Background
Clear cell renal cell carcinoma (ccRCC) is the most common subtype of renal cell carcinoma, with a currently increasing incidence [
1‐
3]. Approximately 30% of patients develop metastases and, despite the implementation of targeted therapies, the 5 year survival rate of patients with metastatic disease remains below 20%. Thus, stratification of patients with ccRCC into different molecularly defined groups to identify patients at risk of worse outcome is increasingly important in the perspective of personalized medicine. With this in mind, several prognostic scores have been developed based on, for example, pathological features, gene expression, or DNA methylation status [
4‐
6]. One of the most widely applied score established on clinicopathological data is the SSIGN (stage, size, grade, and necrosis) score [
7,
8], whereas the ClearCode34 score, which predicts two ccRCC subtypes (ccA/ccB), has been suggested for prediction of survival using gene expression data [
9,
10]. Moreover, Rini et al. [
11] proposed a 16-gene score to predict recurrence in ccRCC patients. In general, prognostic signatures using RNA-seq data hold great promise for precision oncology, as previously demonstrated for lung adenocarcinoma [
12]. We recently developed an in silico prediction score (named S3-score) for ccRCC, based on the gene expression of 97 signature genes and the similarity of gene expression between tumor cells and their proposed normal cell of origin in the nephron [
13]. The S3-score outperforms several other scores [
13], including the ClearCode34 model, and significantly improves the predictive value of the SSIGN score and the original ccA/ccB assignment based on clustering [
14]. Moreover, compared with the ccA/ccB signature, the S3-score is slightly less dependent on the tumor section investigated [
13] and, in consequence displays little intra-tumor heterogeneity. This is of importance because, in a recent study investigating the ccA/ccB signature [
10], approximately one-quarter of metastatic tumors (two of nine patients) displayed intra-tumor heterogeneity and, in 43% of the cases, patient-matched primary and metastatic tumors displayed different molecular ccA/ccB subtypes. In this context, a recent multiregion sampling process using a protein-based prognostic model was described, enabling the study of the impact of intra-tumor heterogeneity on risk stratification of sunitinib-treated metastatic patients [
15].
As our S3-score was evaluated only in silico using data from The Cancer Genome Atlas (TCGA), we now intended to verify the performance of the score using newly generated whole transcriptome data of an independent cohort of ccRCC patients, including metastases derived from ccRCC, to determine the concordance of the score prediction in primary tumors and ccRCC-derived metastases. Moreover, we evaluated whether the score predicts outcome in sunitinib-treated ccRCC patients. Finally, our objective was to improve the clinical applicability of the S3-score by reducing the number of genes necessary for calculation of the score and by using the more cost-effective real-time PCR technology.
Discussion
Several risk scores based on gene expression data have been developed for prediction of patient survival in ccRCC [
4]. We recently developed a novel prediction score, named the S3-score, based on the similarity of gene expression in the tumor to its cell of origin in the nephron region [
13,
21]. Thus, in contrast to other scores, risk prediction using the S3-score is related to biologic alterations of the cell of origin of ccRCC. The S3-score outperformed other scores or signatures based on gene expression data or clinicopathological variables [
13,
21] and was even able to improve the predictive value of the clinically validated SSIGN score [
7,
8]. Moreover, evaluation of tumor heterogeneity of our S3-score showed that only a few samples displayed heterogeneity [
13], which indicates that risk prediction with our score is largely independent from the tumor region investigated.
Generally, most of the scores developed using gene expression data are thus far not introduced into clinical practice because they have not been generated to evaluate individual patients. Thus, for clinical application, the prediction scores need to be validated in several studies defining optimal cut-off values for classification of individual patients into subtypes. Moreover, prediction scores, typically developed using genome-wide gene expression data, need to be evaluated using different technologies and gene expression platforms. Since our S3-score, which is based on the expression of 97 signature genes, was originally developed using RNA-seq data from the TCGA, we first evaluated its predictive ability in the present work using gene expression data generated through microarray technology in our own ccRCC cohort. Here, we showed not only that a platform transfer to microarray data is possible, but also that the S397-score significantly predicts CSS in our cohort.
