Abstract:
|
Spatially resolved transcriptomics (SRT) has revolutionized biomedical research, providing spatial context to gene expression (GE). Some SRT technologies measure genes at “spots” containing multiple cells. Thus, to localize cell types in tissue, we developed SpaDecon, a transfer learning-based SRT cell-type deconvolution method. First, SpaDecon locally smooths the SRT GE matrix through multiplication with an adjacency matrix representing the spatial and histological similarity between each pair of spots. SpaDecon then trains a stacked autoencoder to identify a mapping of an annotated scRNA-seq GE matrix to a latent feature space. The weights of the network are optimized by finding hidden features relevant for cell-type classification. These weights are then transferred to a network for which the smoothed SRT GE matrix is used as input. After optimization, cell-type proportions for a given spot are estimated using the distances between that spot’s embedding and those of cell types in the scRNA-seq data. SpaDecon is effective in deconvolving SRT datasets from cancerous tissue sections and outperforms competing methods when deconvolving benchmark datasets resembling real SRT data.
|