acceptodds
Under review as a conference paper at ICLR 2027

Partner-Conditioned RNA Representations for Full-Sequence snoRNA–Target Interaction Prediction

Abstract

Predicting small nucleolar RNA (snoRNA) interactions requires identifying both the target RNA and its interaction site among many sequence-compatible candidates. Existing supervised snoRNA formulations largely score preselected local regions, while interaction predictors based on RNA language models (RNA-LMs) typically combine representations obtained independently for each RNA. We introduce SnoBench, an experimentally grounded benchmark that extends evaluation from controlled ranking of predefined candidate windows to target identification and site localization from complete target sequences. Built from transcriptome-wide RNA-RNA contact measurements, SnoBench contains 9,156 snoRNA-target pairs and 31,790 observed interaction regions, preserves variable-width annotations, keeps closely related targets within the same data split, and samples cross-target controls from a genome-wide target pool. We further develop a partner-conditioned interaction model that adapts frozen pretrained RNA representations to the candidate partner before interaction scoring. Bidirectional cross-attention allows the two RNA representations to inform one another, while a signed nucleotide-pair attention bias uses biologically motivated base-pairing information to guide cross-RNA attention. Granularity-matched prediction heads separately model target-level compatibility and site-level interaction evidence. Across representative thermodynamic, snoRNA-specific, and RNA-LM-based baselines, our model performs best in both evaluation settings. In full-sequence prediction with 100 candidate targets, it improves Target Recall@5 from 10.34% to 15.72% and Target MRR from 0.0911 to 0.1183 over the strongest baseline, while more than doubling Joint Recall@5, which requires correct target identification and site localization. SnoBench further reveals that strong discrimination among predefined local candidates does not necessarily transfer to complete sequences.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA language models, RNA-RNA interaction prediction, snoRNA, cross-attention, biological inductive bias, benchmark, computational biology

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_dPA4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies interactions between RNAs, specifically predicting the existence of interactions and the location of interactions between snoRNAs and their target RNAs. There are two major contributions: 1. propose SnoBench, an experimentally grounded benchmarking framework for both controlled candidate ranking (rank the interaction windows within a target RNA) and full-sequence target identification with site location (rank target RNAs according to their interaction probability with snoRNA and predicts site of interaction for each target RNA) tasks; 2. a novel partner-conditioned interaction model for two tasks by adding a learned signed nucleotide-pair attention bias and granularity-match dual heads. The two major contributions are both well-motivated and leads to conclusions that 1. strong discrimination among local candidate windows doesn't necessarily transfer to complete sequences; 2. target identification and site localization need to learn jointly.

Critical strengths

This paper proposes a more realistic benchmarking setup for RNA-RNA interaction prediction models when both the targets and its interaction sites are unknown. The new benchmarking and the conclusion from the benchmarking are impactful for this field to motivate new methods that can be used in practice. The paper also proposes a new methods for predicting targets and sites at the same time. Moreover, the paper is well written, especially the benchmarking part.

Critical weaknesses

Although the benchmarking part is well written, the partner-conditioned snoRNA interaction model section is less well written with several missing details (see my questions).

In the experimental section, some of the standard deviation are missing which makes it hard to justify the significance of the conclusions.

Questions for authors
  1. It seems that the partner-conditioned snoRNA interaction model takes a complete snoRNA sequence q and a candidate window w on target t (section 4.1). It is not clear how this model is applied on the whole sequence.
  2. Is the signed nucleotide-pair bias added to the trainable lookup table (B in eq.4) or is it an initialization of it? If the bias is only an initialization, how would it affect the results when optimization is converged?
  3. Where are H^w and H^q from? Do they have the same length as the window size and the query? What does \Omega_k mean in eq.6?
  4. In experimental results, the standard deviation represents variances across different runs. However, it would be more informative to show the variance across different train-test splits. Could the author show the variances of all results across different splits to show the statistical significance of the proposed methods?
6 comments

[1/2]

We sincerely thank you for recognizing the practical impact of SnoBench and our partner-conditioned model, as well as the clarity of our benchmark formulation. We address each of your questions (Q1–Q4) and the concerns in W1–W2 below, and we will incorporate these clarifications and complete three-seed standard deviations into Sections 4–6 of the revised manuscript. All newly reported ERNIE-based results use the corrected native attention-bias layout (see our response to Reviewer f6GR, Q2) and validation-selected checkpoints across three training seeds.

Q1. How the window-level model is applied to complete target sequences (Addresses W1)

The same candidate-local scorer connects SnoBench's two settings: it ranks supplied windows in controlled evaluation and scans complete candidate targets to produce both target rankings and interaction sites (Appendix C.1):

Candidate window generation: Given a complete candidate target t of length L, we generate candidate windows Iₜ across widths w ∈ {21, 42, 70} nt (for w ≤ L) advancing by stride ⌊w/2⌋, with an additional window starting at L − w when needed to cover the target end.

Contextual scoring: Each candidate window k ∈ Iₜ is taken together with up to 64 nt of flanking sequence on both sides (clipped at target boundaries) to form the target context, which is paired with the full snoRNA q and scored by the model to produce zₖ = aₖ + ℓₖ (Eq. 7).

Target-level reduction and site localization (Eq. 10): Across all N = 100 complete candidate targets, each target t is ranked by its maximum window score Sₜ = max of zₖ over k ∈ Iₜ, and the window achieving this maximum (ŵₜ = argmax of zₖ over k ∈ Iₜ) is output as its predicted interaction site.

Q2. Parameterization of the signed nucleotide-pair bias and its behavior after training (Addresses W1)

1. Initialization vs. additive term: In Eq. (4), the cross-attention bias for head h is αₕB⁽ʰ⁾, where B⁽ʰ⁾ᵢⱼ = Tₕ[xᵢ, yⱼ]. The canonical pairing prior (Watson–Crick = 3.0, G–U/U–G wobble = 1.5, others = 0) initializes each head's nucleotide-pair lookup table Tₕ, while the gates αₕ are initialized to +1, −1, or 0 (four heads per group). Both Tₕ and αₕ remain trainable.

2. Behavior of the effective bias in the trained models: In the submission's Table 3, "w/o pair prior" removes the additive pairing-bias term αₕB⁽ʰ⁾ from cross-attention (Eq. 4; Appendix B.4), lowering controlled MRR. In the corrected three-seed models, the relative Frobenius-norm change across the concatenated effective bias tables (αₕTₕ) is below 0.185%. Positive- and negative-group gates retain their initialized signs, and their effective tables retain their respective initialized pairing orderings; neutral-group gates remain near zero (max |αₕ| ≤ 0.0045). These results indicate that the signed structural prior is preserved after training.

Q3. Origin and lengths of HqH^q and HwH^w, and meaning of Ωₖ (Addresses W1)

Origin and lengths of HqH^q and HwH^w: In Section 4.1, the frozen encoder produces nucleotide representations Xq=E(q)X^q = E(q) for the full snoRNA q and Xc=E(t~w)X^c = E(\tilde{t}_w) for the target context (the candidate window w plus flanking sequence on both sides). Both XqX^q and XcX^c are passed into the bidirectional cross-attention block (Section 4.2, Eq. 4):

  • HqH^q: The partner-conditioned representation from the snoRNA branch, with the exact same length as the query q (|q| nucleotides).
  • HwH^w: On the target branch, after contextual encoding and partner conditioning over the target context, we retain the central slice corresponding to the candidate window w (so flanking tokens provide contextual input without entering the candidate-local interaction map), yielding HwH^w with the exact same length as the candidate window w (|w| nucleotides).

