acceptodds
Under review as a conference paper at ICLR 2027

CellHIST: A Cell-Centric Dataset and Benchmark for Weakly-Supervised Virtual Spatial Transcriptomics from Histopathology

Abstract

Spatial Transcriptomics (ST) at single-cell resolution has revolutionized our understanding of spatial biology. However, high experimental costs impede its widespread adoption, driving virtual ST from histopathology. Current approaches either require expensive single-cell labels for training or rely on paired spot labels at inference. An emerging paradigm, weakly-supervised learning (WSL), overcomes this by training a model on cost-effective spot-level ST data to infer single-cell spatial gene expression solely from images. Although offering a highly scalable alternative, a single-cell-level dataset and benchmark framework are still lacking. To bridge this gap, this paper presents CellHIST. It builds the first *cell-centric* dataset for virtual single-cell ST from histopathology. CellHIST-Data curates 169 samples across 14 organs, encompassing around 36 million cells and 1 million image patches, standardized to support model development across various paradigms. Furthermore, a comprehensive benchmark, CellHIST-Bench, is introduced to systematically evaluate weakly-supervised virtual single-cell ST in terms of predictive performance, data scaling behavior, single-cell prediction specificity, biological relevance, and downstream utility. Results uncover the practicability and limitations of WSL for this task. This cell-centric study can bridge an important gap and pave the way for histopathology-driven virtual single-cell ST.

Primary area:
datasets and benchmarks
Keywords:
Spatial Transcriptomics, Histopathology, Single-Cell Spatial Gene Expression, Weakly Supervised Learning, Benchmark Dataset

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_Uaq1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CellHIST, a standardized dataset and benchmark for weakly supervised single-cell gene expression prediction from histopathology. It collects 169 samples across 14 organs and evaluates existing methods under different supervision settings, with additional analyses of data scaling, cell specificity, and downstream biological applications.

Critical strengths

The paper addresses an important gap by providing a large-scale, standardized dataset covering 169 samples across 14 organs. This could be a useful resource for future research on virtual single-cell spatial transcriptomics.

The paper compares several weakly supervised methods with fully supervised and super-resolution approaches.

Critical weaknesses

The number of evaluation samples varies considerably across organs. In particular, Ovary and Skin have only two and three single-cell samples, respectively. This makes it difficult to draw strong conclusions about model performance and generalization on these tissues.

Questions for authors

The cell-matching coverage is relatively low for some tissues, especially Lung. Could this affect the evaluation results or introduce bias toward certain cell populations?

Given the differences in gene panels and sequencing platforms across datasets, how much do these factors influence the benchmark results?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Uaq1(Part 1 / n)

Thank you for your encouraging comments and constructive feedback. We appreciate your positive assessment of CellHIST’s value as a large-scale, standardized resource for virtual single-cell spatial transcriptomics and of our experimental design comparing weakly supervised, fully supervised, and super-resolution approaches.

Next, we provide a detailed point-to-point response to address your questions and concerns.

AuthorsOpenReview ↗

Response to Reviewer Uaq1 (Part 2 / n)

W1: The number of evaluation samples varies considerably across organs. In particular, Ovary and Skin have only two and three single-cell samples, respectively. This makes it difficult to draw strong conclusions about model performance and generalization on these tissues.


AW1: Thanks for pointing this out and for your thoughtful assessment of its potential limitations. We really appreciate your feedback that helps strengthen this paper.

Response summary: More samples for some organs (e.g., Skin) would help strengthen the generalizability of the organ-specific findings. It is also our ultimate goal. However, the number of available single-cell-resolution Xenium samples is growing slowly because of the high cost (approximately $10k per sample). This work represents an initial step in this area by contributing the first standardized cell-centric dataset for virtual single-cell ST. We will continue collecting additional samples from public resources in our future work. We have discussed this issue in our revised paper.

