New directions in genetic association studies

点击翻译按钮获取中文摘要

Overview: The general goal of this proposed research program is development of methods for working with high dimensional data in genetic and genomic studies. This methodological work is highly relevant to colleagues who are increasingly working with datasets containing large sample sizes and many measured variables (large feature space), sometimes including data from different tissues or measurement types. The proposed ideas are a logical continuation of several active projects including my current NSERC DG, loosely organized around three Themes.******Theme 1: Inference and prediction in interaction models with one high dimensional feature set. For numerous human traits, small and approximately additive effects have been precisely estimated for each of thousands of genetic variants (usually single nucleotide polymorphisms (SNPs)). Recently, linear combinations of these SNPs are now being used to create genetic measures that collectively explain quite substantial proportions of trait variance. Here we focus on penalization and variable selection approaches that will optimize sensitivity for finding interactions between a high dimensional feature set (e.g. SNPs) and a single covariate such as an exposure. We will build on past progress in Theme 1 (Bhatnagar, Yang et al. 2018, Bhatnagar, Yang et al. 2018, Jolicoeur-Martineau, Wazana et al. 2018) by improving capabilities to cope with larger datasets and incorporating external information into penalty terms.******Theme 2: Estimating and exploiting constrained and penalized projections. Causal inference methods based on instrumental variables rest on an often violated assumption: that the instrumental variable influences the outcome only through the exposure of interest. However, when using genetic variants as instruments, the variants often demonstrate horizontal pleiotropyinfluencing multiple traitsthereby violating this key assumption. We have developed a variable selection tool that, combined with a constrained linear projection, minimizes horizontal pleiotropy(Jiang, Oualkacha et al. submitted). In Theme 2, we propose to improve this estimator by making it more robust, and to continue this line of research into higher dimensional settings.******Theme 3: Network models for genomic data: We have recently built a model for analysis of network associations in microbiome data where precision matrices vary by group membership (McGregor, Labbe et al. 2018). We are using Laplace priors on the off-diagonal elements of precision matrices to estimate sparse networks. We propose to continue this line of research to allow external annotations to influence the network structure, and furthermore to use concepts in time-varying networks to examine how networks change with covariates. ******Impact: All three themes involve development of statistical methods and software that will not only be useful to researchers in genetics, but will also generate statistical theory elements applicable to multiple domains. ***