Correcting for Background Noise Improves Phenotype Prediction from Human Gut Microbiome Data

The ability to predict human phenotypes accurately from metagenomic data is crucial for developing biomarkers and therapeutics for diseases. However, metagenomic data is commonly affected by technical or biological variables, unrelated to the phenotype of interest, such as sequencing protocol or host sex, which can greatly reduce or, when correlated to the phenotype of interest, inflate prediction accuracy. We perform a comparative analysis of the ability of different data transformations and existing supervised and unsupervised methods to correct microbiome data for background noise. We find that supervised methods are limited because they cannot account for unmeasured sources of variation. In addition, we observe that unsupervised approaches are often superior in addressing these issues, but existing methods developed for other omic data types, e.g., gene expression and methylation, are restricted by parametric assumptions unsuitable for microbiome data, which is typically compositional, highly skewed, and sparse. We show that application of the centered log-ratio transformation prior to correction with unsupervised approaches improves prediction accuracy for many phenotypes while simultaneously reducing variance due to unwanted sources of variation. As new and larger metagenomic datasets become increasingly available, background noise correction will become essential for generating reproducible microbiome analyses.