A Molecular Classification of Papillary Renal Cell Carcinoma

Kidney cancer is a heterogenous disease consisting of various subtypes with diverse genetic, biochemical, and morphologic features. Epithelial renal cell carcinoma (RCC) accounts for the vast majority of renal malignancies in adults. Based on morphologic features defined in the WHO International Histological Classification of Kidney Tumors, RCC can be divided into clear cell (conventional), papillary (chromophil), chromophobe, collecting duct, and unclassified subtypes (1, 2). Papillary RCC (PRCC) is the second most common subtype comprising 10% to 15% of kidney cancers (3), with an estimated annual incidence of between 3,500 and 5,000 cases in the United States, based on overall statistics for kidney cancer (4). PRCC is histologically characterized by the presence of fibrovascular cores with tumor cells arranged in a papillary configuration. The majority of PRCC tumors show indolent behavior and have a limited risk of progression and mortality, but a distinct subset displays highly aggressive behavior (5). The biological and clinical aspects of this cancer have been reviewed recently (6).

Delahunt and Eble have proposed that PRCC can be morphologically classified into two subtypes (7). Type 1 is characterized by the presence of small cuboidal cells covering thin papillae, with a single line of small uniform nuclei and basophilic cytoplasm. Type 2 is characterized by the presence of large tumor cells with eosinophilic cytoplasm and pseudostratification. Generally, type 2 tumors have a poorer prognosis than type 1 tumors (8). However, the morphologic classification remains controversial, and there is limited molecular and biochemical evidence to support this morphologic classification. The relatively high incidence of mixed type 1 and 2 tumors poses additional difficulties for such a method of classification. As a result, some recent studies of PRCC do not stratify PRCC into type 1 and 2 tumors (9, 10).

Despite the moderate incidence of PRCC, comparable with that of chronic myeloid leukemia, there is a disproportionately limited knowledge about the underlying molecular basis for development and progression of PRCC. To date, no effective therapy is available for patients with advanced PRCC (11), and patients with PRCC may be excluded from clinical trials that are usually designed for the more common clear cell RCC. It is thus imperative to identify new molecular markers for establishing an accurate diagnosis and prognosis and for developing effective medical therapies for this cancer. Gene expression profiling is a technique that has shown promise in addressing these issues in RCC (12). Recently, we and several other groups of investigators have reported molecular signatures specific for several subtypes of kidney cancer, including PRCC (13–18). PRCC can be effectively distinguished from the other major subtypes of RCC using gene classifiers, from which α-methylacyl-CoA racemase has been additionally validated as a useful immunohistochemical marker (19). However, no distinct molecular subclasses of PRCC were identified in any study possibly because of limited numbers of tumors in previous expression studies (between 2 and 9). We therefore did gene expression profiling on 34 cases of PRCC to search for distinct molecular subtypes of PRCC that were both biologically and clinically relevant.

Patient samples and tissue processing. Institutional review board approval was obtained from each participating institution. Frozen samples of 43 primary tumor specimens with a diagnosis of PRCC after routine pathologic review at each medical center were initially collected following nephrectomy. All tumor specimens were collected from participating institutions in the United States, except one case from Japan. Tumor tissue was flash frozen in liquid nitrogen immediately after nephrectomy and stored at −80°C. Portions of the tumors were fixed in buffered formalin, and H&E-stained slides for all cases were centrally reviewed, except for one case (P30), where slides were not available and histologic description from the pathology report was used for subclassification. We extracted total RNA from homogenized samples using Trizol reageant (Invitrogen, Carlsbad, CA) as described previously. Total RNA was subsequently purified with a RNeasy kit (Qiagen, Montgomery Country, MD), and quality was assessed on denaturing gel electrophoresis. Nine specimens were excluded because of degraded RNA quality. Information on metastatic status at surgery was derived by review of pathologic, radiologic, and intraoperative findings. Clinicopathologic features of the final 34 cases have been provided in Table 1. Twelve noncancerous kidney cortical specimens were also obtained for comparison of gene expression profiling. For histologic evaluation and immunohistochemical analysis, formalin-fixed, paraffin-embedded tissue blocks and sections were obtained from a total of 34 cases. Nineteen of these cases had undergone expression profiling and the additional 15 cases were derived from independent patients, whom did not have tumor tissue profiled.

