irRECIST for the Evaluation of Candidate Biomarkers of Response to Nivolumab in Metastatic Clear Cell Renal Cell Carcinoma: Analysis of a Phase II Prospective Clinical Trial

Cancer immunotherapy targeting programmed cell death protein 1 (PD-1) signaling improves overall survival (OS) in several tumor types with manageable toxicity and durable responses in a subset of patients (1). In patients with previously treated metastatic clear cell renal cell carcinoma (mccRCC), nivolumab, a fully human monoclonal antibody against PD-1, demonstrated superior OS and fewer serious adverse events than everolimus in the CheckMate-025 trial, leading to its FDA approval (2). Although nivolumab's favorable therapeutic index makes it an appealing consideration for earlier disease settings, the lack of predictive biomarkers for selecting patients likely to achieve durable benefit limits the ability to establish the value of anti–PD-1s' monotherapy in treatment-naïve mccRCC patients.

World Health Organization (WHO) tumor response criteria and the most recent RECISTv1.1 are surrogates of survival routinely used by oncologists for clinical decision-making (3, 4). Compared with targeted agents and conventional chemotherapy, immune-checkpoint inhibitors can display an atypical pattern of response, where new lesions develop or established lesions grow before an objective response or stable disease is observed (5–10). Immune-related response criteria (irRC, adapted from WHO criteria) and subsequently irRECIST, immune-based therapeutics RECIST (iRECIST), and immune-modified RECIST (imRECIST; all adapted from RECISTv1.1) were therefore developed to prevent misclassification of atypical responders as early progressors by the conventional WHO and RECISTv1.1 criteria (11–14). Recent analyses demonstrated that compared with RECISTv1.1, immune-related response criteria may more accurately predict long-term survival outcomes in patients with melanoma and lung cancer treated by PD-1 blockade (15, 16). Although it is increasingly accepted that response per immune-related criteria can more accurately assess benefit from immunotherapy, efforts to identify predictive biomarkers for anti–PD-1 agents have exclusively utilized endpoints based on RECISTv1.1, potentially impairing biomarker discovery.

The present article is based on the analysis of the CheckMate-010 trial, a dose-finding study where patients with mccRCC were randomly assigned to three different doses of nivolumab. It should be noted that in the initial publication of the trial, although irRECIST-based endpoints were reported, they were solely used as exploratory efficacy endpoints to demonstrate that, similar to RECISTv1.1-based endpoints, nivolumab efficacy was dose-independent (17). In this study, we first evaluated whether atypical responses to nivolumab, defined by irRECIST, affected clinical outcome of patients with mccRCC enrolled in the trial. We further tested the hypothesis that candidate biomarkers for nivolumab response show improved association with clinical endpoints capturing atypical responders (i.e., irRECIST) compared with standard clinical endpoints (i.e., RECISTv1.1).

We studied mccRCC patients from the CheckMate-010 trial (BMS-936558, ClinicalTrials.gov_NCT01354431; ref. 17). This trial is a multicenter phase II dose-finding study of nivolumab in patients with mccRCC who received previous regimen of agent targeting vascular endothelial growth factor pathway. Formalin-fixed and paraffin-embedded (FFPE) tumor sections were collected by the sponsor at the time of the trial. Institutional Review Board approval and individual written-informed consents were obtained before tissue acquisition, tissue staining, and analysis of clinical information in accord with an assurance filed with and approved by the Department of Health and Human Services.

Progression-free survival (PFS) was defined as the time from randomization to documented disease progression per RECISTv1.1 or death. Patients alive and progression-free were censored at the date of last tumor assessment. Objective response rate (ORR) was defined as the proportion of randomized subjects whose best tumor response is either partial response or complete response per RECISTv1.1 (3). The definition of irRECIST has been previously published, and these criteria have been utilized in previous studies (12, 16–19). Briefly, irRECIST include the following major modifications from RECISTv1.1: (1) requirement to confirm progression ≥ 4 weeks after initial radiological progression and (2) not scoring new small nontarget lesions as progression but using the net tumor burden to define progression.

irRECIST definitions for PFS (irPFS) and ORR (irORR) are similar to PFS and ORR and were derived from modified RECIST (12, 17).

