Fig. 1

Overview of MIOSTONE. a MIOSTONE designs the neural network with layered architecture, explicitly mirroring the taxonomic hierarchy. Each neuron represents a distinct taxonomic group, while connections between neurons signify hierarchical subordination relationships among these groups. By contrast, a conventional neural network functions as a black box, lacking explicit encoding of knowledge within its architecture. b Each internal neuron in MIOSTONE has the capability to discern whether taxa within the corresponding taxonomic group provide a more effective explanation of the trait when assessed either holistically (i.e., additively) as a group or individually (i.e., non-linearly) as distinct taxa. c MIOSTONE establishes a versatile microbiome data analysis pipeline, applicable to a variety of tasks including disease status prediction, microbiome representation learning, microbiome-disease association identification, and enhancement of predictive performance in tasks with limited samples through knowledge transfer