For example, PAX-8 was listed as over-expressed in clear cell tumors relative to papillary. PAX-8 has been implicated in Wilm��s tumors. SCRN1, a marker of colorectal cancer, was also over-expressed in ccRCC relative to papillary. Some of the genes in the signatures may be observers rather than drivers. For example, the over-expression of aquaporin 6 in chromophobe tumors relative to oncocytomas is unlikely to have any direct effect on tumor growth or invasion, but is nonetheless a good marker this is the second study to show differential expression of this gene between chromophobe RCC and oncocytomas. We acknowledge several limitations of this study. First, while the study includes the most common types of renal cortical tumors, not all histologic subtypes are represented in our model. For example, we excluded collecting duct RCC, clear cell RCC with papillary features, and tubulo-papillary RCC. We felt that most of these sub-types are either very rare or represent heterogeneous histologies. We also did not separate papillary tumors into type I and type II. These tumors may represent a continuum rather than separate entities, and papillary type II may include eosinophilic tumors of many origins. We also did not performed analysis of normal renal parenchyma. Normal renal parenchyma can be readily distinguished from a solid renal tumor on H&E stain. Therefore, we focused on differentiating different types of RCC. Furthermore, while we validated the robustness of this signature by examining the performance of our signature in outside datasets, the true performance of the signature will have to be confirmed in a set of unclassified renal cortical neoplasms. Indeed, this may prove to be the most useful utility of these signatures. Finally, we did not incorporate stage and grade of the tumors into our algorithm as this information was not available for all studies. While the lack of centralized pathology review may be considered as a limitation of our study, it is an inherent feature of this study that may also be its strength, since the samples analyzed were derived from multiple institutions in two countries. Our 94% accuracy when validated on external datasets is likely due to the use of data from multiple sources making our results more GDC-0941 Abmole PIK3CA hotspot mutations differentially impact responses to MET targeting in MET-driven and non-driven preclinical cancer models generalizable. The strong performance of our signature even in the non-clear histologic subtypes is best explained by the fact that we were not evaluating the genes that correlated with a histologic subtype defined by one pathologist, but rather the integration of molecular profiling with morphology as determined by multiple pathologists at various academic institutions. We hope that our future studies will help address the current shortcomings in subclassification of renal cortical tumors, support the clinical utility of our algorithm, and move the field closer towards personalized medicine for patients with renal cortical neoplasms. In summary, the use of meta-analytic techniques has facilitated the creation of signatures that have accurately differentiated renal cortical neoplasms. Sequential application of three signatures, according to an algorithm that utilized the natural differences in gene expression between tumor subtypes, correctly classified renal epithelial tumors from five institutions with 94% accuracy. Our algorithm may potentially be used as a adjunct for pathologists when the diagnoses are not obvious in order to improve the diagnostic accuracy of percutaneous renal biopsies and to help direct treatment options.