SpaJoint: a transfer learning method for spatial transcriptomics deconvolution

Currently, many widely used spatial transcriptomics (ST) technologies do not achieve single-cell resolution, with each spot capturing signals from multiple, potentially heterogeneous cells. As a result, a key challenge is to resolve the spatial distribution of distinct cell types within tissues, which is fundamental for understanding tissue architecture and biological function. Here we present a deconvolution method based on transfer learning, SpaJoint. This method integrates gene expression derived from single-cell RNA sequencing (scRNA-seq) and ST, taking into account the spatial correlation across locations of different spots. Comprehensive experiments demonstrate that SpaJoint achieves excellent performance in predicting the cell-type composition of spatial spots and identifying the spatial regions of cell types, thus highly effective and broadly applicable among various scRNA-seq and ST datasets. Additionally, it exhibits remarkable robustness to hyperparameters and provides significant advantage in computational efficiency.