We will benchmark and improve polygenic risk scores (PRS) for common complex, psychiatric, and selected behavioral and cognitive traits using UK Biobank’s genetic and phenotypic data.
Our key questions are: (1) How well do state-of-the-art PRS methods predict risk within families, especially among siblings? (2) How much additional liability-scale variance and rank-ordering accuracy can be achieved beyond current benchmarks, including through newly developed deep neural network models? (3) How does PRS performance and calibration vary across traits and subgroups relevant to clinical or public-health use?
We will focus on highly heritable traits with robust case numbers in UK Biobank, including coronary artery disease, type 1 and type 2 diabetes, hypertension, atrial fibrillation, ischaemic stroke, breast, prostate and colorectal cancer, asthma, inflammatory bowel disease, major depressive disorder, bipolar disorder, schizophrenia, and anxiety disorders. Where available, we will also analyse quantitative traits such as BMI, blood pressure, lipid levels, cognitive test performance and educational attainment.
Objectives are to: (i) implement a unified evaluation framework for leading PRS methods and neural network-based predictors across many traits; (ii) develop and train novel deep neural network models directly on UK Biobank genotypes and covariates; (iii) perform both between-family and within-family validation; (iv) quantify clinically relevant metrics (liability R², discrimination, calibration, and rank-based measures such as Spearman correlation); and (v) synthesise results into a publicly available white paper and peer-reviewed publications that clarify the strengths and limitations of current PRS tools.