SNP selection and classification of genome-wide SNP data using stratified sampling random forests

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Genome-wide association studies (GWAS) generate high-dimensional data with many irrelevant single-nucleotide polymorphisms (SNPs).
• Standard random forest methods with default parameters may select suboptimal feature subspaces, reducing accuracy in GWAS.
• Exhaustive search for optimal parameters in random forests is computationally intensive for high-dimensional GWAS data.

Purpose of the Study:

• To propose a novel stratified sampling method for feature subspace selection in random forests for high-dimensional GWAS data.
• To enhance the accuracy and efficiency of identifying disease-associated SNPs in complex diseases.
• To provide a computationally feasible alternative to exhaustive search for optimal random forest parameters in GWAS.

Main Methods:

• Developed an equal-width discretization scheme to group SNPs based on informativeness.
• Implemented a stratified sampling procedure to select an equal number of SNPs from each group for feature subspaces.
• Applied the method to Parkinson's and Alzheimer's case-control GWAS datasets.

Main Results:

• The proposed stratified sampling method generated more accurate random forests with lower error bounds compared to standard methods.
• Demonstrated effectiveness on large-scale Parkinson's and Alzheimer's GWAS datasets.
• Identified potentially relevant genes associated with neurological disorders in the Parkinson's data.

Conclusions:

• Stratified sampling is an effective strategy for feature subspace selection in high-dimensional GWAS data.
• The method balances the need for informative SNPs with computational efficiency, improving random forest performance.
• This approach offers a promising tool for genetic association studies and biological discovery.