A generalized distance covariance framework for genome-wide association studies

TL;DR

This paper introduces a new statistical framework for detecting complex dependencies between genetic variants and phenotypes in genome-wide association studies, overcoming limitations of traditional additive models.

Contribution

The proposed method guarantees detection of all dependencies expressed by classical genetic models and extends to regression settings with covariates, with efficient computation and strong statistical power.

Findings

Method accurately detects diverse genetic-phenotype dependencies.

Finite sample p-values are efficiently computed using hypergeometric series.

Simulation studies show superior performance over existing approaches.

Abstract

When testing for the association of a single SNP with a phenotypic response, one usually considers an additive genetic model, assuming that the mean of of the response for the heterozygous state is the average of the means for the two homozygous states. However, this simplification often does not hold. In this paper, we present a novel framework for testing the association of a single SNP and a phenotype. Different from the predominant standard approach, our methodology is guaranteed to detect all dependencies expressed by classical genetic association models. The asymptotic distribution under mild regularity assumptions is derived. Moreover, the finite sample distribution under Gaussianity is provided in which the exact p-value can be efficiently evaluated via the classical Appell hypergeometric series. Both results are extended to a regression-type setting with nuisance covariates, enabling hypotheses testing in a wide range of scenarios. A connection of our approach to score tests is explored, leading to intuitive interpretations as locally most powerful tests. A simulation study demonstrates the computational efficiency and excellent statistical performance of the proposed methodology. A real data example is provided.