acceptodds
Withdrawn from ICLR 2027

Single-Cell Spatial Transcriptomics Prediction from Histology with Multi-Scale Experts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial transcriptomics (ST) offers complementary insights to underlying tissue morphology, yet the acquisition cost limits adoption in large cohorts. Virtual ST has demonstrated its cost-effective potential by predicting ST from H&E morphology, with recent methods extending prediction to single-cell resolution. At the cell level, sparse and heterogeneous gene expression creates a trade-off between preserving localized signals and capturing spatially coherent expression patterns. In this work, we introduce TIGER, a single-cell virtual ST framework that leverages pathology foundation model token embeddings at the spatial scale of individual cells. TIGER integrates local morphologic evidence with tissue-level context through two complementary branches: a focal branch that independently captures fine-grained morphomolecular signals and a context branch that integrates local cellular and global tissue context. A gene- and cell-specific routing mechanism adaptively combines the two. We evaluate TIGER on 43 H&E-paired Xenium tissue sections across diverse cancer types, comprising over 10 million cells. TIGER improves correlation by over 30% and significantly outperforms five state-of-the- art virtual ST baselines across a comprehensive set of conventional and sparsity-aware metrics. Furthermore, TIGER shows stronger generalization performance when evaluated on two external breast and lung cohorts (36 Xenium samples). The application of TIGER-predicted ST on downstream tasks, such as cell phenotyping and treatment response prediction, shows improved performance, underlining the biological and clinical utility of single-cell virtual ST. Code and preprocessed data are available: https://anonymous-hf.com/a/xgn0n1j4ylek/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, computational pathology, virtual ST

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_ZoPTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TIGER for predicting single-cell spatial gene expression from histology. It constructs cell representations from pathology foundation model tokens and combines a cell-wise focal predictor with a spatial-context predictor through cell- and gene-specific gating. Evaluation includes internal cross-validation, external cohorts, and downstream applications.

Critical strengths
  1. The paper addresses an important problem and evaluates the method across multiple cancer types, including independent external cohorts.
  2. Sparsity-aware metrics complement conventional correlation measures and provide a more informative assessment of single-cell predictions.
  3. Patient-level statistical comparisons and computational efficiency measurements strengthen the experimental evaluation.
Critical weaknesses
  1. Limited methodological novelty. TIGER mainly combines established designs: token-based cell representations, local/global context modeling, and adaptive gating. The paper does not sufficiently establish what methodological advance this combination offers beyond existing approaches.

  2. Limited contribution of the additional modules. The focal-only predictor—a simple MLP on extracted cell features—already outperforms all reported baselines in Table 2. The full model provides only small further gains, increasing PCC from 0.286 to 0.295 and AUC-PR from 0.320 to 0.324. These results suggest that much of the advantage comes from the cell representations, while the contribution of the context and gating modules remains limited.

  3. Comparisons with recent single-cell ST prediction methods, such as GHIST+[1] and CELLO[2], should be included.

[1] Zhang, A., Yu, L., Bian, B., et al. Generalisable tissue-wide molecular reconstruction from histology.

[2] Gao, Z., Gu, C., Xiang, J., Luo, X., and Heng, P.-A. Towards Scalable Context-Aware Single-Cell Spatial Transcriptomics Prediction from Histology Images

Questions for authors

see weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZoPT (1/3)

W1. Methodological novelty

TIGER architecture. While TIGER builds on established components in computational pathology, its contribution lies far beyond simply combining them; the cohesive framework was designed to carefully address the gene sparsity, an innate property of single-cell ST, while incorporating tissue spatial context. Our extensive ablation experiments show that each and every design choice matters and that simpler alternatives fail. Such delicate balancing is what we achieve, a design desideratum overlooked in other previous single-cell virtual ST frameworks [1,2,3].

Cell representation. Most prior single-cell methods represent each cell by the CLS token of a pathology foundation model (PFM) of a re-encoded or centered image crop or use a CNN encoder. In contrast, TIGER pools the PFM tokens that overlap each cell, weighted by area, levering local morphology encoded by PFM typically overlooked in previous works. In the new cell-embedding ablation (Table 4b; below), this outperforms CLS-based alternatives (PCC 0.220 vs. 0.117 for nucleus crops and 0.171 for cell-centred crops), as well as simpler token-based cell representations. Among existing methods, only the concurrent CELLO also uses token-based cell embeddings. TIGER outperforms it on 83% of genes (significantly on 30.5%; see Q3), which shows that token-based embeddings alone do not explain TIGER's performance.

Cell featureSupportPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
Tight crop CLSnucleus crop0.117 (±0.016)0.344 (±0.039)0.113 (±0.017)0.462 (±0.010)0.511 (±0.012)0.232 (±0.020)0.226 (±0.020)
Cell-centred crop CLS112 µm window0.171 (±0.015)0.453 (±0.038)0.154 (±0.009)0.447 (±0.011)0.498 (±0.013)0.264 (±0.021)0.253 (±0.022)
Nearest tokencentroid0.214 (±0.025)0.520 (±0.043)0.205 (±0.023)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.027)
Uniform poolingboundary polygon0.213 (±0.025)0.521 (±0.043)0.203 (±0.022)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.026)
Area-weighted pooling (ours)boundary polygon0.220 (±0.026)0.528 (±0.044)0.205 (±0.025)0.435 (±0.013)0.490 (±0.014)0.287 (±0.026)0.276 (±0.027)

 

Adding context naively removes sparse signal and a fully adaptive gate solves this issue. Context aggregation helps spatially coherent genes but smooths out sparse, cell-intrinsic expression. For example, the context-only branch collapses on rare markers of scattered cell populations, such as sparse immune cells, whereas the focal branch retains their signal. TIGER's gate is learned for each cell and gene, so it can rely more on the focal branch for sparse genes and on context for spatially coherent ones. In the fusion ablation (Table 5), it outperforms four other aggregation mechanisms. At the gene level, TIGER improves PCC over the focal branch on 69% of genes and over the context branch on 85%, and the gate value relates to gene sparsity (Fig. A8), indicating that the gate combines the complementary strengths of both branches rather than defaulting to either one.  

Biologically grounded architectural design. TIGER's biologically motivated architecture consistently outperforms technically elaborate methods, including PHEONIX and sCellST, across diverse cohorts and evaluation metrics. These results highlight that carefully designed architectures grounded in biological principles can be more effective than architectural complexity alone. TIGER architecture ranks first on every metric across the three internal cohorts, with significant gains over every baseline (Table 2, Table A3). Without retraining, it improves PCC over the strongest baseline by up to 73% (breast) and 109% (lung) on the external cohorts, and it transfers remarkably better to Atera, the new 10x Genomics whole-transcriptome platform (PCC improvement of 64%; Appendix A.7.4, Table A9).

Task-specific evaluation metrics. Beyond our methodological contributions, we introduce several novel, sparsity-aware metrics that comprehensively assess prediction error, correlation, and expression detection in single-cell virtual ST, addressing the limitations of PCC, which remains the predominant but insufficient evaluation standard in the field.

 

[1] Zhang, Daiwei, et al. "Inferring super-resolution tissue architecture by integrating spatial transcriptomics with histology" Nature biotechnology (2024)

[2] Chadoutaud, Loïc, et al. "sCellST predicts single-cell gene expression from H& E images" Nature Communications (2026)

[3] Fu, Xiaohang, et al. "Spatial gene expression at single-cell resolution from histology using deep learning with GHIST" Nature methods (2025)

 

AuthorsOpenReview ↗

Response to Reviewer ZoPT (2/3)

W2. Contribution of the context and gating modules

Cell representation. We agree that the focal expert alone achieves strong performance. However, we clarify that the focal expert is not simply an MLP applied to conventional cell features; its performance stems from our cell representation, which pools PFM tokens inside the cell boundary and is not a conventional cell feature. The ablation above (W1) evaluates this contribution separately: token-based pooling clearly outperforms CLS-based crops and simpler token variants.

Gene-level analysis. Although the overall gains over the focal expert look modest for metrics averaged across genes and patients (PCC 0.295 vs. 0.286, AUC-PR 0.324 vs. 0.320; Table 5), they become clear gene by gene. The full TIGER outperforms the focal expert on 69.2% of genes (626/904), with more pronounced benefits for immune genes than for non-immune genes (65.7%, 471/717). In lung, the complete architecture improves the PCC of immune genes by 6.6% (0.198 → 0.211), twice the gain on non-immune genes (3.2%). Spatial context thus contributes information for genes whose expression individual-cell morphology cannot explain. These analyses have been included in Section 5.2 of the revised manuscript: Context matters differently for different genes. 

