We sincerely thank the reviewer for their constructive feedbacks and recognizing our work’s novelty and robust performance across downstream tasks. We would like to clarify and elaborate on the following points of concern:

[W1,L1] Model Novelty

As our submission is under the Machine Learning for Sciences area, our goal is to tailor the architecture to scATAC-seq characteristics for biologically meaningful representation and practical applications. We clarify the design of ChromFound’s each component as follows:

• Unified OCR tokenization (Sec 3.1) encodes chromosomal identity, genomic coordinates, and continuous-valued accessibility, providing a generalizable architecture for large-scale pretraining of heterogeneous scATAC-seq datasets.
• WPSA (Sec 3.2.1) captures local cis-regulatory dependencies within ±200 kb of transcriptional start sites, aligning with biological enhancer-promoter interactions.
• Mamba (Sec 3.2.2) complements WPSA by modeling long-range interactions across hundreds of thousands of OCRs per cell, enabling ChromFound to learn both local and global chromatin accessibility patterns.
• Pretrain objective (Sec 3.3) reconstructs masked both zero and non-zero OCRs using MSE loss, which ensures robustness to the high sparsity of scATAC-seq data.

[Q1] Shallow Model Baseline Comparison

We thank the reviewer for the helpful suggestion of including shallow baselines for both cell type annotation and cross-omics prediction.

1. Cell Type Annotation

   We implement two logistic regression models (sklearn.linear_model.LogisticRegression) using full or HVG-filtered OCRs (5,000 peaks), trained with default hyperparameters. These baselines are evaluated across the same datasets in Fig. 3 and compared against ChromFound. Results are reported in the format of Accuracy/F1 score as below:

2. Cross-Omics Prediction

   For cross-omics prediction, we implement a shallow two-layer MLP that maps OCR inputs to genome-wide gene expression profiles. Given the complexity of predicting over 30,000 gene targets, we used a hidden layer size of 4096, ReLU activation, learning rate of 1e-4, and 50 training epochs. Results are reported in the format of PCC/CCC as below:

Key Conclusions:

1. ChromFound consistently outperforms logistic regression, a strong baseline for annotation.
2. Full OCR input significantly outperforms HVG-filtered input, reinforcing ChromFound's necessity of modeling genome-wide regulatory elements.
3. The MLP baseline performs poorly on cross-omics prediction, revealing the limitations of simple linear or shallow models in capturing the complex cis-regulatory logic underlying chromatin-to-expression mapping. ChromFound is specifically designed to model this complexity through deep representation learning and large-scale self-supervised pretraining.

[W2,L2] Training Loss Suitability

We appreciate the reviewer’s concern regarding sparsity in scATAC-seq data. Actually, we have mentioned that ChromFound selectively reconstructs both non-zero and zero OCRs (Sec. 3.3). The masking strategy ensures the training procedure is not affected by the dropout events. Followed by xTrimoGene, we use the MSE loss for reconstructing the masking peaks, which is suitable for continuous-valued accessibility data. Fig. 2 further demonstrates that ChromFound maintains stable performance under severe downsampling, showcasing its robustness to sparsity.

[W3,Q2] OCR Format Unification Across Datasets

Thank you for the question. ChromFound adopts a coordinate-based tokenization strategy (Sec. 3.1), where each OCR is represented by its chromosome ID, start/end positions on GRCh38, and continuous accessibility. This enables ChromFound to unify peaks across datasets without requiring pre-aligned peak sets. Notably, the cross-tissue OCR overlap is quite low, highlighting the necessity of position-aware encoding for robust generalization.

[W4,W5] References to Related Works

We sincerely thank the reviewer for pointing out these three important related works: Epifoundation, ATAC-Diff, and GET.

• Epifoundation is quite impressive work, building a foundation model by leveraging paired scATAC-seq and scRNA-seq data to learn peak-gene alignments. However, its reliance on paired multi-omics data limits scalability as the paired data remains limited. In contrast, ChromFound adopts a self-supervised approach that does not require paired modalities, enabling pretraining on over 2.6 million cells from diverse tissues and disease conditions. Furthermore, ChromFound extends beyond Epifoundation’s scope by supporting biological applications such as enhancer-gene link inference and perturbation response prediction.
• ATAC-Diff is designed for identifying differential chromatin accessibility across conditions instead of supporting zero-shot transfer learning.
• GET, as we have already acknowledged in Section 2, Line 83 of the manuscript, is a transcriptomic foundation model using pseudo-bulk pairs of chromatin accessibility and sequence data, differing in resolution and modality from ChromFound.