Our detailed responses are as follows:

(1) Data size comparison between CellHIST and existing studies

We would like to note that single-cell ST datasets remain limited in scale because of the high cost and technical complexity of data generation. Related studies have also used limited numbers of samples for model development and evaluation, as summarized in Table R1.

Concretely, for Skin, GHIST used one melanoma sample and SPoTLIghT used two, whereas CellHIST incorporates these public sources and an additional Xenium Prime 5K sample, yielding three slides. For Ovary, UTOPIA analyzed one Xenium 5K sample, while MIST includes two FFPE ovarian cancer datasets and one fresh-frozen ovarian dataset; CellHIST incorporates the two FFPE datasets. For Bowel, MISO evaluated high-resolution expression prediction using one colorectal cancer Xenium sample, whereas our evaluation pool includes five slides.

Concretely, for skin, GHIST [A] used one melanoma sample and SPoTLIghT [B] used two, whereas CellHIST incorporates these public sources and an additional Xenium Prime 5K sample, yielding three slides. For bowel, MISO [C] evaluated high-resolution expression prediction using one colorectal cancer Xenium sample, whereas our evaluation pool includes five slides. For ovary, UTOPIA [D] analyzed one Xenium 5K sample, while MIST [E] includes two FFPE ovarian cancer datasets and one fresh-frozen ovarian dataset; CellHIST incorporates the two FFPE datasets.

Table R1. Data size comparison between CellHIST and existing studies.

Method/resourcePublication venueOrganNumber of samplesSlides in CellHIST evaluation pool
GHISTNature Methods, 2025Skin13
SPoTLIghTnpj Precision Oncology, 2024Skin23
MISONature Communications, 2025Bowel15
UTOPIAbioRxiv preprint, 2026Ovary12
MISTICML 2026Ovary32

These studies differ in their prediction tasks, evaluation settings, and sample inclusion criteria. The comparison therefore contextualizes organ-specific sample availability rather than predictive performance or the number of independent patients. It does not establish that the available samples are sufficient for broad generalization.

The high cost and technical complexity of spatial transcriptomic profiling remain barriers to assembling large patient cohorts. Suitable data for our evaluation additionally require paired high-resolution histology and reliable cell-level correspondence. CellHIST addresses this resource gap through systematic curation and standardized processing. Nevertheless, these efforts do not eliminate the limitations associated with small and uneven organ-specific cohorts. In particular, the Ovary and Skin results may be sensitive to the characteristics of individual slides, and differences in sample size and cohort composition warrant caution when comparing performance across organs. We therefore restrict our organ-specific conclusions to the evaluated cohorts and plan to incorporate additional independent patient samples, prioritizing underrepresented organs. We also plan to publicly release the expanded resources, subject to data-sharing permissions.

(2) Our revisions

We have added the following discussion of sample availability and generalizability in Section 6 (Discussion):

Limited single-cell ST samples, particularly for Ovary, Skin, and Bowel, may constrain the generalizability of our organ-specific results. Single-cell-resolution samples are expensive, and the data scale increases slowly in this field. This study has curated as many high-quality cancer samples as possible from public resources. We will gather more ST data in the future to improve CellHIST.

Please refer to page 10, where this discussion is colored in red.

AuthorsOpenReview ↗

Response to Reviewer Uaq1 (Part 3 / n)

Q1: The cell-matching coverage is relatively low for some tissues, especially Lung. Could this affect the evaluation results or introduce bias toward certain cell populations?


AQ1: We thank the reviewer for their careful reading. This prompts us to examine how matching coverage relates to prediction performance and whether the matching process preferentially retains certain cell populations, thus helping us better understand its potential influence on the evaluation results. Insightful comments.

Response summary: Our Lung analyses show that cell matching changes the evaluated cell population, with T and B cells enriched in the matched subset. Lower matching coverage also coincides with lower PCC across all seven methods. These findings indicate the differences in cell-type representation caused by cell matching and an association between matching coverage and prediction performance.

Next, we describe the experimental setup, results, and our revisions.

