BAC

BAC identifies enriched genomic regions in ChIP-chip data by using a Bayesian hierarchical model to detect loci bound by DNA-binding proteins while accounting for spatial dependence and probe-specific variance.

Key Features:

• Bayesian hierarchical modeling: Captures dependencies among neighboring probes without assuming fixed characteristics such as bound-region length.
• Spatial dependence integration: Incorporates spatial dependence between probes to reflect correlation along the genome.
• Exchangeable prior for variances: Allows different variance levels across probes and mitigates the impact of extreme empirical variances to handle probe outliers.
• Parameter estimation via MCMC: Uses Markov chain Monte Carlo for parameter estimation and bases inference on the joint posterior distribution.
• Posterior probability-based detection and FDR estimation: Identifies bound regions using posterior probabilities of neighboring probes and provides well-calibrated estimates of false discovery rate.

Scientific Applications:

• Transcription factor binding detection: Detects transcription factor–bound regions from ChIP-chip experiments to map protein–DNA interactions.
• Validation and comparative performance: In two publicly available ChIP-chip datasets containing 18 experimentally validated regions, BAC outperformed Wilcoxon's rank sum test, TileMap, HGMM (Hierarchical Gaussian Mixture Model), and MAT (Model-based Analysis of Tiling-arrays), showing greater robustness to probe outliers.
• Simulation-based assessment: Simulation studies demonstrated enhanced statistical power under various scenarios and resilience to model misspecification.

Methodology:

Employs a Bayesian hierarchical model that integrates spatial dependence among neighboring probes, uses an exchangeable prior for probe variances, performs parameter estimation with MCMC, and conducts inference from the joint posterior distribution with posterior-probability-based detection and FDR estimation.