Spatial transcriptomics (ST) technology allows scientists to measure all the gene activity in a tissue sample while also knowing its spatial location (spot) in the tissue. However, several of these technologies do not have a single-cell resolution but rather produce a group of potentially heterogeneous cells at each spot. One of the challenging problems associated with the spatial transcriptomics data has been to deconvolve these heterogeneous mixtures of cells.

In this work, we develop a reference-free approach based on Bayesian non-negative matrix factorization to deconvolve spatial transcriptomics data and obtain the cell type composition of each spot. Using simulations, we show that our method is more accurate in detecting the cell-type compositions than existing deconvolution techniques in case of varying spot size, heterogeneity, and imperfect single-cell reference. We illustrate the usefulness of our method using Mouse Brain Cerebellum data and Human Intestine Developmental data.