Programmed cell death 1 ligand 1 (PD-L1) and programmed cell death 1 ligand 2 (PD-L2) IHC were performed using a verified assay developed by Dako (clone 28-8) and an in-house validated assay (clone D7U8C, Supplementary Materials and Methods and Supplementary Fig. S1), respectively. Membranous PD-L1 expression and membranous and/or cytoplasmic PD-L2 expression were independently scored as a percentage of positive tumor cells (TC) by S. Signoretti and J.-C. Pignon that were blinded to patient outcomes. Interscorer discrepancies were resolved by consensus review. Overall tumor PD-L1 expression was quantified using ImageScope Membranous v9-algorithm (Leica).

Performance of previously validated antibodies for cluster of differentiation 8 (CD8), PD-1, T-cell immunoglobulin and mucin-domain containing-3 (TIM-3) and lymphocyte activation protein-3 (LAG-3) was assessed by staining nonneoplastic tonsil tissue and obtaining a staining pattern overlapping with published data (20–24). CD8, PD-1, TIM-3, and LAG-3 multiplex immunofluorescence (IF) was performed by serial staining using the Opal tyramide signal system from Perkin Elmer as described in the Supplementary Materials and Methods and Supplementary Table S1.

Multiplex IF-stained tissue sections were visualized with the Mantra Quantitative Pathology Workstation (Perkin Elmer). Multispectral images were acquired with the Mantra microscope using a 20x objective. Inform 2.2 software was then used in order to deconvolute the multispectral images, and to segment and phenotype cells (Supplementary Fig. S2). Briefly, for each fluorochrome, a spectral library was created on single stained tissue sections; this library was used to deconvolute the multispectral images in spectral images corresponding to a specific fluorochrome (Supplementary Fig. S2A). Cell segmentation based on nuclear DAPI signal, cell size, and CD8, PD-1, TIM3, and LAG-3 membranous staining signal was then applied (Supplementary Fig. S2B). The phenotyping step, based on machine-learning recognition algorithm, was performed by developing three algorithms recognizing: (1) cells monostained for CD8 or costained for CD8 and PD-1, (2) cells monostained for CD8 or costained for CD8 and TIM-3, and (3) cells monostained for CD8 or costained for CD8 and LAG-3 (Supplementary Fig. S2C). For each case, positive cells were manually identified until the learning recognition algorithm was concordant with visual count (concordance >90%). Of note, a cell was called CD8⁺ only if it was recognized as CD8⁺ by all three different algorithms. For each segmented cell, we therefore obtained information about the presence or the absence of CD8, PD-1, TIM3, or LAG3 staining. Unique phenotyping was performed for each tumor to account for intersample variability of signal intensities as previously described (25). The percentage and the density of CD8⁺ cells expressing PD-1, TIM-3, LAG-3, either alone or in different combinations were calculated on a minimum of 5 different 0.36 mm² images acquired from tumor area containing the highest density of tumor-infiltrating CD8⁺ cells (26). Additional images were taken for tumors with low CD8⁺ cells to reach a minimum of 100 CD8⁺ cells counted. Tumor specimens with tumor area below 3.6 mm² were not scored to minimize tissue selection bias.

RNAs were extracted using the AllPrep DNA/RNA FFPE Mini Kit (Qiagen) from a 5 cm² tumor-enriched area macrodissected from 4-μm-thick FFPE tissue sections prepared form a single tumor block. The tumor-enriched area contained an estimated percentage of TCs that ranged from 30% to 70%. Transcriptome capture analysis was performed as described in the Supplementary Materials and Methods. The cytolytic activity (CYT) signature, the effector-T cell (T_eff) signature, and 18 IFNγ-responsive genes signature were defined based on previously published associations with the respective biology, and/or with clinical outcome (27–29). For each gene signature, an expression score for each patient sample was calculated as the geometric mean–normalized count of all component genes in the signature. The geometric mean of the signatures was then log₂ transformed to obtain the final signature expression score.

