Fig. 2

MIOSTONE provides accurate predictions the host’s disease status. The evaluation was performed on three simulated and seven real microbiome datasets with varying microbial taxa sizes, covering different proportions of taxonomic levels. MIOSTONE is compared against nine baseline methods, divided into two categories: tree-agnostic methods and tree-aware methods. The former category comprises random forest (RF), support vector machine (SVM) with a linear kernel, XGBoost, and multi-layer perceptron (MLP), while the latter includes DeepBiome, Ph-CNN, PopPhy-CNN, TaxoNN, and MDeep. Each model was trained by times using different train-test splits, and reported by the average performance along with 95% confidence intervals. The models’ performances are measured by the area under the precision-recall curve (AUPRC). Because Ph-CNN is not scalable for processing the HMP2 dataset, the result is denoted as N/A. For scientific rigor, the performance comparison between MIOSTONE and any other baseline method is quantified using one-tailed two-sample t-tests to calculate p-values: \(****\ p\text {-value}\le 0.0001\); \(***\ p\text {-value}\le 0.001\); \(**\ p\text {-value}\le 0.01\); \(*:\ p\text {-value}\le 0.05\); \(\text {ns}:\ p\text {-value}> 0.05\)