acceptodds
Under review as a conference paper at ICLR 2027

Every Gene Count Has a Platform and a Slide: Measurement-Aware Observation Model for H&E-to-Spatial-Transcriptomics Prediction

Abstract

Predicting spatial transcriptomics (ST) from H&E images is moving toward foundation models trained by pooling data across cohorts and platforms, on the assumption that a larger pool makes the model more generalizable in a new cohort. We tested this assumption with five H&E-to-ST models and found that the larger pool gave no consistent gain on the HEST benchmark or across downstream tasks. The cause, we argue, lies in the target, not the model. We first clarify that gene counts, the prediction target, are measurements rather than expression itself, and show how far this shifts the target: expression is read at a scale that differs by platform and gene, at a level that varies from slide to slide, and as counts whose noise varies by gene and platform. The standard objective, squared error on log counts, has a term for none of these, so the output must account for everything the measurement contributed. We then propose an observation model that attaches to any regressor without changing its architecture, with a term for each of the three: the per-gene scale of each platform, estimated once and held fixed; the variation of each slide, absorbed in training by an intercept of its own; and the counts themselves, modeled with a negative binomial likelihood. Under our observation model, the same five models all improve with pooling on the benchmark. The improvement carries over to three further tasks: detection of tertiary lymphoid structures, spatial clustering, and evaluation on a new platform, with a gain over the standard objective in nearly every comparison. Our code is available at https://anonymous.4open.science/r/ST-Obs/.

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

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xaiVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This study proposes an observational model that integrates (1) the capture area, (2) the platform, (3) the slide to account for batch effects when pooling data across cohorts in virtual ST applications. The proposed observational model can be integrated with any virtual ST regressor. The model is evaluated with five backbones (Probe, Probe+LoRA, Simple-TF, STPath, STORM) in expert (only Visium or Xenium datasets), and pooled (mixing both modalities) training settings. The evaluation is performed on the HEST Benchmark and also additional tasks including external validation on the newly released Atera platform and spatial clustering. The observation model provides moderate improvements for integrating ST data across modalities. The ablation shows that the platform and slide term drive the gain, not the likelihood alone.

Critical strengths
  • The study is clearly written and allows readers to easily follow it.

  • Easy to be adopted. The observation model proposed in this study can be integrated to any virtual ST regression framework which could benefit existing and future virtual prediction strategies.

  • External validation on newly released ST platforms. The authors assess their observation model on the Atera platform which improves by 10 to 28% on all the five backbones.

Critical weaknesses
  • Contextualization of the work and motivation of the study. Only a few virtual ST works have pooled data across platforms (which is the pooling tested on this work) since batch effects and platform differences are well known issues of spatial transcriptomics. From the cited papers, only STORM (arXiv,2026) integrates data across modalities. Phoenix only trains on Xenium data to explicitly avoid batch effects from pretraining. Another framework that recently incorporates data across multiple models is TissueFormer (Qitian Wu et al, 2026) which should be evaluated with the proposed observation model if possible. In addition, the claim that the sensitivity of Xenium is greater than Visium (section 2.1) has also been previously observed and should be acknowledged (e.g., STP-BENCH, Y. Chung et al., 2026).

  • Methodological innovation. While the proposed observation model is well motivated, some of the components have already been utilized in prior works. As pointed out by the authors in Section 6, count likelihoods (negative binomial likelihood) have already been incorporated in prior studies such as STimage, and also more recent works such sCellST. The authors should also move Related Work (Section 6) to the beginning of the manuscript, to make this explicit and clear for the readers what are the contributions of the work.

  • Moderate and inconsistent improvements. In Table 1 under the Xenium expert setting, the observation counts does not provide consistent improvements and often results in lower performance. While the platform here is just Xenium, the sample and the proposed negative binomial likelihood does not seem to improve.

Questions for authors

Q1. Since the observation model can be applied to any virtual ST regressor, the study would benefit from incorporating common virtual ST framework suchs as STNet (Nat BME, 2020), BLEEP (NeurIPS, 2023), TRIPLEX (CVPR, 2024)… The authors assess Probe, Probe+LoRA, Simple-TF and their implementation of STPath and STORM but conventional ST methods should also be included as part of the benchmarking.

Q2. While Pearson Correlation Coefficient (PCC) is one of the most widely used metrics in virtual ST applications, other metrics are also used. Reporting metrics such as MAE or RMSE would make the claims stronger.

Q3. Are the moderate improvements statistically significant?

Q4. ST data from different modalities is aggregated at different resolutions. Visium is left at the native 55-um spots, while Xenium is pooled in 100 um squares. Many virtual ST methods are now being adopted at the single-cell level (e.g., sCellST, DeepSpot2Cell, GHIST…). Does the benefit of the observation model also holds when Xenium data is pooled into smaller bins or left at its native cell-level?

9 comments
AuthorsOpenReview ↗

Response to Reviewer xaiV (1/6)

We sincerely thank Reviewer xaiV for the detailed evaluation and constructive comments on our manuscript. We are pleased that the reviewer found the paper clearly written and easy to follow. We also appreciate the recognition that the observation model is easy to adopt in any virtual ST regression framework and can therefore benefit both existing and future methods. We are glad as well that the reviewer valued the external validation on the newly released Atera platform, where the observation model improves all five backbones by 10–28% on the two samples in the manuscript.

We have carefully considered all of the reviewer's comments and address each of them below.


