acceptodds
Under review as a conference paper at ICLR 2027

SpliceMamba: Tissue-Conditioned State-Space Modeling of Cassette-Exon Inclusion

Abstract

Cassette exon inclusion is the predominant form of alternative splicing in mammals and a central mechanism of tissue-specific gene regulation. Predicting cassette exon inclusion remains challenging because cassette exons vary substantially in length, while their splicing outcomes are determined by surrounding cis-regulatory sequences, and existing sequence models are largely optimized for classifying individual splice sites rather than quantifying inclusion at the level of the complete splicing event. We introduce SpliceMamba, a selective state-space architecture for tissue-specific prediction of cassette exon Percent Spliced In (PSI). SpliceMamba constructs a unified token representation integrating nucleotide sequence, genomic annotation, and tissue identity; encodes four splice-site-centered windows using a shared selective state-space block; and integrates information across these windows to generate a single tissue-conditioned PSI prediction. On a 56-tissue benchmark, SpliceMamba achieves a Pearson correlation of 0.94 across event–tissue pairs using only 0.181M parameters, compared with 0.82–0.85 for splice-site models adapted to the same task. Because pooled correlation can be dominated by between-event variation, we additionally evaluate pairwise tissue-differential inclusion, for which SpliceMamba achieves a correlation of 0.50, compared with 0.23–0.37 for the baselines. In silico mutagenesis recovers canonical donor and acceptor motifs and identifies positions whose substitution shifts predicted inclusion between tissues. Ablations within a fixed model scaffold isolate the contributions of genomic features, cross-window integration, and the sequence encoder. Together, these results show that event-level, tissue-conditioned modeling of splice-site-proximal sequence provides an effective and parameter-efficient framework for quantitative PSI prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tissue-Conditioned Splicing, Sequence Modeling, State-Space Model

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

The paper proposes AsyGA, a personalized federated learning framework for multispectral palmprint verification. Following federated classifier training, each client uses a locally trained diffusion model to generate single-spectrum images for personalization. Contrastive learning, feature distillation from a frozen global model, fixed batch normalization statistics, and backbone freezing are intended to preserve cross-spectrum performance. The authors train on CASIA and evaluate open-set verification on PolyU, reporting substantial reductions in equal error rate and an improvement in true acceptance rate at a low false acceptance rate.

Critical strengths
  1. Event-level PSI is an appropriate target for cassette-exon modeling, and the four-window representation provides a compact way to combine the relevant splice-site contexts.
  2. The paper recognizes that pooled PSI correlation can obscure tissue-specific performance and includes a differential-PSI analysis as well as component ablations.
Critical weaknesses

The evaluation and model-selection protocol is not sufficiently sound. Appendix Table 8 explicitly says that baseline checkpoints are selected by Pearson correlation on the test split. This makes their reported test metrics unsuitable as unbiased baseline estimates, even though the direction of this bias may favor the baselines. The paper also refers to validation-based hyperparameter selection without defining a validation split. These issues require a corrected protocol and rerunning the comparisons. The positioning and comparisons omit a close predecessor. TSplice/MTSplice already modeled tissue-specific cassette-exon PSI using the same 61,823-exon ASCOT resource across 56 tissues, yet this paper does not provide a direct comparison or explain clearly what the new task formulation adds. Moreover, SpliceMamba receives conservation and other genomic annotations that its main sequence-model baselines do not; the large drop when conservation is removed shows why the reported advantage cannot be attributed to the Mamba architecture alone. The six-tissue dataset is GTEx-derived, while most ASCOT tissues also draw on GTEx, so it should not be described as an independent biological cohort.

Questions for authors

Was the test split used to select checkpoints or hyperparameters for any model? Please specify the validation partition and report results after model selection is performed without test-set access.

How does SpliceMamba compare with TSplice/MTSplice under a matched ASCOT split? What are the gains over feature-matched non-Mamba encoders and a genomic-feature-only model?

What overlap exists between the ASCOT and six-tissue GTEx data at the donor, event, and gene levels? A precise answer would clarify the extent of external validation.

5 comments
AuthorsOpenReview ↗

Response to Reviewer xeqM – Part1

Response to Weakness & Question 1:

The six-tissue dataset is GTEx-derived, while most ASCOT tissues also draw on GTEx, so it should not be described as an independent biological cohort.

What overlap exists between the ASCOT and six-tissue GTEx data at the donor, event, and gene levels? A precise answer would clarify the extent of external validation.

We thank the reviewer for requesting clarification regarding the overlap between ASCOT 56-tissue and GTEx 6-tissue datasets. To solve these concerns, we firstly summarized the overlaps at the donor, event, and gene levels in the Table below.

