Statistical models of gene-environment interactions

Thesis

Despite longstanding interest in gene-by-environment (GxE) interactions, the contribution of GxE interactions to complex disease in humans remains poorly characterised. Studies of GxE interaction are statistically challenging due to the inherent high dimensionality of environmental exposures, the increased sample size required to robustly detect GxE effects and the historic lack of large datasets that combine genetic and environmental data. In this thesis I introduce a new approach for GxE interactions with multiple environmental variables on biobank scale datasets called LEMMA (Linear Environment Mixed Model Analysis). LEMMA uses a Bayesian whole genome approach to jointly model SNP marginal effects and SNP GxE interaction effects with an Environmental Score (ES). The ES is a linear combination of multiple environmental variables, and 'interaction weights' used to construct the ES are learnt in the same Bayesian framework. The ES can subsequently be interpreted directly, used in single SNP hypothesis or used for heritability estimation. Using variational inference to fit the model and a distributed computational architecture allows LEMMA to be computationally tractable on Biobank scale datasets.

LEMMA is compared to existing single-SNP approaches that jointly model the effects of multiple environmental variables at each locus. Results suggest that LEMMA has substantially greater power to detect GxE effects when the underlying model assumptions hold.

I have applied LEMMA to model GxE interactions in body mass index, systolic, diastolic and pulse pressure, using 42 environmental variables and approximately 280,000 white British individuals from the UK Biobank. LEMMA uncovered 3 loci with GxE interactions at genome wide significance levels, and estimated that 9.3%, 3.9%, 1.6% and 12.5% of phenotypic variance is explained by GxE interactions, and that rare variants explain most of this variance.

In my final chapter I suggest a new model, referred to as RHE-NLS, that promises to achieve many of the same objectives as LEMMA with improved computational complexity.

Authors: Kerin, M

• DOI: 10.5287/ora-gw8ebr2zm
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Bayesian statistics
• Subjects: Human genetics--Statistical methods