acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Masked Reconstruction for Multi-Scale circRNA-miRNA Representation Learning

Abstract

Predicting circRNA-miRNA interactions requires learning from incomplete association networks and RNA sequence-derived evidence. We present SMASH, a framework thatcombines multi-scale hypergraph learning with adaptive masked reconstruction. SMASH represents each RNA using its training association profile and fused sequence, structural, and composition similarities. These features are propagated through interaction-derived and similarity-derived structures. The model retains representations from successive propagation depths and learns RNA-specific weights to combine them, allowing each RNA to draw on different neighborhood ranges. To provide auxiliary supervision, a reconstruction branch recovers masked RNA representations. Its masking probabilities are adjusted using historical reconstruction errors, assigning higher probabilities to RNAs with larger errors while maintaining the average masking probability. The encoder is trained jointly on interaction prediction and masked reconstruction. The resulting RNA representations are then used by a CatBoost classifier to score candidate interactions. Under association-level five-fold cross-validation, SMASH achieves mean AUC/AUPR values of 0.9470/0.9487, 0.9604/0.9596, and 0.9572/0.9566 on CMI-9905, CMI-9589, and CMI-20208, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
circRNA–miRNA interaction, hypergraph neural networks, multi-scale representation learning, masked representation reconstruction

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

This paper proposes SMASH for circRNA-miRNA interaction prediction. The model combines multi-scale hypergraph representation learning with adaptive masked reconstruction. RNA representations are constructed from training association profiles together with sequence, structural, and composition similarities, and different propagation depths are adaptively combined at the node level. The reconstruction branch adjusts node masking probabilities according to historical reconstruction errors, while the learned RNA representations are used by a CatBoost classifier for final prediction. Experiments on three CMI datasets show competitive AUC and AUPR results, together with ablation, cold-start, negative-sampling, parameter, and efficiency analyses.

Critical strengths

S1. The method combines multi-scale hypergraph learning and masked reconstruction in a coherent framework, and the adaptive masking mechanism is reasonably motivated by heterogeneous reconstruction difficulty across RNA nodes.

S2. The experimental study is relatively comprehensive. The paper reports three datasets, component ablations, parameter analyses, cold-start experiments, negative-sampling controls, and computational cost.

Critical weaknesses

W1. The baseline comparison is not fully controlled, as several results are taken from original papers using different splits, negative sampling, and preprocessing. This makes the reported performance gains difficult to attribute solely to SMASH.

W2. The benefit of adaptive masking is not sufficiently isolated. The ablation removes the whole reconstruction branch rather than directly comparing adaptive and uniform masking under the same architecture.

W3. The cold-start evaluation is transductive rather than fully inductive, since held-out RNAs remain in the graph and their similarity features are available during representation learning.

W4. The dense adaptive propagation has O(N3)O(N^3) time and O(N2)O(N^2) memory complexity, which may limit scalability to substantially larger RNA datasets.

Questions for authors

Q1. Could the authors rerun the strongest competing methods using exactly the same folds, negative samples, and preprocessing as SMASH? This would make the performance comparison more convincing.

Q2. Could the authors provide a direct adaptive-masking versus uniform-masking comparison using the same reconstruction architecture and masking ratio?

Q3. How does SMASH perform in a fully inductive setting where test circRNAs or miRNAs are completely absent during representation learning and inserted only at inference time?

Q4. How sensitive are the results to the IntaRNA-based negative sampling strategy? Would the ranking of SMASH and competing methods remain similar under alternative negative-sampling schemes?

Q5. The final predictor uses CatBoost on top of learned embeddings. Could the authors report the performance of the neural prediction head alone, or a simpler classifier, to clarify how much of the final gain comes from representation learning versus the downstream classifier?

8 comments

We sincerely thank you for your thoughtful review and for recognizing SMASH's coherent framework and comprehensive evaluation. We addressed your concerns through matched baseline reruns, adaptive-versus-uniform masking controls, fully inductive cold-start tests, alternative negative-sampling and classifier comparisons, and scalability measurements. We hope these new results address your questions, and would be grateful if you would consider revisiting your rating. Thank you again for your valuable feedback.

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

Thank you for highlighting the limitations of the original baseline comparison. We agree that differences in data partitions, negative-sample construction, and preprocessing can affect the interpretation of reported performance gains. To address this concern, we have rerun MRHRL, BEROLECMI, and BioDGW-CMI on CMI-9905, CMI-9589, and CMI-20208 under the evaluation protocol used for SMASH. The resulting comparisons are presented in Tables R1-R3.

For each dataset, all four methods use exactly the same positive pairs, IntaRNA-screened negative pairs, and five-fold assignments. Thus, training and held-out pairs are identical across methods in each fold, controlling the sample composition and partitions that differed among the original literature-reported results.

Shared preprocessing includes data cleaning, RNA index correspondence, and preparation of common base inputs. Each baseline retains its own representation construction and prediction mechanism, preserving its model definition while using a common processed dataset.

Table R1. Five-fold comparison on CMI-9905 using identical samples, fold assignments, and shared data preprocessing.

MethodAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
BioDGW-CMI0.8953±0.00610.8964±0.006182.16±0.8184.33±1.0979.83±0.7281.80±0.6664.27±1.64
BEROLECMI0.9043±0.00470.9022±0.004183.83±0.9184.20±1.2883.44±0.6384.54±0.6667.64±1.81
MRHRL0.9309±0.00570.9342±0.004885.82±0.6788.57±1.5782.86±1.5584.76±1.0271.65±1.35
SMASH0.9469±0.00320.9487±0.002588.47±0.5888.47±0.7688.48±0.9289.20±0.7776.93±1.16

Table R2. Five-fold comparison on CMI-9589 using identical samples, fold assignments, and shared data preprocessing.

MethodAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
BioDGW-CMI0.9459±0.00240.9419±0.002987.90±0.3887.53±0.5488.26±0.7888.18±0.6775.79±0.75
BEROLECMI0.9440±0.00300.9412±0.004187.32±0.3386.90±0.5387.74±0.9187.64±0.7674.65±0.67
MRHRL0.9441±0.00220.9390±0.003287.79±0.4687.32±0.5988.26±0.8388.15±0.7375.59±0.93
SMASH0.9604±0.00370.9596±0.005290.55±0.6792.03±0.9888.90±0.5090.21±0.4681.04±1.36

Table R3. Five-fold comparison on CMI-20208 using identical samples, fold assignments, and shared data preprocessing.

MethodAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
BioDGW-CMI0.8863±0.00320.8915±0.006981.88±0.8477.32±1.3786.45±1.5485.11±1.4264.05±1.72
BEROLECMI0.9083±0.00200.9034±0.003484.10±0.2383.56±0.8784.63±0.7184.47±0.4968.21±0.46
MRHRL0.9283±0.00250.9237±0.003885.84±0.4785.22±1.2186.45±1.0886.29±0.8471.70±0.94
SMASH0.9572±0.00520.9566±0.006790.10±0.3991.52±0.4588.51±0.9089.86±0.7080.14±0.79

Under this matched protocol, SMASH achieves the highest mean AUC and AUPR among the evaluated methods on all three datasets. On CMI-9905, MRHRL is the strongest competing method on these two metrics. SMASH increases AUC from 0.9309 to 0.9469 and AUPR from 0.9342 to 0.9487, corresponding to absolute differences of 0.0160 and 0.0145. On CMI-9589, BioDGW-CMI has the highest baseline AUC and AUPR. The corresponding SMASH advantages are 0.0145 AUC and 0.0177 AUPR. On CMI-20208, SMASH exceeds MRHRL by 0.0289 AUC and 0.0329 AUPR. These results provide direct evidence for the observed ranking-performance advantages under a common evaluation protocol.

The threshold-based metrics provide additional context. On CMI-9905, MRHRL has slightly higher mean sensitivity than SMASH, at 88.57% and 88.47%, respectively. SMASH has higher mean accuracy, specificity, precision, and MCC on that dataset. On CMI-9589 and CMI-20208, SMASH has the highest reported mean for each metric shown in the tables.

For CMI-20208, we evaluated the same three competing methods used in the other two dataset comparisons. We could not obtain public implementations of WSCD and CA-CMA, so these two methods were excluded from the controlled reruns. We will present the matched results for MRHRL, BEROLECMI, and BioDGW-CMI and explain the resulting baseline coverage in its caption and accompanying text.

We will update the main experimental section and Tables 2 and 17-19 with these matched comparisons. Captions will identify rerun results and the shared samples, fold assignments, and preprocessing.

Response to W2 and Q2: Adaptive vs. Uniform Masking under a Controlled Architecture

Thank you for emphasizing the need to isolate the contribution of adaptive masking. We agree that removing the entire GraphMAE branch evaluates the contribution of the reconstruction branch as a whole. A direct comparison between adaptive and uniform masking is needed to evaluate the masking policy within that branch. We have therefore supplemented the ablation study with a uniform-masking variant of SMASH that retains the same reconstruction architecture.

In this comparison, both configurations retain the reconstruction encoder, decoder, and masking-token mechanism. The reconstruction objective and its loss weight remain unchanged. The remaining SMASH components, downstream CatBoost classifier, and training settings are also held fixed. Both configurations use the same positive and negative samples and exactly the same five-fold assignments on CMI-9905. The experimental change is confined to the rule used to select masked nodes.

Both configurations use the same average masking probability, ρ=0.5\rho = 0.5. In the uniform variant, each node is independently selected for masking with probability 0.5 at each masking step. In the adaptive variant, node-selection probabilities are updated using historical reconstruction errors while retaining the same configured average masking probability. This matches the expected masking budget across the two configurations. Under Bernoulli sampling, the realized number of masked nodes can vary between iterations. We will clarify this distinction in the revised description of the masking procedure.

Table R4. Direct comparison of adaptive and uniform masking within the same SMASH reconstruction architecture on CMI-9905 at ρ=0.5\rho = 0.5.

ConfigurationAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Uniform masking0.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49
Adaptive masking0.9469±0.00320.9487±0.002588.47±0.5888.47±0.7688.48±0.9289.20±0.7776.93±1.16
Difference0.03230.03613.756.291.212.617.39

With this controlled comparison, adaptive masking achieves an AUC of 0.9469 ± 0.0032, compared with 0.9146 ± 0.0065 for uniform masking. AUPR increases from 0.9126 ± 0.0084 to 0.9487 ± 0.0025. These correspond to absolute improvements of 0.0323 AUC and 0.0361 AUPR. Accuracy increases by 3.75 percentage points, and MCC increases by 7.39 percentage points. Adaptive masking also achieves higher mean sensitivity, specificity, and precision under the same evaluation conditions.

The uniform-masking variant retains the reconstruction objective and the capacity of the reconstruction network. Its comparison with the original SMASH configuration therefore directly evaluates the effect of the masking policy within the same model. The observed differences support the empirical benefit of error-feedback masking on CMI-9905. We will state this conclusion within the scope of the evaluated dataset and configuration.

For completeness, Table R5 provides the five-fold results of the newly added uniform-masking experiment.

Table R5. Five-fold performance of SMASH with uniform Bernoulli masking at ρ=0.5\rho = 0.5 on CMI-9905.

FoldAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Fold 10.92340.924385.4682.6988.2487.5571.03
Fold 20.90780.903383.5780.3186.8285.9167.28
Fold 30.91030.906184.4082.1886.6286.0068.87
Fold 40.91910.916685.2383.4487.0386.5470.51
Fold 50.91260.912984.9682.2887.6386.9370.01
Mean ± SD0.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49

We will add the uniform-masking configuration to the main ablation table and revise the accompanying discussion to distinguish the questions answered by the two experiments. The branch-removal ablation measures the contribution of the reconstruction branch. The adaptive-versus-uniform comparison measures the contribution of the masking policy while preserving that branch. The revised text will report the direct AUC and AUPR differences and use this comparison as the evidence supporting the benefit of adaptive masking.

We will also expand the appendix to include the complete fold-level results for both masking configurations and the corresponding mean and standard deviation. The experimental description will specify the shared reconstruction architecture, the masking probability, and the settings held fixed across the comparison. It will explicitly identify the common positive and negative samples and fold assignments.

