Reference-Based Cell Type Deconvolution Using Self-Supervised Contrastive Learning in Spatial Transcriptomics with Ctdecon

Abstract It is important to analyze the cell types in the spatial domain at the single cell resolution, especially in spatial transcriptomics. It is also key to present the cell types distributed in the spatial domain. Here, we introduce ctdecon, a module constructed by neural network algorithm, which can solve the problem of heavy mixing of multiple cell type signals within individual spatial spots, and obtain the reasonable proportion of cell types in each spatial site. It uses the cell types referenced by scRNA-seq to quantify the proportion of cell types in spatial transcriptomics, which greatly improves the performance. Different from the existing methods that only rely on gene expression, ctdecon integrates the spatial correlation of cell types to improve the performance of deconvolution, and yields consistent results relative to traditional methods in solving mixed cell types. ctdecon can minimize the reconstruction error while maintaining biological rationality and ensure the generalization ability of technical noise. In addition, ctdecon can directly provide a cell type map with spatial resolution, which can directly identify the location and proportion of cell distribution.