Abstract Spatial transcriptomic (ST) technologies allow researchers to examine high-quality RNA-sequencing data along with maintained two-dimensional positional information as well as a co-registered histology image. A popular use of ST omics data is to provide insights about tissue structure and spatially unique features. However, due to the technical nature unique to most ST data, the resolution varies from a diameter of 2-10 μm to 50-100 μm instead of single-cell resolution, which brings uncertainty into cell number and cell mixture within each ST spot. Motivated by the important role for spatial arrangement of cell types within a tissue in physiology and disease pathogenesis, several ST deconvolution methods have been developed and are being used to explore gene expression variation and identification of spatial domains. The aim of this work is to review state-of-the-art methods for ST deconvolution, while comparing their strengths and weaknesses. Specifically, we use four real datasets to examine the performance of eight methods across different tissues and technological platforms. Key Points Cell mixture inference is a critical step in the analysis of spatial transcriptomics (ST) data to prevent downstream analysis suffering from confounding factors at the spot level. Existing ST deconvolution methods can be classified into three groups: probabilistic-based, non-negative matrix factorization and non-negative least squares based, and other deep learning framework-based methods. We compared eight ST deconvolution methods by using two single cell level resolution datasets and two spot level resolution ST datasets. We provided practical guidelines for the choice of method under different scenarios as well as the optimal subsets of genes to use for each method.

A comprehensive comparison on cell type composition inference for spatial transcriptomics data