In response to your professional suggestion, we will add citations to Epifoundation and ATAC-Diff, and update Section 2 to include a more thorough comparison. In addition, we clarify the positioning of ChromFound as the first genome-wide, self-supervised foundation model for scATAC-seq with zero-shot generalization across tissues.

[Q3] Performance Variation Across Tissues

Thank you for the insightful question. We have thoroughly evaluated ChromFound across various tissue types on all downstream tasks. While minor variations in performance exist due to inherent differences in tissue complexity and data sparsity, ChromFound consistently achieves strong results across all tasks and tissue types. If any specific aspect of the results remains unclear, please do not hesitate to let us know. We would be happy to provide further clarification.

[Q4] Comparison Between WPSA and Linear Attention

Thank you for the question. ChromFound’s WPSA is implemented using FlashAttention, significantly improving efficiency compared to vanilla self-attention. To further compare with linear attention, we implement Performer and Linformer under the same conditions, using identical hidden dimensions, batch size of 4, and FP16 mixed-precision training. The following results are obtained on the cell clustering task using the PBMC VIB_10xv1_1 dataset. As shown in the table below, WPSA outperforms both alternatives in terms of model performance and computational efficiency.

[Q5] Statistical Significance and Random Seed Control

We thank the reviewer for this professional and important question.

For cell clustering (Table 1) and low-count denoising (Figure 2), we perform each evaluation 20 times with different random seeds and report the mean and standard deviation to reflect performance stability. In the perturbation response prediction task (Figure 4), we report p-values for the PCC metric to establish the statistical significance of the predictions. For other tasks, we ensure that all benchmark methods are implemented and evaluated under the same random seed settings for fair comparison.

Together, these strategies ensure that all reported performance gains are both robust and statistically meaningful. We hope that all these clarifications have addressed reviewer's questions.

[W6,L4] Ablation Study of less OCR

We thank the reviewer for raising this important point. We try to understand the comment on "less OCR" in two possible ways, and we have conducted targeted ablation experiments to address both:

1. Reduce OCR input per cell: In Section 4.4 (Question 3), we reduce the maximum number of OCRs per cell to one-half and one-quarter of the full set by highly variable peaks filtering. We observe a clear performance drop in both clustering and annotation tasks, demonstrating that genome-wide OCR input is critical for capturing rich regulatory information and ensuring robust generalization.
2. Reduce pretraining data size: In Section 4.4 (Question 2), we also perform controlled experiments by reducing the pretraining corpus to one-tenth and one-hundredth of the original size. We observe a progressive decline in downstream performance, confirming the benefit of scaling up training data size.

[L3] More Downstream Tasks

We sincerely thank the reviewer for this suggestion. Actually, we have performed perturbation response prediction and presented the results in Figure 4. As the reviewer mentioned, temporal scATAC-seq prediction is quite challenging, and we are currently working on extending ChromFound to support this task.

Conclusion

We sincerely thank the reviewer for your valuable time and constructive feedback. We hope our responses have addressed your concerns. If there are any further questions, we would be happy to provide further clarification.

Dear reviewer:

We sincerely thank the reviewer for your detailed and thoughtful feedback. Your comments have helped us better articulate the design motivations behind ChromFound, clarify technical choices, and position our work within the broader context of scATAC-seq modeling. We appreciate your recognition of the strengths of our approach, as well as your critical perspective on its limitations.

In this rebuttal, we have carefully addressed each of your concerns through additional experiments, expanded comparisons, and clearer explanations. We have also updated our discussion of related work to reflect your suggestions, and we will revise the manuscript accordingly. We hope these responses provide a clearer understanding of our contributions and the rationale behind our methodological decisions.

If there are any remaining questions or further points for clarification, please do not hesitate to reach out. Thank you again for your time, expertise, and engagement.

We sincerely thank you for your thoughtful and constructive feedback. We are encouraged by your recognition of the technical soundness, comprehensive evaluation, and potential impact of ChromFound. Below we address your comments in detail.

[W1] Source Code

We appreciate your comment regarding code availability. We are fully committed to promoting transparency and reproducibility. Upon acceptance, we will release the complete source code, pretrained weights and tutorials for all experiments.

