Abstract
Recent advances in spatial transcriptomics technologies have enabled gene expression profiling across the transcriptome in spots with subcellular resolution, but high sparsity and dimensionality present significant computational challenges. We present STHD for probabilistic cell typing of single spots in whole-transcriptome spatial transcriptomics with high definition. With a machine learning model combining count statistics with neighbor regularization, STHD accurately predicts cell type identities of subcellular spots, revealing both global tissue architecture and local multicellular neighborhoods. We demonstrate STHD in spatial analyses of cell type-specific gene expression and immune interaction hubs in tumor microenvironment, and its generalizability across samples, tissues, and diseases.