Penalized variance components for association of multiple genes with traits

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Variance component models are popular for genetic analyses due to their flexibility in analyzing multiple genetic variants and accounting for population stratification.
• Standard maximum-likelihood methods struggle with convergence and statistical testing in exploratory analyses with modest sample sizes and numerous variance components.

Purpose of the Study:

• To develop a penalized-likelihood model to overcome limitations of standard methods for estimating variance components.
• To improve the power of genetic analyses and balance false-positive and false-negative results.

Main Methods:

• Developed a penalized-likelihood model incorporating elastic-net (L1 and L2 penalties) for variance components.
• Utilized simulations to evaluate the model's performance and determine optimal penalty parameter distribution (80% L1, 20% L2).

Main Results:

• Simulations demonstrated increased statistical power using combined L1 and L2 penalties.
• An 80/20 L1/L2 penalty ratio showed a good balance between false-positive and false-negative rates.
• Larger sample sizes improved method performance but increased computation time.

Conclusions:

• The penalized-likelihood elastic-net model offers an effective approach for variance component estimation in genetic analyses.
• The method aids in prioritizing findings for further functional studies, as demonstrated in a DNA methylation and cortisol response study.