[Q2] DNA sequence modeling

Thank you for this insightful suggestion. We totally agree that DNA sequence information can complement chromatin accessibility data by providing base-resolution regulatory context. We note that scBasset also incorporates local DNA sequence to predict binary chromatin accessibility and has demonstrated strong performance in cell representation and batch correction.

While ChromFound currently does not explicitly encode DNA sequence, our tokenization schema can be extended to include local sequence features around each OCR. This would allow the model to jointly learn from static genomic sequences and dynamic chromatin accessibility profiles. One possible way of integrating DNA sequence is to fuse the DNA sequence embeddings from DNA foundation models (e.g., Evo2, Nucleotide Transformer, Enformer, etc.) into the ChromFound architecture. This approach would allow ChromFound to benefit from both dynamic chromatin activity and static regulatory code, potentially improving its ability to resolve fine-grained enhancer-gene interactions. We will explore this direction in future work.

[Q2] scRNA-seq modeling

Thank you for your professional suggestion. We totally agree that incorporating scRNA‑seq is both natural and valuable. Many current methods jointly model scATAC‑seq and scRNA‑seq by using scATAC‑seq as input to predict scRNA‑seq expression (e.g., GET) or perform multi‑omics integration through shared embeddings (e.g., GLUE, scMODAL).

However, these approaches typically treat peaks and genes as separate modalities linked only in downstream alignment. A current limitation in existing approaches is the lack of modeling under a unified genomic scale. We hope that the unified genomic scale will place peaks and genes on the same coordinate framework to enable end‑to‑end training from the genome sequence to transcriptomic output.

In our future work, we aim to bridge this gap by pretraining ChromFound in a truly genomic-centered manner: jointly embedding OCRs and genes in the same latent space from the start, enabling the model to internalize the central dogma (i.e., the information flow from DNA to RNA to protein) via self‑supervised objectives. While paired single-cell scATAC‑seq and scRNA‑seq data at sufficient scale remain limited for pretraining, we are actively pursuing large‑scale collection and simulation strategies to realize this vision.

[Q1] GRN prediction

We appreciate the reviewer’s suggestion on GRN prediction. ChromFound is currently designed to infer cis-regulatory interactions, such as enhancer-gene links and perturbation responses. Looking forward, we aim to expand ChromFound to support broader gene regulatory network (GRN) inference.

As discussed in the response regarding scRNA-seq modeling, we envision extending ChromFound to jointly model scATAC-seq and scRNA-seq data in a truly genomic-centered manner. This unified framework would allow the model to capture not only local cis-regulatory dependencies, but also trans-regulatory interactions. While data limitations currently constrain pretraining across paired multi-omics at single-cell resolution, we are actively exploring strategies to realize this integrative modeling paradigm in future work.

[Q3] Minor comments

Thank you for pointing this out. We will correct the table reference to Table 4 and revise Figure 6 in the revised manuscript to maintain consistent panel ordering.

[L1] Limitations on Human Data

We appreciate your suggestion to more explicitly discuss limitations. While we acknowledge that mouse datasets are often more abundant and widely available, our motivation is to prioritize human data due to its direct relevance for studying disease mechanisms and interpreting non-coding variants associated with human traits. We believe that this choice has greater significance for advancing biomedical research and understanding human-specific regulatory programs.

Nevertheless, extending ChromFound to other species is an important future direction. Our genome-aware tokenization is inherently adaptable, but cross-species modeling will require additional algorithmic revisions, such as species-aware OCR tokenization and the alignment of genomic coordinates across organisms. We plan to explore multi-species pretraining and cross-species transfer learning in future work.

Conclusion

Once again, we truly thank the reviewer for the valuable comments and strong endorsement of our work. We hope our responses have addressed your concerns, and we are happy to provide further clarifications if needed.

We sincerely thank the reviewer for the thoughtful and detailed feedback. We appreciate your recognition of ChromFound’s novelty, the biologically informed architectural design, and the breadth of downstream evaluations. Below, we address each of your suggestions and concerns.

[W3,L1] Discussion on Peak-calling Methods

Thank you for the important point. We agree that different upstream data pipelines can lead to variability in the cell-by-peak matrix and peak sets, particularly in sparsity and peak boundary definitions. ChromFound is designed to be robust to these variations.

• To assess robustness to such sparsity differences, we perform downsampling experiments (Figure 2) simulating varying levels of accessible peaks per cell. ChromFound maintains stable performance across all conditions, indicating strong resilience to sparsity variation introduced by upstream preprocessing.

