acceptodds
Under review as a conference paper at ICLR 2027

ZoomST: Pretraining Spatial Transcriptomic Representations with Learnable Cross-Granularity Transformations

Abstract

Tissue function depends on spatial organization across scales, from cellular neighborhoods to anatomical regions. Expanding spatial context around a focal location yields related but distinct representations of tissue organization at different granularities. Although existing spatial transcriptomics pretraining models incorporate multiscale context for representation learning, they do not explicitly learn how representations change as spatial context expands. We introduce ZoomST, a self-supervised framework that jointly learns granularity-specific, cross-scale-compatible representations and transformations among them. ZoomST combines molecular-spatial dual-view alignment within each granularity with a soft distributional equivariance constraint across granularities. Across held-out mouse sections spanning all sampled developmental stages, frozen ZoomST representations outperform the evaluated pretrained baselines and remain competitive with target-fitted methods for tissue mapping. Analyses in mouse and human developmental atlases further illustrate the complementary information captured by fine and coarse spatial context. Continuous transformations learned only from discrete training granularities predict representations at unseen intermediate spatial scales and reveal relative differences in context sensitivity across tissue locations. By making spatial granularity adjustable after pretraining, ZoomST turns changes in spatial context into both a source of self-supervision and an axis for tissue analysis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Self-supervised learning, Multiscale representation learning, Cross-scale learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9oUcOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Zoom ST, a framework to learn spatial-transcriptomic representations from gene expression and coordinates (spatial component) using nested neighborhoods and neural ODE to predict representations at intermediate sizes not seen during training. It tackles a relevant problem with impact for the community; how spatial transcriptomic representations change as tissue context expands. The authors performed the evaluation on held-out mouse and human tissue sections showing promissing results; The novelty of this work consists of providing an axis for analyzing tissue-context dependence, for that they treat the spatial granularities as a lernable component of spatial transcriptomics pre-training.

Critical strengths

• The paper tackles a relevant problem for the scientific community by modeling and proposing a framework to understand how spatial transcriptomic representations changes as tissue context expands, instead of treating each scale independently. This gives an axis for analyzing tissue-context dependence.

• Using a neural ODE, to query intermediate neighborhood sizes is interesting and potentially useful, however I would expect stronger validation against a simple interpolation using conventional methods.

• The evaluation protocol regarding data broad and coverage is reasonable well-defined, the use of mouse and human samples gives a stronger support in terms of generalizability.

• The method (ZoomST) is conceptually coherent, and the paper is technically detailed and reproducible. The supplementary material provides detailed descriptions to support paper reproducibility.

Critical weaknesses

• The method has some limitations in terms of comparability across sections and sampling densities. Since granularity is not a precise state the neighborhoods are defined by 8, 32 and 128 nearest observations (p.3, lines 156 - 161), which depends on tissue sampling.

• The neural ODE as far as I understood is supervised at only three granularities, which leaves many possible trajectories between the observed scales. Comparisons with classical approaches like linear interpolation approaches are missing.

• The fact that the intermediate states are evaluated using the same encoder demonstrates internal consistency but l it would be great to see Molecular and Spatial encoders using different approaches.

• The context sensitivity metrics is weakly validated since the score measures the latent representation displacement and not a directly observed biological quantity. The majority of biological examples come from the training sections, and the correlation with tissue composition change is low, as measured by the whole-section Spearman correlation.

• The generalization claim should be more conservative, the experiments demonstrate within-atlas generalization to held-out sections, not cross-atlas transfer. Moreover, the same improvements are marginal, (e.g. On MOSTA mean ARI is 0.345 for ZoomST, compared with 0.343 for STAGATE and 0.342 for BANKSY - Table C2).

Questions for authors
  1. The benefit of the neural ODE over simpler interpolation remains unclear, does it outperform simpler alternatives? Baselines are missing, compare against linear interpolation or splines, using the same encoder and evaluation protocol (training splits).

  2. The biological relevance of the intermediate representationsis is only partially established. It would be great to see the impact on an independent task, such as tissue annotation or boundary detection, or any other relevant task.

  3. Neighbor count may not represent a comparable spatial scale across tissues. Is granularity comparable across tissues? It would be great to see the physical radius covered by each neighbor count, and a comparison between count-based and fixed-radius neighborhoods. This can give us more details about the dependecy on how densely the tissue was sampled.

  4. What is the practical advantage of querying an intermediate scale instead of directly encoding the corresponding neighborhood? Would be useful a comparison of computational cost and downstream performance to clarify whether the ZoomST provides practical value beyond interpolation, also to understand the tradeoff.

10 comments
AuthorsOpenReview ↗

Response to 9oUc (1/4): W2, W3 & Q1 - Intermediate-state validation and interpolation comparisons

Intermediate-state validation and interpolation comparisons — Weaknesses 2–3 and Question 1

Summary of concerns. (1) encoder design and validation beyond encoder-derived targets; (2) comparison with conventional interpolation under matched conditions.

