acceptodds
Under review as a conference paper at ICLR 2027

Task-Conditioned Dual-Graph Residual with Marginal Constrained Refinement for circRNA-miRNA Interaction

Abstract

Accurate circRNA-miRNA interaction prediction requires integrating heterogeneous biological similarities with unevenly distributed interaction evidence. Different RNAs may rely on different similarity views, while recorded interaction networks vary substantially in endpoint connectivity. TAPER learns RNA-specific preferences over sequence, structural, and composition similarities from interaction-recovery utility estimated within each training fold. These preferences are used to construct task-conditioned biological graphs. A degree-based bipartite reference is then fitted to the training interaction network, and the standardized interaction residual is separated into graph-smooth and complementary components. The graph-smooth component is combined with a graph-regularized factor model to obtain base interaction probabilities, while the complementary component provides neighborhood evidence for candidate-specific refinement. A Bernoulli projection preserves the row and column marginals of the base predictor during this refinement. In five-fold transductive evaluation, TAPER achieves AUCs of 0.9721, 0.9827, and 0.9691 on CMI-9905, CMI-9589, and CMI-20208, respectively. Ablations on CMI-9905 show lower AUC and AUPR after each component removal. TAPER outperforms the tested mechanism alternatives, while literature evidence and expression profiles contextualize its top-ranked predictions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
circRNA–miRNA interaction, Link prediction, Multi-view graph learning, Degree-referenced residual learning, Bernoulli marginal projection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_awggOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper details the development of TAPER, which can be used to predict circRNA-miRNA interactions. Although other predictors exist, TAPER achieves the highest reported AUC, AUPR, and ACC. Additionally, TAPER takes a multi-step approach with task-conditioned graph construction, degree-conditioned residual decomposition, global prediction, marginal-constrained local refinement, and local to global refinement. Of these steps, the ablation of dual-graph residual decomposition was found to have the largest effect on accuracy.
Critical strengths
  • Method description behind the development of TAPER, including the equations used for the construction of the graphs and the prediction, in much detail.

  • Highly relevant problem in biology, finding regulatory post-transcriptional interactions between circRNA and miRNA has great potential for identifying novel elements driving disease and cellular states in general

Critical weaknesses
  • One of the weaknesses of the paper is that, although it is briefly mentioned that circRNA-miRNA interactions regulate gene expression, the further real-world significance and relevance of TAPER is not explicitly expanded upon.

  • Not clear why they did not use a circRNA database for testing and reporting plots AUROC/AUPR over full graphs. This was partly done in the supplement, so please consider moving it to main text and consider a test data set unseen by the model. Also, the AUC of 0.97 over the full set of interactions may indicate a leakage, potentially through a degree bias in the graph. A degree-preserving control has to be done for faithful comparison.

  • Incremental contributions over ML architecture, especially considering recent advances such as a graph attention etc.

  • Although the data on which the model was evaluated is specified, the data on which the model was trained is not explicitly mentioned in the paper.

Questions for authors
  • Is it possible to do graph prediction conditional on a cell type/state? The concern here is that these interactions are context-specific.

  • It seems that sequence basically drives these interactions. Structure is predicted from sequence, so wouldn’t a sequence-only predictor help learn accurate graphs. It would be interesting for example to report the co-folding scores between cirRNA and miRNA as a baseline.

8 comments

We sincerely thank you for your careful review and for recognizing TAPER's detailed methodology and biological relevance. We value your constructive feedback. We added full-pool and cross-dataset evaluations, degree-preserving controls, an RNAcofold baseline, and cell-background analyses, while clarifying training data, novelty, and practical significance. We address each concern below. If these responses resolve your concerns, we would be grateful if you would reconsider your assessment. Thank you again.

Response to W1: Biological Significance and Practical Relevance of TAPER

Thank you for highlighting this gap in the presentation. We agree that the manuscript should explain more explicitly how circRNA-miRNA interactions affect downstream gene regulation and why identifying these interactions is useful in biological research. TAPER is intended to help researchers select candidate interactions for experimental investigation and develop testable hypotheses about the functions of circRNAs.

MiRNAs regulate gene expression by binding to target mRNAs and promoting their degradation or repressing their translation. Some circRNAs bind miRNAs and function as molecular sponges. Under conditions that permit competition for miRNA binding, a circRNA can reduce miRNA activity toward its target transcripts, thereby relieving repression and changing the expression of downstream genes. The extent of this effect depends on the abundance of the interacting RNAs and the availability of binding sites. Identifying the relevant circRNA-miRNA pair therefore provides a possible mechanistic connection between a change in circRNA expression and a change in cellular function.

TAPER’s practical role is to help identify which candidate relationships deserve experimental attention. For example, a researcher may observe that a circRNA is expressed differently between treatment-sensitive and treatment-resistant cells, while the miRNA partners responsible for its possible regulatory effects remain unknown. TAPER provides a ranked list of candidate partners within the available RNA set. Researchers can examine these predictions alongside expression measurements from their experimental system and select pairs for direct binding assays. This gives the model’s output a concrete use in deciding which interactions to investigate when the number of possible pairs exceeds the available experimental capacity.

