Cross-species cell-type assignment of single-cell RNA-seq by a heterogeneous graph neural network

Abstract Cross-species comparative analyses of single-cell RNA sequencing (scRNA-seq) data allow us to explore, at single-cell resolution, the origins of cellular diversity and the evolutionary mechanisms that shape cellular form and function. Here, we aimed to utilize a heterogeneous graph neural network to learn aligned and interpretable cell and gene embeddings for cross-species c ell type a ssignment and gene m odule e xtraction (CAME) from scRNA-seq data. A systematic evaluation study on 649 pairs of cross-species datasets showed that CAME outperformed six benchmarking methods in terms of cell-type assignment and model robustness to insufficiency and inconsistency of sequencing depths. Comparative analyses of the major types of human and mouse brains by CAME revealed shared cell type-specific functions in homologous gene modules. Alignment of the trajectories of human and macaque spermatogenesis by CAME revealed conservative gene expression dynamics during spermatogenesis between humans and macaques. Owing to the utilization of non-one-to-one homologous gene mappings, CAME made a significant improvement on cell-type characterization cross zebrafish and other species. Overall, CAME can not only make an effective cross-species assignment of cell types on scRNA-seq data but also reveal evolutionary conservative and divergent features between species.