acceptodds
Under review as a conference paper at ICLR 2027

Enhancing the expressivity of single-cell chromatin accessibility representations through adaptive region embeddings

Abstract

Single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) enables the characterization of epigenomic regulatory landscapes and cellular heterogeneity at single-cell resolution. However, learning effective representations from scATAC-seq data remains challenging because of its extreme sparsity, high dimensionality, and variability in genomic regions across datasets. Here, we introduce scChromHEARTs, a representation learning framework for single-cell Chromatin accessibility based on High-Expressivity Adaptive Region-aware embedding Transformations and a tailored self-supervised learning strategy. Rather than assigning fixed embeddings to predefined genomic regions, scChromHEARTs generates adaptive embeddings for arbitrary genomic intervals from their chromosomal identities and genomic coordinates. Combined with subset-based contrastive learning, this design enables efficient representation learning across the full genomic region space. scChromHEARTs achieved strong performance across diverse downstream applications, including cell-type annotation, perturbed cell generation, batch correction, and cross-dataset transferability. Ablation studies further demonstrate that input embeddings designed to finely encode genomic structure substantially improve representation quality. These results highlight the importance of designing input embeddings for effective representation learning of epigenomic data. Code is available at https://github.com/anonymous/ChromHEARTs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell chromatin accessibility, Representation learning, Self-supervised learning, Genomic representation learning, Genomic region embedding

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_ne6KOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces scChromHEARTs, a model for learning cell representations from scATAC-seq. The encoder represents each genomic interval by rotating a chromosome embedding at its two boundaries and taking the difference. It then applies a chromosome-specific gate and bias based on the interval midpoint. A Transformer combines accessible intervals into a cell embedding. Pre-training uses a subset-view contrastive objective adapted from scRobust, along with region-accessibility prediction. The experiments cover cell-type annotation, batch-associated clustering, transfer between datasets from the same tissue, and a perturbation case study. The model achieves strong annotation results on several datasets and outperforms EpiAgent on most of the five transfer targets.

Critical strengths
  1. The encoder uses chromosome identity and interval coordinates without a fixed region lookup table. This is useful when datasets have different peak definitions.

  2. The paper evaluates several downstream tasks and includes component ablations. Table 1 shows competitive cell-type annotation results. Table 4 also shows promising transfer to unseen datasets from the same tissue.

Critical weaknesses
  1. The advantage of the interval encoder remains unclear. Its main contribution is the combination of boundary rotations and midpoint gate/bias. The contrastive objective is adapted from scRobust, which already included a limited scATAC-seq application. Figures 3 and A7 provide useful ablations, but the token-plus-bias variant performs similarly on several datasets. Comparisons with simple coordinate encoders of similar size are still needed to justify this design. Figures 4 and A9 show organization by position and chromosome, but do not establish functional regulatory relationships. The biological interpretation of the learned “genomic structure” therefore needs stronger evidence.

  2. Table 4 supports transfer between datasets from the same tissue, but does not directly test the claimed flexibility on new genomic regions. The paper reports neither the overlap between source and target peaks nor separate results for shared and unseen intervals. EpiAgent is the only comparator, and differences in pre-training make the source of the gains hard to assess. ARI and NMI are also selected at the best Leiden resolutions in a search range. This uses target labels to select the reported results and leaves performance under a fixed, label-free protocol unclear.

Questions for authors
  1. How does the interval encoder compare with simpler alternatives, such as Fourier features of midpoint and width or a coordinate MLP? Please use the same Transformer, losses, input budget, and a similar parameter count. Does the lookup-table ablation reproduce scRobust, or just replace the region embedding in your pipeline?

  2. In Table 4, what fractions of target intervals exactly match or overlap source intervals? Do all datasets use the same reference assembly? Results for shared and unseen intervals, or for altered peak boundaries, would help test the encoder's flexibility.

  3. Could you add another representation baseline to Table 4 and report results under a fixed Leiden resolution? Please also explain how differences in pre-training may affect the comparison.

2 comments

Q1) Comparison with simpler coordinate-based embedders

Thank you for your suggestion. We conducted additional experiments comparing scChromHEARTs with two simpler coordinate-based embedders, Coordinate MLP and Fourier MLP, while maintaining the same Transformer backbone, pre-training objectives, and input budget, with closely matched embedding parameter counts.