(1) Experimental setup

Cell-type representation. We compared cell-type proportions between the full Xenium cell population and the successfully matched subset, separately for the Xenium and post-Xenium groups.

Performance across alignment groups. We compared matching coverage and prediction performance across the two groups, comprising four Xenium slides and 16 post-Xenium slides. Performance was measured using gene-wise PCC on TargetH50 with assay-CPM log1p normalization. We calculated equal-weight slide averages and report the mean ± SD across three training seeds.

(2) Results

Cell matching changes cell-type representation. As shown in the table below, the matched subset had higher proportions of T and B cells in both alignment groups. In contrast, the proportions of epithelial, stromal, myeloid, and endothelial cells each differed by less than two percentage points.

Table R2: Cell-type proportions before and after matching in the Lung cohort.

Cell typeXenium all → matchedPost-Xenium: all → matched
Epithelial30.81 → 31.6622.35 → 23.62
Stromal15.97 → 16.2232.98 → 31.10
Myeloid8.31 → 8.1813.26 → 14.05
Endothelial7.15 → 7.342.27 → 2.01
T cells6.63 → 8.052.65 → 3.82
B cells3.32 → 3.994.02 → 4.87
Unknown26.70 → 23.2821.92 → 19.92

Lower matching coverage coincides with lower prediction performance. The post-Xenium group had lower matching coverage than the Xenium group (38.19% versus 75.33%) and lower PCC across all seven methods, as shown below.

Table R3: Matching coverage and prediction performance across Lung alignment groups.

MeasureXeniumPost-Xenium
Slides416
Matching coverage75.33%38.19%
ABMIL0.237 ± 0.0020.075 ± 0.003
ABMIL-gated0.236 ± 0.0070.076 ± 0.004
GraphMIL0.217 ± 0.0030.063 ± 0.016
MS-GraphMIL0.103 ± 0.0070.023 ± 0.002
Transformer0.212 ± 0.0020.085 ± 0.010
DeepSpot2Cell0.227 ± 0.0000.061 ± 0.004
sCellST0.229 ± 0.0030.058 ± 0.008

PCC: TargetH50, assay-CPM log1p normalization; equal-weight slide averages, reported as mean ± SD across three seeds.

Residual alignment errors can push nuclear centroid distances beyond our 5 μm threshold and affect retained image–expression correspondences. Consistent with the nuclear-geometry-based matching used by Xenium and GHIST, post-Xenium samples showed fewer distance-threshold exclusions for T/B cells (3.84%/4.57%) than for epithelial/stromal cells (17.98%/16.89%).

Together, these results show that matching changes cell-type representation and that the lower-coverage cohort also yields poorer prediction scores. The cohort comparison supports an association with alignment quality, rather than establishing causality.

(3) Our revisions

In response, we have made the following revisions on Pages 20 and 21:

  • Add the cell-type proportion comparison to show how matching changes the composition of the evaluated cell population.

  • Report matching coverage and prediction performance separately for the Xenium and post-Xenium groups.

  • Add the corresponding discussion of potential alignment-related effects and clarify that the reported performance is evaluated on the matched cell population.

Reviewer_Uu5zOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces CellHIST, a dataset and benchmark for histopathology-driven virtual single-cell spatial transcriptomics. It focuses on weakly supervised learning that uses spot-level gene expression labels for training and predicts single-cell spatial gene expression solely from histopathology images at inference. CellHIST-Data curates and standardizes 169 samples across 14 organs from public resources, comprising approximately 36 million detected nuclei and one million image patches. It provides cell embeddings, patch–cell correspondence information, and spatial gene expression labels at different resolutions. Using this dataset, CellHIST-Bench evaluates weakly supervised methods alongside fully supervised and spatial super-resolution references across five organs, and investigates data scaling, cell specificity, and downstream utility.