• To address potential peak shift caused by different peak calling methods, ChromFound encodes OCRs using sinusoidal positional embeddings of their genomic coordinates. This design enables the model to capture relative spatial relationships between peaks, enhancing robustness to peak boundary shifts. Unlike dictionary-based approaches, our coordinate-based tokenization reduces reliance on exact peak definitions. We indeed conduct experiments aligning peaks across datasets to validate ChromFound’s robustness to peak shift. Due to space limitations, we kindly refer the reviewer to our response to Reviewer dMbP’s comment 2 for detailed results. We appreciate your understanding.

In summary, while the pretraining and evaluation datasets are from different upstream analytical tools, ChromFound still achieves strong performance and generalization. We will include a discussion of these points in the revised manuscript. Please feel free to discuss further if you have additional concerns.

[W3,Q5] Generalization to Disease Contexts

We appreciate your interest in ChromFound's generalization to disease contexts. As detailed in Appendix B Table 5, our pretraining corpus does span 6 disease types, including Alzheimer’s, Parkinson’s, leukemia, glioma, myocardial infarction and breast cancer.

Moreover, we have already included two disease benchmark datasets in downstream tasks: Morabito130K, derived from Alzheimer’s disease, and Kuppe139K, a heart dataset from patients with myocardial infarction. ChromFound demonstrates strong performance on both datasets (Table 1, Figure 2, Table 2), supporting the model’s robustness in disease contexts.

We hope these results address the reviewer’s concerns. In addition, We have not found a scATAC-seq dataset of autoimmune disorders with explicit cell labels. If the reviewer has specific datasets in mind, we would be happy to explore them further.

[W2] More Comprehensive Ablation Study

Thank you for the insightful suggestion. As requested, we compare ChromFound to both a pure Mamba and a pure Transformer baseline under the same pretraining corpus.

1. The pure Mamba model (Row 3 of Table 4) removes WPSA and uses Mamba alone. This results in a ~15% ARI drop, confirming that local attention is critical and that the hybrid design contributes synergistically to performance.
2. Training a vanilla Transformer on scATAC-seq inputs containing nearly one million peaks is practically infeasible due to its quadratic scaling in memory and computation. To approximate a pure Transformer baseline under practical settings, we train two representative single-cell foundation models, Geneformer and scGPT, on the same pretraining corpus with increasing peak lengths (4k–32k) and the same hyperparameters as WPSA. The results of comparison on the cell clustering task using the PBMC169K (batch VIB_10xv1_1) dataset are detailed in the table below.
   • Geneformer sorts peaks by accessibility value before input, which disrupts the native genomic order. As a result, it fails to learn meaningful representations.
   • scGPT applies a highly variable OCR selection strategy, limiting inputs to a small subset of peaks. Performance saturates at ARI ≈ 0.39, constrained by the maximum input length that Transformers can handle efficiently.

[W1,Q1,Q3] Model Sensitivity to Hyperparameters

Thank you for this helpful suggestion. We have evaluated both hyperparameters during early-stage model development on the cell clustering task using PBMC169K (batch VIB_10xv1_1) dataset. The results and findings are summarized below.

1. Positional embedding temperature: This parameter controls the frequency of sinusoidal position encodings and can be viewed as a proxy for “genomic resolution.” As shown below, performance slightly drops when temp is within 1e3 to 1e5, and degrades more substantially when temp is too large, likely due to loss of relative position sensitivity.

2. Mamba projection dimension: This parameter controls the compression within the Mamba block. As shown below, $D_{low}$ = 32 achieves a strong balance between performance and efficiency. Our choice of 32 is primarily motivated by the trade-off between performance and computational efficiency. Larger values lead to marginal gains but significantly increase FLOPs and memory, with $D_{low}$ = 128 exceeding the memory limits on A100 80G GPUs.

We hope these results address the reviewer's concerns. We will include this analysis in the revised manuscript.

[Q2] Interpretation of Fine-tuning Benefits

Thank you for the thoughtful and professional insight, and we strongly agree with your interpretation. The strong zero-shot performance indeed suggests that ChromFound learns generalizable representations of cellular states and underlying regulatory structures during pretraining. These representations are already informative for unsupervised clustering and can reveal biologically relevant subpopulations.

