Work

Statistical Methods for Aggregation of Sequence Data and Multiple Testing Correction in Common and Rare Variant Analysis

Year: 2020
Type: dissertation
Abstract: Over the last fifteen years, there have been substantial improvements in how we study the association between trait and genetic variations in the human genome. Genome-wide association studies (GWAS) now routinely test millions of variants in hundreds of thousands of individuals and the advance of genome sequencing technology allows us to examine the role of genetic variants across the full allele-frequency spectrum. However, with these changes come new challenges in analyzing and interpreting genetic results. In this dissertation, we present methods to aggregate sequence data and identify significant associations in common and rare variant analysis. In chapter two, we compare two strategies to aggregate sequence data from multiple studies: joint variant calling of all samples together versus calling each study individually and then combining the results using meta-analysis. Although joint calling is the gold standard, single-study calling can be more appealing due to fewer privacy restrictions and smaller computational burden. We use deep- and low-coverage sequence data on 2,250 samples from the GoT2D study to compare the two strategies in terms of variant detection sensitivity, genotype accuracy, and association power. We show single-study calling to be a viable alternative to joint calling for deep-coverage sequence data but show them to have noticeable discrepancies in rare variant calling and association results for low-coverage sequence data. In chapter three, we revisit the common variant P-value significance threshold of 5e-8 and explore the rates of true and false discoveries that can be expected using less restrictive P-value thresholds and three other multiple testing procedures: Benjamini-Hochberg (BH) and Benjamini-Yekutieli (BY) for controlling false discovery rate (FDR), and Bayesian false discovery probability for controlling Bayesian FDR. Using data from the Global Lipids and GIANT consortia, we show for large sample common variant GWAS that using a less stringent P-value threshold of 5e-7 or use of the BH procedure at target FDR threshold of 5% substantially increases the number of true positive discoveries while only modestly increasing false positive discoveries compared with the 5e-8 threshold. The latter threshold remains appropriate for modest-sized studies or for resource-intensive follow-ups such as constructing animal models where a stringently curated list of significant loci is desired from GWAS. In the chapter four, we propose a Bayesian method for multiple testing correction in rare variant studies that calculates the posterior probabilities using an approximation of the Bayes factor and estimates prior parameters from summary statistics using an Expectation-Maximization algorithm. Using simulations analyses of ~400,000 individuals and ~107 million variants from the TOPMed-imputed UK Biobank study, we show that our Bayesian method discovers more true positive loci than P-value-based methods such as the P-value threshold, BH, and BY procedures at equivalent false positive rates. In addition, we show that the Bayesian method controls empirical FDR among discovered loci. Finally, we estimate the genome-wide significant P-value threshold for testing ~107 million variants from the TOPMed imputation reference panel to be 1e-9.
Language: English
Subfield: Genetics
Open Access status: green
OpenAlex ID: W4389321043
Cites: 107
Cited by:
Related to: 10