Critical strengths
  1. The motivation is clear and practically relevant. Training with spot-level supervision and predicting single-cell spatial gene expression solely from histopathology images offers a potentially scalable approach that reduces reliance on costly single-cell spatial transcriptomics measurements.
  2. The experimental evaluation is comprehensive and well designed. It includes comparisons across supervision paradigms, data scaling analyses, cell-specificity tests, and downstream applications, providing a thorough assessment of the potential and limitations of weakly supervised single-cell expression prediction.
Critical weaknesses
  1. The incremental contribution of the dataset relative to existing resources is not sufficiently established. All samples are curated from public resources, with 121 of the 169 samples obtained from HEST-1K. Although the additional samples and standardized cell-centric processing are useful, a clearer comparison with existing datasets and benchmarks is needed to establish the added value of CellHIST.
  2. The limited number of independent samples in some organs constrains the generalizability of the organ-specific findings. The single-cell evaluation pool contains only two Ovary slides, three Skin slides, and five Bowel slides. The large number of cells does not compensate for the small number of independent patients or slides, and variation across training seeds does not capture uncertainty across patients.
Questions for authors
  1. I was unable to access the dataset link provided in the manuscript. Could the authors clarify the current access procedure and whether all data and patient-level splits used in the benchmark will be made available?
  2. CellHIST combines data from different ST platforms and applies platform-dependent processing, including pseudo-spot construction, expression normalization, and histology–transcriptomics alignment. How does the experimental design account for potential measurement differences and processing errors? In particular, how should the performance differences between the standard-spot and pseudo-spot settings be interpreted, given that these settings differ in training cohorts and measurement platforms?
7 comments
AuthorsOpenReview ↗

Response to Reviewer Uu5z (Part 1 / n)

Thanks for your encouraging comments and feedback. We appreciate your recognition of this work’s clear motivation, practical relevance, comprehensive and well-designed experiments, and clear presentation.

Next, we provide a detailed point-to-point response to address your questions and concerns.

AuthorsOpenReview ↗

Response to Reviewer Uu5z (Part 2 / n)

W1: The incremental contribution of the dataset relative to existing resources is not sufficiently established. All samples are curated from public resources, with 121 of the 169 samples obtained from HEST-1K. Although the additional samples and standardized cell-centric processing are useful, a clearer comparison with existing datasets and benchmarks is needed to establish the added value of CellHIST.


AW1: Thanks for your thoughtful comment that helps to make the added value of CellHIST clearer. We really appreciate your wonderful suggestion.

Response summary: Following your suggestion, we have added a clear comparison with existing datasets and benchmarks in Section 2, Table 1, on page 3. The comparison distinguishes virtual ST resolution, data standardization, cell image embeddings, and benchmark scope. CellHIST provides standardized cell-centric data and cell image embeddings, together with a benchmark of weakly and fully supervised single-cell SGE prediction. The revised discussion also clarifies that existing single-cell resources may lack paired histopathology images or available standardized data with quality control, or may be processed at pseudo-spot resolution for benchmarking.

Our detailed responses are presented below.

(1) Clearer comparisons between CellHIST and existing public datasets and benchmarks

We provide a detailed table for comparisons, as follows:

Table R1. An overview of comparison between CellHIST and existing public datasets and benchmarks.

NameVirtual ST resolutionData standardizationCell image embeddingsBenchmark scope
HEST-1K [A]Spot✓✗Pathology foundation models
STimage-1K4M [B]Spot✓✗Spot-level SGE prediction
SpaRED [C]Spot✓✗Spot-level SGE prediction and completion
DeepSpot2Cell [D]Single cell✗✗Spot-supervised single-cell SGE deconvolution and prediction
CellHIST (ours)Single cell✓✓Weakly and fully supervised single-cell SGE prediction

Note: Data standardization and cell image embeddings follow the comparison in Table 1 of the revised manuscript. ✓/✗ denotes yes/no.

