Exact Bayesian Screening for Rapidly Identifying Uninformative Features from High-Dimensional Biomedical Arrays

• Activity Number: 467 - Modeling, Design Strategies and Assessment of Biomarkers
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): A Lawrence Gould* and Richard Baumgartner
• Companies: Merck Research Laboratories and Merck Research Laboratories
• Keywords: mixture model; microarray; metabolomics

Abstract

Biomedical applications such as genome-wide association studies, metabolomics, and clinical safety screen large databases with high-dimensional features such as genes and adverse event names to identify important features for subsequent detailed analysis when there is noise and interference from unimportant features and when, in addition, the important features are rare and weakly expressed. We describe a Bayesian screening approach based on a 2-stage gaussian random mixture model analogous to a method described earlier for identifying potentially important adverse events from large clinical or observational trials based on frequency of occurrence. We focus on the missed rather than false discovery rate because it is important not to miss potentially informative features. The method indicates how likely a feature is worth further investigation and measures the likelihood of various differences. The critical values for the screening process reflect clinical and regulatory priorities and the diagnostic properties of the method can be assessed directly. The calculations are “exact” and rapid: the key computation for a database with over 7200 features took less than 20 seconds.