A single-cell cell-of-origin audit of the lncRNA RMST: bulk expression under-represents its broad neural expression and over-attributes a liver biomarker.

Long non-coding RNAs (lncRNAs) are widely nominated as biomarkers from bulk-tissue data, which cannot reveal which cell type carries a transcript or whether a disease change is cell-intrinsic or compositional. Using the lncRNA RMST, I show that bulk expression can misrepresent where a lncRNA resides in both directions, and that cell-of-origin attribution should precede biomarker claims. I combined GTEx with > 40 million single cells and nuclei from the CZ CELLxGENE Census. (i) Under-representation: at single-cell resolution RMST is most broadly detected in brain (17.2% of ~ 29 million cells; 35% of neurons, up to 74% of dopaminergic neurons), confirming its neural role, yet bulk brain expression is low (0.7 TPM) because RMST is broadly but lowly expressed and nuclear-enriched. (ii) Tissue-dependent peripheral carriers: across six non-brain tissues RMST-positive cells are 85.5% stromal/epithelial (hepatocyte 0.02%), and the dominant carrier switches by tissue (cardiac/vascular fibroblasts vs renal/pulmonary/hepatic epithelium); in heart, fibroblasts carry it (79-91% of positive cells), but a disease per-cell change is not robust (bootstrap CI spans zero; within-study flat-to-negative), so I report attribution, not direction. A broad control (KCNQ1OT1: 6.2 vs RMST 0.5 lineages/tissue) confirms RMST's restriction is gene-specific. (iii) Over-attribution: RMST is undetected in all 411,742 diseased-liver nuclei (NEAT1 and MALAT1: ~ 90% of the same nuclei) and near-absent across liver cells, so a reported liver biomarker - including my own prior report - is unattributable. Cell-of-origin attribution, which both rescues an under-counted neural transcript and retracts an over-counted liver biomarker, belongs in the standard workflow for lncRNA biomarkers.