acceptodds
Under review as a conference paper at ICLR 2027

Hi-Cell: A Cell-Level Foundation Model for Single-Cell Hi-C Data

Abstract

Single-cell Hi-C (scHi-C) technology provides a direct view of three-dimensional genome organization in individual cells, but its high dimensionality, extreme sparsity, and multiscale contact patterns create an urgent need for a sophisticated and versatile computational framework. Here we introduce Hi-Cell, a cell-level foundation model pretrained on Human-Hi-C-Corpus, our curated large-scale collection of 324,102 single cells from 17 human tissues and 357 bulk Hi-C experiments. Hi-Cell represents each cell as a hierarchical "cell sentence” composed of cell-, chromosome-, and multiscale patch-level tokens, and uses global bidirectional self-attention to integrate local contact patterns with chromosome-wide and whole-cell organization. A three-stage progressive pretraining strategy combining single-cell pretraining, bulk-derived pseudo-single-cell supervision, and cross-depth refinement enables Hi-Cell to learn transferable cellular representations and structural priors. Across six unseen scHi-C datasets, zero-shot embeddings are competitive with dedicated methods, while label-free fine-tuning achieves the best mean clustering performance among the evaluated baselines. Frozen Hi-Cell representations also support accurate cell-type annotation, achieving the highest mean macro-F1 among baselines. Beyond cell representation, Hi-Cell performs zero-shot contact imputation across sequencing depths, leading all evaluated methods in held-out and leave-one-out structural correlation on all six datasets and preserving independently validated A/B compartment profiles. Together, Hi-Cell provides a unified and transferable framework for analyzing cellular identity and reconstructing multiscale 3D genome organization from sparse scHi-C data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell Hi-C, 3D genomics, foundation models, self-supervised learning, representation learning, chromatin conformation, transfer learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_PUthOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Hi-Cell, a cell-level foundation model for single-cell Hi-C analysis. The method represents chromatin contact maps using hierarchical cell, chromosome, and multiscale patch tokens, which are jointly processed by a Transformer to capture local and global chromatin organization. The authors also construct Human-Hi-C-Corpus, comprising approximately 324k single cells and 357 bulk Hi-C experiments, and propose a three-stage pretraining strategy incorporating real single-cell data, bulk-derived pseudo-single cells, and cross-depth reconstruction. The model is evaluated on six external scHi-C datasets for cell clustering, cell-type annotation, and contact imputation. Results demonstrate competitive zero-shot representations, improved clustering after label-free fine-tuning, and strong zero-shot contact reconstruction across sequencing depths. Additional experiments investigate the benefits of pretraining and multiscale representations.

Critical strengths
  • Large-scale foundation modeling for an underexplored modality.
  • Strong and comprehensive empirical results.
Critical weaknesses
  • Incomplete comparison with closely related methods. Although the paper discusses HiCFoundation and other recent Hi-C modeling approaches, the main experiments primarily compare against traditional scHi-C methods such as Higashi and scHiCluster. A direct comparison with HiCFoundation, or a clearer explanation of why its single-cell applications cannot be evaluated under the same protocols, would better establish the contribution relative to existing foundation models.

  • Insufficient ablations of the progressive pretraining strategy. The authors compare the initial single-cell-only checkpoint against the final model, but do not separately isolate the contributions of bulk-derived pseudo-single-cell supervision and cross-depth refinement. Consequently, the relative importance of the three training stages remains unclear. Matched-budget ablations would help establish whether the proposed progressive strategy is necessary.

  • Potential reliability issues in baseline evaluation. The supplementary annotation results reveal unusually large performance drops for scHiClassifier in individual folds, such as Kim2020 and Wu2024, which substantially reduce its aggregate performance. These results raise questions about baseline training stability and whether the reported comparisons accurately reflect its achievable performance. Further clarification or verification of these folds would strengthen the evaluation.

  • Limited resolution and computational efficiency analysis. The model operates at 1 Mb resolution and considers only intrachromosomal contacts, limiting its applicability to finer genomic structures and interchromosomal interactions. Moreover, the approximately 594M-parameter architecture introduces substantial computational overhead, but inference cost, memory consumption, and scalability are not systematically compared with existing approaches. These limitations affect the practical significance of the proposed framework.

Questions for authors
  • Could the authors clarify why HiCFoundation is not included in the main evaluation? Is there an applicable pretrained configuration or evaluation setting that would allow a more direct comparison?

