acceptodds
Under review as a conference paper at ICLR 2027

RiboBridge: protein-conditioned de novo RNA binder structure design

Abstract

Designing RNA molecules that bind to target protein structures requires generating both an RNA conformation and its bound complex geometry, without assuming a predefined binding site. Existing approaches generally model these coupled processes within a unified representation, entangling RNA conformation generation with protein-relative placement. In this work, we introduce RiboBridge, a structure-first framework that separates RNA conformational generation from protein–RNA assembly through cross-frame interaction geometry. RiboBridge first proposes multiple candidate RNA-binding interfaces on the protein surface, generates canonical RNA conformations within a pretrained RNA structural latent space (RiboSphere), and then predicts protein–RNA interaction constraints to bridge the independently defined RNA and protein coordinate frames. The resulting interaction geometry is converted into a geometrically consistent protein-relative pose through rigid placement. We evaluate RiboBridge on a temporally split, sequence-filtered, and redundancy-controlled benchmark of 124 protein–RNA complexes. Under post-hoc interface-correspondence evaluation, where native information is used only after generation to select among already generated predicted-pocket candidates, RiboBridge achieves the highest fNAT and DockQ and the lowest interface and ligand RMSDs among compared methods. Component ablations and stage-wise diagnostics further characterize the contributions of RNA conformational generation, interaction prediction, and rigid placement, identifying the main sources of complex-recovery error.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA design, Protein–RNA complex generation, Biomolecular structure generation, Protein–RNA interactions

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pT3eOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper focuses on the problem of protein-conditioned RNA generation, an important and active area of research that remains computationally unsolved. In contrast to standard approaches, the authors propose generating the RNA structure within coordinate frames independent of the protein target. Their framework, RiboBridge, operates in four main stages: 1) Identification of potential binding sites on the protein: 2) Selected binding site protein residues conditioned generation of RNA structure; 3) Cross-frame interaction prediction; 4) Guided rigid assembling.

Critical strengths
  1. The targeted problem is of high practical importance, and novel methodological approaches are welcome.
  2. The evaluation set preparation protocol is fully clear, sound and fair.
  3. The manuscript is well-written, easy to follow, and the results are clearly presented.
  4. The authors conduct systematic ablations across the pipeline stages to validate their architectural choices.
Critical weaknesses
  1. A significant portion of the evaluation (Table 1) relies on direct comparison to ground-truth structures. Given the generative nature of the task, where a single protein can bind diverse RNA sequences and structures, this metric is not suitable.
  2. The reported values for fNAT, iRMSD, lRMSD, scLRMSD, scTM, and designability are very poor across all baselines as well as RiboBridge itself, especially considering preemptive oracle and DockQ based filtering. Restoring at most 5% of contacts alongside an scLRMSD around 40 Å indicates that none of the evaluated methods are currently near practical usability. Comparing methods under such baseline performance is uninformative, particularly without error bars.
  3. Relying on oracle RNA length and oracle-based candidate selection deviates significantly from real-world application scenarios.
  4. The manuscript does not analyze or discuss the novelty and diversity of the generated entities.
Questions for authors
  1. Is RNA length used to condition the generation comes from ground truth structure, or was there another protocol for its selection?
  2. Is the number of identified binding sites the only reason there are multiple candidate structures?
  3. Could you please release the code to improve paper's reproducibility?
  4. Could you double check the references for correctness? Some authors names are inaccurately reported.
6 comments

Dear Reviewer pT3e,

Thank you for the detailed feedback. We respond below.

W1. Native-reference metrics

Table 1 assesses recovery of experimentally observed protein–RNA binding modes. DockQ, fNAT, and RMSD measure agreement with experimental structures and are appropriate for this goal. Multiple valid modes do not diminish observed-mode recovery, and low native-reference similarity cannot rule out other plausible modes.

To capture what reference-based metrics cannot, we added conditional diversity, structural novelty, and reference-free structural quality analyses. Different predicted-pocket conditions alter RNA intrinsic conformation, not merely rigid placement (W4). We will distinguish known-complex recovery from other generative behaviors.