Regarding the contextualization and motivation of the study (Weakness 1)

We agree that batch effects and platform differences are well-known issues in spatial transcriptomics, as we also note in Sections 1 and 2. The Phoenix study mentioned by the reviewer illustrates this well. It trained only on Xenium data and explicitly excluded all Visium data, citing Visium's lower sensitivity, lateral diffusion, and other artifacts, and it also reported reduced performance when trained on Xenium 5K. Given these issues, training on a single platform is one reasonable approach. In our view, however, such an approach leaves the large number of publicly available Visium slides unused and thus limits the size of the training data. By modeling the differences between platforms and slides explicitly rather than excluding data, our method aims to let such data contribute to training together with data from other platforms.

Revised manuscript: Sections 1 and 2 → Sections 1 and 3.

We also thank the reviewer for pointing out that Phoenix is trained only on Xenium data. In the Introduction, we cited it as an example of pooling slides across cohorts and organs at scale. We agree that it does not combine platforms, and we apologize for the confusion. In the revised manuscript, we have replaced Phoenix in that sentence with TissueFormer, which the reviewer mentioned, and STPath, so that all works cited there (STORM, TissueFormer, and STPath) pool slides across cohorts, organs, and platforms.

Revised manuscript: this sentence is in the first paragraph of the Introduction (Section 1).

Despite these known measurement differences, a growing number of studies use data from multiple platforms together, even when they do not directly predict ST from H&E. MINT [1] fine-tunes a pathology foundation model with ST supervision, using separate expression heads for Visium and Xenium. scGPT-spatial [2] continues the pretraining of the single-cell foundation model scGPT on spots and cells from Visium, Visium HD, Xenium, and MERFISH, with modality tokens and a mixture-of-experts decoder. These approaches handle platform differences with learned embeddings or separate heads, whereas our observation model accounts for them in the training objective, outside the backbone.

We agree that the higher sensitivity of Xenium relative to Visium has been reported before. Section 2.1 cites Janesick et al. (2023) for this, and the revised manuscript also cites STP-Bench, which the reviewer mentioned and which reports the same tendency.

Revised manuscript: Section 2.1 → Section 3.1; STP-Bench is cited in Section 2 (Related Work).

Regarding TissueFormer, which the reviewer suggested evaluating with the observation model, please see our response to Q1, where we report its results together with those of the additional baselines.

[1] Lee, Minsoo, et al. "MINT: molecularly informed training with spatial transcriptomics supervision for pathology foundation models." International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2026.

[2] Wang, Chloe, et al. "scGPT-spatial: continual pretraining of single-cell foundation model for spatial transcriptomics." biorxiv (2025): 2025-02.

AuthorsOpenReview ↗

Response to Reviewer xaiV (2/6)

Regarding the methodological novelty (Weakness 2)

As noted in Section 6, count likelihoods have already been used for H&E-to-ST prediction, including the NB likelihood in STimage. The NB distribution is also used in the statistical analysis of ST data; for example, Zhao et al. [1], whom we cite in Section 3, report that ST counts can be modeled with count distributions such as the NB without zero inflation. We therefore agree with the reviewer that the NB likelihood itself is not new, and we do not claim its use as a contribution of this work.

Revised manuscript: Section 6 (Related Work) → Section 2; Section 3 (Method) → Section 4.

In addition, we do not claim that the NB likelihood is the main source of the performance gain. In our ablation (Section 5.3; Appendix H, Table 29), the NB likelihood alone improves over MSE on only 32 of the 50 task–backbone cells, compared with 44 for the full observation model, and the gain comes mainly from the platform and slide terms, as the reviewer also noted in the summary.

Revised manuscript: Section 5.3 → Section 6.3; Appendix H, Table 29 → Appendix I, Table 38.

We would also like to clarify that we chose the NB likelihood not only because the target is a count, but also because it can model how count variability differs across genes and platforms. As described in Section 3, its dispersion has a gene factor and a platform factor, so these differences enter the training objective directly. This makes it well suited to training on data pooled across platforms. We have clarified this motivation in the Introduction of the revised manuscript.

Revised manuscript: Section 3 → Section 4; the motivation is stated in the Introduction (Section 1).

Following the reviewer's suggestion, we have moved the Related Work section to right after the Introduction in the revised manuscript, so that prior work is presented before our analysis and method. In that section, we discuss how prior studies have reported platform differences, addressed batch effects in ST, and used count likelihoods for H&E-to-ST prediction, and then explain what our work adds by modeling platform and slide effects together in the training objective for pooled data. Since this shifts the numbering of the later sections, the section numbers in our responses refer to the submitted manuscript, and the corresponding location in the revised manuscript is given in a quoted line below the paragraph.

Revised manuscript: Related Work is now Section 2, and the extended related work is Appendix A.

[1] Zhao, Peiyao, et al. "Modeling zero inflation is not necessary for spatial transcriptomics." Genome biology 23.1 (2022): 118.


Regarding the moderate and inconsistent improvements (Weakness 3)

We agree that the observation model does not provide consistent improvements in the Xenium expert setting. Over the five Xenium tasks, it yields a significantly lower PCC than MSE training for two of the five backbones: Probe+LoRA (0.515 → 0.497; difference −0.019 before rounding, 95% CI [−0.028, −0.009]) and STORM† (0.449 → 0.431; difference −0.018, 95% CI [−0.028, −0.008]). For the other three backbones (Probe, Simple-TF, and STPath†), the difference is not statistically significant, as the 95% confidence interval includes zero. All intervals are computed with the same bootstrap as in our response to Q3.

