Ensemble-Based Feature Selection for Bayesian Integration Models to Improve Biomarker Panel Identification

• Activity Number: 409
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 2:00 PM to 2:45 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Bobbie-Jo Webb-Robertson*+ and Marian Rewers and Qibin Zhang and Katrina Waters and Thomas Metz
• Companies: Pacific Northwest National Laboratory and University of Colorado and Pacific Northwest National Laboratory and Pacific Northwest National Laboratory and Pacific Northwest National Laboratory
• Keywords: Bayesian ; Data integration ; Feature selection ; Biomarker discovery

Abstract

High-throughput technologies can capture data at both global and targeted scales for the transcriptome, proteome and metabolome. The goal of integrated biomarker discovery is to use these disparate data streams to acquire a more predictive and diagnostic biomarker panel than any single data source can provide. However, in a space of tens of thousands of variables (e.g., genes, proteins), feature selection approaches often yield over-trained models with poor predictive power. Moreover, feature selection algorithms are typically focused on single sources of information and do not evaluate the effect on downstream statistical integration models. Bayesian statistics has been shown to be an effective approach for statistical integration across multiple data streams. We present an ensemble-based feature selection approach that optimizes the outcome of interest in the context of the integrated posterior probability. We demonstrate that this approach improves sensitivity and specificity over existing selection routines based on individual datasets.