Specificity-driven graph learning for the identification of rare cell states

Multicellular organisms comprise a diverse array of cell types and cell states, some of which exist at extremely low abundance and appear only transiently within restricted spatial niches or in response to stress and injury. Despite their scarcity, these cells can nonetheless contribute to lineage commitment, tissue regeneration, disease progression, immune evasion, therapy tolerance, and relapse. Single-cell RNA sequencing has provided a high-resolution window for dissecting this heterogeneity, thereby enabling low-abundance populations to be captured in principle (Choi 2019 ). For cells that have already been captured, the challenge of preserving and reliably identifying rare signals while controlling technical variation remains a central issue in single-cell and spatial transcriptomic analyses. Many existing methods construct neighborhood graphs based on intercellular similarity and propagate information among similar cells. When rare cells lack similar neighbors, their representations are prone to over-smoothing and are ultimately absorbed into high-abundance populations. In multi-sample analyses, batch correction may further weaken signals from rare cell states (Argelaguet et al. 2021 ; Haghverdi et al. 2018 ).