High-precision cell-type mapping and annotation of single-cell spatial transcriptomics with STAMapper

Recent advances in single-cell spatial transcriptomics (scST) have enabled the analysis of gene transcription levels in individual cells while preserving their spatial positions. Cell-type mapping and annotation are crucial in understanding the complex interactions between cells and their microenvironments within a spatial context. To this end, we develop a heterogeneous graph neural network, STAMapper, to transfer the cell-type labels from single-cell RNA-seq data to scST data. STAMapper captures both the expression similarity among cells and the expression relationships between cells and genes and adopts a graph attention classifier to conduct semi-supervised learning for more accurate cell-type prediction. We collected 81 scST datasets consisting of 344 slices and 16 paired scRNA-seq datasets from eight technologies and five tissues to validate the efficiency of STAMapper. STAMapper achieved the best performance on 75 out of 81 datasets compared to competing methods in accuracy. STAMapper demonstrated enhanced performance over manual annotations, particularly at the boundaries of cell clusters, enabled the unknown cell-type detection in scST data, and exhibited precise cell subtype annotations. Additionally, STAMapper provided biologically meaningful gene embeddings, facilitating the identification of shared or unique gene modules across datasets.