Response

We sincerely thank the reviewer for their thoughtful assessment and constructive suggestions, and for recognizing the value of studying how representations change as tissue context expands. Following these suggestions, we have clarified the roles of the molecular and spatial encoders and expanded the developmental-age evaluation to compare the neural ODE with conventional interpolation methods under matched conditions.

1. Dual-view design and downstream validation (Weakness 3).

The molecular view encodes explicitly aggregated neighborhood expression, providing a direct basis for learning cross-granularity transformations. The spatial view combines spot-level expression and relative spatial information to capture granularity-specific tissue structure. Both views are jointly learned through within-granularity VICReg alignment and expression reconstruction.

To assess the usefulness of intermediate representations beyond agreement with encoder-derived targets, we evaluated developmental-age prediction on held-out MOSTA sections using observed embryonic ages as labels.

2. Comparison with conventional interpolation (Weakness 2 and Question 1).

Using the same encoder and training/evaluation splits, we compared the neural ODE with linear interpolation, quadratic spline interpolation, and natural cubic spline interpolation. The interpolation methods used representations at the trained granularities (r=8,32,128r=8,32,128) as anchors. All methods were evaluated at the same six unseen intermediate granularities using separately fitted ridge readouts, with matched training samples, regression settings, and held-out evaluation locations.

The neural ODE achieved lower developmental-age prediction MAE than all three interpolation methods at every unseen granularity. Compared with linear interpolation, which had the lowest mean MAE among the interpolation baselines, the ODE reduced mean MAE from 0.9165 to 0.8565 days, a 6.5% reduction (Table R8.5).

Developmental-age prediction at unseen granularities.

This table reproduces the interpolation comparisons from Table R8.5 (Appendix H.8, p. 53).

Values are MAE in embryonic days; lower is better. Each method and target granularity uses a separately fitted ridge readout. Interpolation uses the same log-granularity coordinate as the ODE, u(r)=log⁡2(r/8)/4u(r)=\log_2(r/8)/4.

Target granularityNeural ODELinear interpolationQuadratic splineNatural cubic spline
120.87600.92110.92660.9236
160.87020.89750.91920.9163
240.86200.89200.90370.9032
480.84930.91740.91430.9146
640.84440.93000.92130.9240
960.83710.94080.92610.9285
Mean0.85650.91650.91850.9184
AuthorsOpenReview ↗

Response to 9oUc (2/4): Q2, Q4 & W5 - Intermediate-scale utility and generalization

Intermediate-scale utility and generalization — Questions 2 and 4; Weakness 5

Summary of concerns. (1) scope of generalization and interpretation of comparative performance; (2) downstream usefulness and computational cost of intermediate-scale predictions.

Response

Thank you for these thoughtful questions and constructive suggestions. We agree that our experiments support within-atlas generalization to held-out sections and that the MOSTA ARI results should be described as comparable to STAGATE and BANKSY. We will revise the corresponding claims accordingly.

Following your suggestions, we have further evaluated the downstream usefulness and computational efficiency of intermediate-scale predictions. The additional results provide evidence for their practical value in the following two aspects.

1. Predicted intermediate-scale representations support tissue annotation.

We evaluated ODE-predicted representations at six intermediate granularities not used during training, using the same seven held-out HESTA sections, coarse anatomical labels, and evaluation protocol as Table 2. Mean AMI ranged from 0.5295 to 0.5346, exceeding the comparator means reported in Table R7.1. These results support the usefulness of intermediate-scale predictions for an independent anatomical task.

Table R7.1. Tissue-annotation performance of ZoomST at intermediate granularities on HESTA.

AMI (higher is better) on the same seven held-out HESTA sections and coarse anatomical labels as Table 2. ZoomST uses molecular representations predicted from r = 8 by its frozen neural ODE. Other methods retain their Table 2 values. Entries average clustering seeds 42, 43, and 44; Mean weights sections equally. Bold indicates the best score in each column.

MethodCS12-13CS14-15CS17CS18CS19CS20CS23Mean
ZoomST (intermediate r = 12)0.51680.52380.49970.54310.53240.56010.53740.5305
ZoomST (intermediate r = 16)0.53640.53730.50250.53850.52720.56530.53530.5346
ZoomST (intermediate r = 24)0.53000.53360.49920.54880.51800.56560.54670.5345
ZoomST (intermediate r = 48)0.51900.53050.49410.53510.53560.56140.53750.5305
ZoomST (intermediate r = 64)0.53060.52920.49150.54430.52160.56000.53270.5300
ZoomST (intermediate r = 96)0.51800.52920.49470.53930.52460.56160.53900.5295
Novae0.33830.33070.37620.28280.38170.41920.42300.3646
SToFM0.17050.20520.27030.17490.28900.25730.31980.2410
scGPT-spatial0.34220.38410.41640.36560.44730.43830.49000.4120
TERRA Cell0.22820.27890.35830.26730.38230.33760.39270.3208
TERRA Nbh.0.32680.41030.41690.38920.46390.43780.47770.4175
TERRA S-cell0.31780.41600.41650.41080.47120.45720.48600.4251
CellWorld0.28470.22760.32250.25210.28930.32140.37640.2963
DECIPHER Ctr.0.40220.41590.47590.33060.48310.46470.46870.4344
DECIPHER Ctx.0.47490.52330.41350.45510.34340.42360.42150.4365

2. Additional scales can be queried efficiently.

ZoomST applies its pretrained model directly to held-out sections, avoiding the section-specific training used by methods such as GraphST and STAGATE. Once the r=8r=8 representations are cached, the frozen ODE queries additional scales without repeated neighborhood construction and encoding.

