Fig. 3

Comparison of NEMoE on simulation dataset and real dataset. a An illustration of a non-separable case where nutrition intake does not show a difference between two nutritional-ecotypes, but each subcohort shows a different relationship between microbiome taxa and health state. b An illustration of a separable case where nutrition intake is significantly different between two nutritional-ecotypes and relationships in each model are similar to the illustration in a. Simulation studies showed that NEMoE can identify both case a and case b. c Receiver operating characteristics curve of different methods (see Table 1) in predicting Parkinson’s disease using LOOCV. NEMoE showed the best LOOCV-AUC (AUC = 0.78). d ROC plot of NEMoE at different taxonomic levels using LOOCV. Genus level showed the best predictive performance (AUC = 0.78)