STHD: probabilistic cell typing of single Spots in whole Transcriptome spatial data with High Definition

Recent spatial transcriptomics (ST) technologies have enabled single- and sub-cellular resolution profiling of gene expression across the whole transcriptome. However, the transition to high-definition ST significantly increased data sparsity and dimensionality, posing computational challenges in identifying cell types, deciphering neighborhood structure, and detecting differential expression - all are crucial steps to study normal and disease ST samples. Here we present STHD, a novel machine learning method for probabilistic cell typing of single spots in whole-transcriptome, high-resolution ST data. Unlike the current binning-aggregation-deconvolution strategy, STHD directly models gene expression at single-spot level to infer cell type identities without cell segmentation or spot aggregation. STHD addresses sparsity by modeling count statistics, incorporating neighbor similarities, and leveraging reference single-cell RNA-seq data. We show in VisiumHD data that STHD accurately predicts cell type identities at single-spot level, which achieves precise segmentation of both global tissue architecture and local multicellular neighborhoods. The high-resolution labels facilitate various downstream analyses, including cell type-stratified bin aggregation, spatial compositional comparisons, and cell type-specific differential expression analyses. Moreover, STHD labels further reveal frontlines of inter-cell type interactions at immune hubs in cancer samples. STHD is scalable and generalizable across diverse samples, tissues, and diseases, facilitating genome-wide analyses in various spatial organization contexts. Overall, computational modeling of individual spots with STHD facilitates discoveries in cellular interactions and molecular mechanisms in whole-genome spatial technologies with high resolution. STHD is available at https://github.com/yi-zhang/STHD.