Critical Evaluation of a microRNA-Based Risk Classifier Predicting Cancer-Specific Survival in Renal Cell Carcinoma with Tumor Thrombus of the Inferior Vena Cava

Simple Summary Renal cell carcinomas (RCCs) can build a so-called tumor thrombus and grow into the renal vein and the vena cava, thereby representing a high-risk situation for affected patients. Nevertheless, there are substantial differences in the clinical courses of RCC patients with a tumor thrombus. Currently, there is no established biomarker which helps identifying patients with a substantial risk of tumor relapse or even cancer-related death. Previously, members of our group discovered a signature of three small RNA molecules (specifically: microRNAs) in RCC tissue, which significantly predicted cancer-related survival. In this study, we validated this signature in a larger cohort of patients suffering from RCCs with a tumor thrombus. Notably, stratifying our patients according to this microRNA signature nearly separated our cohort into two halves, which significantly differed in terms of clinical risk. Our research could help identifying high-risk patients in need for additional therapy, while sparing others from unnecessary treatments. Abstract (1) Background: Clear cell renal cell carcinoma extending into the inferior vena cava (ccRCCIVC) represents a clinical high-risk setting. However, there is substantial heterogeneity within this patient subgroup regarding survival outcomes. Previously, members of our group developed a microRNA(miR)-based risk classifier—containing miR-21-5p, miR-126-3p and miR-221-3p expression—which significantly predicted the cancer-specific survival (CSS) of ccRCCIVC patients. (2) Methods: Examining a single-center cohort of tumor tissue from n = 56 patients with ccRCCIVC, we measured the expression levels of miR-21, miR-126, and miR-221 using qRT-PCR. The prognostic impact of clinicopathological parameters and miR expression were investigated via single-variable and multivariable Cox regression. Referring to the previously established risk classifier, we performed Kaplan–Meier analyses for single miR expression levels and the combined risk classifier. Cut-off values and weights within the risk classifier were taken from the previous study. (3) Results: miR-21 and miR-126 expression were significantly associated with lymphonodal status at the time of surgery, the development of metastasis during follow-up, and cancer-related death. In Kaplan–Meier analyses, miR-21 and miR-126 significantly impacted CSS in our cohort. Moreover, applying the miR-based risk classifier significantly stratified ccRCCIVC according to CSS. (4) Conclusions: In our retrospective analysis, we successfully validated the miR-based risk classifier within an independent ccRCCIVC cohort.

To further assess the miR-based risk classifier as a prognostic tool in ccRCC IVC patients, we retrospectively evaluated it within an independent cohort (n = 56) from the Department of Urology, University of Regensburg (Regensburg, Germany). Cut-off values for miR expression levels, as well as internal classifier weights, were transferred from the previous pilot study [10], in order to test its transferability to independent study cohorts. Figure 1 illustrates the course of our study.
To further assess the miR-based risk classifier as a prognostic tool in ccRCC IVC patients, we retrospectively evaluated it within an independent cohort (n = 56) from the Department of Urology, University of Regensburg (Regensburg, Germany). Cut-off values for miR expression levels, as well as internal classifier weights, were transferred from the previous pilot study [10], in order to test its transferability to independent study cohorts. Figure 1 illustrates the course of our study. Figure 1. Course of the study-using a previously established microRNA (miR)-based risk classifier [10], we examined the prognostic impact of miR-21, miR-126, and miR-221 expression in an independent cohort of clear cell renal cell carcinoma samples with infiltration of the inferior vena cava (ccRCC IVC ; n = 54). To assess the transferability of the miR-based risk classifier, cut-off values and weights were identical to the previous study.