Prentice–Wilcoxon and McNemar tests were used to assess difference between correlated pairs of irPFS and PFS, and irORR and ORR endpoints, respectively. The χ² test of equal proportion was used to estimate the difference between the rate of responses using RECISTv1.1 or irRECIST.

Association of TC PD-L1 and TC PD-L2 positivity with Fuhrman nuclear grade (FNG) was assessed using the χ² test. Correlation of TC PD-L1 and TC PD-L2 positivity was estimated by calculating Spearman rank correlation coefficient. PD-L1 and PD-L2 expression scores were categorized with frequently used PD-L1 cutoff points in immuno-oncology trial and correlated with clinical endpoints. The Fisher exact test was used for binary endpoints of irORR and ORR and the log-rank test for irPFS and PFS. Superiority of immune-related endpoints (irPFS and irORR) over nonimmune-related endpoints (PFS and ORR) was assessed graphically and also numerically via the appropriate test statistic. For example, log-rank test statistic comparing irPFS by PD-L1/PD-L2 was compared with log-rank test statistic comparing PFS by PD-L1/PD-L2.

Multiplex IF assay was performed in a total of five batches—Kruskal–Wallis one-way ANOVA test was used to assess differences in median expression score across batches. Correlations of the different fractions or cell densities of CD8⁺ cells expressing PD-1, TIM-3, or LAG-3 either alone or in different combination were estimated by calculating Pearson rank correlation coefficient. Differences between the median percentages or densities of CD8⁺ cells expressing PD-1, TIM-3, or LAG-3 either alone or in different combination were estimated using the Sign test.

Expression scores of the proportion and density of CD8⁺ cells expressing PD-1, TIM-3, and LAG-3 either alone or in different combinations were correlated with PFS/irPFS and ORR/irORR. The observed expression scores were positively skewed and therefore a cube root transformation of expression scores was taken to reduce the influence of extreme biomarker expression scores on model parameter estimates. There were a total of 33 different combinations of proportion and density of CD8⁺ PD-1, TIM-3, and/or LAG-3 expression scores. Transformed expression scores were correlated with PFS and irPFS using univariable Cox proportional hazards (PH) model. This approach to correlate continuous expression scores with clinical outcomes better utilizes the information contained in the continuous measurements and allows for assessment of a linear relationship between expression scores and clinical endpoints. In addition, it reduces the chances of false-positive results as a consequence of using various approaches to classify expression scores. The Holm–Bonferroni approach was also used to control the family-wise error rate (FWER; Type I error) at 5% for the hypotheses tests conducted (30). The Holm–Bonferroni approach is uniformly more powerful than the rather conservative classic Bonferroni approach. For interpretation and visualization purposes, each of the biomarkers that meet the FWER control screening threshold (i.e., declared statistically significant) specified above was also classified into binary variables by selecting an optimal threshold that maximizes specificity and sensitivity with respect to irORR (31). At optimized cutoffs, the Fisher exact test was used for binary endpoints of irORR and the log-rank test for irPFS.

The difference in gene signature expression score distribution between groups was evaluated using a Wilcoxon rank-sum test. Signature expression scores were correlated with irPFS and irORR using univariable Cox PH and binary logistic regression analysis, respectively.

One-sided P values ≤0.05 were considered statistically significant for all analysis.

From May 2011 to January 2012, 167 previously treated mccRCC patients were randomly assigned to nivolumab 0.3, 2, or 10 mg/kg intravenously once every 3 weeks as previously described (17). Median clinical follow-up of the present study was 38.4 months (min–max: 0.6–43.3 months). Baseline demographics, treatment arms, and clinical characteristics are presented in Table 1. Because no significant dose–response relationship of nivolumab with primary endpoint (PFS) and secondary endpoints (ORR, irPFS, and OS) were previously found, patient cohorts were pooled together in the present study (17). PD-L1 and PD-L2 expressions and positivity of the immune-checkpoint inhibitors PD-1, TIM-3, and LAG-3 on CD8⁺ tumor-infiltrating cells (TIC) were assessed in 140, 127, and 98 patients, respectively (Supplementary Fig. S3).

