Table 3. Challenges presented by microbiome data as input for ML models and the approaches taken by discussed methods to tackle these challenges.

Large feature space; small sample size

MetAML feature selection with Lasso, ENet, or RF n most important

PopPhy-CNN feature selection with novel alg; network regularization

DeepMicro autoencoder for low-dimensionality representation; early stopping

MVIB stochastic probabilistic encoders

MicroPheno shallow subset of 16S k-mers; early stopping; dropout hidden layers

MetaPheno select k-mer counts from 1000 most significant k-mers

Met2Img convert profile to binned image; early stopping

Presence of novel species

MicroPheno & MetaPheno raw sequence input data (k-mers)

Temporal fluctuations in microbe abundances

MetAML include multiple samples from a single test subject

MVIB combine abundance and marker profiles for each sample