Bayesian network-guided sparse regression with flexible varying effects

Area of Science:

• Statistical modeling
• Genomics
• Biostatistics

Background:

• Complex datasets from genomics and imaging studies require advanced feature selection methods.
• Existing regression models may not fully capture the intricate relationships between predictors and covariates.

Purpose of the Study:

• To propose a novel Bayesian method, Varying Effects Regression with Graph Estimation (VERGE), for feature selection in regression.
• To leverage complex data structures by distinguishing predictors and subject-level covariates.
• To infer networks among predictor variables for enhanced feature selection.

Main Methods:

• Developed a varying coefficients modeling framework.
• Employed variable selection spike-and-slab priors for selecting network-linked predictors and modifying covariates.
• Inferred a network among predictor variables to encourage selection of related predictors.

Main Results:

• VERGE demonstrated superior performance over existing methods in simulation studies for both feature selection and predictive accuracy.
• The method successfully identified microbial taxa and their ecological dependencies in a gut microbiome and obesity study.
• Subject-level covariates (sex, diet) were shown to modify the effects of microbiome predictors.

Conclusions:

• VERGE is a powerful Bayesian approach for feature selection in regression, particularly for complex, high-dimensional data.
• The model effectively identifies important predictors and their interrelationships, offering insights into biological systems.
• Application to microbiome data highlights its utility in uncovering complex interactions influencing health outcomes.