Studies of comorbidities and genomic risks of cardiovascular diseases in the UK Biobank cohort

Abstract

Congenital heart disease (CHD) is the most common congenital malformation seen in live births. Vast improvements in the care of CHD patients over recent decades has given rise to significant growth in the population of adults with CHD. As such, there is an urgent need to understand the burden of comorbidity in this nascent population, in order to inform later-life care. Recent studies have demonstrated that adults with even mild CHD defects are at an increased risk of a range of later-life comorbidities, including several which manifest outside of the cardiovascular system. However, the mechanisms underlying these comorbid associations remain poorly understood, and more work is needed to fully characterise the spectrum of comorbidity in mild CHD cases. This project used the UK Biobank resource to explore novel comorbidities of mild CHD, before comparing the distribution of comorbidity manifesting across different body systems in CHD cases, with cases of other congenital disorders. Additionally, a series of complementary genome-wide approaches including polygenic risk score (PRS) analyses, and approaches based on LD-score regression and genomic restricted maximum likelihood (GREML), were used to investigate whether shared genetic aetiologies could underlie the increased risk of several known comorbidities of mild CHD. The recent decade has seen proliferation in the use of PRSs, with considerable investment made into studies focused on investigating the use of PRSs as disease screening and risk stratification tools. In the context of CHD, highly predictive PRSs could significantly improve our ability to assess the risk of recurrence in the children of those affected, which is currently difficult in the majority of cases. Whilst PRSs have been shown to provide considerable utility in aetiological studies, recent results have cast doubt upon their use as predictive tools, demonstrating that significant methodological improvements are likely required if PRSs are to provide real clinical utility. In light of this, a handful of recent studies have explored the use of deep learning approaches for the estimation of PRSs, which could improve upon current PRS methodologies that are based on linear models via improved modelling of epistatic effects across the genome. However, attempts to use deep learning for PRS estimation have yielded mixed results, in particular when trained using larger sets of genomic variants. As such, this project aimed to further explore the use of deep learning for PRS estimation, with a focus on facilitating the incorporation of large numbers of variants in training without loss of performance.

Date of Award	21 Mar 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Bernard Keavney (Main Supervisor) & David Talavera (Co Supervisor)