CHORD: a framework for cross-species single-cell integration across gene, cell and cell-type levels

Quantifying cross-species relationships among cell types from single-cell transcriptomic data can reveal both conserved and divergent patterns of cell-type hierarchies. However, existing cross-species integration methods can be limited in modeling genes beyond orthologs by leveraging cell-type-resolved transcriptional context, or in learning explicit type-level representations. Here we present CHORD, a cross-species integration framework that jointly learns representations of genes, cells and cell types. We demonstrate that CHORD can integrate cross-species single-cell atlases and support cell-type annotation with unknown cell-type detection. In the frog-zebrafish embryogenesis and mammalian motor cortex atlases, CHORD infers cell-type trees that place conserved cell types from different species in relative proximity and summarize hierarchical relationships among cell types. CHORD also supports cross-species comparison of continuous phenotypic variation by placing embryonic cells along an aligned developmental timeline. CHORD further yields gene embeddings that capture orthologous and functional relationships, and gene importance scores linking genes to cell types.