As shown in the newly added Table A7 (Appendix G), scChromHEARTs consistently outperformed both alternatives across four datasets. For example, on PBMC, scChromHEARTs achieved an accuracy of 0.95 and a Macro-F1 of 0.84, compared with 0.67 and 0.38 for Fourier MLP, respectively. These results support the effectiveness of our structured interval embedding design beyond simpler coordinate encoding approaches.

We also clarify that the lookup-table ablation does not reproduce the full scRobust framework; it replaces only the genomic region embedding module within our pipeline.

We have included these findings in the revised manuscript (Page 9, Lines 472–479; Appendix G, Table A7).

W2 & Q2–Q3) Cross-dataset transferability and genomic interval sharing

Thank you for your suggestions. We expanded the cross-dataset transferability evaluation by adding PeakVI and PoissonVI as additional baselines, using shared cCRE features to ensure consistent input spaces for these VAE-based methods. To avoid target-label-based optimization, Leiden resolutions were selected exclusively on the training datasets and then fixed when applied to unseen evaluation datasets without further tuning.

We also clarified the differences in pre-training: EpiAgent was initialized from a large-scale human scATAC-seq pretrained model and subsequently trained on the tissue-matched dataset, whereas scChromHEARTs was trained only on the tissue-matched dataset without human corpus pre-training. As shown in the updated Table 4, scChromHEARTs achieved the highest average ARI/NMI on Lareau2019 (0.671) and Hocker2021 (0.803), demonstrating competitive cross-dataset transferability despite using less pre-training data.

Furthermore, we quantified genomic interval sharing between source and target datasets. As shown in Table A8 (Appendix H), 70.09–83.56% of active target intervals exactly matched source intervals, while 16.44–29.91% were source-unseen. All datasets used the same GRCh38/hg38 reference assembly and a shared cCRE dictionary. While these results characterize the extent of genomic region differences across datasets, performance on source-unseen intervals and robustness to altered peak boundaries remain to be evaluated separately.

We have included these analyses and clarifications in the revised manuscript (Page 8, Lines 398–427; Table 4; Appendix H, Table A8).

I thank the authors for their response. They have addressed most of my concerns. I have raised my score to 3.

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

The paper proposes scChromHEARTs, a representation model for single-cell chromatin accessibility (scATAC-seq) data. The model learns genomic region embeddings using chromosome identity, coordinates, and chromosome-specific gate/bias components. Specifically, each chromosome has a learned embedding, which is rotated by the region's start and end positions via rotary positional encoding (RoPE), and the region embedding is the difference of the two rotations, modulated by learned gate and bias embeddings. The cell embeddings are obtained by aggregating embeddings of accessible genomic regions using a Transformer pretrained on ~4.6M cells with a multi-positive contrastive objective and genomic region accessibility prediction. The authors validate on different tasks, including cell type annotation F1 on 7 datasets against foundation-model baselines (EpiAgent, ChromFound), classical methods (LSI), and dataset-specific deep methods; an ablation of the proposed embedder against a lookup table and a ChromFound-style additive encoding; batch-correction and cross-dataset transfer experiments; and a perturbation-response analysis on one dataset. The conclusion discloses that the model requires target-specific pretraining, limiting its direct use as a general-purpose foundation model.

Critical strengths

The core idea targets a real problem with a sensible mechanism. Computing region embeddings from genomic coordinates avoids the fixed-vocabulary problem (up to ~1M possible regions, extreme sparsity, and region sets that differ across datasets) and, by construction, shares information across nearby regions. The embedder-vs-lookup ablation shows that the proposed embedder trains in a regime where the lookup table does not, supporting the motivation. This matters to my assessment because it makes the paper worth fixing rather than dismissing.

Broad evaluation, candidly reported. Seven annotation datasets, several task types, and baselines spanning foundation models to classical methods, with a conclusion that openly concedes the need for target-specific pretraining. This matters because the breadth is what makes the mixed results interpretable.

