|Home | About | Journals | Submit | Contact Us | Français|
Renal cancers account for more than 3% of adult malignancies and cause more than 13,000 deaths per year in the US alone. The four most common types of kidney tumors include the malignant renal cell carcinomas; clear cell, papillary, chromophobe and the benign oncocytoma. These histological subtypes vary in their clinical course and prognosis, and different clinical strategies have been developed for their management. In some kidney tumor cases it can be very difficult for the pathologist to distinguish between tumor types on the basis of morphology and immunohistochemistry (IHC). In this publication we present the development and validation of a microRNA‐based assay for classifying primary kidney tumors. The assay, which classifies the four main kidney tumor types, was developed based on the expression of a set of 24 microRNAs. A validation set of 201 independent samples was classified using the assay and analyzed blindly. The assay produced results for 92% of the samples with an accuracy of 95%.
Renal cell carcinomas (RCC) are a family of carcinomas that arise from the epithelium of the renal tubules. Renal cancers account for more than 3% of adult malignancies. The estimated number of new cases for 2013 is ~65,000 and the estimated number of deaths per year is over 13,000 in the US alone.
There are 11 subtypes of RCC according to the WHO classification of 2004 (Lopez‐Beltran et al., 2006) and the 2009 update (Lopez‐Beltran et al., 2009), as well as 4 types of benign renal tumors. The four most common types of kidney tumors are Clear cell RCC (most common subtype), Papillary RCC, Chromophobe RCC and Oncocytoma. These histological subtypes vary in their clinical courses and in their prognosis, and different clinical strategies have been developed for their management.
The differential diagnosis between the different subtypes of kidney tumors based on morphology can be challenging and is subject to intra‐observer variability (Ficarra et al., 2006). Even when utilizing immunohistochemistry (IHC) markers, the ability to differentiate between the different subtypes can be challenging, especially in the setting of uncommon morphology and small biopsies. Allory et al. recently described a subset of 12 antibodies as a basis for classification of renal cell carcinomas. In this report AMACR, CK7, and CD10 were the most powerful classifiers with 78–87% of carcinomas correctly classified (Allory et al., 2008).
Historically, the kidney cancers were characterized by a lack of early warning signs, diverse clinical manifestations, and resistance to radiation and chemotherapy. However when the disease is diagnosed at earlier stages, prognosis is often improved. Since conventional chemotherapy is not highly effective, targeted treatments open up a new direction in the treatment of RCC (Vasudev et al., 2012).
For the purpose of targeted therapy it is especially important to classify the different subtypes of RCC. The histological types arise from: different cells of origin in the kidney, different constellations of genetic alterations (Lopez‐Beltran et al., 2008), and expression or mutation in different oncogenic pathways. Therefore, different subtypes offer different molecular candidates for targeted therapy, such as Tyrosine Kinase Inhibitors, Sorafenib & Sunitinib, mTOR inhibitors, Everolimus & Tersirolimus etc. There is growing evidence that variability in response rates may be linked to sub‐classification (Tazi el et al., 2011).
Therefore, new biomarkers are needed in order to improve the identification and diagnosis of renal tumor subtypes. The tissue‐specificity of microRNA expression opens a window to new types of molecular diagnostic assays. MicroRNAs are small non‐coding RNA molecules with an important role in the regulation of gene expression (Calin and Croce, 2006; Farazi et al., 2011; Lu et al., 2005), that have been established as strong molecular biomarkers, especially defining tissue origin, differentiation status of cells, and histological types (Gilad et al., 2012; Barshack et al., 2010; Lebanony et al., 2009; Rosenfeld et al., 2008; Bentwich, 2005). The clinical utility of microRNA in kidney cancer classification and diagnosis was explored by us (Fridman et al., 2010) and others (White and Yousef, 2010) and it was shown that microRNAs have a great potential in diagnosis and therapeutics of RCC. We have previously shown that microRNAs can differentiate between the four main types of kidney tumors (Fridman et al., 2010). Here we present the development and validation of a 24‐gene microRNA‐based assay (miRview® kidney, Rosetta Genomics Inc. Philadelphia, PA) for classification of subtypes of primary kidney tumors.
