Purpose: Gene expression profiling followed by unsupervised machine learning (ML) has failed to provide specific, clinically-actionable subgroups. Standard strategies sample and profile tumours, but this produces a mixed signal from cancer cells, cancer-infiltrating cells and unwanted adjacent tissue variably captured during surgery. As most ML approaches rely on highly variable features, patient classifiers built from such heterogeneous yet unreproducible profiles do not significantly improve patient outcomes. Such strategies also ignore decades of mechanistic cancer biology when deciding on the most informative features. We focused on muscle-invasive urothelial carcinoma of the bladder (MIBC; 40% 5-year survival). To derive cell-type-specific profiles, we performed RNAseq on histologically normal, pure human urothelium across three differentiation states (n=88). We constructed a co-expression network using the most variable 5000 genes expressed in over 90% of samples, prioritising the connections of transcription factors (TFs). Perturbation experiments revealed 98 consistently hyper-connected TFs which we used to stratify MIBC samples from The Cancer Genome Atlas.