Profiling and classification tree applied to renal epithelial tumours.

AIMS: Selection of the relevant combination from a growing list of candidate immunohistochemical biomarkers constitutes a real challenge. The aim was to establish the minimal subset of antibodies to achieve classification on the basis of 12 antibodies and 309 renal tumours. METHODS AND RESULTS: Seventy-nine clear cell (CC), 88 papillary (PAP) and 50 chromophobe (CHRO) renal cell carcinomas, and 92 oncocytomas (ONCO) were immunostained for renal cell carcinoma antigen, vimentin, cytokeratin (CK) AE1-AE3, CK7, CD10, epithelial membrane antigen, alpha-methylacyl-CoA racemase (AMACR), c-kit, E-cadherin, Bcl-1, aquaporin 1 and mucin-1 and analysed by tissue microarrays. First, unsupervised hierarchical clustering performed with immunohistochemical profiles identified four main clusters-cluster 1 (CC 67%), 2 (PAP 98%), 3 (CHRO 67%) and 4 (ONCO 100%)-demonstrating the intrinsic classifying potential of immunohistochemistry. A series of classification trees was then automatically generated using Classification And Regression Tree software. The most powerful of these classification trees sequentially used AMACR, CK7 and CD10 (with 86% CC, 87% PAP, 79% CHRO and 78% ONCO correctly classified in a leave-one-out cross-validation test). The classifier was also helpful in 22/30 additional cases with equivocal features. CONCLUSION: The classification tree method using immunohistochemical profiles can be applied successfully to construct a renal tumour classifier.