Gene Expression Signature Correlates with Outcomes in Metastatic Renal Cell Carcinoma Patients Treated with Everolimus Alone or with a Vascular Disrupting Agent

This article is featured in Highlights of This Issue, p. 1325

Studies focused on the molecular underpinnings of renal cell carcinoma (RCC) have identified the VEGF and mTOR pathways as promising therapeutic targets. Frequent dysregulation of the von Hippal–Lindau (VHL) gene, particularly in tumors with clear cell histology leads to accumulation of the hypoxia-inducible factor (HIF) subunits, resulting in an upregulation of VEGF and other proangiogenic factors (1, 2). The mTOR is a conserved serine/threonine kinase that orchestrates cell growth, metabolism, survival, and angiogenesis (3, 4). mTOR is composed of two distinct protein complexes, mTOR complex 1 (mTORC1) and mTORC2 (5). Increased mTORC1 activity leads to multiple metabolic and anabolic derangements that have an impact on synthesis machinery of nucleotides, lipids, amino acids, and biogenesis of ribosomes (6–8). Overall, these mTOR-related functions are relevant to RCC and implicate mTOR inhibition as a mechanism of tumor vulnerability.

Large randomized clinical trials on rapalogs in the metastatic renal cell carcinoma (mRCC) setting led to the FDA approval of temsirolimus and everolimus as single-agent therapy in the first-line and post-VEGF inhibitor settings, respectively (9, 10). However, rapalog monotherapy has only shown modest efficacy and limited success in the mRCC setting. As such, prior trials attempted to capitalize on the unique vulnerabilities of RCC by exploiting and targeting both the angiogenesis and mTOR signaling pathways, for example, lenvatinib combined with everolimus (11–15). Following the advent of first-line PD1/L1 inhibitor therapy combined with CTLA-4 inhibition (nivolumab plus ipilimumab) or VEGF inhibitor therapy (pembrolizumab/avelumab plus axitinib), the optimal second-line therapy is unclear (15–19).

We hypothesized that given a relatively favorable therapeutic index and lack of effective options following immune checkpoint and VEGF inhibitors, a continued role of salvage mTOR inhibitor monotherapy or in combination with novel agents may exist in selected patients with mRCC. A trial conducted by Pal and colleagues (12) explored the efficacy and safety of everolimus and a vascular disrupting agent (VDA), BNC105P, following VEGF inhibitors. This was later compared with another arm of patients randomized to everolimus alone. BNC105P (Bionomics Limited) is the disodium phosphate ester prodrug of BNC105, a tubulin polymerization inhibitor that exhibits selective actionability against tumor vasculature endothelium (20, 21). BNC105P leads to prominent occlusion of tumor vasculature and induces hypoxic stress that can be lethal to tumor cells, especially cells with aberrant VHL signaling (20). Unfortunately, this combination therapy did not improve outcomes compared with rapalog therapy alone. Understanding why this combination did not achieve added benefit remains an important area that warrants further investigation. Herein, we used primary specimens from the trial and explored for the first time the association between transcriptomic gene signatures and clinical outcomes in patients treated with either rapalog therapy alone or in combination with BNC105P. We then validated our findings in a publicly available phase III trial dataset (CheckMate 025) using a subset of 130 patients with RNA-sequencing (RNA-seq) data and treated with everolimus.

We identified a discovery cohort of 107 patients with mRCC from whom we obtained written informed consent to participate in the phase I/II randomized trial of everolimus ± BNC105P for progressive disease following VEGF inhibitors conducted by the Hoosier Cancer Research Network (HCRN). We included all subjects from both arms, for whom pretreatment formalin-fixed paraffin-embedded (FFPE) tissue was available, and sufficient RNA was obtained for sequencing analysis. Clinical data were obtained from the trial through a data transfer agreement. This study was approved by the UAB Institutional Review Board (Protocol No. X120917005) and conducted in accordance with ethical guidelines including the Good Clinical Practice Guidelines, the United States Code of Federal Regulations, and the Common Rule, which was informed by the Belmont Report and the Declaration of Helsinki. Progression-free survival (PFS) was the primary endpoint and was defined as the time from treatment initiation to disease progression or death; patients were censored at the date of the last disease assessment. Secondary endpoints include best response and overall survival (OS). OS was calculated from the time of initiation of therapy to death or last follow-up. Best response was evaluated using RECIST version 1.0. (RECIST v1.0). Patients that derived clinical benefit (CB) were defined as those that achieved complete remission (CR), partial response (PR), or stable disease (SD) for at least 6 months prior to progression or death. Patients with no clinical benefit (NCB) were those with progressive disease (PD) or SD lasting less than 6 months.

