Deciphering the Influence of Gut Microbiome Disruption on Human Health Through Multi-Algorithm Approaches

Introduction:: The human gut virome modulates bacterial communities and may contribute to dysbiosis, yet diagnostic utility is underexplored, and legacy taxonomy complicates interpretation. Methods:: We analysed a curated gut virome cohort with health status, age, and 28 legacy viral families mapped to current ICTV placements. Abundances were treated as compositional with zero handling and centred log-ratio transform. Extra Trees, Random Forest, and LightGBM were evaluated using nested, repeated, stratified cross-validation. Class imbalance was handled with SMOTE-NC applied strictly within training folds. Explainability used SHAP. Association-rule mining followed light feature reduction. A compact Self-Organising Maps (SOM) module supported topology-aware visual analytics. Robustness was probed on a condensed metagenomic dataset. Results:: Best single models achieved accuracy 0.86-0.87, AUC 0.92-0.94, and MCC 0.62-0.65. A hard-voting ensemble reached an accuracy of 0.866. SHAP consistently highlighted Microviridae and Caudoviricetes-associated groups (T7-like, T4-like) as major contributors. SOMs revealed coherent departures from putative homeostasis that aligned with supervised predictions. On the metagenomic dataset, accuracies were 0.74-0.78 with AUC 0.84–0.87. Discussion:: Composition-aware modelling with interpretable ensembles identifies a compact, stable set of virome features associated with health status. The SOM module quantifies departures from cohort homeostasis and agrees with SHAP and association rules on the directionality of key signals. Conclusion:: An interpretable, leakage-safe multi-algorithm workflow predicts health state from virome composition with competitive performance and transparent drivers. Findings are associative and dataset-specific. External prospective validation, disease-specific multiclass modelling, and integrated virome-bacteriome analyses are priorities.