W2. Performance, utility, and statistical reliability

RiboBridge contributes a testable decomposition of RNA intrinsic conformation generation and protein-relative placement. Stage-wise controls identify each stage's effect and improvement targets.

For 54 targets, generated RNA reaches mean DockQ 0.464 with oracle pose. Exact interaction distances plus RBL retain 0.446 (down only 0.018), while predicted InteractionMap yields 0.253. Thus, RBL preserves recovery with accurate cross-chain constraints; the drop with predicted distances identifies cross-frame interaction prediction as a key bottleneck. Oracle-pose performance separately measures RNA conformation generation.

To quantify statistical uncertainty, we used 5,000 unique-native-RNA cluster bootstrap replicates to obtain 95% CIs for mean DockQ:

MethodMean DockQ [95% CI]
ODesign0.1156 [0.0919, 0.1418]
RNAFlow0.0693 [0.0548, 0.0858]
RFDpoly+NA-MPNN0.0883 [0.0746, 0.1033]
NACraft0.0737 [0.0568, 0.0928]
RiboBridge+NA-MPNN0.1308 [0.1055, 0.1593]

Under Table 1's method-specific conditioning protocols, RiboBridge has the highest mean DockQ among listed methods. The CIs quantify uncertainty, while stage-wise controls identify the sources of error; together they support the structural modeling contribution and guide improvements.

W3 & Q1. Oracle RNA length and candidate selection

RNA length is user-controlled; no native RNA sequence or structure is required. We specify native length on 124 test complexes solely to compare recovery at matched lengths; with protein and predicted pocket fixed, varying requested length generates different RNA structures.

Crucially, our complete Blind Top-1 site-agnostic evaluation uses only the full protein structure and specified RNA length, ranking predicted pockets by confidence without native binding-site information or DockQ-based selection. It therefore tests the entire pipeline without oracle site selection and directly demonstrates end-to-end complex generation and native-like recovery for a subset of targets.

W4. Novelty and diversity of generated RNAs

We add complementary analyses of pocket-conditioned structural diversity and similarity to existing structure collections:

AnalysisMetricResult
Conditional DiversityIntrinsic RNA C4′ RMSD ↑4.296 Å
FSQ token Hamming fraction ↑0.818
Structural NoveltyWhole-fold-near to RiboBridge train ↓8.42%
Whole-fold-near to RiboSphere pretrain ↓5.45%

Conditional Diversity: Across 79 candidate pairs from different predicted-pocket conditions, median intrinsic RNA C4′ RMSD and FSQ token Hamming fraction are 4.296 Å and 0.818. These differences occur within RNA structure and its discrete representation, showing that pocket conditioning affects conformation rather than merely binding pose.

Structural Novelty: Exhaustive US-align nearest-neighbor analysis finds that only 8.42% and 5.45% of generated structures meet the whole-fold-near threshold relative to RiboBridge training and RiboSphere pretraining sets, respectively. Thus, under this defined whole-fold criterion, most candidates lack identified structural neighbors in these collections.

Together, these analyses complement native-reference recovery and demonstrate predicted-pocket-dependent RNA conformation generation.

Q2. Multiple candidate generation.

Candidates represent different predicted binding-pocket hypotheses, not repeated stochastic sampling at one pocket. Of 124 targets, 60, 50, and 14 retain one, two, and three pockets, respectively (202 candidates); 64 targets explore multiple predicted regions. At fixed length, each pocket yields one RNA structure.

Q3. Code availability.

Anonymous inference code: https://anonymous.4open.science/r/ribobridge_anonymous-EC34/

Q4. Reference correctness.

We thank the reviewer for identifying reference errors. We will verify original publication records and correct author names and order, paper titles, and publication details.

Thank you for the detailed response.

I consider the novelty and diversity issues to be resolved. I also thank the authors for sharing their code.

However, I do not agree that comparisons with native structures and related metrics (RMSD, DockQ) are suitable for evaluating generative models. Additionally, generating a single sample per condition is not standard practice. I recommend that the authors focus on AF3-based metrics and generate a suite of samples per target, which will allow for selecting the best candidate (thus improving performance) and provide a clearer picture of the model’s overall performance.

