acceptodds
Under review as a conference paper at ICLR 2027

Predicting Spatial Gene Expression from Histology with a Single-Cell Reference and Orthogonal Residuals

Abstract

H&E slides are routinely available, whereas spatial transcriptomics (ST) is costly and limited in scale. This has motivated models that predict spatial gene expression from histology. Most existing methods, however, predict spot-level expression directly from image features without modeling the cell types that contribute to each spot. We introduce ScComp, which predicts composition weights over a fixed single-cell reference and combines the reference profiles accordingly. A residual captures expression variation not represented by the reference and is restricted to the orthogonal complement of the reference space. Across three patient-disjoint ST benchmarks, ScComp achieves the best equal-dataset macro performance among the compared methods with only 0.87M trainable parameters. More importantly, the composition weights can be applied beyond the genes used during training. They predict approximately 11,000 reference-covered genes never used as training targets, at both the gene and pathway levels, and agree with cell-type proportions measured by Xenium without composition supervision. Ablation results further show that the orthogonal projection helps preserve this zeroshot behavior by keeping the residual separate from variation represented by the reference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial Transcriptomics, SingleCell Reference, Gene Expression Prediction

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_HycSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ScComp, a framework for predicting spatial gene expression from H&E histology. ScComp predicts composition weights over a fixed single-cell reference matrix and combines the reference profiles to obtain a reference component of the predicted expression. A reference-orthogonal residual captures additional expression variation outside the space spanned by the reference profiles.

Critical strengths
  1. The paper presents a complete framework for histology-to-spatial-transcriptomics prediction, with evaluation across three patient-disjoint benchmarks. ScComp achieves strong predictive performance with a relatively small number of trainable parameters.
  2. The learned composition weights can be reused to predict reference-covered genes outside the training gene panel. The gene- and pathway-level evaluations provide useful evidence of this capability.
Critical weaknesses
  1. The overall narrative could be clearer and more focused. The introduction is somewhat lengthy, while the methodological novelty is not explained in sufficient depth. In particular, the distinction from existing single-cell reference-informed prediction methods could be articulated more explicitly.
  2. The motivation for the residual component is not sufficiently clear. The description suggests that it captures biologically meaningful expression variation not represented by the reference, but the paper does not establish which specific biological situations this component addresses. The evidence primarily supports its role as a mathematical correction constrained to the orthogonal complement of the reference space. A clearer distinction between this mathematical role and its possible biological interpretation would make the motivation easier to understand.
Questions for authors
  1. Could the authors explain more explicitly what constitutes the “single-cell reference”? In particular, please clarify what each reference profile represents and how these profiles are constructed and used to predict spot-level expression.
  2. Could the authors clarify what variation the reference space can represent and what its limitations are? A concrete example of expression variation that cannot be represented by the reference component would help explain the need for the residual and the rationale for the orthogonal constraint.
4 comments

We sincerely thank the reviewer for the careful review and constructive suggestions. For clarity, all changes in the revised manuscript are highlighted in blue. We address Q1 here and discuss W1 and Q2/W2 in the following comment.

Q1. Could the authors explain more explicitly what constitutes the “single-cell reference”? In particular, please clarify what each reference profile represents and how these profiles are constructed and used to predict spot-level expression.

Motivation and reference construction. In spot-based ST, each spot contains multiple cells, and its measured gene expression reflects their combined contributions. Single-cell RNA sequencing provides gene expression profiles for individual cell types and states, which can serve as a reference for estimating these contributions.

We construct the reference matrix SS using publicly available single-cell data from the CZ CELLxGENE Census. Each row corresponds to a (normal-source or cancer-source, annotated cell type) pair and represents its average log-normalized expression profile. The source category describes the donor diagnosis and does not imply that every cancer-source cell is malignant.

We first average expression within each donor and then average equally across donors, preventing donors with more cells from dominating a profile (Eqs. 2–3; Appendix C). This yields 68 reference rows for the breast datasets and 76 for the skin dataset.

