Powerful and Adaptive Testing for Multi-trait and Multi-SNP Associations with GWAS and Sequencing Data

Abstract

Testing for genetic association with multiple traits has become increasingly important, not only because of its potential to boost statistical power, but also for its direct relevance to applications. For example, there is accumulating evidence showing that some complex neurodegenerative and psychiatric diseases like Alzheimer's disease are due to disrupted brain networks, for which it would be natural to identify genetic variants associated with a disrupted brain network, represented as a set of multiple traits, one for each of multiple brain regions of interest. In spite of its promise, testing for multivariate trait associations is challenging: if not appropriately used, its power can be much lower than testing on each univariate trait separately (with a proper control for multiple testing). Furthermore, differing from most existing methods for single-SNP-multiple-trait associations, we consider SNP set-based association testing to decipher complicated joint effects of multiple SNPs on multiple traits. Because the power of a test critically depends on several unknown factors such as the proportions of associated SNPs and of traits, we propose a highly adaptive test at both the SNP and trait levels, giving higher weights to those likely associated SNPs and traits, to yield high power across a wide spectrum of situations. We illuminate relationships among the proposed and some existing tests, showing that the proposed test covers several existing tests as special cases. We compare the performance of the new test with that of several existing tests, using both simulated and real data. The methods were applied to structural magnetic resonance imaging data drawn from the Alzheimer's Disease Neuroimaging Initiative to identify genes associated with gray matter atrophy in the human brain default mode network (DMN). For genomewide association studies (GWAS), genes AMOTL1 on chromosome 11 and APOE on chromosome 19 were discovered by the new test to be significantly associated with the DMN. Notably, gene AMOTL1 was not detected by single SNP-based analyses. To our knowledge, AMOTL1 has not been highlighted in other Alzheimer's disease studies before, although it was indicated to be related to cognitive impairment. The proposed method is also applicable to rare variants in sequencing data and can be extended to pathway analysis.KEYWORDS adaptive association test; ADNI; default mode network; gene-based test; imaging genetics; multiple traits A LZHEIMER'S disease (AD) (MIM 104300) is the most common neurodegenerative disease, and every 67 sec, someone in the United States develops AD (Alzheimer's Association 2015a). Currently there is no cure for AD, and most cases are diagnosed in the late stage of the disease. It is projected that the number of Americans of age 65 years and older with AD will increase from 5.1 million in 2015 to 13.5 million in 2050, a growth from an estimated 11% of the U.S. senior population in 2015 to 16% in 2050, costing .$1.1 trillion in 2050 (Alzheimer's Association 2015b). To advance our...

Imaging-wide association study: Integrating imaging endophenotypes in GWAS

“…Furthermore, the daSPU test identified 44 and 63 significant genes with a common set of 37 genes in the two analyses. Hence, in spite of the relatively small sample sizes and some differences between the two ADNI cohorts (Kim et al 2016), the two analyses yielded largely overlapping sets of the significant genes.…”

“…by weighting) the SNPs and endophenotypes will largely determine the statistical power of the test. The situation is similar to pathway-based association testing, which critically depends on the unknown association patterns of both the genes (in a pathway) and SNPs (in each gene) (Pan et al 2015), and to gene-based association testing on multiple traits, for which one has to account for possibly varying association patterns of the SNPs across the traits (Kim et al 2016). Accordingly, to maintain high statistical power with various unknown association patterns, as in the above previous works, we use two non-negative integers, γ 1 and γ 2 , to control the degrees of weighting over the SNPs and over the endophenotypes respectively, and the two parameters will be chosen data-adaptively as to be shown later.…”

An Adaptive Test on High-dimensional Parameters in Generalized Linear Models

“…Due to a relatively small sample size and usually small genetic effect sizes, applying a univariate test to the ADNI data failed to identify any SNP passing the genome-wide significance level at 5 × 10 −8 (Kim et al, 2016), and even a much larger meta-analysis of 74,046 individuals only identified very few genome-wide significant SNPs (Lambert et al, 2013). Hence, it is natural to consider possible associations at the pathway or even chromosome level, which may be more powerful through effect aggregation and a reduced burden of multiple testing, and shed light on the underlying genetic architecture.…”

Powerful rare variant association testing in a copula‐based joint analysis of multiple phenotypes

“…Most of these approaches assume a very specific and restricted Gaussian dependence structure between traits while there exist many empirical data sets with non-Gaussian dependencies. There exist some empirical comparisons of multivariate tests for the analysis of common (Kim et al, 2016;Liang et al, 2016;Zhu, Zhang, & Sha, 2015) and rare genetic variants (Broadaway et al, 2016;Dutta et al, 2019;B. Most of the above tests are multi-marker and "multivariate" tests (i.e., testing the association of all variants in a region with all traits jointly), hence they may not have optimal power for testing variants with large effect sizes and when the tested variants are only associated with a few of the tested traits.…”