As you rightly pointed out, fine-tuning further improves performance by enabling the model to specialize in cell-type-specific regulatory features. Such features often involve subtle patterns or rare combinations of regulatory elements that are better resolved with supervision. This fine-tuning complements the global regulatory landscape already captured during pretraining. In addition, we have provided a detailed analysis of the cell type annotation confusion matrix in Appendix F.2, demonstrating ChromFound’s advantage in recognizing rare and long-tail cell types and cell-type-specific regulatory patterns.

We appreciate your expert perspective and will incorporate this interpretation into our revised discussion. Your feedback greatly enhances the clarity of our presentation and strengthens our understanding of how pretraining and fine-tuning synergize in ChromFound.

[Q4] Fixed WPSA Window

Thank you for raising this important point. Both Hi-C[1] and CRISPRi[2] have shown that the vast majority (over 90%) of validated enhancer-gene interactions occur within 200 kb of the transcription start site (TSS). Theoretically, the Mamba block modeling genome-wide dependencies allows ChromFound to infer potential distal enhancer-gene links. That said, we acknowledge that experimental validation of distal enhancer–gene links (>200 kb) remains limited, we view this as an important direction for future work. If the reviewer is aware of suitable datasets, we would be happy to incorporate such experiments.

[1] Bonev, B. & Cavalli, G. Organization and function of the 3D genome. Nat Rev Genet 17, 661–678 (2016).

[2] Fulco, C. P. et al. Activity-by-contact model of enhancer–promoter regulation from thousands of CRISPR perturbations. Nat Genet 51, 1664–1669 (2019).

[L2] Compute Resources for Fine-tuning

We thank the reviewer for highlighting the importance of computational efficiency. We provide profiling results on cell type annotation and cross-omics prediction tasks using a single A100 80G GPU:

We note that the primary driver of memory consumption is the length of the input peak sequence, which can reach several hundred thousand tokens per cell. In contrast, the model architecture itself remains lightweight (parameter count: 450305). All current resource metrics will be included in the appendix to guide users in estimating compute requirements. In future work, we aim to further optimize memory efficiency and training speed for broader applications.

Conclusion

We hope these responses address your concerns and questions. We sincerely appreciate your valuable feedback and constructive suggestions. If you have any additional questions or suggestions, please feel free to let us know.

Dear reviewer:

We sincerely thank the reviewer for the thoughtful, detailed, and constructive feedback. We appreciate your recognition of ChromFound’s novelty, biologically grounded architecture, and the comprehensive downstream evaluations. In this rebuttal, we have carefully addressed each of your suggestions, including analysis of upstream dependencies, generalization to disease contexts, expanded ablation studies, and sensitivity to key hyperparameters.

We hope our responses have addressed your concerns and clarified the rationale behind our design choices. If any questions remain or if you would like to discuss any points further, we would be very glad to continue the conversation. Your feedback has been instrumental in strengthening our work, and we are truly grateful for your time and insight.

We appreciate the reviewer for recognition of ChromFound’s novelty and constructive feedbacks. Below, we summarize our key revisions and additional analyses in response:

[1]. Benchmark completeness

1. We expand the clustering benchmarks (Table 1) by adding Signac (Sig), SnapATAC2 (SAT2), peakVI (pkVI), and poissonVI (psVI). All methods use the same preprocessing (Appendix C). Signac, peakVI, and poissonVI skip log normalization; scBasset uses binarized input.

2. We expand batch correction evaluation (Table 2), including peakVI (pkVI) and poissonVI (psVI) to replace scVI and scANVI.

3. The descriptions for all benchmark methods are provided in Appendix H. The APIs and tools used are as follows:

   • scBasset, peakVI, poissonVI, scVI: scvi.model
   • Liger: the pyliger Python package
   • Scanorama, Harmony: scanpy.external
   • scANVI: scib.integration
   • Cross-omics benchmarks: DANCE.modules.multi_modality.predict_modality
   • Others: official source codes and tutorials from their respective publications

   To directly address the reviewer’s question:

   • L2 regularization parameter for scBasset is set to 1e-8;
   • Input representation passed to Harmony is PCA.

4. We include logistic regression results in cell type annotation (Fig. 3) using Signac (Sig), peakVI (pkVI), and SnapATAC2(SAT2) embeddings. The results of baselines are much worse than ChromFound's.