Response to W3 and Q3: Fully Inductive Cold-Start Evaluation

Thank you for identifying the distinction between our original cold-start evaluation and fully inductive prediction. We agree that the original RNA-disjoint setting is transductive: held-out RNAs remained in the graph, and their similarity features were available during representation learning, although their interaction labels were withheld. This evaluates prediction for a fixed node set rather than entirely unseen RNAs. We will label these results transductive RNA-disjoint evaluation in the main text.

To address this request, we conducted supplementary fully inductive experiments for unseen circRNAs and miRNAs on CMI-9905, CMI-9589, and CMI-20208. These evaluate an inductive SMASH extension using training RNAs as fixed feature references and uniform masking. This differs from the main method's error-feedback masking.

For unseen-circRNA evaluation, circRNAs are partitioned into five outer folds. In each fold, held-out circRNAs are excluded from the representation-learning graph and input feature coordinates, as well as training similarity-fusion statistics and the reconstruction objective. All positive and negative pairs involving them are excluded from encoder training and CatBoost fitting, including internal validation. Training uses the remaining circRNAs and available miRNAs. The unseen-miRNA experiment applies the same restrictions with RNA types exchanged.

Because similarity and association profiles depend on the node set, each outer fold uses fixed reference coordinates defined by training RNAs. Training and test nodes are represented relative to these same references. Similarity-fusion coefficients are calculated using training-node submatrices and kept fixed at inference. Association-derived inputs and the initial graph use only training associations. Unavailable association entries for new RNAs are zero; their true labels are reserved for evaluation.

After training, held-out RNAs are inserted in a batch, and their features become available to the frozen encoder. Feature-derived connections among newly inserted nodes are permitted during inference. Encoder and CatBoost parameters remain frozen, without gradient updates or classifier refitting. This protocol evaluates batch insertion of previously unseen RNAs. Fixed reference coordinates preserve the feature dimensions required by the trained model.

Table R6. Fully inductive prediction for unseen circRNAs.

DatasetAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
CMI-99050.8199±0.02450.8333±0.026375.46±1.8777.03±3.8573.77±2.9775.99±1.8950.91±3.71
CMI-95890.8168±0.02360.8324±0.022074.93±2.3877.41±2.8572.26±4.8775.13±3.2549.83±4.83
CMI-202080.8482±0.01500.8548±0.016676.25±0.9384.40±3.6867.48±4.8173.73±2.0652.94±1.72

Table R7. Fully inductive prediction for unseen miRNAs.

DatasetAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
CMI-99050.6870±0.03920.7309±0.032364.19±1.7748.36±4.2381.21±4.0873.67±3.5231.27±3.76
CMI-95890.7011±0.03000.7269±0.023064.95±2.4049.15±6.9681.94±3.0674.57±1.5432.82±3.74
CMI-202080.7079±0.01560.7310±0.017264.84±0.9248.38±3.4282.53±3.5275.05±2.7432.82±2.19

For unseen circRNAs, mean AUC ranges from 0.8168 to 0.8482 across the three datasets, and mean AUPR from 0.8324 to 0.8548. For unseen miRNAs, the corresponding AUC range is 0.6870–0.7079, with AUPR values of 0.7269–0.7310. These estimates apply when evaluated RNA identities and their features are unavailable during representation learning and become available only at inference.

The unseen-miRNA setting shows lower predictive performance. At the fixed threshold of 0.5, mean sensitivity ranges from 48.36% to 49.15%, versus specificity of 81.21% to 82.53%. This indicates that many positive interactions involving unseen miRNAs remain undetected at this operating point. We will acknowledge this limitation in the discussion.

The original transductive and supplementary inductive results will be reported separately. Their comparison also involves changes in node availability, input representation, masking configuration, and candidate-negative construction. Therefore, score differences cannot be attributed solely to removing test RNAs from representation learning.

We will revise the main experimental section to distinguish transductive RNA-disjoint evaluation from inference on newly introduced RNAs. The main text will summarize supplementary AUC and AUPR results and acknowledge weaker performance for unseen miRNAs. Appendix G will include complete fold-level metrics and, for each fold, counts of held-out RNAs and evaluated positive and negative pairs. It will document fixed training-reference construction, batch insertion, and frozen inference conditions. Claims will remain within this experimental scope and distinguish these results from the original adaptive-masking configuration.

Reviewer_rHUnOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SMASH, a framework that combines multi-scale hypergraph learning with adaptive masked reconstruction to predict circRNA-miRNA interactions. The SMASH method represents each RNA using its training association profile and fused sequence, structural, and composition similarities. And then these similarity informations are propagated through interaction-derived and similarity-derived structures. The main contribution is that the model retains representations from successive propagation depths and learns RNA-specific weights to combine them, allowing each RNA to draw on different neighborhood ranges. At the same time a reconstruction branch recovers the masked RNA representations with masking probabilities adjusted using historical reconstruction errors. As a result, under association-level five-fold cross-validation, SMASH achieves a good performance.

Critical strengths

This paper addresses two limitations of existing CMI prediction methods: structural priors that are rigid and a propagation depth that is fixed. It learns RNA-specific weights over successive propagation depths, so that each RNA draws on a different neighborhood range rather than sharing one receptive field across all nodes. A reconstruction branch adds auxiliary supervision. There, each RNA's masking probability is set from its past reconstruction errors, so RNAs with larger errors are masked more often while the average masking probability stays unchanged.

Critical weaknesses

The pipeline of SMASH has two minor weaknesses. Its encoder and CatBoost classifier are trained in separate stages rather than end-to-end, so they are not optimized against the same objective. Its reconstruction target is produced by the trainable backbone and recomputed every iteration. Gradients flow back into that target, which can weaken the reconstruction constraint, and the adaptive mechanism may degenerate. It squares cosine loss averages over adaptively selected nodes without inverse-probability correction, so the error-feedback sampler is biased.

Questions for authors

The authors report that removing GraphMAE causes the largest performance drop, but they give numeric deltas only for channel attention and multi-scale HGCN. Please provide a complete ablation table, includes mean and standard deviation across folds. In addition, please add an ablation in which the adaptive sampler is replaced by uniform masking.

