Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome

Table 4 Performance when predicting metabolite (class-level) abundance from microbiome (genus-level) abundance

Method	Top 10 Acc Corr.Coef	Active links	Stability index
Linear Regression (LR)	0.39	9741	_
+ Sparsity (LASSO)	0.62	898	0.57
+ Non-neg. weights	0.58	685	0.61
CCA	0.61	2601 (E) + 9741 (D)	_
Linear NED	0.59	7150 (E) + 84600 (D)	_
+ Nonlinear activation (NED)	0.63	7150 (E) + 84600 (D)	_
+ Sparsity (Sparse-NED)	0.67	551 (E) + 702 (D)	0.80
+ Non-neg. weights (Nonneg-Sparse-NED)	0.66	*360 (E) + 428* (D)**	0.66

Acronyms: NED neural encoder-decoder; E encoder; D decoder. The stability indices of the fully connected models are naturally 1.0 and not worth including

ISSN: 1471-2164