scPCIP: cross-species annotation of plant cell types with conserved gene programs

Single-cell atlases now cover many plant species and organs, yet their cell types are still annotated with marker genes that often fail to transfer between species. Here we show that plant cell identities are captured by gene programs conserved across species. We learned programs for root, leaf, flower and seed cell types on ortholog groups from 112,905 author-labelled cells and implemented them in scPCIP, which annotates cells of species absent from training without a reference. Across 26 held-out datasets, scPCIP outperformed marker-based and plant deep-learning methods, matched supervised classifiers trained on the same atlas and transferred, at lower accuracy, to an independent study; its programs followed tissue position in spatial transcriptomes. The conserved core of each program separated cell types better than its periphery in all 11 held-out species, whereas most species-specific markers turned over. Across organs, leaf bundle-sheath cells carried a root endodermis program in 8 of 12 studies.