How many hyperparameters does the method require, and how were they selected? Please confirm that the same hyperparameters are used for all datasets, and show the corresponding sensitivity analysis.

Finally, regarding the scalability limitation noted above: what runtime and peak memory were measured, and how do both grow with the number of RNAs?

7 comments

We sincerely thank you for your thoughtful review and encouraging assessment of SMASH. We are especially grateful for your recognition of RNA-specific multi-scale propagation and error-feedback reconstruction. Your constructive comments have helped us improve the work. We have clarified the two-stage optimization and reconstruction gradients, and added inverse-probability weighting controls, complete ablations, hyperparameter analyses, and scalability measurements. Thank you again for your time, insight, and support.

Response to W1: Two-Stage Optimization, Reconstruction Gradients, and Sampling Bias (Part 1/2)

We thank you for examining the optimization procedure and the reconstruction objective. We agree that the distinction between supervised representation learning and subsequent classifier fitting should be stated more clearly. During representation learning, the encoder is optimized using

L=LBCE+0.1Lrec\mathcal{L}=\mathcal{L}_{\mathrm{BCE}}+0.1\mathcal{L}_{\mathrm{rec}}

where the association prediction loss directly supervises the learned representations. CatBoost is subsequently fitted to the fixed pair representations using binary log loss. Both stages address association prediction using the training labels, but they optimize different model parameters. CatBoost fitting provides no further update to the encoder. We will clarify this separation in the Method section. The additional reconstruction experiments described below examine the target-gradient and sampling-weight choices within the current architecture.

To investigate these choices, we evaluated five configurations on CMI-9905. These comprise the original adaptive formulation, uniform masking, target stop-gradient variants under both masking schemes, and an adaptive variant with inverse-probability-weighted reconstruction. The configurations use the same evaluation partitions and retain the remaining experimental settings. The mean masking probability is 0.5 throughout. Table R1 summarizes the results, and Tables R2-R6 provide the complete fold-level measurements.

Table R1. Comparison of masking, reconstruction-target gradients, and inverse-probability weighting on CMI-9905.

ConfigurationMaskingTarget receives gradientsIPWAUCAUPR
Original SMASHAdaptiveYesNo0.9469±0.00320.9487±0.0025
Uniform maskingUniformYesNo0.9146±0.00650.9126±0.0084
Adaptive masking with target stop-gradientAdaptiveNoNo0.9157±0.00700.9152±0.0053
Uniform masking with target stop-gradientUniformNoNo0.8901±0.00420.8974±0.0044
Adaptive masking with IPWAdaptiveYesYes0.9449±0.00340.9456±0.0045

Regarding the reconstruction target, the original formulation recomputes the backbone representation at each iteration and allows the reconstruction loss to propagate through both the reconstructed representation and its target. The stop-gradient operation used in the error-history update applies to the feedback state. The target representation retains its gradient path. We agree that allowing the target to change during optimization creates a potential route for reducing reconstruction error through changes in the target itself. We will make this distinction explicit in the revised objective description.

The new target stop-gradient controls directly examine the predictive effect of this gradient path. Under adaptive masking, stopping target gradients reduces mean AUC from 0.9469 to 0.9157 and mean AUPR from 0.9487 to 0.9152. The corresponding decreases are 0.0312 and 0.0335. Under uniform masking, the decreases are 0.0245 AUC and 0.0152 AUPR. Thus, retaining target-side gradients is associated with higher downstream scores under both masking schemes in these experiments.

Adaptive masking also retains higher mean scores when target gradients are stopped. Its mean AUC and AUPR exceed the corresponding uniform stop-gradient configuration by 0.0255 and 0.0178. The predictive advantage of adaptive selection is therefore observed under both target-gradient settings. These comparisons characterize the consequences of the gradient choice for prediction. We will preserve this distinction in the revised discussion.

Table R2. Five-fold performance of the original SMASH configuration on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.94300.945587.5288.0486.9787.9075.01
Fold 20.94450.947188.3687.4389.3689.8376.74
Fold 30.94860.948888.7088.5488.8789.5477.38
Fold 40.95070.951988.8389.3088.3389.1677.63
Fold 50.94830.950288.9689.0588.8789.5977.90
Mean ± SD0.9469±0.00320.9487±0.002588.47±0.5888.47±0.7688.48±0.9289.20±0.7776.93±1.16

Response to W1: Two-Stage Optimization, Reconstruction Gradients, and Sampling Bias (Part 2/2)

Table R3. Five-fold performance of uniform masking with target-side gradients on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.92340.924385.4682.6988.2487.5571.03
Fold 20.90780.903383.5780.3186.8285.9167.28
Fold 30.91030.906184.4082.1886.6286.0068.87
Fold 40.91910.916685.2383.4487.0386.5470.51
Fold 50.91260.912984.9682.2887.6386.9370.01
Mean ± SD0.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49

Table R4. Five-fold performance of adaptive masking with target stop-gradient on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.92130.915385.0184.1585.8785.6270.03
Fold 20.91330.914284.2381.0287.4386.5768.59
Fold 30.90450.907183.2281.0285.4184.7466.50
Fold 40.91980.918684.4582.8486.0785.6068.94
Fold 50.91950.920984.5583.1985.9285.5269.13
Mean ± SD0.9157±0.00700.9152±0.005384.29±0.6682.44±1.3986.14±0.7685.61±0.6568.64±1.31

Table R5. Five-fold performance of uniform masking with target stop-gradient on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.89360.901183.1981.1785.2184.5966.43
Fold 20.88640.893782.2679.6684.8684.0364.60
Fold 30.88550.891781.8580.4183.2982.8063.73
Fold 40.89510.899883.2481.7884.7084.2466.51
Fold 50.88980.900783.2982.3384.2583.9466.59
Mean ± SD0.8901±0.00420.8974±0.004482.77±0.6781.07±1.0684.46±0.7483.92±0.6765.57±1.32

Regarding inverse-probability correction, we agree that the original loss changes the effective contribution of nodes through adaptive selection. For each RNA type, the original reconstruction term is

Lrec,r=1max⁡(1,∣Ωr∣)∑i∈Ωrℓr,i,ℓr,i=(1−cos⁡(Ur[i,:],Ur[i,:]))2.\mathcal{L}_{\mathrm{rec},r}=\frac{1}{\max\left(1,\left|\Omega_r\right|\right)}\sum_{i\in\Omega_r}\ell_{r,i},\qquad\ell_{r,i}=\left(1-\cos\left(U_r[i,:],U_r[i,:]\right)\right)^2.

