Abstract
Late-onset Alzheimer disease (AD) is a highly complex disease with multiple subtypes, as demonstrated by its disparate risk factors, pathological manifestations, and clinical traits. Discovery of biomarkers to diagnose specific AD subtypes is a key step towards understanding biological mechanisms underlying this enigmatic disease, generating candidate drug targets, and selecting participants for drug trials. Popular statistical methods for evaluating candidate biomarkers, fold change (FC) and area under the receiver operating characteristic curve (AUC), were designed for homogeneous data and we demonstrate the inherent weaknesses of these approaches when used to evaluate subtypes representing less than half of the diseased cases. We introduce a unique evaluation metric that is based on the distribution of the values, rather than the magnitude of the values, to identify analytes that are associated with a subset of the diseased cases, thereby revealing potential biomarkers for subtypes. Our approach, Bimodality Coefficient Difference (BCD), computes the difference between the degrees of bimodality for the cases and controls. We demonstrate the effectiveness of our approach with large-scale synthetic data trials containing nearly perfect subtypes. In order to reveal novel AD biomarkers for heterogeneous subtypes, we applied BCD to gene expression data for 8,650 genes for 176 AD cases and 187 controls. Our results confirm the utility of BCD for identifying subtypes of heterogeneous diseases.