Furthermore, using the oracle RNA length prevents us from understanding the model’s performance in a real-world scenario, where such information is unavailable. I recommend generating samples across a reasonable range of lengths, which is fully compatible with AF3-based evaluation.

The current state of the evaluation is the main reason preventing me from raising my score.

Dear Reviewer pT3e,

Thank you for your thoughtful follow-up and valuable suggestions. We agree that generative evaluation should extend beyond native-reference recovery. We are therefore conducting additional experiments on multi-sample generation, variable RNA lengths, and AF3-based self-consistency.

We respectfully believe that native-reference metrics remain valuable for measuring recovery of experimentally observed binding modes, while reference-free metrics assess complementary aspects of generative quality. Similarly, RNA length is a user-controlled design variable, although evaluating multiple lengths would better characterize practical applications.

We are running these additional experiments and will share the results shortly. We sincerely appreciate your constructive feedback.

Best,

The authors

Reviewer_AqS2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RiboBridge, a modular framework for generating protein-conditioned RNA complex structures from a target protein structure and a requested RNA length, without requiring a user-specified binding site or native RNA sequence. The central methodological idea is to separate intrinsic RNA conformation generation from protein-relative placement. The interaction module uses independently rigid-transform-invariant within-chain features and predicts protein–RNA distance constraints without observing pre-aligned cross-chain coordinates. These constraints are then converted into a rigid transformation through a range-based solver and nonlinear refinement. The method is evaluated on a temporally split benchmark of 124 protein–RNA complexes. Under the paper’s post-hoc interface-correspondence protocol, RiboBridge obtains the best mean complex-recovery metrics among the reported methods.

Critical strengths
  • The structural decomposition is well motivated and technically coherent. Separating intrinsic RNA conformation generation from protein-relative pose recovery avoids requiring the RNA generator to learn arbitrary joint coordinate frames from a relatively small set of protein–RNA complexes. Predicting cross-frame distance constraints from independently invariant within-chain representations is a sensible way to connect the two coordinate systems.
  • The paper provides unusually useful stage-wise diagnostics.
  • The benchmark construction and reporting are relatively careful.
Critical weaknesses
  • The main comparison is not an end-to-end evaluation of the claimed site-agnostic task. RiboBridge retains one to three predicted interface hypotheses and generates a complete candidate for each. The native interface is then used to select the candidate with the greatest native-interface overlap. In an actual site-agnostic application, the native interface is unavailable, so the method must rank or select its candidates using model-accessible information.
  • The methods in Table 1 do not receive comparable information or candidate-selection budgets. ODesign receives a native-derived hotspot, RNAFlow uses a native-RNA-guided receptor crop, RFDpoly receives no explicit site, NACraft operates from protein sequence, and RiboBridge uses multiple predicted sites followed by native-interface-based candidate selection. These are materially different tasks and conditioning regimes. Following each released pipeline is understandable, but the resulting table cannot isolate whether the advantage comes from the proposed structural decomposition, the available conditioning, the number of generated candidates, or the oracle correspondence rule.
  • The evidence does not yet support a strong protein-binding RNA design claim. Under the post-hoc protocol, the mean DockQ is only 0.1308, and 20/124 targets (16.13%) achieve DockQ ≥0.23\geq 0.23. More critically, only 8.06% of targets pass the paper’s AlphaFold 3 “designability” threshold.
  • The temporal split does not establish decontamination of the frozen pretrained components. The RiboBridge train/validation/test complexes are separated by PDB release date, but the method uses a RiboSphere decoder pretrained on a separate RNASolo/gRNAde-derived collection. The paper does not state whether the 124 test RNAs, structurally similar RNAs, or later PDB entries were excluded from RiboSphere pretraining.
  • The baseline coverage is incomplete for the claimed SOTA. The paper discusses Conditional MMDiff and RNA-EFM but does not include them in the experiments. It also omits the recent joint diffusion framework for RNA sequence–structure co-design that reports protein-conditioned and complex-design experiments.
Questions for authors

Please refer to the weaknesses.