We would like to clarify that the proposed model is primarily designed as a training recipe for data pooled across multiple platforms, and that improving single-platform training is not its main target. Even so, it can also improve single-platform training: in the Visium expert setting, it achieves a higher PCC than MSE on all five backbones (Table 1), with significant improvements for Probe and STPath†.

Why the observation model is less effective in the Xenium expert setting than in the Visium expert setting deserves further investigation. Since Xenium is widely used for ST at single-cell resolution, we believe that improving the observation model for training on Xenium alone, as in the Phoenix scenario, is a valuable direction for future work.

AuthorsOpenReview ↗

Response to Reviewer xaiV (3/6)

Regarding additional baselines (Q1)

We agree that including conventional virtual ST frameworks would strengthen our benchmark. In response to this comment, we trained STNet and TRIPLEX on the pooled set, once with MSE and once with the observation model, and evaluated them on the HEST benchmark (Table R5). Because of the limited time of the discussion period, we could not repeat every experiment of the manuscript for these models, and we report the pooled setting only.

We followed the released architectures and hyperparameters of both models wherever applicable, and used the same data split, gene space, and evaluation protocol as in our main experiments. There are a few departures from the released settings: we increased the batch sizes to speed up training and used UNI2-h features for TRIPLEX's neighbor and global views. Each model was trained with five seeds per objective, with all other settings matched between MSE and the observation model.

Table R5. STNet and TRIPLEX trained on the pooled set with MSE or with the observation model (Obs.): PCC over the 50 benchmark genes on the ten HEST benchmark tasks, mean over five seeds. Mean: the ten-task mean, mean ± sd over seeds. Bold: the better objective per model and task (both when tied at three decimals).

ModelObjectiveREADCCRCCHCCLYMPHPRADCOADIDCLUNGPAADSKCMMean
STNetMSE−0.0000.1170.0340.1440.2070.3510.4230.4960.3890.2850.244 ± 0.019
Obs.−0.0010.1900.0460.1820.3190.3690.4650.5100.3500.3760.281 ± 0.004
TRIPLEXMSE0.2380.2100.0940.2500.3310.5300.6220.6280.4870.6610.405 ± 0.004
Obs.0.2560.2180.1180.2480.3250.5300.6180.6340.4890.6680.410 ± 0.003

For STNet, replacing the released squared-error loss with the observation model raised the ten-task mean PCC from 0.244 to 0.281, in all five seeds and in 8 of the ten tasks. For TRIPLEX, it raised the ten-task mean PCC from 0.405 to 0.410, in all five seeds and in 7 of the ten tasks.

Reviewer_HC9qOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies H&E-to-spatial-transcriptomics prediction when training data are pooled across cohorts and ST platforms. Through a variance-decomposition analysis on HEST-1k, the authors argue that pooling fails to help under the standard MSE objective because gene counts carry platform-specific scales, slide-level measurement levels, and count noise that the objective does not model. They propose a plug-in observation model with a fixed per-gene platform offset, a slide-specific intercept integrated out at inference, and a negative binomial likelihood. Experiments are conducted with five backbones on the HEST benchmark, with additional downstream evaluations on TLS detection, spatial clustering, and transfer to a new platform.

Critical strengths
  1. The studied problem is important. Whether H&E-to-ST foundation models actually benefit from large-scale data across cohorts and platforms is a question of broad current interest. The paper offers a systematic analysis of this question, and these analyses provide useful insights into the generalization of such models.
  2. The proposed method is simple yet intuitive and can be plugged into any existing regressor.
  3. The evaluation is comprehensive. Experiments cover the ten HEST benchmark tasks spanning multiple cancer types, organs, and platforms, together with five backbones.
Critical weaknesses
  1. My main concern is that HEST-1k is relatively small, so it is unclear whether conclusions about generalization drawn from it justify claims about foundation-model-scale pretraining. After filtering, the training pool contains only 546 slides. The number of distinct patients is not reported. Besides, many organ–platform combinations contain only one to a few slides (Table 9). This is orders of magnitude smaller than the data used by recent H&E-to-ST foundation models (e.g., 22.2M Xenium cell pairs for Phoenix, as cited in the paper).
  2. Constructing both training and evaluation sets from HEST-1k carries a risk of data leakage, since slides from different organs, studies, or tasks may come from the same patient.
  3. Estimating the platform offsets from the training set is risky. The offset c_gp is computed as a stratified ratio of mean counts on the training slides and then held fixed. Yet the platforms are applied to very different mixtures of organs and patients, and the overlap between Visium and Xenium is limited to essentially two organs. The estimates may therefore encode tissue composition or patient-specific signal rather than a pure platform effect. Besides, only 83.8% of (gene, platform) pairs receive one, and on Atera only 1,539 of 18,023 genes do.
Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer HC9q (1/3)

We sincerely thank Reviewer HC9q for the thorough evaluation and constructive comments on our manuscript. We are pleased that the reviewer considered the problem important and of broad current interest, and found that our systematic analysis provides useful insights into the generalization of H&E-to-ST models. We also appreciate the recognition that the proposed method is simple, intuitive, and applicable to any existing regressor, and that the evaluation is comprehensive, covering the ten HEST benchmark tasks across multiple cancer types, organs, and platforms with five backbones.