Reference-based prediction. Given the histology image of a spot and its local neighborhood, ScComp predicts nonnegative weights π\pi that sum to one. Each weight corresponds to a reference profile, and the weighted combination πS\pi S produces the reference component of spot-level expression in log-normalized space.

The reference matrix remains fixed throughout training, so each predicted weight retains its association with the corresponding reference cell type or state. The final prediction combines this reference component with a learned gene-wise bias and an orthogonal residual, whose role we explain in the following comment.

W1. The overall narrative could be clearer and more focused. The introduction is somewhat lengthy, while the methodological novelty is not explained in sufficient depth. In particular, the distinction from existing single-cell reference-informed prediction methods could be articulated more explicitly.

Q2 & W2. Could the authors clarify what variation the reference space can represent and what its limitations are? A concrete example of expression variation that cannot be represented by the reference component would help explain the need for the residual and the rationale for the orthogonal constraint.

Reference-based prediction and its limitations. The reference consists of a finite set of donor-averaged profiles and may not capture all expression patterns in a tissue. For example, breast tumors exhibit patient-associated malignant expression patterns and distinct fibroblast states, illustrating variation that average reference profiles may not fully represent (Wu et al., 2021).

Residual and orthogonal constraint. To account for expression variation outside the reference space, ScComp introduces a residual component. Without an orthogonal constraint, the residual could reproduce expression patterns already represented by the reference, making it unclear which component explains a given signal. We therefore constrain the residual to the orthogonal complement of the row space of SS. This ensures that the residual is orthogonal to every reference profile and to πS\pi S (Eqs. 8–10), preventing it from duplicating reference-aligned expression patterns. The final prediction combines the reference component, a learned gene-wise bias, and the orthogonal residual.

Appendix I provides concrete spatial examples (lines 965–970; Figure 4). For COL1A2 in HER2ST section B5, PCC increases from −0.01 for the reference component to 0.68 for the final prediction. The corresponding improvements are 0.06 to 0.72 for KRT16 in cSCC section P2_ST_rep2 and 0.06 to 0.83 for EVL in STNet section BC23287_D1. These examples show spatial expression patterns that the reference component captures poorly and that the final prediction captures more accurately after adding the residual.

Distinction from existing methods. SciSt combines a single-cell-based expression estimate with learned image features, whereas CPNN updates its reference profiles during training. In contrast, ScComp keeps the reference fixed and constrains the residual to its orthogonal complement. This maintains a consistent correspondence between the predicted weights and the original reference profiles while modeling additional expression variation separately.

These properties support two capabilities demonstrated in our experiments: the predicted weights correlate with Xenium-derived cell-type proportions without composition supervision (Sec. 4.4), and they can be used to predict reference-covered genes outside the training panel without additional training (Sec. 4.3).

We have clarified the motivation and complementary roles of the fixed single-cell reference and the orthogonal residual in the Introduction (lines 067–074). We have also made the distinction from SciSt and CPNN explicit in Sec. 2.2, explaining how the fixed reference and orthogonal residual preserve the correspondence between the weights and the original reference profiles (lines 132–138). Concrete spatial examples are provided in Appendix I (lines 965–970; Figure 4).

I thank the authors for their response and revisions, which have improved the manuscript to some extent. However, substantial concerns remain that, in my view, cannot be adequately addressed within the rebuttal period. I encourage the authors to further develop and revise the manuscript before resubmission.

  1. The proposed residual formulation primarily constitutes a mathematical contribution, and its justification focuses largely on improvements in predictive performance. Its biological significance has not been sufficiently established. This concern partly reflects how the manuscript motivates and presents the method.

  2. The figures and the narrative structure of the Introduction still lack clarity. Beyond the residual formulation, the manuscript does not clearly articulate other compelling contributions. Since the residual formulation is primarily a mathematical operation, its novelty alone appears insufficient to support the overall contribution.

  3. I encourage the authors to develop a more biologically grounded motivation, strengthen the methodological contributions, and provide more direct validation of the proposed formulation’s potential biological significance. A substantially revised Introduction and clearer figures would also help communicate the contribution more effectively.