The predicted pairs can also guide the development of more specific mechanistic hypotheses. By combining a candidate circRNA-miRNA interaction with independently supported miRNA-mRNA target relationships, researchers can nominate downstream genes whose expression might respond to circRNA perturbation. They can then test whether manipulating the predicted miRNA reverses the observed effect of altering the circRNA. Such experiments examine whether the proposed interaction mediates the biological response. In studies of cancer or treatment resistance, this workflow can help identify a regulatory relationship through which an RNA affects cell proliferation or drug sensitivity. Establishing that relationship provides a basis for subsequent studies of whether manipulating the pathway changes the disease-related phenotype.

Our case studies illustrate these connections. Among the 20 highest-ranked predictions, 19 have supporting publications in Table 3. Published experiments linked hsa_circ_0004771 and miR-339-5p to CDC25A regulation and esophageal squamous cell carcinoma progression. The hsa_circ_0025202 and miR-182-5p interaction was linked to FOXO3a regulation and tamoxifen sensitivity in breast cancer. These examples connect predicted interactions to experimentally studied mechanisms with relevance to disease progression and treatment response.

In the revised manuscript, we will expand the Introduction to explain the regulatory connection from circRNA binding to downstream gene expression. We will also describe how TAPER’s ranked predictions can inform candidate selection and the formulation of mechanistic hypotheses. Section 5.5 will be revised to connect the representative predictions to their reported biological functions, with a concise explanation of their relevance to experimental follow-up. We will clarify that the functional consequences of a newly predicted interaction require validation in the relevant biological context.

Response to W2: Full-Graph Evaluation, Data Leakage, and Degree Bias

Thank you for highlighting the need for database-based testing, evaluation over the full candidate graph, and controls for degree effects. Our benchmarks include database-derived circRNA–miRNA records. In particular, CMI-9589 is obtained from circBank, as described in Section 5.1. We will make this provenance more explicit and move the full candidate pool evaluation from Appendix Table 13 into the main experimental section. We will also add the corresponding ROC and precision-recall curves and specify the candidate pool composition.

Table R1. TAPER performance under balanced and full candidate pool evaluation on CMI-9905.

Evaluation settingAUROCAUPR
Balanced evaluation, positive-to-unlabeled ratio of 1:10.9721±0.00230.9794±0.0014
Full eligible unlabeled test pool0.9710±0.00250.7102±0.0089

We added cross-dataset testing using models trained on CMI-9905 and evaluated on additional CMI-20208 interactions. Biological identifier matching established 841 shared miRNAs and 2,113 shared circRNAs. After removing 9,342 overlapping interactions, we retained 7,320 additional positive interactions among these shared RNAs and evaluated them with 8,784 eligible unlabeled pairs. The external test pair identities were reserved before source fitting and excluded from source training supervision. Each source model was fixed before external scoring.

The five source fold models achieve an external AUROC of 0.8993 ± 0.0035 and an AUPR of 0.8975 ± 0.0023. Table R2 provides the detailed results. This experiment assesses prediction of additional interaction labels in a separate benchmark among RNA entities shared with the source dataset.

Table R2. Cross-dataset evaluation of TAPER trained on CMI-9905 and tested on nonoverlapping CMI-20208 interactions among shared RNAs.

FoldACC(%)SEN(%)SPE(%)PRE(%)MCC(%)AUCAUPR
Fold 180.4462.1295.7192.3462.490.90380.8996
Fold 277.1151.9898.0595.7057.780.89530.8958
Fold 380.6163.0295.2891.7562.640.89840.8971
Fold 479.7759.7496.4693.3661.620.90200.9001
Fold 579.8060.8695.5891.9961.290.89700.8951
Mean ± SD79.55±1.4159.54±4.4196.22±1.1293.03±1.6261.16±1.980.8993±0.00350.8975±0.0023

All five source models are evaluated on the same fixed external test set. Classification thresholds are selected exclusively using CMI-9905 validation data. The summary describes variation across the five source models.

We further conducted a controlled assessment of degree effects on CMI-9905 using identical five-fold partitions and evaluation pairs. The degree-only predictor uses the probabilities Πij\Pi_{ij}, fitted exclusively from the training-positive graph. For the degree-preserving control, bipartite edge swaps change interaction partners while preserving the training degree of every circRNA and miRNA. Validation and test pairs remain fixed and are excluded from the admissible swaps. All association-dependent quantities are rebuilt from the rewired training graph. Biological similarity inputs, model architecture, and training hyperparameters are retained.

Table R3. Controlled assessment of degree effects on CMI-9905.

ConditionAUROCAUPR
Degree-only predictor0.8157±0.00550.8361±0.0043
TAPER with degree-preserving rewired training graphs0.8528±0.00320.8541±0.0013
TAPER0.9721±0.00230.9794±0.0014

For the rewired condition, ten repetitions are averaged within each fold before calculating statistics across folds. All fifty runs reached the prescribed swap target.

