acceptodds
Under review as a conference paper at ICLR 2027

GLoDe: Global–Local Decomposition for Histology-based Spatial Gene Expression Prediction

15 views

Abstract

The high cost of spatial transcriptomics motivates scalable prediction of spatial gene expression from readily available hematoxylin and eosin (H&E) images. However, existing methods typically predict absolute gene expression without explicitly disentangling slide-level molecular shifts from within-slide spatial variation, which may cause models to capture slide-specific biases rather than transferable spatial patterns, limiting their generalization across tissue sections. To address this limitation, we propose GLoDe, a global–local decomposition framework that decomposes gene expression into a slide-shared molecular baseline and a spot-specific residual, explicitly accounting for their distinct sources of variation. The global component captures tissue-level molecular characteristics shared across spots within each tissue section, while the local component employs a gene-conditioned residual decoder to integrate gene priors, local morphology, and spatial context for predicting gene-specific residual expression. Extensive experiments on HEST-1K and HER2ST demonstrate consistent performance gains, including a 16.9% relative improvement in average Pearson correlation under out-of-fold evaluation over the strongest evaluated baseline on HEST-1K.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, Computational Pathology

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_7JKWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes GLoDe, a histology-to-spatial-transcriptomics prediction framework that decomposes gene expression into a slide-level baseline and spot-specific residuals. It combines pretrained pathology and gene embeddings with spatial attention and gene-conditioned decoding.

Critical strengths

Addresses a relevant biological prediction problem.

Critical weaknesses
  1. Limited methodological novelty. Hierarchical modeling on this problem has been long existing.

  2. Small datasets and weak generalization evidence. Several cohorts contain only 2–4 slides. Despite the large number of spatial spots, the number of independent biological samples is very limited, raising concerns about overfitting and the reliability of the reported generalization results.

  3. Limited gene coverage. The main HEST-1K experiments evaluate only 50 selected genes, a critical detail relegated to the appendix. Performance on such a small panel does not adequately characterize transcriptome-wide prediction quality or generalization across genes.

  4. Questionable baseline comparability. STPath supports more than 30k genes but is evaluated without fine-tuning, whereas GLoDe is specifically trained on small gene panels. Although the authors report matched evaluation panels, differences in training protocols and gene coverage complicate the comparison. The fairness of the reported improvements is insufficiently established.

Questions for authors
  1. How reliable are the generalization claims given that several cohorts contain only 2–4 slides?
  2. How does GLoDe perform on substantially larger gene panels, including low-expression and less-variable genes?
  3. Why is STPath evaluated without fine-tuning, and how are differences in training protocols, gene coverage, and expression preprocessing controlled?
Reviewer_8KNDOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GLoDe, a global–local decomposition framework for histology-to-spatial-transcriptomics prediction. The authors argue that most existing methods directly predict absolute spot-level expression without explicitly distinguishing slide-shared expression levels from within-slide spatial variation, and that shared spot representations may not fully capture gene-specific cellular and morphological patterns. GLoDe decomposes expression into a slide-shared molecular baseline and a spot-specific residual. The authors further introduce a Gene-Conditioned Residual Decoder that combines gene priors, local morphology, and spatial context to construct gene-specific visual features for residual prediction through a parameter-shared decoder. Experiments on HEST-1K and HER2ST report improvements in average prediction performance over the evaluated baselines.

Critical strengths
  1. Clear presentation. The paper is generally well organized, with a readable architecture diagram and clear descriptions of the global and local branches, gene-conditioned decoder, and training objectives.
  2. Explicit and testable formulation. The decomposition into a slide-shared baseline and spot-specific residual is clearly formulated. The baseline supervision and residual-centering objective provide concrete constraints for examining the intended roles of the two branches.
  3. Evaluation across diverse settings. The experiments cover ten HEST-1K cancer cohorts and HER2ST, spanning multiple spatial transcriptomics technologies and gene-panel sizes. Patient-level splits evaluate generalization to unseen patients, and the baselines represent several established modeling approaches.
  4. Useful supporting analyses. The paper includes ablations of the global branch, decomposition supervision, gene conditioning, and pretrained encoders, together with biomarker prediction and representation analyses. These provide preliminary evidence about the contributions of the proposed components.
Critical weaknesses
  1. Unverified baseline–residual separation. Prediction gains do not establish that the branches learn their intended roles on unseen slides. Baseline errors, residual errors, and slide-wise mean predicted residuals should be evaluated. It also remains unclear whether the global branch captures biological variation or technical batch effects.

  2. Incomplete comparison with STPath. STPath outperforms GLoDe in MF PCC on five HEST-1K cohorts, yet its OOF results are missing without clear justification. Using a pretrained checkpoint without fine-tuning does not explain this omission, particularly when the headline improvement relies on OOF PCC.

  3. Insufficiently differentiated motivation. The paper does not sufficiently explain why existing spatial-context methods cannot capture slide-level variation or why explicit decomposition is preferable. The alleged bottleneck of a “shared spot representation” also needs clarification, since shared embeddings can support gene-specific output mappings.

  4. Uncertain statistical reliability. An exploratory paired sign-flip test on the ten reported cohort-level MF PCC values does not detect a significant difference between GLoDe and STPath (two-sided p ≈ 0.592). This does not replace patient-level paired analysis, which is needed to assess robustness. Tables 3–4 and Figure 3 also lack variability estimates despite modest component gains.

  5. Ambiguities affecting reproducibility. Full-model results differ between Table 1 (0.392/0.270 MF/OOF PCC) and Tables 3–4 (0.386/0.280) without explanation. The paper should also clarify whether fold-specific gene panels are identical and how differing gene coverage is handled in OOF evaluation.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.