Fig. 3

Runtime analysis. a MIOSTONE demonstrates comparable training efficiency to other tree-aware methods. Tree-agnostic methods, with their highly optimized implementations, are the most efficient models for small microbiome datasets. However, their training times escalate significantly faster than those of tree-aware methods as the sample size increases. b MIOSTONE utilizes a fully connected DNN with additional pruning, offering significantly greater efficiency than naive models based on customized taxonomic connections. The setting used by MIOSTONE is marked by \(\bigstar\)