Averaged over the seven held-out HESTA sections and the same six intermediate granularities, querying required 0.092 seconds per scale per section, compared with 77.49 seconds for direct neighborhood construction and encoding (Table R7.2). This low incremental cost supports repeated multiscale analysis as spatial atlases grow.

Table R7.2. Runtime summary on held-out HESTA sections.

MethodMean time (s)
ZoomST(query scales)0.092
Pretrain model
ZoomST direct encoding77.49
Novae872.83
SToFM1,405.61
TERRA2,642.63
scGPT-spatial603.52
CellWorld30.85
Target-fitted models
GraphST684.53
STAGATE265.12
DECIPHER314.28
BANKSY199.02

Legend. Values are mean times across seven equally weighted held-out HESTA sections (n = 7). Bold indicates the lowest mean runtime. ZoomST(query scales) queries from e8 which need not encoding new neighborhoods. Shared raw-input preparation is excluded. ZoomST direct encoding includes neighborhood pooling, tokenization and encoding. Pretrained-model timings include input encoding and inference; GraphST, STAGATE and DECIPHER timings include fitting and inference. GPU runs primarily used NVIDIA L20 48GB; BANKSY measures CPU feature construction. Following L20 memory limitations, STAGATE (1 section) and DECIPHER (4 sections) used NVIDIA A800 80GB

Added evidence: Revised manuscript, Introduction, p. 2 (generalization scope); Appendix H.7, p. 50 (Tables R7.1–R7.2).

AuthorsOpenReview ↗

Response to 9oUc (3/4): W4 - Biological interpretation of context sensitivity

Biological interpretation of context sensitivity — Weakness 4

Summary of concerns. Biological interpretation of context-sensitivity scores and validation on held-out tissue.

Response

We appreciate your suggestion to strengthen the biological validation of context sensitivity. The score measures how strongly a location’s representation changes as the surrounding tissue context expands, and we assess its biological relevance through correspondence with anatomical organization.

The original whole-section comparison combines responses averaged over 8–128 with composition changes based on coarse tissue-region labels (Fig. 5b–c). These summaries can obscure differences among finer structures and between scale intervals. Following your suggestion, we therefore examined a held-out HESTA section using finer anatomical annotations from the original atlas (Fig. R6.1A–B), separating responses over 8–32 and 32–128.

1. Responses distinguish local anatomical compartments.

Over 8–32, the choroid plexus epithelium and its surroundings are distinguished in both the spatial response map and the compartment-specific density curves (Figs. R6.1C and R6.2A). Epithelial responses are lower and more concentrated, whereas surrounding responses are higher and more broadly distributed, with clearly separated density peaks. This contrast corresponds to the annotated spatial organization: epithelial spots cluster together, whereas erythrocyte-labelled spots are interspersed among other tissue components in the surrounding compartment (Fig. R6.1B).

2. The response contrast depends on the scale interval.

Over 32–128, responses in both compartments are lower and more concentrated, and their density peaks are less clearly separated (Figs. R6.1D and R6.2B). The weaker separation indicates that 8–32 is more informative for resolving this local anatomical contrast. Examining responses over separate intervals therefore helps identify the observation ranges in which local tissue organization is most apparent, illustrating the value of following representation changes across scales.

Added evidence: Appendix H.6, p. 49 (Figures R6.1–R6.2).

Reviewer_6AXtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ZoomST is a self-supervised framework that jointly learns granularity-specific, cross-scale-compatible representations and transformations among them. It combines molecular–spatial dual-view alignment within each granularity with a soft distributional equivariance constraint across granularities. The framework is evaluated on human and mouse developmental atlases. Frozen representations generalize to held-out tissue sections and support tissue mapping and developmental organization. Continuous transformations learned only from discrete training granularities predict representations at unseen intermediate spatial scales and reveal relative differences in context sensitivity across tissue locations.

Critical strengths
  1. Cross-granularity transformation learning is well motivated. ZoomST jointly learns granularity-specific representations and transformations among them. Making spatial granularity adjustable after pretraining provides a useful way to examine tissue-context sensitivity.
  2. The tissue-mapping results and objective ablations are convincing. Frozen ZoomST representations outperform the evaluated pretrained baselines and remain competitive with target-fitted methods. On MOSTA, No Cross, CS, and ZoomST use the same initialization, training data, and training budget, differing only in their loss terms. These comparisons support the contribution of cross-granularity supervision.
Critical weaknesses
  1. The benefit of continuous cross-granularity transformations needs stronger validation. Predictions are mainly evaluated against directly encoded states from the jointly trained encoder. The comparison with a granularity-conditioned residual MLP is helpful, but interpolation between predictions at trained granularities is not evaluated. The shared age readout also makes it difficult to separate improved developmental-stage prediction from better agreement with the representations used to fit that readout.

