acceptodds
Under review as a conference paper at ICLR 2027

STAGE: Deciphering Tissue Organization and Predicting Perturbation Responses through Transferable Spatial Representations

Abstract

Spatial transcriptomics measures gene expression while preserving tissue architecture, providing a basis for investigating tissue microenvironments and their functional responses. Learning shared representations that capture spatial organization and expression information across heterogeneous measurements is a key challenge for tissue analysis and downstream prediction. We present STAGE, a spatial transcriptomics foundation model for spatial domain analysis and expression recovery. STAGE jointly encodes gene expression and spatial geometry in cellular neighborhood graphs and uses hierarchical prototypes to organize representations at multiple resolutions. The model is pretrained on approximately 32.44 million cells spanning 18 tissue types and six spatial transcriptomics platforms. Across cross-platform, technical-replicate, and normal–tumor evaluations, both zero-shot and fine-tuned STAGE outperform six comparison methods in scenario-averaged spatial label consistency and embedding separation. Task-specific modules extend these representations to masked-cell expression recovery and spatial expression completion under known perturbation conditions. On 1,314 masked validation cells from colon Xenium, STAGE spatial reconstruction improves nonzero-expression Pearson correlation from 0.2305 for CIFM to 0.4139. On Perturb-map, conditioning CONCERT on STAGE expression representations reduces nonzero-gene MSE by 5.9% and energy distance by 5.6% relative to CONCERT. These results establish spatial representations as a shared basis for tissue organization analysis and expression reconstruction under known perturbation conditions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
spatial transcriptomics, spatial omics, foundation models, self-supervised learning, graph representation learning, spatial domain identification, hierarchical representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EkcsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces STAGE, a spatial transcriptomics foundation model that combines RNA expression and spatial information within cell neighbourhood graphs with hierarchical prototype learning. The model is pretrained on 32.44 million cells from 109 samples. The authors evaluate zero-shot and fine-tuned representations for spatial-domain analysis. Task-specific modules further extend the framework to masked-cell expression recovery, spatial expression completion under known perturbation conditions, and population-level response prediction for held-out perturbation targets. The results show improvements in most of these scenarios.

Critical strengths

The paper targets a meaningful problem in spatial transcriptomics: learning transferable representations across heterogeneous tissues and measurement platforms while adding cellular spatial context. The motivation for connecting tissue organisation with expression embedding is reasonable.

Critical weaknesses
  1. The proposed STAGE mainly combines existing components: graph-attention, SwAV prototype learning, hierarchical clustering, and consistency regularisation. This paper does not show a clear technical innovation over existing ST foundation models. Basically, the core design of this work overlaps with Novae, and the changes are incremental. A controlled comparison with Novae under matched pretraining data and postprocessing, together with targeted ablations of the added modules, is needed to determine whether the gains really from these modifications.

  2. The evaluation is dominated by spatial-domain consistency and embedding-separation metrics, with few assessment of general ST foundation model evaluation tasks such as cell-type annotation, cross-dataset label transfer, biologically faithful integration, and annotated spatial-niche or neighbourhood-composition prediction. UCE (Rosen et al., Nature, 2026), the SOTA single-cell FM that has been reported to perform much better than scGPT, is not evaluated. scGPT, as well as UCE, produces cell-level representations, whereas STAGE explicitly encodes spatial neighbourhoods. The paper does not clearly describe whether the scGPT baseline receives comparable neighbourhood context. Controls using scGPT and UCE embeddings aggregated over the same neighbourhoods, like using mean pooling of centre-cell and neighbourhood embeddings, are needed for fair comparison.

  3. ASW of the proposed method is calculated after a domain-centroid transformation, whereas baseline scores use their respective embeddings. This transformation changes the embedding geometry whose separation ASW measures, confounding representation quality with model-specific postprocessing. The reported ASW advantage cannot be attributed solely to the proposed pretraining method.

Questions for authors

Can the authors provide a comparison between STAGE and Novae using the same pretraining corpus and identical downstream evaluation and postprocessing protocols?

How does STAGE compare with scGPT and UCE when their cell embeddings are aggregated over the same physical neighbourhoods using mean pooling for example?

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

STAGE is a foundation model pre-trained on over 30 million cells. It uses a graph attention encoder to combine gene expression and spatial coordinates, and uses SwAV and Hierarchical consistency strategies for embedding learning.

STAGE is applied to spatial domain identification, masked-cell recovery, spatial completions of perturbations, and perturbation response prediction. Both zero shot and fine-tuned models outperform other methods on these tasks. Ablation studies examine the contribution of spatial information and hierarchical consistency.

Critical strengths
  1. The model is pre-trained on a large spatial corpus, which qualifies it as a foundation model from the data view in this area.
  2. Zero-shot model outperforms 6 other methods on the spatial domain identification task. The results also show model’s generalizability across different platforms, technical replicates, and disease conditions.
  3. STAGE achieved substantial improvement on perturbation spatial completion and response prediction.
Critical weaknesses
  1. Compared with Novae, the novelty of the method is limited. Novae already used GNN and SwAV. A sufficient ablation study on the Hierarchical consistency strategy is needed.