The degree-only result shows that endpoint degree carries predictive information in this benchmark. Preserving those degrees while changing interaction partners decreases TAPER’s AUROC by 0.1193 and AUPR by 0.1253. This comparison supports a contribution from the observed interaction structure beyond the information retained by node degrees.

To address the leakage concern, test pairs are excluded from supervised training. Every computation that depends on interaction labels uses only the current training partition. Model checkpoints, the global evidence scale, and classification thresholds are selected using validation data. Test interaction labels are reserved for final evaluation. In the cross-dataset experiment, target labels are excluded from feature construction, fitting, and model selection.

We will state this information separation explicitly in the revised manuscript. The appendix will provide the individual fold records and document overlap removal and degree-preservation checks.

Response to W3: Architectural Novelty Beyond Graph Attention Methods

Thank you for highlighting the need to position TAPER in relation to recent advances in graph attention. We recognize that these methods already provide substantial modeling capabilities. GATv2 supports query-dependent attention, and Graphormer incorporates degree and structural information into graph representations. Recent circRNA-miRNA predictors also combine information at different scales.

TAPER contributes a specific relationship between connectivity information and the evidence used for prediction. In Graphormer, degree information enters the representation through centrality encodings. In TAPER, degrees define a reference interaction probability against which observed interactions are evaluated. The resulting residual expresses how a pair departs from the connectivity expected for its endpoints. This gives degree information a statistical role in defining the prediction signal. The biological graphs, constructed from measured interaction-recovery utilities, determine how this residual evidence is organized.

TAPER also specifies how global and local evidence should interact at the prediction level. The residual decomposition assigns the graph-smooth component to global prediction and the remaining component to local refinement. The global predictor then establishes the marginal probability budgets that constrain the local update. Under the projection objective, increasing one candidate’s probability requires compensating adjustments among other candidates sharing its endpoints. This makes local refinement a coordinated allocation of probability across possible partners. The contribution is the explicit coupling between the residual decomposition and the constrained prediction objective, which gives global and local evidence defined responsibilities.

We will revise the Introduction and Related Work to discuss these recent attention-based approaches and make this contribution more precise. We will acknowledge the established foundations of adaptive weighting and graph regularization, and explain how TAPER uses them within its degree-referenced, marginal-constrained formulation. This will clarify the modeling assumptions introduced by TAPER and their relevance alongside advances in attention-based representation learning.

Reviewer_vkt3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies circRNA–miRNA interaction prediction under heterogeneous biological similarities and uneven endpoint connectivity. It proposes TAPER,which jointly performs RNA-specific preference learning, bipartite residual decomposition, global graph-regularized prediction, and candidate-specific residual refinement while preserving interaction probability margins.

Critical strengths
  1. The paper develops a mathematically principled global–local decomposition in which each component has a distinct role, and key results are established theoretically rather than merely asserted.
  2. Each major component is ablated, alternative mechanisms are inserted into the surrounding pipeline, and the appendices report several design choices, clarifying why the proposed decomposition is effective 3.The implementation is documented in detail, improving reproducibility and practical verifiability.
Critical weaknesses
  1. The baseline results in Table 1 are taken from their original publications, but it is unclear whether negative-sample construction, data partitioning, and evaluation procedures are consistent with those used for TAPER. For example, MFERL follows a different five-fold splitting procedure. Consequently, the reported margins do not isolate the contribution of TAPER under matched conditions.
  2. The comparison omits DAMFCMI (https://doi.org/10.1093/bioadv/vbag109), a recent method for circRNA–miRNA interaction prediction that reports results on the same benchmark datasets.
  3. The node-specific view weights are learned from a proxy recovery model whose utility need not align with the final TAPER objective. The Shapley values are defined using similarity propagation and coalition-averaged probabilities, whereas downstream prediction uses factorization, residual decomposition, local gating, and constrained projection. The manuscript does not directly test whether proxy-derived node utilities correlate with downstream gains.
  4. The factor-regularization graph is not permutation invariant. After directed row-wise top-K selection, the implementation retains an edge only when the selected source index is smaller than the target index and an edge selected solely in the opposite direction is discarded. Because RNA indices are arbitrary, node relabeling can change the regularization graph and potentially the predictions. This rule should be corrected or justified with node-permutation tests.
Questions for authors
  1. Can the authors provide a controlled comparison with the strongest baselines under exactly the same experimental protocol?
  2. Can the authors validate that the proxy recovery utilities used to derive node-specific view weights are aligned with the downstream TAPER objective?
  3. Is the index-dependent factor-edge construction in Appendix D.1 intentional, and how sensitive are the results to this rule?
4 comments

We sincerely thank you for your review and for recognizing TAPER's principled design, theoretical grounding, and reproducibility. We addressed your concerns through matched baseline reruns, inclusion of DAMFCMI, tests of proxy–downstream alignment, and permutation tests of a corrected factor-edge rule. We discuss these findings. If our responses resolve your concerns, we would be grateful if you would reconsider your assessment. Thank you again for your valuable feedback.