5. Clarification on non-pretrained ChromFound:

   In manuscript, we have already included non-pretrained ChromFound as a baseline in Fig. 3 and Table 3. We emphasize that for tasks such as cell clustering and denoising, ChromFound is evaluated in zero-shot settings. Non-pretrained ChromFound loading randomly initialized weights in these scenarios is not a meaningful or fair baseline. We hope the reviewer understands our rationale, and we are happy to further discuss this point if needed.

[2]. Other peak calling methods and VAE pretraining

We sincerely thank the reviewer for insightful suggestions. In response, we perform an experiment to directly compare ChromFound with VAE-based models trained on reference-peak-aligned data. Specifically, we adopt the cPeaks reference set proposed by the recent work "A generic reference defined by consensus peaks for scATAC-seq data analysis", which defines 1,657,194 peaks across the human genome. We use a 20,000-cell PBMC subset (same as in Table 4, Row 6) as the training data. After mapping peaks to the cPeaks reference and filtering out peaks and cells with all-zero values, we obtain a training set of 306,784 peaks and approximately 18,000 cells.

We train four VAE-based baselines on the aligned data and compare their performance on zero-shot cell clustering (Table 1) against ChromFound trained on the same aligned data (Ours). Evaluation metrics and datasets follow Table 1, all aligned to 306,784 peaks. Results are summarized below:

We highlight two key observations from this experiment:

1. Baseline models perform poorly on unseen tissues. We find that only 30% of peaks in retina, 35% in cortex, and 70% in heart overlap with the cPeaks-aligned PBMC training peaks, suggesting that cross-tissue peak heterogeneity is a major factor limiting generalization.
2. Using cPeaks reduce ChromFound's performance compared to its native OCRs, suggesting that reference-based harmonization cannot capture novel or shifted OCRs introduced by batch effects or disease-specific variation. In contrast, ChromFound’s dynamic OCR tokenization allows the model to implicitly learn genomic distance and neighborhood structures, enhancing both robustness and extensibility.

[3]. Minor questions

1. OCR filtering when cell type labels are unavailable

   When cell labels are unavailable, we apply TF-IDF + LSI to infer pseudo labels for OCR filtering. We will clarify this in the manuscript.

2. Choice of loss function

   Due to the high sparsity of scATAC-seq data, we adopt a dual masking strategy (Sec 3.3) during pretraining, masking equal number of both nonzero and zero peaks. Under this setting, Poisson/ZINB losses often cause unstable optimization due to gradients dominated by nonzero entries. Instead, MSE loss on log-transformed, normalized signals ensures stability. This choice is also supported by xTrimoGene discussed in both their manuscript and rebuttal.

3. Ground truth in Fig. 4

   We would like to clarify that the ground truth labels in Fig. 4 are not ABC predictions, but experimental CRISPRi perturbation data in the Fulco4K dataset, as stated in lines 235–236 of the manuscript. Additionally, the citation of Fulco4K dataset will be corrected as "Activity-by-contact model of enhancer–promoter regulation from thousands of CRISPR perturbations".

4. Ablation study on the Mamba layer

   Thank you for the comment. To clarify, Question 3 in our ablation study is designed to assess the importance of long input peak sequences, not the effectiveness of the Mamba layers. Many existing scATAC-seq methods apply dimensionality reduction (e.g., highly variable peaks or LSI), but we aim to show that modeling long-range chromatin context leads to better cell representations and cross-omics performance.

   The effectiveness of the Mamba layers is directly evaluated in Question 1. The significant performance drop in this setting highlights the contribution of each component, especially Mamba layer.

Conclusion

We sincerely thank the reviewer for these thoughtful and professional suggestions. We hope our responses have addressed the reviewer’s concerns. Please let us know if further clarification is needed.

Dear reviewer:

We sincerely thank you for the time, expertise, and thoughtful suggestions provided throughout the review process. In this rebuttal, we have done our best to carefully address every concern raised, including expanded benchmarking, detailed comparisons with alternative peak harmonization strategies, and clarification of modeling and implementation details. These revisions were made with the goal of strengthening the work and making our contributions clearer.

We greatly value your feedback and would be very glad to further discuss any remaining questions or concerns. Please do not hesitate to reach out. We truly welcome continued dialogue and are committed to engaging with all comments in depth. Thank you again for your time, insight, and consideration.

We sincerely thank the reviewer for the follow-up questions and for acknowledging the additional experiments. We address each point below.

