Bayesian Variable Selection for Compositional Microbiome Data

Ongoing advancements in microbiome profiling have provided unprecedented insights into the
molecular dynamics of microbial communities, sparking a surge of interest in uncovering the
microbiome’s critical role in human health. Identifying microbial features linked to clinical
outcomes, however, remains challenging due to the high-dimensional, sparse, and compositional nature of microbiome data. Additionally, many microbial taxa, although classified as distinct, may share functional roles, complicating traditional variable selection methods.
To address this, we developed two complementary Bayesian frameworks. Our first
approach, BRACE, provides a unified nonparametric approach for data-adaptive feature
aggregation and selection. By clustering regression coefficients, BRACE identifies groups of taxa with similar functional impacts, yielding interpretable models and effective dimension reduction without relying on external information. In contrast, our second approach, GRACE, investigates the role of external feature similarity graphs. It combines variable selection with outcome-driven graph learning, characterizing how recovery of outcome-relevant networks changes when guided by external, outcome-agnostic graphs such as phylogeny or co-occurrence. Both methods explicitly respect compositionality and have shown strong performance in oral microbiome applications, together offering new strategies for interpretable regression models in compositional data settings.