In contrast to other prediction scores such as the 16-gene signatures [
11], the 97 marker genes were not selected based on pathway analyses (e.g., including genes related to inflammation or immune response) and subsequent optimization for prediction of prognosis, but were originally selected to show that tumor aggressiveness in RCC correlates with the level of divergence from its cell of origin within the nephron region. Noticeably, we observe an overlap of one vascular pathway gene (PPAP2B) in the 97 marker genes and those genes from the 16-gene signature described by Rini et al. [
11]. Further studies are warranted to compare the predictive ability of both scores.
Because metastases might represent the most aggressive phenotypes of a heterogeneous tumor, herein, we were interested in inter-tumor or metastases heterogeneity, using gene expression data generated by microarray technology once again. Interestingly, the predictive S3
97-score was comparable between matched tumor and metastases, or matched metastases pairs. In only one case (Additional file
1: Table S5) classification differed between metastases and tumors.
Since data on treatment outcome are limited in the TCGA cohort originally used to develop the S3-score, we were not previously able to evaluate the effect of tyrosine kinase inhibitor (TKI) treatment on outcome prediction. Therefore, herein, we investigated the S3
97-score using microarray data from a cohort of sunitinib-treated ccRCC patients. In this cohort, the S3
97-score was significantly associated with progression-free survival of patients, indicating that our score enables even prediction of sunitinib outcome. Whether the same holds true for immunotherapy in the form of T cell immune checkpoint inhibitors like nivolumab needs to be investigated in future studies. Preliminary investigation of the S3
97-score in metastatic RCC patients treated with nivolumab [
22] shows that the S3
97-score did not differ significantly in pre- and post-treatment biopsies (Additional file
1: Figure S6), indicating that there was no influence of treatment with nivolumab on the S3-score.
Taken together, we provide evidence that the S397-score is more widely applicable than originally intended. To provide a more cost-effective approach for clinical application of the S3-score in individual patient samples, such as even formalin-fixed paraffin-embedded samples, we improved the S397-score by reducing the number of signature genes from 97 to 15 especially for expression analyses through RT-PCR. Our improved S315-score was validated using RT-PCR technology in a cohort of 108 ccRCC cases, clearly indicating that the S315-score was associated with CSS in the complete cohort, as well as non-metastatic and metastatic subsets. Moreover, the S315-score improves prediction of CSS by the currently clinically applied SSIGN score, which is based on clinical parameters and pathologic features. Finally, the S315-score allows risk prediction in tumor and metastases tissue.
In summary, we found that our score enables valid prediction of patient outcome even if applied to different sample types (e.g., primary and metastatic tissue) and independent cohorts (e.g., patients treated with TKIs). Moreover, different platforms (RNA-seq, microarray) and technologies more appropriate for clinical utility (qRT-PCR) can be used for prediction of patient risk by the S3-score. Further prospective studies are warranted to assess the implementation of the score into clinical practice with consequences on personalized patient care.
Conclusions
Since the stratification of patients to identify those with worse prognosis is increasingly important, especially for treatment selection, the molecular subtyping through gene expression signatures may be promising for ccRCC patients. In the present work, the clinical utility of the gene expression-based S3-score, which reflects the similarity of the tumor to its cell of origin in the nephron, was assessed in independent cohorts. The 97 gene-based S397-score and a simplified 15-gene RT-PCR-based S315-score are significantly associated with CSS or progression-free survival in non-metastatic and metastatic ccRCC patients, as well as in TKI-treated patients. As a result, this score, as a promising, cost-effective, and robust diagnostic tool, enables the risk stratification of patients with ccRCC in clinical practice in the non-metastatic, metastatic, and sunitinib-treated setting.