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

Cell-type composition and tissue architecture jointly influence gene expression patterns. Although single-cell and spatial transcriptomic technologies can characterize both factors, their cost limits widespread use. Bulk RNA deconvolution estimates cell-type composition but lacks spatial context. In contrast, routine H&E whole-slide images provide a scalable source of explainable cell-type-specific spatial features describing tissue architecture. However, the extent to which the spatial features explain gene expression variability beyond cell-type composition remains largely unknown, and in this study, we aim to address this gap. The study leveraged a dataset of 471 normal prostate tissue samples with bulk RNA gene expression and paired WSIs. The WSIs are processed for nuclei segmentation and cell-type classification to extract cell-type-specific proportions ( n = 7) and spatial features ( n = 38) reflective of local neighborhood and tissue architecture. We model the gene expression as a function of cell-type-specific features to assess their contribution to gene expression variability, and evaluate the incremental explanatory effect of spatial features over the cell type compositions. Likelihood ratio testing comparing the full model (cell-type-specific spatial and proportion features, with the proportions-only model serving as the null model) to the proportions-only model identified 2330 of 24,075 genes with significantly improved model fit after adjustment for multiple testing (FDR-adjusted p-value < 0.05). Among significant genes, the full model shows the highest ability to explain gene expression variability (median adjusted \(\:{R}²\) = 0.195 [IQR 0.122, 0.289]), compared to the spatial features only model (0.176 [0.109, 0.263]), and the proportions only model (0.068 [0.031, 0.128]). Finally, gene expression was modelled using XGBoost with cell-type-specific features. 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. These findings demonstrate 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. Such features provide a scalable and cost-effective approach for explaining gene expression variability in bulk tissue samples.