Formalin Fixed Paraffin Embedded (FFPE) samples of primary kidney tumors were collected from several sources (Sheba Medical Center, Tel‐Hashomer, Israel; Beilinson Hospital, Rabin Medical Center, Petah‐Tikva, Israel; ABS Inc., Wilmington, DE, USA; Temple University, Philadelphia, PA, USA; BioServe, Beltsville, MD, USA; Soroka University Medical Center, Beer Sheva, Israel; ProteoGenex, Culver City, CA, USA). Institutional review board approvals were obtained for all samples in accordance with each institute's guidelines. The patient's original diagnosis was one of the 4 subtypes (clear cell, papillary, chromophobe and oncocytoma) and was set as a reference diagnosis for each sample. The diagnosis was based on all available data at the time of patients' diagnosis and IHC and/or colloidal iron stains were performed for some of the cases in the different institutions as decided by the pathologists at the time of diagnosis. Prior to its inclusion in the study, the specimens were reviewed again by pathologists from the different institutes (E.F, S.Z and Y.H) according to morphology and available IHC data and only samples for which the pathologist agreed with the reference diagnosis were included in the study. 181 samples were used for training of the assay.
A blinded validation of the assay was performed on 201 independent samples. Samples were collected in the different institutes as described above and were sent to Rosetta Genomics together with accompanied data. An additional pathologist, (a medical director in Rosetta's CLIA certified and CAP accredited laboratory‐ M.S or TBE) reviewed the H&E slide of all validation cases (except 2 cases for which H&E slides were not available) together with IHC and colloidal iron data that accompanied 55 of the validation set samples. Only samples for which the second pathologist agreed with the reference diagnosis were included in the study. Micro‐dissection was performed, on 6 samples, to reach >50% tumor cell content.
Samples were classified and analyzed using the microRNA‐based assay and the results were compared to the reference diagnosis. Finally, a set of 38 samples was used for inter‐laboratory reproducibility testing.
High‐quality total RNA, including the well‐preserved microRNA fraction, was extracted from the FFPE samples using a proprietary protocol, as previously described (Rosenfeld et al., 2008; Nass et al., 2009).
Expression levels of >700 known microRNAs (corresponding to the Sanger miRBase, version 13), as well as >260 predicted microRNAs sequences (MIDs) were measured using custom‐designed arrays from Agilent Technologies (Santa Clara, CA) which harbor 8 identical sub‐arrays each. RNA was labeled and hybridized as previously described (Meiri et al., 2012). Arrays are scanned using the Agilent DNA Microarray Scanner Bundle and signals were analyzed as previously described (Meiri et al., 2012).
An assay that classifies kidney tumors was developed based on the expression of different microRNAs in the four renal tumor types. Following extraction, seven RNA samples together with a positive control (PC) undergo labeling and hybridization to one array. The PC is an RNA sample that was set as a reference and should meet defined QA criteria: Pearson correlation to the reference hybridization, median of differences from reference and the number of the expressed microRNAs in the dynamic range (expression above 300). QA for each sample is based on several parameters such as the number of microRNAs in the dynamic range, the 98th percentile expression level of the microRNA, the Pearson correlation between the hybridization spikes and the reference, the expression of the negative control probes, and the number of microRNAs with consistent triplicate signals. The signal values of the 24 test microRNAs for each sample are obtained following normalization, and used as input to the test classifier. The assay uses a K‐Nearest‐Neighbor (KNN) classifier with k = 5, that searches for the 5 samples in the training database (181 samples used for assay development) that are most similar to the tested sample. Similarity between the tested sample and the samples in the training database is defined by the pearson correlation coefficient over the expression of the 24 microRNAs. The result for the tested sample is determined by the subtype that appears most often among these 5 closest neighbors. Unless there are four or five neighbors sharing a common subtype, the assay results in no classification and reports that the miR expression pattern of the sample does not match any of the expression patterns in the panel closely enough and does not generate a result. PPV calculation: Positive predictive value (PPV) was calculated in two ways. First, as is widely done, it was calculated based on the validation set by taking the ratio of true positives (e.g. true calls of oncocytoma) to true positives plus false positives (e.g. all cases the test gave oncocytoma as a result). Second, since unlike sensitivity and specificity, the PPV is very sensitive to the prevalence of the different classes, and since the class distribution in the validation set does not accurately reflect the distribution in the patient population expected to use the test, we also calculated the PPV of each class by taking into account the determined sensitivity and specificity values of the classes along with their relative prevalence in the patient population expected to use the test. More precisely, let C be the counting matrix of the validation set results, where C ij is the number of cases that the test gave the answer class j to a case with reference diagnosis of class i. We define the normalized counting matrix, , to take into account the prevalence expected in the patient population of the test, i.e., , where P i is the expected prevalence of class i and N i is the number of cases in the validation set with reference diagnosis of class i. The PPV of class k is then defined as . Since the test is assumed to be used only upon diagnosis (once per patient), we used incidence rates as estimates for the prevalence of the classes in the test population.