We thank the reviewer for the follow-up comments and for recognizing the improvements in the manuscript. We also appreciate that the initial review acknowledged the evidence for predictive performance, parameter efficiency, and prediction beyond the training gene panel. The comments helped us explain the role of the residual within the overall framework more clearly.

Our contribution is the combination of reusable, cell-type-associated reference weights and an orthogonal residual that complements their predictive capacity. Cell composition describes which cells are present and in what proportions, but it does not fully describe expression differences among cells of the same type. To test whether the residual complements prediction of these differences, we used Xenium measurements to split the expression of each tile into the part expected from cell-type counts and within-section mean profiles (the composition component) and the remainder (the within-type component). The analysis is performed at the level of 13 cell types.

Adding the residual increases the PCC with the within-type component from 0.170 to 0.257. The full model also exceeds a model trained separately without a residual (0.201), and both comparisons improve in all 12 evaluated donors (p < 0.001). The increase for the composition component is not significant (0.386 to 0.436, p = 0.11). These results support that the residual complements prediction of expression differences that are not captured by the expression expected from cell composition.

The gene-level results give these differences a biological context. Of the 308 evaluated gene–cell-type pairs, 78 improve significantly and one worsens. In luminal cells, 5 of the 17 evaluated Hallmark gene sets improve more than the remaining luminal genes. The largest effects are for E2F targets and G2M checkpoint, in which 15 of 19 and 14 of 17 genes improve significantly (q < 0.001 for both sets). In fibroblasts, four of the six evaluated ECM regulators improve significantly, including MMP11, which was first reported in stromal cells surrounding invasive breast carcinomas (PCC −0.20 to 0.38; Basset et al., 1990). These patterns connect the benefit of the residual to proliferation-related gene sets and to individual matrix-remodeling genes such as MMP11.

Together with the cell-type correspondence of the reference weights and their reuse for unseen genes and pathways, these results support the complementary value of the two components. We have added this analysis as Sec. 4.5 (lines [453–494], Figure 4) and provide the definitions, statistical procedures, and full results in Appendix K (Figures 7–8). We have also stated in the Introduction which expression differences the residual complements (lines [072–073, 083–085]). Figure and line numbers refer to the latest revision.

Reference

Basset, P., et al. (1990). A novel metalloproteinase gene specifically expressed in stromal cells of breast carcinomas. Nature, 348(6303), 699–704. doi:10.1038/348699a0

Reviewer_h47hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of predicting spot-level spatial gene expression from H&E histology images and proposes predicting composition weights over a fixed single-cell reference matrix of cell-type expression profiles. The predicted expression is formed by combining the weighted reference profiles with a residual restricted to the orthogonal complement of the reference space. Experiments are conducted on three spatial transcriptomics datasets against 11 baselines.

Critical strengths
  1. The studied problem is meaningful. Predicting spatial gene expression from routinely available H&E slides could greatly extend the reach of costly ST assays, and making such predictions interpretable is equally important.
  2. The proposed method is simple yet intuitive. Introducing cell-type information through a fixed single-cell reference gives the predicted composition weights a clear biological meaning.
Critical weaknesses
  1. The benchmarks used are relatively small. Large-scale benchmarks such as HEST-1k, which span many more organs, patients, and genes, would make the results more convincing and better test the generality of the single-cell reference across tissues.
  2. Another concern I have is the ablation study in Table 3, which suggests that the core design contributes little to in-panel performance. Replacing the reference–residual prediction with a direct expression head only reduces the macro HVG-50 PCC from 0.287 to 0.282. In contrast, removing the local neighborhood context or the patient-balanced update causes larger drops. This suggests that much of the improvement over baselines comes from components shared with the direct head, such as the frozen UNI features.
  3. All three gene-expression prediction datasets come from the first-generation ST platform (100 µm spots). The Xenium experiment evaluates only cell-type proportions, not gene-level prediction accuracy, and it also uses 100 µm tiles. It would be great to see gene-level results at higher resolution, such as Xenium with smaller tiles or Visium HD.
Questions for authors

See weaknesses.