Expression profiling. For oligonucleotide expression profiling, total RNA (5-20 μg) was used to prepare antisense biotinylated RNA. A subset of cases was added to external poly(A) RNA-positive controls (Affymetrix, Santa Clara, CA). Synthesis of single-stranded and double-stranded cDNA was done with the use of T7-oligo(dT) primer (Affymetrix). In vitro transcription was done using Enzo Bioarray Transcript Labeling kit (Enzo, Farmingdale, NY). The biotinylated cRNA was subsequently fragmented, and 10 μg were hybridized to each array at 45°C over 16 hours. The HGU133 Plus 2.0 GeneChips contain 54,675 probe sets, representing ∼47,000 transcripts and variants. Scanning was done in a GeneChip 3000 scanner. Quality assessment was done in GeneChip Operating System 1.1.1 (Affymetrix) using global scaling to a target signal of 500. Quality assessment was done using denaturing gel electrophoresis. The manufacturer's recommended protocol (GeneChip Expression Analysis Technical Manual, Affymetrix, April 2003) was followed for expression profiling. Median background was 73, median scaling factor was 3.06, and median GADPH 3′/5′ ratio was 1.03, indicative of a high overall array and RNA quality. The data discussed in this publication have been deposited in NCBI's Gene Expression Omnibus (GEO)¹⁹

Data analysis. Statistical analyses were done in the statistical environment R 2.0.1 using packages from the Bioconductor project (20). The robust multichip average algorithm was used to perform preprocessing of the CEL files, including background adjustment, quartile normalization, and summarization. Principal component analysis was used to visualize the 34 expression profiles. Significance analysis of microarrays (SAM) based on two-class unpaired analysis, assumption of unequal group variances, and 10,000 permutations was used to derive a list of genes differentially expressed between tumor subclasses and ordered by relative fold change (21). We did pathway analysis on these genes using Ingenuity Pathway Analysis (Ingenuity Systems, Mountain View, CA), and enrichment of canonical pathways was assessed for significance by a hypergeometric algorithm that did not correct for multiple testing. For derivation of a small gene classifier, we used prediction analysis of microarrays (PAM), a R implementation of nearest shrunken centroids methodology with 10-fold cross-validation over 30 gene thresholds and an offset percentage of 30% (22). Gene predictors corresponding to a minimum misclassification error were obtained, with class discriminant scores calculated for class 1 and 2 tumors as described previously. We inferred cytogenetic profiles for the tumors through the use of a refinement of the comparative genomic microarray analysis (CGMA) algorithm (23), which predicts chromosomal alterations based on regional changes in expression. Relative expression profiles (R) were generated from the single-channel tumor expression profiles (T) and the mean expression values of the 12 single-channel kidney cortical expression profiles (N), such that R = log₂(T) − log₂(N). Survival analysis was done by fitting to a Cox proportional hazards model, and significance was determined by the likelihood ratio test. Two-tailed Student's t test and Fisher's exact testing was used to evaluate correlation between variables and tumor subclassification. For the purpose of this analysis, tumor grade and stage was classified into two categories corresponding to low grade or stage (1 and 2) versus high grade and stage (3 and 4).