We compared responses evaluated using RECISTv1.1 with those evaluated using irRECIST in the whole cohort (n = 167). RECISTv1.1 were used to calculate PFS and ORR, and irRECIST were used to calculate irPFS and irORR. irORR (22.8%) was not significantly different from ORR (21%). However, the percentage of patients with immune-related progressive disease (irPD) (24.6%) was significantly lower than the percentage of patients with progressive disease (PD) (35.3%, P = 0.03), and the percentage of patients with immune-related stable disease (irSD) (50.3%) tended to be higher than the percentage of patients with stable disease (SD) (41.3%; Supplementary Table S2). Median irPFS (5.5 months) was significantly longer than median PFS (3.3 months; P < 0.001; Fig. 1A).

In the initial study of the Checkmate-010, TC PD-L1 expression was evaluated in 107 patients using a prototype IHC assay based on the 28-8 clone (17). Here, PD-L1 was independently assessed in larger number of patients (n = 140; Supplementary Fig. S3) using a verified assay developed by Dako using the 28-8 clone. Tissue analyzed included 95 (67.9%) primary tumors, 36 (25.7%) metastases, and 9 (6.4%) lesions from unknown sites. TC PD-L1 expression ≥1% was observed in 19 (13.6 %) patients (Table 2; Supplementary Fig. S4) and was associated with high FNG (P = 0.003; Supplementary Table S3).

TC PD-L1 expression was not significantly associated with improved PFS (P = 0.09; Table 2 and Fig. 1B). In contrast, median irPFS was significantly longer in the TC PD-L1 expression ≥1% group (15.4 months) compared with the TC PD-L1 <1% group (4.3 months; P = 0.04; Table 2 and Fig. 1C). Although ORR tended to be higher in the TC PD-L1 expression ≥1% group (36.8%) compared with the TC PD-L1 expression <1% group (18.8%), this difference did not reach statistical significance (P = 0.07). On the other hand, irORR was significantly higher in the TC PD-L1–positive ≥1% group (47.4%) compared with the TC PD-L1 <1% group (19.3%; P = 0.01; Table 2). Although the patients randomly assigned to nivolumab 0.3, 2, or 10 mg/kg achieved similar clinical efficacy, we also aimed to demonstrate that the association between TC PD-L1 expression and clinical outcome was independent of drug dosing. Indeed, no significant difference in TC PD-L1 positivity score among the different arms of nivolumab-treated patients was observed. Of note, baseline demographics and clinical characteristics were also similar between TC PD-L1 expression groups (Table 1). When higher cutoffs were used to define TC PD-L1 positivity (5% and 10%), the proportion of patients with positive tumors became progressively smaller (8% and 5%, respectively), preventing meaningful correlations with clinical endpoints.

PD-L1 expressed on tumor immune cells (IC) can also bind to PD-1 and inhibit the antitumor immune response (1, 32). Therefore, we further evaluated overall PD-L1 expression by measuring membranous PD-L1 staining on both TC and IC in 131 patients (Supplementary Fig. S3). Using an arbitrary cutoff of ≥1% PD-L1–positive cells, overall (TC and IC) PD-L1 expression was observed in 29% of patients. No significant association between overall PD-L1 expression and clinical endpoints was found at any tested cutoff (Supplementary Table S4).

To test if TC expression of the second PD-1 ligand, PD-L2, could improve the identification of tumors that respond to anti–PD-1 therapy, TC PD-L2 expression was assessed by IHC in 127 patients (Supplementary Figs. S3 and S4; ref. 33).TC PD-L2 expression ≥1% was observed in 46 (36.5%) patients (Table 2) and was not associated with FNG (Supplementary Table S3). No significant association between overall PD-L2 expression and clinical outcomes was found at 1% positivity cutoff (Table 2) or any other tested cutoff (Supplementary Table S5). TC PD-L1 expression did not correlate with TC PD-L2 expression (Supplementary Fig. S5), and only 8 of 126 (6.34%) patients had tumors expressing both PD-L1 (≥1%) and PD-L2 (≥1%). Therefore, we evaluated whether combined TC PD-L1 and/or TC PD-L2 expression could better identify patients that respond to nivolumab. When TC positivity was defined as expression of either or both PD-L1 (≥1% cutoff) and PD-L2 (≥1% cutoff), 55 of 126 patients (43.65%) were categorized as positive. No significant association between PD-L1/PD-L2 positivity and clinical endpoints was found (Supplementary Table S6).