  • Could the authors provide separate results after Stage II and Stage III, or a comparison that removes pseudo-single-cell supervision while maintaining a comparable training budget? This would help establish the contribution of the individual pretraining components.

  • Could the authors clarify the unusually low scHiClassifier scores in certain annotation folds, particularly Kim2020 and Wu2024? Were these caused by training instability, convergence failures, or other factors?

  • Could the authors report approximate inference time and GPU memory requirements for Hi-Cell relative to the strongest baselines? This would help assess the practical tradeoff between model complexity and performance.

3 comments

Response to Reviewer PUth (Part 1/2)

We appreciate the reviewer's positive feedback on the value of foundation modeling for scHi-C and the breadth of our empirical evaluation. In response to the reviewer's constructive comments, we have conducted a direct comparison with HiCFoundation, expanded our pretraining ablations, systematically audited the scHiClassifier results, and performed additional analyses of inference efficiency and GPU memory usage. We also provide further clarification of Hi-Cell's 1-Mb intrachromosomal modeling scope.

1. Direct comparison with HiCFoundation

We thank the reviewer for highlighting the importance of comparison with existing Hi-C foundation models. HiCFoundation provides official pretrained configurations that support evaluation on our downstream tasks, and we have now completed this direct comparison.

Our original benchmark focused on methods specifically designed for cell-level scHi-C analysis, whereas HiCFoundation primarily learns genomic-region representations from contact windows and aggregates them for downstream applications. Hi-Cell instead directly learns whole-cell representations by jointly modeling cell-, chromosome-, and multiscale patch-level information. This difference in pretraining focus motivated our initial baseline selection.

Following the reviewer's suggestion, we evaluated HiCFoundation using its official pretrained encoder for clustering and annotation and its official single-cell checkpoint for contact imputation, preserving its native preprocessing and representation aggregation.

Metric ↑HiCFoundationHi-Cell
Clustering NMI0.24680.6423
Clustering ARI0.15720.5925
Annotation macro-F10.69110.8569
Imputation AUPRC0.13920.2893
Imputation held-out SCC0.19830.3235
Imputation LOO SCC0.36270.7296

Clustering and annotation use frozen pretrained representations, with the same supervised readout protocol for annotation. For imputation, both methods are evaluated at 10k, 25k, and 50k contacts per cell on the same 14-chromosome support; Hi-Cell predictions and references are recomputed accordingly.

Hi-Cell achieves higher mean performance across all six metrics, demonstrating its effectiveness in transferring cellular representations and structural priors to unseen datasets. Notably, the same pretrained Hi-Cell checkpoint supports clustering, annotation, and zero-shot contact reconstruction without target-dataset encoder fitting.

2. Contributions of pseudo-single-cell supervision and cross-depth refinement

We thank the reviewer for the insightful suggestions. We conducted additional stage-wise and matched-budget ablations to separately examine the contributions of bulk-derived pseudo-single-cell supervision and real-cell cross-depth refinement.

All continuations start from the same Stage-I checkpoint pretrained on real single-cell Hi-C data. Stage II introduces pseudo-single-cell-to-bulk reconstruction, and Stage III introduces binomial thinning of real-cell inputs for cross-depth reconstruction.

Training sequenceNMI ↑Held-out SCC ↑LOO SCC ↑
Stage I0.59040.30310.6536
I → II (real-only)0.61350.30180.6532
I → II (real + pseudo)0.61050.34300.7392
I → II (real-only) → III (real-only)0.62450.31880.6843
I → II (mixed) → III (no thinning)0.63350.34570.7432
I → II (mixed) → III (Full Hi-Cell)0.64230.34780.7460

All results are six-dataset means; imputation averages 10k/25k/50k contacts per cell on the original 23 chromosomes. Stage II and Stage III use 125,600 and 30,720 updates, respectively.

Bulk-derived pseudo-single-cell supervision provides substantial structural reconstruction benefits. Under the same Stage-II training budget, mixed real/pseudo supervision achieves a mean LOO SCC of 0.7392, compared with 0.6532 for real-only training. After the complete continuation, Full Hi-Cell reaches 0.7460, compared with 0.6843 when pseudo-single-cell supervision is removed throughout Stages II and III. This demonstrates that the structural gains cannot be explained simply by additional real-cell optimization.

Real-cell cross-depth refinement provides further improvements. Removing binomial thinning from Stage III, while retaining the mixed-supervision Stage-II initialization, yields held-out SCC and LOO SCC of 0.3457 and 0.7432, respectively, compared with 0.3478 and 0.7460 for Full Hi-Cell. The complete recipe also improves mean clustering NMI from 0.6105 after Stage II to 0.6423 after Stage III.

