'Unlocking therapeutic innovation in heart failure through genomic data science

Existing classification systems do not adequately describe the diversity of human disease, nor do they account for common mechanisms across diseases. Since underlying mechanism often predicts therapeutic efficacy, a mechanism-based taxonomy of disease is required to unlock therapeutic progress. A data-science approach provides the framework to address this unmet need by integrating diverse patient data from genotype to disease outcomes. Heart failure (HF) is a complex clinical syndrome for which rising prevalence stands in contrast to a lack of effective therapeutic options. This project seeks to reclassify HF patients according to underlying causal mechanisms through analysis of harmonised patient data from HDR London, UK Biobank and the HERMES Consortium (a major global collaboration in HF genomics founded by the applicant) using emerging genome bioinformatics and unsupervised learning methods. The overall aim is to identify and validate molecular pathways as candidate targets for novel therapeutic intervention and to provide an exemplar within HDR UK for novel approaches to the study of complex disease.

Objective 1. Define the contribution of common genetic variation to risk of HF and clinically established sub-phenotypes.
A genome-wide association meta-analysis (GWAMA) of >32,000 HF cases from 25 studies has been completed by the HERMES Consortium, yielding variants at a number of loci. To investigate sub-significant association signals, I will perform an extended analysis with >60,000 HF cases (HERMES, HDR London and UK Biobank) and investigate HF subgroups defined by comorbidity (atrial fibrillation, coronary heart disease) and left ventricular phenotypes by stratified analysis. The results of these analyses may enable the identification of novel causal genes and will form the starting point for investigating clinical and genetic heterogeneity (Objective 2). In addition, they will provide the HF outcome dataset for target validation Mendelian randomisation analyses (Objective 3).

Objective 2. Identify and characterise new HF sub-phenotypes using clinical, EHR and genomic data.
Preliminary analyses indicate that HF risk variants exhibit allelic heterogeneity among HF subtypes defined by comorbid disease status. The observed clinical and genetic complexity suggest an opportunity to define a new, data-driven taxonomy. As well as studying established HF subtypes, I will use unsupervised learning methods (machine learning) to uncover new HF disease clusters based on phenotypic similarity. To explore genetic heterogeneity of HF subgroups, I will test for differences in HF-associated allele frequencies from Objective 1 using locally available, individual participant level data. To validate these findings, I have established a collaboration with investigators from Harvard/ MIT and Boston University. Using these approaches, I will seek to define a data-driven taxonomy of HF subtypes and to provide insights into mechanisms through genomic analysis.

Objective 3. Discover and validate therapeutic targets for heart failure subgroups through genetic causal inference analysis.
Genetic variants that alter the abundance or function of gene transcripts or their cognate protein can serve as instruments to explore the causal role of a protein in a disease outcome using Mendelian randomisation analysis. The genetic determinants of a many circulating proteins are known and can be used to investigate potential therapeutic targets in HF and/or HF subtypes (from Objective 2). I will prioritise proteins for study based on the presence of prior observational associations or other evidence of a causal role in disease. For proteins with evidence of a causal role, I will seek to characterise the mediators of effect by exploring the association of the genetic tool variants with related imaging and EHR phenotypes, to inform the conduct of clinical trials.

Objective 1. Define the contribution of common genetic variation to risk of HF and clinically established sub-phenotypes We will update the existing HERMES analysis to incorporate imputation to the Haplotype Reference Consortium panel. Analysis will be limited to European ancestry participants. For case-control analysis, we will compare all HF (prevalent + incident) with non-HF and perform stratified analysis of subtypes defined by comorbidity. For ejection fraction stratified analysis, we will use randomly selected proportional controls. We will include sex and age as covariates and results will be combined using METAL.

Objective 2. Identify and characterise new HF sub-phenotypes using clinical, EHR and genomic data. We will select clinical characteristics a priori based on past publications and expert judgment. We will test a range of methods for feature selection and dimensionality reduction to identify the most predictive and will use principal components and factor analysis to further reduce dimensionality. To cluster participants, we will use machine learning algorithms including connectivity models, centroid models, and expectation-maximisation. To investigate the genetic heterogeneity of disease clusters we will use BUHMBOX and other approaches and, where powered, GWA.

Objective 3. Discover and validate therapeutic targets for heart failure subgroups through genetic causal inference analysis. To investigate a potential causal role of serum proteins and other candidate drug targets, we will derive tool variants from expression quantitative trait loci (QTL), protein QTL from multiplexed aptamer- and affinity-based assays and downstream intermediate phenotypes. We will select variants in cis- with the cognate protein for each LD block using cluster analysis. We will perform Mendelian randomisation analysis using a range of regression models and will characterise effect mediators using multi-omic datasets enriched for imaging and EHR.