Ranking causal variants and associated regions in genome-wide association studies by the support vector machine and...

Area of Science:

• Genetics and Bioinformatics
• Statistical Genomics
• Machine Learning in Biology

Background:

• Identifying causal genetic variants is crucial for understanding complex diseases.
• Traditional methods like chi-squared statistics rely on single SNP associations.
• Machine learning offers potential for improved variant and region identification.

Purpose of the Study:

• To compare the performance of chi-squared statistics, support vector machine (SVM), and random forest (RF) in identifying causal variants and associated regions.
• To evaluate the impact of different SNP selection cutoffs on the performance of SVM and RF.
• To assess the utility of these methods for real-world applications, including type 1 diabetes genetics and disease risk prediction.

Main Methods:

• Utilized simulated and real genetic datasets.
• Ranked single nucleotide polymorphisms (SNPs) using 1 df chi-squared statistic, SVM, and RF.
• Applied SVM and RF to top SNPs selected based on chi-squared ranking and Bonferroni correction (2r, 5r, 10r cutoffs).

Main Results:

• SVM and RF applied to top 2r chi-squared ranked SNPs improved causal variant and region rankings and increased power in simulated data.
• Performance and stability of SVM and RF rankings decreased with higher cutoffs (5r, 10r).
• Comparison of previously replicated SNPs in type 1 diabetes data showed differences in rankings across methods.

Conclusions:

• SVM and RF demonstrate superior performance over traditional chi-squared statistics for identifying causal variants and associated regions, especially with optimized SNP selection.
• The effectiveness of machine learning methods is sensitive to the stringency of SNP selection criteria.
• The developed methods show promise for genetic association studies and disease risk prediction, with available software and a webserver.