  2. The relation between hierarchical prototype learning and downstream tasks such as perturbation prediction is not clear.

  3. Baseline methods are not sufficient. Other spatial foundation model works such as SToFM[1] and HEIST[2] should be included.

[1] Zhao, Suyuan, et al. "SToFM: a multi-scale foundation model for spatial transcriptomics." arXiv preprint arXiv:2507.11588 (2025).

[2] Madhu, Hiren, et al. "HEIST: a graph foundation model for spatial transcriptomics and proteomics data." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors
  1. What are the key differences between STAGE and Novae in prototype learning and hierarchical consistency?

  2. How to use the hierarchy prototype embeddings on downstream tasks? How much do prototype probabilities contribute to downstream tasks?

  3. How does STAGE compare with SToFM and HEIST?

  4. Why zero-shot model achieves higher performance than the fine-tuning model in Table 1?

  5. In Table 2, why does spatial KNN have a higher NZr and NZρ? More justification for these results is needed.

  6. Why are the local and extended views defined as non-overlapping regions?

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

The paper proposes STAGE, a self-supervised model for spatial transcriptomics pretrained on about 32M cells from 18 tissues. A graph attention network encodes each cell’s spatial neighborhood from shared gene embeddings and normalized relative coordinates. Training uses SwAV-style swapped prediction against a three-level prototype tree, plus a view-consistency loss, an expression–coordinate alignment loss and a cross-level path loss. The main evaluation is unsupervised spatial-domain identification across cross-platform, technical-replicate and normal–tumor pairs, where zero-shot and fine-tuned STAGE report lower PAS and higher ASW than six baselines. Task-specific heads on top of the encoder are also evaluated on three downstream tasks: masked-cell expression recovery, spatial completion on Perturb-map, and population-level response prediction on Perturb-Multi.

Critical strengths
  • The problem is relevant. A single model that produces comparable neighborhood representations across platforms with different gene panels would be useful, and the shared gene-embedding design lets cells from different panels enter one feature space.
  • Pretraining is at a meaningful scale (32.4M cells, 109 samples, 18 tissues). The evaluation covers three realistic sources of heterogeneity: platform, technical replicate, and disease state.
  • Zero-shot STAGE achieves the best PAS and ASW in all three scenarios without any pair-specific training, which supports the claim that the representations transfer.
  • The appendix documents the protocols in detail, including the source of variance behind each reported standard deviation. The downstream claims are worded carefully (e.g., “known perturbation conditions”).
Critical weaknesses
  • Limited novelty relative to Novae. Much of STAGE’s design follows Novae [1]:

    • trainable gene embeddings with L2 normalization to handle different panels;
    • a GAT encoder with attention aggregation over Delaunay neighborhoods with a 100 µm cutoff;
    • SwAV swapped prediction with Sinkhorn-balanced prototypes between neighboring views;
    • the same pseudo-batch-effect and gene-panel-subsetting augmentations.
  • The hierarchical prototype losses are not new and are weakly motivated. Prototype hierarchies built by bottom-up k-means appear in [2], and multi-granularity prototypes in Prototypical Contrastive Learning [3]. Neither is cited or compared against. Lglobal (a hard parent target from the top-scoring child) and Lpath (a soft parent target from summed child probabilities) also constrain the same quantity, and the paper does not explain why both are needed.

  • The coordinate branch is not invariant to rotation or reflection. Coordinates are centered and RMS-scaled (Eq. 11) but otherwise enter as raw xy values, and Appendix B.7 lists no rotation augmentation. Embeddings can therefore depend on section orientation, which is arbitrary. The authors should use invariant features such as pairwise distances, add rotation and reflection augmentation, or show that the embeddings stay stable when an input section is rotated.

  • Fine-tuning makes results worse, and the paper does not explain why. In Table 1, zero-shot STAGE has better PAS than fine-tuned STAGE in all three scenarios: 0.224 vs. 0.263, 0.137 vs. 0.152, and 0.158 vs. 0.221. It also has better ASW in two: 0.195 vs. 0.179 cross-platform, and 0.194 vs. 0.159 normal–tumor. This weakens the case for the fine-tuning procedure and should be discussed; it may reflect overfitting to the reference section or a problem with the four-epoch protocol.

  • The ablation cannot attribute gains to individual components. Table 3 uses one trained model per configuration, and its “spatial information fusion” arm adds the coordinate branch, Lgeo and Lalign together. The only component tested on its own is Lpath (weight 0.0015). It changes PAS from 0.119 to 0.108 while JSD gets slightly worse (0.0504 to 0.0517), which may be within seed variation. Multiple seeds, a separate ablation of each loss term, and a Novae-equivalent configuration trained on the same corpus would make the contribution clear.

[1] Blampey, Quentin, et al. "Novae: a graph-based foundation model for spatial transcriptomics data." Nature Methods 22.12 (2025): 2539-2550.

[2] Guo, Yuanfan, et al. "Hcsc: Hierarchical contrastive selective coding." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

[3] Li, Junnan, et al. "Prototypical contrastive learning of unsupervised representations." arXiv preprint arXiv:2005.04966 (2020).

Questions for authors

Please see the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.