ASCOT TissueGTEx TissueDonor samplesShared ASCOT eventsShared GTEx eventsShared genes
LungLung3744,6304,6303,127
Atrial Appendage - HeartHeart (Atrial_Appendage)2185,5855,5843,568
Cerebellum - BrainBrain (Cerebellum)1455,0965,0963,346
LiverLiver1364,2944,2942,840
SpleenSpleen1184,1134,1122,818
EBV-xform lymphocytes - CellsEBV-transformed lymphocytes1394,0954,0952,773

Subsequently, all shared PSI observations were removed from the ASCOT dataset to create a clean and independent 50-tissue ASCOT dataset, while the GTEx 6-tissue dataset remained unchanged. In this way, both the 50-tissue ASCOT and 6-tissue GTEx datasets are independent for further data partitioning and model training.

Response to Weakness & Question 2:

The evaluation and model-selection protocol is not sufficiently sound. Appendix Table 8 explicitly says that baseline checkpoints are selected by Pearson correlation on the test split. This makes their reported test metrics unsuitable as unbiased baseline estimates, even though the direction of this bias may favor the baselines. The paper also refers to validation-based hyperparameter selection without defining a validation split. These issues require a corrected protocol and rerunning the comparisons.

Was the test split used to select checkpoints or hyperparameters for any model? Please specify the validation partition and report results after model selection is performed without test-set access.

We thank the reviewer for identifying this important issue regarding the data partitioning and the protocol of model training, selection, and validation. To address these concerns, we have revised the experimental protocol to strictly separate model training, checkpoint selection, and final evaluation on the independent dataset.

Firstly, both the 50-tissue ASCOT and 6-tissue GTEx datasets are partitioned into training, testing, and validation sets at a 7:1:2 ratio using the random seed of 42. Specifically, we now explicitly define three independent data partitions, which are applied consistently to all baseline models and SpliceMamba: Train set: Used exclusively for model training; Test set: Used for hyperparameter tuning and checkpoint selection. The optimal checkpoint is selected based on the Pearson correlation. Validation set: Used exclusively for final performance evaluation.

The detailed dataset partitioning statistics are summarized in the Table below:

Independent DatasetSplitEventsGenesevent--tissue PSI observations
ASCOT (50 tissues)Train43,01015,4041,935,875
Test6,1444,822276,738
Validation12,2898,033551,838
Total61,44328,2592,764,451
GTEx (6 tissues)Train12,0535,65740,439
Test1,7211,4675,699
Validation3,4442,55911,612
Total17,2189,68357,750

Subsequently, we have rerun the model comparisons under this corrected protocol and updated the corresponding performance results in the table below (performance is predicted from the 50-tissue ASCOT validation set with the random seed of 42). Specifically, the publicly pretrained models are loaded for Pangolin and SpliceTransformer, while SpliceAI is not a tissue-specific prediction model and is trained from scratch. w/o Cons: model using sequence and tissue information, without conservation score. w Cons: model using sequence, tissue, and conservation score information.

ModelMAERMSEPearsonSpearmanR2R^2
SpliceMamba (w/o Cons)0.07840.17930.92460.81980.8509
SpliceMamba (w Cons)0.07470.16120.93830.82140.8795
SpliceAI (w/o Cons)0.19020.28570.78850.70870.6212
SpliceAI (w Cons)0.09220.17710.89440.80190.8445
Pretrained Pangolin (w/o Cons)0.13060.1840.81770.73670.8114
Pretrained Pangolin (w Cons)0.11320.15010.86410.78310.8543
Pretrained SpliceTransformer (w/o Cons)0.12820.22510.87470.77020.7651
Pretrained SpliceTransformer (w Cons)0.09790.1890.91350.79710.8345
AuthorsOpenReview ↗

Response to Reviewer xeqM – Part2

Response to Weakness & Question 3:

TSplice/MTSplice already modeled tissue-specific cassette-exon PSI using the same 61,823-exon ASCOT resource across 56 tissues, yet this paper does not provide a direct comparison or explain clearly what the new task formulation adds.

How does SpliceMamba compare with TSplice/MTSplice under a matched ASCOT split? What are the gains over feature-matched non-Mamba encoders and a genomic-feature-only model?

We thank the reviewer for highlighting the importance of comparing SpliceMamba with existing tissue-specific models, which also utilize the ASCOT resource. To address this concern, we have conducted additional experiments comparing SpliceMamba with MTSplice using a matched ASCOT dataset and consistent data partitioning and evaluation protocols.

The table below shows the results, with metric values reported on the 50-tissue ASCOT independent validation set using random seed 42. In summary, the results demonstrate that Full SpliceMamba consistently outperforms Full MTSplice across all evaluation metrics.

ModelMAE (↓\downarrow)RMSE (↓\downarrow)Pearson (↑\uparrow)Spearman (↑\uparrow)R2R^2 (↑\uparrow)
Full SpliceMamba0.07470.16120.93830.82140.8795
Genomic-feature-only SpliceMamba0.28760.3770.43440.33150.1369
Full MTSplice0.10110.19390.90940.76860.8256
Genomic-feature-only MTSplice0.17950.27520.80350.72390.645

1. Model comparisons under a matched ASCOT split: Following the revised data partitions and model-selection protocol, the training set is used for model training, and the test set is used for hyperparameter tuning and checkpoint selection. The optimal checkpoint is selected based on the Pearson correlation. The validation set is used for the final performance evaluation across all comparison models. Both models are evaluated on the same cassette-exon prediction task, with identical training, test, and validation partitions, ensuring a fair comparison.

