The impact of genetic structure on sequencing analysis

Genome-wide association studies have made substantial progress in identifying common variants associated with human diseases. Despite such success, a large portion of heritability remains unexplained. Evolutionary theory and empirical studies suggest that rare mutations could play an important role in human diseases, which motivates comprehensive investigation of rare variants in sequencing studies. To explore the association of rare variants with human diseases, many statistical approaches have been developed with different ways of modeling genetic structure (ie, linkage disequilibrium). Nevertheless, the appropriate strategy to model genetic structure of sequencing data and its effect on association analysis have not been well studied.

We investigate 3 statistical approaches that use 3 different strategies to model the genetic structure of sequencing data. We proceed by comparing a burden test that assumes independence among sequencing variants, a burden test that considers pairwise linkage disequilibrium (LD), and a functional analysis of variance (FANOVA) test that models genetic data through fitting continuous curves on individuals' genotypes.

Through simulations, we find that FANOVA attains better or comparable performance to the 2 burden tests. Overall, the burden test that considers pairwise LD has comparable performance to the burden test that assumes independence between sequencing variants. However, for 1 gene, where the disease-associated variant is located in an LD block, we find that considering pairwise LD could improve the test's performance.

The structure of sequencing variants is complex in nature and its patterns vary across the whole genome. In certain cases (eg, a disease-susceptibility variant is in an LD block), ignoring the genetic structure in the association analysis could result in suboptimal performance. Through this study, we show that a functional-based method is promising for modeling the underlying genetic structure of sequencing data, which could lead to better performance.