As a validation cohort, we analyzed a publicly available dataset, CheckMate 025, which was a phase III study of patients randomized to nivolumab or everolimus therapy (CM-025; NCT01668784; ref. 22). Patients enrolled in the CheckMate 025 dataset had mRCC with clear cell histology and had progressed on at least one prior systemic antiangiogenic therapy. This phase III trial demonstrated an improved overall survival in patients treated with nivolumab compared with everolimus. Of 397 patients treated with everolimus, 130 had available bulk RNA-seq data and were included in this analysis (23, 24). Patients with CR, PR with tumor shrinkage, or SD with PFS of at least 6 months were classified as having clinical benefit. All other patients were classified as having nonclinical benefit.

For all the patients included in the discovery cohort, the cases were reviewed and one representative slide was selected for hematoxylin and eosin (H&E) staining. Six of 107 samples were not processed due to the lack of an H&E slide or demarcated tumor regions. For the remaining 101 samples, a genitourinary pathologist identified by HCRN reviewed all the slides and selected tumor-enriched sections from representative FFPE tissue that was adequate for RNA extraction. tRNA was then extracted from two to 14 sections of 4-μm-thick FFPE sections, depending on the surface area of the tumor, using a High Pure FFPE RNA Isolation Kit (Roche). Concentrations of extracted RNA were determined using a DeNovix DS-11 Spectrophotometer, and samples of less than 20 ng/L were further concentrated using the RNA Clean & Concentrator-5 Kit (Zymo Research Corp.). Among the 101 samples, an additional eight samples of less than 2 ng/L or yielding negative A_260/280 ratios were considered inadequate and were excluded from the NanoString analysis.

The nCounter FLEX Analysis System (NanoString Technologies) was used for the discovery cohort. Gene expression analysis was performed for each sample using the PanCancer Pathways Panel. This panel comprises 790 genes, among which: (i) 720 genes are involved in essential oncogenic pathways and driver genes of clinical significance, (ii) 30 genes are known to be involved in the mTOR pathway, and (iii) 40 housekeeping genes exhibiting minimal variation across RCC specimens. Each reaction consisted of 100 ng of tRNA, reporter and capture probes, and six positive control and eight negative control probes. Analysis and normalization of the raw NanoString data was performed using nSolver Analysis Software 3.0 (NanoString Technologies). Advanced analysis of the raw data files identified 11 of 40 housekeeping genes that could be discarded by using the geNorm program. Raw expression data were normalized on the basis of the internal levels of the remaining 29 housekeeping genes. A background detection level was calculated using the geometric mean of the eight negative control probes in every reaction plus two SDs. In this case, the background level for gene detection was 14 counts. Among the 750 nonhousekeeping genes, we found that 233 (31%) non-mTOR pathway genes exhibited expression values below the limits of detection, and thus, they were filtered out and excluded, leaving a total of 517 genes for downstream analysis. Moreover, one sample yielded a low binding density, which was flagged for low counts and was thus removed from the analysis. In addition, PFS information could not be obtained for 10 patients; and thus, these were also excluded from our analysis, resulting in a total of 82 patients.

GSEA (http://software.broadinstitute.org/gsea/index.jsp) analysis was performed to test whether any biologically relevant gene sets were differentially expressed between patients with clinical benefit and those without clinical benefit. Hallmark gene sets (N = 50) were selected on the basis of previously published methods (25). A FDR q value of 0.25 was used as a significance threshold for all analyses.

Statistical analyses were conducted using R software. Analysis was performed to first identify predictive, or otherwise prognostic, gene signature. A biomarker is predictive if the treatment outcome of one treatment group compared with another varies for biomarker-positive patients compared with biomarker-negative patients. In contrast, a biomarker was considered prognostic if the P value of the biomarker is significantly associated with clinical outcomes irrespective of treatment groups (26). We used the modified covariates approach (27, 28) as the main method of identifying predictive features. This approach is used in the context of high-dimensional data (as is our case with the >500 list of genes) and implemented in R biospear package (25). The modified covariates approach fits a model containing modified covariates defined as the product of each predictor (X_i) and treatment (T) represented as ±0.5 for the study control and experimental arm. As a sensitivity analysis, the group LASSO method also implemented in the R biospear package (28) was also fitted; results are shown in the Supplementary section. The group LASSO method imposes the hierarchy constraint (i.e., it included treatment, biomarker, and treatment by biomarker interaction terms in the model). The modified covariates method was prioritized because it estimates less coefficients than the group LASSO method. The gene scores were log-transformed prior to analyses.