Human tumors have been shown to contain severely exhausted T cells expressing multiple immune checkpoints, and it has been proposed that these cells mediate resistance to PD-1 blockade (34, 35). For this reason, we evaluated CD8⁺ TIC for the expression of the immune checkpoints PD-1, TIM-3, and LAG-3 in 98 patients (Supplementary Fig. S3). For each tumor, we measured both the absolute number (cell density per mm²) and the relative amount (percentage) of CD8⁺ TIC expressing the three immune checkpoints either alone or in various combinations by multiplex IF.

PD-1 was the immune checkpoint most frequently expressed on CD8⁺ TIC followed by TIM-3 and LAG-3 (Fig. 2A and C). In line with these findings, the median percentage (and median density) of CD8⁺ TIC expressing PD-1 alone was significantly higher than the median percentage (and median density) of CD8⁺ TIC expressing PD-1 and TIM-3 or PD-1 and LAG-3. Of note, TIM-3 and LAG-3 were detected more frequently on CD8⁺PD-1⁺ TIC compared with CD8⁺PD-1⁻ TIC (Fig. 2B and D).

We next investigated whether the expression of PD-1, TIM-3, and LAG-3 (in various combinations) on CD8⁺ TIC (Supplementary Table S7) was associated with outcome on nivolumab therapy. Biomarker analysis included both the mean percentage and mean density measurements as they were found to be weakly correlated with each other and therefore provide different information on the composition of the tumor microenvironment (Supplementary Fig. S6). To correlate these biomarkers with clinical endpoints, we utilized a stringent approach in which the candidate biomarkers were screened for the presence of a linear relationship between biomarker levels and PFS or irPFS. Chances of false-positive results were controlled using the Holm–Bonferroni criteria. Whereas no biomarker was found to be significantly correlated with PFS, three biomarkers (percentage of CD8⁺ TIC that are PD-1⁺TIM-3⁻LAG-3⁻; density of CD8⁺ TIC that are PD-1⁺TIM-3⁻LAG-3⁻; percentage of CD8⁺ TIC that are PD-1⁺) significantly correlated with irPFS (Supplementary Table S8). Using a cutoff that maximizes sensitivity for irORR, the percentage of CD8⁺ TIC that are PD-1⁺TIM-3⁻LAG-3⁻ (hereafter called CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC) exhibited the best sensitivity and specificity (Table 2 and Supplementary Table S9) and was further explored. At the optimized cutoff of 36%, high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC was observed in 74 of 98 (75.5 %) patients. The median irPFS was significantly longer in the group with high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (6.9 months) compared with the group with low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (2.6 months; P = 0.005; Table 2; Fig. 3A). irORR was significantly higher in the group with high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (28.8%) compared with the group with low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (0%; P = 0.001; Table 2). Of note, no significant difference in biomarker expression scores among the three-dose arm of nivolumab-treated patients was observed. Baseline demographics and clinical characteristics were also similar between groups of patients having a high or a low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (Table 1).

The high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC associated to nivolumab response might identify patients with an active antitumor immune response in their pretreatment tumors. To test this hypothesis, we performed RNA sequencing on a subset of tumors with available tissue (Supplementary Fig. S3) and interrogated three gene expression signatures associated with T-cell activation: the CYT signature, the T_eff signature, and the IFNγ–responsive gene signature (27–29). Thirty-nine of 47 patients with gene expression signature data had also IF data. In line with our hypothesis, high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (≥36%) was found to be associated with high expression levels of the CYT, T_eff, and IFNγ signatures (P = 0.004, 0.001, and 0.007, respectively; Fig. 3B). None of the three signatures were significantly associated with clinical outcomes.

