SMART: reference-free deconvolution for spatial transcriptomics using marker-gene-assisted topic models

Abstract

Abstract Spatial transcriptomics (ST) offers valuable insights into gene expression patterns within the spatial context of tissue. However, most technologies do not have a single-cell resolution, masking the signal of the individual cell types. Here, we present SMART, a reference-free deconvolution method that simultaneously infers the cell type-specific gene expression profile and the cellular composition at each spot. Unlike most existing methods that rely on having a single-cell RNA-sequencing dataset as the reference, SMART only uses marker gene symbols as the prior knowledge to guide the deconvolution process and outperforms the existing methods in realistic settings when an ideal reference dataset is unavailable. SMART also provides a two-stage approach to enhance its performance on cell subtypes. Allowing the inclusion of covariates, SMART provides condition-specific estimates and enables the identification of cell type-specific differentially expressed genes across conditions, which elucidates biological changes at a single-cell-type resolution.