Fully adaptive gate. TIGER integrates context selectively, per-gene and per-cell. Our results show that, once the focal expert has a strong cell representation, context improves performance if it is incorporated adaptively. Simpler integration strategies (averaging, late fusion, gene-level gating and the newly added cell-level gating) do not improve over the focal expert (PCC 0.280–0.287 vs. 0.286). Only the fully adaptive gate does (0.295), so the benefit of context depends jointly on the gene and on the cell's local environment.

FocalAverageLate fusionGene adaptiveCell adaptiveFully adaptive (TIGER)
PCC ↑0.286 (±0.027)0.280 (±0.026)0.280 (±0.025)0.287 (±0.026)0.283 (±0.025)0.295 (±0.024)
AUC-PR ↑0.320 (±0.026)0.319 (±0.025)0.320 (±0.024)0.319 (±0.025)0.320 (±0.024)0.324 (±0.025)

Table 5, macro-average over the three internal cohorts (all gating variants trained on the same splits).

 

External datasets and cell phenotyping. We also compared the full TIGER with its individual experts beyond internal cross-validation. On the external breast and lung cohorts, TIGER significantly outperforms the focal expert on PCC ( q < 0.001), Recall and AUC-PR (q < 0.01), with no difference on PCC+, and significantly outperforms the context expert on all metrics except PCC+. In cell phenotyping on the colorectal cohort (16 slides), TIGER reaches the highest macro precision (0.452), significantly above the context (0.421) and focal (0.432) experts (p < 0.01). Its accuracy matches the focal expert (0.581 vs. 0.579) and exceeds the context expert (0.558, p < 0.05).

 

GHIST (strongest baseline)Context expertFocal expertTIGER
Macro precision ↑0.351***0.421**0.432**0.452
Accuracy ↑0.510**0.558*0.5790.581

Colorectal cohort, 16 slides. Stars: TIGER significantly better.

Together, these results show two complementary contributions: token-based cell representations that give strong prediction even with a simple MLP, and adaptive context that adds gene-dependent gains. The strong focal expert thus demonstrates the cell representation rather than diminishing TIGER's contribution.

 

AuthorsOpenReview ↗

Response to Reviewer ZoPT (3/3)

W3. Comparison with GHIST+ and CELLO

New baselines. We thank the reviewer for pointing us to these recent methods. GHIST+ was preprinted in June and CELLO was released after our submission. We have now added both as baselines on the three internal cohorts. Results are reported here and will be included in the final manuscript version. For a controlled comparison, both were re-implemented in our framework from their released code. They use the same inputs as TIGER: UNIv2 token embeddings and CellViT nuclei; and were evaluated under the same protocol as all other baselines (splits, validation-fold checkpoint selection, early stopping, metrics). For GHIST+ we excluded the cross-panel module (not applicable to our setting) and the TMA-based molecular anchoring (it requires measured expression of the test tissue, which would be data leakage). 

ModelPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
GHIST+0.224 (0.018)0.462 (0.024)0.185 (0.015)0.429 (0.008)0.568 (0.009)0.309 (0.008)0.272 (0.010)
CELLO0.266 (0.017)0.496 (0.027)0.237 (0.014)0.430 (0.009)0.498 (0.009)0.321 (0.009)0.309 (0.010)
TIGER0.283 (0.027)0.526 (0.036)0.237 (0.020)0.430 (0.013)0.497 (0.014)0.329 (0.022)0.319 (0.025)

Macro-average over lung, breast and colorectal. Best in bold.

 

TIGER is best on PCC, PCC50, bRMSE, Recall and AUC-PR, and ties CELLO on PCC+. GHIST+ attains a marginally lower bMAE, but with the highest bRMSE and the lowest correlation scores. CELLO is the strongest baseline. Like TIGER it uses cell-located token embeddings, which supports token-level PFM representations at single-cell resolution. TIGER's gate, which weighs local morphology and context for each cell and gene, gives higher patient-level PCC than CELLO in every cohort (lung 0.221 vs. 0.206, breast 0.384 vs. 0.374, colorectal 0.245 vs. 0.219).

 

Gene-level comparison with CELLO. The gap is clearer gene by gene: TIGER has higher PCC on 83.0% of genes (significantly on 30.5%), and CELLO is not significantly better on any gene.

CohortGenesTIGER betterSignificantly better (q < 0.05)Significantly worse (q < 0.05)
Lung (16 patients)302251 (83.1%)101 (33.4%)0
Breast (11 patients)280217 (77.5%)13 (4.6%)0
Colorectal (16 patients)322282 (87.6%)162 (50.3%)0
Total904750 (83.0%)276 (30.5%)0

"Better" = higher mean per-patient PCC. Significance: paired one-sided t-test across patients, BH-corrected across genes.

Where the gap comes from. A closer look at the colorectal cohort shows that most of the gap arises from genes for which CELLO predicts zero expression in all cells: ~42 genes per slide (81 on at least one slide), mainly rare markers of scattered cell populations such as sparse immune cells (CD40LG, CTSG, S100A12). TIGER is significantly better on 84% of these genes (68/81) vs. 39% of the remaining genes. Our context-only branch shows the same collapse, whereas the cell-anchored focal branch retains the signal. The fully adaptive gate lets TIGER rely on the focal branch for these genes, instead of averaging their signal away with the neighbourhood context.

 

We thank the reviewer again for the constructive feedback. We have added both requested baselines and the analyses that separate the contributions of the cell representation, the context branch and the gate. We hope that these additions address the concerns, and we would be grateful if the reviewer would consider revisiting the score.

Dear Reviewer,

Thank you once again for your thoughtful review and valuable feedback. We have provided a detailed rebuttal addressing the concerns raised in your review.

With the discussion period drawing to a close, we would greatly appreciate it if you could review our responses at your convenience. We would be happy to address any remaining questions or provide further clarification.

Thank you once again for your thoughtful consideration.

Reviewer_Tc63OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TIGER for predicting single-cell spatial transcriptomics from H&E histology. It constructs cell representations by pooling pathology foundation model tokens according to their overlap with segmented cell boundaries. A focal MLP and a context branch incorporating neighboring cells and slide-level information are combined through cell- and gene-specific gating. Evaluation includes patient-level cross-validation on 43 internal patients across three cancer types, transfer to 36 external tissue sections, cell phenotyping, and retrospective breast cancer molecular-status and treatment-response prediction. The reported results show consistent improvements over existing methods, while the ablations indicate that the token representation and focal branch already account for much of the overall gain.

Critical strengths
  1. TIGER achieves an internal macro-average PCC of 0.295 versus 0.225 for GHIST, with improvements also covering PCC+, AUC-PR, recall, and balanced errors. Cohort-specific results support the overall trend. The internal significance analysis uses patient-level differences rather than treating millions of cells as independent observations. Without external fine-tuning, PCC increases from the strongest baseline's 0.091 to 0.193. This is meaningful evidence of transfer, although the uncertainty analysis requires revision (Tables 2–3 and A3–A4, pp. 7 and 23).
  2. Reusing a fixed token grid avoids separately encoding every cell. The method can use H&E-derived CellViT boundaries, with performance close to that obtained using Xenium boundaries. Comparisons of focal/context experts, local/global context, and alternative fusion strategies help explain the trade-off between spatial smoothing and localized expression recovery. These analyses strengthen the contribution beyond a model combination alone (Table A1, p. 16; Tables 5 and A9–A11, pp. 8 and 33–34).
Critical weaknesses
  1. The external lung cohort contains 17 sections from eight patients, and the breast cohort contains 19 sections from 17 patients, yet Tables A6/A7 report paired t-tests on section-level differences without describing patient-cluster adjustment (Appendix A.4.2, p. 19; pp. 26–27). The clinical analysis uses 50 repeated 80%/20% Monte-Carlo splits of the same 166 or 62 patients, followed by a paired t-test across those splits and standard errors across test folds (Appendix A.10, p. 36; Table A12, p. 38). Overlapping splits do not provide 50 independent replications. This affects the evidential strength of the treatment-response AUROC increase from 0.620 to 0.658 with reported p=0.006p = 0.006. Patient-level clustering and an uncertainty procedure appropriate for repeated splitting are needed. This is a concern about inference, not an allegation of patient leakage or a claim that the numerical improvement disappears.

  2. In Table 5, the focal branch achieves PCC, PCC+, and AUC-PR of 0.286, 0.239, and 0.320, respectively, versus 0.295, 0.242, and 0.324 for the full model; the context branch has slightly higher PCC at 0.296. The authors acknowledge the contribution of the PFM-derived representation. Table A11 supports adaptive gating over alternative fusion strategies, but does not provide patient-level statistical comparisons against the standalone experts. Those experts are also absent from the external and clinical comparisons. Thus, the evidence supports the complete system more strongly than it establishes routing as the source of its principal advantages. A comparison against both experts on the existing external cohorts would clarify whether routing contributes beyond the strong underlying representation (Tables 5, A9, and A11; Tables 3 and A12).