Critical weaknesses
  1. The authors pass both predicted and real references through the same trained decoder to obtain the reconstructed profile. The ground truth obtained from real references this way would inflate the results by design. I think the authors should instead compare decoder predictions of the evaluation set directly against measured accessibility changes in each cell type (Page 6 and 19-20; Table 2, Appendix E8).
  2. Lower ARI/NMI between clusters and batch labels does not necessarily indicate good batch correction. For example, the score would also decrease if two distinct cell types are merged together (Table A3, page 21) - GSE194122 cell type ARI is 0.532 for scChromHEARTs, whereas it is 0.664 for EpiAgent and 0.583 for LSI. It would be useful for authors to evaluate batch correction within each cell type, which would show batch mixing. Cell type structure should be evaluated independently using biological conservation metrics.
  3. Cross-dataset transferability is evaluated against only a single baseline method EpiAgent (Page 7). This does not fully establish the advantage of scChromHEARTs against other baselines. ChromFound uses genomic coordinates and would be particularly useful. The authors should include this and a traditional method with matched genomic regions to strengthen the claim about transferability.
  4. The authors indicate disabling cell aggregation procedure for the baseline method ChromFound. However, removing an architectural component could hurt its performance. Example: F1 of 0.02 on Ameen2022 5.The positional encoding has an unaddressed aliasing problem. With the rotary base of 10^6 (Table A1), the slowest rotary frequency has wavelength 2π × 10^6 ≈ 6.3 Mb, while chr1 is ~248 Mb, so even the slowest dimension wraps ~40 times and genomically distant positions receive near-identical rotations. Figure A9 (before pretraining) shows the expected signature: locally smooth but globally scrambled, while ChromFound's panel shows a cleaner global gradient. Figure 4 suggests pretraining partially restores continuity, but the wraparound is a property of the encoding itself, and the base is never justified or ablated. This is consequential because the encoding scheme is itself the paper's contribution.
  5. The empirical evidence does not support the headline claims. (a) The central ablation is confounded: Table A1 lists one shared hyperparameter configuration with no baseline-specific tuning, and the lookup baseline's loss shows "little to no decrease" (the authors' words), which reads as untuned rather than inherently worse. (b) The benchmark results are mixed: on Table 1 F1, scChromHEARTs wins 3 of 7 datasets, LSI wins 2, EpiAgent wins 2; ChromFound's 0.02 on Ameen2022 appears to be a failed run, and since the authors disabled ChromFound's cell aggregation, no ChromFound column reflects the published method. (c) "w/o corpus PT" ties or beats "w corpus PT" on 4 of 7 datasets (exception: Li2023, 0.58 vs 0.79), and the conclusion itself concedes the model "requires target-specific pre-training," which "limits its direct use as a fully general-purpose foundation model." Together these leave the actual contribution, a coordinate-based region embedder, supported by a confounded ablation and mixed benchmarks, while the framing claims substantially more.
Questions for authors
  1. The authors jointly optimize a multi-positive contrastive objective and a genomic region accessibility prediction objective (Page 4). However, the current benchmarking does not evaluate the contribution of each training objective. How much does contrastive learning add? What is the performance with accessibility prediction only?
  2. It is unclear what the authors mean by cross-validation folds (Page 5). Is the evaluation split at the cell type or donor level? If it is stratified by cell type, it might be useful to additionally include performance evaluations using donor-stratified cross-validation folds to demonstrate cross-donor transferability.
  3. The authors indicate a disabling cell aggregation procedure for the baseline method ChromFound. However, removing an architectural component could hurt its performance. Have the authors quantified the effect of this modification?
  4. The rotary base of 10^6 implies a slowest-frequency wavelength of ~6.3 Mb, roughly 1/40 of chr1. How was the base chosen? Please report an ablation over the base, or another mechanism preventing positional aliasing at the chromosomal scale, with the Figure A9 visualization repeated for each setting.
  5. Were the baselines in the embedding ablation (lookup table, ChromFound-style additive encoding) tuned separately from the proposed method? If not, please report results with per-baseline tuning, given that the lookup baseline's loss showed little to no decrease under the shared configuration.
  6. ChromFound's F1 of 0.02 on Ameen2022 suggests a failed run. Please rerun it, and quantify the performance effect of disabling ChromFound's cell aggregation so the comparison reflects the published method.
  7. Given that "w/o corpus PT" matches or exceeds "w corpus PT" on 4 of 7 annotation datasets, what concrete benefit does the 4.6M-cell corpus pretraining provide beyond the Li2023 case, and how does that support the foundation-model positioning?
3 comments
AuthorsOpenReview ↗

Response 1

W1) Evaluation against measured accessibility changes

Thank you for pointing this out. To address this concern, we revised the evaluation to compare predicted accessibility changes directly against perturbation effects calculated from experimentally measured scATAC-seq profiles, without using decoder-reconstructed ground truth.

As shown in the revised Table 2 (Page 7), scChromHEARTs achieves the highest PCCs for both CD8 (0.623) and NK (0.569) cells, substantially outperforming all competing methods. These results demonstrate that scChromHEARTs effectively captures experimentally observed perturbation-associated accessibility changes.