In the phase II portion of this study, patients were randomized to either of everolimus + BNC105P or everolimus only; therefore, patients in both arms are expected to be balanced with respect to observed and unobserved characteristics (29). However, among the entire cohort, 15 patients were derived from phase I of the study and included in the analysis. Given that this subset may introduce imbalance between patients in the experimental and control arms, a propensity score covariate adjustment was used to control for the potential imbalance. The propensity score was obtained by fitting a logistic model predicting experimental arm randomization conditional on two factors: (i) number of prior tyrosine kinase inhibitor (TKI) therapies received, and (ii) log baseline LDH. The predicted probability was included as a covariate, alongside the RNA expression values in the predictive model. The modified covariate model was fitted for: (i) PFS using the aforementioned package, and (ii) CB by first calculating the modified covariates and then fitting an adaptive LASSO-penalized logistic regression model. In the absence of a predictive signature, iterative sure independent screening (ISIS; refs. 30, 31) was used to identify a prognostic signature. Model discrimination performance was assessed using the area under the ROC curve (AUC). The optimal cutoff that maximizes sensitivity and specificity was identified. In the validation cohort (CheckMate 025), signature genes identified in the training model were used to predict clinical benefit and the performance of the model was reassessed using AUC.

All codes for data cleaning and analysis associated with the current submission are available at https://codeocean.com/capsule/3160935.

In the discovery cohort, as there was no difference in clinical outcomes between both arms in the phase I/II trial (12) (everolimus + BNC105P vs. everolimus only), and given that all samples sequenced by Nanostring were treatment-naïve at the time of sequencing, data were combined for both arms for a total of 82 patients (Supplementary Table S1.0; Supplementary Fig. S1). Among the entire cohort, 35 (43%) were treated with everolimus alone and 47 (57%) received a combination of everolimus + BNC105P. The tissue specimens were derived from primary nephrectomies (82/82, 100%). The median age at trial registration was 62 years (range, 45–84 years; Table 1). The majority of the patients were male (62/82, 76%; Table 1). Sixty-nine (84%) patients had received one prior TKI therapy while the remaining received two lines of TKI. The overall CB rate was 45.1% (37/82), and median PFS for the entire cohort (95% CI) was 4.9 (3.4–6.3) months. Baseline characteristics and clinical outcomes were similar in the two groups of patients (Table 1) and were also similar to the entire set of patients enrolled in the trial (12).

In the Checkmate 025 dataset, 130 patients were treated with everolimus and had available bulk RNA-seq data. The median age at trial enrollment was 63 years (range, 31–86; Supplementary Table S1.1), and 71% (92/130) of patients were males. Among the 130 patients, 43 (33%) achieved CB and the median PFS was 3.7 (95% CI, 3.4–5.7) months.

Within the discovery cohort, we studied the association between the 517 genes and clinical benefit in each of the two treatment arms. None of the genes was predictive for clinical benefit. However, a signature of four genes (ASXL1, DUSP6, ERCC2, and HSPA6) correlated with clinical benefit in both of the treatment arms (Table 2). Among 37 patients with high expression of this four-gene signature, 30 (81.1%) displayed clinical benefit, while among 45 patients with low expression, seven patients (15.6%) displayed clinical benefit. A linear predictor of clinical benefit was calculated for every patient using the sum of the products of gene expression levels and coefficient estimates. The threshold point score that maximizes the AUC was −0.06. Thirty of 37 (81%) patients with a point score >−0.06 derived clinical benefit. Notably, the estimated area under the ROC curve (AUC) (95% CI), a measure of discrimination, was 86.9% (79.2–94.7; Fig. 1A). Other clinical characteristics including age, gender, prior number of TKI lines, baseline calcium, hemoglobin, platelet count, absolute neutrophil count (ANC), and Karnofsky performance status (KPS) were not associated with clinical benefit (Supplementary Table S1.0).

We then performed pathway enrichment analysis of differentially expressed gene sets in patients with CB versus NCB. Of 50 “hallmark” gene sets representing major biological processes, none including the mTOR pathway was significantly associated with clinical benefit (q < 0.1).