Questions for authors
  1. After accounting for repeated sections within patients and dependence between Monte-Carlo splits, what are the effect sizes and confidence intervals relative to the strongest external baseline and the image-only clinical model? In particular, how much statistical support remains for the 0.038 treatment-response AUROC improvement?

  2. What are the patient-level differences between the full model and the independently trained focal and context experts? Can their external-cohort results be provided to establish whether the transfer advantage depends on adaptive routing? This concerns the contribution of existing components, rather than requesting a new dataset.

  3. Do Tables A3 and A4 use the same predictions, gene sets, and runs? If so, please explain their aggregation. For sCellST, weighting the A4 cohort bMAEs of 0.476, 0.415, and 0.500 by the corresponding cohort sizes of 16, 16, and 11 patients gives approximately 0.459, whereas A3 reports 0.518; corresponding bRMSE estimates are approximately 0.544 versus 0.619. For DeepSpot2Cell, PCC50\mathrm{PCC}_{50} similarly gives approximately 0.433 versus 0.482. These are estimates from reported cohort summaries, not exact recalculations from patient-level outputs, and fold averaging may differ. Nevertheless, the captions do not sufficiently explain the discrepancies. Please provide the aggregation details, corrected entries, or patient-level results (p. 23).

5 comments
AuthorsOpenReview ↗

Response to Reviewer Tc63 (1/4)

We thank the reviewer for their careful statistical assessment and for recognizing TIGER's consistent performance improvements, external generalization, and computational efficiency. We also appreciate their recognition of our methodological contributions and extensive ablation studies. We address the reviewer's comments in detail below.

 

W1 + Q1. Patient-level clustering and inference for repeated splits

External cohorts. We repeated Tables A6/A7 at the patient level. Metrics are first averaged over each patient's sections. We then compared TIGER against each baseline by one-sided paired t-tests on per-patient differences, BH-corrected across the 42 baseline × metric comparisons per cohort. We now also report 95% CIs (paired t) against the strongest baseline: DeepSpot2Cell (breast) and Phoenix (lung). A patient-level bootstrap (10,000 resamples) gives the same conclusions.

Table. Mean per-patient difference with 95% CI and number of patients where TIGER wins. Positive values favour TIGER (bMAE and bRMSE are sign-flipped).

MetricTIGER vs. DeepSpot2Cell (Breast: n = 17)TIGER vs. Phoenix (Lung: n = 8)
PCC+0.108 [0.079, 0.138] (16/17)+0.073 [0.046, 0.101] (8/8)
PCC++0.088 [0.062, 0.114] (15/17)+0.052 [0.030, 0.073] (8/8)
bMAE+0.037 [0.026, 0.048] (16/17)−0.001 [−0.007, 0.004] (3/8)
bRMSE+0.023 [0.008, 0.038] (13/17)+0.007 [0.002, 0.013] (7/8)
Recall+0.056 [0.042, 0.070] (17/17)+0.042 [0.022, 0.062] (8/8)
AUC-PR+0.056 [0.041, 0.071] (17/17)+0.039 [0.021, 0.057] (8/8)

We verify that every comparison significant at section level stays significant. Against the strongest baseline, PCC improves 1.7× (breast) and 2.1× (lung), and the intervals of all correlation and detection metrics lie above zero. The only exceptions are error metrics in lung: a tie with Phoenix on bMAE and with GHIST on bRMSE, and a slightly lower bRMSE for iStar. Tables A6 and A7 now report patient-level significance, and the effect sizes are in Table A8.

 

Clinical task. We thank the reviewer for this important point: overlapping Monte-Carlo splits are not independent replications, so the original paired t-tests may overstate significance. We re-analysed the clinical comparisons with a patient-level paired bootstrap (2,000 resamples, same patients for both models, one slide per patient, see Appendix A.10: Statistical analysis):

Task (n)MetricTIGERImage-onlyΔ (95% CI)one-sided p
Treatment response (62)AUROC0.6580.620+0.038 (−0.009, +0.080)0.058
Treatment response (62)Bal. acc.0.5900.570+0.020 (−0.024, +0.063)0.212
HER2 status (166)AUROC0.7120.655+0.057 (−0.026, +0.140)0.082
HER2 status (166)Bal. acc.0.6530.626+0.027 (−0.041, +0.101)0.214
ER status (166)AUROC0.7470.712+0.036 (−0.039, +0.107)0.172
ER status (166)Bal. acc.0.7030.633+0.069 (+0.004, +0.134)0.017

We now present the clinical results as consistent but not statistically established improvements, and revised Table 6 and the text accordingly. We note, however, that every point estimate favors TIGER over the image-only model, and that the width of the intervals can be partly explained by the small size of these cohorts (62–166 patients). 

AuthorsOpenReview ↗

Response to Reviewer Tc63 (2/4)

Continuation of W1+Q1

Independent validation. Since the internal cohorts are relatively small, we additionally evaluated HER2 status on a larger external cohort, the Yale HER2 cohort [2] (n = 189 patients, 92 HER2-positive; one slide per patient), applying models trained only on Sammut at al. [1] without retraining. Besides the image-only model, and due to the limited rebuttal time, we included the best- and worst-performing baselines in single-cell ST prediction (GHIST and Img2ST-Net). The remaining baselines will be included in the final manuscript version.

 

HER2 status (n = 189)AUROCBAcc.ΔAUROC (95% CI)pΔB Acc. (95% CI)p
Image-only0.5930.524––––
Img2ST-Net0.4350.480−0.158 (−0.258, −0.053)– (worse, p = 0.003)−0.044 (−0.068, −0.019)– (worse, p < 0.001)
GHIST0.6560.550+0.063 (−0.031, 0.155)0.097+0.026 (−0.022, 0.071)0.147
TIGER0.7170.615+0.124 (0.042, 0.204)0.0015+0.091 (0.043, 0.140)< 0.0005

 

On this larger cohort, TIGER is significantly better than the image-only model, with a larger effect than in the internal splits. In contrast, GHIST does not differ significantly from the image-only model and Img2ST-Net is significantly worse than it, indicating that the benefit depends on the quality of virtual ST rather than on the downstream model alone. We note that this check concerns HER2 status; a larger-cohort validation of the treatment-response endpoint itself remains future work, and we state this limitation in the revised manuscript.

 

[1] Stephen-John Sammut et al. Nature (2022)

[2] Farahmand, Saman, et al. (2022). Dataset: The Cancer Imaging Archive. https://doi.org/10.7937/E65C-AM96

AuthorsOpenReview ↗

Response to Reviewer Tc63 (3/4)

W2 + Q2. Contribution of TIGER components

Patient-level differences between the full model and focal and context experts in the internal cohorts. We performed patient-level paired tests over the 43 internal patients (one-sided paired t-test, BH-corrected). This analysis showed that TIGER significantly outperforms the context expert on all seven metrics. PCC+, Recall and AUC-PR improve in all 43 patients, and PCC improves by +0.012 [0.005, 0.018] (32/43 patients). Against the focal expert, TIGER significantly improves PCC (+0.010 [0.007, 0.012]; 37/43 patients), PCC50 (32/43), Recall (32/43) and AUC-PR (35/43), while the differences in PCC+ and the error metrics (bMAE, bRMSE) are not significant.

Analysis in the external cohorts. To test whether the transfer advantage depends on adaptive routing, we evaluated on both external cohorts the independently trained focal and context experts and the four alternative fusions of Table 5 (average, late fusion, gene adaptive, cell adaptive). The 36 external slides come from 25 patients (17 breast, 8 lung). We averaged the slides of each patient and report patient-level means, averaged over the two cohorts with equal weight.

 

ModelPCC ↑PCC+ ↑bMAE ↓Recall ↑AUC-PR ↑
Context0.170***0.1800.444**0.238***0.227***
Focal0.205***0.1820.4340.260**0.248***
TIGER (average)0.203***0.1820.4380.258***0.247***
TIGER (late fusion)0.198***0.178***0.4390.257***0.245***
TIGER (gene adaptive)0.203***0.1820.4390.258***0.247***
TIGER (cell adaptive)0.205***0.1820.4390.259***0.247***
TIGER (fully adaptive)0.2100.1840.4400.2620.250

External breast and lung, patient-level means. Best in bold (Table A13, new)

Following Tables A6/A7, we tested whether the fully adaptive TIGER is better than each of the six alternatives. We used one-sided paired t-tests on the 25 per-patient differences, with Benjamini–Hochberg correction over all 30 comparisons (6 models × 5 metrics).