Meaning of Ωₖ in Eq. (6): For candidate k, the 2D interaction map P (Eq. 5) and feature map Fₖ are indexed over snoRNA positions i ∈ {1, …, |q|} and window positions j ∈ {1, …, |wₖ|}. Because query and window lengths vary across instances, Ωₖ = {1, …, |q|} × {1, …, |wₖ|} denotes the set of valid, unpadded 2D nucleotide-pair positions (i, j) for candidate k (|Ωₖ| = |q|·|wₖ|). Pooling over Ωₖ in Eq. (6) (and Eq. 17) aggregates features strictly over valid, unpadded positions.

We will clarify the transition from (XqX^q, XcX^c) to (HqH^q, HwH^w), their lengths |q| and |w|, and the definition of Ωₖ in Sections 4.2–4.3.

[2/2]

Q4. Standard deviations in full-sequence evaluation and variance across data splits (Addresses W2)

1. Complete 3-seed standard deviations for full-sequence N100 evaluation (W2):

We have completed full N100 evaluations across three training seeds for retrained RIME (NT), matched-backbone ERNIE-RIME, No-partner (Ours w/o partner conditioning), and Ours across the entire 986-pair cohort.

In controlled ranking, 3-seed MRR is 0.3982 ± 0.0044 for Ours, 0.3843 ± 0.0071 for No-partner, 0.3729 ± 0.0024 for ERNIE-RIME, and 0.3688 ± 0.0028 for RIME (NT). Complete 3-seed N100 full-sequence results (mean ± SD; recall in %) are:

N100 metricRIME (NT)ERNIE-RIMENo-partnerOurs (Full)
Target R@12.54 ± 0.186.09 ± 0.375.19 ± 1.286.30 ± 0.33
Target R@510.85 ± 0.4611.73 ± 1.9314.82 ± 1.4114.52 ± 2.33
Target MRR0.0915 ± 0.00120.1139 ± 0.00510.1210 ± 0.00360.1255 ± 0.0044
Center Localization R@110.99 ± 1.0213.83 ± 0.3613.22 ± 0.9314.36 ± 1.26
Coverage Localization R@19.36 ± 0.7712.37 ± 0.2012.68 ± 0.9713.37 ± 1.10
Center Joint R@10.81 ± 0.354.60 ± 0.413.72 ± 1.134.81 ± 0.40
Center Joint R@52.43 ± 0.376.22 ± 0.566.60 ± 0.607.11 ± 0.55
Coverage Joint R@10.81 ± 0.354.19 ± 0.523.45 ± 1.244.45 ± 0.43
Coverage Joint R@52.27 ± 0.265.85 ± 0.696.28 ± 0.536.75 ± 0.50

Across three seeds, Ours achieves higher three-seed mean performance than both RIME (NT) (nearly tripling Joint R@5 from 2.43%/2.27% to 7.11%/6.75%) and matched-backbone ERNIE-RIME on all reported endpoints.

2. Sampling uncertainty and consistency across query subsets:

SnoBench uses a fixed target-component-disjoint split and shared N100 candidate lists to standardize comparisons across methods (Appendices A.1 and A.3). Under this protocol, three training seeds quantify training variability, target-component-clustered bootstrap quantifies sampling uncertainty over held-out targets, and Query-Macro and stratified evaluations assess consistency across query weightings and subsets:

Updated target-component-clustered bootstrap: Using paired target-component bootstrap on seed-averaged per-pair metrics, the 95% confidence intervals for Ours minus matched-backbone ERNIE-RIME exclude zero on Target MRR (+0.0116 [+0.0041, +0.0195]), Target R@5 (+2.79 [+0.97, +4.96] pp), Center Joint R@5 (+0.89 [+0.18, +1.59] pp), and Coverage Joint R@5 (+0.90 [+0.20, +1.65] pp); against RIME (NT), the clustered 95% CIs on Center and Coverage Joint R@5 are +4.67 [+0.71, +10.36] pp and +4.49 [+0.53, +10.21] pp.

Consistency across query subsets: Evaluating 3-seed N100 performance under equal query weighting (Query-Macro across all 46 queries) and stratified by Non-U3 vs. U3 subsets confirms positive mean gains across all regimes:

Evaluation regimeTarget MRR: ERNIE-RIMETarget MRR: OursTarget R@5 (%): ERNIE-RIMETarget R@5 (%): OursCoverage Joint R@5 (%): ERNIE-RIMECoverage Joint R@5 (%): Ours
Query-Macro (46 queries)0.1490 ± 0.00920.1787 ± 0.019517.43 ± 7.9229.94 ± 7.819.13 ± 1.2811.61 ± 0.46
Non-U3 (186 pairs)0.1646 ± 0.00080.1856 ± 0.017517.92 ± 5.4925.45 ± 5.9511.11 ± 1.2413.35 ± 0.56
U3 (800 pairs)0.1021 ± 0.00610.1115 ± 0.002910.29 ± 1.2511.98 ± 1.604.62 ± 0.575.22 ± 0.51

We will add the inference details to Section 4 and the complete three-seed statistics to Table 2 in the revised manuscript, and we would appreciate your reassessment with these additions.

Reviewer_dPA4OpenReview ↗

Official Comment

Thank you for your effort in rebuttal and thank you for incorporating the clarifications! About the standard deviations of the performance, could you please report the statistical significance level of the proposed method to the baseline methods, such as using a t-test?

Reviewer_5HTvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SnoBench, a benchmark for snoRNA–target interaction prediction that evaluates both controlled candidate-window ranking and full-sequence target identification with interaction-site localization. Constructed from experimentally observed snoRNA–RNA contacts, SnoBench contains 9,156 snoRNA–target pairs and 31,790 interaction regions, with target-component-disjoint splits and matched cross-target distractors.

The authors further propose a partner-conditioned interaction model combining frozen ERNIE-RNA representations, bidirectional cross-attention with a learnable nucleotide-pairing bias, a 2D interaction module, and separate target and site prediction heads.

Experiments demonstrate improvements over representative thermodynamic, snoRNA-specific, and RNA language model-based baselines. In the full-sequence setting with 100 candidate targets, the proposed method improves Target Recall@5 from 10.34% to 15.72% and Target MRR from 0.0911 to 0.1183 compared with retrained RIME.

Critical strengths

1. A meaningful benchmark addressing an important evaluation gap.

The distinction between controlled candidate ranking and full-sequence target identification is well motivated. SnoBench provides a standardized framework for examining whether methods that discriminate among predefined local candidates can also identify interacting targets and localize interaction regions when candidate sites are not supplied. The observed differences between the two evaluation settings are interesting and potentially valuable for future RNA–RNA interaction prediction research.