  2. The developmental interpretation is limited by the evaluation split. The split holds out sections without establishing separation by embryo, and each evaluated developmental stage is represented by a single section. DPT–stage correlation and developmental-stage prediction may therefore be influenced by specimen or section differences. This limits the evidence that the representations capture developmental organization beyond these effects.

Questions for authors
  1. Do held-out sections come from embryos that also contribute pretraining sections? How might this affect developmental ordering and developmental-stage prediction?
  2. How does the neural ODE compare with interpolation between predictions at trained granularities, using the same source representation? Does the improvement in developmental-stage prediction remain when separate readouts are fitted to unchanged source states and predicted states?
  3. Are the main improvements stable across independent pretraining runs? The reported downstream clustering seeds do not capture variability in representation learning.
3 comments
AuthorsOpenReview ↗

Response to 6AXt (1/2): W2, Q1 & Q3 - Embryo-level independence and pretraining stability

Embryo-level independence and pretraining stability — Weakness 2 and Questions 1 and 3

Summary of concerns. (1) embryo-level independence and the interpretation of developmental results; (2) stability across independent pretraining runs.

Response

We sincerely thank the reviewer for the thoughtful feedback and positive assessment. Following your comments, we have clarified the original evaluation split, repeated the developmental evaluation with all test embryos excluded from pretraining, and assessed stability across independent pretraining runs.

1. Embryo-level independence of the developmental evaluation.

(a) Clarification of the original split.

The original split retained all eight developmental stages in pretraining while reserving one section per stage for evaluation. At three stages—E10.5, E12.5, and E15.5—the test embryos were entirely excluded from pretraining. At the other five stages, additional sections from the same embryos were retained in pretraining. This choice balanced embryo-level independence with developmental-stage coverage, particularly because E11.5 and E13.5 were each represented by only one embryo in the available atlas. Accordingly, the original evaluation primarily supports generalization to held-out sections within MOSTA, with embryo-level holdouts at three stages.

(b) Developmental evaluation after excluding all test embryos from pretraining.

Following your suggestion, we excluded all sections belonging to the eight original test embryos and retrained ZoomST from scratch using the remaining 23 sections. The resulting evaluation cohort comprised 30 sections from eight embryos entirely unseen during pretraining, spanning all eight developmental stages.

For comparison with the original developmental-ordering results, we first evaluated the newly pretrained model on the original eight held-out sections. ZoomST achieved a DPT–stage Spearman correlation of 0.797, close to the original 0.816 and higher than all six evaluated external methods (Table R8.1).

Table R8.1. Developmental ordering on the original eight held-out MOSTA sections.

Brain DPT–stage Spearman correlations. The additional ZoomST model was pretrained after excluding all sections from the eight test embryos.

MethodSpearman ρ\rho
ZoomST0.816
ZoomST (test embryos excluded)0.797
Novae0.714
SToFM0.341
GraphST0.491
STAGATE0.229
BANKSY-0.267
MENDER0.458

Under the same embryo-excluded pretraining split, ZoomST also achieved lower developmental-stage prediction MAE than the matched No Cross model at all three training granularities: 0.731 versus 0.935 days at r=8r=8, 0.778 versus 0.859 at r=32r=32, and 0.847 versus 0.859 at r=128r=128 (Table R8.2). These results support the preservation of developmental ordering and the stage-prediction benefit of cross-granularity supervision when all test embryos are excluded from pretraining.

Table R8.2. Developmental-stage prediction with all test embryos excluded from pretraining.

Stage MAE in embryonic days, following the evaluation protocol in the main text. Lower is better.

Modelr=8r=8r=32r=32r=128r=128
ZoomST (test embryos excluded)0.7310.7780.847
No Cross (test embryos excluded)0.9350.8590.859

Added evidence: Appendix H.8, p. 51 (Tables R8.1–R8.2).

AuthorsOpenReview ↗

(Continued) Response to 6AXt (1/2): W2, Q1 & Q3 - Embryo-level independence and pretraining stability

(Continued) 2. Stability in the expanded evaluation cohort and across independent pretraining runs.

(a) Developmental ordering on the expanded evaluation cohort.

To assess both section coverage and pretraining variability, we extended the DPT evaluation to all 30 sections from the eight held-out embryos and evaluated three independently pretrained ZoomST models (using different random seeds) under the same embryo-excluded split. Using the same brain spots, the three runs achieved DPT–stage Spearman correlations of 0.725, 0.716, and 0.716 (mean 0.719), consistently exceeding all evaluated external methods (Table R8.3). The developmental-ordering signal therefore persists in the expanded evaluation cohort and remains consistent across the three independent pretraining runs.

Table R8.3. Developmental ordering across 30 held-out MOSTA sections from eight embryos spanning eight stages.

DPT–stage Spearman correlations using the same brain spots. The evaluation cohort contains one held-out embryo per stage. The three ZoomST values correspond to independent pretraining runs under the embryo-excluded split.