Together, these experiments demonstrate the complementary roles of real-cell pretraining, bulk-derived structural supervision, and real-cell cross-depth refinement.

Response to Reviewer PUth (Part 2/2)

3. Reliability of scHiClassifier evaluation

We appreciate the reviewer's careful examination of the fold-level annotation results. We used the authors' publicly released scHiClassifier implementation and its training procedure, with the same canonical cells and stratified five-fold test splits as the other annotation baselines.

To investigate the reported results, we audited the saved predictions, fold assignments, labels, and training histories across all 30 scHiClassifier folds and reproduced the reported scores. The two folds highlighted by the reviewer exhibit highly concentrated predictions:

Dataset / foldPrediction behaviorAccuracyMacro-F1
Kim2020, fold 41,367/1,579 cells predicted as HFF0.13050.0966
Wu2024, fold 3All 86 cells predicted as BJ0.15120.0657

The saved training history for Wu2024 shows a sharp increase in training loss after approximately 380 epochs, while the reported predictions were obtained from the fixed 420-epoch checkpoint. This behavior is consistent with late-stage training instability under the adopted training procedure, although its precise cause remains undetermined.

Hi-Cell's annotation performance is also supported by an independent comparison with logistic regression. Under the same five-fold evaluation framework, frozen Hi-Cell embeddings with an MLP readout achieve a mean macro-F1 of 0.8569, compared with 0.8047 for logistic regression, an improvement of 5.22 percentage points. This provides additional evidence for the effectiveness of Hi-Cell's transferable cell representations.

We will include these audit findings in the revised manuscript.

4. Computational efficiency and scalability.

We thank the reviewer for highlighting the importance of computational efficiency and practical applicability. We profiled Hi-Cell, HiCFoundation, and Higashi across six external datasets using matched 10k-contact inputs and the same GPU within each dataset.

The following table reports whole-cohort contact imputation times. Hi-Cell and HiCFoundation reuse pretrained checkpoints without target-dataset fitting, whereas Higashi follows its native target-specific fitting and neighbor-imputation pipeline.

DatasetCellsHi-Cell inference (s) ↓HiCFoundation inference (s) ↓Higashi fitting (s)Higashi inference (s)
Ramani2017640.3032.233,250.6514.84
Lee20193,89618.233,176.842,442.721,041.42
Kim2020970.8645.633,137.1267.74
Chang2024840.3137.042,323.5018.09
Wu20244281.56224.092,598.4298.67
Zhou20248377.11391.212,936.39624.06

Hi-Cell achieves the lowest native-pipeline inference time on all six cohorts. For the 3,896-cell Lee2019 dataset, Hi-Cell completes imputation in 18.23 seconds, compared with 3,176.84 seconds for HiCFoundation. Higashi requires 2,442.72 seconds of target fitting followed by 1,041.42 seconds of prediction, in addition to setup. These results highlight the deployment advantage of reusable pretraining, enabling efficient analysis of new cohorts without dataset-specific encoder fitting.

GPU memory requirements are also practical: on Lee2019, Hi-Cell uses 3.01 GiB of peak allocated memory at cell batch size 64, compared with 1.39 GiB for HiCFoundation at window batch size 16 and 1.13 GiB for Higashi's required native workflow. The methods use their respective batching interfaces.

Regarding resolution, Hi-Cell operates at 1 Mb across chromosomes 1–22 and X, modeling genome-wide intrachromosomal contacts. This shared coordinate system enables transferable cellular representations and chromosome-scale structural reconstruction across heterogeneous datasets. We will clarify this modeling scope and discuss extensions to finer-resolution and interchromosomal contact modeling in the revised manuscript.

We will incorporate the resource measurements and their evaluation protocols into the revised manuscript.

Thank you for the rebuttal. I've raised my score accordingly.

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

This paper presents Hi-Cell, a foundation model for single-cell Hi-C data analysis. The model aims to address the challenges of data sparsity and complex chromatin structures through hierarchical representations and a Transformer-based architecture. The authors introduce a progressive pretraining strategy that integrates single-cell and bulk Hi-C data to learn transferable representations. The proposed framework is evaluated on multiple external datasets across three downstream tasks: cell clustering, cell-type annotation, and contact imputation. Experimental results demonstrate competitive performance against existing methods, suggesting that Hi-Cell provides a unified framework for learning cellular representations and reconstructing chromatin contact patterns across different datasets.