RNA Extraction and qRT-PCR
Using the RecoverAll TM Total Nucleic Acid Isolation Kit (Thermo Fisher Scientific, Waltham, MA, USA), total RNA from paraffin-embedded samples was isolated according to the manufacturer's instructions. RNA concentrations and 260/280 ratios were analyzed by Spark ® 10M (TECAN, Männedorf, Switzerland). cDNA was synthesized from total RNA with stem-loop reverse transcription primers (TaqMan microRNA Assay protocol, Applied Biosystems, Birchwood, UK). The TaqMan microRNA Assay kit was used to quantify miR expression according to manufacturer's protocols. Table 1 lists the specific primers and kits used in this study.
Samples exhibiting a standard deviation > 0.5 were excluded (all reactions performed in triplicates). Small nuclear RNA (RNU6B) expression was used for the normalization of relative miR expression values. Samples with expression levels of RNU6B > 30 Ct were excluded from further analyses. Relative miR expressions were calculated using the ∆Ctmethod (∆Ct sample = Ct miR of interest − Ct RNU6B). To calculate fold changes in miR expression between samples, we used the 2∆∆Ct method (hereafter referred to as the ∆∆Ct method). To select the best fitting Cox model, the relative goodness-of-fit was measured based on the Akaike information criterion (AIC) for different variable combinations. The combination of miR-21, -126, and -221 displayed the best prediction properties. • Finally, the miR-based risk classifier was calculated based on the publication by Lossos et al. [22]. Hereby, the z-factor calculated by the multivariable Cox model (R package 'survival') for miR-21, -126, and -221 was multiplied with the relative expression levels (∆Ct) of the respective miR. This approach resulted in the following formula: (4.592 × ∆Ct miR-21) + (−3.892 × ∆Ct miR-126) + (−1.938 × ∆Ct miR-221). • Subsequently, a receiver operating characteristic (ROC) curve was plotted (R package pROC), showing the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings. The risk score cut-off/threshold of 18.7 ∆Ct was determined from ROC using the Youden's J statistics [23] (integrated in pROC package [24], default settings).

Clinical Validation of microRNA-Based Risk Classifier
miR-specific weights for individual ∆Ct values and the cut-off value (≥18.7 ∆Ct = "unfavorable subgroup", <18.7 ∆Ct = "favorable subgroup") were transferred from the pilot study [10] and applied in order to stratify the ccRCC IVC study cohort from Regensburg and perform Kaplan-Meier analyses. Within the risk classifier formula, a negative factor indicated that higher expression levels correlated with longer survival, whereas a positive factor correlated with shorter survival. For further analysis, we also transferred cut-off values for miR-21 (8.17 ∆Ct), miR-126 (3.57 ∆Ct), and miR-221 expression (1.84 ∆Ct) to evaluate their predictive potential in the new cohort using the Kaplan-Meier survival analysis. Table 2 summarizes the basic clinical and pathological characteristics of our study cohort. Follow-up information was available for 56 patients, the precise date of death for two patients was unknown. To further characterize our patient cohort, we performed Kaplan-Meier analyses regarding the clinicopathologic parameters nodal status, synchronous metastasis and Fuhrman grade. As illustrated in Figure 2a, nodal status significantly stratified the patients regarding CSS (p = 0.00024). For synchronous metastasis (Figure 2b), stratification was not statistically significant. In contrast, CSS according to Fuhrman grade 2 vs. 3 (Figure 2c) was significantly different (p = 0.011). While being statistically significant for nodal status and Fuhrman grade, stratification based on these parameters only reached sensitivities of 46% and 54%, respectively. To further characterize our patient cohort, we performed Kaplan-Meier analyses regarding the clinicopathologic parameters nodal status, synchronous metastasis and Fuhrman grade. As illustrated in Figure 2a, nodal status significantly stratified the patients regarding CSS (p = 0.00024). For synchronous metastasis (Figure 2b), stratification was not statistically significant. In contrast, CSS according to Fuhrman grade 2 vs. 3 (Figure 2c) was significantly different (p = 0.011). While being statistically significant for nodal status and Fuhrman grade, stratification based on these parameters only reached sensitivities of 46% and 54%, respectively.

