Using Two Microbiome Classifiers Instead of One Uncovers Hidden Gut Signals of Aging

Gut microbiome studies of aging depend heavily on the software used to turn raw DNA sequences into lists of microbes. Two widely used tools take different approaches. MetaPhlAn uses curated marker genes and is more specific. Kraken/Bracken uses all sequence data via k-mer matching and is more sensitive but prone to false positives. Most studies pick one tool, which could bias conclusions about which microbes track with age and longevity.

The researchers sequenced stool metagenomes from 220 participants in the Integrative Longevity Omics (ILO) study, including 78 centenarians and 142 of their offspring. They replicated analyses in a public dataset of 348 Han Chinese individuals aged 50-105. Both datasets were run through MetaPhlAn4 and Kraken2/Bracken. They compared alpha diversity (within-sample), beta diversity (between-sample), and differential abundance of species with age. They also introduced AdjMaxP, a correlated meta-analysis that accounts for the same individuals being measured by two classifiers, and benchmarked it by simulation against two alternatives.

The classifiers identified quite different species lists. Kraken2/Bracken found more species (1,044 in ILO versus 787 for MetaPhlAn4), and only about 21-22% of species overlapped within a cohort. The most abundant species were similar across tools, so the disagreement was mainly in low-abundance taxa. Normalized alpha diversity rose with age at higher taxonomic levels (phylum, class, order) in both cohorts and both tools, but not consistently at genus or species level. Richness conclusions were more classifier-dependent. Beta diversity patterns were highly concordant (Procrustes correlation 0.83-0.87), indicating similar community-level shifts with age.

For individual species, each classifier alone found roughly 29-59 age-associated species, with many found by only one tool. AdjMaxP identified 87 age-associated species in ILO and 71 in the Han Chinese cohort. Only 16 and 8, respectively, would have been found by a single-classifier approach, and 13 and 19 were not significant in either tool alone. Of 141 species found in at least one cohort, 17 were age-associated in both. Examples with prior aging links included Mediterraneibacter gnavus, Faecalibacterium prausnitzii, and Anaerostipes hadrus. In simulations, AdjMaxP had high specificity and precision for conserved signals and the best overall area under the curve, at some cost in sensitivity versus one comparator.

The practical message is methodological: conclusions about microbes tied to aging can depend on the tool chosen, and combining tools can improve discovery and robustness. Caveats include that the findings are cross-sectional associations, not causal, and the two cohorts differ in ancestry, diet, and environment. Only two classifiers were tested, no ground truth exists for real samples, and some alpha diversity findings were sensitive to outliers and to how unclassified genome bins were handled.