Response to W1 and Q1: Fair Baseline Comparison Under a Unified Evaluation Protocol

Thank you for highlighting this comparability issue. We have re-evaluated CMAGN and MFERL as representative strong baselines on CMI-9905, CMI-9589, and CMI-20208. These evaluations use exactly the same positive and negative sample partitions as the TAPER main experiments. In particular, the new MFERL evaluations use TAPER’s five-fold assignments, directly aligning its training and test examples with those used for TAPER.

For each dataset, we reuse the actual sampled negative pairs and their fold assignments. Consequently, each method receives the corresponding training pairs and is evaluated on the same held-out pairs within each fold. This controls the composition of the negative examples and the partitioning of the prediction task across methods. Predictions are assessed against the same test labels using consistent metric definitions. Table R1 presents the newly obtained baseline results alongside the original TAPER results from these matched partitions.

Table R1. Controlled comparison of CMAGN, MFERL, and TAPER using identical positive and negative sample partitions across three benchmarks.

DatasetMethodAcc%)Sen(%)Spe(%)Pre(%)Mcc(%)AUCAUPR
CMI-9905CMAGN87.12±0.5085.70±0.8988.53±0.6388.20±0.5674.27±0.990.9094±0.00610.9259±0.0052
MFERL88.95±0.5986.78±1.6791.11±0.7490.71±0.5777.98±1.100.9540±0.00410.9579±0.0034
TAPER93.23±0.6490.55±2.4395.90±1.7895.72±1.6886.63±1.140.9721±0.00230.9794±0.0014
CMI-9589CMAGN87.14±0.8987.15±2.3787.12±3.0187.21±2.2774.35±1.720.9231±0.00600.9332±0.0067
MFERL88.22±0.9887.04±1.4489.41±1.0689.15±1.0276.47±1.960.9512±0.00540.9544±0.0044
TAPER94.09±0.1592.68±1.0995.50±1.1795.40±1.1188.24±0.340.9827±0.00120.9860±0.0007
CMI-20208CMAGN86.47±0.4289.55±1.4583.38±1.9484.38±1.3173.10±0.740.9261±0.00320.9283±0.0033
MFERL90.01±0.3488.85±0.5591.17±0.7890.97±0.6980.04±0.690.9610±0.00200.9630±0.0024
TAPER91.54±0.3789.98±1.1893.10±1.2392.90±1.1483.14±0.750.9691±0.00200.9732±0.0020

TAPER achieves higher mean AUC and AUPR than both re-evaluated baselines on all three datasets. MFERL achieves the higher mean AUC and AUPR among the two baselines throughout the comparison. Relative to MFERL, TAPER improves AUC by 0.0181 and AUPR by 0.0215 on CMI-9905. The corresponding improvements are 0.0315 and 0.0316 on CMI-9589, and 0.0081 and 0.0102 on CMI-20208. TAPER also achieves the highest mean across the five reported classification metrics on each dataset. These directly measured differences support its performance advantage over the two selected baselines with the interaction examples and fold assignments held fixed.

We will replace the literature-derived CMAGN and MFERL entries in the main comparison table with these controlled results. Any retained literature-reported results will be presented separately and identified by their source. The revised experimental section will specify the shared negative examples and five-fold assignments. The appendix will identify the baseline implementations and provide the individual fold results. It will also document preprocessing, model selection, and metric computation, including the classification-threshold selection rule. These revisions will make the experimental conditions underlying the reported margins explicit and tie the comparative conclusions to the controlled evaluations.

Response to W2: Inclusion of DAMFCMI in Benchmark Comparisons

Thank you for bringing DAMFCMI to our attention. We agree that it is a relevant baseline because it evaluates circRNA–miRNA interaction prediction on CMI-9905, CMI-9589, and CMI-20208. We will cite this work in the revised related-work section and include it in the benchmark comparison. Its published results are presented alongside the original TAPER results in Table R2.

Table R2. Reported five-fold benchmark results (mean ± SD) for DAMFCMI and TAPER on three circRNA–miRNA interaction datasets.

DatasetMethodAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)AUCAUPR
CMI-9905DAMFCMI93.35±0.2694.40±1.0992.30±1.2292.48±1.0586.74±0.500.9799±0.00080.9746±0.0027
CMI-9905TAPER93.23±0.6490.55±2.4395.90±1.7895.72±1.6886.63±1.140.9721±0.00230.9794±0.0014
CMI-9589DAMFCMI93.24±0.2692.99±1.2893.49±1.5293.49±1.3786.51±0.560.9801±0.00060.9773±0.0018
CMI-9589TAPER94.09±0.1592.68±1.0995.50±1.1795.40±1.1188.24±0.340.9827±0.00120.9860±0.0007
CMI-20208DAMFCMI90.77±0.2589.55±1.2091.99±0.8791.80±0.7481.59±0.480.9623±0.00140.9629±0.0021
CMI-20208TAPER91.54±0.3789.98±1.1893.10±1.2392.90±1.1483.14±0.750.9691±0.00200.9732±0.0020

On CMI-9905, DAMFCMI reports an AUC of 0.9799, slightly higher than TAPER’s 0.9721. TAPER reports a higher AUPR of 0.9794 compared with DAMFCMI’s 0.9746. On CMI-9589, TAPER’s AUC of 0.9827 and AUPR of 0.9860 exceed DAMFCMI’s corresponding values of 0.9801 and 0.9773. On CMI-20208, TAPER reports an AUC of 0.9691 and an AUPR of 0.9732, compared with 0.9623 and 0.9629 for DAMFCMI, respectively. TAPER also reports higher values for six of the seven evaluation metrics on CMI-9589 and all seven metrics on CMI-20208.

