Empirical Bayes Methods for Biomarker Identification in Metabolomics

• Activity Number: 365
• Type: Contributed
• Date/Time: Wednesday, August 6, 2008 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Cheng Zheng*+ and Olga Vitek+ and Haiwei Gu and Zhengzheng Pan and Daniel Raftery
• Companies: Purdue University and Purdue University and Purdue University and Roche Diagnostics Co. and Purdue University
• Keywords: Mass Spectrometry ; Direct Analysis in Real Time (DART) ; Desorption Electrospray Ionization (DESI) ; empirical Bayes ; Linear Model for Microarray Analysis (LIMMA) ; metabolomics

Abstract

Mass spectrometry-based metabolomics studies measure the abundance of metabolites in a sample at hundreds of mass to charge ratios (m/z). Determination of features that are differentially abundant between cases and controls is the primary goal. Empirical Bayes approaches (e.g., LIMMA) efficiently determine such features by assuming a common prior distribution of feature variances. However, the assumption is inappropriate for ionization techniques such as DART and DESI which produce a relationship between the features m/z and its variance. We extend the model in LIMMA by relating feature variances to m/z by a spline, and integrate the estimation into the empirical Bayes framework. We show by simulation that the proposed approach outperforms t-test, LIMMA, and EBarrays in terms of FDR and power. We illustrate the method on a metabolomic profiling experiment of breast cancer.