2. Genomic-feature-only ablations:

To further investigate the source of these improvements, we performed genomic-feature-only experiments for both architectures. Specifically, we removed the sequence-derived representations while retaining genomic features, allowing us to assess the contribution of sequence information and feature integration. For SpliceMamba, removing sequence information substantially degrades performance, and MTSplice exhibits a similar performance decrease. These experiments demonstrate that genomic features alone are insufficient to achieve the predictive performance of the complete models, highlighting the importance of incorporating sequence-derived information.

The comparison with MTSplice and the genomic-feature-only ablations therefore provide complementary evidence of the benefits of the proposed architecture and multimodal input representation. In addition, we will integrate these experiments and corresponding analyses into the revised manuscript to provide a more comprehensive comparison with existing approaches and clarify the methodological contributions of SpliceMamba.

We sincerely thank the reviewer again for the constructive comments and valuable suggestions, which have helped us substantially improve the completeness of our work. We hope that the additional experiments, revised analyses, and clarifications adequately address the concerns raised and demonstrate the contributions of SpliceMamba. In light of these substantial revisions, we would be grateful if the reviewer could reconsider the overall assessment and score. We would also be pleased to provide any further clarification or conduct additional analyses should any questions or concerns remain. Thank you again for your time and careful consideration.

Hi reviewer xeqM,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

Reviewer_wj4XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SpliceMamba, a predictive framework based on state-space models. SpliceMamba is designed to predict tissue-condition Percent Spliced In (PSI) of cassette exons in alternative splicing. The method first constructs four nucleotide sequence windows and employs a shared Mamba encoder to extract representations, then combines them by a cross-window integration module from different splice sites. Using the integrated representations SpliceMamba finally constructs exon inclusion and skipping pathways to predict PSI. The proposed method is evaluated on the ASCOT and GTEx datasets to verify the effectiveness in both overall PSI prediction and the tissue-conditioned PSI predictions with differences analysis.

Critical strengths
  1. The paper addressed tissue-condition PSI prediction of alternative splicing, which is a relative important problem in biological research. The same exon may exhibit different PSI values across tissues. Clarify the differences can help study gene regulation under different tissue conditions.
  2. SpliceMamba follows a clear design. The representations are encoded using mamba from different sequence windows with auxiliary additive embeddings using. Then two different path are constructed for PSI prediction. The structure is well motivated align the tissue-condition splicing prediction task.
Critical weaknesses
  1. Conservation level is incorporated as an auxiliary feature in input representation building and has a huge impact on prediction performance as shown in appendix. That suggest it could be an important source of predictive information. However, the baselines do not use that conservation features, which means the performance advantage of SpliceMamba is not clearly distinguished from the model architecture or purely more information.
  2. SpliceMamba divides the input into four windows of 394 nucleotides each, rather than modeling a long history sequence. In that case, the computational complexity and sequence modeling will not be the bottleneck of Transformer. Therefore, the paper does not clearly demonstrate whether Mamba offers advantages over Transformer under the structure used.
Questions for authors
  1. The cross-window module is introduced in Section 3.3, but the specific key network architecture and implementation are not specified. Could the authors provide the detailed network of the module?
  2. The impact of conservation level is not clearly assessed yet. Could the authors provide experiments in which the baselines are given the same conservation level information?
  3. The existing ablation replaces Mamba with LSTM but does not conduct experiments with Transformer-based architectures. Could the authors provide ablation experiments with Transformer-based architectures to further establish the necessity and performance advantages of using Mamba?
  4. Equation (1) combines different types of embeddings through addition. But addition may cause interference or information loss. Could the authors assess the rationality of that design?
5 comments
AuthorsOpenReview ↗

Response to Reviewer wj4X – Part1

Response to W1 & Q2:

We sincerely thank the reviewer for the constructive suggestion. Specifically, we have addressed the reviewer's concerns through the following improvements:

  1. Adaptation of pretrained models: We incorporated pretrained Pangolin and SpliceTransformer as additional baselines. Their original prediction heads were adapted for PSI prediction, enabling a more direct and informative comparison with SpliceMamba. SpliceAI is trained from scratch with the published architecture because it does not provide tissue-specific predictions.
  2. Conservation score integration: To ensure a fair comparison, we incorporated sequence conservation scores into the pretrained comparison models. We also conducted dedicated ablation experiments by removing conservation scores from each adapted model, allowing us to quantify the contribution of evolutionary conservation information.
  3. Statistical reliability: All reported performance results are now obtained from three random seed runs of 42, 43, and 44, and are reported as mean ± standard deviation (SD), providing a more robust assessment of model performance.

