Variable Selection in Whole-Genome Association Studies Using Lasso-Based Techniques

In genome-wide association studies (GWAS), the high-dimensionality of genomic data has raised new statistical and computational challenges. The control of false discovery rate is one of them, which is basically a problem of variable selection. Existing methods for this issue are usually unsatisfactory, as the collinearity and high-dimensionality of genomic data tend to reduce the test power in multiple testing dramatically, and the permutation tests in multivariate regression are usually computationally prohibitive. In this study, we evaluated two Lasso-based methods for their performance in variable selection and computational efficiency. One is to estimate empirical critical values through permutation testing (permLasso), and another is the adaptive Lasso (adaLasso), which is an extension of the ordinary Lasso. These two methods were evaluated in an association analysis with real data for a pork meat quality trait, thawed muscle color b* (TMCOL b*), for records of 1943 pigs genotyped using Illumina Porcine60 SNP Beadchip v2.0. In this study, a subset of 211 candidate markers was obtained using a single run of the ordinary Lasso, and then permLasso and adaLasso were used to remove those false positives in these markers. The results of the two methods are in good agreement. Of the 211 markers, five were statistically significant in permLasso, and seven markers were confirmed by adaLasso. Four markers were in common between the two methods. Both methods are computationally feasible, and adaLasso is relatively more time-efficient.