2018/11/08 by Hélène Ruffieux, Ruffieux, Hélène, A. C. Davison +11 · 1 citation
Biochemistry, Genetics and Molecular Biology · #Applications (stat.AP) #FOS: Computer and information sciences #Gene expression and cancer classification #Genetic Mapping and Diversity in Plants and Animals #Genetic and phenotypic traits in livestock
paper · pdf · doi:10.48550/arxiv.1811.03334
openalex publication_date 2018/11/08 · openalex created_date 2022/08/02 · openalex updated_date 2026/07/28
We tackle modelling and inference for variable selection in regression\nproblems with many predictors and many responses. We focus on detecting\nhotspots, i.e., predictors associated with several responses. Such a task is\ncritical in statistical genetics, as hotspot genetic variants shape the\narchitecture of the genome by controlling the expression of many genes and may\ninitiate decisive functional mechanisms underlying disease endpoints. Existing\nhierarchical regression approaches designed to model hotspots suffer from two\nlimitations: their discrimination of hotspots is sensitive to the choice of\ntop-level scale parameters for the propensity of predictors to be hotspots, and\nthey do not scale to large predictor and response vectors, e.g., of dimensions\n103-105 in genetic applications. We address these shortcomings by\nintroducing a flexible hierarchical regression framework that is tailored to\nthe detection of hotspots and scalable to the above dimensions. Our proposal\nimplements a fully Bayesian model for hotspots based on the horseshoe shrinkage\nprior. Its global-local formulation shrinks noise globally and hence\naccommodates the highly sparse nature of genetic analyses, while being robust\nto individual signals, thus leaving the effects of hotspots unshrunk. Inference\nis carried out using a fast variational algorithm coupled with a novel\nsimulated annealing procedure that allows efficient exploration of multimodal\ndistributions.\n