Abstract
Abstract Alzheimer’s disease is a progressive, currently irreversible, neurodegenerative disorder associated with the accumulation of β-Amyloid (Aβ) plaques between cells and tau (τ) tangles within cells. The gut-brain axis has been implicated in Alzheimer’s disease (AD) pathogenesis, yet experimental results regarding a gut microbial signature associated with AD remain inconsistent. In this study, I reprocessed a publicly uniformly reprocessed four publicly available fecal 16S rRNA cohorts (PRJNA554111, PRJNA489760, PRJNA633959, PRJNA734525; 413 samples, 218 AD / 195 controls) through a common amplicon-sequence-variant pipeline and evaluated cross-cohort generalizability using leave-one-dataset-out (LODO) validation across five machine-learning algorithm families. Cohort of origin was predictable from genus composition with 92.7% accuracy (chance 25%), whereas disease status reached only 0.74–0.81 AUC within studies and 0.57–0.64 AUC under LODO, a consistent generalization gap of approximately 0.17 AUC across all five models. Naive per-cohort standardization did not rescue LODO performance. Of 168 genera tested, 12 showed consistent direction across all four cohorts and survived FDR correction and showed depletion of butyrate-producing taxa in AD. The strongest signal was Lachnospira, detected in 46.8% of AD versus 75.9% of control samples (Fisher p = 1.2 × 10⁻⁹), consistent in direction across all four cohorts. SHAP feature attribution showed only Lachnospira and Akkermansia important in all four folds, corroborating the differential-abundance result. Together these results indicate that technical variation between studies exceeds disease-associated variation in AD microbiome data and that certain taxa, primarily butyrate producers whose prevalence was reduced, replicate across independent cohorts.