2 comments

We sincerely thank the reviewer for the careful review and constructive suggestions. We appreciate the opportunity to clarify and strengthen our work. Changes made in response to the reviews are highlighted in blue in the revised manuscript.

W1. The benchmarks used are relatively small. Large-scale benchmarks such as HEST-1k, which span many more organs, patients, and genes, would make the results more convincing and better test the generality of the single-cell reference across tissues.

W3. All three gene-expression prediction datasets come from the first-generation ST platform (100 µm spots). […] It would be great to see gene-level results at higher resolution, such as Xenium with smaller tiles or Visium HD.

We appreciate the reviewer's suggestions regarding benchmark diversity and spatial resolution. To address these concerns, we conducted additional experiments on HEST-Bench to evaluate ScComp across additional tissue types and at finer spatial resolution.

For W1, we evaluated ScComp on three HEST-Bench datasets—CCRCC, IDC, and LYMPH_IDC—covering kidney, breast, and lymph node tissues. We selected datasets with at least four patients to ensure that at least two patients remained for training after assigning one patient each to validation and testing. All splits were patient-disjoint.

Regarding resolution (W3), we additionally evaluated the IDC Xenium task using square bins with a side length of 55 µm.

Table. Gene-level Pearson correlation on HEST-Bench tasks and the additional IDC evaluation with 55 µm bins

ModelCCRCCLYMPH_IDCIDC (100 µm)IDC (55 µm)
TRIPLEX0.1760.2320.4810.419
STFlow0.2220.2690.5430.485
ScComp (Ours)0.2980.2780.5790.514

ScComp achieves the highest PCC among the compared methods on all three HEST-Bench tasks and in the additional 55 µm evaluation. These results extend the evaluation beyond the breast and skin tissues used in our original benchmarks and show consistent performance at the finer spatial resolution.

We have included these additional HEST-Bench experiments in Appendix J, including the patient-disjoint evaluation protocol and the finer-resolution Xenium evaluation (lines 1089–1101; Table 11).

W2. Another concern I have is the ablation study in Table 3, which suggests that the core design contributes little to in-panel performance. […] This suggests that much of the improvement over baselines comes from components shared with the direct head, such as the frozen UNI features.

We thank the reviewer for highlighting the similar in-panel performance of ScComp and the direct head. We agree that their predictive accuracy is comparable (HVG-50 PCC: 0.287 vs. 0.282; exact two-sided Wilcoxon signed-rank test across 20 seed-averaged folds, p=0.35p = 0.35). Both heads benefit from spatial neighborhood context and patient-balanced updates. Importantly, ScComp offers two additional advantages: interpretable composition weights and the ability to predict genes beyond the training panel.

  • Interpretable composition weights: The predicted weights π\pi correlate with Xenium-derived cell-type proportions despite receiving no composition supervision (Spearman ρ=0.66\rho = 0.66; Sec. 4.4).
  • Prediction beyond the training panel: The predicted weights can also be combined with the fixed reference to predict approximately 11,000 genes outside the 250-gene training panel, without additional training. ScComp achieves a gene-level PCC of 0.187 and a pathway-level PCC of 0.209, outperforming the compared zero-shot baselines on average (Sec. 4.3, Table 2). In contrast, the direct head cannot predict genes outside its output panel.

Thus, the decomposition maintains competitive in-panel accuracy while enabling cell-type-associated interpretation and prediction beyond the training genes.

Regarding the frozen UNI features, we note that DKAN and STFlow also use the same frozen UNI encoder in our experiments. Thus, the performance differences cannot be explained by the image encoder alone.

We have added the direct-head statistical comparison in Sec. 4.5 (lines 461–467) and Appendix E (lines 859–863). The accompanying discussion clarifies that the decomposition retains comparable in-panel accuracy while supporting cell-type-associated interpretation and prediction beyond the training panel (lines 471–475).

Thanks for your explanation and experiments. I have no more questions at this point.