We have updated the results in the revised manuscript (Page 7, Table 2, Lines 329–344; Appendix I.8.3).

W2) Batch mixing and biological conservation

Thank you for your comment. We agree that lower batch-label ARI/NMI alone does not necessarily indicate better batch correction. To address this concern, we evaluated batch mixing using Batch ASW and kBET (Table 3), with Batch ASW assessing batch mixing within each cell type. We additionally evaluated biological conservation using cell-type ARI/NMI (Table A10).

Furthermore, we adopted a unified Leiden resolution selection strategy that jointly considers cell-type preservation and batch mixing, rather than independently optimizing each metric. As shown in Table A10, scChromHEARTs achieved the second-best batch-label ARI and NMI on both datasets while maintaining competitive cell-type preservation.

These results provide complementary evidence of batch mixing and biological conservation. We have clarified the evaluation in the revised manuscript (Pages 7–8, Lines 365–390; Appendix F, Table A10).

W3) Cross-dataset transferability comparison

Thank you for your suggestion. We expanded the cross-dataset transferability evaluation by including two additional baselines, PeakVI and PoissonVI, using shared cCRE features to ensure consistent input spaces across datasets. All models were evaluated on unseen datasets without target-specific training, with hyperparameters selected exclusively on the training datasets.

We also considered ChromFound, as suggested by the reviewer. However, its official GitHub repository does not provide the pre-training code required to train the model on the designated source datasets, preventing us from including it in this evaluation.

As shown in the updated Table 4, scChromHEARTs achieved the highest average ARI/NMI on Lareau2019 (0.671) and Hocker2021 (0.803), while remaining competitive on the other datasets.

We have included these additional comparisons and clarified the evaluation protocol in the revised manuscript (Page 8, Lines 398–427; Table 4).

W4 & Q3) Impact of cell aggregation on ChromFound

Thank you for your comment. We conducted additional experiments to quantify the impact of cell aggregation on ChromFound across six datasets. As shown in the newly added Table A5 and Fig. A2 (Appendix E), cell aggregation consistently improved clustering performance, increasing ARI by 0.013–0.121 and NMI by 0.033–0.103. These results confirm that disabling aggregation reduces ChromFound's clustering performance. We further evaluated downstream cell-type annotation performance with and without aggregation (Table A6). Notably, on Ameen2022, enabling aggregation improved Macro-F1 from 0.48 to 0.77.

Cell aggregation incorporates information from LSI-derived neighboring cells, which can provide an additional advantage for cell-type annotation independent of the learned representations. Moreover, constructing neighborhoods using both training and test cells can introduce potential information leakage, as information from test cells may influence the aggregated representations of training cells. Therefore, we disabled aggregation in our main comparisons to evaluate representation quality under a consistent protocol.

We have clarified these findings in the revised manuscript (Page 5, Lines 245–251; Appendix E, Tables A5–A6).

AuthorsOpenReview ↗

Response 2

W5 & Q4) Effect of rotary base on positional encoding

Thank you for your comment. To investigate the effect of the rotary base and address the concern regarding positional aliasing, we conducted an additional ablation study with rotary bases ranging from (10^4) to (10^8). We note that genomic coordinates were already scaled by a factor of 1/10 in our original implementation before applying RoPE, increasing the effective positional wavelengths in the original genomic coordinate space by tenfold. Consequently, the slowest-frequency wavelength for the default base of (10^6) is approximately 63 Mb rather than 6.3 Mb, reducing the number of rotations across chromosome 1 from approximately 40 to 4. This scaling was maintained consistently across all rotary base settings in our additional ablation study.

As shown in the newly added Table A2 (Appendix C.4), downstream annotation performance remained relatively stable across different rotary bases on three datasets, with the default base of (10^6) achieving competitive results. We further repeated the positional embedding visualization for each base before pre-training (Fig. A1). Larger bases generally produced smoother positional structures, while (10^6) exhibited reasonably well-organized representations while maintaining competitive downstream performance.

These results support our choice of (10^6) as a balanced configuration between positional organization and downstream performance. We have added these analyses to the revised manuscript (Page 10, Lines 516–520; Appendix C.4, Table A2, Fig. A1).

Q5) Correction and re-evaluation of embedding ablation baselines

Thank you for pointing this out. Upon re-examining our experimental implementation, we identified an error in the previous code that prevented the lookup-table and ChromFound-style additive encoding baselines from being properly trained and evaluated. We corrected this issue and reran the experiments for both baselines. The updated results show reasonable performance and improved convergence behavior compared with the previous results.