8 comments

Dear Reviewer AqS2,

Thank you for engaging with our work and for the opportunity to clarify RiboBridge's novelty and contributions.

W1. End-to-end site-agnostic evaluation

We added a complete Blind Top-1 evaluation, directly demonstrating complex generation without a known binding site. For 124 targets, RiboBridge receives only the full protein structure and a specified RNA length. Before generation, it selects the Top-1 pocket by predicted pocket confidence and produces one RNA–protein complex candidate. No native RNA sequence, coordinates, binding site, or DockQ-based candidate selection is used.

Of 124 outputs, 116 are evaluable (mean DockQ 0.1261), and 19/124 reach the structural recovery threshold. This provides direct evidence of end-to-end site-agnostic complex generation and native-complex recovery for a subset of targets.

W2. Baseline conditioning and candidate budgets

RiboBridge's structural decomposition is supported not only by final performance comparisons but also by independent component ablations and stage-wise error diagnostics. Table 1 reports recovery under each method's actual conditioning protocol. We appreciate the reviewer's observation that methods differ in input information and candidate budgets. We will explicitly describe the original Table 1 as a method-specific conditioning comparison, not a strictly matched site-agnostic SOTA ranking, and will not attribute its differences directly to structural decomposition itself.

W3. Validity of RNA binder design

An important contribution of RiboBridge is decomposing protein-conditioned RNA complex generation into independently evaluable stages of conformation generation, interaction prediction, and rigid placement, with quantitative diagnostics identifying error sources.

In the 54-target stage-wise analysis, generated RNAs achieve mean DockQ 0.464 under an oracle pose. RBL assembly with exact interaction distances achieves 0.446, decreasing to 0.253 when they are replaced by the predicted InteractionMap. This controlled comparison supports the feasibility of separating RNA conformation generation from rigid assembly and identifies cross-frame interaction prediction as the primary target for improvement.

The 8.06% AF3 designability rate provides an independent diagnostic of structure-to-sequence realization: it measures complete-pipeline self-consistency after designing one RNA sequence per generated backbone with NA-MPNN and refolding with AF3, rather than the RNA structure generation module alone.

We will distinguish complex structure generation, sequence realizability, and experimental binding validity in the revision, aligning the paper's structural modeling contribution with the distinct evaluation objectives of functional RNA binder design.

W4. RiboSphere pretraining data overlap

Excluding all eight exact-overlap targets leaves a mean DockQ of 0.1285 on 116 targets, only 0.0023 below the original 0.1308; 18 targets still meet the recovery threshold. This sensitivity analysis shows that these identified overlaps do not dominate the reported structural recovery.

Our audit identified 8/124 test targets with exact RNA overlap in RiboSphere pretraining. RiboSphere learns RNA structural representations and reconstruction without the corresponding proteins or binding interfaces; thus, exposure to RNA structures differs from directly learning protein-conditioned complex generation. We will report this audit and exclusion analysis, limiting its interpretation to sensitivity to the identified exact overlaps.

W5. Conditional MMDiff, RNA-EFM, and recent joint diffusion baselines

To ensure reproducible baseline comparisons, we examined public code, model weights, and inference requirements for each method:

  • RNA-EFM: Code is public, but we have not verified training weights needed to directly reproduce its reported results.
  • Conditional MMDiff: MMDiff provides code and a joint-generation checkpoint, but the fixed-protein-conditioned model and reproduction procedure remain unverified.
  • RiboDiff: It studies protein-conditioned joint RNA design, but we have not found a verifiable complete official inference implementation and corresponding checkpoint.

We therefore cannot yet reliably reproduce these methods under a unified protocol on 124 test complexes. We will discuss their methodological relevance in the related work and retain verifiable comparisons. RiboBridge's modular design contribution is directly supported by its controlled ablations and stage-wise diagnostics, without relying on performance rankings across benchmarks.