2. Thoughtful benchmark construction and evaluation design.

The benchmark incorporates several careful design choices, including target-component-disjoint splits, preservation of variable-width interaction annotations, and distractor matching based on target-level sequence properties. These measures help mitigate potential leakage and shortcut learning. The use of complementary center-based and coverage-based localization metrics is also reasonable given the heterogeneous prediction widths of different methods.

3. Representative baseline comparisons and encouraging empirical results.

The authors compare multiple families of interaction predictors and retrain the main learned baselines on SnoBench. The matched-backbone comparison in Table 3 and the paired and target-component-clustered bootstrap analyses are particularly appreciated. The improvements in full-sequence target retrieval and joint prediction are encouraging, and the benchmark highlights the difficulty of moving beyond controlled local candidate ranking.

Critical weaknesses

W1. The main full-sequence improvements are not sufficiently disentangled from the pretrained backbone and individual architectural contributions.

The central methodological claim is that partner-conditioned representations improve full-sequence interaction prediction. However, the strongest baseline in Table 2, retrained RIME, uses Nucleotide Transformer, whereas the proposed model uses ERNIE-RNA.

The authors partially address this confounding factor through a matched-backbone comparison in Table 3. Replacing RIME's backbone with ERNIE-RNA improves controlled-ranking MRR from 0.3688 to 0.3778, compared with 0.3918 for the proposed model. While this is a useful control, it is only evaluated in the controlled-ranking setting, rather than the full-sequence setting where the paper reports its most important improvements.

Furthermore, the controlled-ranking ablations provide somewhat limited evidence for the independent contribution of partner conditioning. The variant without partner conditioning achieves an MRR of 0.3834, while cross-attention without the pairing prior achieves 0.3814, compared with 0.3918 for the full model. Although the complete architecture performs best, the relatively modest differences and variability across runs make it difficult to determine the magnitude and stability of the contribution from partner conditioning itself.

Consequently, the current experiments support the effectiveness of the overall model configuration but do not fully establish how much of the main full-sequence improvement is attributable to the proposed interaction architecture rather than the backbone and other design choices.

W2. The target-level supervision is not fully aligned with the granularity of its input representations.

The proposed dual-head architecture is an important methodological component, but the motivation for its target-level supervision requires further justification. Specifically, Eq. (8) assigns the same positive target label to all candidate windows belonging to an interacting target, including windows that are treated as negative candidates for site localization. Although this is not inherently inconsistent with training separate target and site predictors, it creates a potential mismatch between the supervision and the information available to each prediction head.

The target logit is computed from a local candidate window and its limited surrounding context, rather than an aggregated representation of the complete target. It is therefore unclear why every window from a positive target should independently provide positive evidence of target-level compatibility, particularly when the actual interaction may be confined to a small region. This design may encourage the target head to learn indirect sequence characteristics associated with positive targets rather than evidence of a specific interaction. The site-level and joint objectives may compensate for this issue, but the existing ablations do not fully clarify the roles of the two heads.

An explicit multiple-instance learning formulation, in which target-level supervision is applied after aggregating window-level evidence, would provide a useful conceptual and empirical comparison.

W3. The full-sequence comparison with snoGloBe may be sensitive to the choice of target-level score aggregation.

One of the paper's important observations is that snoGloBe, despite performing reasonably well in controlled candidate ranking, achieves nearly random target-identification performance in the full-sequence setting. This finding is used to motivate the limitations of existing local interaction predictors.

However, the full-sequence evaluation uses the maximum predicted site probability as the target-level score for snoGloBe (Appendix Table 8). This differs from the more conservative prediction filtering employed in the original snoGloBe study (Deschamps-Francoeur et al., 2022), which required multiple consecutive high-scoring windows when making transcriptome-wide interaction predictions.

The original filtering rule does not directly define a target-ranking score, so it cannot be substituted into the current evaluation without adaptation. Nevertheless, the distinction matters because taking the maximum over many local predictions can amplify isolated false-positive scores, particularly when scanning long target sequences.

It is therefore unclear to what extent snoGloBe's poor full-sequence performance reflects an inherent limitation of its local prediction formulation, as opposed to the aggregation strategy used to construct target-level rankings. Evaluating alternative aggregation rules, such as top-k pooling or contiguous high-scoring window aggregation, would strengthen the comparison and the paper's interpretation of the local-to-full-sequence performance gap.

W4. The biological interpretation and practical generalizability of SnoBench require more careful qualification.

SnoBench is derived from snoKARR-seq measurements, which can capture both direct RNA base-pairing interactions and spatially proximal RNA associations. As discussed in the original experimental study (Liu et al., 2025), these observations do not always establish direct duplex formation at single-nucleotide resolution. This distinction is relevant because the proposed model explicitly incorporates nucleotide-pairing information.

Furthermore, the benchmark treats unobserved candidate interactions as negatives, although the absence of experimental evidence does not necessarily imply the absence of a biological interaction. The extent to which incomplete experimental coverage affects the evaluation remains uncertain.

The practical scope of the reported results is also relatively restricted. The main full-sequence comparison uses 986 box C/D snoRNA–target pairs and a closed-world ranking task containing one annotated positive target among 100 candidates. This provides a meaningful controlled evaluation but does not fully capture the scale, potentially multiple true targets, or uncertainty encountered in transcriptome-wide interaction discovery.

These limitations do not invalidate SnoBench as a benchmark for predicting experimentally observed RNA contacts. However, they warrant a clearer distinction between predicting observed contacts, identifying direct base-pairing interactions, and discovering functionally relevant targets in realistic biological settings.

Questions for authors

Q1. Matched-backbone evaluation in the full-sequence setting. Could the authors evaluate RIME with the ERNIE-RNA backbone in the full-sequence setting? Even a comparison on a representative subset of the N100 test cohort would help determine how much of the reported improvement is attributable to the backbone versus the proposed interaction architecture.

Q2. Target-level supervision and the contributions of the two heads. Have the authors considered applying target-level supervision after aggregating candidate-window representations or scores, for example through multiple-instance learning? Additionally, could the authors report full-sequence performance using the target logits (a_k), site logits (\ell_k), and combined scores (a_k+\ell_k) separately? This would help clarify whether the two heads learn complementary information and whether their combination is necessary.

Q3. Alternative target-level aggregation for local predictors. How sensitive is snoGloBe's full-sequence performance to the aggregation of local predictions? Could the authors compare maximum-score ranking with alternatives based on top-k scores or consecutive high-scoring windows, inspired by the original snoGloBe prediction protocol? This would help establish whether the observed degradation primarily reflects local model limitations or the target-level aggregation procedure.

Q4. Role and interpretation of the signed nucleotide-pair attention bias. Could the authors further investigate the contribution of the signed pairing bias, for example by comparing it with fixed pairing priors, positive-only biases, or randomly initialized trainable biases? Since both the attention gates and pairwise lookup tables are unconstrained and trainable, their product can also be viewed as an effective learned pair-bias table. An analysis of the learned effective biases, or a comparison without separately parameterized gates, would help clarify whether the signed initialization provides a meaningful inductive bias.

Q5. Sensitivity to target length, search-space size, and candidate-set size. Since full-sequence target scores are obtained by maximizing over candidate-window scores (Eq. 10), targets with more scanned windows may have different score distributions. How does performance vary with target length and the number of scanned windows, and to what extent does the distractor-matching procedure control this effect? It would also be informative to examine retrieval performance across different candidate-set sizes, if computationally feasible.

