Identification of feature combinations in genome-wide association studies

Abstract

Association of single nucleotide polymorphisms (SNPs) with traits is the most popular method used in genome- wide association studies. SNPs with high association are often considered as a feasible locus for searching SNP combinations. However, this approach has a potential pitfall: correlated SNPs are usually not good partners to improve associations because their combinations do not enhance the quality of trait prediction. Therefore, a computational approach that could reduce the redundancy of SNPs is required. To solve this issue, an approach to reducing the SNP redundancy is proposed in this study. The feature relevance approach was used to select an optimized feature set which could generate the enhanced prediction per- formance. This approach was applied for the identifi cation of mutations in Mycobacterium tuberculosis strains resistant to drugs. It was found that the combination of 2-4 SNPs could achieve an accuracy range from 65% to 90% to predict resistance for some drugs applied for the tuberculosis treatment.