BGmix

BGmix identifies differentially expressed genes using fully Bayesian hierarchical mixture models to estimate differential effects and false discovery rates in gene expression data.

Key Features:

• Bayesian hierarchical model: Employs a hierarchical Bayesian framework with a 3‑component mixture prior that classifies genes as over‑expressed, under‑expressed, or non‑differentially expressed.
• Mixture prior on differential effects: Represents differential effects via a mixture prior to capture multiple classes of expression change.
• Exchangeable gene variances: Models gene variances as exchangeable across genes to accommodate heterogeneity in variance.
• Fully Bayesian estimation: Estimates the proportion of differentially expressed genes and mixture parameters within a fully Bayesian inferential framework.
• False discovery rate estimates: Produces estimates of false discovery rates for assessing reliability of identified differential expression.
• Parametric flexibility: Supports different parametric families for mixture components to allow alternative component distributions.
• Predictive model checks: Incorporates predictive model checks to guide selection of mixture priors, with evidence from Affymetrix knockout versus wildtype mouse data that allowing extra variability around zero better fits the data than a point‑mass null.

Scientific Applications:

• Knockout studies: Analysis of differential expression between knockout and wildtype samples, including Affymetrix microarray data.
• Cancer genomics: Identification of genes with altered expression in cancer versus control conditions.
• Developmental biology: Detection of gene expression changes across developmental stages or conditions.

Methodology:

Uses a hierarchical Bayesian model with a 3‑component mixture prior on differential effects, exchangeable gene variances, fully Bayesian estimation of proportions and mixture parameters, support for alternative parametric families for mixture components, and predictive model checks (including experiments on Affymetrix knockout/wildtype data showing preference for extra variability around zero over a point‑mass null).