The experimental results are summarised in the table below, and they have been added to the new Appendix C of the manuscript in red. w/o Cons: model using sequence and tissue information, without conservation score. w Cons: model using sequence, tissue, and conservation score information.

Model + ablationMAE (↓\downarrow)RMSE (↓\downarrow)Pearson (↑\uparrow)Spearman (↑\uparrow)R2R^2 (↑\uparrow)Param. (M)
SpliceMamba0.0784±0.00170.1565±0.00030.9412±0.00030.8153±0.00110.8856±0.00060.181
SpliceAI (w Cons)0.0865±0.00220.1782±0.00390.9242±0.00280.8105±0.00100.8526±0.00650.699
SpliceAI (w/o Cons)0.1374±0.00130.2605±0.00200.8332±0.00180.7538±0.00040.6851±0.00480.698
Pangolin (trained from scratch)0.1451±0.00140.2687±0.00140.8212±0.00190.7435±0.00250.6651±0.00340.698
Pretrained Pangolin (w Cons)0.0696±0.00040.1527±0.00110.9249±0.00060.8149±0.00070.8718±0.00160.699
Pretrained Pangolin (w/o Cons)0.0910±0.00060.1943±0.00010.9095±0.00010.8107±0.00020.8249±0.00020.699
SpliceTransformer (trained from scratch)0.1301±0.00180.2449±0.00130.8525±0.00160.7681±0.00150.7218±0.002910.939
Pretrained SpliceTransformer (w Cons)0.1198±0.00630.2187±0.00930.8833±0.01180.7856±0.01140.7803±0.020817.090
Pretrained SpliceTransformer (w/o Cons)0.1181±0.01230.2179±0.01220.8836±0.01420.7854±0.01010.7809±0.025017.087

These experiments demonstrate that SpliceMamba achieves an overall balance between predictive performance and computational efficiency. In particular, it obtains the highest Pearson correlation (0.9412) among all evaluated models while requiring only 0.181M parameters, substantially fewer than the pretrained alternatives.

We sincerely appreciate the reviewer's suggestions, which have helped us improve the completeness of this work.

Response to W2 & Q3:

We sincerely thank the reviewer for the insightful suggestion. We agree that comparing Mamba with the Transformer can demonstrate whether Mamba offers advantages over the Transformer under the structure used. To address this concern, we conducted architectural ablation experiments by replacing the Mamba layers with standard Transformer layers while keeping all other model components, input representations, and experimental configurations unchanged. We evaluated both architectures on the ASCOT 56-tissue and GTEx 6-tissue datasets, comparing predictive performance, running time per epoch, and model parameter counts. The results show that Mamba consistently outperforms Transformer on both datasets in the table below. Metrics values are reported as mean ± standard deviation (SD) over three independent runs with random seeds 42, 43, and 44.

Model VariantDatasetMAE (↓\downarrow)RMSE (↓\downarrow)Pearson (↑\uparrow)Spearman (↑\uparrow)R2R^2 (↑\uparrow)Time/epoch(Min)Param.(M)
MambaASCOT 56-tissue0.0784±0.00170.1565±0.00030.9412±0.00030.8153±0.00110.8856±0.000635.70.181
TransformerASCOT 56-tissue0.1293±0.04610.2227±0.04560.8762±0.05770.7681±0.05020.7636±0.097435.50.182
MambaGTEx 6-tissue0.1399±0.00030.2446±0.00330.8104±0.00390.7144±0.00340.6367±0.00991.80.181
TransformerGTEx 6-tissue0.1444±0.00470.2654±0.00590.7852±0.00920.6996±0.00610.5723±0.01891.60.182
AuthorsOpenReview ↗

Response to Reviewer wj4X – Part2

Response to Q1:

We thank the reviewer for the constructive suggestion. To address this, we have streamlined and clarified the architectural details of the cross-window module in the new Appendix A.6 of the manuscript, highlighted in red. The module integrates four sequence windows through the following key components:

  • Attention Pooling: Dynamically aggregates window sequences into compact vector representations while masking out padding positions.
  • Window-Offset Projection: Encodes sequence overlaps and normalized distances between adjacent windows, injecting vital spatial-geometric priors.
  • Window Dropout: Randomly drops individual window vectors during training (preg=0.1p_{\mathrm{reg}} = 0.1) to prevent overfitting and improve generalization.
  • CrossSite Block: Employs a causal selective-SSM (Mamba) architecture with customized step-size initialization (Δ0∼LogUniform(10−1,1)\Delta^0 \sim \text{LogUniform}(10^{-1}, 1)) to model directional information flow across all four windows.
  • Prediction Head: Combines inclusion and skipping path read-outs with length-bin embeddings and CDS vectors, passing them through the MLP to predict the final PSI value.

We hope this concise breakdown effectively clarifies the implementation and design of the cross-window module.

Response to Q4:

We sincerely thank the reviewer for raising this important point. We evaluated the additive embedding design from both theoretical and empirical perspectives.

