A versatile and upgraded version of the LundTax classification algorithm applied to independent cohorts

Abstract

Stratification of cancer into biologically and molecularly similar subgroups is a cornerstone of precision medicine and transcriptomic profiling has revealed that urothelial carcinoma (UC) is a heterogeneous disease with several distinct molecular subtypes. The Lund Taxonomy classification system for urothelial carcinoma aims to be applicable across the whole disease spectrum including both non-muscle invasive and invasive bladder cancer. For a classification system to be useful it is of critical importance that it can be applied robustly and reproducibly to new samples. Many transcriptomic methods used for subtype classification are affected by the choice of expression platform, data preprocessing, cohort composition, and tumor purity. The application of a subtype classification system across studies therefore comes with a degree of uncertainty regarding whether the predictions in a new cohort accurately recapitulate the originally intended stratification. Currently, only limited data has been published evaluating the transferability and applicability of existing stratification systems and their respective classification-algorithms to external datasets. In the present investigation we develop a single sample classifier based on in-house microarray and RNA-sequencing data intended to be broadly applicable across datasets, studies, and tumor stages. We evaluate the performance of the proposed method and the Lund Taxonomical classification across 10 published bladder cancer cohorts (n=2560 cases) by examining the expression of characteristic subtype associated gene signatures, and whether complementary data such as mutations, clinical outcomes, response, or variant histologies are captured by our classification. Effects of varying sample purity on the classification results were also evaluated by generating low-purity versions of samples in silico. We show that the classifier is robustly applicable across different gene expression profiling platforms and preprocessing methods, and less sensitive to variations in sample purity.The classifier is available as the ‘LundTaxonomy2023Classifier’ R package onGitHub.