We will incorporate DAMFCMI’s published results into the method comparison table in the main experimental section, covering CMI-9905, CMI-9589, and CMI-20208. The revised table will clearly identify the source of these results. We will also update the accompanying discussion to describe the reported AUC and AUPR comparisons with TAPER and include the corresponding reference in the revised manuscript.

Response to W3 and Q2: Alignment Between Proxy Utilities and Downstream Prediction

Thank you for identifying the need for direct validation of the proxy utilities. We conducted an additional analysis that relates the utilities used to derive node-specific view weights to contributions measured from the complete TAPER predictor.

For each outer fold, we obtained the proxy utilities from the training partition and evaluated all eight task graph view coalitions, including the empty coalition. Each coalition was evaluated through TAPER’s complete prediction pipeline, including factorization, residual decomposition, local gating, and constrained projection. We defined downstream utility using the negative binary cross-entropy of the final projected test probabilities. Exact three-view Shapley values quantified the loss reduction attributable to each view. Averaging these pair-level contributions over the test pairs incident to each node produced downstream contributions directly comparable with the corresponding proxy utilities.

We centered both quantities within each node across its three views and computed Spearman correlations separately for miRNA and circRNA nodes. Table R3 reports the mean and standard deviation across the five outer folds.

Table R3. Correlation between proxy utilities and downstream TAPER contributions

DatasetmiRNA Spearman ρ\rhocircRNA Spearman ρ\rho
CMI-99050.5873±0.05730.4933±0.0327
CMI-95890.6102±0.01740.5673±0.0114
CMI-202080.5655±0.01820.5010±0.0135

Positive mean correlations were observed for both RNA types on all three datasets, ranging from 0.4933 to 0.6102. These results provide empirical support for alignment between the training-derived proxy utilities and view contributions measured after the complete downstream prediction pipeline.

We also evaluated the seven non-empty view coalitions to examine their predictive performance. Table R4 summarizes AUC and AUPR. The complete three-view coalition achieves the highest mean AUC and AUPR on each dataset, providing complementary evidence for jointly incorporating the three views.

Table R4. Prediction performance under different view coalitions for task graph construction

View coalitionCMI-9905 AUCCMI-9905 AUPRCMI-9589 AUCCMI-9589 AUPRCMI-20208 AUCCMI-20208 AUPR
Seq0.9275±0.00480.9428±0.00410.9224±0.00420.9393±0.00340.9327±0.00240.9456±0.0025
Str0.9322±0.00280.9434±0.00310.9110±0.00620.9304±0.00410.9276±0.00270.9408±0.0020
Com0.9212±0.00360.9375±0.00220.9296±0.00400.9431±0.00250.9330±0.00450.9442±0.0027
Seq+Str0.9365±0.00510.9491±0.00440.9309±0.00370.9467±0.00230.9391±0.00360.9516±0.0022
Seq+Com0.9380±0.00390.9481±0.00310.9335±0.00170.9462±0.00110.9426±0.00140.9523±0.0012
Str+Com0.9371±0.00560.9499±0.00380.9425±0.00360.9519±0.00300.9441±0.00250.9529±0.0019
Seq+Str+Com0.9552±0.00430.9654±0.00290.9601±0.00210.9679±0.00130.9612±0.00240.9690±0.0017

We will incorporate the direct alignment results and their interpretation into the appendix section of the revised manuscript. The appendix will provide the downstream contribution definition, coalition construction procedure, node aggregation rule, eligible-node coverage, and per-fold results. It will also include the complete prediction metrics and experimental configurations underlying Table R2.

Response to W4 and Q3: Permutation Invariance and Factor-Edge Construction

Thank you for identifying this issue. Appendix D.1 explicitly describes the factor-edge rule used in the original TAPER experiments. After row-wise top-kfk_f selection, the condition i<ji<j retains a single orientation of the selected entries. Applying this condition directly to directed selections also determines which neighbor pairs are retained. We acknowledge that the resulting factor graph depends on RNA ordering and have evaluated a symmetric union variant to quantify this sensitivity.

The union variant retains an undirected edge whenever either endpoint selects the other. Its weight is the maximum of the selected directional weights. We use kf=10k_f=10 and retain all positive neighbors tied at the Kth similarity threshold. Each undirected edge is stored once after symmetrization. The weighted normalization in Equation D.3 and the objective coefficients in Equation D.4 are retained. This variant provides a controlled comparison with the original factor-edge construction.