Association of miR-21, -126, and -221 Expression with Clinicopathological Characteristics
To investigate the impact of miR-21, -126, and -221 within our ccRCC IVC cohort, we associated expression levels of miR-21, miR-126, and miR-221 with the relevant clinical parameters. Figure 3 illustrates the results.
At time of surgery, 11 of 56 ccRCC IVC patients (19.6%) were diagnosed with lymphonodal metastasis. In cases with a positive nodal status, trends towards the up-regulation of miR-21 (p = 0.065) and a significant down-regulation of miR-126 (p < 0.01) were observed. For miR-221, there was no statistically significant association to nodal status.
Of 56 patients with ccRCC IVC , 13 (23.2%) died during the follow-up period due to cancer (cancer-related death, CRD). Regarding miR expression levels, we found a significant up-regulation of miR-21 (p < 0.001) and a down-regulation of miR-126 (p < 0.01) in CRD cases. Instead, miR-221 expression did not show a statistically significant association with CRD in this analysis (p = 0.27).
At time of surgery, 11 of 56 ccRCC patients (19.6%) were diagnosed with lymphon-odal metastasis. In cases with a positive nodal status, trends towards the up-regulation of miR-21 (p = 0.065) and a significant down-regulation of miR-126 (p < 0.01) were observed. For miR-221, there was no statistically significant association to nodal status.
For comparison, we also investigated the expression levels of miR-21, miR-126, and miR-221 depending on survival within the ccRCC cohort of the TCGA database ( Figure S1). In line with our findings, the upregulation of miR-21 was significantly associated with death in patients with pT3b tumors. In contrast, miR-126 was significantly downregulated in these cases, whereas miR-221 levels did not differ significantly.

Single-Variable and Multivariable Cox Regression Analysis
Next, we performed single-variable Cox regression analysis to further assess the prognostic potential of miR-21, -126, and -221 expression levels as predictors of CRD. Detailed follow-up information was available for 54 of 56 cases, with a median of 94 months.
As summarized in Table 3a, miR-21 and miR-126 significantly predicted the occurrence of CRD in our study cohort (p = 0.003, hazard ratio (HR) 3.79 for miR-21, p = 0.00003, HR 0.19 for miR-126). In contrast, miR-221 expression in tumor tissue did not display significant prognostic potential (p = 0.22). Regarding further clinical parameters, significant results were also observed for nodal involvement, metastatic status, and Fuhrman grade.
Additionally, as shown in Table 3b and Figure 4, we performed multivariable Cox regression to investigate whether the miR candidates remained as relevant predictors of CRD in attendance of nodal involvement and Fuhrman grade. Again, miR-21 and miR-126 significantly predicted the occurrence of CRD (p = 0.02, HR 4.94 for miR-21, p = 0.01, HR 0.27 for miR-126). miR-221 expression again did not meet statistical significance as a predictor of CRD (p = 0.12) ( Table 3 and Figure 4). No significant results were observed for the clinical parameters nodal status (p = 0.71) and Fuhrman grade (p = 0.13), either. To further understand and illustrate the prognostic impact of single miRNAs on CSS (based on previously fitted multivariable Cox regression model), we plotted survival curves according to isolated miR expression levels ( Figure S2). For miR-21, higher relative expression levels were associated with lower CSS ( Figure S2a). In contrast, higher expression levels of miR-126, as well as miR-221, were associated with higher CSS (Figure S2b,c). expression levels were associated with lower CSS ( Figure S2a). In contrast, higher expression levels of miR-126, as well as miR-221, were associated with higher CSS (Figure S2b,c).

Kaplan-Meier Analyses for Single miR Expression and the Risk Classifier
To investigate the prognostic validity of single miRNA expression levels, survival analyses using the identical cut-offs from Vergho et al. [10] were performed. Both miR-21 (p = 0.006) and miR-126 (p = 0.00028) showed a strong predictive significance in the

Kaplan-Meier Analyses for Single miR Expression and the Risk Classifier
To investigate the prognostic validity of single miRNA expression levels, survival analyses using the identical cut-offs from Vergho et al. [10] were performed. Both miR-21 (p = 0.006) and miR-126 (p = 0.00028) showed a strong predictive significance in the Kaplan-Meier survival analysis (Figure 5a,b). However, differences regarding CSS of miR-221 high vs. low expressing tumor specimens (Figure 5c) did not reach statistical significance (p = 0.25). As illustrated in Figure 5d, combing all three miRs within the risk classifier and maintaining the established cut-off level of 18.7 ∆Ct (≥18.7 ∆Ct = unfavorable subgroup, <18.7 ∆Ct = favorable subgroup) revealed the most significant stratification (p = 0.000075).

