A Genetic Association Study Comparing Kernel-based Methods, with Application to Crohn's Disease

Abstract

Advancements in high-throughput genotyping technologies have made it easier to identify both rare and common genetic variants in the human genome. With these improvements, our capability to further research the genetic basis of human disease has been greatly enhanced. In fact genetic association studies, which are used to test for association between genetic variation and a phenotype of interest, have become more prominent. Recently, kernel-based statistical methodologies have been proposed for genetic association studies. Kernel approaches first require specification of a kernel function, which maps the degree of genetic similarity between pairs of individuals, followed by the application of a kernel statistic. Many kernel functions have been proposed with strategies ranging from scoring genotype similarity to tree-based approaches. Similarly, many kernel statistics have been proposed such as: Multivariate Distance Matrix Regression (MDMR), Gene trait similarity regression, and Sequence Kernel Association Test (SKAT). There has been no study that has described and compared the performances of all the different combinations of kernel-based association statistics with the different kernel functions. The purpose of this will be to compare the performance of different kernel definitions and kernel-based association statistics. Using a simulation approach, we apply kernel methods to monitor their relative abilities to detect significance in data generated from a construct of true genotype - phenotype association. Under a common causal genetic variant model, we find that power is best when using kernels that score genotypic similarity. Under a multiple rare causal variant model, we find that power is best when using the tree or SKAT kernels. Finally, we also compare performance of kernel-based approaches on real data collected on families having a child with Crohn’s disease. We show that the results depend on the choice of kernel and kernel-based statistic.