A Penalized Mixture Model Approach in Genotype/Phenotype Association Analysis for Quantitative Phenotypes

A mixture normal model has been developed to partition genotypes in predicting quantitative phenotypes. Its estimation and inference are performed through an EM algorithm. This approach can conduct simultaneous genotype clustering and hypothesis testing. It is a valuable method for predicting the distribution of quantitative phenotypes among multi-locus genotypes across genes or within a gene. This mixture model's performance is evaluated in data analyses for two pharmacogenetics studies. In one example, thirty five CYP2D6 genotypes were partitioned into three groups to predict pharmacokinetics of a breast cancer drug, Tamoxifen, a CYP2D6 substrate (p-value = 0.04). In a second example, seventeen CYP2B6 genotypes were categorized into three clusters to predict CYP2B6 protein expression (p-value = 0.002). The biological validities of both partitions are examined using established function of CYP2D6 and CYP2B6 alleles. In both examples, we observed genotypes clustered in the same group to have high functional similarities. The power and recovery rate of the true partition for the mixture model approach are investigated in statistical simulation studies, where it outperforms another published method.