The square is applied to each node's cosine error before averaging. Nodes with larger historical reconstruction errors receive higher selection probabilities and consequently have more opportunities to contribute to this term. This formulation defines a curriculum reconstruction objective conditioned on the error history. It provides no established guarantee of unbiased estimation of an equally weighted reconstruction risk over all nodes. We will clarify the intended objective and the absence of inverse-probability correction in the Method section.

To evaluate the effect of node weighting, the additional IPW configuration uses

LIPW,r=1nr∑i∈Ωrℓr,ipr,i\mathcal{L}_{\mathrm{IPW},r}=\frac{1}{n_r}\sum_{i\in\Omega_r}\frac{\ell_{r,i}}{p_{r,i}}

where pr,ip_{r,i} is the actual Bernoulli inclusion probability used for that iteration. The reconstruction target continues to receive gradients. The feedback state is updated using the original, unweighted per-node errors, so this control changes the loss aggregation while retaining the error-history update rule.

Table R6. Five-fold performance of adaptive masking with inverse-probability-weighted reconstruction on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.94100.939987.9988.2987.6888.5175.96
Fold 20.94250.942888.0287.1388.9889.4876.06
Fold 30.94850.950588.8389.2588.3889.2077.63
Fold 40.94820.949788.5289.3087.6888.6377.00
Fold 50.94450.945287.8689.5086.1087.3875.70
Mean ± SD0.9449±0.00340.9456±0.004588.24±0.4188.69±0.9987.76±1.0888.64±0.8176.47±0.81

The IPW configuration achieves a mean AUC of 0.9449 and a mean AUPR of 0.9456. The corresponding differences from the original SMASH configuration are 0.0020 AUC and 0.0031 AUPR. The weighted adaptive configuration also exceeds uniform masking with target gradients by 0.0303 AUC and 0.0330 AUPR. Thus, higher mean prediction scores than uniform masking are observed after introducing node-level inverse-probability weighting.

The interpretation of node-level IPW also requires care in graph reconstruction. A node’s reconstruction error depends on the complete masking configuration. Weighting selected nodes by their inverse inclusion probabilities therefore does not automatically recover the full reconstruction objective under uniform masking. We will describe this configuration as a node-weighting sensitivity control and specify its normalization and estimation scope. The original adaptive loss will be presented explicitly as a curriculum objective with unequal node contributions. We will revise the Method section to distinguish the two training stages and identify the gradient paths in the reconstruction objective and feedback-state update.

Response to Q1: Complete Ablation Results and Uniform Masking Control

We thank you for requesting the complete ablation results. The original discussion reported numerical decreases for channel attention and multi-scale HGCN, while describing the GraphMAE ablation qualitatively. Omitting the corresponding numerical decrease for GraphMAE made the relative contribution of the reconstruction branch difficult to assess. We have now completed the ablation summary and added the requested uniform-masking control. The tables below report the mean and sample standard deviation across five folds on CMI-9905.

For the additional sampling ablation, we replace the error-feedback adaptive sampler with independent uniform Bernoulli masking. Each RNA node is selected with a constant probability of ρ=0.5\rho = 0.5 throughout training. The GraphMAE reconstruction branch remains active, and the remaining model components and training settings are retained. This comparison evaluates the effect of the node-selection rule within the same reconstruction framework. Both sampling configurations use the same mean masking probability, which controls the expected number of masked nodes. The realized number can vary between iterations because masking is sampled independently for each node.

The uniform-masking configuration achieves an AUC of 0.9146 ± 0.0065 and an AUPR of 0.9126 ± 0.0084. Table R7 provides its complete five-fold results.

Table R7. Five-fold performance of uniform Bernoulli masking at ρ=0.5\rho = 0.5 on CMI-9905.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
Fold 10.92340.924385.4682.6988.2487.5571.03
Fold 20.90780.903383.5780.3186.8285.9167.28
Fold 30.91030.906184.4082.1886.6286.0068.87
Fold 40.91910.916685.2383.4487.0386.5470.51
Fold 50.91260.912984.9682.2887.6386.9370.01
Mean ± SD0.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49

Table R8 combines this new control with the complete model and the three component ablations. CAM denotes channel attention, and MHGCN denotes multi-scale hypergraph convolution. The w/o GraphMAE configuration removes the auxiliary reconstruction branch, whereas the uniform-masking configuration retains that branch and replaces its adaptive node-selection rule.

Table R8. Complete ablation results on CMI-9905.

ConfigurationAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
w/o CAM0.9041±0.00550.9069±0.006484.14±0.8283.75±1.1784.56±0.9685.11±0.8468.29±1.64
w/o MHGCN0.9154±0.00410.9171±0.004484.69±0.5483.58±1.0385.90±1.1386.45±0.8869.43±1.09
w/o GraphMAE0.8871±0.00470.8973±0.005182.04±0.5780.12±0.9383.95±0.9383.32±0.7964.12±1.14
Uniform masking ρ=0.5\rho = 0.50.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49
SMASH0.9469±0.00320.9487±0.002588.47±0.5888.47±0.7688.48±0.9289.20±0.7776.93±1.16

Removing GraphMAE reduces mean AUC from 0.9469 to 0.8871 and mean AUPR from 0.9487 to 0.8973. These changes correspond to decreases of 0.0598 AUC and 0.0514 AUPR. Removing CAM produces decreases of 0.0428 AUC and 0.0418 AUPR. Removing MHGCN produces decreases of 0.0315 AUC and 0.0316 AUPR. GraphMAE removal therefore yields the largest mean decrease in both metrics among the evaluated component ablations. We will revise the original paragraph to report all three numerical decreases explicitly and make this ranking directly verifiable.