Q6. Full-sequence robustness and statistical dependence. Could the authors provide full-sequence results across multiple training seeds, perhaps on a smaller shared subset? The current bootstrap analyses characterize evaluation-set uncertainty while keeping the trained checkpoints fixed. Additionally, are some snoRNA queries represented by multiple test pairs? If so, query-macro metrics or snoRNA-query-clustered uncertainty estimates would help assess the robustness of the reported improvements beyond target-component-level dependence.

Q7. Absolute localization performance and stricter joint metrics. The absolute full-sequence localization performance remains relatively low. While Appendix Table 6 reports localization results under IoU thresholds, could the authors also report joint target-retrieval and localization performance under IoU ≥ 0.5? A simple random-window or complementarity-based localization baseline evaluated on the same search space would also help contextualize the reported localization accuracy.

Q8. Biological confidence of the interaction annotations. Does the source data permit distinguishing high-confidence direct-duplex interactions from contacts that may reflect spatial proximity? If so, could the authors report performance separately on these subsets, potentially using the duplex-related classifications from the original experimental study? It would also be helpful to discuss how incomplete observation of true interactions and cell-type-dependent RNA expression might affect the construction of negative targets.

Q9. Scope of full-sequence representation and generalization. Appendix A.1 describes target representatives as contiguous strand-oriented genomic reference spans, which may include intronic sequences, while the proposed model scores windows with limited flanking context. Could the authors clarify how this representation relates to mature transcript isoforms and the biological interpretation of "full-sequence" prediction? Additionally, how broadly do the conclusions extend beyond box C/D snoRNAs, and does the benchmark support evaluating generalization to previously unseen snoRNA queries?

8 comments

[1/4]

We sincerely thank you for the positive assessment of our paper's presentation, informativeness, and soundness, and for recognizing SnoBench's task design, leakage controls, and empirical comparisons. Below, we address Q1–Q9 in order, cross-referencing the corresponding concerns (W1–W4). All newly reported ERNIE-based results use the corrected native attention-bias layout (see our response to Reviewer f6GR, Q2) and validation-selected checkpoints across three training seeds.

Q1. Matched-backbone evaluation in the full-sequence setting (Addresses W1; see also Q6)

To separate the contribution of our interaction architecture from the pretrained encoder and quantify multi-seed stability on the full-sequence task, we evaluated three training seeds each for ERNIE-RIME, No-partner (which replaces cross-RNA attention with matched-capacity within-sequence refinement; Appendix B.4), and Ours across the complete 986-pair N100 test cohort. All three models share the exact same frozen ERNIE-RNA encoder.

In controlled candidate ranking, 3-seed MRR is 0.3982 ± 0.0044 for Ours, 0.3843 ± 0.0071 for No-partner, and 0.3729 ± 0.0024 for ERNIE-RIME. Complete 3-seed N100 full-sequence results (mean ± SD; recall in %) are:

N100 metricERNIE-RIMENo-partnerOurs (Full)
Target R@16.09 ± 0.375.19 ± 1.286.30 ± 0.33
Target R@511.73 ± 1.9314.82 ± 1.4114.52 ± 2.33
Target MRR0.1139 ± 0.00510.1210 ± 0.00360.1255 ± 0.0044
Center Localization R@113.83 ± 0.3613.22 ± 0.9314.36 ± 1.26
Coverage Localization R@112.37 ± 0.2012.68 ± 0.9713.37 ± 1.10
Center Joint R@14.60 ± 0.413.72 ± 1.134.81 ± 0.40
Center Joint R@56.22 ± 0.566.60 ± 0.607.11 ± 0.55
Coverage Joint R@14.19 ± 0.523.45 ± 1.244.45 ± 0.43
Coverage Joint R@55.85 ± 0.696.28 ± 0.536.75 ± 0.50

With the frozen ERNIE-RNA backbone held fixed, Ours achieves higher three-seed mean performance than ERNIE-RIME on all reported full-sequence metrics, and improves over No-partner on Target R@1, Target MRR, and all Localization and Joint metrics.

Q2. Target-level supervision, MIL considerations, and contributions of aₖ, ℓₖ, and aₖ + ℓₖ (Addresses W2)

1. Target-level labels and site-specific supervision:

We considered MIL-style target supervision over sampled same-target windows. In SnoBench's controlled lists, the positive target can contribute multiple windows, whereas cross-target controls come from distinct target components (Appendix A.2), making group cardinality label-informative: as Reviewer f6GR's diagnostic shows, ranking windows by component window count alone yields 0.3176 MRR (random: 0.2023). To avoid constructing supervised scores from this sampler-defined grouping, we apply the target loss directly to candidate-local logits, with 1/|Iₜ| averaging losses within each target (Eq. 8). Each candidate score is independent of the other listed windows, so component counts do not enter the forward score. The benchmark methods in Table 1 follow the same independent-candidate scoring protocol (Appendix C.1). Target labels supervise association, while the site and joint objectives (Eqs. 14–15) provide location-specific supervision; the same candidate scorer is used unchanged in full-sequence scanning.

2. Separate full-sequence evaluation of aₖ, ℓₖ, and aₖ + ℓₖ:

Across our three runs on the 986-pair N100 cohort (mean ± SD; recall in %):

N100 metricaₖ onlyℓₖ onlyaₖ + ℓₖ (Default)
Target R@16.50 ± 0.335.81 ± 0.286.30 ± 0.33
Target MRR0.1347 ± 0.00610.1178 ± 0.00630.1255 ± 0.0044
Center Localization R@113.82 ± 0.8814.66 ± 1.0614.36 ± 1.26
Coverage Localization R@112.97 ± 0.8213.74 ± 1.0013.37 ± 1.10
Center Joint R@14.89 ± 0.444.43 ± 0.064.81 ± 0.40
Coverage Joint R@14.57 ± 0.423.99 ± 0.134.45 ± 0.43

The target and site readouts emphasize retrieval and localization, respectively. The aₖ-only readout achieves Center/Coverage Localization R@1 of 13.82%/12.97%, versus 6.93%/6.41% for uniform selection over the same scanned windows. Under the tie handling in Appendix A.4, the latter is also the expected localization performance of a score that is constant within each target. These results demonstrate informative within-target ranking in the jointly trained target head. The pre-specified combined score zₖ = aₖ + ℓₖ further improves Center/Coverage Localization R@1 to 14.36%/13.37%. All three columns are inference-time readouts of the same model trained jointly with target, site, and joint supervision (Eq. 9).