We have updated the corresponding ablation results in the revised manuscript (Fig. 3, Figs. A8–A9). We apologize for the error and appreciate the reviewer for bringing this issue to our attention.

Q1) Contribution of contrastive learning

Thank you for your suggestion. We conducted an additional ablation study to evaluate the contribution of multi-positive contrastive learning by comparing the full scChromHEARTs model with a variant trained using only the accessibility prediction objective.

As shown in the newly added Table A3 (Appendix C.5), the accessibility-only variant achieved an accuracy of 0.7308 and a Macro-F1 of 0.4757 on Granja2019-PBMC, whereas the full model achieved 0.9500 and 0.9100, respectively. Thus, incorporating contrastive learning improved accuracy by 0.2192 and Macro-F1 by 0.4343.

These results highlight the importance of contrastive learning in learning discriminative cell representations. We have included these findings in the revised manuscript (Page 8, Lines 431–434; Appendix C.5, Table A3).

Q2) Donor-level cross-validation and cross-donor generalization

Thank you for your suggestion. Our original evaluation used cell-level five-fold cross-validation rather than donor-level splitting. To further assess cross-donor generalization, we conducted additional donor-level cross-validation experiments, ensuring that cells from held-out donors were excluded from supervised annotation training.

As shown in the newly added Table A9, scChromHEARTs achieved competitive performance across six datasets, obtaining the highest or tied-highest accuracy on three datasets (Kanemaru2023, Granja2019 bone marrow, and Li2023). In particular, it achieved an accuracy of 0.92 and Macro-F1 of 0.79 on Granja2019 bone marrow, and 0.94 and 0.74 on Li2023, respectively. These results demonstrate that scChromHEARTs retains useful cell-type information when evaluated on unseen donors.

We have clarified the evaluation protocol and added these results to the revised manuscript (Page 5, Lines 228–231 and 258–262; Table A9).

AuthorsOpenReview ↗

Response 3

Q6) ChromFound re-evaluation and impact of cell aggregation

Thank you for your comment. We reran ChromFound on Ameen2022 with cell quality control (QC). Previously, without QC, all input cells were retained, but the model exhibited poor training performance. After applying QC, ChromFound achieved substantially improved performance in three of the five folds, although the remaining two folds still showed poor performance. The fold-wise results are summarized below.

FoldTest AccuracyTest Macro-F1
10.82400.7976
20.14500.0127
30.82780.8069
40.14500.0127
50.80890.7799
Mean ± SD0.5501 ± 0.36990.4820 ± 0.4285

These results indicate that cell quality control improved performance in several folds, although substantial fold-dependent variability remained. We have incorporated these updated results into the revised manuscript (Table 1).

We also conducted additional experiments to quantify the impact of cell aggregation across six datasets. As shown in the newly added Table A5 and Fig. A2 (Appendix E), enabling aggregation consistently improved clustering performance, increasing ARI by 0.013–0.121 and NMI by 0.033–0.103. The corresponding downstream cell-type annotation performance with and without aggregation is reported in Table A6.

We have clarified the rationale for disabling aggregation in our main comparisons in the revised manuscript (Page 5, Lines 245–251; Appendix E, Tables A5–A6).

Q7) Contribution of large-scale corpus pretraining and foundation-model positioning

Thank you for your comment. We acknowledge that large-scale corpus pretraining does not consistently improve annotation performance across all datasets. However, as shown in Table 1, corpus pretraining generally yields more consistent performance across diverse datasets, suggesting its potential to improve representation stability and generalizability.

These findings also suggest that the benefits of large-scale corpus pretraining may depend more strongly on the pretraining methodology than on the proposed genomic region embedding architecture itself. We therefore believe that developing pretraining strategies specifically designed for large-scale biological foundation models could further enhance the generalization capabilities of scChromHEARTs.

We have clarified this interpretation in the revised manuscript (Page 5, Lines 252–257).

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

The paper introduces scChromHEARTs, a framework for learning cell representations from chromatin accessibility data. scChromHEARTs generates interval embeddings using chromosome-specific token, bias, and gate vectors combined with rotary transformations of genomic coordinates. A transformer model aggregates these embeddings, trained through contrastive learning and accessibility prediction.

