NEMoE: a nutrition aware regularized mixture of experts model to identify heterogeneous diet-microbiome-host health interactions

Table 1 Summary of methods for comparison

Method	Input data of identified subcohort	Input data of modeling within each subcohort	Model
sLR		Microbiome	Sparse logistic regression
SVM		Microbiome	Support vector machine
RF		Microbiome	Random forest
sLR K^a	Nutrition	Microbiome	Two-stage sLR with K latent class
SVM II	Nutrition	Microbiome	Two-stage SVM
RF II	Nutrition	Microbiome	Two-stage RF
NEMoE K^b	Nutrition	Microbiome	NEMoE with K latent class
MMMoE^c	Microbiome	Microbiome	RMoE
NNMoE	Nutrition	Nutrition	RMoE
MNMoE	Microbiome	Nutrition	RMoE
Comb-MoE	Microbiome+nutrition	Microbiome+nutrition	RMoE

^aTwo stage sparse logistic regression fitted with two, three four latent classes were denoted as sLR II, sLR III, and sLR IV
^bNEMoE fitted with two, three four latent classes were denoted as NEMoE II, NEMoE III, and NEMoE IV. When not explicitly including the number of latent classes, we refer to NEMoE II
^cOur NEMoE is easy to extend to partition the population with different types of data. We also investigate the different types of data as input of the NEMoE model. Results showed using nutrition to split the population obtained the best performance in our dataset

ISSN: 2049-2618