Modeling X Chromosome Data Using Random Forests: Conquering Sex Bias.

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Machine learning, including Random Forests (RF), is vital for genetic data analysis.
• Standard RF algorithms exhibit bias when analyzing X chromosome single nucleotide polymorphisms (SNPs).
• This bias leads to inaccurate variable importance estimates, particularly in sex-stratified genetic studies.

Purpose of the Study:

• To develop and validate extended RF methods for accurate X chromosome SNP analysis.
• To address and correct the biased variable importance estimates in standard RF for X-linked genetic variants.
• To provide an unbiased, powerful multimarker approach for genetic association studies incorporating X chromosome data.

Main Methods:

• Proposed extensions to the standard RF algorithm to correctly model X chromosome SNPs.
• Incorporated a stratified approach and an X chromosome inactivation model into RF.
• Applied novel and standard RF methods to alcohol dependence case-control data (SAGE study).
• Conducted simulation studies to compare the performance of different RF approaches.

Main Results:

• Standard RF produced inflated variable importance for X SNPs in alcohol dependence data, even with sex as a covariate.
• The proposed RF extensions yielded results consistent with univariate regression, correctly modeling X chromosome data.
• Simulations confirmed that new RF methods eliminate bias in X SNP importance when sex is associated with the trait.
• The extensions successfully detected both autosomal and X chromosome causal SNPs.

Conclusions:

• Extended RF methods provide an unbiased approach for analyzing genetic data involving the X chromosome.
• These methods enhance the accuracy of variable importance estimation in genetic association studies.
• The developed methods are implemented in the R package "snpRF" for broader accessibility.