ABSTRACT
Current interest in the potential for clinical use of new tools for improving human health are now focused on techniques forthe study of the human microbiome and its interaction with environmental and clinical covariates. This review outlines theuse of statistical strategies that have been developed in past studies and can inform successful design and analyses ofcontrolled perturbation experiments performed in the human microbiome. We carefully outline what the data are, theirimperfections and how we need to transform, decontaminate and denoise them. We show how to identify the importantunknown parameters and how to can leverage variability we see to produce efficient models for prediction and uncertaintyquantification. We encourage a reproducible strategy that builds on best practice principles that can be adapted for effectiveexperimental design and reproducible workflows. Nonparametric, data-driven denoising strategies already provide the beststrain identification and decontamination methods. Data driven models can be combined with uncertainty quantification toprovide reproducible aids to decision making in the clinical context, as long as careful, separate, registered confirmatorytesting are undertaken. Here we provide guidelines for effective longitudinal studies and their analyses. Lessons learnedalong the way are that visualizations at every step can pinpoint problems and outliers, normalization and filtering improvepower in downstream testing. We recommend collecting and binding the metadata and covariates to sample descriptors andrecording complete computer scripts into an R markdown supplement that can reduce opportunities for human error andenable collaborators and readers to replicate all the steps of the study. Finally, we note that optimizing the bioinformatic andstatistical workflow involves adopting a wait-and-see approach that is particularly effective in cases where the features suchas ‘mass spectrometry peaks’ and metagenomic tables can only be partially annotated.