This study describes the classification of the 4 most common kidney tumor types (clear cell, papillary, chromophobe and oncocytoma) based on microRNA expression. In the past we have shown the ability to classify renal tumor types using six microRNAs by two microRNA quantification platforms; microarray and qRT‐PCR (Fridman et al., 2010). In this study we demonstrate the robustness of using custom‐designed Agilent microarrays to develop a 24‐gene microRNA‐based assay, and KNN algorithm that classifies kidney tumor subtypes.
Identification of markers and training of the assay was done on 181 samples of the main histological subtypes of RCC: clear cell (51 samples), papillary (51 samples), chromophobe (40 samples) and oncocytoma (39 samples). After comparing the expression of all microRNAs on the Agilent microarray platform, 22 microRNAs were found to be differential between at least one of: 1) oncocytoma and chromophobe vs clear cell and papillary, 2) oncocytoma vs chromophobe, and 3) clear cell vs papillary. In order to improve upon this list, other microRNAs were added in a forward stepwise algorithm. This resulted in the addition of three microRNAs (miR‐146a‐5p, miR‐192‐5p, MID‐00536) to the 22 differential microRNAs. Each of these 25 microRNAs was tested to see whether its removal improved or did not change classification within the training set. Following this analysis, miR‐125b‐5p was removed, resulting in the final list of 24 microRNAs that were chosen as features in the KNN algorithm. Figure 1 shows an unsupervised clustering of the 181 samples based on the expression of the 24 microRNAs. As seen in the figure there is a very distinct pattern of microRNAs differentiating the groups of clear cell RCC and papillary RCC from the groups of chromophobe RCC and oncocytoma. Specific microRNAs can also distinguish between clear cell and papillary RCC (miR‐31‐5p, miR‐126‐3p, miR‐195‐5p, miR‐200a‐3p, miR‐200b‐3p, and miR‐122‐5p) and between chromophobe and oncocytoma (miR‐200a‐3p, miR‐200b‐3p, miR‐200c‐3p, and miR‐141‐3p). Not surprisingly, the microRNAs that were found to be differentially expressed in a previous cohort that we studied (Fridman et al., 2010), and in another study (Youssef et al., 2011) were also shown to be differential in this study. Previously we demonstrated the ability to differentiate the 4 types using a binary tree with 2 microRNAs in each node, differentiating first the more resembling groups (clear cell and papillary from chromophobe and oncocytoma) and then differentiating the 4 subtypes. Based on these initial results, we developed an assay that utilizes a microarray platform and can use the optimal number of microRNAs needed for the classification. In the initial steps, microRNAs were chosen based on their ability to separate between histological types in terms of p‐value, fold‐change, and classification performance. KNN was chosen as the classification algorithm, based on its performance on the training set as evaluated by Leave‐One‐Out Cross Validation.
MicroRNA expression in main renal tumor subtypes – unsupervised clustering of 181 samples used for the training of the assay using the expression of the 24 assay microRNAs. Expression was measured using the Agilent microRNA microarray platform. ...
A blinded independent validation set was used to study the assay performance. The validation set included only samples that met the Gold Standard criteria: over 50% tumor cell in the sample, and concordance in diagnosis between two pathologists. The validation set included a balanced number of samples from each type (see Table 1). Out of the 201 samples used for validation, one sample failed QA due to insufficient RNA and additional 16 samples completed processing but did not result in classification, as explained in methods. For 184 samples (92%) the assay was able to produce results. 174 samples out of the 184 samples (95%) were classified accurately. Figure 2 is a confusion matrix showing all validation results and Table 1 presents the overall performance of the assay. Sensitivity and specificity are shown per tumor type indicating the very high accuracy of the assay. Out of the 10 misclassified samples, 7 were due to confusion between chromophobe RCC and oncocytoma (Figure 2).
