The gut microbiome is increasingly recognized as a contributing factor in host metabolism, immune function, and energy regulation. Although previous studies have used machine learning (ML) to predict obesity from microbiome data [1–3], the present approach emphasizes model interpretability and reproducibility using SHAP-based feature ranking in publicly available datasets. In this study, interpretable machine learning techniques were applied to publicly available 16S rRNA amplicon data (ENA accession ERP003612) corresponding to individuals from the MetaHIT cohort [6], with the aim of predicting obesity-associated phenotypes from operational taxonomic unit (OTU) profiles. Several classifiers were evaluated (RandomForest, CatBoost, LightGBM, XGBoost, GBM, Logistic Regression, Naive Bayes), class imbalance was addressed via SMOTE within a nested 5 \(\times \) 5 cross-validation framework, and model decisions were interpreted using SHAP. RandomForest achieved the highest ROC AUC ( \(0.678 \pm 0.039\) ), with CatBoost close behind ( \(0.674 \pm 0.056\) ). SHAP analysis identified top OTUs mostly from the phyla Firmicutes and Bacteroidetes, including genera such as Clostridium, Coprobacillus, Odoribacter, and Blautia, and families like Clostridiaceae, Lachnospiraceae, and Ruminococcaceae. This study demonstrates that the application of machine learning to publicly available datasets can reveal predictive microbial signatures that might be overlooked by conventional analytical approaches. The proposed framework facilitates systematic hypothesis generation, selects relevant taxa for subsequent experimental validation, and provides reproducible insights into microbiome patterns associated with obesity.