Xener: Bidirectional evidence propagation on a plant cell-type knowledge graph for traceable cross-species annotation

The cross-species annotation of plant single-cell data is limited by sparse reference atlases, one-to-many homology relationships, and context-dependent marker specificity, particularly in nonmodel species. We developed Xener, a reference-independent framework integrating a heterogeneous plant cell-type knowledge graph with bidirectional evidence propagation. Xener-Anno propagates evidence from cluster-level differentially expressed genes through protein homology, marker–cell-type associations, and Plant Ontology (PO) relationships to identify the most specific supported identity. Xener-Refine projects identity evidence onto individual cells, resolving within-cluster heterogeneity and recovering minority populations without reclustering. Across 62 datasets from 24 species, Xener achieved the highest hierarchical concordance in 46 datasets and outperformed the reference-independent methods GPTCelltype and PubMed-SKILL in terms of mean and median concordance. Compared with reference-dependent SATURN, Xener achieved 3.47-fold higher median concordance. In a Brassica rapa atlas, Xener resolved bundle sheath annotations into transcriptionally distinct candidate populations supported by traceable gene–homolog–cell-type paths. We further compiled Xener-Markers, an orthogroup-linked cross-species resource that classifies candidate markers as Core, Broad, Narrow, or Unique across 38 PO cell types. With respect to the spatial data for bamboo, Xener integrated evidence from Arabidopsis thaliana and rice, reaching 87.5% spot-weighted agreement with published labels; RNA in situ hybridization supported the localization of candidate marker clrGene018032. Thus, Xener extends traceable cell-type annotation beyond well-characterized model species.