Evaluation includes cell-type annotation on seven datasets, perturbation prediction for two held-out cell types, batch-effect analysis on two datasets, and transfer across datasets from matching tissues. Ablations compare embedding components and alternative embedding schemes. Annotation performance is competitive, although improvements over other methods vary across datasets.

Critical strengths

(1) The embedding design is a useful contribution. scChromHEARTs combines chromosome identity and genomic position through rotary transformations, biases, and gates.

(2) The ablations make a strong case for the embeddings. Keeping the transformer and training procedure fixed helps isolate the benefits of the embedding design.

Critical weaknesses

(1) The ChromFound comparison needs clarification. Cell aggregation is disabled to prevent information sharing between training and test cells, but its effect on performance is not measured. This makes the comparison difficult to assess.

(2) The annotation setup is not fully clear. The paper does not clearly state whether target-dataset pretraining and graph construction include held-out cells. This affects how the results should be interpreted for new cells.

(3) The perturbation results need stronger validation. The experiment covers one treatment in one dataset. The paired meta-cell analysis uses measured accessibility, but the prediction metrics use reconstructed targets. Evaluation against measured treatment effects and a simple average-effect baseline would strengthen the conclusions.

Questions for authors

(1) Does pretraining use all cells or only the training cells in each fold? Does the graph include test cells?

(2) How much does removing aggregation affect ChromFound’s performance? Can aggregation be applied separately to training and test cells?

(3) How does the predictor compare with using the average treatment effect from the training cell types? Does it still perform better when predictions are compared with measured changes in accessibility?

2 comments

W1 & Q2. Impact of Cell Aggregation on ChromFound

Thank you for your comment. We conducted additional experiments to quantify the impact of cell aggregation on ChromFound across six datasets.

As shown in the newly added Table A5 and Fig. A2 (Appendix E), enabling aggregation consistently improves clustering performance, increasing ARI by 0.013–0.121 and NMI by 0.033–0.103 across datasets. Additionally, Table A6 presents a direct comparison of downstream cell-type annotation performance with and without cell aggregation. These results confirm that removing aggregation affects ChromFound's performance.

Aggregation can be applied separately to training and test cells within each fold by constructing independent cell neighborhoods for each subset. However, this approach requires repeating the aggregation procedure for each fold, introducing additional computational overhead.

More importantly, cell aggregation modifies the input data by incorporating information from LSI-derived neighboring cells, which can provide an additional advantage for cell-type annotation independent of the learned representations. Therefore, we disabled aggregation in our main comparisons to isolate the quality of the learned representations under a consistent evaluation protocol.

Manuscript updates: Page 5, Lines 245–251; Appendix E.


W2 & Q1. Clarification of the Transductive Evaluation Setting

Thank you for raising this question. Pretraining was performed using all cells, including held-out test cells, without accessing their cell-type labels. Similarly, SANGO-based graph construction included test cells. Thus, our evaluation follows a transductive setting.

Manuscript updates: Page 5, Lines 242–244.


Q3. Comparison with the Average Treatment Effect Baseline

Thank you for your suggestion. We conducted additional cell-level evaluations against an average-effect baseline estimated from training cell types, excluding the evaluated CD8 and NK cells.

We compared predictions with experimentally measured post-perturbation accessibility profiles using three cell-pairing strategies: optimal transport (OT), random pairing, and scChromHEARTs-derived pairing.

As shown in the newly added Table A13 (Appendix I.8.4), scChromHEARTs consistently outperformed the average-effect baseline across Jaccard similarity, F1 score, and Pearson correlation (PCC) for both CD8 and NK cells.

Specifically, scChromHEARTs achieved:

  • F1 score: 0.2338 (CD8) and 0.2238 (NK), compared with 0.2287 and 0.2132 for the strongest average-effect baseline.
  • PCC: 0.0950 (CD8) and 0.0934 (NK), compared with 0.0038 and 0.0089 for the baseline.

These results demonstrate that scChromHEARTs achieves better agreement with measured post-perturbation accessibility profiles under inferred cell pairing.

Manuscript updates: Page 7, Lines 344–350; Appendix I.8.4, Table A13.

Reviewer_iXXrOpenReview ↗

Acknowledgement of authors' responses

Thank you to the authors for responding to all the comments. After reviewing them carefully, it seems the reported improvement is not significant enough to warrant revisiting the scores.

Comments on the paper

1 comment
Area_Chair_89vSOpenReview ↗

Please Join the Discussion

Dear Reviewers,

The authors have submitted their responses. Please read them and provide your feedback. Thank you!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.