TIGER vs. Context: TIGER is significantly better on PCC, Recall and AUC-PR (q < 0.001) and on bMAE (q < 0.01).

TIGER vs. Focal: TIGER is significantly better on PCC and AUC-PR (q < 0.001) and on Recall (q < 0.01). 

TIGER vs. other fusion strategies (average, late fusion, gene adaptive, cell adaptive): TIGER is significantly better than all four on PCC, Recall and AUC-PR (all q < 0.001). On PCC+ it is significantly better than late fusion (q < 0.001).

Thus, the adaptive gate yields significant gains in correlation and detection on unseen cohorts, suggesting that adaptive routing of the two experts contributes to the transfer.

 

Gene-analysis and focal/context in cell phenotyping. The gains of the full TIGER model, that adaptively integrates context in a per-gene and cell basis, become clear when inspecting genes individually. The full TIGER architecture outperforms the focal expert on 69.2% of genes (626/904), with more pronounced benefits for immune genes than for non-immune genes. In lung, for example, the complete architecture improves the PCC of immune genes by 6.6% (0.198 → 0.211), twice the gain on non-immune genes (3.2%). Spatial context thus contributes information for genes whose expression individual-cell morphology cannot explain. These analyses have been included in Section 5.2 of the revised manuscript: Context matters differently for different genes.

Finally, and as suggested by the reviewer, we compared the focal and context experts with the full TIGER model on cell phenotyping, to assess how their predictions translate to a downstream biological task. TIGER achieves the highest macro precision (0.452), significantly above both the context (0.421) and focal (0.432) experts (p < 0.01). Its accuracy matches the focal expert (0.580 vs. 0.579) and significantly exceeds the context expert (0.558, p < 0.05):

 GHIST (strongest baseline)Context expertFocal expertTIGER
Macro precision ↑0.351***0.421**0.432**0.452
Accuracy ↑0.510**0.558*0.5790.580

 

Overall, the focal expert performs strongly owing to our proposed PFM token-based cell representation, and adaptively integrating context in TIGER brings further gains, particularly for sparse and scattered genes such as immune markers. These gains carry over to the external cohorts and to cell phenotyping.

 

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

This paper introduces TIGER, a framework for predicting single-cell spatial transcriptomics from H&E-stained histological images. The framework was evaluated on 43 Xenium tissue sections across three cancer cohorts, with additional external validation performed on 36 samples. The experimental design includes comparisons with existing "virtual" spatial transcriptomics methods, ablation studies, the use of sparsity-aware evaluation metrics, and downstream applications such as predicting cell phenotypes and therapeutic responses.

Critical strengths
  1. The distinction between spatially localized gene expression and spatially coherent gene expression provides a sound rationale for integrating cell-specific morphological features with broader spatial context.

  2. By aggregating token embeddings from a pre-trained pathology foundation model based on cell boundaries, TIGER avoids redundant encoding of cell-centered image patches while preserving fine-grained morphological information.

  3. The experimental evaluation is comprehensive. The authors compare TIGER against five existing virtual spatial transcriptomics methods across three internal cancer cohorts and further investigate the model's generalizability using two independent external cohorts. The use of patient-level data splitting and multiple sparsity-aware metrics strengthens the evaluation.

Critical weaknesses
  1. The contribution of the proposed "dual-expert" architecture appears relatively limited compared to the "focal" branch alone. Table 5 shows that the TIGERfocal component—on its own—outperforms all five baselines presented in Table 2. The fully adaptive model improves the Pearson Correlation Coefficient (PCC) from 0.286 to 0.295, PCC+ from 0.239 to 0.242, and the Area Under the Precision-Recall Curve (AUC-PR) from 0.320 to 0.324. While these improvements are positive, their magnitude is relatively small. This suggests that the majority of TIGER’s performance gains may stem from its cell-level morphological characterization, rather than the proposed multi-scale expert integration. Further controlled experiments would help clarify the individual contributions of feature extraction, spatial context modeling, and adaptive gating.
  2. The innovation is somewhat limited, as the approach primarily consists of a combination of existing modules. The focal branch employs an MLP for independent cell-level regression, while the context branch integrates both local and global Transformer-based representations. Final predictions are integrated via an adaptive gating network. While this combination effectively facilitates the handling of sparse gene expression, the individual components are largely built upon established architectural mechanisms.
  3. As noted in Table 1 and Appendix A.5, several baselines rely on spot-level supervision, whereas TIGER directly utilizes cell-level expression labels. Furthermore, not all baselines employ the same foundational pathology models or feature extraction strategies. These discrepancies complicate the attribution of performance gains specifically to the proposed architecture. Although the authors acknowledge these differences and have retrained the baselines on the same datasets, further comparisons conducted under more strictly controlled supervision and feature settings would strengthen their conclusions. 4.Although the external evaluation was conducted on independent breast and lung cancer cohorts, these cancer types were already represented in the internal training cohorts. Consequently, these results primarily reflect the model's generalizability for the same cancer types across different cohorts, rather than its ability to generalize to previously unseen cancer types. Furthermore, as noted in Appendix A.2, the external evaluation for lung and breast cancer relied on only 93 and 61 shared genes, respectively—a significant reduction compared to the 302 and 280 genes included in the corresponding internal training cohorts. This substantial decrease in gene coverage may complicate the interpretation of the generalization results, particularly for genes that differ in terms of expression frequency or biological function. 5.While the downstream clinical evaluation shows some potential, its scope remains limited. In the treatment response experiment, only 62 HER2-positive patients were included, and the performance improvement over models based solely on imaging was modest (with the AUROC increasing from 0.620 to 0.658). Although statistical analysis indicates that this improvement is significant, the experiment relied on multiple random train/test splits within the same patient cohort. Further validation on an independent clinical cohort would provide more robust evidence for the reported performance gains and the model's generalizability. Therefore, caution should be exercised when interpreting the clinical implications of these findings.
Questions for authors

The author needs to address weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer UCGH (1/5)

We thank the reviewer for the thoughtful review and for valuing our biologically-grounded architecture, the efficient use of foundation-model token features, and the comprehensive evaluation across internal and external cohorts with novel sparsity-aware metrics. We answer each point below.   

W1. Contribution of TIGER components

Two contributions. We agree that the focal expert is strong and that the additional gains from multi-scale expert integration are relatively modest when averaged across all genes. However, we would like to emphasize that TIGER makes two complementary methodological contributions: constructing fine-grained cell-level morphological representations from PFM tokens and adaptively integrating spatial context according to individual cell–gene characteristics. To better distinguish these contributions, we conducted additional ablations:

1. Cell embedding (new, Table 4b). We trained the focal expert alone on the lung cohort with alternative cell embeddings, keeping the PFM (UNIv2) and everything else fixed. CLS embeddings of cropped cells are much worse (PCC 0.117–0.171 vs. 0.220) and need one PFM pass per cell, which become computationally expensive in slides with millions of cells. The gain comes mainly from reading cell-localized PFM tokens, with a further benefit from area-weighted pooling, which is best on nearly all metrics. The cell representation is therefore an important contribution of TIGER independent of context and gating.

Cell featureSupportPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
Tight crop CLSnucleus crop0.117 (±0.016)0.344 (±0.039)0.113 (±0.017)0.462 (±0.010)0.511 (±0.012)0.232 (±0.020)0.226 (±0.020)
Cell-centred crop CLS112 µm window0.171 (±0.015)0.453 (±0.038)0.154 (±0.009)0.447 (±0.011)0.498 (±0.013)0.264 (±0.021)0.253 (±0.022)
Nearest tokencentroid0.214 (±0.025)0.520 (±0.043)0.205 (±0.023)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.027)
Uniform poolingboundary polygon0.213 (±0.025)0.521 (±0.043)0.203 (±0.022)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.026)
Area-weighted pooling (ours)boundary polygon0.220 (±0.026)0.528 (±0.044)0.205 (±0.025)0.435 (±0.013)0.490 (±0.014)0.287 (±0.026)0.276 (±0.027)

2. Gene-level analysis. Gene-level analyses show biologically relevant benefits. The full TIGER architecture outperforms the focal expert on 69.2% of genes (626/904), more for immune genes (82.9%, 155/187) than for non-immune genes (65.7%, 471/717). In lung, the complete architecture improves the PCC of immune genes by 6.6% (0.198 → 0.211), twice the gain on non-immune genes (3.2%), with significant improvements for 37% of those genes. These genes have sparse and scattered spatial patterns, and their identity is often informed by the surrounding environment. Spatial context thus contributes information that individual-cell morphology cannot explain. Expanded analyses are in Section 5.2 (Context matters differently for different genes) of the revised manuscript.

