Top-Scoring Pair Methods for Uncovering Metabolite Markers for Discriminating Diabetic Kidney Disease Stage

• Activity Number: 29 - Advances in Methods for Microbiome and Omics Data
• Type: Contributed
• Date/Time: Sunday, August 7, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Brian Kwan* and Daniel Montemayor and Karen Messer and Minya Pu and Jing Zhang and Loki Natarajan
• Companies: University of California, Los Angeles and University of Texas Health San Antonio and University of California, San Diego and Moores Cancer Center, University of California, San Diego and Moores Cancer Center, University of California, San Diego and University of California at San Diego (UCSD)
• Keywords: biomarker; classification; feature selection; kidney disease; metabolomics; order statistics

Abstract

The top-scoring pair (TSP) algorithm has notable potential in deriving biologically interpretable single pair decision rules that are accurate and robust in disease discrimination. However, existing TSP methods do not consider confounding covariates that could heavily influence feature selection. In this work, we provided a novel extension to TSP that uses the residuals from a regression of features on the covariates for identifying the top feature pairs in disease discrimination that are largely decorrelated from the covariates. We demonstrated by simulation studies as well as metabolomics and diabetic kidney disease (DKD) real data application that TSP yields new insights into biomarkers for disease discrimination via our residualizing extension. These residualized top-scoring pairs represent metabolite features, uncorrelated from clinical covariates, for discriminating DKD stage based on the relatively ordering between two features, and thus provide insights into the order reversals in the early vs advanced disease states. Notably, TSP achieved competitive classification accuracy to conventional methods, i.e., LASSO and random forests, while serving as more parsimonious models.