Testing the Composition and Robustness of the Risk Classifier
In line with the results from the Würzburg cohort investigated in a previous publication [10], patient stratification based on single miR-221 expression did not demonstrate significant results. Therefore, we checked whether adding miR-221 expression to our Cox regression model contributed to its performance. Measuring the goodness-of-fit with the Akaike information criterion (AIC), the combination of miR-21, miR-126, and miR-221 again emerged as the best fitting Cox model with the lowest partial AIC (Table S1). The coefficient for distant metastases (Table 3) was not estimable within Cox regression (positively infinite); therefore, we did not include this variable in our analysis.
Finally, we investigated the robustness of the risk classifier in terms of the varying cut-off values applied. We calculated sensitivities, specificities, and p values for cut-offs ranging from 16.0 ΔCt to 25.0 ΔCt (Table S2). In general, all cut-offs produced highly sig- (d) Combined miR-based risk classifier (miR-21, -126, -221) using identical cut-offs and weights from a previous publication [10]. p values from log-rank tests are shown within each plot.

Testing the Composition and Robustness of the Risk Classifier
In line with the results from the Würzburg cohort investigated in a previous publication [10], patient stratification based on single miR-221 expression did not demonstrate significant results. Therefore, we checked whether adding miR-221 expression to our Cox regression model contributed to its performance. Measuring the goodness-of-fit with the Akaike information criterion (AIC), the combination of miR-21, miR-126, and miR-221 again emerged as the best fitting Cox model with the lowest partial AIC (Table S1). The coefficient  (Table 3) was not estimable within Cox regression (positively infinite); therefore, we did not include this variable in our analysis.
Finally, we investigated the robustness of the risk classifier in terms of the varying cutoff values applied. We calculated sensitivities, specificities, and p values for cut-offs ranging from 16.0 ∆Ct to 25.0 ∆Ct (Table S2). In general, all cut-offs produced highly significant results. Higher values (21.0 ∆Ct to 25.0 ∆Ct) revealed a patient stratification with even lower p values; this rise in significance was accompanied by a decline in sensitivity-from 92% to 85% for the cut-off values 21.0-24.0 ∆Ct and 62% for the cut-off value 25.0 ∆Ct. Kaplan-Meier curves for the cut-off values 16.0, 17.0, 20.0, and 22.0 ∆Ct are shown in Figure S3.

Discussion
ccRCCs infiltrating the inferior vena cava represent a clinically relevant high-risk subgroup. However, there was substantial clinical heterogeneity within this distinct subgroup-biomarkers were needed to assess the individual risk of progression. To address this, researchers have investigated the influence of clinical characteristics-showing the prognostic impact of variables such as perinephric fat invasion, body mass index, metastasis, and Fuhrman grade [8,25,26]. Regarding our in-house cohort, synchronous metastasis did not meet statistical significance-probably due to small sample size and limited follow-up. Although nodal status and Fuhrman grade demonstrated statistical significance, low sensitivities severely limited clinical applicability in our patient cohort.
In line with nomograms and prediction models in metastasized RCC [27,28], researchers have also investigated the prognostic role of blood counts in RCC cases with tumor thrombus. Specifically, serum gamma-glutamyltransferase (GGT) [29], as well as neutrophil-tolymphocyte ratio [30,31] and lymphocyte-to-monocyte ratio [32], were reported to significantly predict survival in patients suffering from RCC with tumor thrombus.
In general, adjuvant therapy with tyrosine kinase inhibitors (TKIs) or immune checkpoint blockers [33] could be a promising therapeutic option after nephrectomy-especially for patients suffering from high-risk RCC. However, European kidney cancer guidelines currently do not contain strong recommendations towards adjuvant therapies due to the mixed outcome in clinical trials [34]. For the TKI sunitinib, one trial indicated improved disease-free survival (DFS) for patients-although showing no significant differences in overall survival (OS) [35]. Additionally, another phase 3 trial did not identify significant survival effects for adjuvant sunitinib or sorafenib in nonmetastatic high-risk renal cell carcinoma [36]. Due to the sobering TKI results, research efforts have mainly shifted towards immune checkpoint blockers. For the PD-1 (Programmed Cell Death Protein 1) inhibitor pembrolizumab, the KEYNOTE-564 trial identified improved progression-free survival (PFS) in an adjuvant setting after nephrectomy [37].

