Augmenting clinical risk predictors through multi-omics

Risk profiling is at the heart of disease delay and prevention. At present, most disease risk prediction tools use self-report or clinical data, but it can likely be improved using multi-omic measures. They include genomic data describing liability to disease, dynamic epigenomic changes affecting gene expression, and proteomic data. In this project, I will explore how the incorporation of omic data can help to improve clinical risk prediction tools such as QRISK (stroke), ASSIGN (cardiovascular disease) or CAIDE (dementia) using data of 20,000 participants of Generation Scotland study collected over 14 years of follow-up. The disease data are linked to health records from both GP and hospital records.

First, I will extract information about factors considered by the analysed prediction tools (such as smoking status, age, or sex) from the GS dataset (null models). I will use Cox proportional hazards regression and logistic regression models to determine how well these risk scores classify disease risk. Second, I will build a series of additional risk predictors that will assess (1) levels of around 300 proteins from a mass spectrometry analysis; and (2) 800,000 CpGs from DNA methylation arrays. Third, I will use the aforementioned DNAm and proteomic data to train omics-based predictors of the clinical measures that feed into the null model. I will determine if these omics predictors can augment or supersede the basic model with only clinical predictors. The second and third steps mentioned above will require the application of machine learning methods, such as penalised linear regression (to derive DNAm and proteomic surrogates for continuous outcomes) and penalised logistic and Cox regression for binary and time-to-event outcomes. I will split my data into training and testing sets to avoid overfitting. This will allow me to compare the clinical and omics tools and provide recommendations on how to improve the clinical prediction tools.