We have carefully considered all of the reviewer's comments and address the three concerns below.


Regarding the dataset scale and the number of patients (Weakness 1)

We first acknowledge that HEST-1k is relatively small compared with the training data of the largest recent H&E-to-ST foundation models. At the same time, we would like to clarify that the aim of this work is not to increase the absolute amount of training data, but to address the cross-cohort and cross-platform variation that grows as training data are pooled from more sources. We chose HEST-1k because it combines hundreds of slides from many cohorts, platforms, and organs, and therefore reflects this problem setting well. We also acknowledge that many organ–platform combinations in HEST-1k contain only one to a few slides (Table 9).

Revised manuscript: Table 9 → Table 11.

The benefit of the observation model in pooled training was verified at the scale of HEST-1k, and the same trained models also outperformed pooled MSE training on all six Atera samples in the external validation (Table R1). Its effect at larger training scales remains a subject for future work. Although we have not tested this experimentally, we expect larger training datasets to affect our setting in two ways. First, additional cohorts would add slides to the sparse organ–platform combinations, which could make the estimates of the platform offsets more reliable, since the offsets are estimated within organs and image clusters. Second, each additional cohort is likely to bring its own measurement level (e.g., library size), which would add to the slide-to-slide variation that our slide term addresses.

We would also like to note that the scale of our training set may be larger than the number of slides suggests. The 546 training slides contain approximately 1.03M spots, each with its own measured expression profile, which is comparable in scale to the training data of recent H&E-to-ST foundation models such as STORM (632 sections, 1.2M spots). According to the CellViT nuclei segmentation provided with HEST-1k, about 20M cells lie within the spot areas used as training targets (44M over the whole slides), although our expression is measured per spot rather than per cell.

Regarding the number of distinct patients, patient identifiers are available for only 185 of our 546 training slides, which makes a reliable count difficult. Based on the available metadata, the number of distinct patients in our training set lies between 172 and 454. The lower bound counts each study without patient identifiers as a single patient, and the upper bound counts every unlabeled slide as a different patient. Because this range is too wide to be reported as a reliable number, we did not include it in the manuscript.


Regarding the risk of data leakage (Weakness 2)

We agree that this concern is well founded, as HEST-1k assembles public data released by several vendors together with datasets released by prior studies.

As described in Appendix C, we used the HEST-1k metadata to separate the training pool from the benchmark test slides used in our main evaluation. To prevent leakage, we removed every slide that the HEST-1k metadata records as sharing a patient with a test slide. We also removed the slides that HEST-1k assigns to the same study as a test slide, even when they have a different patient identifier or no patient identifier. The single exception to the latter rule, and the relation of each downstream evaluation set to the training pool, are documented in Appendix C.

Revised manuscript: Appendix C → Appendix D (the split in D.1, the downstream evaluation sets in D.3).

AuthorsOpenReview ↗

Response to Reviewer HC9q (2/3)

Regarding the estimation of platform offsets from the training set (Weakness 3)

We agree that the organ overlap between platforms is limited, and that a simple offset estimate, such as one stratified by organ alone, could therefore be risky. As the reviewer notes, only two organs (breast and lung) have at least three training slides on each of the two platforms (Visium and Xenium), the criterion used in Appendix A.1, whereas nine organs have at least one training slide on each platform (Table 9). An organ with only a few slides can still contribute an organ stratum to the Xenium offset, but we agree that not all of these organs overlap sufficiently, and we expect that more slides per organ would make the estimates more stable. For this reason, we do not rely on the organ strata alone: we also estimate the offset within morphological clusters of image embeddings (with 20 random seeds), which are formed regardless of organ, and average the two estimates.

Revised manuscript: Appendix A.1 → Appendix B.1; Table 9 → Table 11.

Regarding the concern that the offsets may encode biology, we expect the biological information that the offsets can carry to be limited by their construction, as discussed in our response to Reviewer gJ5U. The offset is a single constant fixed for each (gene, platform) pair, and every slide and every spot of that platform uses the same value. Therefore, even if some tissue differences were mixed into the offset, they would amount to a single platform-wide average. This one number can neither track the tissue and disease states that differ from slide to slide nor vary across spots within a slide. Our response to Reviewer gJ5U also reports two further analyses. First, a leave-one-study-out analysis shows that the Xenium and Visium HD offsets are stable across cohorts. Second, a validation on sections of the same tissue measured on two platforms shows that, at the platform level, the offsets are close to the directly measured ratios (median deviation of 1.6-fold, compared with 7.0-fold without an offset).

Regarding the concern that platform offsets are estimated for only 83.8% of (gene, platform) pairs, we provide Tables R2 and R3 below. As shown in Table R2, 99.9% of (gene, platform) pairs receive an offset on Xenium, Xenium Prime 5K, and Visium HD, and 99.8% of the pairs without an offset lie on Spatial Transcriptomics and Visium HD 3′. Moreover, as shown in Table R3, 70.4% of these pairs involve genes in the lowest expression quartile, whose mean count per spot across the training slides that measure the gene is below 0.002, whereas 99.2% of pairs involving genes in the upper half of expression receive an offset. For such lowly expressed genes, counts are too sparse to estimate the cross-platform ratio reliably. Lowering the per-stratum spot and count thresholds to their minimum raises coverage to 93.8%, but most of the newly obtained values rest on only one or two counts, too few for a reliable estimate. Even at this minimum threshold, about two-thirds of the pairs that still lack an offset correspond to genes that the platform's training slides measured but never detected, leaving no signal from which a ratio could be estimated.

