Four failure modes in detecting pathogenic cell states from single-cell RNA sequencing, and how to avoid them

Single-cell RNA sequencing is routinely used to identify pathogenic cell states that drive disease and serve as therapeutic targets and stratification biomarkers, most often through two simple operations: scoring a cell-state gene signature and quantifying the fraction of cells in the state. Both are easy to run and easy to get wrong. Drawing on a uniform screening pipeline applied to fibrotic, vascular, and autoimmune diseases across many public human datasets, this working paper documents four recurring failure modes, each with a worked example and real numbers. (1) Baseline saturation: per-cell signature scoring fails where resting cells natively express the program, while compositional counting still separates disease from control — shown in kidney fibrosis, where a periostin signature was null (p=0.34) but the myofibroblast fraction gave perfect donor separation (p=0.0143), and in endometriosis and morphea. (2) Pseudoreplication: per-cell tests on a few donors manufacture extreme significance that dissolves under per-donor testing. (3) Subtype-versus-disease confounding: a state can mark a disease subtype without separating disease from control at the patient level, illustrated by the inflammatory synovial fibroblast in rheumatoid arthritis, which is strong at the cell level (CD74, HLA-DRA up) yet fails to separate disease from control per donor even by the field's own annotation because it tracks the leukocyte-rich subtype. (4) Spatial spot-composition inversion: a two-sided cell-state-transition signature nulls or reverses at multi-cell Visium spot resolution because its poles belong to different cell types that mix within a spot, shown for an endothelial-to-mesenchymal transition axis in scleroderma skin. Each remedy follows one principle: the unit of statistical inference must match the biological claim. The paper argues that honestly reported negatives are an underweighted contribution to a single-cell literature that selects strongly for positive associations. Discovery-stage, methods-focused computational study; no new data generated.