Distinguish IBD from non-IBD (control) gut microbiome samples using cross-cohort, pooled shotgun metagenomic species and pathway abundance data, via machine-learning classifiers that are validated not just by pooled cross-validation but by explicit out-of-cohort generalization testing.
Characterize cohort structure before touching disease signal. Quantify how much of the variance in microbial community composition is explained by cohort of origin versus IBD status (PERMANOVA on Bray-Curtis dissimilarity), so that downstream conclusions can be interpreted against that baseline rather than assumed to be disease-specific.
Identify differentially abundant species and pathways between IBD and control, using three statistically independent methods — MaAsLin2 (all-samples model), MaAsLin2 (age-adjusted subset), and a cohort-blocked Wilcoxon rank-sum test — and require 3-way agreement plus direction agreement before treating a feature as a validated hit. This is deliberately conservative: any single test can produce false positives driven by one dominant cohort; requiring convergence across three differently-specified tests is the project’s main defense against that.
Characterize the pathway (functional) layer separately from the species (taxonomic) layer, since a priori it is not obvious that pathway-level function will show the same cross-cohort consistency as species-level composition — this project treats that as an empirical question, not an assumption, and the two layers are analyzed in fully parallel but independent tracks.
Reduce the validated feature sets to compact, ranked panels via Recursive Feature Elimination (RFE, random-forest-based, AUC-optimized), producing three panel sizes per feature type (species and pathway) that trade off predictive performance against parsimony.
Train and compare multiple ML model families (Random Forest, XGBoost, LightGBM, ElasticNet-penalized Logistic Regression) on each feature panel, using nested cross-validation with Bayesian hyperparameter search, so that model choice — not just feature choice — is empirically justified rather than assumed.
Stress-test generalization with leave-one-cohort-out (LOCO) validation. Pooled cross-validation cannot detect a model that has learned cohort-specific rather than disease-specific signal; LOCO — training on three cohorts and testing entirely on a fourth, held out cohort — directly probes this failure mode, following the same design used in Ning et al. 2023 (Ning et al. 2023).
Interpret the winning model, using SHAP values to identify which individual features drive its predictions, and cross-reference those features against the independently validated differential-abundance signature — convergence between an ML model’s learned importances and an independent statistical test is stronger evidence than either alone.