A Regression-based Framework for Scalable Pathway-guided Search in Genome-wide Association Studies

Traditional unbiased genome-wide association studies (GWAS) have successfully identified thousands of loci associated with various complex diseases but there is evidence to suggest that many variants were missed at stringent genome-wide thresholds. Fortunately, there is a rapidly increasing amount of prior knowledge in publicly available genomic datasets and biological databases that can be harnessed to enhance the power of discovering SNPs/Genes from existing or new GWAS datasets. For most diseases, many of the identified loci tend to cluster into a few specific biological pathways/networks. From the point of view of disease etiology, such clustering is generally to be expected. This phenomenon can be exploited to conduct a more powerful genome-wide scan that is tailored to identify loci that are interconnected in pathways. We propose a scalable regression-based analytical framework to enable such a pathway-guided GWAS and demonstrate that it provides significant gains in power to detect disease associated SNPs. Our method requires two inputs, namely a) genome-wide summary level data (e.g., SNP p-values) and b) a grouping of genes into biologically meaningful categories (e.g., a database of pathways). It automatically adjusts the input p-values by incorporating the knowledge derived adaptively from the data and the pathways specified. The method involves a regularized logistic regression analysis to derive priors of each SNP and then re-weights the p-values of SNPs so as to maximize overall power of making discoveries. It increases the power to discover SNPs co-clustering into some of these pathways, while maintaining the global type-1 error (FWER) at the desired level. We used whole-genome simulations and summary data from real GWA studies of psoriasis, SLE, coronary artery disease and type-2 diabetes to illustrate the power improvement achieved by pathway-guided search. Our pipeline implemented as an R package can flexibly handle large number of prior annotations possibly derived from multiple databases.