r2VIM: A new variable selection method for random forests in genome-wide association studies.

Area of Science:

• Genetics
• Bioinformatics
• Statistical genomics

Background:

• Machine learning, specifically random forests (RFs), offers an alternative to single nucleotide polymorphism (SNP) analyses in genome-wide association studies (GWAS).
• RFs provide variable importance measures (VIMs) to rank SNPs by predictive power.
• A lack of clear criteria for selecting SNPs for downstream analysis hinders the application of RFs in GWAS.

Purpose of the Study:

• To introduce a novel variable selection method, recurrent relative variable importance measure (r2VIM), for GWAS.
• To establish objective criteria for selecting relevant SNPs identified by RFs.
• To control the number of false-positive results in GWAS SNP selection.

Main Methods:

• The r2VIM method calculates importance values relative to a minimal importance score across multiple RF runs.
• SNPs with consistently high relative VIMs across all runs are selected.
• The method was evaluated using simulated GWAS data and an experimental GWAS dataset.

Main Results:

• r2VIM effectively controls the number of false-positive SNPs under the null hypothesis.
• The method demonstrates comparable power to logistic regression under a simple alternative hypothesis with independent main effects.
• In an underpowered experimental GWAS dataset, r2VIM identified the true signal while selecting no SNPs, unlike other methods.

Conclusions:

• The r2VIM method provides a robust extension to standard RF for objective SNP selection in GWAS.
• r2VIM enhances the reliability of GWAS by controlling false-positive findings.
• This approach facilitates more accurate and interpretable results from genome-wide association studies.