Impact of renal cell carcinoma molecular subtypes on immunotherapy and targeted therapy outcomes

The introduction of immune checkpoint inhibitors (ICIs) has revolutionized the therapeutic management of metastatic renal cell carcinoma (RCC)¹ with FDA-approved regimens including ICI combinations or ICI with an anti-angiogenic tyrosine kinase inhibitor (VEGF-TKI). A critical challenge in the RCC field is predicting responses to these varied treatments. Previously, the molecular profiling of 823 tumors in the IMmotion151 (IM151) trial identified seven transcriptomically-defined subgroups, influencing treatment response²: (1) angiogenic/stromal, (2) angiogenic, (3) complement/oxidation, (4) T-effector/proliferative, (5) proliferative, (6) stromal/proliferative, (7) snoRNA. Clusters 4, 5, and 7 showed significantly better outcomes with atezolizumab+bevacizumab compared to sunitinib, unlike clusters 1 and 2. Clusters 3 and 6 exhibited non-significant trends towards improved outcomes with atezolizumab+bevacizumab. This work has already come to the clinic, motivating clinical trials including the phase 2 OPTIC RCC trial³, a pioneering biomarker-driven study that assigns patients to either cabozantinib+nivolumab or ipilimumab+nivolumab based on the specific molecular subtype. Given the rapid translation to the clinic, it is crucial to validate these molecular subtypes and examine their predictive capabilities in larger clinical trial datasets.

To evaluate the biology and predictive nature of these molecular clusters, we analyzed the transcriptomic and clinical landscapes of patients included in the phase III JAVELIN Renal 101 (JR101) trial⁴, which evaluated the (now FDA-approved) regimen of avelumab+axitinib (ICI+VEGF-TKI) against sunitinib. Our study investigated whether these molecular subsets represent truly distinct biological states, and whether they are predictive of differential clinical outcomes to the two treatment regimens (Supp.Fig.1A).

We first trained a machine learning model (Supp.Methods) on tumor transcriptomic data from the IM151 trial then applied it to classify 734 tumors (472 primary, 261 metastatic, 1 unspecified) from the JR101 trial into one of seven molecular subtypes.

Baseline demographics were comparable between IM151 and JR101 (Supp.Table 1A). The distribution of patients across the molecular clusters was similar between the two trials (Chi-squared test, excluding cluster 7, p=0.28; Supp.Fig.1B, Supp.Table 1B). Similar to prior reports⁵., we observed some differences in the distribution of molecular clusters depending on the origin of the tumor specimen that was used for sequencing (primary vs metastatic), highlighting the issue of tumor heterogeneity in molecular classifications (Supp.Table 1C).

In both IM151 and JR101 trials, cluster 6 (stromal/proliferative) had the highest rate of sarcomatoid component (36.79% in IM151 and 26.83% in JR101), while the angiogenic clusters 1 and 2 had the lowest (Supp.Fig.1B–C, Supp.Table 1B). In concordance with IM151 dataset, sarcomatoid differentiation in JR101 closely correlated with the molecular clusters’ transcriptional states, with no significant difference in per-cluster proportions of sarcomatoid differentiations between the two trial datasets (Chi-squared test, p=0.52). We further explored the association between the molecular clusters and IMDC prognostic risk groups⁶. Confirming the IM151 findings, the IMDC favorable risk group predominantly had angiogenic tumors (clusters 1 and 2: 20.5% and 39.2% in IM151, 17.7% and 34.2% in JR101; Supp.Fig.1D–E, Supp.Table 1B). The poor risk group showed a prevalence of immune/proliferative signature tumors (clusters 4, 5, and 6). Patient distribution across molecular clusters in favorable, intermediate, and poor risk groups showed high concordance in both IM151 and JR101 (Chi-squared tests, p=0.721, p=0.59, and p=0.96, respectively).

We next evaluated whether the gene expression patterns of the JR101 clusters supported the distinct biological states proposed by Motzer et al through DGE and GSEA (Supp.Fig.1F; Supp.Table 1E–R). Tumors in clusters 1 (angiogenic/stromal) and 2 (angiogenic) showed high levels of vascular pathway genes (e.g. VEGFA), and low cell-cycle gene expression. Cluster 1 was rich in fibroblast/collagen genes (e.g., FAP, COL1A1), while cluster 2 had catabolism genes (e.g. CPT2, PPARA). Cluster 3 (complement/oxidation) had genes linked to the complement cascade (e.g, F2, C1S) and CYP450 family (e.g. CYP4F3). Clusters 4 (T-effector/proliferative), 5 (proliferative), and 6 (stromal/proliferative) expressed high levels of cell cycle (e.g. CDK2, CCNE1) and anabolism genes (e.g. FAS, G6PD). Cluster 4 also showed T cell activation genes (e.g., CD8A, IFNG).Clusters 5 and 6 were rich in myeloid inflammation genes (e.g. CXCL1, IL6). Cluster 6 additionally exhibited fibroblast and collagen-associated gene enrichment. Overall, these findings support the distinct biological states of each molecular subtype.

We observed a distinct somatic mutational landscape for each molecular cluster (Supp.Fig.1G–H). Angiogenic cluster 2 showed a significant depletion in TP53 (q=0.029) and PTEN (q=0.015) mutations. On the other hand, proliferative clusters 4 and 5 exhibited a marked depletion of respectively PBRM1 (q=0.0003) and VHL (q=3.99e-8) mutations. Cluster 3 displayed the highest enrichment of VHL (q=0.0039). Overall, the pattern of distinct somatic alteration and transcriptional profiles for each molecular subtype that was identified in IM151 was again observed in our analysis of JR101, further supporting the distinct biology of different molecular clusters.