Immunohistochemistry. Immunostaining was done on 5 μm thick formalin-fixed, paraffin-embedded sections using the biotin-avidin system (19) with mouse monoclonal antibodies specific for cytokeratin 7 (CK7; 1:50 dilution, DAKO, Carpinteria, CA) and DNA topoisomerase IIα (TopIIα; 1:20 dilution, Vector Laboratories, Burlingame, CA) as described previously. To verify the differential value of CK7 and TopIIα, we studied 19 PRCC samples that had undergone microarray analysis (10 class 1 tumors and 5 class 2 tumors) as well as an independent set of 15 tumors (10 class 1 tumors and 5 class 2 tumors). The 21 class 1 tumors were composed of histologic type 1 (n = 15), low-grade type 2 tumors (n = 3), and mixed type 1/low-grade type 2 tumors (n = 8). The 13 class 2 tumors were all high-grade type 2 tumors. The CK7 immunoreactivity was graded as negative (<0.1% positive tumor cells), focally positive (0.1-10% positive tumor cells), or positive (>10% positive tumor cells). The TopIIα immunoreactivity was graded as negative (<0.1% positive tumor cells), focally positive (0.1-10% positive tumor cells), or positive (>10% positive tumor cells). The Mann-Whitney test was used to evaluate significance of the differential staining.

Morphologic characteristics. Based on histologic features and Fuhrman grading system, we designated four categories of tumors: type 1 (n = 14; Fig. 1A) characterized by small tumor cells of low Fuhrman grade (≤2); type 2A (n = 4; Fig. 1B) characterized by large eosinophilic tumor cells of low Fuhrman grade (≤2); combined type 1 and 2A (n = 5; Fig. 1C); and type 2B (n = 11; Fig. 1D) characterized by large eosinophilic tumor cells with Fuhrman grade (≥3; Table 2).

Molecular characteristics. We visualized the 34 expression profiles by principal component analysis. We noted overlap between histologic type 1 and 2 tumors, contrary to our expectation of distinct molecular subtypes (Fig. 1E). Tumors with mixed type 1 and 2 components (n = 5) grouped with type 1 tumors. PAM with 10-fold cross-validation persistently classified three of four low-grade type 2 tumors with type 1 tumors over a wide range of shrinking gene thresholds (Supplementary Fig. S1A). The only low-grade type 2 tumor that persistently classified with the high-grade type 2 tumors was P30 (the only tumor we were unable to personally evaluate histologically to confirm a reported grade of 2). These results supported a hypothesis that type 2 tumors were molecularly heterogenous. We analyzed the profiles based on this morphologic subtyping into two classes (class 1 corresponding to type 1, low-grade type 2, and mixed type 1/low-grade type 2 tumors and class 2 corresponding to high-grade type 2 tumors) from a molecular viewpoint. Visualization of principal components now showed distinct differentiation between expression profiles of class 1 and 2 tumors, consistent with distinct tumor subclasses (Fig. 1F). Transcripts (n = 796) differentially expressed between class 1 and 2 tumors were identified using SAM at a δ of 1.8, with a false discovery rate of 0.01. We list the top 50 transcripts relatively upexpressed in each subclass (Table 3) and show a hierarchical clustering of the tumor samples based on these 100 transcripts (Fig. 2A). We were able to identify multiple gene classifiers that effectively differentiated class 1 and 2 tumors at 97% accuracy at multiple shrinkage thresholds using PAM (between 7 and 3,881 transcripts) using nearest shrunken centroids methodology (Supplementary Fig. S1B). We report here the seven-transcript predictor that achieved this accuracy (Table 4). Only the tumor of P30, initially reported as a type 2 tumor with grade 2, which we were unable to confirm histologically, persistently classified as a class 2 tumor, rather than as a class 1 tumor, throughout these multiple shrinkage thresholds.

Survival characteristics. Survival analysis (Fig. 1G and H) showed that this refined morphologic and molecular classification system showed a survival prediction that showed a statistically insignificant edge over the previous morphology-based classification approach (Nagelkerke's R² = 0.505 and P = 0.001 versus R² = 0.389 and P = 0.005). Class 2 tumors were larger in tumor dimension (P = 0.003), of higher grade (P < 0.001), of higher stage (P < 0.001), and were more likely to exhibit distant metastases at initial surgery (P < 0.001) than class 1 tumors. Indeed, all tumors metastatic at initial surgery were class 2 tumors (n = 7). No significant difference in age (P = 0.37) or gender (P = 0.70) was found between the two classes.