[2/4]

Q3. Alternative target-level aggregation for local predictors (snoGloBe) (Addresses W3)

In the original snoGloBe study (Deschamps-Francoeur et al., 2022), transcriptome-wide predictions required at least 3 consecutive windows scoring above a high probability cutoff (≥ 0.98 in their stringent setting). To test the sensitivity of full-sequence ranking to target-level aggregation, we evaluated a target-ranking adaptation of the contiguous ≥ 0.98 filter and three continuous multi-window aggregators on the 986-pair N100 cohort (holding local scores and the native 13-mer scan fixed):

  • Contiguous ≥ 0.98: requiring ≥ 3 consecutive windows each scoring ≥ 0.98, ranked by the mean score of the best qualifying segment;
  • 3-consecutive min-max: taking the minimum score across 3 consecutive antiparallel windows and maximizing over the target (a threshold-free counterpart of requiring 3 consecutive high-scoring windows);
  • Top-3 mean and Top-5 mean: averaging the top 3 or top 5 window scores per target to smooth isolated single-window spikes.
Aggregation ruleRetrained R@1R@5R@10MRRReleased R@1R@5R@10MRR
Max (Default)1.066.0911.620.05751.045.3310.700.0552
Top-3 mean0.965.9812.580.05741.616.9412.790.0639
Top-5 mean1.526.5911.460.06100.917.1013.790.0613
3-consecutive min-max0.515.0910.300.05191.915.7911.350.0632
Contiguous ≥ 0.98 (≥ 3 win.)0.814.649.510.04921.096.1910.420.0558

Across all five tested aggregation rules, the highest snoGloBe Target MRR is 0.0610 (retrained) and 0.0639 (released), compared with 0.1139 for ERNIE-RIME and 0.1255 for Ours. These results support SnoBench's finding of a gap between local-window discrimination and full-sequence target discovery.

Q4. Role and interpretation of the signed nucleotide-pair attention bias (Also relates to W1)

In the submission's Table 3, "w/o pair prior" removes the additive pairing-bias term αₕB⁽ʰ⁾ from cross-attention (Eq. 4; Appendix B.4), lowering controlled MRR. In the corrected three-seed models, the relative Frobenius-norm change across the concatenated effective bias tables (αₕTₕ) is below 0.185%. Positive- and negative-group gates retain their initialized signs, and their effective tables retain their respective initialized pairing orderings; neutral-group gates remain near zero (max |αₕ| ≤ 0.0045). These results indicate that the signed structural prior is preserved after training.

Q5. Sensitivity to target length, scanned window count, and candidate-set size N

1. Distractor matching reduces marginal length and window-count imbalance: Across the 986 N100 lists, 92.93% of distractors lie within 0.5×–2× of the positive target length (standardized mean difference: 0.0151 for log-length and 0.0149 for window count C(L)). Ranking candidate targets by window count alone yields 0.0534 MRR and 5.04% R@5, matching random expectation (0.0519 MRR, 5.00% R@5).

2. Performance across target-length / window-count quartiles and candidate-pool sizes N: We report Target MRR across target-length quartiles and candidate-set sizes, with smaller candidate sets evaluated using exact expectations over uniform subsets of the N100 distractors (3-seed mean ± SD):

Target length quartile (nt)Pairs (med. win.)ERNIE-RIMEOurs
≤ 3,855250 (119)0.2956 ± 0.01080.3121 ± 0.0049
3,856–19,899243 (1,814)0.0489 ± 0.00490.0639 ± 0.0036
19,900–59,584247 (6,002)0.0414 ± 0.00090.0491 ± 0.0043
> 59,584246 (23,646)0.0662 ± 0.00740.0734 ± 0.0075
Candidate pool size NERNIE-RIMEOurs
N = 100.3772 ± 0.01670.4192 ± 0.0099
N = 250.2302 ± 0.01330.2612 ± 0.0105
N = 500.1597 ± 0.00930.1798 ± 0.0085
N = 1000.1139 ± 0.00510.1255 ± 0.0044

Ours achieves higher three-seed mean Target MRR than ERNIE-RIME in all four length quartiles and across all four candidate-set sizes.

[3/4]

Q6. Full-sequence robustness across training seeds and snoRNA query dependence (Addresses W1; see Q1 for 3-seed N100 table)

Multi-seed full-sequence results across the entire 986-pair N100 cohort are reported in Q1 (3 seeds per method). In addition, because the 986 test pairs span 46 distinct snoRNA queries (with five U3 queries accounting for 800 pairs), we evaluated 3-seed N100 performance weighting all 46 queries equally (Query-Macro) alongside separate U3 and Non-U3 breakdowns (mean ± SD; recall in %):

EvaluationMethodTarget MRRTarget R@5Center Joint R@5Coverage Joint R@5
Query-Macro (46 queries)RIME (NT)0.1557 ± 0.039319.90 ± 3.621.57 ± 0.430.87 ± 0.13
ERNIE-RIME0.1490 ± 0.009217.43 ± 7.929.71 ± 0.859.13 ± 1.28
Ours0.1787 ± 0.019529.94 ± 7.8112.04 ± 0.9111.61 ± 0.46
Non-U3 (186 pairs, 41 queries)RIME (NT)0.1210 ± 0.028715.41 ± 3.462.15 ± 0.541.25 ± 0.31
ERNIE-RIME0.1646 ± 0.000817.92 ± 5.4912.01 ± 1.1211.11 ± 1.24
Ours0.1856 ± 0.017525.45 ± 5.9513.98 ± 0.0013.35 ± 0.56
U3 (800 pairs, 5 queries)RIME (NT)0.0846 ± 0.00529.79 ± 0.622.50 ± 0.382.50 ± 0.33
ERNIE-RIME0.1021 ± 0.006110.29 ± 1.254.88 ± 0.454.62 ± 0.57
Ours0.1115 ± 0.002911.98 ± 1.605.51 ± 0.675.22 ± 0.51

The three-seed mean improvements over both baselines remain positive under equal query weighting and in both the U3 and Non-U3 subsets.

Q7. Absolute localization performance, stricter IoU ≥ 0.5 joint metrics, and random baseline

To contextualize full-sequence localization on the shared 21/42/70-nt search grid, we compare our 3-seed models against a uniform random baseline (uniform target ranking and uniform selection over all scanned windows within each target on the same search grid, with localization evaluated on the annotated positive target t*), and report Joint target-retrieval + localization under IoU ≥ 0.5 (3-seed mean ± SD on the 986-pair N100 cohort; values in %):

MethodLoc. Center@1Loc. Coverage@1Loc. IoU≥0.5 @1Joint (IoU≥0.5) R@1Joint (IoU≥0.5) R@5Joint (IoU≥0.5) R@10
Uniform target + window6.936.412.780.030.140.28
ERNIE-RIME13.83 ± 0.3612.37 ± 0.204.87 ± 0.371.76 ± 0.682.23 ± 0.512.70 ± 0.36
No-partner13.22 ± 0.9312.68 ± 0.975.57 ± 0.411.54 ± 0.612.74 ± 0.333.23 ± 0.42
Ours (Full)14.36 ± 1.2613.37 ± 1.106.05 ± 0.191.88 ± 0.172.86 ± 0.153.62 ± 0.17

Ours achieves 14.36%/13.37% Center/Coverage Localization R@1, more than doubling the same-grid uniform baseline (6.93%/6.41%) and exceeding the thermodynamic baseline IntaRNA reported in Table 2 (9.229%/8.925%). At IoU ≥ 0.5, Ours achieves 6.05% localization recall versus 2.78% for the uniform baseline, and higher three-seed mean Joint (IoU ≥ 0.5) recall than both neural controls at R@1, R@5, and R@10.

