Transcriptomic and proteomic profiling classify bladder cancers into luminal and basal molecular subtypes, with controversial prognostic and predictive associations. The complexity of published subtyping algorithms is a major impediment to understanding their biology and validating or refuting their clinical use. Here, we optimize and validate compact algorithms based on the Lund taxonomy, which separates luminal subtypes into urothelial-like (Uro) and genomically unstable (GU). We characterized immunohistochemical expression data from two muscle-invasive bladder cancer cohorts ( n=193, n=76) and developed efficient decision tree subtyping models using 4-fold cross-validation. We demonstrated that a published algorithm using routine assays (GATA3, KRT5, p16) classified basal/luminal subtypes and basal/Uro/GU subtypes with 86%–95% and 67%–86% accuracies, respectively. KRT14 and RB1 are less frequently used in pathology practice but achieved the simplest, most accurate models for basal/luminal and basal/Uro/GU discrimination, with 93%–96% and 85%–86% accuracies, respectively. More complex models with up to eight antibodies performed no better than simpler two- or three-antibody models. We conclude that simple immunohistochemistry classifiers can accurately identify luminal (Uro, GU) and basal subtypes and are appealing options for clinical implementation.

No abstract

Transcriptomic and proteomic profiling reliably classifies bladder cancers into luminal and basal molecular subtypes. Based on their prognostic and predictive associations, these subtypes may improve clinical management of bladder cancers. However, the complexity of published subtyping algorithms has limited their translation into practice. Here we optimize and validate compact subtyping algorithms based on the Lund taxonomy. We reanalyzed immunohistochemistry (IHC) expression data of muscle-invasive bladder cancer samples from Lund 2017 (n=193) and 2012 (n=76) cohorts. We characterized and quantified IHC expression patterns, and determined the simplest, most accurate decision tree models to identify subtypes. We tested the utility of a previously published algorithm using routine antibody assays commonly available in surgical pathology laboratories (GATA3, KRT5 and p16) to identify basal/luminal subtypes and to distinguish between luminal subtypes, Urothelial-Like (Uro) and Genomically Unstable (GU). We determined the dominant decision tree classifiers using four-fold cross-validation with separate uniformly distributed train (75%) and validation (25%) sets. Using the three-antibody algorithm resulted in 86-95% accuracy across training and validation sets for identifying basal/luminal subtypes, and 67-86% accuracy for basal/Uro/GU subtypes. Although antibody assays for KRT14 and RB1 are not routinely used in pathology practice, these features achieved the simplest and most accurate models to identify basal/luminal and Uro/GU/basal subtypes, achieving 93-96% and 85-86% accuracies, respectively. When translated to a more complex model using eight antibody assays, accuracy was comparable to simplified models, with 86% (train) and 82% (validation). We conclude that a simple immunohistochemical classifier can accurately identify luminal (Uro, GU) and basal subtypes and pave the way for clinical implementation.