3. Fully adaptive gate. Context only helps if it is incorporated selectively. Averaging, late fusion, gene-level gating and the new cell-level gating (one gate value per cell) do not improve meaningfully over the focal expert (PCC 0.280–0.287 vs. 0.286). The fully adaptive gate is the only strategy that does (0.295), so the benefit of context depends jointly on the gene and on the cell's local environment. As we include in the newly added Fig. A8, the learned gate is also interpretable: within spatially coherent expression domains, the gate shifts weight toward the context expert (gate= 0.60 inside vs. 0.70 outside the domain in lung; 0.63 vs. 0.76 in colorectal; consistent in 16/16 patients per cohort, Wilcoxon p < 10−4). The full model also improves over the focal expert on the external cohorts (newly added Table A13) and in cell phenotyping (newly added Table A17).

Table. Ablation on gating mechanisms (cell adaptive gating newly added).

FocalAverageLate fusionGene adaptiveCell adaptiveFully adaptive (TIGER)
PCC ↑0.286 (±0.027)0.280 (±0.026)0.280 (±0.025)0.287 (±0.026)0.283 (±0.025)0.295 (±0.024)
AUC-PR ↑0.320 (±0.026)0.319 (±0.025)0.320 (±0.024)0.319 (±0.025)0.320 (±0.024)0.324 (±0.025)

 

Summary. Taken together, the contribution of TIGER lies in both its effective cell-level morphological representation and its ability to selectively leverage spatial context, in particular for sparse and scattered genes.

AuthorsOpenReview ↗

Response to Reviewer UCGH (2/5)

W2. TIGER innovation

Design follows from a property of the data. While TIGER builds on established components in computational pathology, its contribution lies far beyond simply combining them; the cohesive framework was designed to carefully address the gene sparsity, an innate property of single-cell ST, while incorporating tissue spatial context. Our extensive ablation experiments show that each and every design choice matters and that simpler alternatives fail. Such delicate balancing is what we achieve, a design desideratum overlooked in other previous single-cell virtual ST frameworks [1,2,3]. TIGER's biologically motivated architecture consistently outperforms other technically elaborate methods, highlighting that carefully designed architectures grounded in biological principles can be more effective than architectural complexity alone. 

Cell representation. Most prior single-cell methods use the CLS token of a re-encoded or centred crop, or a CNN encoder. In contrast, TIGER pools the PFM tokens that overlap each cell, weighted by area, levering local morphology encoded by PFM typically overlooked in previous works. In the new cell-embedding ablation (as discussed in W1), this substantially outperforms CLS-based alternatives as well as other simpler token-based cell representations. 

Adding context naively removes sparse signal and a fully adaptive gate solves this issue. Context aggregation helps spatially coherent genes but smooths out sparse, cell-intrinsic expression. For example, the context-only branch collapses on rare markers of scattered cell populations, such as sparse immune cells, whereas the focal branch retains their signal. TIGER's gate is learned for each cell and gene, so it can rely more on the focal branch for sparse genes and on context for spatially coherent ones (Fig. A8 in the updated manuscript) . In the fusion ablation (Table 5), it outperforms four other aggregation mechanisms, including a new cell-adaptive gating baseline (fix gate for all genes in a cell), which fail in incorporating context from the context branch and fall below the focal component.

Benchmark. Beyond our methodological contributions, we aimed to make evaluation of virtual single-cell ST more robust by introducing sparsity-aware metrics that assess prediction error, correlation and expression detection, addressing the limitations of PCC, which remains the predominant but insufficient standard in the field. TIGER ranks first on every metric across the three internal cohorts, with significant gains over every baseline (Table 2, A3). Furthermore, without retraining, it substantially outperforms other architectures on two external cohorts.

 

[1] Zhang, Daiwei, et al. "Inferring super-resolution tissue architecture by integrating spatial transcriptomics with histology" Nature biotechnology (2024)

[2] Chadoutaud, Loïc, et al. "sCellST predicts single-cell gene expression from H& E images" Nature Communications (2026)

[3] Fu, Xiaohang, et al. "Spatial gene expression at single-cell resolution from histology using deep learning with GHIST" Nature methods (2025)

AuthorsOpenReview ↗

Response to Reviewer UCGH (3/5)

W3. Supervision and vision encoders across baselines

We thank the reviewer for highlighting the importance of controlled benchmarking. 

Controlled as far as each design allows. Virtual ST has evolved with the technology. Before high-resolution platforms such as Xenium and Visium HD became widely available, methods inferred single-cell expression from spot-level measurements by weak supervision (e.g., iStar, DeepSpot2Cell). With cell-resolved data (e.g., with initiatives like HEST-1k), direct cell-level supervision has become possible (e.g., GHIST). In this study, we aimed for the broadest benchmark of published high-resolution methods, with comparisons as controlled as each method's design allows. All baselines were retrained on the same datasets, patient-level splits, and checkpoint-selection protocol as TIGER. We adapted baselines to UNIv2 encoder as TIGER whenever possible. Every ViT-based method was adapted while CNN-based architectures had to keep their original backbones, since their decoders depend on intermediate convolutional feature maps and adaptation to a PFM is not possible. 

New fully controlled baselines: GHIST+ and CELLO. As part of this rebuttal, we have added two additional methods that have been recently released: GHIST+ (bioRxiv 2026), and CELLO (NeurIPS 2026). These two methods also match TIGER's setting exactly: both are fully supervised with cell-level expression and use the same frozen UNIv2 token embeddings as TIGER. This setting isolates the contribution of the architecture, and TIGER outperforms both of them (e.g., macro-averaged PCC 0.283 vs. 0.266 for CELLO and 0.224 for GHIST+ using common train/val/test splits). TIGER also has higher patient-level PCC than CELLO for each cohort (lung 0.221 vs. 0.206, breast 0.384 vs. 0.374, colorectal 0.245 vs. 0.219). Results are reported below and will be incorporated into the final manuscript:

Table. Newly added baselines that match TIGER full supervision with cell-level expression.

ModelPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
GHIST+0.224 (0.018)0.462 (0.024)0.185 (0.015)0.429 (0.008)0.568 (0.009)0.309 (0.008)0.272 (0.010)
CELLO0.266 (0.017)0.496 (0.027)0.237 (0.014)0.430 (0.009)0.498 (0.009)0.321 (0.009)0.309 (0.010)
TIGER0.283 (0.027)0.526 (0.036)0.237 (0.020)0.430 (0.013)0.497 (0.014)0.329 (0.022)0.319 (0.025)
Reviewer_LndXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework (TIGER) for predicting single-cell spatial transcriptomics from H&E images. The method extracts cell-level representations using a pathology foundation model and combines two complementary modules: a Focal Expert that captures individual cell morphology and a Context Expert that incorporates local cellular neighborhoods and global tissue information. An adaptive gating mechanism then combines their predictions for each gene and cell.

The authors evaluate TIGER on multiple Xenium datasets across colorectal, lung, and breast cancers. The proposed framework outperforms existing virtual ST prediction methods and demonstrates promising results on downstream tasks, including cell phenotyping and treatment response prediction.

Critical strengths
  1. Well-motivated problem and method design. The paper addresses an important challenge in single-cell virtual ST prediction. The authors clearly explain the trade-off between preserving sparse gene expression and incorporating spatial context. The motivation for combining Focal and Context Experts is reasonable and biologically meaningful.

  2. Clear and easy-to-follow methodology. The proposed framework is well presented. The roles of the Focal Expert, Context Expert, and Adaptive Gating are clearly explained. The qualitative examples also help readers understand the motivation behind each component.

  3. Comprehensive experiments and promising results. The paper evaluates TIGER on multiple cancer datasets with both internal and external validation. The comparisons with existing methods are comprehensive, and the introduction of sparsity-aware evaluation metrics is useful for single-cell ST prediction. The downstream experiments further demonstrate the potential practical value of the framework.

Critical weaknesses
  1. Focal expert already performs well
  • The Focal Expert alone already achieves strong performance, even without incorporating spatial context or adaptive gating. For example, Table 5 shows that the Focal Expert achieves a PCC of 0.286 and PCC+ of 0.239, compared with 0.295 and 0.242 for the full TIGER model. Furthermore, the Context Expert alone achieves a slightly higher PCC (0.296) than the full model (0.295).

  • These results suggest that the Focal Expert may already be sufficient for achieving competitive performance. Given the additional complexity introduced by the Context Expert and Adaptive Gating, it remains unclear whether the relatively small improvements justify the full framework. A more detailed analysis of which types of genes benefit from the complete model would help clarify its advantages.