First, theoretical justification. For categorical inputs, summing embeddings is not inherently less expressive than concatenating them. Each embedding lookup can be expressed as Ek(xk)=Ek1xkE_k(x_k) = E_k \mathbf{1}_{x_k}. Therefore, the sum in Eq. 1 is equivalent to a single linear transformation applied to the concatenated one-hot representations:

Similarly, concatenating learned embeddings and projecting them back to the model dimension yields ∑k(WkEk)1xk\sum_k (W_k E_k)\mathbf{1}_{x_k}, which belongs to the same function class. Thus, addition does not inherently discard information compared with concatenation followed by a linear projection, a commonly used embedding fusion strategy in Transformer architectures, where token, positional, and segment embeddings are combined additively.

Second, empirical validation. Our existing ablation experiments demonstrate that individual embedding components contribute meaningful information even when combined through additive fusion. Removing individual components consistently degrades predictive performance (Table in Appendix B.8): Pearson correlation decreases from 0.8137 to 0.5561 without conservation embeddings, to 0.7815 without exon/intron embeddings, and to 0.7982 without region embeddings. These results indicate that the different embedding components provide complementary predictive signals that remain beneficial after additive fusion.

Furthermore, relocating the tissue embedding from the additive input representation to the prediction head reduces Pearson correlation from 0.8137 to 0.7797 (Table in Ablation Study), supporting the effectiveness of early additive tissue conditioning.

Taken together, the theoretical equivalence and empirical ablation results support our additive embedding design as an effective and parameter-efficient strategy for integrating sequence, genomic, and tissue-related information.

We sincerely thank the reviewer again for the thoughtful and constructive feedback. We believe that the additional experiments and clarifications have improved our work and adequately addressed the concerns raised. We hope these revisions provide sufficient evidence of the validity and contributions of SpliceMamba, and we would greatly appreciate it if the reviewer could reconsider the overall evaluation in light of these changes. We would be happy to provide further clarifications or engage in additional discussion to address any remaining concerns. Thank you again for your time and consideration.

Thank you for the authors' response. The responses have addressed most of my concerns. However, some results (e.g., Pretrained Pangolin with conservation level) still outperform SpliceMamba on certain metrics. Although Pretrained Pangolin has more parameters, its overall parameter count remains acceptable. Therefore, I am inclined to maintain my current rating.

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

This work presents a new method for predicting alternative splicing in cassette exon events. The authors propose a bespoke architecture with fixed windows around relevant genomic positions, mamba-based modeling, and attention pooling to collect sequence information. The proposed method gives state of the art performance predicting alternative splicing probability when compared to other splicing model architectures trained from scratch with the same data and compute budget.

Critical strengths

Architectural Design

The proposed architecture has several elements that explicitly encode the alternative splicing mechanism:

  1. The sequence inputs are limited to small windows around the splice junctions, which are the most important regions that govern splicing.
  2. The inclusion and exclusion paths are computed separately with the information from only the junctions that would be used in their corresponding isoforms. These approaches can inspire future works with specific biological problems to include functional biases in their sequence modeling and architecture choices. Additionally, limiting the windows to be small and local reduces the overall number of parameters significantly vs if the entire sequence was used as input.

Performance

SpliceMamba achieves the best performance against the presented comparisons in the main benchmark. In both the ASCOT 56 tissue and GTEx 6 tissue datasets the proposed method scores well above any competitors. This suggests that the authors present a promising method for alternative splicing prediction for future downstream tasks or applications.

Critical weaknesses

Model Comparisons

The comparisons with other models are weakened in my opinion because the other models are trained from scratch. Even though the other models are splice site predictors and do not predict alternative splicing, they have significant relevant pretraining and could be fairly easily adapted to predict PSI. In particular, pangolin and spliceformer are trained with tissue specificity, and therefore could be directly compared to the proposed method with a simple layer on top of their final layer. This could be an alternative section to the existing comparisons, but would prove that the proposed method indeed achieves state of the art performance.

Additionally, per table 6, much of performance gains vs. other methods seems to come from the conservation score. This strong drop

These concerns are heightened through inconsistent language. Section 4.2 says:

Because SpliceMamba uses a richer event-level representation and auxiliary genomic/tissue features, this comparison should not be interpreted as an architecture-only evaluation.

while appendix B.10 says:

None incorporated pre-trained or publicly released model weights, ensuring that performance differences reflect architectural distinctions rather than disparities in pretraining data or optimization history. This experimental design deliberately controls for data budget and training protocol to enable a fair and architecture-centric comparison.

Tissue Transfer

While the proposed architecture shows strong performance across various tissues, since it uses static cell type labels it is unable to generalize to unseen tissue types. While a full custom encoder would involve significant architecture/design changes as well as more data, a simple test could use a pretrained gene expression model and a generic gene expression sample for each tissue type in place of the existing learned tissue ID embedding. The work would be much stronger if the pretrained SpliceMamba could transfer to new tissues, either zero shot or with limited or low data fine tuning.

