A meta-refined human Alzheimer's disease-associated gene subset shows partial mouse-model pathway correspondence and limited cross-cohort machine-learning transportability.

INTRODUCTION: Common Alzheimer's disease (AD) mouse models are widely used, but their molecular correspondence with human AD remains uncertain. METHODS: We analyzed 15 post mortem human brain datasets, four human non-brain or in vitro sensitivity datasets, and nine AD-related mouse-model molecular-profiling datasets; GSE222494 was analyzed separately for single-nucleus localization. Refined107, a meta-refined 107-gene human AD-associated gene subset, was evaluated using recurrence-matched control sampling, independent pathway-level comparison, within cohort out-of-fold (OOF) classification, and ordered cross-cohort transportability. RESULTS: Refined107 comprised 37 Tier 1 and 70 Tier 2 genes and showed greater statistical consistency than recurrence-matched controls. Functional annotation implicated inflammatory, blood-brain barrier, metabolic, neuronal, and developmental programs. Among 636 recurrent human pathways, 226 showed same-direction support in at least one evaluated mouse-model family, indicating partial and model-dependent human-mouse correspondence. Post-selection within-cohort OOF areas under the curve (AUCs) were 0.833 to 0.980, whereas ordered cross-cohort AUCs were 0.288 to 0.720. DISCUSSION: Refined107 represents a statistically consistent, meta-refined human AD-associated gene subset rather than a comprehensive disease signature. Human-mouse pathway correspondence was partial and model-dependent, and within-cohort classification did not translate into robust cross-cohort transportability.