AnnoAudit: a marker-based protocol for auditing single-cell atlas annotations reveals annotation-driven artifacts in a widely used traumatic brain injury resource

Abstract

Single-cell atlas annotations are routinely treated as ground truth but rarely validated before use. We present AnnoAudit, a marker-based audit protocol that combines four convergent checks - marker scoring, unsupervised clustering, margin-gated module scoring, and applicability-gated pretrained models - into a composite Annotation Contamination Score (ACS), plus an independent-gene discrimination step that distinguishes genuine mis-assignment from ambient-RNA detection artifacts, and a trajectory-correlation fingerprint tracing suspicious signals to their cell type of origin. Applied to CEREBRI (GSE269748), a widely used single-cell TBI atlas (73 citations; 65 citing works; six re-analysis studies, none of which validated the annotations), the audit shows that contamination is systematic: the official "glutamatergic neuron" label is 97.8% non-excitatory by markers, with at least 67.4% of its cells independently confirmed as non-neuronal (microglia-dominated); the GABAergic (46.6%), OPC (66.1%), and astrocyte (37.7%) labels are likewise contaminated by marker-confirmed microglia, whereas the microglial, oligodendrocyte, endothelial, and pericyte labels are largely clean (85-96% self-confirmed). The contamination is state-dependent: astrocyte and OPC labels are largely correct in uninjured controls (76.0% OPC self-marker; 18.5% discrimination-confirmed contamination for the astrocyte label) but collapse in the acute 24-h window (14-16% self-marker; 57-71% microglia), recovering by 6 months - so the annotation failure concentrates precisely in the window of maximal injury response. This temporal structure generates coherent false signals - a biphasic trajectory for 40 of 307 ion-channel genes, a KCNC3-specific OXPHOS signature, and an inverted KCNC3 trajectory at 7 days - whereas the corrected response is a sustained acute KCNC3 up-regulation conserved across three independent datasets and three injury models. Simulation-calibrated ACS is 82.6% for CEREBRI. In a human ALS atlas (GSE330130), an ambient-aware re-analysis shows that the official neuron labels are largely correct (1.8-6.1% contamination after discrimination), confirming that the audit does not flag well-annotated resources, and that marker-only estimates on snRNA data with high ambient RNA must be interpreted with the discrimination step. AnnoAudit needs only the deposited count matrix and a canonical panel; we propose it as a routine quality step for atlas-based re-analysis.