Questions for authors

Useful additional comparisons:

  • Please attempt simple adaptation of pretrained models (see first concern above). Trunks can be left alone with just an additional head trained, or the entire model can be fine tuned.
  • Test pangolin/spliceformer in their existing setting on any tissues they are trained on, with small adapter head for PSI.
  • Consider adding a simple conservation score adapter to existing models/architectures, and evaluate ablations (specifically removing conservation score) on the ASCOT dataset.

Consider swapping the mamba layers for CNN/hyena or transformer. Would this change performance? Runtime? I think this could be a strong additional ablation over the existing LSTM/bidirectional mamba swaps.

I think it would be good to cite some more transformer based foundation/splice models, and specifically to compare against Alphagenome which contains tissue-specific splice strength.

It would additionally strengthen the work to include confidence intervals for the main results, either via multiple seeded runs or bootstrap CIs.

Can you briefly give validation sweeps for window size, and explain how the candidate windows were chosen?

Small clarity issue: 4.4 first paragraph is confusing/out of order/duplicated?:

The upper panel summarizes nucleotide-level attribution across four 394-nt splice-site-centered sequence windows, while the lower panels provide enlarged sequence-logo representations around the splice sites. The ISM heatmap in Figure 3 (Top) reveals that the influence of ISM effects is most pronounced in proximity to splice sites and attenuates progressively with increasing distance from these sites.”

11 comments
AuthorsOpenReview ↗

Response to Reviewer Vjws – Part1

Response to Weaknesses and Questions in Model Comparisons:

The comparisons with other models are weakened in my opinion because the other models are trained from scratch. Even though the other models are splice site predictors and do not predict alternative splicing, they have significant relevant pretraining and could be fairly easily adapted to predict PSI. In particular, pangolin and spliceformer are trained with tissue specificity, and therefore could be directly compared to the proposed method with a simple layer on top of their final layer. This could be an alternative section to the existing comparisons, but would prove that the proposed method indeed achieves state of the art performance.

Useful additional comparisons: • Please attempt simple adaptation of pretrained models (see first concern above). Trunks can be left alone with just an additional head trained, or the entire model can be fine tuned. • Test pangolin/spliceformer in their existing setting on any tissues they are trained on, with small adapter head for PSI. • Consider adding a simple conservation score adapter to existing models/architectures, and evaluate ablations (specifically removing conservation score) on the ASCOT dataset. I think it would be good to cite some more transformer-based foundation/splice models, and specifically to compare against Alphagenome which contains tissue-specific splice strength. It would additionally strengthen the work to include confidence intervals for the main results, either via multiple seeded runs or bootstrap CIs.

We sincerely thank the reviewer for the insightful and constructive suggestions. To address these concerns, we have conducted additional experiments and substantially expanded our comparative evaluation on the 56-tissue ASCOT test set (n = 867,155) through the following improvements:

  1. Adaptation of pretrained models: We incorporated pretrained Pangolin, SpliceTransformer, and AlphaGenome as additional baselines. Their pretrained representations were adapted for PSI prediction by modifying their prediction heads, enabling a more direct and informative comparison with SpliceMamba.
  2. Conservation score integration: To ensure a fair comparison, we incorporated sequence conservation scores into the pretrained comparison models. We also conducted dedicated ablation experiments by removing conservation scores to quantify the contribution of evolutionary conservation information.
  3. Statistical reliability: All reported performance results are now obtained from three random seed runs of 42, 43, and 44, and are reported as mean ± standard deviation (mean ± SD), providing a more robust assessment of model performance.
  4. Additional foundation model comparison: We added AlphaGenome to further strengthen the comparison against state-of-the-art pretrained genomic models.

The expanded experimental results are summarised in the table below, and they have been added to the new Appendix C of the manuscript in red. w/o Cons: model using sequence and tissue information, without conservation score. w Cons: model using sequence, tissue, and conservation score information.

Model + ablationMAE (↓\downarrow)RMSE (↓\downarrow)Pearson (↑\uparrow)Spearman (↑\uparrow)R2R^2 (↑\uparrow)Param. (M)
Full SpliceMamba0.0784 ± 0.00170.1565 ± 0.00030.9412 ± 0.00030.8153 ± 0.00110.8856 ± 0.00060.181
Pangolin (trained from scratch)0.1451 ± 0.00140.2687 ± 0.00140.8212 ± 0.00190.7435 ± 0.00250.6651 ± 0.00340.698
Pretrained Pangolin (w Cons)0.0696 ± 0.00040.1527 ± 0.00110.9249 ± 0.00060.8149 ± 0.00070.8718 ± 0.00160.699
Pretrained Pangolin (w/o Cons)0.0910 ± 0.00060.1943 ± 0.00010.9095 ± 0.00010.8107 ± 0.00020.8249 ± 0.00020.699
SpliceTransformer (trained from scratch)0.1301 ± 0.00180.2449 ± 0.00130.8525 ± 0.00160.7681 ± 0.00150.7218 ± 0.002910.939
Pretrained SpliceTransformer (w Cons)0.1198 ± 0.00630.2187 ± 0.00930.8833 ± 0.01180.7856 ± 0.01140.7803 ± 0.020817.090
Pretrained SpliceTransformer (w/o Cons)0.1181 ± 0.01230.2179 ± 0.01220.8836 ± 0.01420.7854 ± 0.01010.7809 ± 0.025017.087
Pretrained AlphaGenome (w Cons)0.0953 ± 0.00360.1831 ± 0.00160.9190 ± 0.00780.8947 ± 0.00890.8309 ± 0.0201259.560
Pretrained AlphaGenome (w/o Cons)0.0981 ± 0.00430.1880 ± 0.00860.9155 ± 0.00810.8907 ± 0.01070.8308 ± 0.0028259.550

