Reproducible Analysis of Cross-Site Distribution Shift in High-Dimensional Microbiome Classifiers

This repository contains the complete reproducible analysis pipeline supporting the manuscript: “Site-Level Distribution Shift Dominates Disease Signal and Undermines Cross-Site Generalization in High-Dimensional Microbiome Classifiers.” The study evaluates cross-site portability of Parkinson’s disease (PD) microbiome classifiers across three geographically distinct cohorts (Finland, Malaysia, USA; n = 682). Key findings: • Strict within-cohort cross-validation yields high discrimination (AUC up to 0.944).• Leave-one-cohort-out validation results in substantial degradation (AUC as low as 0.506).• Cohort identity explains 68.9% of compositional variance (PERMANOVA).• PD status explains 0.92%.• Disease-associated effect vectors are nearly orthogonal to the dominant ecological axis (PC1 explaining 53.5% of total variance). The repository includes: • Harmonized genus-level count matrices (108 shared genera)• Centered log-ratio transformed feature matrices• Complete R analysis pipeline (scripts C21–C33)• Within-cohort and leave-one-cohort-out performance outputs• PCA structure analyses and PD effect alignment quantification• All summary tables used in the manuscript All analyses were conducted in R 4.4.2 with deterministic random seeds. This version corresponds to the IEEE JBHI submission version (v1.0-portability).