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

The approach aims to predict gene expression in spatial transcriptomics from histology (H&E). The authors specifically propose to leverage single-cell data to build a reference set of gene expression. Then, the framework first (i) extracts features from the H&E locally (at the spot location) and in the surrounding area (5x5 neighbourhood) using UNI (a frozen foundation model) and (ii) predicts the softmax weights π\pi over the reference profiles using a lightweight network. A residual component is added to capture signal that may not be contained in the reference set of single-cell gene expressions (orthogonal complement of the reference set obtained via SVD).

Experiments follow DKAN's protocol. The method is evaluated on three different datasets, reaching competitive performance. Experiments mostly focus on gene prediction with metrics such as the Pearson correlation coefficient and not on biologically relevant downstream tasks (e.g. domain identification, tumour-region detection or biomarker status.)

Critical strengths
  1. Leveraging single-cell data to build a decomposition basis of the ST gene profiles is a novel and interesting direction.
  2. The evaluation is rigorous: patient-disjoint folds, per-section PCC, 3 seeds, and a statistical test, 11 baselines.
Critical weaknesses
  1. The estimated ST profile is a convex combination π\pi of log-normalized profiles. A spot aggregates counts from several cells. Therefore, the physical mixture is linear in the count space not in the log domain. What is the justification of this choice? How does this approach scales to different signal amplitudes?

  2. Results are inconsistent with previously published ones. DKAN (AAAI 2026, Table 1) and TRIPLEX (CVPR 2024, Table 1) were evaluated with the same datasets, patient-disjoint folds and per-section PCC as in this work. For example, on HER2ST/cSCC/STNet, PCC on ALL DKAN reports 0.330/0.407/0.219 while this work reports 0.301/0.350/0.173. Similar observations hold for TRIPLEX. These changes are large, and ScComp scores are lower than these published results.

  3. The proposed decomposition contributes little in-panel. Against a parameter-matched direct head it adds 0.005 PCC (0.287 vs. 0.282), with no statistical test reported. Overall gains of the proposed approach seem to come from the 5x5 neighbourhood, as it adds 0.066 PCC.

  4. I am not convinced by experiments 4.4. Predicting cell composition from H&E is well-established. Showing that π\pi, which is predicted from H&E, retains this information is therefore expected rather than informative. What would be interesting is to see if a probe trained on real ST for tasks that are typically predictable (survival, tumour growth rate etc) can be used on the predicted ST.

  5. The approach is light but relies on UNI. I agree that the number of learnable parameters is low, but it leverages UNI parameters (300 million parameters)

Questions for authors
  1. Which components, inputs and settings of DKAN and TRIPLEX were changed that could explain the performance gap mentioned earlier?
  2. Which test supports "the difference on HER2ST alone is not significant". This sentence implies the cSCC and STNet differences are significant.
  3. How were the training hyperparameters chosen?
  4. How stable is π\pi for the 68 and 76-row references? Many rows are redundant. Depending on the seeds, how stable is π\pi f?
  5. How do you account for the amplitude of the signal? The proposed method heavily relies on log normalisation. Are there differences in terms of MAE depending on the depth?
  6. The experiment in 4.4 trains a separate model with a 13-row reference on Xenium. Why aren't the 68-row used in the main results not validated?
8 comments

We thank the reviewer for the careful review and constructive suggestions. We first clarify the training and evaluation procedures, statistical comparisons, and the role of the decomposition. We then address the stability and interpretation of the composition weights, the log-space formulation, downstream utility, and model size and training time. For clarity, all changes in the revised manuscript are highlighted in blue.

W2. Results are inconsistent with previously published ones. […] These changes are large, and ScComp scores are lower than these published results.

Q1. Which components, inputs and settings of DKAN and TRIPLEX were changed that could explain the performance gap mentioned earlier?

Q3. How were the training hyperparameters chosen?

We appreciate the reviewer’s important comment regarding the performance differences. We first clarify the training and evaluation procedures below.

The official implementations of DKAN and TRIPLEX do not provide pretrained checkpoints, so we trained both models from scratch using their official code and original architectures. For the main benchmark, we followed DKAN’s gene panel, expression targets, and patient-disjoint test folds. Because DKAN’s original protocol did not include a separate validation set, we created one for model selection by holding out the last patient in each fold’s ordered training-patient list, while keeping the test patients unchanged. All methods used the same resulting training, validation, and test patients.