MethodSpearman ρ\rho
ZoomST (test embryos excluded), run 10.725
ZoomST (test embryos excluded), run 20.716
ZoomST (test embryos excluded), run 30.716
Novae0.616
GraphST0.481
SToFM-0.278
STAGATE0.086
BANKSY0.082
MENDER0.017

(b) Tissue-mapping performance across pretraining runs.

We also evaluated the three independently pretrained models on the original eight held-out sections using the annotation-mapping protocol of Table 1. Their AMI scores were 0.574, 0.573, and 0.573, showing consistent performance across runs and remaining close to the original model’s score of 0.577. All three runs remained competitive with the evaluated methods, including BANKSY (0.567), STAGATE (0.564), and MENDER (0.553) (Table R8.4).

Table R8.4. Annotation mapping across independent pretraining runs with all test embryos excluded.

AMI on the original eight held-out MOSTA sections, following the evaluation protocol of Table 1.

MethodAnnotation AMI ↑
ZoomST (original split)0.577
ZoomST (test embryos excluded), run 10.574
ZoomST (test embryos excluded), run 20.573
ZoomST (test embryos excluded), run 30.573
BANKSY0.567
STAGATE0.564
MENDER0.553
GraphST0.468
Novae0.435
SToFM0.319

Added evidence: Appendix H.8, p. 52 (Tables R8.3–R8.4).

AuthorsOpenReview ↗

Response to 6AXt (2/2): W1 & Q2 - Separate-readout evaluation and interpolation comparisons

Separate-readout evaluation and interpolation comparisons — Weakness 1 and Question 2

Summary of concerns. (1) developmental-stage prediction with separately fitted readouts; (2) interpolation comparisons using predictions from the same source representation.

Response

Thank you for your careful reading and constructive suggestions. Following your recommendations, we have evaluated developmental-stage prediction with separately fitted readouts and compared the neural ODE with conventional interpolation methods under the same protocol.

1. Developmental-stage prediction with separately fitted readouts.

To test whether the improvement persists with separately fitted readouts, we fitted one ridge regressor to the unchanged source states (e8e_8) and a separate regressor to the ODE-predicted states at each target granularity. We kept the training locations, age labels, held-out evaluation locations, and preprocessing and regression settings identical across representations. Preprocessing parameters and regression coefficients were estimated independently for each representation using only the readout-training data, while the encoder and ODE remained frozen.

Under this protocol, ODE-predicted states achieved lower MAE than unchanged e8e_8 at all six unseen granularities. Mean MAE decreased from 0.9432 to 0.8565 days, a 9.2% reduction relative to unchanged e8e_8 (Table R8.5). The prediction advantage therefore persists when each representation is evaluated with its own fitted readout.

2. Interpolation comparisons under the same readout protocol.

Using the same separate-readout protocol, we then compared the neural ODE with linear interpolation, quadratic spline interpolation, and natural cubic spline interpolation. The interpolation methods used representations at the trained granularities (r=8,32,128r=8,32,128) as anchors. Interpolation was performed in the same log-granularity coordinate used by the ODE, u(r)=log⁡2(r/8)/4u(r)=\log_2(r/8)/4, with each method’s intermediate representations evaluated under the matched conditions described above.

The neural ODE achieved lower MAE than all three interpolation methods at every unseen granularity. Compared with linear interpolation, which had the lowest mean MAE among the interpolation baselines, the ODE reduced mean MAE from 0.9165 to 0.8565 days, a 6.5% reduction (Table R8.5).

Table R8.5. Developmental-stage prediction at unseen granularities.

Values are MAE in embryonic days; lower is better. Each representation was evaluated using a separately fitted ridge readout.

Target granularityUnchanged source (e8e_8)Neural ODELinear interpolationQuadratic splineNatural cubic spline
120.94320.87600.92110.92660.9236
160.94320.87020.89750.91920.9163
240.94320.86200.89200.90370.9032
480.94320.84930.91740.91430.9146
640.94320.84440.93000.92130.9240
960.94320.83710.94080.92610.9285
Mean0.94320.85650.91650.91850.9184

Added evidence: Appendix H.8, p. 53 (Table R8.5).

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

This paper introduces ZoomST, a self-supervised framework for learning spatial transcriptomics representations across different neighborhood sizes. The method learns molecular and spatial representations at three granularities (r = 8, 32, and 128 neighbors) and aligns them using VICReg. A neural ODE is then used to learn continuous transformations between molecular representations at different granularities, guided by coarse-support correspondence and distributional consistency losses. The authors evaluate ZoomST on MOSTA and HESTA atlases, demonstrating its performance in 4 complementary tasks.