We evaluated both constructions on the same five CMI-9905 folds, with matched training examples and initial factors. Model hyperparameters and optimization settings were held fixed. Table R5 presents the original manuscript result alongside the two controlled re-evaluations.

Table R5. Original manuscript result and controlled five-fold factor-edge comparison on CMI-9905.

Method and result sourceAcc%)Sen(%)Spe(%)Pre(%)Mcc(%)AUCAUPR
D.1 rule93.52±0.2890.16±0.9296.32±0.4995.34±0.5786.99±0.560.9710±0.00190.9757±0.0014
Corrected union rule93.37±0.2989.82±1.9596.32±1.2495.36±1.4586.72±0.480.9728±0.00190.9764±0.0017
TAPER93.23±0.6490.55±2.4395.90±1.7895.72±1.6886.63±1.140.9721±0.00230.9794±0.0014

Values are mean ± standard deviation across five test folds. Acc, Sen, Spe, Pre, and MCC are expressed as percentages. The effect of changing the edge rule is assessed using the two controlled re-evaluation rows.

Within the controlled comparison, the union variant increases mean AUC by 0.001824 and mean AUPR by 0.000728. The classification metrics show mixed changes; mean accuracy changes from 93.52% to 93.37%. These measurements quantify the predictive effect of changing the factor-edge construction under matched settings.

We also performed ten random RNA relabelings in each fold, yielding fifty relabelings per construction. We compared edge sets and weights after restoring the original RNA identities. To isolate factor-edge construction, other fold-local inputs were retained by RNA identity and mapped consistently with each permutation. Initial factors and ordered sample pairs were mapped in the same way. Final TAPER predictions were compared for three relabelings of the prespecified first fold.

Table R6. Graph consistency and prediction sensitivity under RNA relabeling on CMI-9905.

Factor-edge ruleMean miRNA edge JaccardMean circRNA edge JaccardMaximum absolute AUC changeMaximum absolute AUPR changeMaximum absolute probability difference
Original D.1 rule0.60110.65980.0033070.0012090.5371
Corrected union rule1.00001.00000.0000000.0000001.43×10−61.43\times10^{-6}

Edge Jaccard values are averages over fifty relabelings per RNA type. Prediction columns report the largest absolute change across three relabelings of the first fold, relative to the corresponding construction’s original-order run.

The original construction changes both the graph and the predictions under relabeling. Its maximum observed probability difference reaches 0.5371, demonstrating sensitivity at the level of individual predictions. The union variant produces identical edge sets and weights after restoring RNA identities in every tested graph comparison. Its three prediction runs retain identical test AUC and AUPR, with a maximum probability difference of 1.43×10−61.43\times10^{-6}.

We will retain the originally reported TAPER results with their factor-edge construction explicitly identified. We will expand Appendix D.1 to state the original rule’s dependence on node ordering and describe the symmetric union variant used in the additional analysis. The factor-fitting step referenced by Algorithm 4 will be linked to the corresponding edge definition. We will incorporate the comparison section into the appendix and provide detailed ranking results in the appendix.

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

This paper proposes TAPER, a framework for circRNA-miRNA interaction prediction that integrates task-conditioned biological graph construction, degree-conditioned residual decomposition, and marginal-constrained prediction refinement. The method estimates RNA-specific preferences over sequence, structure, and composition similarities using training-fold interaction recovery. It then decomposes interaction residuals into graph-smooth and complementary components, which support global interaction prediction and candidate-specific local refinement, respectively. A Bernoulli information projection preserves the row and column marginals of the global predictor during refinement. Experiments on three circRNA-miRNA interaction benchmarks report AUCs of 0.9721, 0.9827, and 0.9691, respectively. Ablation studies and mechanism-level comparisons further demonstrate the contributions of the proposed components.

Critical strengths
  1. Principled integration of global and local interaction evidence.

The proposed framework provides a coherent approach to combining heterogeneous biological similarities and interaction-network information. In particular, separating degree-dependent interaction propensity from residual evidence, and subsequently decomposing the residual into graph-smooth and complementary components, provides a meaningful distinction between global interaction patterns and candidate-specific information. The marginal-constrained refinement further offers a principled mechanism for adjusting individual predictions without changing the interaction budgets established by the global model.

  1. Strong empirical performance supported by component-level validation.

TAPER reports consistently strong predictive performance across three circRNA-miRNA benchmarks. The ablation study shows performance degradation when each major component is removed, and the mechanism-level comparisons examine alternative graph fusion, residual modeling, and probability refinement methods. These experiments provide useful evidence that the proposed components contribute to the observed predictive performance.

Critical weaknesses
  1. Lack of directly comparable evaluations against existing state-of-the-art methods.