Critical strengths

1.The paper provides comprehensive evaluations across multiple datasets and downstream tasks, supported by baseline comparisons and ablation studies.

2.The authors construct a large-scale Hi-C corpus covering diverse tissues and experimental conditions, providing a solid foundation for transferable representation learning.

3.The paper addresses important challenges in single-cell Hi-C analysis, particularly data sparsity and limited transferability. The proposed unified framework has potential practical value for various downstream applications.

Critical weaknesses

1.Although the paper compares several established methods, it lacks direct comparisons with recent approaches such as HiCFoundation and Hi-Cformer, making it difficult to fully assess its relative advantages.

2.The ablation studies do not fully isolate the contributions of key components, such as pseudo-single-cell supervision and cross-depth refinement. Additionally, sensitivity to important hyperparameters, including loss weights and contact-thinning settings, is not systematically evaluated.

Questions for authors

1.Could the authors explain the absence of comparisons with recent methods such as HiCFoundation and consider adding them where feasible?

2.Could the authors further evaluate the individual contributions of key training components and the sensitivity to important hyperparameters?

1 comment

We appreciate the reviewer's positive feedback on our evaluation, diverse pretraining corpus, and the practical value of transferable cell-level modeling. We provide new comparisons with HiCFoundation and Hi-Cformer and controlled experiments on progressive pretraining and hyperparameter sensitivity.

1. Comparisons with HiCFoundation and Hi-Cformer

We thank the reviewer for this constructive suggestion. We have now conducted direct comparisons with both HiCFoundation and Hi-Cformer on cell-level scHi-C tasks.

Our original benchmark focused on cell-level scHi-C methods. HiCFoundation pretrains genomic contact windows and aggregates regional features for single-cell applications, whereas Hi-Cformer trains multiscale cell-level representations on target datasets. Hi-Cell instead pretrains a whole-cell representation with a cell token encoding genome-wide cis contacts and global attention integrating chromosome and multiscale patch information.

Across six external datasets, we evaluated HiCFoundation's official pretrained encoder, Hi-Cformer's target-trained embeddings (depth-normalized training variant on Kim2020), and Hi-Cell's frozen pretrained encoder. Annotation uses a shared supervised MLP readout.

MethodMean NMI ↑Mean ARI ↑Mean macro-F1 ↑
HiCFoundation0.24680.15720.6911
Hi-Cformer (target-trained)0.59310.55480.8126
Hi-Cell (zero-shot)0.64230.59250.8569

Hi-Cell achieves the highest means without target-dataset encoder fitting: NMI rises from 0.5931 to 0.6423 and macro-F1 from 0.8126 to 0.8569 versus Hi-Cformer, supporting transferable cell-type-associated representations.

Imputation uses 10k/25k/50k inputs on six datasets, with matched Hi-Cell predictions and references: 14-chromosome support versus HiCFoundation and 23-chromosome support versus Hi-Cformer.

Matched benchmarkMethodAUPRC ↑Held-out SCC ↑LOO SCC ↑
6 datasets, 14 chrHiCFoundation0.13920.19830.3627
6 datasets, 14 chrHi-Cell0.28930.32350.7296
6 datasets, 23 chrHi-Cformer0.13340.14810.3988
6 datasets, 23 chrHi-Cell0.29730.34780.7460

Hi-Cell achieves higher reconstruction scores in both matched comparisons. Against Hi-Cformer, held-out SCC increases from 0.1481 to 0.3478 and LOO SCC from 0.3988 to 0.7460. Together, the results support transferable structural priors and cellular representations from one pretrained checkpoint without target-specific encoder fitting.

2. Progressive pretraining and hyperparameter sensitivity

We thank the reviewer for the constructive suggestions. All variants below start from the same real-cell-pretrained Stage-I checkpoint. Stage II adds pseudo-single-cell-to-bulk supervision; Stage III adds reconstruction of original real-cell profiles from binomially thinned inputs. We evaluate six-dataset means; imputation averages 10k/25k/50k contacts per cell on the original 23-chromosome support.

