µGWAS on a Grid Enabling Small Sample Screening for Common Complex Conditions

Genome Wide Association Studies (GWAS) based on traditional univariate statistical approaches, where p-values are computed from individual single nucleotide polymorphisms (SNPs), have identified loci conferring disease risk for some rare diseases, but have had limited success when applied to complex diseases. Multivariate approaches are more appropriate, in principle, but have rarely been used because parametric methods require unrealistic assumptions about linearity, additivity, and independence to be made, and non-parametric methods have not been fully developed due to their high computational demand.

In the spirit of computational biostatistics, we propose a novel, non-parametric, multivariate statistical method (µGWAS), which is based on recent extensions (Morales 2008, http://www.bepress.com/sagmb/vol7/iss1/art19) of a statistical approach, termed u-statistics, which were originally proposed in the 1940s (Hoeffding 1948), but so far mostly applied to univariate and censored data only (Mann-Whitney 1947, Gehan 1965).

When empowered by a grid of hundreds of computers, µGWAS can detect epistasis within diplotypes (neighboring SNPs of unknown phase) as well as between intragenic regions and genes. In an application of this computational statistics approach to several common diseases, we demonstrate its unique ability to pinpoint specific regions within a haplotype block and to detect functionally relevant interactions. While single SNP GWAS is often believed to require (tens of) thousands of subjects, a study of childhood absence epilepsy, for instance, generated biologically plausible hypotheses based on only 185 cases and 824 controls from a publicly available data base.

Having a more powerful multivariate approach available to identify epistasis in common diseases, has implications for the use of HapMap tag SNPs and has potential to advance comparative effectiveness research and personalized diagnostics or treatment.