A parsimonious three-gene mitochondrial classifier from frontal cortex for Alzheimer's disease: Nested cross-validation and multicenter external validation

Background Mitochondrial dysfunction is a hallmark of Alzheimer's disease (AD). The frontal cortex is among the earliest and most severely affected regions, yet no AD diagnostic classifier based on frontal cortical mitochondrial genes exists. Objective We aimed to construct a parsimonious diagnostic classifier from frontal cortex mitochondrial gene expression and rigorously validate it across independent cohorts. Methods Four public brain transcriptomic datasets were downloaded from the NCBI Gene Expression Omnibus (GEO) database: GSE48350, GSE5281, GSE44770, and GSE33000. The training set (n = 103) was restricted to superior frontal gyrus samples. After ComBat correction, differentially expressed genes (adj. p < 0.2, |log 2 FC|>0.3) were intersected with MitoCarta 3.0. A 10-fold nested cross-validation (CV) was applied to eliminate pre-selection bias: mitochondrial DEGs were re-selected in each fold, followed by LASSO regression. The final model was externally validated on GSE44770 (n = 230) and GSE33000 (n = 467). Results From 25 candidate genes, LASSO retained three: PRKACA , ACSM3 , and BIK . Nested CV AUC was 0.635 (95% CI 0.527–0.744); external AUCs reached 0.714 (GSE44770) and 0.753 (GSE33000). The 25 genes were enriched in oxidative phosphorylation. Calibration was acceptable, and decision curve analysis indicated net benefit for guiding confirmatory testing. In external cohorts, BIK was upregulated and PRKACA downregulated, while ACSM3 showed a non-significant downward trend. With an optimized threshold, the model achieved a sensitivity of 0.726 and a specificity of 0.682 on the largest cohort. Conclusions A three-gene mitochondrial signature derived from the frontal cortex provides a highly parsimonious, externally reproducible diagnostic tool for AD, with biological plausibility and clinical potential.