Comparisons of statistical genetics methods for improved prediction of quantitative phenotypic traits related to individual CVD risk Seminar

Abstract
My project explores further the potential of genetics to improve individual risk prediction, and in particular the development of statistical methodology for the genetic based prediction of quantitative phenotypic traits related to disease risk.

We explore ways to use much larger sets of single-nucleotide-polymorphisms (SNPs), allowing for both common variants of small effect and rare variants of large effect, so that a larger amount of genetic variation is exploited.

Initially, simple risk score models are constructed, based on sums of minor allele counts, and predictive accuracy is compared between weighted and un-weighted scores, as well as investigating any differences between internal and external weighting.

The benefits of different measures of predictive accuracy, such as Mean Square Error and Correlation, are also compared for all models.

These will later be compared to risk prediction models using shrinkage estimation methods, adapted from GWAS, e.g. Lasso and Hyper-Lasso models, which relate to different priors on the effect size, and are expected to yield greater accuracy.

This methodological work is applied to the genetic prediction of lipid levels, which are established risk factors for CVD. Our methods are tested using the Whitehall II and British Women's Health & Heart cohort studies, which contain data for almost 50,000 SNPs from the cardiochip, for over 5,000 and 3,000 individuals, respectively.