Machine Learning Techniques for Predictive Modeling in Pharmacogenetics: Modeling Warfarin Dose Response in African Americans

• Activity Number: 566
• Type: Topic Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Erdal Cosgun*+ and Nita A. Limdi and Christine Woods Duarte
• Companies: University of Alabama at Birmingham and University of Alabama at Birmingham and University of Alabama at Birmingham
• Keywords: Machine Learning ; Warfarin ; Data Mining ; Random Forest Regression ; Support Vector Regression ; Boosted Regression Tree

Abstract

With most complex traits and diseases having expected genetic contributions of many hundreds or even thousands of genetic factors,and with genotyping arrays consisting of hundreds of thousands SNPs,powerful high dimensional statistical techniques are needed to comprehensively model the genetic variance.We have previously applied 3 machine learning (ML) approaches:Random Forest Reg.,Boosted Reg. Tree and Support Vector Reg. to the problem of prediction of warfarin maintenance dose in a sample of African Americans. We have developed a multi-step approach for SNP selection,model building, and model assessment in a cross-validation framework,and our results indicate that our modeling approach gives much higher accuracy than previous models for warfarin dose prediction,with an average R2 between predicted and actual square root of warfarin dose as measured in the test samples ranging from 52.4% to 68.2% depending on the method used.We are currently investigating larger prediction models with up to a thousand pre-selected SNPs,improved methods for marker pre-selection.In summary,ML approaches for high-dimensional pharmacogenetic prediction hold great promise and warrant further research.