  1. Insufficient analysis of the proposed cell-level representation
  • The Focal Expert achieves strong performance even without incorporating any spatial context, suggesting that the cell-level representations extracted from the pathology foundation model may be an important contributor to the overall performance. However, the paper does not sufficiently investigate the effectiveness of the proposed cell-level feature construction compared with simpler alternatives.

  • For example, how would the performance change if the authors used the feature at each cell's center or a simpler aggregation of patch tokens instead of the proposed cell-level representation? A controlled comparison using the same foundation model and prediction head would help clarify the contribution of this design.

  1. Limited evaluation on spot-level ST datasets and downstream tasks
  • Although TIGER achieves strong performance at the single-cell level, it remains unclear whether the model generalizes well to spot-level ST datasets, such as Visium. The authors evaluate spot-level performance by aggregating cell-level predictions on Xenium data (Appendix A.7.4), but this is not equivalent to evaluating the model on actual Visium datasets.

  • For example, the authors could apply TIGER to Visium H&E images, aggregate the predicted cell-level gene expression within each spot, and compare the results with measured spot-level expression and existing spot-level prediction methods. This would help determine whether strong cell-level performance also translates into good spot-level performance across platforms.

  • Furthermore, it would be valuable to investigate whether TIGER-generated virtual single-cell ST improves downstream tasks on Visium datasets. For example, how much does the downstream performance improve when incorporating predicted single-cell ST compared with using H&E features alone or predicted spot-level ST?

  1. Lack of baselines on downstream tasks.
  • For downstream experiments, comparisons with existing virtual ST prediction methods would also be important to demonstrate the additional benefits of TIGER.
Questions for authors

Is there a statistical relationship between gene sparsity and the learned adaptive gating weights?

  • One of the main motivations of TIGER is to balance the preservation of sparse gene expression and the modeling of spatially coherent expression patterns. However, it remains unclear whether the learned gating weights are systematically related to these gene characteristics.

  • For example, suppose Gene A is expressed in only 5% of cells, while Gene B is expressed across a large, spatially coherent tissue region. Does the model assign higher Focal Expert weights to the expression-positive cells of Gene A, while relying more on the Context Expert within the coherent expression region of Gene B?

4 comments
AuthorsOpenReview ↗

Response to Reviewer LndX (1/3)

We thank the reviewer for the constructive feedback and for their positive assessment of our motivation, methodological clarity, and comprehensive evaluation. We also appreciate their recognition of the sparsity-aware metrics and promising downstream results. The reviewer's questions led us to analyse the cell representation and the gate more closely. We answer the reviewer's comments below.

 

W1 + Q1. Focal expert; which genes benefit; gate vs. sparsity

We agree that the focal expert alone achieves strong performance, highlighting the effectiveness of our PFM-derived cell representations. In Weakness 2 below, we provide an ablation on this representation and show how the area-weighted token average is an important contribution of our overall method. However, the benefit of the full TIGER framework extends beyond improvements in average performance, particularly in its ability to adapt to gene-specific spatial expression patterns.

Which genes benefit? The benefit of the full model lies in gene-specific adaptation. The full TIGER outperforms the Focal Expert on 69.2% of genes (626/904), with more pronounced benefits for immune-related genes (82.9%, 155/187) than for non-immune genes (65.7%, 471/717). For example, in lung cancer, the complete TIGER architecture improves PCC of immune genes by 6.6% (0.198 → 0.211), twice the gain on non-immune genes (3.2%). Immune genes have sparse, scattered patterns, but their identity is informed by the surrounding tissue, and context adds information that individual-cell morphology cannot. This analysis is included in Section 5.2: Context matters differently for different genes. Fig. A6 in our initial submission already shows an example for CXCL13, an immune gene marking tertiary lymphoid structures (TLS). The context expert smooths its expression to near zero, while the focal expert detects it but with many false positives. TIGER adaptive gating judiciously balances the two and recovers the sparse expression pattern.

We next investigate whether the learned gating weights systematically reflect gene sparsity and spatial coherence:

Gate and sparsity (Figure A8). We stratified genes by the proportion of cells expressing each gene and compared the mean Focal Expert weight (α) between expression-positive and expression-negative cells (Fig. A8b). For genes expressed in a relatively small fraction of cells (particularly 2–10%), the gate assigned slightly higher Focal Expert weights to expressing cells. In contrast, for broadly expressed genes, the gate tended to assign greater Context Expert weight to expressing cells. 

Gate and spatial coherence. We next selected the 50 genes with the highest spatial autocorrelation (Moran's I) and compared gating weights inside and outside their expression domains (Fig. A8c). The mean Focal Expert weight was significantly lower inside these domains than outside them in both lung (0.60 vs. 0.70) and colorectal cancer (0.63 vs. 0.76). This pattern was consistent across all 16 patients in each cohort (Wilcoxon signed-rank test, p < 10⁻⁴), indicating greater reliance on the context expert within spatially coherent expression regions.

This suggests that the learned gating mechanism adjusts its reliance on local morphology and spatial context according to gene expression prevalence.

 

Additional evaluation of TIGER experts in the external cohorts and in the cell phenotyping task. Beyond the internal cross-validation experiments, we assessed how the full TIGER architecture compares to its individual experts. In the external breast and lung cohorts, the full TIGER model significantly outperforms the focal expert on PCC (0.210 vs. 0.205, q < 0.001), Recall, and AUC-PR (q < 0.01), and with no significant difference on PCC+. It significantly outperforms the context expert on all metrics except PCC+ (q < 0.001). In cell phenotyping on the colorectal cohort (16 slides), the full TIGER achieves the highest macro precision (0.452), significantly above both the context (0.421) and focal (0.432) experts (p < 0.01). Its accuracy matches the focal expert (0.580 vs. 0.579) and significantly exceeds the context expert (0.558, p < 0.05):

 

GHIST (strongest baseline)Context expertFocal expertTIGER
Macro precision ↑0.351***0.421**0.432**0.452
Accuracy ↑0.510**0.558*0.5790.580

 

Summary. Our cell representation already makes the focal expert a strong model on its own. The full TIGER architecture adds to it by adapting, per gene and per cell, how much it relies on context. This matters the most for sparse genes such as immune markers, where TIGER improves over the focal expert on 83% of them. The gains also carry over to the external cohorts and to the downstream cell phenotyping task.

AuthorsOpenReview ↗

Response to Reviewer LndX (2/3)

W2. Analysis of the cell-level representation

We thank the reviewer for highlighting this point. We benchmarked several cell representations during the development of TIGER (now included in Table 4b of the revised manuscript).

Design. Our cell embedding was designed to (1) exploit the fine-grained local morphology encoded in the token-level representations of pathology foundation models (PFMs), which are typically left unused, and (2) be computationally efficient. Most prior methods represent each cell by a CLS token, obtained either from a 224×224-pixel patch centred on the cell (e.g., Phoenix [1]) or from a tight crop around the nucleus resized to the encoder input (e.g., DeepSpot2Cell [2]). Both require one PFM forward pass per cell, which is costly for slides with hundreds of thousands of cells. In contrast, we encode each tile once and derive all its cell embeddings from the same token map at negligible additional cost. For example, on the lung cohort, this reduces foundation-model forward passes ~52-fold (1.11M cells vs. 21K tiles), and embedding extraction for a 161K-cell slide from ~20 min to under a minute.

Controlled comparison (Table 4b). We trained the focal branch with five cell representations on the lung cohort, keeping the PFM (UNIv2), head, loss, splits and checkpoint selection unchanged. Two re-encode the cell as an image (nucleus crop CLS, cell-centred crop CLS). Three read the token map: the token nearest to the centroid, a uniform average of tokens overlapping the cell, and our area-weighted average.

Cell featureSupportPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
Tight crop CLSnucleus crop0.117 (±0.016)0.344 (±0.039)0.113 (±0.017)0.462 (±0.010)0.511 (±0.012)0.232 (±0.020)0.226 (±0.020)
Cell-centred crop CLS112 µm window0.171 (±0.015)0.453 (±0.038)0.154 (±0.009)0.447 (±0.011)0.498 (±0.013)0.264 (±0.021)0.253 (±0.022)
Nearest tokencentroid0.214 (±0.025)0.520 (±0.043)0.205 (±0.023)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.027)
Uniform poolingboundary polygon0.213 (±0.025)0.521 (±0.043)0.203 (±0.022)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.026)
Area-weighted pooling (ours)boundary polygon0.220 (±0.026)0.528 (±0.044)0.205 (±0.025)0.435 (±0.013)0.490 (±0.014)0.287 (±0.026)0.276 (±0.027)

 

The tight-crop embedding performs substantially worse than our token-based embedding (PCC 0.117 vs. 0.220), which we attribute to PFMs being pretrained on larger fields of view rather than isolated cells. The cell-centred crop recovers only part of this gap (PCC 0.171), since its CLS token summarises the entire patch rather than the cell at its centre. Token-based alternatives perform close to ours (PCC 0.214 for the nearest token and 0.213 for uniform pooling), confirming that the main gain comes from reading cell-located tokens rather than re-encoding each cell. Area-weighted pooling over the cell boundary is nonetheless best on all metrics, supporting our final design.

