CEP-IP with CMRF in SEA-AD Microglia: A Mitochondrial Co-expression Module Linked to Cognition

Single-nucleus RNA-seq (snRNA-seq) studies of Alzheimer's disease (AD) routinely filter nuclei by mitochondrial percentage (%MT), yet whether these reads carry information about cognition is rarely examined. In the Seattle Alzheimer's Disease Brain Cell Atlas (SEA-AD) microglia snRNA-seq dataset, a transcriptome-wide screen identified a module of MT-CO3 and eight co-expressed mitochondrial genes. Three microglial populations were analyzed: PVM_3, M4 (Mic.13-like), and PVM_1, with a housekeeping-gene control. The CEP-IP pipeline with a copula monotone recovery filter (CMRF) used per-cell Spearman and Kendall scores, an inflection point (IP), and a cross-validated copula space boundary to isolate MT-CO3 Module cells. The donor percentage of Module cells past the IP (Module post-IP%) was significantly associated with worse cognition (MMSE and CASI; FDR<0.05) in all three populations. In PVM_3 and M4, these associations remained significant after adjustment for tissue quality, plaque burden, tangle burden, and contamination, and after doublet removal. PVM_3 and M4 Module post-IP% repeated in each donor's regions beyond chance, and no per-cell %MT cutoff reproduced the Module cells. In two external datasets without cognitive scores, the MT-CO3 module and its IP replicated. However, at the donor level, Module post-IP% could not be separated from mean %MT, and the source of the mitochondrial transcripts (e.g., retained cytoplasm or a true change in mitochondrial content) remains unresolved. Routine %MT filtering in AD snRNA-seq may therefore remove cells that carry information about cognition. The R package cepip, the pipeline code, and a software container that reproduced the findings are publicly available.