Risk Prediction Using Whole-Genome Data: Methods and Prospects

• Activity Number: 210
• Type: Invited
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Peter Kraft*+
• Companies: Harvard School of Public Health

Abstract

More and more genetic data are available on larger and larger samples: for some traits hundreds of thousands of subjects have been genotyped using genome-wide association arrays and tens of thousands have had their whole exome or even whole genome sequenced. It is now possible to build genetic prediction models for complex traits using clinical variables and a near-complete catalog of genetic variation. This is however a challenging problem involving very high dimensional and sparse data and often sparse and weak signal. I review some of the methods for building genetic prediction models (e.g. ad hoc variable selection, penalized regression) and some of the difficulties (using summary association data when individual-level data is not available, whether and how to combine rare variants, the usefulness of interactions, incorporating genomic annotation). I present empirical results on the performance of genetic prediction models for breast cancer, cardiovascular disease, and height. I review metrics for evaluating the clinical utility of prediction models. I close with a discussion of the limits of genetic prediction models and scenarios where these models may have clinical utility.