acceptodds
Withdrawn from ICLR 2027

Spatial Gene Expression Prediction as Field Reconstruction from Histology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial transcriptomics links gene expression to tissue morphology, but sequencing-based assays remain too costly and labor-intensive for routine profiling at scale. This has motivated a growing line of work on inferring spatial expression from standard hematoxylin-and-eosin (H&E) histology. Current methods, however, remain largely tied to assay-indexed prediction: they estimate expression at measured locations, with spatial context primarily used to inform predictions defined at those locations. This view is misaligned with how molecular signal is organized in tissue. Gene expression is not a set of isolated measurements, but a morphology-conditioned expression field shaped by regional tissue structure and resolved through local transcriptional variation. We formulate histology-based spatial gene expression prediction as *field reconstruction*. **Carta** reconstructs a morphology-conditioned expression field over the measured spatial coordinates of a spatial transcriptomics assay. It first estimates regional expression states from histology and then refines local expression by propagating information across nearby measured locations. To move beyond losses evaluated independently at sampled coordinates, Carta uses a bidirectional conditional transport objective that shapes the distributional structure of the reconstructed expression field across the spatial hierarchy. Across **16** organ-specific prediction tasks — **10** based on HEST-1k and **6** constructed from STImage-1K4M — Carta improves average PCC by **12.9** and **2.3** absolute PCC points over the UNI pathology foundation model baseline and the strongest baseline. It is particularly effective for sparse genes and spatially localized expression patterns, whose weak but structured variations can be under-represented by coordinate-level regression. These results show that morphology-conditioned field reconstruction provides a principled formulation for spatial expression prediction when direct molecular profiling is impractical.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial transcriptomics, field reconstruction, multi-scale modeling, gene expression prediction

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_U97sOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a hierarchical model to predict spatial transcriptomics from H&E images with bidirectional conditional transport.

Critical strengths

None

Critical weaknesses
  1. Poorly justified BCT formulation. The method treats regression-head weights as gene prototypes, directly borrowing from classification-based transport BCT. However, genes are not mutually exclusive classes, and there is no convincing justification for why this alignment improves expression prediction.
  2. Major misalignment between claim and experiments. Line 114-115 claims that “It is particularly effective for sparse genes and spatially localized expression patterns, for which weak but structured variations can be under-represented by coordinate-level regression.”, but in the experiments, only top50 highly variable genes are selected as training panels.
Questions for authors
  1. How does BCT preserve spatial expression structure?
  2. How would the model perform if you include a gene panel that has larger variance, such as the gene pathway benchmark proposed in HistoPrism? (Hu, S., Zeng, Q., Bhasker, N., Kather, J. N., & Speidel, S. (2026). HistoPrism: Unlocking Functional Pathway Analysis from Pan-Cancer Histology via Gene Expression Prediction. ICLR 2026)
Reviewer_d2MgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces Carta for predicting spatial gene expression from H&E histology images. The method constructs a spatial hierarchy from measured coordinates, predicts expression states at coarse scales, and interpolates these states to condition finer-scale predictions. The authors also introduce a bidirectional conditional transport (BCT) loss that regularizes the relationship between node prediction features and gene anchors derived from the output layer across multiple scales.

Critical strengths
  1. Consistent improvements across diverse tasks. Carta achieves the highest average PCC on both datasets. Its comparison with STFlow using the same UNI features supports the practical value of the proposed expression model.

  2. Improvements beyond correlation. Regional error, variance, and high-expression overlap metrics also improve, providing evidence that the gains benefit spatial reconstruction, which is central to the intended application.

Critical weaknesses
  1. BCT constrains cannot support the proposed mechanism for reconstructing spatial distributions and correcting spatial structural errors.

BCT aligns node prediction features with gene anchors, rather than aligning the spatial distributions of predicted and measured expression fields. Its cost does not explicitly incorporate spatial distances, tissue boundaries, or measured expression states. With the node feature set fixed at a given scale, permuting the spatial positions associated with those features leaves the BCT loss at that scale unchanged. Therefore, BCT itself cannot distinguish whether these features form the correct spatial arrangement. Although Table 9 supports such a performance improvement, it does not establish why it should be attributed to the proposed distribution-preserving mechanism.

Furthermore, the authors attribute the problem to insufficient spatial constraints in point-wise supervision, but the added BCT objective does not directly supply those missing spatial constraints. Point-wise MSE already uses the correspondence between expression values and spatial coordinates: predicting a high-expression region at the wrong location incurs a penalty. The authors therefore need to identify which spatial structural errors MSE fails to distinguish adequately, and how BCT provides an additional ability to distinguish them.