In the CheckMate 025 cohort, the clinical benefit rate was 33% (43/130) overall. A corresponding threshold point score was determined at 31.8 with an estimated AUC of 60.2% (95% CI, 49.7–70.7; Fig. 1B) using the same four-gene expression signature (Supplementary Table S1.3). Twenty-two of 49 (44.9%) patients with a point score >31.8 achieved clinical benefit. Among 81 patients with threshold point score ≤31.8, 21 (26%) achieved clinical benefit (Table 3). A continuous linear predictor with all four genes (ASXL1, DUSP6, ERCC2, HSPA6) trended toward clinical benefit; however, it was not statistically significant (P = 0.06; Fig. 1C).

For the discovery cohort, we analyzed the association of all 517 genes with PFS in both treatment arms (Supplementary Table S1.2). A signature composed of 18 genes that correlated with PFS in the treatment arms (Table 4) was identified. A prognostic index (linear predictor) for PFS was calculated for each patient that represents the sum of the product of the 18 modified covariate coefficients (see Materials and Methods for details, Table 4). Patients were divided into two signature score categories (high and low) based on a threshold of −1.2 and 1.2 in the everolimus + BNC105P and everolimus only arms, respectively. Figure 2 shows Kaplan–Meier estimates of PFS by experimental arm between the high and low signature score categories. Within the low signature category (43 patients, 52.4%), patients treated with everolimus + BNC105P had a significantly longer median PFS compared with patients treated with everolimus alone (10.4 vs. 6.9 months; HR, 0.49; 95% CI, 0.24–1.002; P = 0.047; Fig. 2). Within the high signature category, there was no significant difference in PFS between both treatment arms (2.3 vs. 3.1 months; HR, 0.91; 95% CI, 0.50–1.65; P = 0.76; Fig. 2). We also note higher median PFS for both arms in the low signature category (6.9 and 10.4 months) compared with the high signature category (3.1 and 2.3 months), indicating that the signature also has a prognostic value.

The mTOR and angiogenesis pathways are activated in mRCC and have long been recognized as promising targets. Prior biomarker studies correlating rapalogs with clinical outcomes in mRCC have largely focused on mutational data. Unfortunately, such reports have led to conflicting data regarding the association between tumor somatic mutations and response to rapalog therapy in mRCC (6, 32–34). A recent report showed that loss of PTEN protein expression by IHC was strongly associated with improved PFS in patients treated with everolimus compared with those with intact PTEN IHC staining (33). Notably, PTEN loss by IHC did not correlate with the presence of mutations in the PI3K pathway and was much higher than the mutation frequency of PTEN in RCC in past series. Several cancer subtypes have utilized gene expression data to correlate clinical outcomes in patients with cancer treated with rapalogs (35, 36).

In our retrospective analysis of prospective trial data, a four-gene signature was determined to have prognostic value for using everolimus. A greater score on this four-gene signature was associated with a higher likelihood of achieving clinical benefit, regardless of receipt of everolimus alone or with BNC105P in the discovery dataset. This four-gene signature was tested in an independent publicly available dataset of everolimus-treated patients with mRCC and showed a suboptimal AUC. This will require further validation in other datasets before a definitive conclusion can be made. The genes identified are involved in DNA repair, chromatin remodeling, and response to cellular/metabolic stress, tying in tumor responses to mTOR pathway blockade with these pathways, as previously reported. For example, prior work has shown an association between decreased levels of HSP70 and rapamycin treatment (37, 38). A previous study also showed that activation of the PI3K/mTOR pathway leads to degradation of DUSP6 (39). As such, the positive association between everolimus treatment in our study and DUSP6 expression is in line with previous observations. Another study showed that ASXL1 mutants cooperated with BAP1 and led to activation of the Akt/mTOR pathway but the response to everolimus has not been assessed (40).

Overall, this represents the first transcriptomic signature that correlates with clinical benefit in patients with mRCC treated with everolimus. As clinical benefit rates were impressive in patients that exceeded the threshold point score, we believe that this signature, if further validated, may better select for patients likely to derive benefit from everolimus. Counterintuitively, transcripts representing components of the mTOR signaling pathway were not significantly different between responders and nonresponders. This lack of association with mTOR signaling pathway in patients treated with rapalogs has also been reported in other cancer subtypes (41). Overall, this suggests that protein level rather than expression changes, such as with PTEN, may better associate the mTOR pathway with rapalog response (33).

