A novel high-dimensional model for identifying regional DNA methylation QTLs

Area of Science:

• Statistics
• Genomics
• Bioinformatics

Background:

• Varying coefficient models (VCMs) provide interpretable insights into dynamic coefficient changes.
• High-dimensional settings pose challenges for existing VCM estimation, particularly in variable selection for nonlinear/varying effects.
• Identifying genetic variants influencing methylation requires methods that handle complex, high-dimensional data.

Purpose of the Study:

• To develop a novel statistical method for variable selection in high-dimensional VCMs.
• To address the limitations of existing methods in identifying covariates with dynamic effects.
• To apply the method for identifying regional methylation quantitative trait loci (mQTLs).

Main Methods:

• Proposing a composite sparse penalty to enforce both sparsity and smoothness on varying coefficients.
• Developing an efficient proximal gradient descent algorithm for high-dimensional predictor spaces.
• Conducting comprehensive simulation studies to assess estimation, prediction, and selection accuracy.
• Implementing an adaptive penalty version for enhanced performance.

Main Results:

• The proposed composite sparse penalty significantly improves performance over sparsity-only methods.
• Smoothness control in the penalty yields superior results in estimation, prediction, and selection.
• The adaptive penalty version offers further performance gains.
• The method successfully identifies regional mQTLs in the CARTaGENE cohort.

Conclusions:

• The composite sparse penalty is an effective approach for variable selection in high-dimensional VCMs.
• Incorporating smoothness control enhances the accuracy and reliability of varying coefficient estimation.
• The developed methodology provides a valuable tool for genetic association studies, particularly for mQTL identification.