Analysis of gut bacterial communities leveraging AI methods unveils novel relations between species in Alzheimer's disease

<jats:p>The gut microbiome has been increasingly implicated in Alzheimer's disease (AD), with studies reporting numerous species- and genus-level differences. These findings established a growing catalog of AD-associated taxa, yet they typically evaluate taxa individually or in small sets rather than across microbial communities. Gut microorganisms act collectively through cross-feeding, competition, and metabolic exchange. Hence, jointly analyzing co-occurring species can reveal community structure and biological insights that a taxon-by-taxon analysis may miss. In 274 stool metagenomes from 119 older adults (18 with AD), we used our Alzheimer's disease Analysis Model (ADAM) framework to run Latent Dirichlet Allocation (LDA) 1,000 times with other bioinformatics tools, decomposing species abundances into communities and aligning them into 22 reproducible ones. Of the 50 species that define these communities, 15 showed an AD-associated shift in relative abundance (10 depleted, 5 enriched; Cohen's d from -0.91 to +0.67, each 95% CI excluding zero), spanning the depletion of Phocaeicola vulgatus (d -0.91, 95% CI [-1.23, -0.59]) and the enrichment of Bacteroides fragilis (d +0.67, 95% CI [0.35, 0.99]), the two ends of an AD-associated balance. Separately, P. vulgatus competitively excludes its congener Phocaeicola dorei. In AlzBiom, an independent amyloid-defined cohort, the exclusion reproduced (within the Bacteroidaceae, r = -0.43 vs -0.57 in GAINS) and held in both control and AD participants, a conserved, disease-independent property. The P. vulgatus/B. fragilis balance also reproduced but more modestly (d = -0.24, permutation p = 0.038), whereas the substitution toward P. dorei did not.</jats:p>