Another look at microbe-metabolite interactions: how scale invariant correlations can outperform a neural network

Abstract

Many scientists are now interested in studying the correlative relationships between microbes and metabolites. However, these kinds of analyses are complicated by the compositional (i.e., relative) nature of the data. Recently, Morton et al. proposed a neural network architecture called mmvec to predict metabolite abundances from microbe presence. They introduce this method as a scale invariant solution to the integration of multi-omics compositional data, and claim that "mmvec is the only method robust to scale deviations". We do not doubt the utility of mmvec, but write in defense of simple linear statistics. In fact, when used correctly, correlation and proportionality can actually outperform the mmvec neural network.