Addressing Between-Study Heterogeneity and Cross Platform Data Integration for Gene Signature Selection and Clinical Prediction

• Activity Number: 36 - Diagnostic, Prognostic, and Predictive Genomic Biomarkers for Cancer
• Type: Contributed
• Date/Time: Sunday, July 30, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Naim Rashid* and Quefeng Li and Joseph G Ibrahim
• Companies: University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill and UNC
• Keywords: generalized linear mixed models ; RNA-seq ; data integration ; variable selection

Abstract

In the genomic era, gene biomarkers and collections of gene biomarkers are often utilized to subtype cancer patients, determine treatment type, or predict response to therapy. However, often times multiple sets of gene biomarkers ("gene signatures") are published for the purposes of patient subtyping or risk prediction for similar classes of patients within the same cancer. The effects of such gene biomarkers may also vary from study to study. This occurence has practical implications in the the generalizability and clinical applicability of such signatures. We introduce a novel approach to select gene biomarkers from multiple data sets, accounting for study level heterogeneity via high dimensional penalized random effects models. We show that relative to competing approaches that markers selected through this approach result in much more stable gene signatures and prediction. We also introduce a robust platform-independent approach to combine data generated by multiple technologies into a single matrix prior to biomarker selection. We demonstrate the performance of this method on four genomic data sets to predict common subtype across multiple cancers and platforms.