Bayesian network-guided sparse regression with flexible varying effects.

This article presents VERGE (Varying Effects Regression with Graph Estimation), a novel Bayesian hierarchical regression method designed for feature selection in high-dimensional data with complex network structures, such as genomics and microbiome studies. VERGE distinguishes between predictors and subject-level covariates that modulate predictor effects, employing Gaussian process priors for flexible varying coefficients and spike-and-slab priors for sparsity in both predictors and covariates, while simultaneously inferring a network among predictors via a Gaussian graphical model. Simulation studies demonstrate that VERGE outperforms existing methods in both predictor and covariate selection accuracy and predictive performance, particularly by leveraging network information to detect subtle effects. An application to gut microbiome data related to obesity identifies microbial genera and their ecological dependencies, revealing how sex and dietary intake modulate microbiome influences on body mass index. The authors note that VERGE is adaptable to other response types and longitudinal data, with potential extensions to incorporate varying covariate sets per predictor.