Chromosomal aberrations inferred by comparative genomic microarray analysis. Distinct cytogenetic profiles for each tumor were generated using high-resolution CGMA (Fig. 2B). Full-length gains in chromosomes 7, 12, 16, 17, and 20 was found both in class 1 and 2 tumors, consistent with the previously reported trisomies observed by using conventional cytogenetic analysis characteristic of PRCC (24, 25). However, in comparison with class 1 tumors, class 2 tumors exhibited more frequent gains at 1q, 2, and 8q and losses at 3p and 6q and showed fewer gains of chromosome 3, 7, and 16. More frequent losses of 6q and 14q were also evident.

Pathway analysis. Genes (n = 203) derived from the 796 transcripts were eligible for generation of networks in pathway analysis. Ranking of canonical pathways yielded three pathways that were significantly enriched within these differentially expressed genes: G₂-M DNA damage checkpoint regulation (P = 0.007), arginine and proline metabolism (P = 0.011), and G₁-S checkpoint regulation (P = 0.018). Genes involved in G₁-S checkpoint regulation (cyclin D2, cyclin-dependent kinase 6, retinoblastoma-like 2, and p21^Cip1) were relatively upexpressed in class 1 tumors (Supplementary Fig. S2), whereas genes involved in G₂-M checkpoint regulation (cyclin B1, cyclin B2, and TopIIα) were relatively upexpressed in class 2 tumors (Supplementary Fig. S3). Multiple oligonucleotide probe sets corresponding to c-met were identified as being upexpressed in class 1 tumors, ranging between 2- and 3-fold upexpression. Details of individual gene expression in the 796 transcripts may be found in Supplementary Table S1.

Immunohistochemical characteristics. The immunohistochemical findings are reported in Table 5, and are consistent between the sets of profiled and independent tumors. The majority of class 1 tumors (86%), including type 1 (Fig. 3A-C) and type 2A (Fig. 3D-F) tumors, showed strong CK7 immunoreactivity (Fig. 3B and E), whereas the majority of class 2 tumors (Fig. 3G-L) showed absent (77%) or reduced (23%) CK7 immunoreactivity in both the set of profiled tumors (Fig. 3H) and the independent set of tumors (Fig. 3K). In contrast, TopIIα immunoreactivity was focally positive (10%) or negative (90%) in class 1 tumors, including both type 1 tumors (Fig. 3C) and type 2A tumors (Fig. 3F). The majority of class 2 tumors were positive for TopIIα (90% positive and 10% focally positive) in both the set of profiled tumors (Fig. 3I) and the independent set of tumors (Fig. 3L). No TopIIα immunoreactivity was detected in normal kidney tissue. There was no apparent difference between type 1 and low-grade type 2 (type 2A) tumors in CK7 and TopIIα immunostaining. Summarizing the results, CK7 immunoreactivity was significantly higher in class 1 tumors (P < 0.001), and TopIIα immunoreactivity was significantly higher in class 2 tumors (P < 0.001).

Morphologic classification. PRCC is the second most common histologic type of RCC comprising ∼10% to 15% of RCC (5) and is composed of tumor cells characteristically forming papillary or tubopapillary structures. The morphologic classification of PRCC into type 1 and 2 tumors has been supported by several histologic studies, although there is relatively limited molecular evidence to substantiate this subtyping. There remains controversy over the recent proposed morphologic classification system of PRCC, preventing its widespread application. For example, there is no agreement whether a tumor with eosinophilic cytoplasm but low nuclear grade should be classified as type 1 or 2. In the initial proposal outlining this morphologic subtyping (7), 63% of type 2 tumors were assessed as being of low Fuhrman nuclear grade despite pleomorphic nuclei being defined as a characteristic of type 2 tumors. More recently, Allory et al. (26) classified only 1 of 13 (8%) as low-grade type 2 tumors using a modified criteria. The high frequency of tumors with coexisting type 1 and 2 components poses difficulties for such a binary classification, the prevalence of such mixed tumors having been reported as high as 28% (26). Allory et al. chose to classify these tumors with mixed (type 1 and 2) features as type 1 tumors, an approach in line with our molecular classification.

