Parsimonious Models for Classifying Proteomic MS Data

• Activity Number: 681
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Chris Fraley*+ and Bryan Prazen and Mark Seligman
• Companies: Insilicos LLC and Insilicos LLC and Insilicos LLC
• Keywords: proteomics ; mass spectra ; biomarkers ; classification ; prediction ; penalized regression

Abstract

Modern proteomic assays like mass spectra result in thousands of measurements per sample. The problem of translating this data into useful information can be daunting. Goals often include accurate class prediction, and finding small sets of signals that can potentially serve as biomarkers to diagnose disease or guide development of therapies. After initial preprocessing, features are usually selected and/or extracted prior to classification, using methods ranging from univariate significance tests to wavelet transforms. Virtually every statistical pattern recognition technique available has been applied to classification of mass spectra, although no preferred method has so far emerged, even for specific applications. We show that penalized logistic regression, combined with simple dimension reduction strategies, can yield good prediction models with small numbers of potential biomarkers.