Q8. Experimental contact labels and sampled controls (Addresses W4)

Experimental contact ground truth vs. RNAplex-derived duplex classifications: SnoBench retains experimentally observed contact intervals as ground truth. The source study's (Liu et al., 2025) non-rRNA duplex categories are defined using RNAplex predictions with an 8-bp pairing threshold. Using these predictions to select a "high-confidence" subset or define confidence strata would tie the evaluation to RNAplex, a predictor evaluated in this benchmark and also used in snoGloBe's training-window construction, and could favor methods aligned with its predictions. We therefore retain the experimentally defined contact cohort for the primary comparison. Our model's nucleotide-pair bias is incorporated as a learnable attention prior rather than a benchmark label constraint.

Unobserved controls and experimental coverage: The evaluation measures recovery of experimentally observed contacts (target identification and site localization). Unobserved candidates serve as sampled controls (after excluding known interaction partners, aliases, and related target components), as experimental observation depends on cell type, RNA expression, and assay coverage. We will clarify these distinctions in the revised manuscript.

Reviewer_f6GROpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The manuscript introduces SnoBench and a partner-conditioned RNA model for identifying snoRNA interaction targets and localizing interaction sites within complete target sequences. It addresses the gap between ranking preselected local regions and searching full target sequences, where sequence complementarity produces many plausible candidate sites. It also seeks to make pretrained RNA representations responsive to the candidate interaction partner before interaction scoring. The proposed model combines frozen ERNIE-RNA representations with bidirectional cross-attention, a learnable signed nucleotide-pair attention bias, and separate prediction heads for target compatibility and site localization. For full-sequence prediction, it scans candidate windows across each target and uses the highest-scoring window to determine both the target-ranking score and the predicted interaction site. The authors evaluate the approach through controlled ranking of 17 candidate windows and full-sequence prediction over 100 candidate targets, comparing against thermodynamic, snoRNA-specific, and RNA language-model baselines. They report improvements over retrained RIME in controlled-ranking MRR from 0.3688 to 0.3918 and full-sequence Target MRR from 0.0911 to 0.1183, with Target Recall@5 increasing from 10.345% to 15.720%. The benchmark preserves variable-width contact annotations, uses target-component-disjoint splits, and separately evaluates target identification, site localization, and joint prediction. However, full-sequence evaluation uses a single selected checkpoint per method, and the architectural ablations are restricted to controlled candidate ranking. Overall, the manuscript combines partner-conditioned interaction modeling with a benchmark for examining how local interaction scoring transfers to joint target identification and site localization over complete sequences.

Inspection of the supplied supplementary code identified several inconsistencies with the manuscript’s description. The controlled-scoring implementation pools target logits across windows belonging to the same component and applies a component-level cross-entropy loss, whereas Equations 7–8 specify candidate-local scoring and weighted binary cross-entropy; full-sequence evaluation instead uses unpooled target and site logits. The ERNIE wrapper also omits the native permutation of the initial attention bias from [B,L,L,H] to [B,H,L,L] before the downstream reshape, changing the correspondence between bias entries, attention heads, and nucleotide positions. Furthermore, the default controlled-evaluation entry point processes 1,157 test pairs without selecting the stated shared cohort of 986 pairs, and the supplied checkpoint configuration specifies pass 11 rather than the pass 9 reported in Appendix B.2. These findings establish discrepancies in the supplied materials, but their effects on the reported results remain unverified because the corresponding trained checkpoints and execution records were not available.

Critical strengths

A benchmark that distinguishes local discrimination from full-sequence prediction.

The manuscript evaluates both ranking of predefined candidate windows and joint target identification and site localization from complete candidate target sequences. Separate target, localization, and joint metrics help determine whether performance on local candidates transfers to a setting in which the interaction site is unknown. This is a useful contribution to evaluating snoRNA interaction prediction.

Benchmark construction accounts for relevant properties of the experimental data.

SnoBench retains variable-width experimentally observed contact regions and groups related targets into the same data split. These choices address variation in experimental localization resolution and reduce the risk of overly optimistic evaluation caused by closely related targets appearing across splits. Retraining learned baselines on the benchmark also makes the comparisons more informative.

The modeling choices are connected to the prediction task.

Bidirectional cross-attention conditions each RNA representation on its candidate partner, the nucleotide-pair bias introduces a biologically motivated prior, and separate heads distinguish target compatibility from site-level evidence. The manuscript includes matched-backbone comparisons and architectural ablations that investigate these design choices, although those analyses are restricted to controlled candidate ranking.

Critical weaknesses

The supplied scoring implementation and training objective differ from the manuscript.

Equations 7–8 describe candidate-local scoring and candidate-level weighted binary cross-entropy. However, the supplied implementation pools target logits across windows sharing a target component, broadcasts the pooled value back to those windows, and applies cross-entropy over component-level logits. Full-sequence evaluation instead combines unpooled target and site logits. Consequently, the supplied implementation does not follow the stated use of the same candidate-local scoring function in both evaluation settings.

This discrepancy matters because candidate-list construction makes component membership informative about the designated positive. A post-hoc diagnostic was computed on the 986-pair shared evaluation cohort by assigning each window a score equal to the number of windows belonging to its target component. The diagnostic uses no RNA sequence features or trained parameters, and exact ties are evaluated in expectation under uniform tie-breaking.

Component-count diagnosticComputed result
MRR0.3176
Recall@111.34%
Recall@556.26%

The positive window belongs to the uniquely largest component in 983 of the 986 lists. These results establish predictive information in candidate-group structure, but do not establish that the trained model exploits component counts. Normalized log-mean-exp pooling does not automatically reward larger components, and the diagnostic accesses list-level information that would be unavailable to a strictly independent candidate scorer. It also does not explain the reported full-sequence gains, whose inference path uses singleton scoring.

The immediate clarification is to identify the scoring function and training objective actually used for the reported experiments. If the original candidate-level head outputs are available, comparing pooled and candidate-local scores on the same cohort would quantify the inference-time effect of pooling without retraining.

The ERNIE wrapper omits a required attention-bias dimension permutation.

The native implementation permutes the initial projected attention bias from [B,L,L,H] to [B,H,L,L] before attention, whereas the supplied wrapper passes it onward without this permutation. The downstream reshape therefore changes the correspondence between bias entries, attention heads, and nucleotide positions. A native-versus-wrapper output-equivalence check is needed, together with confirmation of which implementation generated the reported results.

Architectural attribution is incomplete for the central full-sequence task.

The matched-backbone comparisons and architectural ablations in Table 3 concern controlled candidate ranking. The principal full-sequence comparison in Table 2 changes both the pretrained encoder and the downstream interaction architecture. Although the controlled experiments provide some evidence for architectural benefits, they do not isolate the contribution of partner conditioning to full-sequence performance. A full-sequence comparison against ERNIE-RNA-based RIME, ideally accompanied by a counterpart without partner conditioning, would directly address this gap.

The evaluation is heavily concentrated in a small number of snoRNA queries.