Evaluating an miR-Based Risk Classifier for RCC with Infiltration of the Vena Cava
To estimate the individual risk of patients suffering from ccRCC IVC , members of our research group have established a risk classifier based on the tissue expression of miR-21, miR-126, and miR-221 [10]. The former cohort contained tumor tissue of n = 37 patients undergoing surgery at the University Hospital of Würzburg, Germany. In this study, we externally validated the prognostic potential of the miR-based risk classifier. Therefore, we examined an independent cohort of ccRCC IVC from the University of Regensburg, Germany (n = 56). To test the transferability and usability of the classifier within an external tissue cohort, we applied identical cut-off values and weights as in the previous pilot study.
In terms of clinicopathological characteristics, miR-21 expression was significantly higher in patients with a positive lymphonodal status, metastatic disease, and in cases of CRD. In contrast, miR-126 expression levels were significantly lower for these clinical scenarios. Within the single-variable Cox regression, both miRs also demonstrated prognostic significance regarding cancer-specific survival (CSS). Lower levels of miR-221 in tumor tissue and its association with CRD did not reach statistical significance. Beyond miR expression, Fuhrman grade, lymphonodal status, and the occurrence of metastasis emerged as prognostically relevant. Next, we added multivariable Cox regression for the three miR candidates. Again, this study identified miR-21 and miR-126 expression levels to significantly predict CRD. Kaplan-Meier analyses based on cut-off values previously determined by Vergho et al. [10] confirmed the significant role of miR-21 as well as miR-126 expression levels for CSS prediction. Finally, the miR-based classifier was applied, using identical cut-off values and weights to split patients into two groups.
Notably, the combined risk classifier nearly stratified the study cohort into two halveswith n = 26 patients belonging to the favorable subgroup and n = 28 patients belonging to the unfavorable subgroup. Regarding the substantial difference in CSS between both groups, adjuvant therapies appear promising especially for the unfavorable subgroup of ccRCC IVC patients.
Given the fact that miR-221 failed to demonstrate statistical significance within singlevariable and multivariable analysis, we re-checked whether combining the expression levels of all three miRs represented the ideal constitution of the risk classifier. As previously demonstrated [10], AIC analysis again confirmed that taking into account miR-221 expression improved the overall performance of the Cox regression model.
We finally challenged the robustness of the applied cut-off value (18.7 ∆Ct) by examining the results of patient stratification in terms of sensitivity, specificity, and significance. Except for the neighboring cut-off 19.0 ∆Ct, higher levels of significance were only reached with lower sensitivities. In conclusion, the risk classifier demonstrated its transferability not only regarding its components, but also in terms of the previously established cut-off value.

