Several groups have during past years produced molecular classification schemes for bladder cancer. Even though no consensus on how to define a subtype exists, one approach has been to base definitions on how tumours cluster according to their mRNA expression profiles. In many cases, obtained profiles, and thus class defining features, are affected by signals from non-tumour cells within the biopsy. To overcome this issue, we combined gene expression analyses with analyses of the actual tumour cells by extensive immunohistochemistry (IHC). By this approach we were able to define tumour cell phenotypes i.e., subtypes defined by features of the tumour cells only, and adjust mRNA-based algorithms accordingly. In the present investigation we address the non-luminal Basal/Squamous-like (Ba/Sq) and Small cell/Neuroendocrine-like (Sc/NE) categories of tumours defined by mRNA-based classification. We make use of IHC data for 15 proteins, all known to be instrumental for defining molecular subtypes of urothelial carcinoma. We show that the UroB type of tumours, frequently grouped together with Ba/Sq, are different from the Ba/Sq entity at several essential features and is a derivative of Urothelial-like tumours (Uro). We show that the Sc/NE tumours are similar to but represents extreme versions of Genomically Unstable (GU) tumours. We apply clustering to 423 cases representing all subtypes using IHC data for 14 proteins and show that the obtained grouping conforms well with the mRNA-based classification. This work describes in detail the molecular pathology of non-luminal RNA-based bladder cancer subtypes and highlight similarities/dissimilarities suggestive of origin.