Assessing mediating effects of high-dimensional microbiome measurements in dietary intervention studies.

Habitual diet can influence health-related outcomes directly, but such effects may also be modulated indirectly by gut microbiota. We consider randomized trials and the question to what extent the effect of diet on an outcome of interest is mediated through the gut microbiome or whether there is a diet-microbiome interaction identifying subgroups of individuals who are more susceptible to specific dietary effects. The baseline microbiome by itself may be a modifier of the effects of diet on health. Yet, the high dimensionality of microbiome data requires innovative statistical approaches to identify potential mediating or moderating effects. To motivate our proposal for an appropriate analysis workflow, we consider a randomized trial that investigates the effect of a 4-week vegan diet on the diversity of gut microbiota and branched-chain amino acid metabolism in healthy omnivorous volunteers. To address the challenge of compositional microbiome data, we consider an adaptation of the lasso for penalized estimation of multivariable regression models with a large number of microbiotic taxa. This is plugged into a classical regression mediation effect analysis strategy. The interaction effects are obtained via an approach that can directly estimate them without having to deal with main effects. As a result we obtain signatures comprised of microbiotic taxa with potential mediating and moderating effects. Some taxa no longer show up as mediating, when taking moderating effects into account. Thus, the proposed analysis strategy allows to identify specific mediating effects, while avoiding potential erroneous conclusions, where moderating effects might have believed to be mediating effects.