The supplied frozen definitions show that the 986-pair full-sequence cohort contains only 46 distinct queries, with five U3 queries accounting for 800 pairs, or 81.14%. The pair-micro results therefore largely reflect these frequently represented queries, leaving the distribution of improvements across other snoRNAs unclear. This concentration does not itself establish data leakage, but it limits the breadth of conclusions supported by the aggregate metrics. Query-macro results and separate U3/non-U3 analyses could be computed from existing predictions to assess whether the improvements extend beyond the dominant queries.

Questions for authors

Which scoring function and training objective generated the reported results?

Equations 7–8 and Section 4.5 specify candidate-local scoring and weighted binary cross-entropy, whereas the supplied implementation uses component-pooled target logits and component-level cross-entropy during training and controlled evaluation. Could the authors identify the exact implementation and checkpoint used for the reported experiments and reconcile this discrepancy? Using the original candidate-level head outputs, how do controlled-ranking results change when applying the manuscript’s candidate-local score instead of the supplied component-pooled score? This comparison would help quantify the inference-time effect of pooling in light of the component-count diagnostic described above.

Was the native ERNIE attention-bias layout preserved in the experimental implementation?

The supplied wrapper omits the native permutation from [B,L,L,H] to [B,H,L,L] before the downstream reshape. Did the implementation used for the reported experiments include this permutation? Could the authors provide an output-equivalence check between the wrapper and the native encoder for identical sequences, including padded inputs? If the supplied implementation was used, which reported results are affected, and how does correcting the layout change them?

Do the full-sequence gains persist when the pretrained backbone is held fixed?

Table 3 provides matched-backbone comparisons and architectural ablations for controlled candidate ranking, while the main full-sequence comparison in Table 2 changes both the encoder and the downstream architecture. Could the authors evaluate ERNIE-RNA-based RIME in the full-sequence setting using the existing trained checkpoints? A corresponding evaluation without partner conditioning would further clarify whether this component contributes to the gains on the manuscript’s central task.

How broadly are the improvements distributed across snoRNA queries?

The frozen evaluation definitions indicate that five U3 queries account for 800 of the 986 full-sequence evaluation pairs. Could the authors report query-macro performance and separate U3/non-U3 results for the proposed model and retrained RIME? These analyses could use the existing predictions and would clarify whether the aggregate gains extend beyond the most frequently represented queries.

11 comments

[1/2]

Thank you for the detailed code inspection and the suggested verification checks, as well as for recognizing SnoBench's separation of local discrimination from full-sequence prediction, its variable-width contact annotations, and its target-component-disjoint splits. We corrected the ERNIE attention-bias layout and completed three-seed matched-backbone evaluations on controlled ranking and the full 986-pair N100 cohort. Below, we clarify the implementation and present the encoder check, matched-backbone comparisons, and query-stratified results.

1. The reported experiments for the proposed model used the candidate-local implementation described in the paper (Q1).

The archive contained an earlier pooled-scoring implementation and its configuration (pass 11); the reported full-sequence scan uses the validation-selected seed 0 / pass 9 checkpoint (Appendix B.2). The reported full model used candidate-local weighted BCE (Eq. 8), together with the site and joint objectives (Eqs. 9, 14–15), and scored each window with zₖ = aₖ + ℓₖ, without cross-window pooling. Scoring with the released submission checkpoints and evaluation script (model_run_1–model_run_3, seeds 0–2) reproduces the Ours row in Table 1 (per-seed MRR 0.3893/0.3914/0.3948; 3-seed 0.3918 ± 0.0028), whereas component pooling on the same checkpoints gives lower MRR in each run (0.3793 ± 0.0008). Component membership is not an input to the window scorer, so the group structure highlighted by your diagnostic does not enter the scores. The reported cross-method evaluations use the shared 986-pair box C/D cohort (Appendix A.3; the archive script defaulted to the unfiltered 1,157-pair split).

Submission checkpoints and the candidate-local evaluation script: https://figshare.com/s/e683e2955f690fe897be

2. Attention-bias layout correction and updated three-seed results (Q2).

The ERNIE wrapper used for the ERNIE-based entries in Tables 1–3 omitted the native .permute(0, 3, 1, 2) attention-bias permutation; non-ERNIE baselines in Tables 1–2 (including NT-based RIME) were unaffected. We restored the native permutation in the wrapper. The corrected wrapper matches native ERNIE-RNA exactly on valid nucleotide outputs in CPU FP32 tests, including padded batches (maximum absolute difference: 0.0). Using the corrected encoder, we retrained Ours, ERNIE-RIME, and the No-partner control (which replaces cross-RNA attention with matched-capacity within-sequence refinement; Appendix B.4) from scratch across three seeds each and evaluated each model on controlled candidate ranking and complete N100 full-sequence scanning. With the corrected wrapper, Ours achieves three-seed controlled MRR of 0.3982 ± 0.0044 (submission: 0.3918 ± 0.0028); the updated three-seed N100 results for Ours, ERNIE-RIME, and No-partner are reported below.

3. Multi-seed full-sequence results support our architectural gains with the backbone held fixed (Q3).

Holding the corrected, frozen ERNIE-RNA encoder fixed across all three methods, 3-seed controlled MRR is 0.3982 ± 0.0044 for Ours, 0.3843 ± 0.0071 for No-partner, and 0.3729 ± 0.0024 for ERNIE-RIME (compared with 0.3688 ± 0.0028 for NT-based RIME). Complete 3-seed N100 full-sequence results across both center and coverage localization criteria are below (mean ± SD; recall in %):

N100 metricRIME (NT)ERNIE-RIMENo-partnerOurs
Target R@12.54 ± 0.186.09 ± 0.375.19 ± 1.286.30 ± 0.33
Target R@510.85 ± 0.4611.73 ± 1.9314.82 ± 1.4114.52 ± 2.33
Target MRR0.0915 ± 0.00120.1139 ± 0.00510.1210 ± 0.00360.1255 ± 0.0044
Center Localization R@110.99 ± 1.0213.83 ± 0.3613.22 ± 0.9314.36 ± 1.26
Coverage Localization R@19.36 ± 0.7712.37 ± 0.2012.68 ± 0.9713.37 ± 1.10
Center Joint R@10.81 ± 0.354.60 ± 0.413.72 ± 1.134.81 ± 0.40
Center Joint R@52.43 ± 0.376.22 ± 0.566.60 ± 0.607.11 ± 0.55
Coverage Joint R@10.81 ± 0.354.19 ± 0.523.45 ± 1.244.45 ± 0.43
Coverage Joint R@52.27 ± 0.265.85 ± 0.696.28 ± 0.536.75 ± 0.50

With the backbone held fixed, Ours achieves higher three-seed mean performance than ERNIE-RIME on all reported full-sequence metrics, and improves over No-partner on Target R@1, Target MRR, and all Localization and Joint metrics under both criteria. Using paired target-component bootstrap on seed-averaged per-pair metrics, the 95% confidence intervals for Ours minus ERNIE-RIME exclude zero on Target MRR (+0.0116 [+0.0041, +0.0195]), Target R@5 (+2.79 [+0.97, +4.96] percentage points), Center Joint R@5 (+0.89 [+0.18, +1.59] pp), and Coverage Joint R@5 (+0.90 [+0.20, +1.65] pp).