Critical strengths
  1. The paper clearly explains the dual-view framework and the constraints on embedding transformations across granularities, making the model architecture and methodological details easy to follow.
  2. The paper includes ablation studies, evaluates transformations at unseen neighborhood sizes, and examines whether predicted representation changes reflect changes in surrounding tissue composition.
Critical weaknesses
  1. The improvement over strong task-specific methods is not convincingly demonstrated. For example, in Table 1 and Table C2, ZoomST shows only marginal improvements over BANKSY and STAGATE on MOSTA especially on ARI. Figure C2 also shows very similar spatial maps of recovered annotation labels. These results leave the practical advantage of the dual-view design over existing methods insufficiently established. The choice of baseline methods also differs across datasets without sufficient justification. TERRA and CellWorld are evaluated only on HESTA, whereas GraphST, STAGATE, and BANKSY are evaluated only on MOSTA.
  2. There is a issue of over-smoothing and loss of local heterogeneity. Figure C1 shows substantially worse recovery of smooth muscle at granularity 128 than at 8, also Figure C2 displayed more smoothed regions than the true annotation. Moreover, the coarse-support loss in Equation (3) encourages fine-scale representations within a neighborhood to approach the same coarse representation, potentially suppressing local heterogeneity.
Questions for authors
  1. Could the authors further demonstrate how the dual-view design improves recovery of annotation maps, including comparisons of misclassification rates for individual anatomical regions against task-specific methods? The authors should also provide visualizations of the recovered labels for HESTA.
  2. In Section 3.3, why are cross-granularity transformations applied only to molecular representations? Could applying them to spatial representations provide additional benefits?
  3. Why is unweighted average expression used to construct the molecular view? Could the authors compare it with distance-weighted aggregation and evaluate whether the spatial encoder provides additional information beyond this alternative?
  4. Could the authors justify the choice of (r={8,32,128}) in Section 3.1 and evaluate sensitivity to alternative neighborhood sizes? How do these sizes correspond to biologically meaningful contexts? This is important as different technologies represent different spatial resolutions.
  5. The coarse-support loss encourages neighboring locations to have similar representations, potentially blurring tissue boundaries and local molecular differences. Although the authors compare ZoomST with existing methods for anatomical clustering, a more focused evaluation of boundary preservation against the No Cross model and methods such as BANKSY and STAGATE would help determine whether cross-granularity learning compromises fine-scale tissue structures.
  6. The benefit of continuous transformation learning for the main downstream tasks remains unclear since the main tissue-mapping results rely on concatenated embeddings across three granularities, rather than representation generated by the learned transformations. Could the authors further justify this?
  7. Figure 5b and 5c show different spatial patterns, with a weak whole-section correlation value. Moreover, the changes in learned representation across spatial granularities does not necessarily indicate a biologically meaningful change in tissue spatial context. The authors should convincingly establish the biological relevance of the proposed context-sensitivity scores.
7 comments
AuthorsOpenReview ↗

Response to iYVA (1/6): W1 & Q1 - Baseline comparisons and dual-view evaluation

Baseline comparisons and dual-view evaluation — Weakness 1 and Question 1

Summary of concerns. (1) clarification of the criteria for baseline selection and expanded comparisons; (2) additional region-level evaluations to further assess the benefit of the dual-view design.

Response

We sincerely thank the reviewer for their careful reading and constructive suggestions, and for recognizing the clarity of our framework. Following these suggestions, we have clarified the rationale for baseline selection, expanded the comparisons, and added region-level evaluations and HESTA label maps to assess the contribution of the dual-view design more directly.

1. Baseline pretraining model selection. TERRA and CellWorld were pretrained on human spatial-transcriptomics data. We therefore kept the comparison within species and evaluated them only on the human HESTA atlas.

2. Additional HESTA comparisons. We have additionally evaluated task-specific methods GraphST, STAGATE, and BANKSY on all seven HESTA sections using the same evaluation protocol. As shown in Table R1.1, ZoomST Joint achieves higher mean AMI and spatial-neighbor retention (R32) than these baselines, while its ARI is comparable to BANKSY.

Table R1.1. Mean performance across seven HESTA sections. Higher is better for all metrics.

MethodAMIARIR32
ZoomST Joint0.55760.32050.4953
BANKSY0.52900.31990.4341
STAGATE0.49820.31190.2702
GraphST0.40690.24100.2251

We also interpret the small ARI differences on MOSTA as comparable performance on this metric. Unlike STAGATE and GraphST, which fit their models separately to each target section, ZoomST directly applies its pretrained model without additional training. The region-level comparisons below provide a more direct assessment of the contribution of the two views to anatomical recovery.

3. Anatomical recovery (Question 1). Following your suggestion, we compared ZoomST Joint with each individual view and the evaluated baselines using region-level FNR, IoU, F1 score, and precision, covering 43 anatomical regions in MOSTA and 38 in HESTA. For each region, scores were first averaged across the held-out sections containing that region. Each metric was then min–max normalized across methods within that region, with FNR reverse-scaled so that 1 and 0 consistently denote the best and worst performance, respectively. Tables R1.2 and R1.3 report the median normalized scores across regions.

Table R1.2. Median normalized regional scores on MOSTA (43 regions).

