BPRMeth

BPRmeth models DNA methylation profiles using a Binomial Probit Regression likelihood to extract higher-order spatial features from methylation data and to predict gene expression.

Key Features:

• Binomial Probit Regression likelihood: Models methylation counts across genomic regions using a Binomial Probit Regression likelihood.
• Higher-order feature extraction: Uses probabilistic machine learning techniques to derive higher-order features that capture spatial correlations and methylation profile shapes beyond average levels.
• Gene expression prediction: Applies higher-order features from promoter-proximal regions to construct machine learning models that predict gene expression with improved accuracy over average-methylation-based methods.
• Clustering of promoter-proximal regions: Employs the Expectation-Maximization (EM) algorithm to cluster promoter-proximal regions by higher-order features, revealing five major methylation patterns across different cell lines and implicating regulation beyond CpG islands.

Scientific Applications:

• Gene regulation analysis: Provides a quantitative framework to analyze spatial correlations in DNA methylation patterns relevant to gene regulation.
• Gene expression modeling: Improves predictive models linking promoter methylation patterns to transcriptional activity.
• Epigenetic pattern discovery: Identifies distinct promoter methylation patterns for investigating epigenetic mechanisms.
• Cellular differentiation and disease studies: Enables investigation of methylation pattern changes across different cell types and disease states.

Methodology:

Models methylation counts with a Binomial Probit Regression likelihood, extracts higher-order spatial features using probabilistic machine learning, clusters promoter-proximal regions with the Expectation-Maximization (EM) algorithm, and uses those features in machine learning models to predict gene expression.