Abstract
Urothelial carcinoma of the bladder is is estimated to have killed over 16,000 people in the United States in 2016. Like breast cancer, bladder cancer is a heterogeneous disease, and characterization of it’s various subtypes can be useful for forecasting prognosis and treatment efficacy. According to The Cancer Genome Atlas (TCGA) project, the mRNA expression profiles of bladder tumours can be used to cluster the tumors into four different categories: I - Papillary-like, II ‐Luminal A, III - basal/squamous-like, and IV - other (similar to III). However it is not clear whether these mRNA expression based clusters correlate with other molecular and genetic features of the tumor cells. In other words, do differences in mRNA expression profile contain the same information as differences in protein expression, micro RNA (miRNA) expression, copy number variation and somatic mutation data. We tried to recreate mRNA based bladder tumor clusters from other multi-omic data for 328 bladder cancer tumor samples using a special deep and wide belief network composed of restricted Boltzmann machines and a multilayer perceptron. For 10-fold cross validation, we got 79% average test accuracy which implies that that differences in mRNA expression between bladder tumor cells can be reliably, though not perfectly, inferred from different molecular and genetic features of the tumors.
