Gut Microbiota and Metabolic Pathway Signatures for Inflammatory Bowel Disease Identified via Subject-Stratified Random Forest Based on the Longitudinal HMP2 Cohort.

Background

Inflammatory bowel disease (IBD) is characterised by severe intestinal microbial dysbiosis. Most machine learning diagnostic models built on the longitudinal HMP2 cohort suffer serious data leakage from random sample-level cross-validation splitting, which leads to artificially inflated AUC values. Additionally, incomplete reporting of microbial preprocessing, random forest hyperparameters and multi-dimensional evaluation metrics reduces the reproducibility of existing research.

Methods

We re-analysed the public HMP2 (IBDMDB) longitudinal metagenomic dataset containing 130 unique subjects (103 IBD/27 healthy controls) and 1627 longitudinal faecal samples. Raw 585 species were filtered by a minimum relative abundance of 1 × 10-5 and sample prevalence ≥20%, retaining 89 taxa; all 1135 metabolic pathways were retained. CLR transformation was applied to compositional abundance data. We performed Wilcoxon differential testing with Benjamini-Hochberg FDR correction, alpha/beta diversity analysis, and three random forest models (filtered species, all FDR-significant pathways, strictly filtered pathways). Critical improvements included subject-ID-stratified 5-fold cross-validation repeated 5 times, within-fold training-set-only feature importance calculation, and class weighting to balance unbalanced IBD/control samples. PERMANOVA with subject stratification and PERMDISP dispersion test were implemented with 999 fixed-seed permutations.

Results

All four alpha diversity indices were significantly lower in IBD patients (all p < 0.0001). Subject-stratified PERMANOVA showed disease status only explained 1.18% of total Bray-Curtis community variance (R2 = 0.0118, p = 1); PERMDISP detected significant group dispersion heterogeneity (p = 0.027). We identified 63 differentially abundant species and 695 perturbed pathways at FDR < 0.05. Canonical butyrate producers Faecalibacterium prausnitzii and Roseburia hominis showed no significant inter-group differences. Bootstrap 1000-resampling AUC 95% CIs indicated moderate classification performance: species model (0.626-0.705, mean AUC = 0.665), all-significant-pathway model (0.645-0.712, mean AUC = 0.679), strict-pathway model (0.620-0.685, mean AUC = 0.654). Alistipes putredinis and peptidoglycan biosynthesis I were the top taxonomic and pathway biomarkers, respectively.

Conclusions

This study established a leakage-free machine learning pipeline for longitudinal microbiome cohorts via subject-level cross-validation splitting. The moderate AUC values eliminate false high performance caused by sample leakage, and we provide reliable candidate microbial and metabolic biomarkers for IBD. Restricted by single-cohort internal validation and unadjusted medication confounders, these markers still require independent multi-centre external verification before clinical translation.