To examine the difference from the published results, we additionally trained and evaluated the official implementations using their original procedure, which selects checkpoints on the test fold rather than on a separate validation set as in our main benchmark. These additional runs produced values close to the published results:

Table. Macro-average PCC under the original test-fold checkpoint-selection procedure

Model and evaluation settingPCC (ALL)PCC (HVG-50)
DKAN, published0.3190.290
DKAN, retrained with original selection procedure0.3210.292
TRIPLEX, reported in DKAN0.2870.252
TRIPLEX, retrained with original selection procedure0.2840.247
ScComp, evaluated with the same selection procedure0.3240.317

The close agreement between the published and reproduced results suggests that differences in checkpoint selection largely explain the reported performance gap. Under the same selection procedure, ScComp achieves comparable PCC (ALL) and higher PCC (HVG-50) than DKAN.

For hyperparameter selection, we used grid search on the held-out validation patients to select the learning rate, ranging from 10−310^{-3} to 10−610^{-6}, and the number of linear layers in the fusion trunk and residual head, considering depths of two, three, or four layers. All other hyperparameters were fixed in advance. The selected configuration and training schedule—60 AdamW epochs followed by 60 patient-balanced iterations—were then used across datasets, folds, and seeds.

We have clarified the validation-patient split and checkpoint-selection procedure in Sec. 4.1 (lines 290–294), with further details in Appendix D (lines 738–743). We also explain that the baseline predictors were trained from scratch using their official implementations (lines 353–355). Hyperparameter selection is summarized in Sec. 4.1 (lines 316–319), with the search ranges and selection procedure provided in Appendix B (lines 644–647).

Q2. Which test supports “the difference on HER2ST alone is not significant”. This sentence implies the cSCC and STNet differences are significant.

We thank the reviewer for pointing out this ambiguity and apologize for the unclear wording. Our original statistical comparison used an exact two-sided Wilcoxon signed-rank test across all 20 seed-averaged folds, yielding a Holm-corrected p=0.014p = 0.014 against DKAN. However, the original sentence did not clearly distinguish this overall comparison from dataset-specific results.

Following the reviewer’s suggestion, we additionally compared ScComp and DKAN within each dataset using exact two-sided Wilcoxon signed-rank tests on the seed-averaged folds. The difference is statistically significant on STNet, where ScComp outperforms DKAN in all eight folds. In contrast, the differences on HER2ST and cSCC are not statistically significant. Although cSCC shows a larger numerical improvement than HER2ST, it contains only four patient-level folds, limiting the statistical power of this comparison.

Table. Comparison with DKAN on HVG-50 PCC

DatasetScCompDKANFoldspp
HER2ST0.2970.29180.740
cSCC0.3070.26840.375
STNet0.2570.20480.008
All datasets (macro-average PCC)0.2870.254200.014

The overall PCC values are dataset-level macro-averages, whereas the overall statistical test uses all 20 seed-averaged folds. Dataset-specific pp-values are unadjusted, while the overall pp-value is Holm-corrected across baseline comparisons.

We have replaced the ambiguous statement in Sec. 4.2 with explicit dataset-specific results (lines 368–371). The test procedure and per-dataset comparisons are provided in Appendix E (lines 844–857; Table 8), distinguishing these unadjusted tests from the Holm-corrected overall comparisons.

W3. The proposed decomposition contributes little in-panel. Against a parameter-matched direct head it adds 0.005 PCC (0.287 vs. 0.282), with no statistical test reported. Overall gains of the proposed approach seem to come from the 5x5 neighbourhood, as it adds 0.066 PCC.

We thank the reviewer for highlighting the limited in-panel performance difference between ScComp and the parameter-matched direct head. We agree that their predictive accuracy is comparable (HVG-50 PCC: 0.287 vs. 0.282; exact two-sided Wilcoxon signed-rank test across 20 seed-averaged folds, p=0.35p = 0.35). Both heads benefit from spatial neighborhood context and patient-balanced updates, consistent with the reviewer’s observation regarding the importance of the 5×55 \times 5 neighborhood.

