Multi-trait methods for genetic association testing

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

The early stages of the genome-wide association study (GWAS) era were dominated by studies focusing on single phenotypes, while in recent years there has been growing interest in multi-trait GWAS. A wide variety of multi-trait GWAS methods have been developed, but publications introducing new methods are highly inconsistent in their evaluation of method performance, obscuring their relative merit. Facilitated by burgeoning national biobank resources, multi-trait analyses are set to become more routinely applied, making understanding their relative performance increasingly important.
We develop a simulation framework to model the complex networks underlying multivariate genetic epidemiology. We exploit our simulation framework to perform a comprehensive comparison study of the leading multi-trait GWAS methods, providing a web application and open-source software program implementing our simulation framework for further benchmarking of multi-trait GWAS methods.
Motivated by our comparison results, we develop novel methodology and present a series of multi-trait analyses. We perform multi-trait genome-wide analyses on publicly available GWAS summary statistics on 19 traits – metabolic, anthropometric and psychiatric. We develop and apply two summary statistic methods: one that has increased power to detect pleiotropic effects on multiple traits, and one that is more powerful for detecting heterogeneous genetic effects.
Polygenic risk scores (PRS) are now a commonly used tool for performing phenotype prediction from genetics, assessing the genetic aetiology underlying diseases, and testing for shared genetic aetiology among traits. Using UK Biobank data, we explore the predictive ability of PRS computed across multiple traits for Major Depressive Disorder (MDD). The MDD PRS itself has so far offered modest prediction of MDD case/control status; we explore the use of PRS built on traits correlated with MDD to improve predictive ability. We build main effect and interaction models, using both AIC and BIC stepwise variable selection, and cross-validation, to establish the most predictive models.
Date of Award2018
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorPaul O'Reilly (Supervisor) & Cathryn Lewis (Supervisor)

Cite this

'