(2) Further clarifications on the difference between CellHIST and DeepSpot2Cell

The comparison between CellHIST and DeepSpot2Cell is further clarified as follows:

  • CellHIST was curated using strict inclusion criteria: cancer specimens, FFPE preservation, and H&E image resolution finer than 0.75 μm/pixel. All 85 single-cell records satisfy these metadata criteria. Applying the same criteria to the reported DeepSpot2Cell collection retains only 9 records, because its 20 lung samples originate from a pulmonary fibrosis and healthy-control cohort.

  • CellHIST adopts cell-centric data QC and standardization. As described in Appendix A.1, our cell-centric processing pipeline organizes histology images, expression matrices, segmentation geometries, image features, and cell correspondences into standardized sample assets. Strict one-to-one correspondences use reciprocal nearest-neighbor matching with a maximum distance of 5 μm and exact-tie rejection. The accompanying QC summaries and visualizations make image scale, spatial layout, segmentation, and coordinate consistency inspectable at the individual-sample level. However, these steps are NOT explicitly given in DeepSpot2Cell.

  • Moreover, CellHIST’s single-cell collection includes 85 samples, with 76 (89.4%) NOT represented in the DeepSpot2Cell benchmark. The overlap only comprises seven breast samples and two pancreatic samples. Moreover, CellHIST covers 13 organ/tissue categories, compared with 3 organs in DeepSpot2Cell.

(3) Our revisions

We have added this comparison as Table 1 in Section 2 on page 3, together with the corresponding discussion of the resource gap for virtual single-cell ST. Both additions are colored in red in the revised manuscript.

[Reference A] Jaume et al. HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis. NeurIPS, 2024

[Reference B] Chen et al. STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics. NeurIPS, 2024.

[Reference C] Ruiz et al. Completing spatial transcriptomics data for gene expression prediction benchmarking. Medical Image Analysis, 2025.

[Reference D] Nonchev et al. DeepSpot2Cell: Predicting Virtual Single-Cell Spatial Transcriptomics from H&E images using Spot-Level Supervision. NeurIPS 2025 Workshop on Imageomics, 2025.

AuthorsOpenReview ↗

Response to Reviewer Uu5z (Part 3 / n)

W2: The limited number of independent samples in some organs constrains the generalizability of the organ-specific findings. The single-cell evaluation pool contains only two Ovary slides, three Skin slides, and five Bowel slides. The large number of cells does not compensate for the small number of independent patients or slides, and variation across training seeds does not capture uncertainty across patients.


AW2: Thanks for pointing this out and for your thoughtful assessment of its potential limitations. We really appreciate your feedback that helps strengthen this paper.

Response summary: More samples for some organs (e.g., Skin) would help to strengthen the generalizability of the organ-specific findings. It is also our ultimate goal. However, the number growth of single-cell-resolution Xenium samples is extremely slow due to the expensive cost (around $10K per sample). This work serves as an initiative in this area by contributing the first standardized cell-centric dataset for virtual single-cell ST. We will continue to collect more samples from public resources in our future work. We have stated this limitation in our revised paper.

Our detailed responses are as follows:

(1) Data size comparison between CellHIST and existing studies

Single-cell ST samples remain limited in scale because of the high cost and technical complexity of data generation. Related studies have also used limited numbers of samples for model development and evaluation, as summarized in Table R2.

Concretely, for skin, GHIST [A] used one melanoma sample and SPoTLIghT [B] used two, whereas CellHist incorporates these public sources and an additional Xenium Prime 5K sample, yielding three slides. For bowel, MISO [C] evaluated high-resolution expression prediction using one colorectal cancer Xenium sample, whereas our evaluation pool includes five slides. For ovary, UTOPIA [D] analyzed one Xenium 5K sample, while MIST [E] includes two FFPE ovarian cancer datasets and one fresh-frozen ovarian dataset; CellHIST incorporates the two FFPE datasets.

Table R2. Data size comparison between CellHIST and existing studies.

