Integrative analysis across metagenomic taxonomic classifiers: A case study of the gut microbiome in aging and longevity in the Integrative Longevity Omics Study.

MetaPhlAn4 and Kraken2 classifiers produced different numbers of classified reads and taxa in the ILO study cohort.

• The study applied both MetaPhlAn4 (marker-gene-based) and Kraken2 (k-mer-based) to stool metagenomic samples from participants in the Integrative Longevity Omics (ILO) study.
• The two classifiers differ fundamentally in their approach: MetaPhlAn uses marker genes while Kraken2 uses k-mer matching.
• Despite shared input data, classifier-specific inferences were identified that would be lost when using only one classifier.

Both classifiers captured similar age-associated changes in alpha and beta diversity across cohorts, but variability in species-level alpha diversity was driven by differences between classifiers.

• Both MetaPhlAn4 and Kraken2 detected consistent age-associated diversity patterns across independent cohorts.
• Species alpha diversity showed variability that was attributable to which classifier was used rather than biological signal alone.
• Beta diversity associations with age were broadly consistent across classifiers.

A correlated meta-analysis approach (AdjMaxP) integrating results from both classifiers identified more age-associated taxa than either classifier alone.

• The AdjMaxP approach is designed to account for correlation between results from multiple classifiers applied to the same samples.
• Using AdjMaxP across classifiers in differential abundance analysis captured more age-associated taxa compared to single-classifier analyses.
• 17 taxa were identified as robustly age-associated across both classifiers and across independent cohorts.
• The meta-analytic integration approach was introduced as a novel method for combining results from multiple metagenomic classifiers.

Classifier-specific inferences were identified that would be missed when using only a single taxonomic classifier.

• Some taxa associations with age were detected by only one of the two classifiers (MetaPhlAn4 or Kraken2) and not the other.
• These classifier-specific findings underscore the limitation of relying on a single classifier in microbiome studies.
• The authors note that 'many results are consistent across the two classifiers' but also that 'classifier-specific inferences would be lost when using one classifier alone.'

The study replicated age-associated microbiome findings in an independent cohort to validate cross-classifier consensus results.

• Analyses of taxonomic diversity and relative abundance associations with age were replicated in an independent cohort beyond the ILO study.
• Replication in an independent cohort was used as a criterion for identifying the 17 robustly age-associated taxa.
• Replication across cohorts was combined with cross-classifier consensus to strengthen confidence in reported associations.

Most microbiome studies use a single taxonomic classifier despite known differences between classification approaches, motivating the development of consensus and meta-analytic integration methods.

• MetaPhlAn and Kraken are described as 'two popular methods at the forefront of many studies,' representing marker-gene-based and k-mer-based approaches respectively.
• The authors note 'calls for the development of consensus methods' exist in the field but most analyses still use a single classifier.
• This study introduces both consensus and meta-analytic approaches as practical tools for integrating results from multiple classifiers.
• The study recommends 'employing multiple classifiers' and 'novel approaches that facilitate the integration of results from multiple methodologies.'