Table R2. Offset coverage by platform.

Platform(gene, platform) pairs measured on both the platform and VisiumCoveragePairs without an offset
Xenium1,58699.9%1
Xenium Prime 5K5,09199.9%4
Visium HD18,27399.9%26
Spatial Transcriptomics29,92277.8%6,635
Visium HD 3′29,73076.2%7,063
Total84,60283.8%13,729

Table R3. Offset coverage by expression level.

Expression quartile (mean count in training)CoveragePairs without an offset
Top 25%99.8%69
50–75%98.4%381
25–50%76.5%3,615
Bottom 25%25.7%9,664

Regarding Atera, where only 1,539 of 18,023 genes receive an offset: as described in Appendix G.3, the pooled model reads Atera as Xenium, the imaging platform of the same family, and uses the Xenium offsets as they are. The number 1,539 is therefore the number of Xenium (v1) genes that carry an offset and that Atera also measures. Each Xenium v1 panel contains only a few hundred genes, and 1,585 of the 1,586 Xenium genes that Visium also measures (99.9%) receive an offset (Table R2). The limited number thus reflects the gene coverage of the Xenium v1 panels rather than a limitation of the estimation method.

Revised manuscript: Appendix G.3 → Appendix H.3.

AuthorsOpenReview ↗

Response to Reviewer HC9q (3/3)

To further address this concern, we estimated Atera-specific offsets from the breast and cervical samples in the manuscript, using the same procedure as for the other platforms: we averaged the organ and image-cluster estimates where both were available and otherwise used the one available. This gave offsets for 17,763 of 18,023 genes (98.6%). We applied them to all six samples at inference using the same trained models as in Table R1.

Table R4. The pooled observation model evaluated on Atera using either the Xenium offsets (as in the manuscript and Table R1) or the Atera-specific offsets, mean ± sd over five seeds. PCC and MSE (mean squared error on log(1 + y)) are computed over all genes. *: the two samples used to estimate the Atera-specific offsets. Bold: the better of the two offsets per backbone and sample.

BackboneOffsetBreast*Cervix*KidneyBrainColorectalBreast (FF)
PCC (all genes, ↑)
ProbeXenium0.635±0.0050.507±0.0040.510±0.0250.402±0.0160.578±0.0100.609±0.008
Atera0.654±0.0080.514±0.0040.523±0.0230.407±0.0160.588±0.0100.619±0.008
Probe+LoRAXenium0.669±0.0030.529±0.0050.566±0.0100.429±0.0110.593±0.0160.615±0.018
Atera0.685±0.0020.535±0.0050.575±0.0080.434±0.0100.602±0.0140.624±0.015
Simple-TFXenium0.693±0.0050.526±0.0080.545±0.0290.425±0.0240.573±0.0280.621±0.007
Atera0.703±0.0050.535±0.0080.554±0.0270.427±0.0230.581±0.0270.626±0.008
STPath†Xenium0.667±0.0130.487±0.0110.513±0.0390.373±0.0390.531±0.0280.586±0.006
Atera0.680±0.0130.493±0.0120.523±0.0390.377±0.0400.543±0.0280.594±0.010
STORM†Xenium0.626±0.0090.452±0.0020.473±0.0070.332±0.0130.502±0.0260.552±0.018
Atera0.643±0.0090.459±0.0030.483±0.0050.337±0.0130.511±0.0250.566±0.017
MSE (all genes, ↓)
ProbeXenium1.323±0.0591.305±0.0291.817±0.2382.090±0.1083.128±0.1663.042±0.194
Atera0.587±0.0480.520±0.0321.082±0.0841.375±0.0961.823±0.1171.780±0.113
Probe+LoRAXenium1.255±0.0311.218±0.0171.731±0.1282.119±0.1343.068±0.2583.028±0.273
Atera0.545±0.0300.547±0.0340.939±0.0681.399±0.1121.831±0.1461.816±0.188
Simple-TFXenium1.048±0.0351.351±0.0351.673±0.1201.590±0.0553.339±0.1402.769±0.105
Atera0.402±0.0070.504±0.0150.901±0.0971.037±0.0361.883±0.0851.540±0.063
STPath†Xenium1.193±0.1241.383±0.1301.993±0.2972.407±0.1584.113±0.5653.664±0.368
Atera0.553±0.0760.627±0.0811.187±0.2591.723±0.1612.634±0.5302.284±0.319
STORM†Xenium1.315±0.0941.440±0.0522.228±0.1192.161±0.1983.257±0.3213.092±0.196
Atera0.620±0.0620.605±0.0341.332±0.0851.412±0.1941.866±0.2291.750±0.158

On the four new samples, which were not used to estimate the offsets, the Atera-specific offsets reduced the MSE by 28–46% for every sample and backbone, and also increased the PCC in every case (Table R4). This suggests that the same procedure can estimate the offsets of a new platform from as few as two slides, and that these offsets further improve predictions on new samples of that platform.

We have added Tables R2–R4 and the accompanying analyses to the appendix of the revised manuscript.

Revised manuscript: Tables R2 and R3 are Tables 9 and 10 in Appendix C.2, and Table R4 is Table 36 in Appendix H.4.