Dear Reviewer AqS2,

Thank you again for your constructive review. We would like to highlight the central ML insight of RiboBridge: protein-conditioned RNA generation need not jointly learn intrinsic RNA conformation and protein-relative placement in a shared coordinate frame. Instead, we leverage a pretrained RNA structural prior, learn protein-conditioned conformations in an independent frame, and recover complex geometry through predicted invariant cross-frame interactions.

Importantly, our controlled ablations and stage-wise interventions provide direct evidence for this decomposition, rather than relying solely on comparisons with existing methods. We have also added blind end-to-end evaluation and pretraining-overlap analyses to address your concerns.

We would greatly appreciate it if you could revisit our rebuttal and share your thoughts on whether these results address your main concerns.

Best,

The authors

Thank you for the additional blind evaluation and overlap analysis, which address my concerns and clarify the contribution. Could you also report the Blind Top-1 results after excluding the eight targets with exact RNA overlap in RiboSphere pretraining?

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

The paper proposes RiboBridge, a structure-first approach to protein-RNA binding modeling. Given a target protein and an RNA length, the method first samples a protein binding pocket, and then generates a canonical RNA structure conditioned on the sampled protein pocket. The two representations are then "bridged" to obtain a protein-RNA complex. The authors evaluate RiboBridge across 124 protein-RNA complexes from the PDB showing performance gains relative to selected baselines for structure modeling. The authors further assess sequence realizability using AlphaFold 3 predictions. In ablation studies, the authors analyze the pocket-conditioned RNA generation and decompose the error across components. In a case study, the authors show sampling of binding pockets and according generations across lengths.

Critical strengths
  1. The aspect that the protein pocket is not predefined but sampled by the method, the subsequent pocket-conditioned RNA structure generation and the final joint complex modeling approach is interesting. The separation between protein-pocket sampling and RNA generation appears reasonable and might be a good idea. To me, this is a novel approach.
  2. The paper is generally well written and readable.
Critical weaknesses
  1. I'm generally concerned about the evaluation protocol and benchmark generation for several reasons: (i) the authors propose a structure-first approach to protein-RNA modeling, yet the benchmark creation does not seem to consider structure similarity between splits at all. For RNA, it is well known that sequence similarity alone is not sufficient to avoid leakage across sets [1,2], and existing RNA benchmarks like RNA3DB (3D data) [3] or RNABench (secondary structure data) [4] try to tackle this problem already by using covariance model based approaches for data splitting. Existing approaches for 3D based design like R3Design [5] consider structure similarity during benchmark generation using TM thresholds, and often family-based evaluations (or better clan-based as RNA families often still share structural similarity) are used to avoid direct leakage. Similarly, temporal splitting is not sufficient either. On the protein side, sequence similarity typically is a better leakage control, however, I'm somewhat concerned that the authors did not consider leakage through similar binding-domains / pockets across sets. This strongly affects the reliability of the shown performance of RiboBridge and its generalization to unseen complexes. (ii) As far as I understood, the selection of candidates for evaluation is based on a comparison to the orcale complex. This strongly favors RiboBridge in my opinion and does not reflect real world scenarios where no orcal might be available. An automated selection process would likely reduce performance of the proposed method. (iii) Minor: The sequence realization uses AlphaFold 3 which is known to be quite dependend on MSA availability [6]. While the protein is unchanged across the compared methods, the RNA is not, and MSA availability might generally bias prediction quality. An analysis on RNA MSA availability would be good typically generated RNA sequences have no or few hits with AF3 database search. The same might be true for protein templates as these might differ across methods which could also be avoided.
  2. My second concern is a more general one: If the method does not consider any RNA sequence at all, how well can it perform on protein-RNA interactions that are clearly sequence-driven (like e.g. Pumilio proteins that recognize well defined RNA sequence motifs through their PUF domains)? This somewhat aligns with 1.; Binding domain specific analysis would be a necessary additional analysis in my opinion.

