A0876
Title: Robust high-dimensional variable selection for GWAS via the Bayesian group lasso
Authors: lanxin li - University of Edinburgh (United Kingdom)
Mayetri Gupta - University of Glasgow (United Kingdom) [presenting]
Vincent Macaulay - University of Glasgow (United Kingdom)
Indranil Mukhopadhyay - University of Nebraska - Lincoln (United States)
Abstract: Genome-wide association studies (GWAS) are a powerful tool for exploring the connections between human disease and genetic variation. The accuracy and reproducibility of association detection in such studies are typically limited due to the weakness of association signals and local correlations between genetic variants. Set-based association tests often tend to be used for their ability to aggregate weak signals and alleviate multicollinearity issues. Statistical methods proposed for set-based association analyses, however, mainly focus on testing the marginal association of each set with a trait at a time, ignoring information across the genome. To improve detection accuracy, the proposal is to examine a set of groups of correlated genetic variants simultaneously via a Bayesian modelling framework, adapting ideas from Bayesian group lasso for high-dimensional regression. In the proposed model, a double-shrinkage hierarchical prior captures the sparsity pattern in GWAS, while a population-based Markov-chain Monte Carlo sampler is used for efficient posterior sampling. The implementation is further sped up by a group-based split-and-merge model-fitting strategy. The model appears to be powerful in locating true association signals both at a group and individual variant level in simulations and a GWAS from an Alzheimer's disease study, and offers improved performance in terms of sensitivity and precision compared to other existing methods.
