acceptodds
Under review as a conference paper at ICLR 2027

EPS quantifies the predictability of gene expression from histology images

Abstract

Histology images and transcriptomic profiles capture complementary aspects of tissue biology: morphology and molecular state, respectively. Recent advances in multimodal learning have enabled models to predict the expression of thousands of genes directly from volumetric H&E images. However, the fundamental predictability of individual genes from histology—and hence the intrinsic limits of such models—remains poorly understood. Here, we introduce the Expression Predictability Score (EPS), a model-agnostic metric for quantifying the extent to which the expression of each gene can be inferred from another modality. Under a constrained Gaussian Markov random field model, we establish a theoretical connection between EPS and mutual information. In simulations where mutual information is analytically tractable, EPS closely tracks the true mutual information between modalities. Across diverse spatial omics datasets and tissue types, EPS consistently distinguishes highly predictable from poorly predictable genes, and its relationship with empirical prediction performance remains stable across substantially different predictive models. Genes with high and low EPS further exhibit distinct enrichment patterns across cellular components, revealing biological determinants of cross-modal predictability. Beyond histology-to-transcriptome inference, EPS shows consistent score–performance relationships when predicting RNA expression from surface proteins or chromatin accessibility. Together, these results establish EPS as a general framework for characterizing cross-modal information content, identifying predictable molecular features, and delineating the intrinsic scope of multimodal inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
spatial omics, H&E image, gene predictability, mutual information

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6AZSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this manuscript, the authors present the Expression Predictability Score (EPS), a model-agnostic metric for quantifying the predictability of gene expression from histology images. The proposed EPS measures the consistency of gene expression within neighbourhoods defined by pretrained image embeddings, using a normalised graph Laplacian energy formulation. The extensive experimental results demonstrate the association between EPS and empirical prediction performance.

Critical strengths
  • This manuscript presents an interesting and useful problem formulation that examines how prediction difficulty varies across genes and tissues, rather than solely proposing a framework for predicting gene expression.
Critical weaknesses
  • The novelty of EPS proposed by this manuscript is insufficiently distinguished from existing statistics. Substituting Eq. (1) into Eq. (2) shows that CgC_g is exactly Geary’s C, evaluated on the image-embedding graph; EPS is its negative logarithm. The manuscript discusses Laplacian Score but does not acknowledge this exact equivalence. Applying this statistic to cross-modal predictability may be useful, but the authors need to clarify what is new beyond the application and empirical evaluation.

  • The evaluation in this manuscript does not really establish that the proposed EPS can predict performance on independently held-out data. EPS is computed using expression values from the entire tissue section, including the test locations, whereas model performance is evaluated on an 80% and 20% within-slice split. This does not demonstrate that a score computed from available training data can reliably guide target selection for unseen observations.

Questions for authors
  1. I suggest the authors clarify further on the contribution of this manuscript. What is the principal advance beyond this statistic and the cited Laplacian Score: the application, theoretical interpretation, aggregation, or empirical findings?

  2. How is the generalizability of the proposed method? Please include additional experiments to evaluate performance on independently held-out data instead of the within-slice split.

Reviewer_hUsJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the Expression Predictability Score (EPS), a model-agnostic metric for quantifying the extent to which the expression of each gene can be inferred from another modality. EPS constructs a nearest-neighbor graph from image embeddings and treats gene expression as a signal on this graph. Experiments show that EPS can reflect changes in crossmodal information dependence.

Critical strengths
  1. Important research question. Given an H&E image to predict gene expression is important and performance varies by gene and models. This paper provides a metric to show when a gene can be predicted.
  2. The results "low-EPS genes exhibit shared limitations in predictive performance, whereas the potential of high-EPS genes is realized to different degrees by different models" show the usage of EPS.
Critical weaknesses
  1. The formula 2 is exactly Geary's C metric. However, this paper doesn't mention this point. The author should mention their novelty but not just apply Geary's C metric.
  2. This paper only uses UNI as patch encoder. Since we know different patch encoders may get different results, the robustness of EPS is unclear.
  3. Figure 2 is low resolution and the scale and axis labels are illegible, making this paper difficult to verify the evidence supporting the core contributation.
Questions for authors

What is the difference between formula 2 and Geary's C metric and what is the novelty beyond just applying Geary's C metric? Can we still get the same results with other patch encoders (CONCH, Gigapath, ...)?

Reviewer_UVv9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Models that predict gene expression from H&E images perform very differently across genes, and it is unclear whether this reflects model limitations or how much information the images actually carry about each gene. The paper proposes the Expression Predictability Score (EPS), a training-free score that measures how consistent a gene’s expression is across image-similar spots on a k-NN graph of pretrained UNI embeddings. Under stated assumptions, a Gaussian Markov random field derivation links EPS to mutual information. EPS tracks analytic mutual information in simulations (r≈0.996r \approx 0.996), correlates with gene-level accuracy of nine predictors across seven tissues and with slice-level accuracy across 510 HEST sections (r=0.87–0.98r = 0.87–0.98), and shows similar trends when predicting RNA from surface protein or chromatin accessibility.

Critical strengths

S1. Useful question. A cheap, model-free way to tell which genes and slices can realistically be predicted from histology would help both method developers and users. This is the paper’s main source of value.

S2. Broad, consistent evidence. The EPS–performance trend holds across tissues, 9 predictors, 510 slices, and 2 non-imaging modalities. That makes it a robust empirical pattern, even though, as argued below, what EPS actually measures is unclear.

Critical weaknesses

W1. Assumption might not hold. Reaching mutual information requires assuming equal marginal entropies across genes, which the authors concede may not hold.

W2. The experiments do not show what EPS adds.

  • Not model-agnostic. EPS closely mirrors the leave-one-out error of a 5-nearest-neighbor regressor on UNI features. Its agreement with Ridge, MLP and the other baselines, likely trained on the same features, is therefore expected. EPS is also computed using the test spots. No simple baselines. Lowly expressed, sparse genes would score low on both EPS and every model’s accuracy simply because of sampling noise. The kNN mutual-information baseline that is tested performs comparably.
  • Leaky evaluation. Random spot-level splits within a slice let spatially smooth genes look easy for both EPS and the models. The cross-patient setting, which matters in practice, is not tested.
  • One encoder, one k. EPS reflects predictability from UNI at k = 5, not “fundamental” predictability from histology as the title and abstract claim.
Questions for authors
  1. How stable are gene rankings across encoders (e.g., CONCH, Virchow, etc.) and values of k?
  2. Do the results hold if EPS is computed on training spots only, and performance is measured on spatially separated regions (a.k.a. from another slide from the same patient)?
  3. Does EPS still predict gene-level performance after controlling for mean expression, sparsity, and spatial Moran’s I? Does SPS still predict slice-level performance after controlling for sequencing depth, spot count, and technology?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.