A major concern is that the main comparisons rely on baseline results taken from previously published papers rather than reimplementing competing methods under a unified experimental protocol. Differences in data partitioning, negative sampling, preprocessing, and evaluation procedures may substantially affect interaction prediction performance. This is particularly important because some reported improvements over strong baselines are relatively small. Consequently, the current comparisons do not conclusively establish the claimed state-of-the-art performance. Re-evaluating representative strong baselines using identical data splits and evaluation settings would be necessary.

  1. Limited evidence of generalization beyond the transductive interaction prediction setting.

The primary evaluation uses pair-level cross-validation, where the same RNA entities can appear in both training and test sets. This setting allows the model to exploit previously observed interaction patterns and may not adequately reflect the difficulty of predicting interactions involving previously unseen RNAs. Although node-held-out experiments are included, they are limited to one dataset and show noticeable performance degradation, particularly when miRNAs are held out. Moreover, no corresponding comparisons with competing methods are provided under these settings. Broader inductive or cold-start evaluations would strengthen the practical relevance of the method.

Questions for authors

See the Weakness section

2 comments

We sincerely thank you for your careful review and for recognizing TAPER's principled global–local design and strong results. We appreciate your concerns about baseline comparability and generalization to unseen RNAs. We reran CMAGN and MFERL with matched samples and folds, and expanded circRNA and miRNA cold-start tests across three datasets. We address both concerns below. If these results resolve your concerns, we would be grateful if you would reconsider your assessment. Thank you again.

Response to W1: Fair Baseline Comparison Under a Unified Evaluation Protocol

Thank you for highlighting the need for directly comparable evaluations. We have re-evaluated CMAGN and MFERL as representative strong baselines on CMI-9905, CMI-9589, and CMI-20208. For every dataset and fold, these evaluations use exactly the same positive and negative sample partitions as the TAPER main experiments reported in the original manuscript. Each method therefore receives the corresponding training pairs and is evaluated on the same held-out pairs. This comparison directly controls the variation arising from independently generated data partitions and negative samples.

Table R1 presents the newly obtained baseline results alongside the original TAPER main-experiment results. Each dataset is evaluated independently using five-fold cross-validation. We report the mean and standard deviation across the five test folds to show both predictive performance and variation across partitions.

Table R1. Comparison of re-evaluated baselines with the original TAPER main-experiment results using identical positive and negative sample partitions.

DatasetMethodAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)AUCAUPR
CMI-9905CMAGN87.12±0.5085.70±0.8988.53±0.6388.20±0.5674.27±0.990.9094±0.00610.9259±0.0052
MFERL88.95±0.5986.78±1.6791.11±0.7490.71±0.5777.98±1.100.9540±0.00410.9579±0.0034
TAPER93.23±0.6490.55±2.4395.90±1.7895.72±1.6886.63±1.140.9721±0.00230.9794±0.0014
CMI-9589CMAGN87.14±0.8987.15±2.3787.12±3.0187.21±2.2774.35±1.720.9231±0.00600.9332±0.0067
MFERL88.22±0.9887.04±1.4489.41±1.0689.15±1.0276.47±1.960.9512±0.00540.9544±0.0044
TAPER94.09±0.1592.68±1.0995.50±1.1795.40±1.1188.24±0.340.9827±0.00120.9860±0.0007
CMI-20208CMAGN86.47±0.4289.55±1.4583.38±1.9484.38±1.3173.10±0.740.9261±0.00320.9283±0.0033
MFERL90.01±0.3488.85±0.5591.17±0.7890.97±0.6980.04±0.690.9610±0.00200.9630±0.0024
TAPER91.54±0.3789.98±1.1893.10±1.2392.90±1.1483.14±0.750.9691±0.00200.9732±0.0020

TAPER achieves higher mean AUC and AUPR than both re-evaluated baselines on all three datasets. MFERL achieves the higher mean AUC and AUPR among the two baselines throughout the comparison. Relative to MFERL, TAPER improves AUC by 0.0181 and AUPR by 0.0215 on CMI-9905. The corresponding improvements are 0.0315 and 0.0316 on CMI-9589, and 0.0081 and 0.0102 on CMI-20208. TAPER also achieves the highest mean across the five reported classification metrics on each dataset. These comparisons provide direct evidence for its performance advantage over the two representative baselines under matched training and test partitions.

We will incorporate these results into the main experimental section and replace the literature-derived entries for CMAGN and MFERL with the re-evaluated values. Any remaining results taken from previous publications will be explicitly identified by their source. The revised experimental setup will describe the shared positive and negative sample partitions. The appendix will provide the individual fold results and document the preprocessing, model-selection procedures, and parameter settings used for each method. We will base the comparative conclusions for these two baselines on the directly measured results presented here.

Response to W2: Generalization Beyond Transductive Prediction and Cold-Start Evaluation

Thank you for highlighting the need for broader evaluation on unseen RNA entities. We expanded the entity-held-out experiments to CMI-9905, CMI-9589, and CMI-20208. We compare CMAGN, MFERL, and TAPER under both settings on all three datasets.

