HetAnn: Cell Type Annotation for Spatial Transcriptomics by Heterogeneous Graph Learning.

Currently, several new spatial transcriptomics(ST) technologies have attained cellular or even subcellular spatial resolution. Effective methods for annotating cell types are essential for the growing volumes of single-cell ST data. Most of automatic annotation methods for ST data use reference single cell RNA-sequencing(scRNA-seq) data, but how to efficiently integrate spatial information from ST data and gene expression information from ST and scRNA-seq data into a unified model remains a challenge. The difficulty in integrating the single cell multimodal data is that different modalities have different feature spaces, which easily leads to cell type mismatching. In this study, we propose HetAnn, a semi-supervised deep learning method for cell type annotation of ST data. HetAnn learns ST-specific representation by cell graphs accounting for both cell-to-cell spatial distance and gene expression similarity in ST data. HetAnn employs a graph attention network to learn multimodal representation with end-to-end joint training on ST and reference scRNA-seq data. These two comprehensive embeddings enable HetAnn to remove the gap between modalities while preserving cell-type heterogeneity. Experiments on public datasets demonstrate the superiority of HetAnn compared with state-of-the-art methods.