We describe a novel strategy (random forest clustering) for tumor profiling based on tissue microarray data. Random forest clustering is attractive for tissue microarray and other immunohistochemistry data since it handles highly skewed tumor marker expressions well and weighs the contribution of each marker according to its relatedness with other tumor markers. This is the first tumor class discovery analysis of renal cell carcinoma patients based on protein expression profiles. The tissue array data contained at least three tumor samples from each of 366 renal cell carcinoma patients. The eight tumor markers explore tumor proliferation, cell cycle abnormalities, cell mobility, and the hypoxia pathway. Since the procedure is unsupervised, no clinicopathological data or traditional classifications are used a priori. To explore whether the tissue microarray data can be used to identify fundamental subtypes of renal cell carcinoma patients, we first carried out random forest clustering of all 366 patients. By analyzing the tumor markers simultaneously, the procedure automatically detected classes that correspond to clear- vs non-clear cell tumors (demonstration of proof-of-principle). The resulting molecular grouping provides better prediction of survival (logrank P=0.000090) than this classical pathological grouping (logrank P=0.023). We then sought to extend the class discovery by searching for finer subclasses of clear cell patients. The procedure automatically discovered: (a) two classes corresponding to low- and high-grade patients (demonstration of proof-of-principle); (b) a subgroup of long-surviving clear cell patients with a distinct molecular profile and (c) two novel tumor subclasses in low-grade clear cell patients that could not be explained by any clinicopathological variables (demonstration of discovery).