 

[1] Tran, Manuel, et al. "Pan-cancer virtual spatial transcriptomics from routine histology with Phoenix." bioRxiv (2026): 2026-04.

[2] Wu, Qitian, et al. "Multi-Modal Foundation Model with Whole-Slide Attention Enables Transferrable Digital Pathology at Single-Cell Resolution." bioRxiv (2026): 2026-07.

AuthorsOpenReview ↗

Response to Reviewer LndX (3/3)

W3. Visium spot-level evaluation and downstream use

We thank the reviewer for these constructive suggestions. To address them, we conducted two additional experiments on spot-level generalizability and downstream utility.

Spot-level performance on real Visium data. We evaluated TIGER on 24 public Visium slides (7 breast from HEST-1k [1], 17 lung from Takano et al. [2]), aggregating its cell-level predictions within each spot (Appendix A.7.4) and comparing them with measured expression. Four spot-level methods (BLEEP, ST-Flow, ST-Net, DeepSpot) were trained on the same Xenium cohorts as TIGER, using the same UNIv2 encoder for a fair comparison, and evaluated in the same way. We report per-gene PCC across in-tissue spots, averaged over genes (254–279 breast and 295–302 lung genes per slide, depending on each slide's panel). Slides from the same patient are averaged first, giving 18 patients (7 breast, 11 lung); we then average within each cohort and across the two cohorts. TIGER is compared with each method by a one-sided paired t-test on per-patient PCC, with BH correction over the four comparisons.

Table. Spot-level performance on real Visium data. 8 patients (24 slides) across lung and breast cancer.

ModelPCC
BLEEP0.168 (±0.099)
ST-Flow0.158 (±0.090)
ST-Net0.202 (±0.103)
DeepSpot0.227 (±0.100)
TIGER0.211 (±0.091)

TIGER achieves competitive performance on real Visium data, significantly outperforming BLEEP, and ST-Flow (higher in 15 of 18 patients against each). It was comparable to ST-Net and DeepSpot, with no significant difference to either (TIGER higher in 12 and 9 of 18 patients, respectively). Notably, TIGER achieved these results without training on spot-level expression targets or retraining for Visium. These findings suggest that TIGER's cell-level predictions can generalize to spot-level ST prediction across platforms, with performance comparable to dedicated spot-level methods. 

Cell-level vs. spot-level virtual ST for downstream use. We used TransNEO treatment-response prediction with ABMIL and the protocol of Table 6. We compared three inputs: (i) H&E image features alone, (ii) image features combined with TIGER predictions aggregated into 128-µm pseudo-spots, and (iii) combined with TIGER's cell-level predictions instead. Both ST-augmented inputs improved over the image-only baseline (AUROC 0.620, bACC 0.570). Cell-level ST achieved the highest AUROC (0.658 vs. 0.637 for pseudo-spot ST), while balanced accuracy was comparable (0.590 vs. 0.596). These results suggest that predicted ST provides complementary information for downstream prediction, and that retaining cell-level resolution may offer a modest additional benefit over spatially aggregated predictions.

InputAUROC ↑Balanced acc. ↑
H&E features only0.6200.570
+ TIGER, 128 µm pseudo-spots0.6370.596
+ TIGER, cell level0.6580.590
 

[1] Jaume, Guillaume, et al. "Hest-1k: A dataset for spatial transcriptomics and histology image analysis." Advances in Neural Information Processing Systems 37 (2024): 53798-53833.

[2] Takano, Yuma, et al. "Spatially resolved gene expression profiling of tumor microenvironment reveals key steps of lung adenocarcinoma development." Nature Communications 15.1 (2024): 10637.  

 

W4. Baselines on downstream tasks

Already in the submission. Cell phenotyping against all baselines is in Section 5.3 / Figure 5. TIGER is significantly better than all the baselines for this task (p < 0.01 vs. GHIST, p < 0.001 vs. iStar, DeepSpot2Cell, sCellST and Img2ST-Net), with +13.9% accuracy over the second-best model. HER2, ER and treatment-response prediction against all five baselines and an image-only model were in Appendix A.10 / Table A12. With the extra page allowed for the revised submission, we have now moved Table A12 into the main manuscript (Table 6 in the revised manuscript).

 

Summary. The new analyses (the cell-embedding ablation, the gate and gene analysis, Visium spot-level results and downstream comparisons) are intended to address the points of the review. Since the reviewer found our single-cell performance to be strong, we hope that these additions ease the remaining doubts about the added value of our full model, and we kindly ask the reviewer to reconsider their score. We would be glad to discuss any remaining point.

Dear Reviewer,

We appreciate the time and effort you put into reviewing our work. We hope our rebuttal has helped clarify the main points raised in your review.

As the discussion period draws to a close, we would welcome any further feedback you might have, particularly if there are concerns that you feel remain insufficiently addressed.

Thank you once again for your thoughtful consideration.

Reviewer_5QHAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes TIGER for predicting single-cell spatial gene expression from H&E histology. The key idea is to separate prediction into a focal expert, intended to capture localized cell-level morphology, and a context expert that incorporates neighboring cells and global tissue information. A per-cell, per-gene gate adaptively combines the two experts. The method is evaluated on lung, breast, and colorectal Xenium cohorts, including external datasets, and is further assessed on cell phenotyping and downstream clinical prediction tasks. The paper reports substantial improvements over several recent virtual spatial transcriptomics baselines.

Critical strengths
  1. The problem formulation is strong.

At single-cell resolution, spatial context is not uniformly beneficial: it can help spatially coherent expression programs while oversmoothing sparse and localized signals. Explicitly modeling this tension is well motivated, and the per-cell, per-gene routing mechanism is a natural way to allow the usefulness of context to vary across both genes and locations.

  1. The system is practical and evaluated at meaningful scale.

The token-to-cell pooling strategy efficiently reuses pathology foundation model features rather than independently encoding every cell crop. The experiments cover multiple cancer types, external cohorts, sparsity-aware metrics, and downstream tasks, which makes the empirical scope stronger than that of many virtual-ST studies.

Critical weaknesses
  1. The focal/context decomposition is not cleanly established

The paper interprets the focal expert as capturing cell-intrinsic morphology and the context expert as capturing the microenvironment. However, the focal branch pools UNI-v2 tokens extracted from a larger histology patch rather than independently encoded cell crops. Since these ViT tokens are already contextualized by self-attention and are not strictly confined to the target cell, the focal representation may already contain neighboring-cell information.

Thus, the method appears to separate implicit/local context from explicit multi-scale context, rather than intrinsic morphology from microenvironment. A true cell-only control, such as a cropped or cell-masked focal encoder, would be important to support the biological interpretation of the adaptive gate.

  1. The evaluation pipeline raises concerns about the reported quantitative results

In the released code, the same held-out fold is used for validation/checkpoint selection and final testing, so the internal cross-validation results are not based on a strictly untouched test set.

In addition, the default PCC50 implementation appears to select genes using PCC computed on the evaluated test sample, and the Average Precision implementation does not correctly handle tied prediction scores. These issues directly affect reported metrics and require corrected evaluation. The external generalization claim should also be scoped carefully, since evaluation is restricted to genes already seen during training (93 for lung and 61 for breast).

Questions for authors

1.Were the reported internal results generated using the released protocol where the same held-out fold is used for validation/checkpoint selection and final testing? If so, please provide corrected results using a separate validation split while keeping the outer test fold fully untouched.

2.How exactly was PCC50 computed for the reported tables? Were the top-50 genes fixed using only training/validation data? If not, this metric should be recomputed.

3.Please recompute AUC-PR using a standard implementation that correctly handles tied prediction scores and report whether the main conclusions change.

4.How do the authors justify interpreting the focal branch as cell-intrinsic morphology, given that it pools ViT tokens already contextualized over a larger histology patch? A true cell-only focal baseline would materially strengthen this claim.

5.Please clarify the intended scope of the generalization claim. The current external experiments evaluate only genes already present in the training vocabulary.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5QHA (1/4)

We thank the reviewer for the constructive review and careful examination of our released code, which helped us identify and correct issues in our evaluation protocol. We also appreciate the reviewer's positive assessment of our problem formulation, computational efficiency, and comprehensive evaluation. We address each point below.

W1 + Q4. Is the focal branch cell-intrinsic?