[1] Szikszai, Marcell, et al. "Deep learning models for RNA secondary structure prediction (probably) do not generalize across families." Bioinformatics 38.16 (2022): 3892-3899. [2] Qiu, Xiangyun. "Sequence similarity governs generalizability of de novo deep learning models for RNA secondary structure prediction." PLOS Computational Biology 19.4 (2023): e1011047. [3] Szikszai, Marcell, et al. "RNA3DB: A structurally-dissimilar dataset split for training and benchmarking deep learning models for RNA structure prediction." Journal of Molecular Biology 436.17 (2024): 168552. [4] Runge, Frederic, et al. "Rnabench: A comprehensive library for in silico rna modelling." bioRxiv (2024): 2024-01. [5] Tan, Cheng, et al. "R3Design: deep tertiary structure-based RNA sequence design and beyond." Briefings in Bioinformatics 26.1 (2025): bbae682. [6] Ludaic, Marko, and Arne Elofsson. "Limits of deep-learning-based RNA prediction methods." BioRxiv (2025): 2025-04.

Questions for authors
  1. How structurally similar are RNAs across train/test splits?
  2. How much of the test protein-domains are shared with training domains?
  3. In how far can RiboBridge generate correct poses for sequence specific protein domains?
  4. I might have missed it, but how different are the samples protein pockets? How many of these pockets hit the oracle ones?
  5. In how far differ RNA MSA and protein templates across the sequence realizability experiment?
  6. How are the length distributions of RNAs in train and test datasets?
8 comments

Dear Reviewer ZDXF,

Thank you for your constructive feedback. We address your questions with additional analyses below.

Q1. How structurally similar are RNAs across train/test splits?

We exhaustively compared train/test RNA structures using US-align. Whole-fold-near requires a symmetric structural score ≥0.45 and bidirectional coverage ≥0.70.

RNA groupTargets (%)fNAT↑iRMSD↓lRMSD↓DockQ↑DockQ≥0.23↑ClashScore↓
Whole-fold-near47 (37.9%)0.12525.146230.42300.167411/47 (23.4%)17.1021
Other77 (62.1%)0.01735.485831.57480.10819/77 (11.7%)13.6295

RiboBridge recovers 9/77 targets outside the whole-fold-near group (mean DockQ 0.1081), showing that recovery is not restricted to RNAs with training-neighbor folds. This analysis also quantifies the effect of structural similarity.

Q2. How much of the test protein domains are shared with training domains?

We analyzed test domains using family annotation and structural similarity to training domains.

Protein-domain groupTargets (%)fNAT↑iRMSD↓lRMSD↓DockQ↑DockQ≥0.23↑ClashScore↓
Shared annotated family33 (26.6%)0.06675.74936.8650.12625/33 (15.2%)20.46
Structural neighbor33 (26.6%)0.11554.12731.9670.17266/33 (18.2%)11.86
No detected neighbor17 (13.7%)0.05135.59621.9670.13163/17 (17.6%)9.12
Unannotated/unresolved41 (33.1%)0.00805.94329.6170.09966/41 (14.6%)15.45

RiboBridge recovers 3/17 targets with no detected training-domain structural neighbor (mean DockQ 0.1316); recovery is not confined to shared annotated domain families. The 41 unannotated/unresolved targets are reported separately, not labeled novel.

Q3. In how far can RiboBridge generate correct poses for sequence-specific protein domains?

We used Blind Top-1 (no native-site-based selection) on six canonical targets (two AAUAAA readers; four Nsp15/poly(U)-related) and eight experimentally resolved PUF-3–RNA complexes.

SetTargetsStrict pocket hits↑DockQ↑DockQ≥0.23↑
Canonical sequence-specific65/60.1362/6 (33.3%)
PUF-3 challenge86/80.3616/8 (75.0%)

Without RNA sequence conditioning, RiboBridge achieves 6/8 strict pocket hits and 6/8 successful recoveries on PUF-3 (mean DockQ 0.361). We also evaluated NA-MPNN sequence realization:

BackboneSamplesOverall recovery↑Contact-position recovery↑Exact contact pattern↑
RiboBridge-generated80.5350.5430/8
Native-backbone oracle80.5960.6040/8

Generated backbones support 0.535 overall and 0.543 contact-position sequence recovery. Even native-backbone controls recover no exact contact patterns (0/8), distinguishing pose recovery from precise nucleotide recognition.