Replacing adaptive sampling with uniform masking reduces mean AUC by 0.0323 and mean AUPR by 0.0361. The complete adaptive configuration also obtains higher mean values for the other reported metrics in this comparison. These results provide empirical evidence for the contribution of the error-feedback selection rule on CMI-9905. The GraphMAE removal experiment evaluates the contribution of the entire auxiliary branch, while the uniform-masking experiment evaluates the sampling rule within the retained branch. We will explain these distinct experimental purposes in the revised ablation subsection.

We will add the complete ablation summary, including the uniform-masking configuration, to the main experimental section. The accompanying text and figure captions will report the numerical decreases consistently and identify GraphMAE removal as producing the largest mean AUC and AUPR decreases among the evaluated component ablations. The revised appendix will provide the complete fold-level results for all configurations, and specify the uniform Bernoulli sampling rule and the settings held fixed in the comparison.

Reviewer_CVcYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SMASH for circRNA–miRNA interaction (CMI) prediction. Three within-type RNA similarity matrices (sequence edit distance, RNAfold structural similarity, dinucleotide-composition Spearman correlation) are fused by a parameter-free channel attention. A joint hypergraph is built from the training association matrix and processed by a multi-scale HGCN whose per-depth outputs are combined with node-specific scale attention; a learnable adaptive-propagation branch modifies the operator. An auxiliary GraphMAE-style branch reconstructs masked backbone representations, with the paper's central novelty being an error-feedback mask sampler: a detached EMA of per-node reconstruction errors modulates per-node Bernoulli mask probabilities. Pairs are scored by CatBoost on frozen embeddings. Evaluation is association-level 5-fold CV with 1:1 IntaRNA-screened negatives on CMI-9905/9589/20208, reporting AUC 0.9470/0.9604/0.9572, numerically above published numbers of MRHRL, BEROLECMI, and BioDGW-CMI. Ablations cover channel attention, multi-scale HGCN, and GraphMAE, plus parameter analyses, a random-negative control, RNA-disjoint cold-start evaluation, and a literature case study (19/20 top-20 pairs verified).

Critical strengths

S1. The evaluation reporting is thorough for this research topic. The authors include a random-negative control quantifying sensitivity to the negative-sampling protocol (AUC 0.9470 → 0.9363 on CMI-9905), an RNA-disjoint cold-start evaluation that exposes how much easier association-level CV is than the practical setting (AUC drops to 0.71–0.83), plus runtime, memory, and per-fold results. This raises my confidence that the reported within-protocol numbers are reproducible, even though the cold-start results themselves document that the headline CV setting overstates practical performance.

S2. The manuscript is written with atypical precision for this genre: explicit conventions for edge cases (empty hyperedges, degree conventions), restrained claims ("numerically higher" rather than "significantly better"), and an explicit AI-use statement. This made the paper straightforward to check.

Critical weaknesses

W1. The central claimed contribution — the error-feedback adaptive mask sampler — is never compared against the obvious baseline. The ablations remove channel attention, multi-scale HGCN, and the entire GraphMAE branch, but nowhere is adaptive masking compared to uniform masking at the same ρ=0.5, even though Appendix C explicitly defines the uniform-masking configuration without reporting its results. Consequently, the titular claim is not experimentally supported: as written, the sampler's benefit could be zero. This is decisive because every other component is an established technique (GraphMAE, hypergraph convolution, SE-style fusion), and the authors themselves position adaptive masking as a variant of prior curriculum-masking work. The paper's incremental novelty rests entirely on this one unvalidated component.

W2. The comparison with state-of-the-art methods is not run under a unified protocol: baseline numbers are quoted from the original papers "under their respective evaluation protocols." The paper's own control shows performance is sensitive to negative construction (ΔAUC ≈ 0.011 between IntaRNA-screened and random negatives), which is the same order as the margins over the strongest baseline (0.009–0.038 AUC). Without rerunning baselines on the same folds with matched negative construction, and without significance testing, the claimed superiority cannot be established from this evidence.

Questions for authors

Q1. What is the result of uniform masking at ρ=0.5 with all other components fixed? If the adaptive sampler shows a statistically significant improvement over this baseline, my assessment of novelty would change substantially.

Q2. How were the hyperparameters (N_scales=3, mask ratio 0.5, T0, k=10, 0.7/0.3 operator mix) selected, and on which data split? Were the Appendix H analyses used to select the final configuration? If model selection drew on the evaluation folds, reported gains are likely optimistic.

Q3. Which negative-sampling protocol did each baseline (MRHRL, BEROLECMI, BioDGW-CMI) use in its original evaluation? This determines whether Tables 2 and 17–19 are interpretable as apples-to-apples comparisons.

Q4. (a) Are the 20 case-study pairs contained in the original 9,905 curated positives of CMI-9905? If so, the 19/20 verification rate largely reflects dataset curation rather than independent validation. (b) Please reconcile the ablation text: it states removing GraphMAE produces the largest decline, yet the largest reported drops (AUC −0.0429) are attributed to removing channel attention. Per-fold ablation numbers would resolve this.

6 comments

We thank you for the careful review and for recognizing SMASH's clarity and thorough reporting. We take your concerns about adaptive masking, baseline comparability, hyperparameter selection, and case-study validity seriously. We added uniform-masking controls, matched baseline reruns, statistical tests, and further analyses. We address each concern below and acknowledge remaining limitations. If these results resolve your concerns, we would be grateful if you would consider revisiting your rating. Thank you again.

Response to W1 and Q1:Uniform Masking Baseline and Statistical Significance

Thank you for identifying the comparison needed to evaluate the proposed sampler. Appendix C specifies uniform masking at ρ=0.5\rho=0.5. The ablation in Section 4.3 removes the entire reconstruction branch, so it leaves the contribution of the sampling rule unresolved. The control requested in Q1 retains the reconstruction branch and assigns every node the same masking probability of 0.5.

In a supplementary five-fold experiment on CMI-9905, we sampled each node independently with probability 0.5 at every training iteration. Uniform masking achieved 0.9146±0.0065 AUC and 0.9126±0.0084 AUPR. Table R1 reports the full results.