Clarification. We agree with the reviewer that UNIv2 tokens used to construct the focal-branch embedding indirectly carry information about the cell's immediate surroundings. However, rather than viewing the focal branch as strictly cell-intrinsic, we describe it as cell-anchored, as it reads only tokens overlapping the cell's segmented boundary. In contrast, the context branch explicitly aggregates information from neighbouring cells, tiles, and the whole slide.

Motivation of the proposed cell representation. The cell embedding proposed in this work is designed with two goals: (1) exploit the fine-grained local morphology already encoded in the tokens of extensively pretrained pathology foundation models (PFMs) which are typically left unused, and (2) be computationally efficient. Most previous methods rely only on the CLS token for cell representation; either by cropping typical 224x224-pixel patches about the centroid of every cell (e.g., Phoenix [1]), or by cropping tight patches around each nucleus and resizing them to the encoder's input size (e.g., DeepSpot2Cell [2]). Both strategies require one foundation-model forward pass per cell, which becomes costly for whole slides containing hundreds of thousands of cells. In contrast, our approach encodes each tile once and derives the embeddings of all cells in that tile from the same token map, at negligible additional cost. For example, on the lung cohort, this reduces PFM forward passes ~52-fold (1.11M cells vs. 21K tiles), and embedding extraction for a 161K-cell slide from ~20 min to under a minute.

Cell-only control and additional ablations (new, Table 4b). We trained the focal branch on the lung cohort with CLS embeddings of a tight nucleus crop (the cell-only control suggested by the reviewer whose embedding does not contain contextualized information) and of a 112 µm cell-centred crop. We have also added two token variants (additional details in section 5.2, Cell embedding ablation).

Cell featureSupportPCC ↑PCC50 ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
Tight crop CLSnucleus crop0.117 (±0.016)0.344 (±0.039)0.113 (±0.017)0.462 (±0.010)0.511 (±0.012)0.232 (±0.020)0.226 (±0.020)
Cell-centred crop CLS112 µm window0.171 (±0.015)0.453 (±0.038)0.154 (±0.009)0.447 (±0.011)0.498 (±0.013)0.264 (±0.021)0.253 (±0.022)
Nearest tokencentroid0.214 (±0.025)0.520 (±0.043)0.205 (±0.023)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.027)
Uniform poolingboundary polygon0.213 (±0.025)0.521 (±0.043)0.203 (±0.022)0.437 (±0.012)0.491 (±0.013)0.284 (±0.025)0.273 (±0.026)
Area-weighted pooling (ours)boundary polygon0.220 (±0.026)0.528 (±0.044)0.205 (±0.025)0.435 (±0.013)0.490 (±0.014)0.287 (±0.026)0.276 (±0.027)

Focal branch only, lung cohort, UNIv2. Best in bold.

The cell-only embedding performs substantially worse than ours (PCC 0.117 vs. 0.220), which we attribute to PFMs being pretrained on large fields of view, not isolated cells. The cell-centred crop recovers only part of this gap (0.171) since the CLS token summarizes the entire patch rather than the cell at its centre. Reading cell-localized tokens from the contextualized token map is therefore both computationally cheaper and more informative than re-encoding each cell as a separate image.

[1] Tran, Manuel, et al. "Pan-cancer virtual spatial transcriptomics from routine histology with Phoenix." bioRxiv (2026): 2026-04.

[2] Nonchev, Kalin, et al. "Deepspot2cell: predicting virtual single-cell spatial transcriptomics from H&E images using spot-level supervision." bioRxiv (2025): 2025-09.

AuthorsOpenReview ↗

Response to Reviewer 5QHA (2/4)

W2 + Q1. Clarification and refinement of the evaluation protocol 

Yes, in the submitted version, the released protocol used the same held-out fold both for checkpoint selection and for final testing, and we thank the reviewer for catching this. We reran the entire internal benchmark, retraining TIGER and all five baselines with separate validation and test folds. For each test fold, the next fold was used exclusively for checkpoint selection (see Section A.11). The updated split files have been added in the released code (ids_validation_set.csv). As shown in the table below and in the new appendix tables Table A19-21, the updated evaluation does not modify the ranking across baselines and TIGER retains the best aggregate performance across all seven metrics under this corrected protocol. The external experiments are unaffected, which use cohorts never seen during training or model selection. 

To further quantify how much the protocol change affects the results, we correlated the per-patient metrics of the submitted protocol with those of the corrected protocol, over the same patients and the six models (258 model–patient pairs). The two protocols agree closely at every level (r = 0.93–0.99). TIGER ranks first on all seven metrics under both protocols. The only rank changes are among weaker baselines (PCC50: iStar and sCellST; bMAE: DeepSpot2Cell and iStar; bRMSE: GHIST and iStar). The direction of the per-patient comparison between TIGER and each baseline is the same under both protocols in 98.5% of the 35 baseline × metric comparisons. The findings of the paper (the advantage of TIGER over all baselines, the contribution of the focal and context experts and of the gating) therefore do not depend on this choice. We commit to re-running all internal results (including the appendix tables) with the new protocol and to updating the paper accordingly in the final version.

 

ModelTraining, resolutionPCC ↑PCC₅₀ ↑PCC+ ↑bMAE ↓bRMSE ↓Recall ↑AUC-PR ↑
Img2ST-NetFull, bin0.040 (±0.010)0.112 (±0.025)0.023 (±0.012)0.493 (±0.013)0.592 (±0.017)0.199 (±0.026)0.198 (±0.027)
iStarWeak, bin0.112 (±0.021)0.271 (±0.032)0.050 (±0.018)0.479 (±0.016)0.517 (±0.019)0.241 (±0.024)0.234 (±0.025)
sCellSTWeak, cell0.126 (±0.016)0.304 (±0.031)0.089 (±0.013)0.467 (±0.013)0.555 (±0.015)0.248 (±0.021)0.243 (±0.023)
DeepSpot2CellWeak, cell0.188 (±0.023)0.360 (±0.030)0.148 (±0.018)0.473 (±0.014)0.523 (±0.017)0.285 (±0.024)0.272 (±0.026)
GHISTFull, cell0.201 (±0.026)0.470 (±0.049)0.164 (±0.018)0.448 (±0.016)0.520 (±0.018)0.284 (±0.021)0.278 (±0.023)
TIGER (ours)Full, cell0.283 (±0.027)0.526 (±0.036)0.237 (±0.020)0.430 (±0.013)0.497 (±0.014)0.329 (±0.022)0.319 (±0.025)

New protocol, macro-average over the lung, breast and colorectal cohorts (Table A19). 

 

AuthorsOpenReview ↗

Response to Reviewer 5QHA (3/4)

W2 + Q2. How was PCC50 computed?

Our original PCC50 evaluated the top 50 genes selected separately for each test sample, reflecting the best achievable correlation rather than performance on a predefined gene set. We clarify that our initial calculation was not data leakage, but a different evaluation choice intended to showcase the peak predictive performance achievable on each test sample. However, we agree with the reviewer that from a practical deployment perspective, the gene set should be fixed using validation data without relying on test sample statistics. We have refined this metric to provide a more rigorous assessment using genes selected exclusively from the validation data (see Appendix A.11: Methods details).

TIGER remains the best model under this definition (PCC50 0.526 vs. 0.470 for GHIST, +12%; submitted protocol: 0.617 vs. 0.556, +11%), and the correlation between the two definitions is high (per-patient Spearman ρ = 0.950). The new appendix table (Table A19-21) uses the validation-based definition, and PCC50 will be reported this way in all tables of the final version. 

We note that the submitted definition was applied identically to all models, with each model's genes selected from its own predictions, so it did not favour TIGER in the comparison between methods, and mainly inflated the absolute PCC50 values of every model.

W2 + Q3. AUC-PR with tied scores

We thank the reviewer for spotting this subtle definition inconsistency. We recomputed AUC-PR for TIGER on all 43 internal samples with the updated definition that handles tied prediction scores and the differences are below 0.001, an order of magnitude smaller than the fold-to-fold standard deviation. They do not change the ranking of the methods or any conclusion of the submitted manuscript. Our original implementation did not group tied scores: cells with identical scores were ordered by their position in the file. We now use the suggested definition (AP = Σ (R_k − R_{k−1}) P_k), evaluated only at distinct score thresholds, which matches scikit-learn's average_precision_score (max deviation 4e-8). All new appendix tables (Table A19-21) use this updated version of AUC-RR. 

CohortnAUC-PR (tie-aware)AUC-PR (original)Δ
Lung160.27590.2762−0.0003
Breast110.45420.4547−0.0004
Colorectal160.22740.2282−0.0008
Macro430.31920.3197−0.0005

 

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

Related papers

Discussion (0)

Sign in to comment.