Molecular classification. Our results provide only partial support for the proposed histologic subtyping of PRCC into type 1 and 2 tumors. Type 2 tumors are molecularly heterogenous, with a subset of type 2 (low-grade) tumors and mixed type 1 and 2 tumors demonstrating molecular profiles more consistent with type 1 tumors. These type 2 tumors were all low-grade tumors and showed excellent clinical outcomes, in contrast with the poor outcomes recorded in high-grade type 2 tumors. Type 2 PRCC is composed of at least two genetically distinct subtypes: one subtype (type 2A) resembles type 1 in terms of indolent tumor behavior, excellent survival, low tumor grade, similar expression profiles, immunoreactivity, and inferred cytogenetic profiles; the other subtype (type 2B) is an highly metastatic, aggressive cancer that is molecularly distinct from type 1 or 2A tumors. Our findings support a view that nuclear grade is the key correlate for a molecular classification with both biological and clinical relevance, with features such as cell size or cytoplasmic eosinophilia being more peripheral. Additional distinctive histopathologic features for these subclasses may be defined with a larger series. In this report, the molecular classification showed a statistically insignificant edge in prognostication over the previously proposed histologic classification. However, the molecular approach with correlation to nuclear grade may be more relevant, as it also accurately classifies mixed type 1 and 2 tumors, which are not well accounted for in the histologic classification. This refined classification of PRCC based on both morphologic features and molecular studies may be more relevant and is likely to benefit diagnosis, prognostication, clinical follow-up, and experimental selection of therapeutic targets.

We successfully generated an internally validated seven-transcript predictor, which was able to classify class 1 and 2 tumors with 97% accuracy, the only misclassification arising from a tumor (P30) that we were unable to personally evaluate. Consistent with our microarray classification, this tumor from P30 behaved in an aggressive fashion, the patient relapsing 2 years after surgery. The patient died of a non-cancer-related cause 10 months after relapse. External validation in a second population is required for assessment of true generalizability of these gene predictors, but these results are very encouraging.

Inferred cytogenetic profiles. Aneuploidy is well established as a key driver of global gene expression, and regional DNA copy number correlates well with regional expression in cancer (27), which we have also shown in RCC classification (23). PRCC typically shows frequent trisomies 7, 12, 16, 17, and 20 (5, 28, 29); our analysis is consistent with Fig. 2A. For PRCC subclassification, our results are strictly not directly comparable with recent cytogenetic studies that have classified their results by the type 1 and 2 classification (30, 31). As expected, our inferred cytogenetic profiles were consistent with previous studies correlating cytogenetic findings with tumor grade; Lager et al. identifying less frequent trisomy of 7 in high-grade tumors (32) and Renshaw and Corless reporting that trisomy of 3 was found in a defined subset of low-grade PRCC tumors (33). In addition to these findings, in demonstrating that loss of 9q occurred more commonly in class 2 tumors, our results support a report that loss of heterozygosity at 9q is associated with reduced survival (33).

Immunohistochemical findings. To validate the gene predictor and to derive immunohistochemical markers for the pathology laboratory, we used immunohistochemistry to confirm high protein expression of CK7 in class 1 tumors and of Topo IIα in class 2 tumors. CK7 immunoreactivity has been reported previously to the vast majority of PRCC (33), but more recent studies suggested that CK may differentiate type 1 and 2 tumors. Our microarray and immunohistochemical findings were generally consistent with findings using the morphologic classification that between 87% and 100% of type 1 tumors showed CK7 positivity and ∼20% of type 2 tumors showed CK7 positivity (7, 34). No immunohistochemical marker has been reported previously as being specifically upexpressed in type 2 tumors; we showed the usefulness of DNA TopIIα as an immunohistochemical marker in class 2 tumors.

