|
Abstract:
|
Currently, methods specifically designed for spatial transcriptomics (ST) data modeling are lacking. Firstly, for most existing methods, cellular and regional expression profiles are typically analyzed first without the spatial information and only later projected back onto the spatial structure for visual inspection of spatial trend. Secondly, similar to single cell RNA-Seq data, ST based gene expression data is also plagued by dropout events, a phenomenon where genes actually expressed in a given cell or region are incorrectly measured as unexpressed. Thirdly, the gene by sample expression matrix is no longer retainable for many spatial methods. To address these challenges, we present a regularized maximum likelihood estimator to recover the noisy observed expression matrix as an approximately low-rank expression matrix under Poisson distribution, which is also spatially smooth. Our method enables spatial clustering by modeling a low-dimensional representation of the count-based gene expression matrix and encouraging neighboring spots to belong to the same cluster via a spatial smoothness penalty term.
|