Method/resourcePublication venueOrganNumber of samplesSlides in CellHIST evaluation pool
GHISTNature Methods, 2025Skin13
SPoTLIghTnpj Precision Oncology, 2024Skin23
MISONature Communications, 2025Bowel15
UTOPIAbioRxiv preprint, 2026Ovary12
MISTICML 2026Ovary32

These studies differ in their prediction tasks, evaluation settings, and sample inclusion criteria. The comparison therefore contextualizes organ-specific sample availability rather than predictive performance or the number of independent patients.

The high cost and technical complexity of spatial transcriptomic profiling remain barriers to assembling large patient cohorts, as also noted by GHIST and Path2Space. Suitable data for our evaluation additionally require paired high-resolution histology and reliable cell-level correspondence. CellHIST addresses this resource gap through systematic curation and standardized processing. Nevertheless, these efforts do not eliminate the uncertainty associated with small organ-specific cohorts. We therefore restrict our organ-specific conclusions to the evaluated cohorts and plan to incorporate additional independent patient samples and publicly release the expanded resources, subject to data-sharing permissions.

(2) Our revisions

We have added the following discussion of sample availability and generalizability in Section 6 (Discussion):

Limited single-cell ST samples, particularly for Ovary, Skin, and Bowel, may constrain the generalizability of our organ-specific results. Single-cell-resolution samples are expensive, and the data scale increases slowly in this field. This study has curated as many high-quality cancer samples as possible from public resources. We will gather more ST data in the future to improve CellHIST.

Please refer to page 10, where this discussion is colored in red.

[Reference A] Fu et al. Spatial gene expression at single-cell resolution from histology using deep learning with GHIST. Nature Methods, 2025

[Reference B] Lapuente-Santana et al. Integrating histopathology and transcriptomics for spatial tumor microenvironment profiling in a melanoma case study. npj Precision Oncology, 2024

[Reference C] Schmauch et al. A deep learning-based multiscale integration of spatial omics with tumor morphology. Nature Communications, 2025

[Reference D] Jin et al. Multiscale confidence quantification for virtual spatial transcriptomics with UTOPIA. bioRxiv, 2026

[Reference E] Kong et al. SPATIA: Multimodal Generation and Prediction of Spatial Cell Phenotypes. ICML, 2026.

Comments on the paper

11 comments
AuthorsOpenReview ↗

Response to Reviewer FvEw (Part 1 / n)

We appreciate your recognition of this study, including the dataset's value and good experimental designs, as well as your constructive comments that help improve this paper.

Next, we provide a detailed point-to-point response, supplemented with additional experiments, to address your questions and concerns.

AuthorsOpenReview ↗

Response to Reviewer FvEw (Part 2 / n)

W1: Most of the correlation in bowel remains after shuffling. The drop suggests the models learn something about individual cells, but recognizing a tumor cell versus a lymphocyte could explain part of it. I’d like to see what happens when expression profiles are shuffled only among nearby cells of the same type.


AW1: Thank you for your careful reading and insightful feedback. It is really a great idea to shuffle only among nearby cells of the same type, because doing this could help to deeply understand the cause of the performance drop and the capability of weakly-supervised models.

Response summary: According to your feedback, we have conducted this experiment. Results show that PCC drops by less than 0.01 for most methods (avg. drop: 0.01), and it is always smaller than the drop of previous unrestricted cell shuffling (avg. drop: 0.032). This comparison suggests that weakly-supervised models mainly recognize expression differences between different cell types.

Next, we provide the details, including experimental setup, results, and our revisions.

(1) Experimental setup

We inferred broad cell-type identities using a panel-matched multinomial logistic regression classifier trained on an external scRNA-seq reference (GSE132465; Lee et al., Nature Genetics, 2020). Nine broad cell types were retained for subsequent analyses; cells failing the annotation criteria were labeled as unknown.

Three subsequent steps are conducted to do local shuffling within the same cell type, as follows:

  • Same-type shuffling. Within each local pseudo-spot, we shuffled complete measured expression profiles among cells of the same inferred cell type. This preserved local cell-type composition while disrupting the correspondence between individual cells and their measured expression profiles.