These additional experiments demonstrate that SpliceMamba achieves an overall balance between predictive performance and computational efficiency. We sincerely appreciate the reviewer's suggestions, which have helped us improve the completeness of this work.

AuthorsOpenReview ↗

Response to Reviewer Vjws – Part2

Response to Weakness in Model Comparisons:

inconsistent language in Section 4.2 and Appendix B.10.

We appreciate the reviewer for identifying this inconsistency, which helps to improve the transparency and methodological robustness of our manuscript. We have revised Section 4.2 and Appendix B.10 of the manuscript in red. Specifically, the Section 4.2 description has been revised to: Since SpliceMamba incorporates both sequence and auxiliary genomic features, we also incorporate genomic conservation information into all baseline models to ensure consistent input representations, thereby facilitating an architecture-centric comparison.

The description of Appendix B.10 has been revised to: “All models in this particular comparison were trained from scratch without using pretrained or publicly released model weights. This controlled setting minimizes potential confounding effects arising from differences in pretraining data and optimization history. Because models may differ in their input representations and auxiliary features, the resulting performance differences should not be attributed exclusively to architectural distinctions. To provide a more direct comparison, we additionally evaluated adapted pretrained models by incorporating genomic conservation information and adapting their prediction heads for PSI prediction.” The experimental results are presented in the table in Part 1.

Response to Weakness in Tissue Transfer:

While the proposed architecture shows strong performance across various tissues, since it uses static cell type labels it is unable to generalize to unseen tissue types. While a full custom encoder would involve significant architecture/design changes as well as more data, a simple test could use a pretrained gene expression model and a generic gene expression sample for each tissue type in place of the existing learned tissue ID embedding. The work would be much stronger if the pretrained SpliceMamba could transfer to new tissues, either zero shot or with limited or low data fine tuning.

We sincerely thank the reviewer for this valuable suggestion. To address this concern, we did an additional external validation on the independent single-cell RNA-seq dataset derived from human hippocampus (accession number: GSE160189) (Ayhan et al., 2021). Specifically, we collected 131,193 cassette exon events from this dataset and directly applied the SpliceMamba model to predict PSI values, without fine-tuning or parameter updates on the target dataset.

In the table below, SpliceMamba achieved a Pearson correlation coefficient of 0.9301 on this unseen dataset, demonstrating its strong ability to generalize learned alternative splicing patterns beyond the original ASCOT dataset. We appreciate the reviewer’s suggestion to show the SpliceMamba's generalization ability.

ModelMAE ↓RMSE ↓Pearson ↑Spearman ↑R² ↑
SpliceMamba0.08270.17100.93010.80880.8633

Ayhan, Fatma, et al. "Resolving cellular and molecular diversity along the hippocampal anterior-to-posterior axis in humans." Neuron 109.13 (2021): 2091-2105.

AuthorsOpenReview ↗

Response to Reviewer Vjws– Part3

Response to ablation experiments:

Consider swapping the mamba layers for CNN/hyena or transformer. Would this change performance? Runtime? I think this could be a strong additional ablation over the existing LSTM/bidirectional mamba swaps.

We thank the reviewer for this constructive suggestion. We agree that comparing Mamba with Transformer architecture provides a valuable assessment of whether the performance improvements originate from the Mamba-based architecture. To address this concern, we conducted additional architectural ablation experiments by replacing the Mamba layers with standard Transformer layers while keeping all other model components, input representations, and experimental configurations unchanged. We evaluated both architectures on the ASCOT 56-tissue and GTEx 6-tissue datasets, comparing predictive performance, running time per epoch, and model parameter counts. The results demonstrate that Mamba consistently outperforms Transformer on both datasets in the table below:

Model VariantDatasetMAE (↓\downarrow)RMSE (↓\downarrow)Pearson (↑\uparrow)Spearman (↑\uparrow)R2R^2 (↑\uparrow)Time/epoch(Min)Param.(M)
MambaASCOT 56-tissue0.0784±0.00170.1565±0.00030.9412±0.00030.8153±0.00110.8856±0.000635.70.181
TransformerASCOT 56-tissue0.1293±0.04610.2227±0.04560.8762±0.05770.7681±0.05020.7636±0.097435.50.182
MambaGTEx 6-tissue0.1399±0.00030.2446±0.00330.8104±0.00390.7144±0.00340.6367±0.00991.80.181
TransformerGTEx 6-tissue0.1444±0.00470.2654±0.00590.7852±0.00920.6996±0.00610.5723±0.01891.60.182

We have included these additional ablation experiments in Section 4.5 and the new Appendix C of the manuscript to provide a more comprehensive architectural comparison. We sincerely appreciate the reviewer’s suggestion, which has helped strengthen the ablation study and provide further evidence supporting the architectural design of SpliceMamba.