Validation confusion matrix – on the y‐axis, the “reference diagnosis” represents the classification given by the two independent pathologists and on the x‐axis, the “classifier answer” is the classification ...
Validation performance per tumor type – the table shows, for each histological class, the total number of samples (“N”), and number of samples with results (in parentheses), the sensitivity, the specificity and the positive predictive ...
For an assay to be robust it is important to show that the results can be reproduced in other laboratories. For this purpose a cohort of 38 samples from RCC patients were studied; 20 of the samples were divided to the two testing laboratories for RNA extraction (extraction was done in each lab from different sections of the same tissue blocks, and the assay was performed in the same laboratory that extracted the RNA), and for an additional 18 samples, RNA was extracted in one laboratory and the assay was performed in both laboratories using that RNA. The first study measured the concordance of the entire process of the 24‐gene microRNA‐based assay (including RNA extraction), and the second study measured the concordance of the RNA labeling and array hybridization in the two laboratories starting from the same 18 RNA samples. 37 out of the 38 samples passed QA criteria of the assay in both laboratories. One of the 20 samples extracted in both laboratories failed QA in one of the sites. Two cases completed processing but resulted in no classification (see Methods) in both laboratories. The remaining 35 samples got the same classification in both laboratories (100% inter‐laboratory concordance) and in agreement with the reference diagnosis.
In this study, a set of 24 differential microRNAs was identified for the accurate classification of kidney tumors. This set was the basis for the development and validation of a standardized diagnostic assay for the classification of renal cell tumors from FFPE resections or biopsy samples. The validation results showed 95% accuracy and demonstrated again the diagnostic power of microRNAs.
The 4 most common subtypes of renal tumor are characterized by different genetic alteration and different chromosomal material loss (Youssef et al., 2011; Bodmer et al., 2002; Steiner and Sidransky, 1996), generating different protein expression and histological markers (Cheng et al., 2009). However those characteristics cannot solely differentiate well between the subtypes. Since histological evaluation by regular light microscopy cannot always distinguish between the various renal tumor types, panels of dozens of biomarkers were chosen for immunohistochemical staining (Shen et al., 2012; Truong and Shen, 2011). For example clear cell RCC is typically positive for vimentin, AE1/AE3 keratins, CD10, RCC marker, and carbonic anhydrase IX (G250), and it has usually diffused immune‐reactivity for CK7, CD117, kidney specific cadherin and parvalbumin. Kidney specific cadherin can show a complex pattern of positive staining or negative staining in RCCs. Papillary RCC is often uniformly positive for vimentin, AE1/AE3 keratins, CK7, AMACR, and RCC marker, and it is usually negative for CD117, kidney‐specific cadherin, and parvalbumin. Chromophobe is positive for kidney‐specific cadherin, parvalbumin, CD117, epithelial membrane antigen, AE1/AE3 keratin, and CK7, and it is usually negative for vimentin, carbonic anhydrase IX, and AMACR. Oncocytoma shares a similar immunoprofile with chromophobe RCC, but in some studies vimentin, S100A1, and CD82, were shown to be helpful in some situations (Truong and Shen, 2011). Hale's colloidal iron stain can be helpful in some situations since it is confluently positive for oncocytoma and diffusely positive for chromophobe RCC.
As can be seen from the list above the similarity between the IHC markers of clear cell RCC and papillary RCC makes IHC a limited tool in differentiating the subtypes. This is also true for differentiation between chromophobe and oncocytoma, which is almost impossible in needle biopsy (Blumenfeld et al., 2010). Moreover, the increasing use of core needle biopsy in preoperative diagnosis limits the amount of material available for a large panel of immuno‐histochemical markers to be tested.
Understanding the molecular mechanism underlying the development of renal carcinomas allowed the development of novel targeted therapies like sorafenib, bevacizumab (with IFN‐α), sunitinib, temsirolimus and more (Najjar and Rini, 2012). In the past, most clinical trials included patients with clear cell or predominantly clear cell histology, but recently more data is accumulating pointing to different molecular pathophysiology of the different subtypes and different responses to targeted therapies. Therefore the ability to correctly classify the different renal tumor subtypes is of great importance. Moreover although FNA can provide useful information in RCC patients the biopsy has poor sensitivity and specificity when assigning Fuhrman nuclear grade, and physicians should be cautious when assigning grade or sarcomatoid elements from biopsy data (Abel et al., 2010).