MethodF1PrecisionReverse-scaled FNRIoU
ZoomST Joint0.74910.64880.81840.7549
ZoomST Spatial0.68610.65160.65930.6511
ZoomST Molecular0.62480.57010.56590.5998
STAGATE0.62010.57860.52810.6177
BANKSY0.61580.57790.52860.6534
MENDER0.55630.52490.51510.5324
Novae0.38690.37300.37110.3179
GraphST0.28740.22840.33370.2865
SToFM0.00000.01230.04290.0000

Table R1.3. Median normalized regional scores on HESTA (38 regions).

MethodF1PrecisionReverse-scaled FNRIoU
ZoomST Joint0.78140.74500.75480.7402
ZoomST Spatial0.74020.69350.68680.7249
ZoomST Molecular0.66060.61770.73270.6276
BANKSY0.59030.56980.60910.5668
STAGATE0.49560.46190.42970.4339
TERRA Spatial-cell0.39570.33230.38820.3278
GraphST0.36350.32280.33960.3167
TERRA Neighborhood0.35490.31880.31500.2839
CellWorld0.08320.06760.08090.0525
TERRA Cell0.05220.08770.06050.0519

Higher is better in all columns; bold marks the highest median.

ZoomST Joint achieves the highest median normalized F1, IoU, and reverse-scaled FNR scores on both datasets. These results demonstrate the advantage of combining the two views for region-level anatomical recovery, complementing the whole-section comparisons.

Following your suggestion, we have also added recovered-label maps for all seven held-out HESTA sections, using the fine organ and substructure annotations provided by the original atlas as anatomical references (Figures R1.1–R1.7).

Added evidence: We have added the HESTA results for GraphST, STAGATE, and BANKSY to Table 2 (lines 325–338). Additional comparisons and visualizations are provided in Appendix H.1, pp. 36-44 (Tables R1.1–R1.3 and Figures R1.1–R1.7).

AuthorsOpenReview ↗

Response to iYVA (2/6): Q2 - Cross-granularity transformation design

Cross-granularity transformation design — Question 2

Reviewer comment.

In Section 3.3, why are cross-granularity transformations applied only to molecular representations? Could applying them to spatial representations provide additional benefits?

Response

Thank you for raising this question and suggesting an extension to the spatial view. We explain the design rationale below and provide an additional ablation to assess its potential benefits.

1. Rationale for molecular-only transformation supervision.

We apply cross-granularity transformations to molecular representations because they are derived from explicitly aggregated expression statistics. As the neighborhood expands, the underlying aggregation changes in a well-defined manner, providing a direct basis for learning relationships between molecular representations across granularities.

In contrast, the spatial view first encodes individual spots and then integrates their representations through spatial attention and set aggregation. By incorporating relative spatial information while retaining location-specific expression differences, it captures richer spatial heterogeneity within each neighborhood. We therefore leave this branch free from an additional cross-granularity transformation constraint, preserving its flexibility for representing spatial patterns specific to each granularity.

The two views are jointly learned through within-granularity VICReg alignment and expression reconstruction. In that way, we kept cross-granularity supervision only for molecular view by design, and the spatial views provide complementary structural information.

2. Additional ablation: applying transformations to both views.

Following the reviewer’s suggestion, we evaluated a variant of ZoomST which applies cross-granularity transformations and the corresponding losses to both views (denoted as M+S). Across eight held-out MOSTA sections, this variant improved molecular-view AMI but reduced spatial-view AMI from 0.5744 to 0.5499 and joint-representation AMI from 0.5772 to 0.5625 (Table R2). These results support retaining molecular-only transformation supervision for joint-representation tissue mapping in the current setting.

Table R2. AMI for ZoomST and M+S.

RepresentationZoomST AMI ↑M+S AMI ↑
Molecular (e)0.5550.565
Spatial (g)0.5740.550
Joint (e+g)0.5770.563

All representations concatenate all scales. M+S uses the same training budget as ZoomST. Results are equally weighted means across eight held-out MOSTA sections; AMI uses three clustering seeds. Values are rounded to three decimal places; bold indicates the higher result within each view and metric.

Added evidence: Appendix H.2, p. 45 (Table R2).

AuthorsOpenReview ↗

Response to iYVA (3/6): Q3 - Molecular aggregation and spatial-view contribution

Molecular aggregation and spatial-view contribution — Question 3

Summary of concerns. (1) rationale for unweighted molecular aggregation; (2) comparison with distance-weighted aggregation; (3) additional value of the spatial view beyond weighted molecular aggregation.

Response

Thank you for this helpful suggestion.

1. Rationale for unweighted molecular aggregation.

The molecular view is intended to describe molecular composition at different observation granularities through simple aggregation over the corresponding neighborhoods. Conceptually, this resembles pooling molecular measurements over tissue regions of different sizes, as in bulk or coarser-resolution spatial measurements. We use unweighted averaging to obtain this neighborhood-level summary without imposing a predefined distance-dependent preference on individual locations.

Spatial information is explicitly considered in the spatial view branch, which encodes individual spots before integrating their representations through spatial attention and set aggregation. By incorporating relative distances and directions while retaining location-specific expression differences, this branch is designed to capture spatial organization and heterogeneity beyond the aggregate molecular summary.

