Spatiotemporal cell type deconvolution leveraging tissue structure

Spot-based spatial transcriptomics (ST) captures aggregated transcriptomic profiles at spatial locations (spots) in tissues. Deconvolution methods are needed to estimate the proportion of each cell type in every spot, and usually leverage a single cell transcriptomic reference (scRNA-seq). Though there are an increasing number of experiments that profile multiple adjacent tissue slices, no deconvolution method leverages 3D tissue structure. Some methods utilize the 2D spatial organization assuming neighboring spots are similar, which is not the case in heterogeneous environments. Moreover, most methods aggregate reference scRNA-seq profiles of the same cell type, missing subtle cell state variations. We present SpaDecoder, a parallelized per-spot deconvolution method for multiple neighboring spatial or temporal ST slices that predicts cell type proportions with a matrix factorization-based objective. SpaDecoder uses slice alignment, per-spot spatio-transcriptomic neighborhood inference, and 3D spatial Gaussian kernel weights to effectively leverage 3D structure and adapt to heterogeneous tissue environments. We model individual scRNA-seq profiles, instead of cell type aggregated, to capture cell state variability. The mathematical framework of SpaDecoder supports several downstream analyses. It uncovers key cell type regions and changing composition across slices, identifies colocalized cell types, imputes spatial gene expression, and predicts 3D spatio-temporal scRNA-seq cell locations. SpaDecoder outperforms other methods on various metrics, datasets, scenarios, and ablations, and yields interpretable biology, showing it harnesses 3D structure and single cell reference profiles to improve deconvolution. SpaDecoder is available at https://github.com/ZhangLabGT/spadecoder.