Reviewer_gJ5UOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates why pooling data across cohorts and platforms does not always improve spatial transcriptomics prediction from H&E images. The authors argue that observed gene counts reflect tissue expression, platform-specific measurement scales, slide effects, and count noise. They propose an observation model that can be added to different regression backbones. It explicitly models the measurement process through pre-estimated gene–platform offsets, low-rank slide effects, and a negative binomial likelihood, while marginalizing over unknown slide effects at inference.

Critical strengths

Modeling platform scales, slide variation, and count noise separately offers a useful perspective on why training with pooled data can yield inconsistent gains. The approach addresses the prediction target rather than relying on a more complex architecture.

The observation model leaves the main prediction architecture unchanged. Comparing different objectives on the same training pools across five backbones helps assess whether the gains extend beyond a particular network design.

Critical weaknesses

In Table 2, “GT” refers to a signature score computed from measured expression, not the TLS annotations themselves. A label such as “measured-expression reference” would be clearer and would avoid suggesting that outperforming this score means exceeding a ground-truth upper bound.

On page 5, platform-specific gene count ratios are estimated relative to a reference platform within strata defined by organ and image-embedding clusters. This controls for some compositional differences, but it does not ensure that tissue states, cell compositions, and disease backgrounds are comparable within strata across platforms. The fixed platform offsets may therefore capture biological differences as well as measurement efficiency. Validation on paired or adjacent tissue sections, together with an assessment of offset stability across cohorts, would strengthen this part of the method.

Questions for authors

How do the authors establish that the fixed platform offsets primarily capture measurement differences rather than differences in tissue composition or disease state? Are results available from paired sections or cross-cohort stability analyses?

Does the low-rank slide term absorb biologically meaningful variation? How is it associated with tissue type, disease state, and cell composition?

4 comments
AuthorsOpenReview ↗

Response to Reviewer gJ5U (1/2)

We sincerely thank Reviewer gJ5U for the careful evaluation and insightful comments on our manuscript. We are pleased that the reviewer found that modeling platform scale, slide variation, and count noise separately offers a useful perspective on why pooled training can yield inconsistent gains. We also appreciate the recognition that the observation model addresses the prediction target while leaving the backbone architecture unchanged, and that comparing objectives on the same training pools across five backbones helps assess whether the gains extend beyond a particular network design.

We have carefully considered all of the reviewer's comments and address each of them below.


Regarding the "GT" label in Table 2 (Weakness 1)

We agree with the reviewer. In the revised manuscript, we have replaced "GT" with "measured-expression reference" throughout the main text and the appendix. We thank the reviewer for this helpful suggestion.


Regarding what the platform offset captures (Weakness 2, Q1)

We agree that the platform offset cannot be fully separated from tissue differences, because different platforms in the training data measure different tissue samples. If the ratio were computed by pooling all spots of each platform, differences in organ and tissue composition across platforms would enter the offset directly. To mitigate this, we compute the ratio of each platform to Visium separately within each organ and within each image cluster (morphologically similar tissue grouped by k-means on UNI2-h embeddings), and use the average of the two estimates as the offset. In the organ estimate, for example, ALB is highly expressed in the liver, but the liver spots of each platform are compared only with Visium liver spots. The high hepatic expression of ALB is therefore present on both sides of the ratio, so the difference in ALB expression between organs does not enter the offset as a platform difference. The image-cluster estimate, in turn, compares regions of similar morphology, such as tumor epithelium or stroma, regardless of organ, which reduces the effect of differences in tissue composition.

Moreover, we expect the biological information that the offset can carry, such as tissue state, to be limited by its construction. The offset is a single constant fixed for each (gene, platform) pair, and every slide and every spot of that platform uses the same value. Therefore, even if some tissue differences were mixed into the offset, they would amount to a single platform-wide average. This one number can neither track the tissue and disease states that differ from slide to slide nor vary across spots within a slide.

To assess whether any single study drives the offsets, we re-estimated the platform offsets after leaving out each of the 91 training studies in turn, keeping the organ and image-cluster strata fixed. The studies differ in organ and disease composition. Even so, among the genes that still received an offset after a study was removed, the fraction whose offset changed by more than two-fold was at most 4.2% for Xenium and 1.1% for Visium HD, the values reached when the most influential study was removed. For every platform, the median of this fraction over the removed studies was at most 0.07%.

To validate the platform-level scale of the offsets, we identified 16 pairs in HEST-1k in which the same section, or sections from the same tissue block, were measured on two different platforms. Because both slides of a pair come from the same tissue, the count ratio measured directly on a pair is little affected by tissue composition and mainly reflects measurement differences. We then re-estimated the offsets after excluding the training slides from these blocks and from other candidate same-tissue blocks (34 slides). For each pair, we took the median over genes of the difference between the offset-predicted and the measured log count ratio; across the 16 pairs, this corresponded to a median factor of 1.6. In contrast, assuming equal counts on the two platforms (i.e., no offset) yielded a median deviation of 7.0-fold, and the offset prediction was closer in 15 of the 16 pairs. Moreover, a ratio estimated without stratification deviated by a median factor of 2.5, consistent with stratification by organ and image cluster reducing the influence of tissue composition.

We have added these analyses to the appendix of the revised manuscript.

Revised manuscript: Appendix C.3.

AuthorsOpenReview ↗

Response to Reviewer gJ5U (2/2)

Regarding the slide term (Q2)

As stated in our Limitations section, we agree that the slide intercept absorbs whatever differs between slides, and that part of it may therefore be biological. With this in mind, the slide term is designed to limit what it can absorb. The intercept is low-rank, shrunk by a quadratic penalty, and centered to zero within each platform, so it cannot carry any component shared by an entire platform. Furthermore, it takes the same value at every spot of a slide, so it cannot absorb the variation across spots within a slide.