Importantly, ScComp provides two additional capabilities:

  • Interpretable composition weights. The predicted weights π\pi correlate with Xenium-derived cell-type proportions despite receiving no composition supervision (Spearman ρ=0.66\rho = 0.66; Sec. 4.4).
  • Prediction beyond the training panel. The predicted weights can also be combined with the corresponding reference expression values to predict approximately 11,000 genes outside the 250-gene training panel, without additional training. ScComp achieves a gene-level PCC of 0.187 and a pathway-level PCC of 0.209, outperforming the compared zero-shot baselines in macro-average (Sec. 4.3, Table 2). In contrast, the direct head cannot predict genes outside its output panel.

Thus, the decomposition enables cell-type-associated interpretation and prediction beyond the training panel while maintaining comparable in-panel accuracy.

We have added the statistical comparison with the parameter-matched direct head in Sec. 4.5 (lines 461–468), with the test details provided in Appendix E (lines 859–863). We also clarify the roles of the decomposition in interpretation and prediction beyond the training panel (lines 471–475).

Q4. How stable is π for the 68 and 76-row references? Many rows are redundant. Depending on the seeds, how stable is π?

We thank the reviewer for raising this important question about the stability of the predicted composition weights. Although similar reference profiles may introduce ambiguity in the predicted weights, the reference matrix SS remains fixed across seeds, allowing direct comparison of the same reference-associated weights.

To assess stability, we compared the weights predicted for the same held-out spots by models trained with different random seeds, without any alignment between models. For each reference profile, we computed the Pearson correlation of its predicted weights across spots between pairs of seeds. We then averaged these correlations over reference profiles, seed pairs, and folds.

The resulting correlations were 0.864 on HER2ST, 0.859 on STNet, and 0.822 on cSCC, compared with 0.044, −0.022, and 0.078 for the random-reference control, respectively. ScComp showed higher seed-to-seed agreement than the control in all 20 folds (exact two-sided paired Wilcoxon signed-rank test, p=1.9×10−6p = 1.9 \times 10^{-6}).

These results indicate substantial seed-to-seed reproducibility of the predicted weights across the breast and skin datasets. However, they do not imply that individual weights are uniquely identifiable, particularly for highly similar reference profiles.

W1. The estimated ST profile is a convex combination π of log-normalized profiles. A spot aggregates counts from several cells. Therefore, the physical mixture is linear in the count space not in the log domain. What is the justification of this choice? How does this approach scales to different signal amplitudes?

Q5. How do you account for the amplitude of the signal? The proposed method heavily relies on log normalisation. Are there differences in terms of MAE depending on the depth?

We thank the reviewer for highlighting the distinction between count-space mixing and our log-space formulation.

We agree that transcript counts aggregate additively and that logarithmic transformation does not preserve this additivity. In ScComp, πS\pi S is a convex combination of log-normalized reference profiles. We use this combination as a predictive representation of cell-type-associated expression patterns, rather than as a physical model of transcript-count aggregation.

Our prediction targets follow the benchmark’s library-size normalization and logarithmic transformation. Log transformation compresses the expression range, although it does not fully remove mean-dependent variability, as discussed by Ahlmann-Eltze and Huber (2023). This statistical motivation is distinct from assuming a physically linear mixture in log space. SpatialDecon illustrates this distinction by retaining a linear-scale mixture structure while modeling variability on the log scale (Danaher et al., 2022).

Regarding signal amplitude, our prediction target is library-size-normalized expression rather than raw transcript counts. Library-size normalization removes the explicit per-spot total-count scale, while logarithmic transformation compresses the remaining dynamic range. However, these transformations do not eliminate depth-dependent measurement noise or guarantee depth-independent prediction error.

In our depth-stratified analysis, MAE decreased with sequencing depth for both ScComp and the parameter-matched direct head. This shared trend suggests that depth dependence is not unique to reference-profile mixing, although it does not rule out an additional effect of the log-space formulation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.