Herein, we also identified a gene expression pattern that correlated with PFS in patients treated with the combination of an mTORC1 inhibitor (everolimus) and a VDA (BNC105P) compared with everolimus alone. We show that an 18-gene signature may have predictive value by which patients with a low signature score are likely to achieve longer PFS in the combination group versus everolimus alone. Moreover, the angiogenesis gene expression program was not associated with better outcomes in patients treated with BNC105P. This is in contrast to previously published transcriptomic data showing that angiogenesis-associated genes can predict response to VEGF TKIs such as sunitinib (42, 43). Our results raise important questions regarding patient therapeutic stratification in mRCC. Patients with a low gene signature score derived a median of 3.5 months of added benefit when treated with a combination of everolimus + BNC105P compared with everolimus alone. As such, future studies are required to validate this gene signature and specifically focus on this subset of patients treated with a combination of everolimus and a VDA.

Several limitations of the reported study should be noted. First, the number of patients is small and not powered to ascertain gene classifiers with robust statistical power. Second, although all patients had received prior VEGF inhibitors, there was heterogeneity in the number of prior lines of therapy rendered. Third, the predominance of primary tumors in this cohort may have limited ability to detect differential gene expression profiles enriched in metastatic specimens, which may correlate better with clinical outcomes. Fourth, inherent limitations of the CheckMate 025 RNA-seq have previously been reported such as batch effect adjustment and known transcriptomic limitations from FFPE tissue (24). Finally, both trials were conducted in an era before the approval of checkpoint inhibitors. Hence, the relevance of these signatures in patients with mRCC receiving an mTOR inhibitor following prior immune checkpoint inhibitors requires evaluation.

In this work, we propose a four-gene signature that is associated with clinical benefit in patients treated with everolimus. If this signature is further validated following immune checkpoint and VEGF inhibitors, it could be used to aid in the selection of patients with progressive disease following prior therapies for mTOR inhibitors. We also discovered an 18-gene signature that is associated with longer PFS in patients treated with combination everolimus and a VDA, BNC105P, compared with everolimus alone. Validation of this signature may allow the rational development of VDAs in combination with mTOR inhibitors.

E.S. Yang reports grants from Novartis and Puma during the conduct of the study and personal fees from Strata Oncology, Bayer, and AstraZeneca outside the submitted work. D.L. Della Manna reports grants from Novartis Corporation during the conduct of the study. D.A. Braun reports personal fees from LM Education and Exchange Services, Schlesinger Associates, Adnovate, LM Education and Exchange Services, Insight Strategy, Imprint Science, Medscape, and Cancer Expert Now outside the submitted work as well as nonfinancial support from Bristol Myers Squibb. S.K. Pal reports personal fees from Pfizer, Novartis, Aveo, Genentech, Exelixis, Bristol Myers Squibb, Astellas Pharma, Eisai, Myriad, and Ipsen outside the submitted work. G.P. Sonpavde reports grants from Novartis during the conduct of the study; grants and personal fees from Sanofi, Immunomedics/Gilead, AstraZeneca, and QED; personal fees from BMS, Genentech, EMD Serono, Merck, Seattle Genetics, Exelixis, Janssen, Bicycle Therapeutics, Pfizer, Scholar Rock, G1 Therapeutics, Debiopharm, Mereo, and Elsevier outside the submitted work. No disclosures were reported by the other authors.

E.S. Yang: Conceptualization, resources, data curation, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing. A.H. Nassar: Data curation, formal analysis, investigation, writing–original draft, writing–review and editing. E. Adib: Software, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. O.A. Jegede: Data curation, software, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. S. Abou Alaiwi: Data curation, writing–original draft, writing–review and editing. D.L. Della Manna: Data curation, software, formal analysis, investigation, writing–original draft, writing–review and editing. D.A. Braun: Data curation, writing–original draft, writing–review and editing. M. Zarei: Data curation, writing–original draft, writing–review and editing. H. Du: Data curation, formal analysis, writing–original draft, writing–review and editing. S.K. Pal: Data curation, writing–original draft, writing–review and editing. G. Naik: Data curation, writing–original draft, writing–review and editing. G.P. Sonpavde: Conceptualization, resources, data curation, formal analysis, funding acquisition, writing–original draft, project administration, writing–review and editing.

This research was supported, in part, by Novartis. Bionomics Limited supported the phase I/II everolimus ± BNC105P clinical trial (NCT01034631).

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.