Mediation Analysis of Microbiome Data and Detection of Causality in Microbiome Studies

Abstract

Microbiome research has basically focused on three factors: environment, microbiome, and host. The interactions among these three factors are dynamic and complicated. Three general hypotheses have been developed to detect these interactions. Among these hypotheses, testing the mediated effects of environmental factors and host mediated by microbiome is the most often designed, because mediation analysis of the human microbiome in these dynamic and very complicated relationships could potentially provide insights into the role of the microbiome in health and the etiology of disease and, more importantly, lead to novel clinical interventions by modulating the microbiome.However, microbiome data are high-dimensional, structured as phylogenetic tree, sparse, non-normally distributed, and are often characterized by the presence of a large portion of zero values and hence are skewed to the right and heteroscedastic. Thus, the suitable methods for mediation analysis of microbiome data are rare. Several methods for mediation analysis of microbiome data were just developed in most current years. In this book chapter, we first introduce traditional mediation models and mediation models in omics studies as backgrounds and then focus on describing and reviewing specifically designed mediation models in microbiome studies. Traditional mediation models include two broad types of frameworks for mediation analysis: one is structural equation modeling (SEM)-based mediation analysis, which covers “product method” or “product of coefficients method” and “difference of coefficients method”, respectively. Another is counterfactual-based mediation analysis, which uses “potential outcomes” or “counterfactual outcomes” method.The data features and statistical issues of microbiome studies are more similar to those in other omics studies, such as high dimensionality and sparsity; thus the mediation models from omics studies provide more insights and motivations to develop the mediation models in microbiome studies, such as how to test multiple putative mediators simultaneously using permutation (MultiMed), how to reduce high-dimensional mediators through regularization or penalization (HIMA), and how to transform high-dimensional mediators into low-dimensional and uncorrelated mediators using the spectral decomposition (CausalMM). This book chapter mainly aims to introduce seven specifically designed mediation models in microbiome studies. They are (1) distance-based omnibus test of mediation effect (MedTest), (2) multivariate omnibus distance mediation analysis (MODIMA), (3) causal compositional mediation model (CCMM), (4) isometric log-ratio transformation for microbiome mediation (IsometricLRTMM), (5) sparse microbial causal mediation model (SparseMCMM), (6) mediation analysis for zero-inflated mediators (MedZIM), and (7) nonparametric entropy mediation (NPEM). All these models were developed to target specific data structure and features of microbiome data (e.g., dimensionality, compositionality, sparsity, zero-inflated) through either SEM-based or counterfactual-based mediation frameworks. We complete this chapter with comments on current mediation models for microbiome data analysis and how to understand establishing causality in microbiome studies.