In light of the above, the use of a robust biomarker like microRNA expression that can be assessed in a single measurement using a very small amount of material is of great appeal.
Our validation shows that an assay based on the expression of 24 microRNAs can distinguish between all 4 common subtypes very well, with an overall accuracy of 95%. 7 out of the 10 misclassifications were confusion between chromophobe and oncocytoma (see Figure 2), which is the most difficult differential diagnosis. Nevertheless the sensitivity for those classes using the 24‐gene microRNA‐based assay is high (93% and 86% respectively). Inter‐laboratory concordance of 100% demonstrates the robustness of the assay. The study uses the classical morphology and IHC diagnosis as a gold standard, despite its limitation, but efforts were made to have an accurate diagnosis by validation by a second pathologist.
Although using microRNA as biomarkers can be unrelated to the downstream biological changes caused by their differential expression or the upstream changes that resulted with differential expression of microRNAs, we looked at the differential microRNAs to see if we can find biological relevance to renal carcinoma. MiR‐126‐3p is known to be associated with endothelial cells, and regulated angiogenesis signaling (Wang et al., 2008). The knockdown of this microRNA is associated with vascular integrity loss. In our assay miR‐126‐3p was found to be lower in papillary RCC, which is relatively hypovascular in comparison to clear cell RCC (Vikram et al., 2009).
The assay also detected the family of miR‐200 and miR‐141‐3p that are known to be associated with Epithelial Mesenchymal transition (EMT) (Mongroo and Rustgi, 2010; Korpal and Kang, 2008) and are characterized as expressed in epithelial cells. This family of microRNAs has higher expression in the chromophobe subtype of RCC, especially miR‐200c‐3p and miR‐141‐3p of this family are higher specifically in this subtype and have very low expression in the clear cell RCC and papillary RCC. Both of those microRNAs are located in the same cluster on chromosome 12 and their transcription is suppressed in several cancers resulting in repression of E‐cadherin and leading to EMT. The down‐regulation of miR‐141‐3p and miR‐200c‐3p in clear cell RCC was proposed to be involved in suppression of CDH1/E‐cadherin transcription via up‐regulation of ZFHX1B (Nakada et al., 2008). RCC is known to have EMT markers that are expressed in correlation to prognosis and resistance to therapy (Harada et al., 2012). Therefore the low expression in the more aggressive subtypes of RCC is in line with previous data.
Mir‐122‐5p, which is highly expressed in liver tissue and is rarely expressed in other tissues (Landgraf et al., 2007), was found here to be expressed (although very low expression relative to liver tissue) specifically in the clear cell subtype of RCC, as was already shown previously (Zhou et al., 2010), but its role in RCC is not yet understood. MiR‐21‐5p that is known to be a widespread marker of cancer and bad prognosis is also higher in clear cell and papillary subtypes compared to chromophobe and the benign oncocytoma. This differential expression is again within the biological rationale since this microRNA is known to be a prognostic marker (Faragalla et al., 2012; Fu et al., 2011) and clear cell RCC and papillary RCC are both the RCC types which are considered more aggressive and therefore have the worst prognosis (Cheville et al., 2003). The same is shown for miR‐210 that also has higher expression in clear cell and papillary subtypes and is yet another microRNA that is associated with several cancers, known to be induced by hypoxia (Devlin et al., 2011; Huang et al., 2010), and also serves as a powerful prognostic marker in breast cancer (Camps et al., 2008) and other cancers.
Therefore the identification of the 24 microRNA biomarkers seems not to be a random occurrence, but to have biological relevance to RCC in general and to the different subtypes in particular.
In summary, the 24‐gene microRNA‐based assay, measured on a microarray platform, can accurately differentiate between the four main types of primary kidney tumors. This assay can serve as a reliable diagnostic tool to aid physicians with the growing unmet need for kidney tumor classification.
We thank Dr. Tina Bocker Edmonston for her help in performing pathological review of validation samples.
Spector Yael, Fridman Eddie, Rosenwald Shai, Zilber Sofia, Huang Yajue, Barshack Iris, Zion Orit, Mitchell Heather, Sanden Mats, Meiri Eti, (2013), Development and validation of a microRNA‐based diagnostic assay for classification of renal cell carcinomas, Molecular Oncology, 7, doi: 10.1016/j.molonc.2013.03.002.