Response to question:

Can you briefly give validation sweeps for window size, and explain how the candidate windows were chosen?

Thanks for this helpful suggestion. To address this concern, we conducted a validation sweep over three candidate window lengths (197, 394, 786) nt. These candidate sizes were selected using a progressive expansion strategy, approximately doubling the sequence context at each step. This allows us to systematically investigate whether incorporating increasing local genomic contexts improves alternative splicing prediction, while assessing the associated computational costs. All other model components, input features, and experimental configurations were kept identical.

As shown in the table below, expanding the window from 197 to 394 nt consistently improves all evaluation metrics, increasing Pearson correlation from 0.7764 to 0.8104, and R2R^2 from 0.5508 to 0.6367, while reducing MAE from 0.1465 to 0.1399 and RMSE from 0.2720 to 0.2446. The additional sequence context therefore provides useful information for characterizing cassette exon events.

Further increasing the window from 394 to 786 nt does not obviously improve prediction performance. Pearson correlation is 0.8077, while the training time per epoch rises from 1.8 to 2.7 min. Therefore, we select 394 nt in all experiments, providing the best overall balance between predictive performance, model complexity, and computational efficiency.

Window SizeMAERMSEPearsonSpearmanR2R^2Param.(M)Time/epoch(Min)
197 nt0.1465±0.00010.2720±0.00010.7764±0.00010.6806±0.00020.5508±0.00040.1681.3
394 nt0.1399±0.00030.2446±0.00330.8104±0.00390.7144±0.00340.6367±0.00990.1811.8
786 nt0.1437±0.00350.2433±0.00250.8077±0.00260.7038±0.00750.6404±0.00740.2062.7

Response to question:

Small clarity issue: 4.4 first paragraph is confusing/out of order/duplicated?

We thank the reviewer for pointing out this issue, and we have removed the duplication in Section 4.4. It has been revised to “Figure3 shows in silico mutagenesis (ISM) attributions of nucleotides and sequence logos.”

We sincerely thank the reviewer again for the constructive comments and valuable suggestions, which have helped us substantially improve our work. We hope that these experiments adequately address the concerns raised and demonstrate the contributions of SpliceMamba. We would be grateful if the reviewer could reconsider the overall assessment and score. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of rebuttal

We thank the Area Chair and all reviewers for their insightful feedback, which has helped us strengthen the manuscript a lot.

In summary, we have addressed all concerns through the following improvements:

1. The advantage of model architecture design. we added ablation experiments to replace the Mamba with the Transformer structure, and the Mamba structure is better than the Transformer (Pearson correlation 0.9412 vs. 0.8762). In addition, we conducted experiments on sequence window size selection (197, 394, 786 nt), and the selected window size of 394 nt provides the overall balance between predictive performance, model complexity, and computational efficiency.

2. Performance advantage: SpliceMamba outperforms the existing pretrained models by adding more comprehensive comparisons. We conducted additional experiments and ablations by comparing SpliceMamba with state-of-the-art methods, including pretrained Pangolin, SpliceTransformer, AlphaGenome, and MTSplice. SpliceMamba achieves stronger prediction performance with a Pearson correlation of 0.9412.

3. Improvement of model comparison robustness and fairness. We ran multiple random-seed experiments to ensure robust prediction performance across all compared models, and added the conservation score information to all compared baselines for a fair comparison to SpliceMamba.

4. Feasibility on unseen tissue. We performed an additional experiment on previously unseen tonsil tissue, mapping its gene expression to a tissue embedding in place of the learned tissue label, without retraining SpliceMamba. This demonstrates the feasibility of applying the model to tissues absent from training. Following the discussion with Reviewer Vjws, we present it as a proof of concept rather than demonstrated generalization, and will include the suggested mean-embedding control in the final version.

5. Additional data partitions. We ran additional experiments on an alternative 50-tissue ASCOT train/test/validation partition, where SpliceMamba retains the highest Pearson correlation (0.9383) among the compared models.

Taken together, these new experiments provide stronger support for the main contribution of SpliceMamba: accurate and parameter-efficient tissue-conditioned splicing prediction for the tissues represented in training.

We thank the Area Chair and Reviewers again for their constructive suggestions, which have helped us improve the quality and clarity of this work. We would be grateful if you could reconsider the overall assessments and increasing recommendations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.