A five-gene mitochondrial–lipid metabolism signature for Alzheimer's disease diagnosis identified by machine learning and validated in independent cohorts

Background Mitochondrial dysfunction and lipid metabolism dysregulation represent early key events in Alzheimer's disease (AD), yet integrated transcriptomic signatures combining these two biological processes for AD diagnosis remain poorly defined. Objective This study aimed to construct and validate a mitochondrial-lipid metabolism (MLM) gene-based diagnostic classifier for AD using machine-learning approaches and explore its underlying biological mechanisms. Methods Two public GEO datasets GSE5281 (training) and GSE28146 (independent validation) were included. Candidate MLM genes were retrieved from MSigDB. Differentially expressed MLM genes were screened by limma, followed by two-step feature selection with LASSO regression and random forest. Four machine-learning algorithms were compared. Model performance was evaluated by AUC, calibration, decision-curve analysis and subgroup analysis across AD severity. GSVA and hypergeometric test were applied for mechanistic exploration. Results A five-gene signature ( SDR16C5, EGR1, ASAH2B, HIP1R, CKMT1B ) was finally identified. Logistic regression model achieved 5-fold cross-validated AUC of 0.953 in training cohort and AUC = 0.892 (95%CI:0.768–1.000) in validation cohort. The signature maintained high diagnostic accuracy across incipient, moderate and severe AD stages. Hypergeometric test demonstrated significant MLM-set enrichment (p = 3.85 × 10 −5 ). GSVA revealed close correlations between this signature with oxidative stress, synaptic dysfunction and apoptosis pathways. Conclusions This validated five-gene MLM signature exhibits favorable diagnostic performance for AD and reflects disturbed mitochondrial-lipid metabolic network. It provides a promising transcriptomic biomarker for AD diagnosis.