Region-aware bridge modeling enables interpretable mesoscale representation of spatial transcriptomic tissue sections

Spatial transcriptomics maps tissue architecture at high resolution, but spot-level maps are difficult to compare across sections, whereas whole-section averages obscure regional organization. We developed region-aware bridge modeling to aggregate interpretable biological program features into compact mesoscale summaries.

Using public colorectal cancer and breast cancer 10x Genomics high-definition Visium sections, we constructed epithelial-like, fibroblast, smooth or myoepithelial, and extracellular matrix bridge features from cell-state indicators, curated gene-program scores, and quality-control summaries. Median-quadrant aggregation produced an eight-region design matrix. Within-section validation showed non-random mesoscale heterogeneity: region-label shuffling reduced the overall mean absolute pairwise regional difference from 0.298 to 0.0020 in colorectal cancer and from 0.328 to 0.0024 in breast cancer, with empirical upper-tail probability values of 0.0002 for both analyses across 5000 permutations. Shifted and rotated partitions changed the overall heterogeneity metric by less than 10%. Exploratory ridge modeling identified fibroblast as the only non-target predictor in the extracellular matrix model, and Bayesian sensitivity analysis supported a positive fibroblast-extracellular matrix association. Supplementary lung, prostate, and ovarian cancer sections supported workflow applicability.

Code and processed outputs are available through GitHub and Zenodo under 10.5281/zenodo.19801620.