A Two-Step Feature Selection Strategy for Large-Scale High-Dimensional Genetic Risk Prediction

• Activity Number: 589
• Type: Topic Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Zhi Wei*+ and Wei Wang
• Companies: New Jersey Institute of Technology and New Jersey Institute of Technology
• Keywords: genetic risk prediction ; model selection ; high-dimensional learning ; large-scale learning ; multiple testing

Abstract

Genome-wide association studies (GWAS) have been fruitful in identifying disease associated loci. However, it remains unclear if these advances can deliver sufficiently accurate genetic risk predictions to make targeted intervention realistically possible. Cursed by the high dimensionality of GWAS data, most existing genetic risk prediction results are modest, if not negative. Here we perform risk assessment for a large-scale high-dimensional GWAS dataset with sample size of 60,000+ and feature size of 150,000+. We employ a two-step feature selection strategy. In the first pre-selection step we reduce the ultra-high dimensionality to a moderate scale using a simple univariate method. Then we conduct penalty-based model selection in the second training step. We evaluate computational feasibility and estimation accuracy in comparison with the strategy without the first step pre-selection. We also study the stability of selected features by re-sampling. Our empirical results shed some insights into the connection between multiple testing and model selection.