A URV-based Biclustering for uncovering Glioblastoma Molecular Signatures

Microarray technology often generates ordinal detection calls (Absent, Marginal, Present) that are frequently treated as continuous or nominal, losing their inherent ranked structure. We propose a model- based biclustering framework based on the Underlying Response Vari- able (URV) approach. By modeling these calls as discretizations of la- tent Gaussian mixtures, we identify localized molecular signatures. Our model employs a flexible factorial covariance structure where gene parti- tions can vary across patient clusters, offering a more nuanced view than rigid Latent Block Models (LBM). Using a pairwise composite likelihood estimation, we analyze 22777 genes from the TCGA-GBM project, iden- tifying distinct clusters that perfectly align with disease status.