[2/2]

4. Gains extend to Query-Macro weighting and Non-U3 snoRNAs (Q4).

Following your suggestion, we computed 3-seed N100 results for retrained RIME (NT), matched-backbone ERNIE-RIME, and Ours with all 46 queries weighted equally (Query-Macro), alongside separate pair-micro results for U3 and Non-U3 (mean ± SD; recall in %):

EvaluationMethodTarget MRRTarget R@5Center Joint R@5Coverage Joint R@5
Query-Macro, all 46 queriesRIME (NT)0.1557 ± 0.039319.90 ± 3.621.57 ± 0.430.87 ± 0.13
ERNIE-RIME0.1490 ± 0.009217.43 ± 7.929.71 ± 0.859.13 ± 1.28
Ours0.1787 ± 0.019529.94 ± 7.8112.04 ± 0.9111.61 ± 0.46
U3, 800 pairsRIME (NT)0.0846 ± 0.00529.79 ± 0.622.50 ± 0.382.50 ± 0.33
ERNIE-RIME0.1021 ± 0.006110.29 ± 1.254.88 ± 0.454.62 ± 0.57
Ours0.1115 ± 0.002911.98 ± 1.605.51 ± 0.675.22 ± 0.51
Non-U3, 186 pairsRIME (NT)0.1210 ± 0.028715.41 ± 3.462.15 ± 0.541.25 ± 0.31
ERNIE-RIME0.1646 ± 0.000817.92 ± 5.4912.01 ± 1.1211.11 ± 1.24
Ours0.1856 ± 0.017525.45 ± 5.9513.98 ± 0.0013.35 ± 0.56

The three-seed mean improvements over both RIME (NT) and ERNIE-RIME remain positive under equal query weighting and in both the U3 and Non-U3 subsets, indicating that the gains are not confined to the five most frequent U3 queries.

Together, these checks and experiments support both contributions of the paper: SnoBench's evaluation of joint target identification and site localization from complete sequences, and the proposed model's full-sequence gains with the pretrained encoder held fixed. We would appreciate your updated assessment in light of this evidence.

Thank you to the authors for their detailed response. I appreciate the substantial effort required to correct the implementation and rerun the experiments within the limited rebuttal period. The clarifications and additional experiments have largely addressed my main technical concerns, and I am raising my soundness assessment to 2.

I nevertheless consider this manuscript a close call. Its main contributions are the partner-conditioned interaction model and SnoBench. The additional experiments strengthen the evidence for the model's effectiveness, but I still regard the methodological contribution and the incremental gains from partner conditioning as modest.

For SnoBench, I appreciate the distinction between controlled candidate ranking and full-sequence target identification with site localization. This provides a useful evaluation framework beyond ranking predefined local candidates. However, the benchmark is constructed from existing snoKARR-seq contact measurements, from a study that prominently investigates non-canonical snoRNA–mRNA interactions. Its evaluation therefore concerns recovery of observed RNA contacts rather than the canonical task of identifying snoRNA-guided rRNA or snRNA modifications. This is a legitimate research direction, and the absence of a modification-prediction task does not invalidate the benchmark. Nevertheless, while I recognize the value of organizing these observations into a standardized evaluation protocol, I remain unconvinced that the benchmark contribution, together with the incremental methodological advances, is sufficient to support acceptance.

Overall, the rebuttal has substantially improved my assessment of the manuscript's technical soundness. My remaining reservations concern the significance of the contribution rather than the implementation discrepancies that motivated my initial review. I am therefore raising my overall rating from 1 to 2 (Weak Reject), while acknowledging the manuscript's strengths and the substantive improvements provided during the discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

The released submission checkpoints, scored candidate-locally, reproduce the Ours row in Table 1; the original checkpoints were trained with the candidate-local target loss in Eq.~8. Separately, we corrected the ERNIE attention-bias layout and completed three-seed N100 evaluations with matched-backbone controls. Mean Target MRR is 0.1255 for Ours versus 0.1139 for ERNIE-RIME; the difference is +0.0116+0.0116, with a paired target-component bootstrap 95% confidence interval of [+0.0041,+0.0195][+0.0041,+0.0195], computed from seed-averaged pair metrics. Ours also achieves higher mean localization and joint retrieval--localization performance. We will update the manuscript tables and text with these three-seed results.

AuthorsOpenReview ↗

Summary of Rebuttal Updates and Current Status

Dear Area Chair and Reviewers,

Thank you for the careful and constructive discussion. Following our individual replies and the additional three-seed experiments, we summarize the key updates and where each point stands.

Main contributions, with the updated three-seed results

  1. SnoBench shows that local discrimination does not transfer to full-sequence discovery. Retrained snoGloBe reaches 0.3059 MRR on controlled candidate ranking but 0.0575 Target MRR over 100 complete candidates (random: 0.0519). This holds across five target-level aggregation rules (best: 0.0610 retrained, 0.0639 released).
  2. The interaction architecture improves full-sequence prediction with the encoder held fixed. On the 986-pair N100 cohort (three seeds), Target MRR is 0.1255 ± 0.0044 for our model versus 0.1139 ± 0.0051 for ERNIE-RIME (difference +0.0116, target-component-clustered 95% CI [+0.0041, +0.0195]), and Center/Coverage Joint R@5 is 7.11% / 6.75% versus 6.22% / 5.85%. The gains persist under query-macro weighting and in both the U3 and non-U3 subsets.

Concerns raised and their status

  • Implementation discrepancies (f6GR). The released submission checkpoints, scored candidate-locally, reproduce the Ours row in Table 1. The ERNIE attention-bias permutation is fixed (outputs match native ERNIE-RNA exactly), and all ERNIE-based models were retrained over three seeds. Reviewer f6GR has noted that the remaining reservations concern significance rather than implementation.
  • Backbone confound, missing seeds, and significance tests (5HTv, dPA4, f6GR). Addressed by the matched-backbone three-seed results above and the statistical tests in the dPA4 thread.
  • Benchmark scope and data sources (f6GR, 5HTv). SnoBench evaluates recovery of experimentally observed snoKARR-seq contacts in the non-rRNA transcriptome. Canonical rRNA guide–site recovery requires a separate protocol and is explicitly scoped out, while curated snRNA controls are reported separately as biological references.

Updated headline numbers and refined framing

  • Headline three-seed results: Against retrained RIME (NT), Target R@5 is 14.52% vs. 10.85% and Target MRR is 0.1255 vs. 0.0915 (single-checkpoint values in the submission: 15.72% vs. 10.34% and 0.1183 vs. 0.0911).
  • Role of early partner conditioning: Relative to the capacity-matched No-partner ablation, early partner conditioning gives consistent but modest gains on 8 of 9 N100 metrics, concentrated in top-1 precision, site localization, and joint discovery; coarse Target R@5 is comparable (14.52% vs. 14.82%).
  • Dual-head readout flexibility: The pre-specified unweighted sum aₖ + ℓₖ achieves the highest Joint R@5 under both localization criteria as a balanced default, while the target head (aₖ) or site head (ℓₖ) can be read out directly when prioritizing standalone transcript ranking or known-target localization.

We remain available for any remaining questions. If these updates address your concerns, we would be grateful if you could reconsider your assessment, or let us know what remains open so it can be accounted for in the reviewer–AC discussion.

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.