Table R1. Five-fold performance of uniform Bernoulli masking at ρ=0.5\rho=0.5 on CMI-9905 under the supplementary evaluation protocol.

FoldAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
10.92340.924385.4682.6988.2487.5571.03
20.90780.903383.5780.3186.8285.9167.28
30.91030.906184.4082.1886.6286.0068.87
40.91910.916685.2383.4487.0386.5470.51
50.91260.912984.9682.2887.6386.9370.01
Mean±SD0.9146±0.00650.9126±0.008484.72±0.7682.18±1.1687.27±0.6686.59±0.6869.54±1.49

Adaptive masking achieved higher AUC and AUPR in all five folds. For AUC, our primary metric, the average paired improvement was 0.0324, with an approximate corrected 95% CI of [0.01830, 0.04646] and a two-sided p=0.003089p=0.003089. We applied a variance-corrected paired t-test to the five fold-level differences to account for overlapping training sets. AUPR increased by 0.0361 on average. These results support an incremental benefit from the error-feedback sampler on CMI-9905 at the same expected masking rate.

Table R2. CMI-9905 scores and nominal variance-corrected calculations obtained by aligning folds by number.

MetricAdaptiveUniformCorrected 95% CICorrected t(4)t(4)Two-sided pp
AUC0.9469±0.00320.9146±0.0065[0.01830, 0.04646]6.38410.003089
AUPR0.9487±0.00250.9126±0.0084[0.01922, 0.05290]5.94590.004013

The contribution addressed in W1 is the feedback rule in Eqs. (20) and (27), which uses a node’s past reconstruction errors to set its future masking probability. A single reconstruction error depends on the current mask and the encoder parameters. To reduce the influence of these fluctuations, SMASH maintains an exponential moving average for each node as an estimate of its reconstruction difficulty. The average is updated only when that node is masked and its reconstruction error is observed. After warm-up, SMASH centers and normalizes the error histories separately for circRNAs and miRNAs. Nodes with larger historical errors receive higher masking probabilities and therefore contribute to the reconstruction loss more frequently. Centering keeps the mean masking probability at 0.5 within each RNA type, so the sampler redistributes reconstruction attempts across nodes while preserving the expected number of masked nodes.

SMASH directly computes each node’s Bernoulli probability from its stored error history. The sampler introduces no trainable parameters. We will clarify this distinction in Section 3.4 and state the contribution in terms of this specific feedback rule.

In the revision, we will add the matched comparison to Section 4.3. We will use the shared candidate-negative pairs with identical fold assignments, construct both graphs from the same training associations, and retain the same reconstruction loss and decoder re-masking in both variants. This will isolate the change from fixed Bernoulli probabilities to error-dependent probabilities. We will report the AUC and AUPR for both variants, with the paired AUC difference as the primary comparison. Appendix C will document the shared protocol, give the fold-level results, and report the 95% confidence interval and two-sided test with a variance correction for overlapping training folds.

Response to Q2: Hyperparameter Selection and Evaluation Bias

Thank you for asking how the hyperparameters were selected. The Appendix H analyses informed the submitted configuration. We used the same procedure for all five settings on the association-level five-fold partition of CMI-9905. We varied one parameter at a time and ranked the candidates by mean validation AUC across the five folds. This procedure selected Nscales=3N_{\mathrm{scales}}=3, ρ=0.5\rho=0.5, T0=5T_0=5 iterations, k=10k=10, and initial and adaptive operator weights of 0.7 and 0.3, respectively. In Appendix H, Table 12 compares scales from 1 to 5, and Table 14 compares masking probabilities of 0.1, 0.3, 0.5, 0.7, and 0.9. The selected values also had the highest mean AUPR in those two comparisons.

The same folds informed the original configuration choice and the reported CMI-9905 scores. We agree that selecting candidates on these folds can favor configurations that happen to perform well on them and can make the reported scores optimistic.

We therefore conducted four separate nested five-fold evaluations of the propagation scale, mean masking probability, adaptive propagation rank, and operator weights. These evaluations used the same CMI-9905 data and the same sample construction procedure as the main experiment. For each search, we varied the parameter under study and held the remaining settings fixed. Within each outer fold, we compared the candidates using only the four outer training folds. Each of those folds served once as the inner validation fold, with the remaining three used for training. We selected the value with the highest mean inner validation AUC, retrained on all four outer training folds, and evaluated the refitted model once on the outer test fold.

The inner searches recovered the submitted value for each of the four parameters in all five outer folds. Tables R3 through R6 report the scale, masking probability, propagation rank, and operator weight results, respectively. Each table reports the hyperparameters selected by inner validation for its respective search as well as the external test scores, which are based on the other settings used in this analysis.

In the revision, we will distinguish the original five-fold estimates from the supplementary outer test estimates in the main experimental section. Appendix C will describe the original and nested partitions, sample construction, and fold-specific hypergraph construction. We will add the original candidate sets and validation results for the warm-up, propagation rank, and operator weights to Appendix H. We will also add the supplementary tables and document the fixed settings and inner validation scores for the four nested searches.

Table R3. Outer test results for propagation scale NscalesN_{\mathrm{scales}}.

Outer foldSelected scaleAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Fold 130.94610.947688.9688.7989.1489.7977.91
Fold 230.94090.940387.3186.0788.6589.0874.67
Fold 330.94830.951688.8687.5890.2390.6077.76
Fold 430.96140.964090.0990.2689.9090.5880.15
Fold 530.94940.947589.3089.7588.8289.6278.57
Mean ± SD3 (5/5)0.9492±0.00760.9502±0.008788.90±1.0188.49±1.7089.35±0.6989.93±0.6577.81±2.00

Table R4. Outer test results for mean masking probability ρ\rho.

Outer foldSelected ratioAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Fold 10.50.94650.948089.1489.0089.3189.9578.27
Fold 20.50.94020.941087.5586.7288.4488.9775.11
Fold 30.50.94870.952388.8387.6890.0790.4777.70
Fold 40.50.96090.963190.1190.3689.8590.5480.20
Fold 50.50.94970.948489.0489.8588.1789.0978.05
Mean ± SD0.5 (5/5)0.9492±0.00750.9506±0.008188.93±0.9288.72±1.5189.17±0.8489.80±0.7477.87±1.82

Table R5. Outer test results for adaptive propagation rank kk.