To examine how the slide term relates to biological factors, we regressed the slide term learned by the Probe model trained on the pooled set on each slide attribute in turn. These multivariate regressions were run separately for the Visium family (Spatial Transcriptomics, Visium, Visium HD, and Visium HD 3′) and the Xenium family (Xenium and Xenium Prime 5K), and the results were averaged over five seeds. In the Visium family (489 slides), tissue type, disease state, and preservation method each explained little of the variation across slides (adjusted R² ≤ 0.03). In the Xenium family (57 slides), the three attributes each explained little (adjusted R² between −0.02 and 0.07). Since HEST-1k does not annotate cell composition, we used the CellViT nuclei segmentation and classification provided with HEST-1k to compute the nuclear density and the fractions of predicted nucleus types within the spot areas of each slide, as proxies for cell composition. On the 454 slides where CellViT detected nuclei in at least half of the spots, cell composition on its own explained little of the slide term (adjusted R² of −0.005 and 0.03 for the two families; permutation p = 0.74 and 0.24, median over seeds). By contrast, in both families the leading axis of the slide term co-varied with the measurement depth of each slide (total counts per target; |ρ| = 0.49 and 0.60, respectively), and in the Visium family this single axis accounted for 98% of the variance of the slide term.

We have also added this analysis to the appendix of the revised manuscript.

Revised manuscript: Appendix C.5.

Reviewer_gJ5UOpenReview ↗

New Official Comment

The author has responded to some of my comments; I am maintaining my positive score.

AuthorsOpenReview ↗

Thank you for your response

Thank you for taking the time to read our response and for maintaining your positive assessment. We sincerely appreciate your constructive comments, which helped us improve the manuscript.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/2)

We sincerely thank all reviewers for their constructive and thoughtful feedback, and for the time and effort they devoted to reviewing our manuscript. Their comments have been invaluable in improving the quality and completeness of this work.

We are encouraged that all reviewers found the paper clear and easy to follow, and that they recognized the following strengths:

  • An important problem and a useful perspective on it. Whether H&E-to-ST models benefit from data pooled across cohorts and platforms is a question of broad current interest, and our systematic analysis provides useful insights into the generalization of such models (HC9q). Modeling platform scales, slide variation, and count noise separately offers a useful perspective on why pooled training can yield inconsistent gains (gJ5U).
  • A simple and broadly applicable method. The observation model is simple and intuitive (HC9q). It addresses the prediction target rather than relying on a more complex architecture, and it can be plugged into any existing regressor without modification, which may benefit both existing and future virtual ST methods (gJ5U, HC9q, xaiV).
  • Comprehensive evaluation. The experiments cover the ten HEST benchmark tasks spanning multiple cancer types, organs, and platforms with five backbones (HC9q). Comparing objectives on the same training pools helps assess whether the gains extend beyond a particular network design (gJ5U). The external validation on the newly released Atera platform shows improvements of 10–28% across all five backbones on the two samples in the manuscript (xaiV).

Section numbering in the revised manuscript

We have uploaded a revised manuscript. Following the suggestion of Reviewer xaiV, we moved the Related Work section to right after the Introduction, and the extended related work to the beginning of the appendix. The analyses added in response to the reviews also introduce new appendix sections and tables. As a result, the section, appendix, and table numbers of the revised manuscript differ from those of the submitted manuscript.

In our responses, section, appendix, and table numbers refer to the submitted manuscript, which is the version the reviews were written on. For convenience, wherever a response refers to such a number, we give the corresponding location in the revised manuscript in a quoted line directly below the paragraph, as in the following example.

PCC, our primary metric, is computed across the spots within each slide (Section 5 and Appendix E.1).

Revised manuscript: Section 5 → Section 6; Appendix E.1 → Appendix F.1.

Tables R1–R7 are the tables of the responses, and Tables 1–3 keep their numbers in the revised manuscript.

We also adjusted the column alignment and spacing of the tables in the revised manuscript. These are formatting changes only and do not alter any reported value.

AuthorsOpenReview ↗

General Response to All Reviewers (2/2)

Extended external validation on Atera

As noted in the Limitations section, the external validation in the submitted manuscript covered two Atera samples (breast and cervical cancer). Since the submission, 10x Genomics has publicly released four additional Atera samples (brain, breast, colorectal, and kidney). We evaluated all six samples using the same pipeline and the same trained models as for the original two.

Table R1. External validation on six Atera samples: PCC over all genes, mean ± sd over five seeds. Breast and Cervix are the two samples in the manuscript, and their results are the same as in Table 2. Kidney, Brain, Colorectal, and Breast (FF, fresh frozen) are the newly released samples. MSE*: the Visium expert trained with MSE. MSE: pooled training with MSE. Obs.: pooled training with the observation model, with Atera treated as Xenium at inference, as in the manuscript. Bold: best per backbone and sample.

