Online Program

Return to main conference page
Wednesday, September 25
Wed, Sep 25, 9:45 AM - 10:30 AM
Marriott Foyer
Poster Session

Novel Statistical Methods for Biomarker Discovery in Pharmocogenomic Studies (300887)

*Juhyun Kim, University of California at Los Angeles 
Devan Mehrotra, Merck & Co., Inc. 
Judong Shen, Merck & Co., Inc. 
Anran Wang, Merck & Co., Inc. 
Hua Zhou, University of California at Los Angeles 
Jin Zhou, University of Arizona 

Keywords: rare variant, minorization-maximization (MM), penalized estimation, variance components model, restricted maximum likelihood (REML), group selection

Improved personalized treatments are key to improving health care. In spite of dramatic developments in both pharmacogenomic knowledge and the techniques used to perform pharmacogenomic studies, multiple challenges still exist that slow the translation of pharmacogenomic discoveries from "bench to bedside". Variation in response to drug therapies results not only from gene sequence variation, ultimately resulting in differences in mRNA and protein expression, but also other patient characteristics such as clinical and environmental factors. However, few statistical methods exist to maximize use of the extensive genomic data and prior biological knowledge in order to unravel the etiology of complex drug-response phenotypes. We proposed to develop innovative statistical analytical tools and supporting software to integrate prior biological knowledge, such as function or known relationship between genes, with genomic data that will facilitate the generation of novel pharmacogenomic hypotheses. Rare genetic variants are thought to be the key to elucidating the genetic architecture of common diseases and complex traits. Single nucleotide polymorphism (SNP) set analysis aggregates both common and rare variants and tests for association between a phenotype of interest and a set. However, multiple genes, pathways, or sliding windows are usually investigated across the whole genome, in which all groups are tested separately followed by multiple testing adjustment. We propose a minorization-maximization (MM) algorithm that selects relevant variance components to prioritize SNP sets. It is achieved by incorporating penalties, including lasso, adaptive lasso, and minimax concave penalty (MCP). Simulation studies demonstrate the superiority of our methods in model selection performance, compared to the traditional marginal testing methods. We apply our method to a real Merck pharmacogenomics study.