These experiments partition RNA entities on one side of the interaction graph before fitting. Strict circRNA cold start excludes held-out circRNAs from fitting and withholds their interaction labels. Partially observed miRNA cold start excludes held-out miRNAs from fitting, but designated support interactions are revealed only at inference. Predictions are evaluated on the remaining masked query interactions.

Tables R2 and R3 report five-fold means ± SD. Acc, Sen, Spe, Pre, and MCC are percentages. Bold indicates the highest mean AUC or AUPR within each dataset.

Table R2. Five-fold comparison under the strict circRNA cold-start protocol. Held-out circRNAs have no revealed support interactions.

DatasetMethodACC(%)SEN(%)SPE(%)PRE(%)MCC(%)AUCAUPR
CMI-9905CMAGN79.63±1.0979.15±4.7780.11±5.9080.21±3.5659.52±1.920.8603±0.00720.8733±0.0052
MFERL76.81±2.5480.28±4.2273.35±9.2275.59±5.1954.11±4.350.8537±0.00890.8718±0.0100
TAPER87.01±1.0878.50±3.7293.77±2.2891.04±2.7673.97±2.140.9243±0.00780.9290±0.0072
CMI-9589CMAGN79.32±0.6280.94±2.0877.70±2.4778.45±1.5358.72±1.210.8578±0.00260.8689±0.0057
MFERL74.64±2.9882.28±4.9566.99±10.7471.96±5.4850.29±4.810.8413±0.00940.8608±0.0077
TAPER89.96±1.0985.82±3.0294.09±1.4693.60±1.3780.24±1.980.9551±0.00560.9638±0.0048
CMI-20208CMAGN79.78±0.8385.50±1.9474.06±1.9676.75±1.1259.98±1.680.8733±0.00690.8761±0.0060
MFERL86.81±1.3485.13±2.2788.50±1.1888.10±1.1673.69±2.630.9342±0.01150.9448±0.0087
TAPER88.73±0.9186.93±2.5290.53±1.6890.21±1.4577.56±1.790.9495±0.00770.9586±0.0055

Under strict circRNA cold start, TAPER achieves higher mean AUC and AUPR than both competing methods on all three datasets. Relative to the strongest baseline for each dataset, its AUC improvements are 0.0640, 0.0973, and 0.0153 on CMI-9905, CMI-9589, and CMI-20208, respectively. The corresponding AUPR improvements are 0.0557, 0.0949, and 0.0138. These results support prediction for circRNAs excluded from fitting on three benchmarks.

Table R3. Five-fold comparison under the partially observed miRNA cold-start protocol. Held-out miRNAs are excluded from fitting, and support interactions are revealed only at inference.

DatasetMethodACC(%)SEN(%)SPE(%)PRE(%)MCC(%)AUCAUPR
CMI-9905CMAGN79.89±1.2882.81±3.6476.97±5.4578.46±3.3560.06±2.340.8654±0.00480.8798±0.0081
MFERL68.36±4.0977.67±13.5359.05±20.8767.76±9.9939.14±7.030.8085±0.01590.8441±0.0113
TAPER83.04±0.5979.74±2.8685.79±2.6882.49±2.2965.81±1.160.8924±0.00430.8969±0.0044
CMI-9589CMAGN80.42±1.1882.93±5.6277.90±6.2179.27±3.6461.24±2.100.8805±0.00560.8844±0.0116
MFERL72.55±2.8279.84±6.5465.27±9.3170.08±4.2445.97±4.970.8204±0.02160.8406±0.0220
TAPER84.43±0.7682.93±4.6085.93±5.0385.79±3.7869.14±1.450.9295±0.00460.9312±0.0067
CMI-20208CMAGN80.08±0.5980.78±3.8479.38±3.8879.79±2.2960.30±1.140.8781±0.00740.8778±0.0112
MFERL79.43±0.9081.25±5.3177.62±4.4778.57±2.3559.13±1.760.8783±0.01100.8898±0.0102
TAPER84.89±1.4485.38±3.5684.39±4.6284.73±3.3669.94±2.840.9293±0.00770.9321±0.0065

For partially observed miRNA cold start, TAPER achieves higher mean AUC and AUPR than both baselines on all three datasets. Compared with CMAGN, the stronger baseline on CMI-9905 and CMI-9589, the AUC improvements are 0.0270 and 0.0490, with AUPR improvements of 0.0171 and 0.0468. On CMI-20208, TAPER exceeds MFERL by 0.0510 in AUC and 0.0423 in AUPR.

The extended evaluation quantifies the performance gap relative to pair-level prediction. TAPER’s miRNA cold-start AUC values are 0.8924, 0.9295, and 0.9293 across the three datasets. These represent decreases of 0.0797, 0.0532, and 0.0398 from the original pair-level results. We will discuss this gap when interpreting performance on miRNAs excluded from fitting.

We will incorporate these comparisons into the experimental section, distinguishing strict circRNA from partially observed miRNA cold start and explaining inference-time information. The appendix will provide individual fold results and document RNA-level partitions, negative sampling, support/query construction, masking ratios, and interaction-derived inputs within each fold. These additions will clarify generalization beyond pair-level prediction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.