[Q1] Details of evaluation datasets

We appreciate your insightful questions. Below we provide detailed information about the evaluation datasets in experiment [1]:

[Q2] Hyperparameters of peakVI/poissonVI

We thank the reviewer for pointing out the hyperparameter settings of peakVI and poissonVI. We follow the default setting in the scvi-tools library. Specifically we detail the hyperparameters you mentioned below:

We observe that peakVI typically runs through 500 epochs, whereas poissonVI usually stops after fewer than 100 epochs due to the early stopping strategy. Thank your again for your professional suggestion. We hope this clarification of the hyperparameter settings will address your question.

[Q3] Non-pretrained ChromFound

Thanks for your suggestions and clarification. We include additional results as below, including pre-trained and continuous training on the dataset of interest (ChromFound with continuous training on evaluation datasets) and training from scratch on the dataset of interest (ChromFound trained from scratch on evaluation datasets).

As shown in the table above, ChromFound with continuous training on evaluation datasets achieves better performance than ChromFound and ChromFound trained from scratch on evaluation datasets. Nevertheless, training on the evaluation datasets incurs significant additional computational cost. We report the inference and training speed of one A100 80G GPU, along with GPU memory usage.

Given the additional computational cost and the limited performance gain, we recommend generating cell representations in a zero-shot setting without further training. Due to time constraints, we will include results for all datasets in the revised manuscript, along with corresponding computational resource requirements, as part of the usage instructions for ChromFound. We once again thank the reviewer for the valuable suggestion.

[Q4] Correction interpretation of the results

We sincerely thank the reviewer for pointing out the correct interpretation of the results. As you mentioned that "VAEs may do poorly on very few cells or too many peaks", we also notice that the used peak number of datasets in the tutorial of peakVI implemented in scvi-tools is 33142, which is much smaller than the number of peaks in our datasets. Therefore, we believe the poor performance of peakVI/poissonVI in our experiments is likely due to the large number of peaks in our datasets. We include the additional experiments to filter peaks with highly variable selection and retrain peakVI/poissonVI on the filtered datasets. The results are summarized below:

As shown in the table above, peakVI and poissonVI perform much better after highly variable peak selection, confirming their strength as baselines under reduced dimensionality. However, as the input dimensionality increases, their performance exhibits a marked decline, suggesting they may struggle to model long-range chromatin dependencies. ChromFound’s hybrid WPSA-Mamba architecture addresses this limitation by integrating local cis-regulatory context with efficient long-range modeling, enabling broader genome-wide applications such as enhancer–gene link inference and perturbation effect prediction (Section 4.3).

Due to time constraints, we are unable to extend this integration experiment to all datasets. We will be glad to include results for all datasets and additional baselines with highly variable peak selection in the revised manuscript. Thank you again for your professional suggestion.

[Q5] Batch effect in the cell type annotation

We sincerely thank the reviewer for this professional and insightful suggestion. We did not integrate the test (query) dataset into the training (reference) dataset in the response before. Following your professional suggestion, we implement scArches for scPoli to compare cell type annotation performance with and without applying scArches-based integration between the training (reference) and test (query) datasets. The integrated setting yields a substantial performance improvement, fully consistent with your observation that batch effects largely account for the poor baseline results.

Due to time constraints, we are unable to extend this integration experiment to all datasets. We will be glad to include results for all datasets and additional baselines with train/test integration by scArches in the revised manuscript. We greatly appreciate your suggestion, which has helped us optimize and strengthen our work.

[Q6] The number of OCRs per cell reduced in ablation study Question 3

We sincerely thank the reviewer for the follow-up questions. For pretraining, we set the maximum OCR sequence length to 440,000. If a dataset contains more than 440,000 OCRs, we retain the top 440,000 most variable OCRs; if it contains fewer, we apply zero padding to match the maximum length. In the ablation for Question 3, we compare maximum lengths of 220,000 and 110,000 OCRs. The reduction in OCRs for these settings is performed via dataset-specific highly variable OCR selection, ensuring that the retained peaks capture the most informative variability within each dataset.

Summary

We are deeply grateful to the reviewer for their thoughtful and professional feedback. Your comments not only demonstrate a thorough understanding of our work, but also reflect a high level of expertise and academic rigor. Your insightful suggestions significantly improve the clarity and depth of our manuscript. We feel privileged to have our work evaluated by someone with such discernment and scholarly acumen. We hope that our response has addressed all your concerns. Thank you again for your invaluable contribution to the refinement of our research.