SpatialTRACE predicts anatomical axes and regions in spatial transcriptomics and microscopy

Abstract

Spatial transcriptomics measures gene expression in tissue sections. However, interpretation requires anatomical maps that link gene expression and cellular composition to tissue structure. Annotating entire sections often requires extensive manual labor. We developed SpatialTRACE (Tissue Region and Axis Coordinate Estimation), consisting of graph- and image-based models, to extend annotations of a few structures into tissue-wide maps of anatomical axes and regions. SpatialTRACE-Graph combines gene-expression profiles with spatial neighborhoods to predict anatomical coordinates or region membership throughout spatial transcriptomic datasets. In held-out mouse small-intestine sections, it predicted crypt-villus and epithelial-distance axis coordinates using as few as 10 annotated training villi and identified Peyer's patches from region annotations. SpatialTRACE-Image predicts the same anatomical axis coordinates and regions across entire tissue images from DAPI alone. This multiscale vision transformer learns from coordinate and region predictions generated by SpatialTRACE-Graph. We applied SpatialTRACE-Image to immunofluorescence images to map the anatomical distribution of antigen-specific P14 CD8 T cells responding to acute systemic infection with lymphocytic choriomeningitis virus (LCMV) Armstrong in the small intestine. Compared with the vehicle-treated section, a section treated with a retinoic acid receptor inhibitor contained fewer P14 CD8 T cells overall, with a smaller fraction in the upper-villus lamina propria and a relative enrichment in the muscularis. Overall, the SpatialTRACE models reduce repeated manual annotation and provide a route to extend anatomical maps learned from spatial transcriptomics to DAPI-containing microscopy data.