BackboneObjectiveBreastCervixKidneyBrainColorectalBreast (FF)
ProbeMSE*0.494±0.0230.341±0.0150.350±0.0610.262±0.0370.324±0.0360.378±0.040
MSE0.495±0.0560.404±0.0250.335±0.0550.311±0.0260.465±0.0380.522±0.039
Obs.0.635±0.0050.507±0.0040.510±0.0250.402±0.0160.578±0.0100.609±0.008
Probe+LoRAMSE*0.538±0.0130.389±0.0080.446±0.0170.298±0.0350.350±0.0160.390±0.043
MSE0.607±0.0230.477±0.0140.521±0.0130.337±0.0330.524±0.0200.530±0.051
Obs.0.669±0.0030.529±0.0050.566±0.0100.429±0.0110.593±0.0160.615±0.018
Simple-TFMSE*0.519±0.0380.363±0.0240.387±0.0260.246±0.0460.375±0.0300.474±0.042
MSE0.612±0.0140.474±0.0280.503±0.0290.344±0.0310.514±0.0190.554±0.025
Obs.0.693±0.0050.526±0.0080.545±0.0290.425±0.0240.573±0.0280.621±0.007
STPath†MSE*0.434±0.0620.274±0.0450.256±0.0340.125±0.0530.260±0.0320.269±0.069
MSE0.594±0.0180.426±0.0260.387±0.0820.262±0.0450.418±0.0360.498±0.050
Obs.0.667±0.0130.487±0.0110.513±0.0390.373±0.0390.531±0.0280.586±0.006
STORM†MSE*0.438±0.0390.277±0.0270.296±0.0650.264±0.0210.326±0.0610.310±0.105
MSE0.548±0.0170.399±0.0470.326±0.0470.283±0.0170.446±0.0140.527±0.012
Obs.0.626±0.0090.452±0.0020.473±0.0070.332±0.0130.502±0.0260.552±0.018

As shown in Table R1, the observation model achieves the highest PCC on every sample for all five backbones. Across the 30 combinations of sample and backbone, its improvement over pooled MSE training ranges from 4.8% to 52.1% (median 15.2%), and the median improvement on the four new samples alone is 17.4% (computed from the unrounded means). The main finding on the original two samples, that the observation model improves over pooled MSE training on every backbone, therefore also holds on the four new samples. This suggests that the benefit of the observation model is not confined to the HEST benchmark but extends to a platform absent from HEST-1k and to a broader range of tissues. We have included the results on all six samples in the revised manuscript.

Revised manuscript: Table 2 now reports the mean over the six samples, and the per-sample results of Table R1 are in Appendix H.3 (Tables 30–35). The Breast and Cervix values of Table 2 in the submitted manuscript are in Tables 30 and 31.

AuthorsOpenReview ↗

Summary of the Discussion

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for evaluating our submission and for taking part in the discussion. The reviewers' comments have substantially improved the manuscript.

All three reviewers rated the submission positively in their initial reviews, and this assessment held through the discussion. After our response, Reviewers gJ5U and xaiV maintained their positive scores, with Reviewer xaiV confirming that the comments had been addressed, and Reviewer HC9q had no further questions. Below, we summarize the main concerns raised and how we addressed them.

Main concerns and how we addressed them

  1. What the platform offset and the slide term capture (gJ5U, HC9q). A leave-one-study-out analysis over the 91 training studies shows that the offsets are stable across cohorts. On 16 pairs of same-tissue sections measured on two platforms, the offsets deviate from the directly measured ratios by a median factor of 1.6, compared with 7.0 without an offset. The slide term is explained little by tissue type, disease state, preservation method, or cell composition, and its leading axis co-varies with the measurement depth of each slide (Appendices C.3 and C.5).

  2. Offset coverage and the Atera offsets (HC9q). The pairs without an offset are concentrated on two platforms and on lowly expressed genes (Tables 9 and 10). Atera-specific offsets estimated from two samples reduce the MSE by 28–46% on the four samples not used for estimation (Table 36).

  3. Dataset scale and train–test separation (HC9q). In our response, we clarified the scale of the training set (546 slides, about 1.03M spots) and how the split separates the training pool from the test slides (Appendix D).

  4. Additional baselines (xaiV, Weakness 1 and Q1). We trained STNet, TRIPLEX, and TissueFormer on the pooled set with MSE and with the observation model. The observation model raises the ten-task mean PCC of all three (Tables R5 and R6 of our response).

  5. Statistical significance and metrics beyond PCC (xaiV, Q2 and Q3). Apart from spatial clustering on two backbones, the observation model improves over pooled MSE training in every evaluation and on every backbone, and 17 of these 23 improvements are statistically significant (Tables 22 and 37). MSE is reported for every backbone and task (Appendix F.2) and is consistent with the PCC results.

  6. Contextualization and novelty (xaiV, Weaknesses 1 and 2). We moved Related Work to Section 2, corrected the citation of Phoenix, and added the works the reviewer mentioned. We also clarified that our contribution is not the NB likelihood itself but the modeling of platform and slide effects in the training objective for pooled data.

  7. Results in the Xenium expert setting and single-cell resolution (xaiV, Weakness 3 and Q4). In our response, we clarified that the method is designed for training on data pooled across platforms, and discussed how its two components carry over to single-cell resolution.

In addition, with four samples released after the submission, we extended the external validation on Atera from two to six samples. The observation model achieves the highest PCC on every sample for all five backbones (Table 2; Tables 30–35).

Following Reviewer gJ5U's suggestion, we also replaced "GT" with "measured-expression reference" throughout. The analyses in items 1, 2, and 5 and the extended Atera validation are included in the revised manuscript.

We thank the Area Chair and the reviewers again for their time and constructive feedback.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.