maps therefore help through the sequence-specific packing information they carry beyond secondary structure. An uninformative prior returns SIRGE to its geometry-only performance (N6q7).
  • Robustness across RNAs (N6q7, AaA2, 7yg7). RNA-level paired bootstrap confidence intervals of SIRGE's lift over every baseline exclude zero on both benchmarks.
  • Simple selection rules on Gen-OOD (AaA2, 7yg7). We report all six fixed-generator rules and the pLDDT and consensus rules. No fixed rule is above chance on both Gen-OOD and CASP15, and the pLDDT and consensus rules are at or below random selection.
  • Additional baseline (AaA2, 7yg7). We retrained RNArank; SIRGE outperforms it on every metric on both benchmarks.
  • Generalisation (N6q7, 7yg7). We split CLS-1 by candidate generator and both benchmarks by bound or free reference state. SIRGE leads every baseline within each generator and on both free and bound CLS-1 RNAs, so its advantage is not confined to bound conformations.
  • Ablations on Gen-OOD (AaA2). The gain appears when the prior and the contact objective are both present, as on CLS-1.
  • Other quality labels (7yg7). The reviewer's question highlights how different metrics can lead to different or wrong conclusions. SIRGE outperforms all baselines under TM-score, the more representative metric for tertiary packing.
  • Compactness (7yg7). SIRGE's objective contains no term that rewards compactness or contact count, and its contact objective explicitly separates native-consistent from spurious contacts.

How the paper has been strengthened

We have updated the paper and highlighted the changes in the text.

  • Table 4 now includes the ViennaRNA and RiNALMo-SS prior ablations, with their construction in Appendix A.4 (N6q7, 7yg7).
  • Tables 1–3 now include RNArank as a seventh baseline (AaA2, 7yg7).
  • Appendix D now reports RNA-level paired confidence intervals for the main comparisons (N6q7, AaA2, 7yg7).
  • Appendix D now reports the CLS-1 results under lDDT and TM-score labels (7yg7).
  • The Conclusion now includes a future direction on radius of gyration (7yg7).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.