2. Comparison with distance-weighted aggregation.

We agree that distance-weighted aggregation is a useful alternative to evaluate. We therefore tested distance-weighted averaging for molecular aggregation, keeping the remaining architecture and training budget unchanged (Table R3). We use Gaussian weights defined as

wij=exp⁡[−dij2/(2hi2)]∑k∈Nr(i)exp⁡[−dik2/(2hi2)],w_{ij} = \frac{\exp[-d_{ij}^{2}/(2h_i^{2})]}{\sum_{k\in\mathcal{N}_r(i)}\exp[-d_{ik}^{2}/(2h_i^{2})]},

where dijd_{ij} is the spatial distance to the center and hih_i is the median strictly positive distance within each center-inclusive neighborhood at each granularity. All-coincident neighborhoods use uniform weights.

Distance weighting did not improve tissue-mapping performance, supporting unweighted averaging as an effective choice.

3. Additional value of the spatial view.

Within the weighted variant, joint representations outperformed the molecular readout across the reported metrics. These results show that spatial representations provide additional downstream value even when molecular aggregation already incorporates distance information.

Table R3. Tissue mapping with unweighted and distance-weighted molecular aggregation.

RepresentationAMI ↑ARI ↑IoU ↑
ZoomST Joint0.5770.3450.327
Weighted-ZoomST Joint0.5750.3410.321
Weighted-ZoomST Molecular0.5570.3290.308

Mean scores across eight held-out MOSTA sections and three shared clustering seeds (42, 43, 44), using concatenated representations at granularities 8, 32, and 128. Bold indicates the best score.

Added evidence: Appendix H.3, p. 45 (Table R3 and the weighting definition).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Brief Summary of Revisions

We sincerely thank the reviewers and the Area Chair for their careful assessment and constructive feedback. We appreciate the reviewers' recognition of the clarity of the dual-view framework (iYVA), the motivation for cross-granularity learning and the convincing tissue-mapping and ablation results (6AXt), and the relevance of studying tissue-context dependence and the conceptual coherence of the method (9oUc). Following their constructive suggestions, we have substantially expanded the evaluation of ZoomST, particularly its cross-granularity learning capabilities, practical utility, and biological interpretation. The major changes are summarized as follows:

1. Continuous transformations and their practical utility. We extended the evaluation of ODE-predicted representations to tissue annotation at six unseen intermediate granularities on held-out HESTA sections, and assessed the computational cost of querying additional scales compared with direct neighborhood encoding. We also compared the learned transformations against linear and spline interpolation through developmental-stage prediction. (Appendices H.7–H.8; responses to Reviewers iYVA, 6AXt, and 9oUc.)

2. Anatomical recovery and preservation of local structures. We expanded the comparisons with task-specific methods on HESTA, added region-level evaluations and anatomical label maps, and conducted focused fine-structure and boundary-preservation analyses against No Cross and other baselines. These evaluations assess how the dual-view design contributes to anatomical recovery and whether cross-granularity supervision preserves fine anatomical structures and tissue boundaries. (Table 2; Appendices H.1 and H.5; response to Reviewer iYVA.)

3. Biological interpretation of context sensitivity. We extended the context-sensitivity analysis to a held-out HESTA section using finer anatomical annotations from the original atlas. By examining responses over different granularity intervals, we investigated their correspondence with local anatomical structures and the dependence of these patterns on the observation range. (Appendix H.6; responses to Reviewers iYVA and 9oUc.)

4. Embryo-level generalization and pretraining stability. We retrained ZoomST from scratch with all test embryos excluded from pretraining. Under this stricter split, we reassessed developmental-stage prediction, expanded the developmental-ordering evaluation to 30 sections from eight held-out embryos, and assessed the stability of developmental ordering and tissue mapping across three independent pretraining runs. (Appendix H.8; response to Reviewer 6AXt.)

5. Model design and neighborhood-size sensitivity. We added ablations of cross-granularity transformations applied to both views and distance-weighted molecular aggregation, and evaluated alternative neighborhood configurations. We further clarified the rationale for the dual-view design and the physical interpretation of neighborhood sizes under the shared bin50 sampling scheme. (Appendices H.2–H.4; responses to Reviewers iYVA and 9oUc.)

We have updated Table 2 and added Appendices H.1–H.8 (pp. 36–53), with new manuscript content marked in red. Detailed responses are provided in our reviewer-specific replies. We sincerely appreciate the reviewers’ thoughtful suggestions and the opportunity to further strengthen our work.

Area_Chair_3cfjOpenReview ↗

Please check authors' rebuttal and participate in the discussion

Dear reviewer,

The discussion phase is coming to an end. Please read the authors’ rebuttal and post a brief follow-up in the discussion thread. Let us know which of your concerns the response addresses and which remain open.

Please post any follow-up questions in the thread visible to the author as soon as possible, so that the author has time to respond and provide clarification. I also encourage you to engage with one another’s assessments—especially where your perspectives differ—and to note whether any counterarguments or discussions lead you to revise your own assessment.

Thank you for your effort and thoughtful review. Thank you again!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.