H&E Slide Spatial Features Explain Gene Expression Beyond Cell-Type Mix

Routine H&E-stained tissue slides, already a staple of pathology labs worldwide, may carry far more transcriptional information than previously assumed. In a study published in BMC Bioinformatics on 28 October 2026, Jagadheshwar Balan and Nicholas B. Larson show that cell-type-specific spatial features extracted from digitized whole-slide images explain gene expression variability beyond what cell-type composition alone can account for. The work addresses a persistent gap: bulk RNA deconvolution can estimate which cell types are present in a tissue sample, but it loses the spatial context in which those cells sit.

The researchers leveraged a dataset of 471 normal prostate tissue samples that had both bulk RNA gene expression measurements and paired whole-slide images. Each slide was processed for nuclei segmentation and cell-type classification, yielding seven cell-type-specific proportion measures and 38 spatial features that describe local neighborhood structure and tissue architecture. These features are described as explainable, meaning they capture measurable geometric and organizational properties of the tissue rather than opaque computational representations.

To quantify how much the spatial features add, the authors modeled gene expression as a function of the cell-type-specific features and compared nested models. A proportions-only model served as the null model, while a full model added the spatial features. Likelihood ratio testing identified 2,330 of 24,075 genes whose model fit improved significantly when spatial features were included, after adjustment for multiple testing with a false discovery rate threshold of 0.05. That is roughly one in ten genes showing a detectable association with tissue architecture beyond cell-type makeup.

The explanatory gains were consistent across model comparisons. Among the significant genes, the full model combining spatial and proportion features achieved a median adjusted R-squared of 0.195, with an interquartile range of 0.122 to 0.289. A spatial-features-only model reached a median of 0.176 (0.109 to 0.263), while the proportions-only model managed only 0.068 (0.031 to 0.128). The pattern suggests that for many genes, the arrangement of cells in tissue carries explanatory signal comparable to, and in combination with composition greater than, the proportions themselves.

The team then turned to machine learning to test whether the regression findings held up. Gene expression was modeled using XGBoost with the cell-type-specific features as inputs. Feature importance analysis identified nine distinct, biologically relevant gene clusters and confirmed the regression-based results, with cell-type-specific spatial features contributing substantially to gene expression variability beyond cell-type composition.

The practical appeal of the approach lies in cost and scale. Single-cell and spatial transcriptomic technologies can characterize both cell-type composition and tissue architecture directly, but their expense limits widespread adoption. H&E whole-slide images, by contrast, are routinely generated in clinical and research settings, making them a scalable and cost-effective source of spatial information about bulk tissue samples.

Several caveats frame the findings. The study population consisted of normal prostate tissue, so the extent to which the same spatial features explain expression variability in diseased tissue remains untested here. The adjusted R-squared values, while clearly higher with spatial features included, indicate that a substantial share of gene expression variability remains unexplained by either feature set. The authors also note that the article was shared early as a peer-reviewed, accepted version of record that is citable and carries a permanent DOI, subject to further edits before the final version appears.

Within those limits, the study demonstrates that explainable cell-type-specific spatial features derived from routine H&E slides capture transcriptionally relevant tissue architecture that is not reflected by cell-type composition alone. For researchers working with bulk tissue datasets that lack matched spatial transcriptomics, the results suggest that the pathology slide already in the drawer may be a useful, computable proxy for a layer of biology that bulk measurements have historically missed.

Subject of Research: Extracting cell-type-specific spatial features from digitized H&E slides to explain gene expression variability in bulk prostate tissue

Article Title: Cell-type-specific explainable spatial features extracted from digitized H&E slides explain gene expression variability beyond cell-type composition

Article References: Balan, J., & Larson, N. B. (2026). Cell-type-specific explainable spatial features extracted from digitized H&E slides explain gene expression variability beyond cell-type composition. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06741-4

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06741-4

Keywords: spatial transcriptomics, H&E histopathology, whole-slide imaging, gene expression, bulk RNA deconvolution, prostate tissue, cell-type composition, tissue architecture, XGBoost, bioinformatics, nuclei segmentation, likelihood ratio test