Use of wrapper algorithms coupled with a random forests classifier for variable selection in large-scale genomic...

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Large-scale genetic association studies produce high-dimensional data, necessitating variable selection for computational efficiency and to prevent overfitting.
• Traditional data analysis methods struggle with the scale of modern genomic datasets.

Purpose of the Study:

• To introduce and evaluate a novel "wrapper" strategy (SIZEFIT) for efficient variable selection in genetic association studies.
• To identify a minimal set of single-nucleotide polymorphisms (SNPs) that retain predictive accuracy for complex traits like coronary heart disease (CHD) and low-density lipoprotein (LDL) cholesterol levels.
• To compare the performance of the proposed method with a statistical approach (SUMSTAT) and refine SNP subsets using a novel method (FIXFIT).

Main Methods:

• A Random Forests classifier was employed within a "wrapper" variable selection strategy (SIZEFIT).
• The methods were applied to a large case-cohort dataset (Atherosclerosis Risk in Communities) with 2,425 individuals and 4,869 SNPs.
• Novel refinement method (FIXFIT) using Extremal Optimization was developed for signal-containing SNPs.

Main Results:

• Most SNPs could be removed without compromising predictive accuracy, with predictive signals concentrated in a small subset of SNPs (often <100).
• Compact optimal predictive SNP subsets were constructed for CHD (<150 SNPs) and LDL (<300 SNPs).
• Significant overlap was observed between SNP rankings generated by different methods, suggesting robustness.

Conclusions:

• The SIZEFIT wrapper strategy is effective for reducing dimensionality in genome-wide association datasets while preserving predictive power.
• A small number of SNPs contain substantial predictive information for CHD and LDL cholesterol.
• The findings provide practical guidelines for generating compact, predictive SNP subsets from large genetic datasets.