Training sequenceNMI ↑Held-out SCC ↑LOO SCC ↑
Stage I0.59040.30310.6536
I → II (real-only)0.61350.30180.6532
I → II (real + pseudo)0.61050.34300.7392
I → II (real-only) → III (real-only)0.62450.31880.6843
I → II (mixed) → III (no thinning)0.63350.34570.7432
I → II (mixed) → III (Full Hi-Cell)0.64230.34780.7460

Pseudo-to-bulk supervision improves structural reconstruction beyond additional real-cell optimization. Under the same Stage-II training budget, mixed supervision achieves 0.7392 LOO SCC versus 0.6532 for real-only training; after full continuation, Hi-Cell reaches 0.7460 versus 0.6843. Real-cell cross-depth refinement adds further gains: removing Stage-III thinning gives held-out/LOO SCC of 0.3457/0.7432, versus 0.3478/0.7460 for Full Hi-Cell.

We additionally varied thinning probability and real/pseudo loss weights during Stage III, starting from the same Stage-II checkpoint and using the same update budget:

Stage-III settingNMI ↑Held-out SCC ↑LOO SCC ↑
Default: q ~ U(0.25, 0.75), weights 9:10.64230.34780.7460
Fixed q = 0.5, weights 9:10.64800.34660.7451
q ~ U(0.25, 0.75), weights 1:10.63290.34760.7463

Both alternatives preserve comparable structural reconstruction performance. Together, the experiments support the complementary roles of real-cell pretraining, bulk-derived structural supervision, and cross-depth refinement.

Comments on the paper

2 comments

Response to Reviewer TBmD (Part 1/2)

We appreciate the reviewer's positive feedback on the value of reusable cell-level modeling, our hierarchical representation, and the complementary use of real and bulk-derived Hi-C data. In response to the reviewer's constructive suggestions, we have conducted direct comparisons with HiCFoundation and Hi-Cformer, controlled experiments on model capacity and pretraining data coverage, and additional ablations of the three-stage pretraining strategy.

1. Comparisons with HiCFoundation and Hi-Cformer

We thank the reviewer for emphasizing comparisons with recent Hi-C models. We have added direct evaluations of HiCFoundation and Hi-Cformer to assess the practical value of reusable cell-level pretraining.

HiCFoundation primarily pretrains regional contact windows and aggregates their representations for single-cell applications. Hi-Cformer provides multiscale cell-level modeling through target-dataset training. Hi-Cell jointly models cell-, chromosome-, and multiscale patch-level contacts, while pretraining a shared backbone to transfer cellular representations and structural priors to unseen datasets.

We compared HiCFoundation's official pretrained encoder, Hi-Cformer's target-trained embeddings (including a depth-normalized training variant on Kim2020), and frozen zero-shot Hi-Cell representations across six external datasets. Annotation uses the same supervised MLP readout.

MethodMean NMI ↑Mean ARI ↑Mean macro-F1 ↑
HiCFoundation0.24680.15720.6911
Hi-Cformer (target-trained)0.59310.55480.8126
Hi-Cell (zero-shot)0.64230.59250.8569

Hi-Cell achieves the highest mean NMI, ARI, and macro-F1 without target-dataset encoder fitting. In particular, its mean NMI/ARI of 0.6423/0.5925 compares with 0.5931/0.5548 for Hi-Cformer, showing that cellular identity information transfers effectively from a shared pretrained model.

For imputation, six-dataset comparisons use matched 10k/25k/50k inputs and references: 14 chromosomes versus HiCFoundation and 23 versus Hi-Cformer.

Matched benchmarkMethodAUPRC ↑Held-out SCC ↑LOO SCC ↑
6 datasets, 14 chrHiCFoundation0.13920.19830.3627
6 datasets, 14 chrHi-Cell0.28930.32350.7296
6 datasets, 23 chrHi-Cformer0.13340.14810.3988
6 datasets, 23 chrHi-Cell0.29730.34780.7460

Hi-Cell achieves higher structural reconstruction scores in both matched settings. Against Hi-Cformer, held-out SCC reaches 0.3478 versus 0.1481 and LOO SCC 0.7460 versus 0.3988. These results support Hi-Cell's central contribution: one pretrained cell-level backbone enables transferable representations and zero-shot reconstruction instead of fitting a new encoder for each target dataset.

2. Model capacity and pretraining data coverage

We thank the reviewer for this valuable suggestion. To examine the contributions of model capacity and pretraining data coverage, we conducted a controlled 2×2 experiment comparing Full (594M parameters) and Small (171M parameters) models trained on 100% or 25% of the real-cell pretraining data. All four models start from random initialization and use the same Stage-I reconstruction objective and 10,000 training updates. We evaluate their zero-shot clustering performance across six external datasets.