As an orthogonal validation, cluster 4 also had the highest CD8 T cell infiltration by immunohistochemistry (ANOVA p.value: 1.722147e-51; Supp.Fig.1I).

In the original IM151 analysis, molecular subtypes were associated with differential response and survival to ICI+VEGF therapy (atezolizumab+bevacizumab) versus VEGF-TKI (sunitinib). Specifically, patients with tumors classified within clusters 4 and 5 demonstrated improved responses to atezolizumab+bevacizumab compared to sunitinib (Supp.Fig.1J; Supp.Table 1D). Conversely, patients characterized by clusters 1 and 2 tumors did not appear to have increased benefit from the combination atezolizumab+bevacizumab regimen compared to sunitinib (Supp.Fig.1J; Supp.Table 1D). In fact, for cluster 2, the proportion of patients with an objective response was significantly higher with sunitinib treatment (46.6%) than with atezolizumab+bevacizumab (33%; Chi-squared, p=0.027). However, in JR101, the response rates to the ICI+VEGF-TKI regimen (avelumab+axitinib) were higher than response rates to sunitinib across all molecular subtypes, even within angiogenic clusters 1 and 2. (Supp.Fig.1K; Supp.Table 1D).

In the IM151 dataset, patients specifically in immune/proliferative clusters 4 and 5 (both individual clusters and the combination) demonstrated longer median PFS with atezolizumab+bevacizumab compared to sunitinib (hazard ratio (HR): Cluster 4: 0.53 (95% CI 0.33–0.84), Cluster 5: 0.47 (95% CI 0.27–0.82), Clusters 4+5: 0.52 (95% CI 0.37–0.74); Supp.Fig.1L). However, in the JR101 dataset, patients treated with avelumab+axitinib tended to have improved PFS versus those treated with sunitinib, irrespective of molecular subtype, though the magnitude (and significance) of improvement varied between clusters. Specifically, patients within the angiogenic/stromal (Cluster 1, HR: 0.55(0.31–0.97)), complement/oxidation (Cluster 3, HR: 0.65(0.43–0.97)), and T-effector/proliferative (Cluster 4, HR: 0.54(0.32–0.91)) demonstrated significantly longer PFS under the avelumab+axitinib regimen compared to sunitinib (Supp.Fig.1M). Of note, we observed similar results when we limited our analyses to include primary tumors only (data not shown).

Overall, the molecular clusters identified in IM151 have concordant clinical and molecular profiles in the JR101 dataset, providing additional insightful for RCC biology and prognosis beyond traditional clinical risk stratification groups. The molecular clusters accurately capture distinct RCC transcriptomic programs, shedding light on the diverse biology underlying differing RCC clinical behaviors. Although our results confirm that the molecular clusters represent distinct biological states, in our study avelumab+axitinib tended to improve clinical outcomes compared to sunitinib across all molecular subtypes. These data suggest that these clusters may not yet be directly applicable as predictive biomarkers for ICI+VEGF-TKI vs VEGF-TKI alone more generally. The discrepancy between IM151 and JR101 trial outcomes may be attributable to several factors. Although both atezolizumab+bevacizumab and avelumab+axitinib are ICI+VEGF combinations, their mechanisms of action are distinct: bevacizumab is a monoclonal antibody that binds to VEGF-A and inhibits its activity, whereas axitinib is a tyrosine kinase inhibitor that binds to and inhibits the VEGF receptors (VEGFR). Clinically, VEGFR-targeting TKIs are now the standard anti-angiogenic therapies used in RCC. It is therefore conceivable that at least part of the superiority of avelumab+axitinib across all molecular subtypes (including angiogenic clusters) could be due to increased activity of axitinib as compared to bevacizumab. Another important possibility is that biomarkers might be specific to individual drugs rather than broadly applicable across diverse classes of drugs (or event different agents within a therapeutic class). It therefore remains on ongoing challenge to effectively develop biomarker-based treatments in RCC.

In this study, we seek to validate the IM151 molecular clusters utilizing an FDA-approved phase 3 trial dataset. Nonetheless, there are inherent limitations that may have influenced our findings. Our predictive model, while highly accurate, does have limitations, particularly in detecting small nucleolar RNA (snoRNA), specifically in the context of two non-entirely overlapping molecular datasets. The multi-center nature of these studies limited our access to crucial information such as sample acquisition details and tumor quality factors. Looking ahead, biomarker-driven therapy will likely require a comprehensive approach, integrating gene signatures, somatic alterations, tumor immunophenotype, and even spatial architecture. To build a robust predictive model, both host and tumor variables may need to be considered, moving beyond a solely tumor-centric approach.

In conclusion, our study sheds light on the complexities of predictive biomarkers in RCC. The IM151 molecular clusters reflect the underlying biology of RCC and are associated with different outcomes in IM151. In the JR01 cohort, the benefit of FDA-approved avelumab+axitinib is maintained over sunitinib across clusters. Further research and validation efforts are essential to refine and extend our understanding of RCC classification and personalized therapeutic strategies.

Supplementary Material

Supplementary Table 1

Table S1. Clinical and transcriptional profiles within each molecular subtype.