  1. The distinct contribution of field reconstruction remains unestablished. Multi-scale and hierarchical modeling have precedents in TRIPLEX and MERGE [1, 2], while next-scale prediction appears in GenAR [3]. Carta’s coarse-scale outputs are interpreted as regional expression states without validating their correspondence to measured regional expression (Sec. 3.2). Table 11 supports expression-state propagation over latent-state propagation, but does not establish that the gains arise from meaningful regional expression reconstruction rather than generic hierarchical refinement. This leaves the novelty of the central contribution insufficiently established.

[1] Chung et al. Accurate Spatial Gene Expression Prediction by Integrating Multi-Resolution Features. CVPR, 2024.

[2] Ganguly et al. MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images. CVPR, 2025.

[3] Ouyang et al. GenAR: Next-scale Autoregressive Generation for Spatial Gene Expression Prediction. Medical Image Analysis, 114:104232, 2026.

Questions for authors
  1. Why should BCT particularly benefit expression signals occurring at only a small number of locations? Do the existing results establish this effect, and how can it be distinguished from general latent-feature regularization?

  2. Do the average improvements in Table 9 and the sparsity-stratified analysis reflect benefits for most patients or slides, or are they driven by large gains on a small subset of samples?

  3. Beyond randomly omitted coordinates, does the method remain effective when contiguous regions are omitted or sampling density changes?

  4. Do coarse-scale outputs correspond to measured regional mean expression, and are they consistent with aggregated finer-scale predictions? Why is direct coarse-scale supervision omitted, and has its effect been evaluated?

Reviewer_PU6kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper formulates histology-based spatial gene expression prediction as reconstructing a morphology-conditioned expression field over measured assay coordinates. Rather than propagating latent features across a coordinate hierarchy, Carta estimates explicit regional expression states at coarser scales and interpolates them to condition finer-scale predictions, pairing coordinate-level regression with a bidirectional conditional transport objective across multiple HEST-1k and STImage-1K4M benchmarks. While the model achieves consistent correlation gains, particularly on sparse or localized genes, its transport regularizer operates purely on similarity between node features and gene anchors and remains invariant to spatial permutation within each scale.

Critical strengths
  1. Passing explicit predicted expression states across coarse-to-fine coordinate scales offers a clean, interpretable alternative to standard latent-feature hierarchies, allowing regional expression priors to directly steer local refinement.
  2. Beyond aggregate correlation metrics, the evaluation examines gene-sparsity strata, partial spatial support, and scale-schedule ablations across diverse organ cohorts, providing a thorough empirical picture of where the model improves over single-scale baselines.
Critical weaknesses
  1. The paper's framing overstates the novelty of formulating spatial transcriptomics as a sparse field over measured coordinates and building an active-bin hierarchy. Constructing a multiscale hierarchy directly on occupied coordinate bins closely parallels established point-set architectures and recent sparse-lattice ST models such as HiST (arXiv:2606.14251), while the within-scale geometry-aware attention is adopted from STFlow. Because the distinct architectural choice in Carta is propagating predicted expression states top-down rather than propagating latent features (as in HiST), the paper should explicitly discuss HiST and isolate whether expression-state propagation actually outperforms feature-level propagation on the exact same coordinate hierarchy.
  2. The BCT objective (Eqs. 11 to 15) does not constrain the spatial organization of the reconstructed field as claimed. Because the transport cost depends only on cosine similarity between node prediction features and gene anchors without using spatial coordinates or neighborhood structure, the loss is completely invariant to permuting node locations within a scale. It therefore acts as a marginal distributional regularizer over expression profiles rather than a spatial regularizer, which conflicts with the paper's central narrative that BCT preserves spatial organization where coordinate-level losses fail.
Questions for authors
  1. How does Carta compare conceptually and empirically against recent hierarchical sparse-field approaches such as HiST? Specifically, if you keep Carta's coordinate binning, local transformer, and BCT fixed, but propagate latent features across scales (in the style of HiST) instead of explicit predicted expression vectors, how much of the performance gain remains? Including a direct comparison with the open-source HiST implementation under the same 50-gene protocol would also help clarify their relative strengths.
  2. Given that BCT is invariant to node ordering within each scale, what evidence supports interpreting it as preserving spatial organization rather than simply matching the non-spatial distribution of expression states? For instance, does BCT behave any differently if node coordinates are randomly shuffled within each scale, or if a spatially aware transport cost is introduced?

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

Related papers

Discussion (0)

Sign in to comment.