Multimorbidity (two or more chronic conditions affecting an individual) is a growing global health concern associated with higher healthcare usage, costs, and mortality. Many common patterns of multimorbidity exist; often a first (index) condition will predispose you to others, as is the case for hypertension (high blood pressure) and its comorbidities (including type 2 diabetes, coronary artery disease, heart failure, and chronic kidney disease). Multimorbidity is associated with multiple demographic factors (e.g. age, sex, socioeconomic status), environment, lifestyle, and individuals’ genetic predisposition to diseases. Common genetic variants explain a significant portion of the variation in the occurrence of many common comorbidities (~50% in coronary artery disease), thus providing a useful starting point to interrogate the molecular etiology of multimorbidity.

This project will initially combine multi-omic and health data from department cohorts (INTERVAL, BELIEVE) and UK Biobank to identify the molecules (proteins, metabolites) and pathways that underlie different patterns of multimorbidity. Genome-wide association studies (GWAS) will be conducted to identify the effects of genetic variants predisposing individuals to multimorbidity and specific comorbidities in the context of an index condition (initially hypertension) using methods to assess effect size heterogeneity. Putative casual molecules will be identified using methods including QTL colocalization from multi-omic data, TWAS, and mediation analyses then mined for novel therapeutic targets.

Polygenic scores (PGS) are a method to predict genetic predisposition to heritable phenotypes and may be informative for targeting preventive therapy in pre-symptomatic individuals. A parallel aim of this project will build models to predict which comorbidity is most likely to develop next in the presence of an index condition, and assess whether polygenic scores are informative to risk stratification in this setting.

This project will use large scale datasets such as UK Biobank, public summary statistics and statistical analysis, including Mendelian Randomisation, to investigate causal relationships between plasma proteome and metabolome, brain structure and function, and mental illness. The candidate will test for druggable targets, and consider repurposing opportunities to influence brain health, mental and neurodegenerative disorders.

Background: Late stage failure of pharmaceutical compounds to show sufficient efficacy and/or an acceptable safety profile is a key challenge of drug development in the 21st century. Evidence from human genetics can guide discovery of drug targets and help inform investigators about the potential impact of intervening on a pathway prior to clinical trials. Within the Cardiovascular Epidemiology Unit (CEU), there is a range of specific causal inference projects available focusing on areas of direct therapeutic relevance to big pharma, e.g. inflammation, metabolism etc.

Approach: The projects will harness several inter-disciplinary resources and approaches:

– Mendelian randomization methods: analytical frameworks have been developed at CEU to enable testing the causality of proteins, metabolites and other biomarkers using Mendelian randomization or factorial Mendelian randomization, in analogy to a randomized controlled trial (RCT) or factorial RCT, respectively.
– Large-scale bioresources: access to in-house data on sizeable biobanks in the UK (e.g. INTERVAL, 50,000 blood donors) and abroad (e.g. BELIEVE, 75,000-person household survey in Bangladesh) with genetic data, multi-omic data and health outcome data.
– Systems medicine: Dense multi-omic data including transcriptomics (e.g. 5000 INTERVAL participants with RNAseq), proteomics (e.g. 7000 plasma proteins in 10,000 INTERVAL participants and 10,000 BELIEVE participants), NMR metabolomics (all INTERVAL and BELIEVE participants), etc. We will also combine with publicly available resources and databases, such as UK Biobank and FinnGen.

Collaboration: The project will involve close collaboration with members of the group, other research groups at the CEU and methodological experts at CEU. The project will also involve working with national and international collaborators, including academic partners and pharmaceutical companies (e.g., AstraZeneca).

Artificial intelligence (AI) and machine learning algorithms for medical image analysis are rapidly approaching implementation in clinical practice. Despite this there is a significant risk of the models using spurious information that is not related to the pathology on the images, but rather to data contained both within the images and metadata. For example, some chest X-ray AI tools have been shown to utilise factors such as image projection (view), the presence of medical apparatus, and whether the film was obtained supine or erect in making its decision. In addition, sex, ethnicity, geographic location, and acquisition parameters may act similarly as undesirable ‘shortcuts’. These shortcuts increase the risk of inaccurate diagnosis when the algorithms are used in other environments and may lead to an over-optimistic assessment of their diagnostic accuracy.

This project will look at how these biases and shortcuts can be removed from the models’ development and testing, and how these will impact model generalisability (how well it performs when it sees data it has never encountered before).

We will look at learning with and without labels, collecting less confounded data, penalised learning (feature disentanglement), and generalised adversarial networks to challenge these algorithms. We will also assess commercial models that are approved for clinical use, and the robustness of their diagnostic accuracy when these markers are stripped out of medical images.

Artificial intelligence (AI) and machine learning algorithms for medical image analysis are rapidly approaching implementation in clinical practice. Despite this there is a significant risk of the models using spurious information that is not related to the pathology on the images, but rather to data contained both within the images and metadata. For example, some chest X-ray AI tools have been shown to utilise factors such as image projection (view), the presence of medical apparatus, and whether the film was obtained supine or erect in making its decision. In addition, sex, ethnicity, geographic location, and acquisition parameters may act similarly as undesirable ‘shortcuts’. These shortcuts increase the risk of inaccurate diagnosis when the algorithms are used in other environments and may lead to an over-optimistic assessment of their diagnostic accuracy.

This project will look at how these biases and shortcuts can be removed from the models’ development and testing, and how these will impact model generalisability (how well it performs when it sees data it has never encountered before).

We will look at learning with and without labels, collecting less confounded data, penalised learning (feature disentanglement), and generalised adversarial networks to challenge these algorithms. We will also assess commercial models that are approved for clinical use, and the robustness of their diagnostic accuracy when these markers are stripped out of medical images.