Model25% Data (NMI / ARI) ↑100% Data (NMI / ARI) ↑
Full (594M)0.6041 / 0.58040.6626 / 0.6628
Small (171M)0.5956 / 0.56160.6201 / 0.5922

Both model capacity and training-data coverage contribute to representation quality. With full data, the larger model improves NMI/ARI by 0.0425/0.0706 over Small. Increasing data coverage from 25% to 100% improves Full by 0.0585/0.0824, compared with 0.0245/0.0306 for Small. These results show that the larger architecture benefits more from broader pretraining data coverage.

The value of pretraining is further supported by our existing initialization ablation (Figure 4A–C). With the same architecture, target data, and label-free adaptation objective, pretrained Hi-Cell achieves a mean annotation macro-F1 of 0.8685 after approximately 1,200 target-training updates, compared with 0.6929 for random initialization even after approximately 4,800 updates. This demonstrates that pretraining provides transferable representations that substantially improve downstream adaptation.

Together with the direct HiCFoundation comparison above, these experiments provide complementary evidence for the benefits of model capacity, large-scale pretraining data, and reusable pretrained representations. We will incorporate the new results into the revised manuscript.

Response to Reviewer TBmD (Part 2/2)

3. Contributions of the three-stage pretraining strategy

We thank the reviewer for the insightful suggestions. We have expanded our ablation studies with intermediate checkpoints, a direct Stage-I-to-III continuation, and controlled real-only and pseudo-only continuations to clarify the contributions of individual training components.

All variants start from the same Stage-I checkpoint pretrained on real single-cell Hi-C data. In the full recipe, Stage II introduces pseudo-single-cell-to-bulk supervision, while Stage III performs cross-depth refinement through binomial thinning of real-cell inputs.

Training sequenceNMI ↑Macro-F1 ↑Held-out SCC ↑LOO SCC ↑
Stage I0.59040.85290.30310.6536
I → II (real-only)0.61350.85450.30180.6532
I → II (mixed)0.61050.84020.34300.7392
I → III (skip II)0.61860.86030.31720.6900
I → real-only continuation0.62450.85570.31880.6843
I → pseudo-only continuation0.56150.80480.33110.7209
I → II → III (no thinning)0.63350.85580.34570.7432
I → II → III (Full Hi-Cell)0.64230.85690.34780.7460

All results are six-dataset means, with imputation evaluated at 10k, 25k, and 50k contacts per cell on the original 23 chromosomes. The real-only and pseudo-only continuations use the same Stage-II/III update budgets as Full Hi-Cell, while I → III skips Stage II entirely. Full Hi-Cell and the no-thinning control use validation-selected checkpoints; the other new controls use terminal checkpoints.

Bulk-derived supervision contributes substantially to structural reconstruction. After the same Stage-II training budget, mixed supervision achieves 0.7392 LOO SCC, compared with 0.6532 for real-only training. After the complete continuation, Full Hi-Cell reaches 0.7460, compared with 0.6843 for real-only and 0.7209 for pseudo-only continuation. These results demonstrate the complementary benefits of real-cell and pseudo-to-bulk supervision.

Stage III further integrates cellular representation learning with structural reconstruction. Compared with the Stage-II checkpoint, Full Hi-Cell improves mean NMI from 0.6105 to 0.6423 and macro-F1 from 0.8402 to 0.8569. Skipping Stage II yields 0.6900 LOO SCC, highlighting the contribution of the intermediate bulk-derived supervision. Moreover, removing real-cell thinning in Stage III yields 0.7432 LOO SCC, compared with 0.7460 for Full Hi-Cell, providing additional evidence for cross-depth refinement.

Together, these experiments support the complementary roles of the three stages: establishing cellular representations from real cells, incorporating bulk-derived structural priors, and refining the shared model across contact depths. We will include these results in the revised manuscript.

4. Fine-scale patches and the full hierarchy

We thank the reviewer for this observation. Importantly, the Fine {8,16} variant retains both the cell- and chromosome-level pathways, including direct access to chromosome-wide and whole-cell contact information. It removes only the coarser patch scales, rather than the hierarchical representation itself.

In our matched label-free adaptation experiments, Fine {8,16} achieves NMI/ARI of 0.7247/0.7146 and macro-F1 of 0.8667, compared with 0.7225/0.7195 and 0.8685 for the Full hierarchy. These results highlight the importance of fine-grained contact patterns for cellular identity, while demonstrating that effective representations can be learned using fewer patch scales together with global cell/chromosome context.

The full hierarchy serves a broader purpose: jointly modeling local, chromosome-wide, and whole-cell structure through a shared representation and multilevel reconstruction framework. We will clarify this distinction in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.