Outer foldSelected rankAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Fold 1100.94760.947689.3389.3089.3690.0378.64
Fold 2100.94110.941287.5886.7788.4488.9875.17
Fold 3100.94910.952288.7087.5889.9090.3277.43
Fold 4100.96080.963589.9090.2189.5890.3079.78
Fold 5100.94900.948688.9989.3088.6589.4377.95
Mean ± SD10 (5/5)0.9495±0.00710.9506±0.008288.90±0.8688.63±1.4189.19±0.6289.81±0.5977.79±1.71

Table R6. Outer test results for initial and adaptive operator weights.

Outer foldSelected weightsAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
Fold 10.7/0.30.94650.947589.0188.9589.0989.7678.01
Fold 20.7/0.30.94100.940187.5286.9388.1788.7675.05
Fold 30.7/0.30.94880.952188.8087.5890.1290.5177.65
Fold 40.7/0.30.96070.963289.8589.9589.7490.4179.68
Fold 50.7/0.30.94850.948088.6289.4087.7988.7377.21
Mean ± SD0.7/0.3 (5/5)0.9491±0.00720.9502±0.008588.76±0.8488.56±1.2788.98±1.0089.63±0.8677.52±1.67

Response to W2 and Q3: Fair Baseline Comparison and Negative-Sampling Protocols (Part 1/2)

Thank you for identifying the limitations of the original comparison. We agree that negative-sample construction can affect measured performance. Our own control, showing an AUC difference of approximately 0.011 between the two negative-sampling settings, reinforces the need for a comparison using shared samples and identical evaluation partitions.

To address this concern, we have supplemented the original literature-based comparison with new five-fold experiments on CMI-9905, CMI-9589, and CMI-20208. On each dataset, MRHRL, BEROLECMI, BioDGW-CMI, and SMASH use exactly the same positive and negative pairs and the same fold assignments. Each method is therefore evaluated on the same held-out pairs in every corresponding fold. This controls the identities of the negative samples and the evaluation partitions in the new comparison.

Regarding Q3, the available materials establish different levels of detail about the original negative-sampling procedures for the three baselines.

For MRHRL, the author-released CMI-9589 example provides predefined positive and negative files for training, validation, and testing, with equal positive and negative sample counts. These files establish the sample composition of that example. We could not verify the procedure used to generate its negative pairs from the available materials. The original negative-sampling procedures for CMI-9905 and CMI-20208 also remain unverified.

For BEROLECMI, the reported procedure for CMI-9905 and CMI-9589 enumerates possible circRNA–miRNA pairs and excludes associations with a confidence score greater than zero in circBank. The authors then randomly select negative samples from the remaining pairs, retaining the same number of negatives as positives. This procedure combines circBank-based exclusion with random sampling and differs from the screening used in the SMASH main experiments.

For BioDGW-CMI, we could not verify either the original negative-sample generation procedure or the positive-to-negative sampling ratio from the accessible materials. We will label both items as unverified in the revised experimental protocol and appendix.

These uncertainties concern the original evaluations reported by the baseline authors. In our supplementary comparisons, all three baselines were rerun using the same concrete positive and negative sample lists and the same fold assignments as SMASH. We will clearly distinguish these supplementary evaluations from the original literature-reported results.

These differences limit the interpretation of the original numbers quoted under their respective evaluation protocols. The new experiments directly evaluate the three baselines using the dataset-specific negative pairs employed by SMASH. All four methods share the corresponding positive pairs and five-fold partitions. The resulting comparisons are reported in Tables R7-R9 Due to the limitation of the space for the reply, we will supplement the complete results of the five-fold cross-validation in the appendix.

Table R7. Five-fold comparison on CMI-9905 using shared samples and identical fold assignments.

MethodAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
BioDGW-CMI0.8953±0.00610.8965±0.006182.16±0.8184.33±1.0979.83±0.7281.80±0.6664.27±1.64
BEROLECMI0.9043±0.00470.9022±0.004083.83±0.9184.20±1.2883.44±0.6384.54±0.6667.64±1.81
MRHRL0.9309±0.00570.9341±0.004885.82±0.6788.57±1.5782.86±1.5584.77±1.0271.65±1.35
SMASH0.9469±0.00320.9487±0.002588.47±0.5888.47±0.7688.48±0.9289.20±0.7776.93±1.16

Table R8. Five-fold comparison on CMI-9589 using shared samples and identical fold assignments.

MethodAUCAUPRAcc(%)Sen(%)Spe(%)Pre(%)MCC(%)
BioDGW-CMI0.9459±0.00240.9419±0.002987.90±0.3887.53±0.5488.26±0.7888.18±0.6775.79±0.75
BEROLECMI0.9440±0.00300.9412±0.004187.32±0.3386.90±0.5387.74±0.9187.64±0.7674.65±0.67
MRHRL0.9441±0.00220.9390±0.003287.79±0.4687.32±0.5988.26±0.8388.15±0.7375.59±0.93
SMASH0.9604±0.00370.9596±0.005290.55±0.6792.03±0.9888.90±0.5090.21±0.4681.04±1.36

Table R9. Five-fold comparison on CMI-20208 using shared samples and identical fold assignments.

MethodAUCAUPRAcc (%)Sen (%)Spe (%)Pre (%)MCC (%)
BioDGW-CMI0.8863±0.00320.8915±0.006981.88±0.8477.32±1.3786.44±1.5485.11±1.4264.05±1.72
BEROLECMI0.9083±0.00200.9034±0.003484.10±0.2383.56±0.8784.63±0.7184.47±0.4968.21±0.46
MRHRL0.9283±0.00250.9237±0.003885.84±0.4785.22±1.2186.45±1.0886.29±0.8471.70±0.94
SMASH0.9572±0.00520.9566±0.006790.10±0.3991.52±0.4588.51±0.9089.86±0.7080.14±0.79

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.