Functional Roles of miR-21, miR-126, and miR-221 in Cancer and Thrombosis
After confirming the prognostic potential of the miR classifier using the validation cohort from Regensburg, we were interested in the previously reported functions of these miRs in RCC and other malignancies. For miR-21, several researchers have demonstrated oncogenic effects in various cancers, including RCC [38,39]. Among the prominent miR-21 target genes are key players of apoptosis induction such as PDCD4 (Programmed Cell Death 4) [40] and genes such as PTEN (Phosphatase and Tensin Homolog) [38]. The latter is an established tumor-suppressor gene best known for regulating PI3K/Akt signaling. In contrast to miR-21, miR-126 acts as a tumor suppressor in tumor tissue, e.g., by targeting ROCK1 (Rho-Associated Coiled-Coil Containing Protein Kinase 1) [41] and VEGFA (Vascular Endothelial Growth Factor A) [42]. For miR-221, oncogenic versus protective functions appear to depend on the underlying cancer entity; researchers have demonstrated both roles [43][44][45]. For RCC, a downregulation of miR-221 appears well in line with previous publications. Specifically, miR-221 is reported to regulate KDR (Kinase Insert Domain Receptor)-also known as VEGFR2 (Vascular Endothelial Growth Factor Receptor 2)-in ccRCC [46] and prostate cancer [47], thereby regulating the sensitivity towards sunitinib. In summary, among diverse tumorigenic functions of these miRs, all three candidates prominently influence angiogenesis-related pathways (so-called AngiomiRs) [48,49]. Given that not all ccRCCs depend on angiogenic signaling to the same degree [50], it is tempting to assume that the unfavorable subgroup identified by our risk classifier could benefit from adjuvant anti-angiogenic therapy.
Notably, expression levels of these three miRs have reportedly been deregulated in thrombotic and embolic events [51]. miR-221 plasma levels, for example, were elevated in patients suffering from pulmonary embolism [52]. Moreover, Wang et al. found miR-21 to be part of an miR signature significantly associated with recurrent thromboembolism [53]. However, further research is needed to elucidate the potential impact of these miRs on a tumor thrombus.

Limitations and Future Directions
Our study has several limitations. Leaving aside the definite RCC subgroup investigated here, sample size of our study is relatively small. Moreover, we purposely did not adjust cut-off values and individual miR weights determined previously in order to check the transferability of the risk classifier to external tissue cohorts. We additionally did not consider the operative duration, which could affect miR expression due to differing manifestations of tissue hypoxia. However, hypoxia does not solely depend on the length of the surgical procedure; it also depends on the surgical technique performed. Moreover, miR expression levels in primary tumors do not necessarily represent expression levels within cancer cells contained in the tumor thrombus. Warsow et al. investigated this potential discrepancy by performing whole-exome sequencing and found substantial heterogeneity regarding mutational signatures [54].
More research-ideally in a prospective setting-could further validate the risk classifier in a clinical setup and elucidate whether the sub-classification of ccRCC IVC is able to identify patients most susceptible towards adjuvant therapy.

Conclusions
Although RCC extending into the inferior vena cava represents a high-risk setting, there is still substantial clinical heterogeneity within this patient subgroup. Previously, Vergho et al. established an miR-based risk classifier-containing miR-21, miR-126, and miR-221 expression-which significantly predicted CSS for patients from this subgroup. To validate this classifier, we examined its impact on an external and independent patient cohort. Using identical cut-off values for single miRs and identical weights within the classifier, we confirmed a highly significant risk stratification within the new cohort. Patients with an unfavorable constellation according to the miR-based classifier could especially benefit from adjuvant therapy and continuous follow-up examinations.
Supplementary Materials: The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/cancers15071981/s1. Figure S1: miR-21 (a), miR-126 (b) and miR-221 (c) expression levels depending on overall survival status in pT3b ccRCC samples (n = 54) of The Cancer Genome Atlas (TCGA) database; Figure S2: Survival curves for varying miR expression levels of miR-21 (a), miR-126 (b) and miR-221 (c)-based on fitted multivariable Cox regression model; Figure S3: (a-d) Kaplan Meier survival analysis for CSS for external independent ccRCC IVC (n = 54) cohort from Regensburg stratified by Combined miR-based risk classifier (miR-21, -126, -221) using various cut-off values.; Table S1: Summary of partial AIC (Akaike information criterion) values for different multivariable Cox model combinations; Table S2: Sensitivities, specificities and associated p values for the combined risk classifier with differing cut-off levels.  Institutional Review Board Statement: All procedures performed in studies involving human participants complied with the ethical standards laid down in the latest declaration of Helsinki of the World Medical Association. The study was approved by the local Ethics Committees (Regensburg: Nr. 08/108, Würzburg: Nr. 136/08). All involved patients gave their written and informed consent for the publication of their anonymized cases.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data presented in this study are available within the article or supplementary material.