  • Evaluation subset. We evaluated cells with valid cell-type annotations that shared a pseudo-spot with at least one other cell of the same inferred type.

  • Controlled comparison. We evaluated both the unrestricted within-spot shuffling strategy (Section 5.5) and the proposed nearby cells of the same type shuffling strategy on exactly the same cell subset, using the unshuffled evaluation as a common reference. This controls for differences in the evaluated cell population when comparing the two shuffling strategies.

(2) Results

There are two key findings from Table R1, as follows:

First, same-type shuffling consistently produced smaller PCC drops than unrestricted within-spot shuffling across all seven models and both training settings. This indicates that broad cell-type differences contribute substantially to the performance degradation observed in our original shuffling experiment.

Second, PCC decreases remained positive for all models even when shuffling was restricted to cells of the same inferred broad type within each pseudo-spot. This suggests that broad cell-type differences do not fully explain the observed performance degradation and weakly-supervised models could also capture expression-related information beyond broad cell-type identity.

Table R1. Same-type shuffling compared with within-spots shuffling.

ModelStandard-spot:within-spot Δ\DeltaPCCStandard-spot: same-type Δ\DeltaPCCPseudo-spot: within-spot Δ\DeltaPCCPseudo-spot: same-type Δ\DeltaPCC
sCellST0.0370.0100.0410.008
ABMIL-gated0.0380.0180.0660.026
ABMIL0.0380.0190.0670.026
Transformer0.0230.0050.0340.007
GraphMIL0.0320.0070.0340.007
MS-GraphMIL0.0310.0030.0330.002
DeepSpot2Cell0.0260.0060.0440.011
Avg.0.0320.0100.0460.012

(3) Our revisions

We have added the above experimental results (Table R1) and their analysis to our revised paper. Please refer to pages 33 and 34 (colored in red) for details.

[Reference A] Lee et al. Lineage-dependent gene expression programs influence the immune landscape of colorectal cancer. Nature Genetics, 2020.

AuthorsOpenReview ↗

Response to Reviewer FvEw (Part 3 / n)

W2: The organ comparisons are hard to read. Bowel has 40 spot-level training samples, and ovary has two. For skin, the cancer type also changes between training and testing. It’s unclear how much of the difference comes from the tissue itself.


AW2: Thank you for pointing this out. Yes, the performance differences across organs may also come from data size and cancer type, besides the tissue itself.

Response summary: We intended to consider tissue heterogeneity, as it is a well-known factor that often affects model performance. However, when data size or cancer type significantly varies across organs, such comparisons are unfair. To remedy this, we have rephrased our analysis regarding the previous Table 1:1) generalization across cancer subtypes (observed from <0.051 in PCC on Skin) and 2) large data size is likely to result in better performance (often > 0.3 in PCC on Bowel). According to your suggestion, the rephrased analysis no longer involves comparisons across different organs.

(1) Rephrased analysis

We have removed the sentence “All WSL methods perform best on Bowel and worst on Skin, reflecting intrinsic heterogeneity across organs.” The revised analysis recognizes the different skin cancer subtypes in training and testing and treats the possible benefit of the larger Bowel dataset as an interpretation rather than an isolated causal effect.

(2) Our revisions

We have replaced the corresponding analysis in Section 5.2 on page 6 with the following text, colored in red:

(1) Generalization across cancer subtypes: All WSL methods obtain low mean PCCs on Skin (≤ 0.051). Skin involves different cancer subtypes in its training and test sets, squamous cell carcinoma and melanoma, respectively. These results indicate the difficulty of prediction across cancer subtypes. In addition to this factor, limited training samples may also contribute to the low performance on Skin. (2) Potential benefits of larger datasets: Most evaluated models achieve PCC values above 0.3 on Bowel. This relatively strong performance may be partially due to its larger dataset size. Additional factors, e.g., tissue characteristics and dataset composition, may also result in the observed performance differences on Bowel.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.