Q4. How different are the predicted protein pockets, and how many match the native interface?

Across 124 targets, RiboBridge retains 202 pockets; 64 targets have multiple candidates. The 92 within-protein pocket pairs have median Dice 0.463 and center distance 20.10 Å, supporting spatially distinct hypotheses. Strict hits require Dice ≥0.30 and center distance ≤6 Å.

BudgetStrict hitsRecall↑
Top-145/12436.29%
Top-253/12442.74%
All retained54/12443.55%

Multiple proposals increase native-interface coverage from 45 to 54 targets. We additionally report candidate-selection outcomes:

SelectionMean DockQ↑DockQ≥0.23↑
Blind Top-10.126119/124 (15.32%)
Post-hoc matched0.130820/124 (16.13%)
Oracle-best DockQ0.132020/124 (16.13%)

Blind Top-1 recovers 19/124 complexes without native-interface selection.

Q5. How do RNA MSAs and protein templates differ across methods?

RNA MSA: 99.4% of inputs are query-only; only three have extra homolog hits (1, 58, and 46).

Protein features: For each target, protein sequences, paired/unpaired MSAs, and all four templates (lists, order, alignments) are identical across methods; the latest template date is 2021-09-30.

Thus, systematic protein-template or RNA-MSA-depth differences are not the main source of the comparison results.

Q6. How are the length distributions of RNAs in train and test datasets?

Train/validation/test median lengths are 117/119/14 nt, with means 92.55/98.04/21.04 nt. In the test set, 74.2% of RNAs are <20 nt, confirming its strong enrichment for short RNAs.

Dear Reviewer ZDXF,

Thank you again for your insightful review. We would like to emphasize that RiboBridge introduces a distinct ML formulation: decoupling RNA structural generation from protein-relative placement, while learning invariant cross-frame interactions to bridge the two representations. This enables us to reuse RNA-only structural priors without learning the entire complex geometry jointly.

Beyond the controlled ablations supporting this formulation, our rebuttal provides additional analyses of RNA structural similarity, protein-domain generalization, blind interface selection, and sequence-specific recognition, directly addressing your questions and clarifying the method's limitations.

We would sincerely appreciate it if you could revisit our detailed response and share any remaining concerns. Your feedback would be very valuable to us.

Best,

The authors

Reviewer_ZDXFOpenReview ↗

Response to author comments

I thank the authors for their detailed response and for addressing my questions. The new blind Top-1 selection results are convincing and largely resolve my concern about oracle-based pose selection. In addition, the PUF-domain results for sequence-specific interactions are encouraging, although limited to a small subset of potentially similar examples. I consider the oracle-selection concern largely resolved and the sequence-specific interaction concern substantially addressed, and will raise my soundness score accordingly.

However, the provided additional analyses support some of my initial concerns about the limited generalization of RiboBridge:

  1. The structural similarity between training and test RNAs is substantial (37.9% of targets have structurally similar training examples), while only 17/124 targets have no detected training-domain structural neighbor on the protein side.
  2. Recovery of structurally dissimilar RNA targets remains limited. Successful modeling is approximately twice as frequent for targets with structurally similar training examples.

Additionally, nearly 3/4 of the test RNAs are shorter than 20 nt, much shorter than the training examples. This limits conclusions about performance on longer, structurally more complex RNAs. A length-stratified evaluation could help clarify this limitation.

To address my remaining concerns, I would like to see whether RiboBridge performs better than existing methods on structurally dissimilar test examples. In particular, a comparison against appropriate baselines on the structurally dissimilar test examples, ideally considering both RNA and protein structural similarity, would be valuable. Alternatively, an evaluation using structurally disjoint train/test splits could provide more convincing evidence of generalization.

I do not necessarily expect state-of-the-art performance across all settings, and I already acknowledged the methodological novelty in my initial review. However, I am still not convinced by the generalization results and therefore cannot recommend acceptance in the paper's current form.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

We thank all reviewers for their time, active participation, and insightful comments. The main concerns involved oracle dependence, structural generalization, and generative evaluation. Below, we summarize our responses and RiboBridge's distinctive contributions.