Pathway analysis. Our study highlighted dysregulation of G₁-S checkpoint genes in class 1 PRCC and dysregulation of G₂-M checkpoint genes in class 2 PRCC as the most highly ranked pathways identified in the differentially expressed genes. In familial studies, mutations of the MET proto-oncogene have been implicated in hereditary type 1 PRCC (35) and a small subset (<10%) of sporadic type 1 PRCCs (36). Interestingly, we showed that c-met was differentially expressed, with higher expression in class 1 tumors (Supplementary Table S1). From a mechanistic point of view, this associative link between MET overexpression/mutation and genes associated with G₁-S checkpoint dysregulation is particularly interesting, as hepatocytes in conditional met-mutant mice exhibit defective exit from quiescence and diminished entry into the S-phase of the cell cycle (37). Further work is required to delineate the role of met signaling in G₁-S checkpoint dysregulation. Differential expression of the FH gene, which is mutated in a group of families with type 2 PRCC (38), was not observed (data not shown).

The implication of dysregulation of the G₂-M checkpoint regulation in class 2 tumors is particularly interesting from a therapeutic point of view. We took a particular interest in DNA TopIIα, which we additionally established as a diagnostic marker for class 2 tumors. As there is no effective medical therapy for advanced PRCC and this enzyme is associated with the more aggressive PRCC subclass, TopII inhibitors are distinct possibilities for a therapeutic trial of PRCC. G₂ arrest occurs in response to these agents (39) and may therefore be particularly appropriate. Although several kidney cancer trials have reported disappointing results for TopII inhibitors (40, 41), these trials have predominantly recruited patients with clear cell RCC, a genetically distinct disease. In further support of this suggestion, we note that we have reported previously in a microarray study that this gene is the most overexpressed gene in pediatric Wilms' tumor (15), for which current therapeutic regimens consisting primarily of TopII inhibitors are very effective.

Clonal origin versus progression. It has been hypothesized previously based on cytogenetic findings that type 1 tumors progress to type 2 tumors (31). Prudent evaluation of our results in the context of this hypothesis is required. Although microarrays of gross tumor tissue show a global expression signature presumably reflective of early clonal events (42), it is plausible that a competitive growth advantage may accrue to the transformation of a single cell into a class 2 within a class 1 tumor, resulting in its expansion at the expense of other class 1 tumor cells. Nonetheless, the additional presence of a distinct group of mixed tumors with coexisting type 1 and 2A histology and presenting with molecular profiles resembling other type 1 tumors strongly suggests that type 1 and 2A tumors are clonally more closely related to each other than to type 2B tumors. We did not note the presence of low-grade components in any of our type 2B tumors. Given the divergent survival outcomes following nephrectomy between class 1 (type 1, type 2A, and mixed type 1/2A tumors) and class 2 tumors, we do not favor the idea of progression between class 1 and 2 tumors.

In conclusion, using gene expression profiling supported by immunohistochemical and morphologic studies, we have identified two distinct classes of PRCC that differ strikingly in their clinical behavior and have dysregulation of genes controlling different parts of the cell cycle. This finding represents a biologically and clinically relevant refinement to previously proposed morphologic criteria for subclassification of PRCC. We summarize our findings that may be practically evaluated in the clinical setting laboratory as follows: class 2 (type 2B) PRCC may be distinguished from class 1 (type 1, mixed type 1 and 2A, and type 2A tumors) by the following characteristics: larger gross tumor size, higher nuclear grade (3-4), decreased CK7 staining and increased TopIIα staining, higher rate of metastases at surgery, and poorer patient survival. Morphologic findings of less specificity include larger cell size and eosinophilic cytoplasm in class 2 tumors. Our findings may benefit further efforts to elucidate the molecular basis of development and progression of PRCC and will be helpful in stratifying patients for additional interventions.

Grant support: Gerber Foundation, Hauenstein Foundation, Fischer Family Trust, and Michigan Technology Tri-Corridor. Four tumor specimens were provided by the Cooperative Human Tissue Network, which is funded by the National Cancer Institute.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

We thank Dr. Chongfeng Gao for constructive discussion.