Statistical challenges of high-dimensional methylation data.

Area of Science:

• Epigenetics
• Bioinformatics
• Statistical Genetics

Background:

• High-throughput DNA methylation profiling generates beta values (proportions 0-1) with non-standard distributions.
• Common statistical methods are often unsuitable for these asymmetric, bimodal, and heteroscedastic beta values.
• Logit transformation to M-values may not fully resolve distribution issues, retaining bimodality and asymmetry.

Purpose of the Study:

• To address the statistical challenges in DNA methylation data analysis.
• To review and compare methods for univariate screening of differential methylation.
• To guide researchers in selecting appropriate statistical approaches for epigenetic data.

Main Methods:

• Overview and discussion of recently proposed statistical methods.
• Comparison of parametric methods (linear, beta regression) and nonparametric methods (rank-based regression).
• Evaluation of methods for identifying differential methylation while adjusting for confounders.

Main Results:

• Standard statistical approaches are often inappropriate for DNA methylation data (beta values).
• Logit transformation (M-values) does not always normalize data distribution, preserving heteroscedasticity.
• Various parametric and nonparametric methods offer potential solutions for analyzing methylation data.

Conclusions:

• Researchers must be aware of the specific statistical properties of DNA methylation data.
• Appropriate statistical methods are crucial for accurate identification of differential methylation.
• This work provides a guide to navigating the complexities of methylation data analysis.