Penalized logistic regression based on L1/2 penalty for high-dimensional DNA methylation data

Area of Science:

• Genomics
• Epigenetics
• Bioinformatics

Background:

• DNA methylation is a critical epigenetic modification influencing gene expression.
• Aberrant DNA methylation patterns, specifically hyper- and hypomethylation at CpG sites, are hallmarks of cancer.
• Identifying disease-associated CpG sites is crucial for understanding disease mechanisms and developing diagnostics.

Purpose of the Study:

• To propose an efficient and accurate computational method for selecting pathogenic CpG sites.
• To utilize the inherent group effect of correlated CpG sites for improved selection.

Main Methods:

• A penalized group constrained logistic regression model was developed.
• The method integrates a L1/2 regularized solver with a central node fully connected network.
• This approach simultaneously enforces sparsity and accounts for the group effect of correlated regression coefficients.

Main Results:

• The proposed method demonstrated superior predictive accuracy and stability compared to existing regularization techniques in simulations.
• Application to over 20,000 CpG sites using ovarian cancer data confirmed higher predictive accuracy.
• The method identified more pathogenic CpG sites with fewer selected sites, outperforming previous approaches.

Conclusions:

• The novel method provides a powerful tool for DNA methylation research.
• It significantly advances the ability to identify pathogenic CpG sites.
• This tool can aid researchers in understanding epigenetic contributions to disease.