1. Native-interface selection and pretraining overlap

In response to all three reviewers, we added Blind Top-1, selecting the pocket by protein-only confidence before generation. It retains 19/124 recoveries, compared with 20/124 under post-hoc selection. At AqS2's request, excluding the eight identified exact RNA pretraining overlaps leaves mean DockQ essentially unchanged (0.1261 → 0.1272), with 17/116 recoveries. Recovery denotes DockQ ≥0.23; means use evaluable outputs and recovery denominators include all targets.

These results support recovery without native-interface-based selection and limited sensitivity to the identified overlaps. ZDXF considered oracle selection largely resolved, and AqS2 confirmed that the exact-overlap analysis addressed that concern.

2. Structural generalization and baseline comparisons

ZDXF and AqS2 requested comparisons on structurally dissimilar targets. We added RNA-similarity-stratified evaluation and a curated 19-system external cohort with no detected RNA whole-fold neighbors in the audited training, validation, and accessible pretraining collections. No targets were excluded based on model performance.

On the original Other RNA group, Blind Top-1 achieves mean DockQ 0.1139 versus ODesign's 0.1073, with both recovering 9/77 complexes. On the 13 shorter external systems, RiboBridge recovers 5/13 versus 3/13, with mean DockQ 0.2081 versus 0.1917. Original ODesign uses native-derived hotspots; external comparisons share predicted pockets and experimental RNA lengths, with technical retries preventing strict compute matching.

Recovery therefore extends beyond RNAs with detected whole-fold neighbors. Longer RNAs remain challenging: neither method recovers the six longer external systems, and protein-domain/pocket structural novelty remains unverified.

3. Generative evaluation, RNA length, and designability

Following pT3e's requests, we added stochastic backbone sampling, generation at prespecified lengths of 12, 20, 32, and 48 nt, and AF3-guided selection followed by reevaluation with a different seed. These lengths are specified independently of native RNA length.

To address designability concerns from pT3e and AqS2, a separate eight-target experiment used four backbone attempts per target and 16 NA-MPNN sequence designs per valid backbone. RiboBridge and ODesign both achieve 14/32 designable backbones (43.75%) and 6/8 targets with a designable candidate (75%); valid-backbone counts are 29/32 and 32/32, respectively.

These evaluations add evidence beyond native-complex recovery and demonstrate sampling without oracle RNA length. The designability results are computational self-consistency evidence; experimental binding remains untested.

4. Sequence-specific interactions and generated diversity

For ZDXF, we added Blind Top-1 evaluation on PUF-domain complexes, achieving 6/8 recoveries and mean DockQ 0.361, while separately examining sequence recovery. For pT3e, we quantified intrinsic conformational diversity and structural novelty and released inference code.

These analyses support pose recovery in the tested sequence-specific domains, while precise nucleotide recognition remains unresolved. ZDXF considered the sequence-specific concern substantially addressed; pT3e considered generated-sample novelty and diversity resolved.

RiboBridge's Distinctive Contributions

  1. A representation-level decomposition. Separating intrinsic RNA conformation from protein-relative placement enables reuse of RNA-only structural priors, while complex data teach pocket conditioning and interaction geometry.

  2. Invariant cross-frame interaction learning. Independently invariant within-chain representations predict cross-chain distance constraints without pre-aligned complex coordinates; a geometric solver recovers the relative rigid placement.

  3. A testable framework with localized bottlenecks. Controlled interventions separate conformation generation, interaction prediction, and assembly. On 54 diagnostic targets, mean DockQ is 0.464 with oracle pose, 0.446 with exact distances and RBL, and 0.253 with predicted interactions and RBL. These diagnostics support the decomposition and identify interaction prediction as a major bottleneck.

We will incorporate these results, protocol clarifications, and limitations into the revision. We thank all reviewers for strengthening the work and the Area Chair for the time and care devoted to evaluating the paper and discussion.

Best,
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.