We next tested whether the combination of TC PD-L1 expression with the percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC could improve the ability to identify patients that respond to nivolumab. Ninety-seven patients had data for both biomarkers. Among these, 11 patients had both high TC PD-L1 expression (≥1%) and high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (≥36%), and 63 patients had low TC PD-L1 expression (<1%) and high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC. All 23 patients with low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (<36%) were found to have low TC PD-L1 expression.

Both median irPFS and irORR were significantly different among three patient groups (P = 0.013 and P = 0.001, respectively; Fig. 3C; Table 3). Specifically, median irPFS was longest in the group with both high TC PD-L1 expression and high percentage of CD8⁺PD-1⁺TIM-3 LAG-3⁻ TIC (15.4 months), intermediate in the group with low TC PD-L1 expression and high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (5.5 months), and shortest in the group with both low TC PD-L1 expression and low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (4.1 months). Similarly, irORR was highest in the group with both high TC PD-L1 expression and high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (54.5%), intermediate in the group with low TC PD-L1 expression and high percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (24.2%), and lowest in the group with both low TC PD-L1 expression and low percentage of CD8⁺PD-1⁺TIM-3⁻LAG-3⁻ TIC (0%).

Our analyses of the CheckMate-010 cohort show that irRECIST enable the capture of patients with mccRCC who displayed atypical responses to nivolumab therapy, which had a significant impact on median PFS. Furthermore, we showed that candidate biomarkers for nivolumab response had an improved association with clinical endpoints defined by irRECIST relative to RECISTv1.1. By utilizing irRECIST, we developed a novel model combining TC PD-L1 expression and percentage of CD8⁺ TIC that express PD-1 (but not the other immune checkpoints TIM-3 or LAG-3). This proposed model allowed stratification of nivolumab-treated patients in three groups with significantly different irPFS and irORR.

Clinical experience increasingly indicates that traditional response criteria cannot adequately assess benefit from immunotherapy. In line with results obtained in melanoma and lung cancer, our study showed that within the CheckMate-010 trial, a subset of patients with mccRCC treated with nivolumab experienced objective response or stable disease by irRECIST, but PD by RECISTv1.1 (14–16). Reclassification of the CheckMate-010 trial patient outcomes by irRECIST resulted in significantly longer irPFS compared with PFS. These data have important implications for clinical management of patients as well as for correlative biomarker studies. Our results indicate that utilization of RECISTv1.1 underestimates the number of mccRCC patients that benefit from nivolumab therapy. Our findings also support the concept that misclassification of atypical responders by RECISTv1.1 has the potential to confound biomarker analyses for immune-checkpoint inhibitors and thus undermines the development of robust patient selection criteria for these agents. Indeed, in line with previously published data, we found that TC PD-L1 expression was not correlated with clinical outcome by RECISTv1.1 in the CheckMate-010 trial patients (17). However, TC PD-L1 expression was found to be significantly associated with both higher irORR and longer irPFS. These associations are notable given that tumor PD-L1 expression has been associated with poor prognosis in patients with RCC treated with either prior ineffective therapies or VEGFR blockers (21, 36–38). Of note, overall PD-L1 expression (including expression in both TC and IC) was not associated with clinical endpoints, suggesting that in ccRCC, PD-L1 expression on TC (but not IC) is a major driver of immune evasion that is reversed by PD-1 blockade. However, further studies are needed to specifically address the role of IC PD-L1 expression in predicting response to nivolumab. In spite of the association of TC PD-L1 expression with clinical outcomes, our data show that this marker alone fails to reliably identify all patients likely to benefit from therapy.

For this reason, we tested other candidate biomarkers to develop a multimarker model that demonstrates a greater predictive value than PD-L1 expression alone. Persistent exposure to antigen during chronic infection or cancer induces T-cell exhaustion, a cell state characterized by coexpression of multiple immune checkpoints, and that has been shown to limit cell proliferation and reduced cytotoxic activities (34). Rejuvenation of exhausted T cells can be achieved by PD-1 blockade, and combined inhibition of PD-1 with TIM-3 or LAG-3 induces synergistic reduction of tumor growth (35, 39, 40). Here, we found that both high percentages of CD8⁺ TIC expressing PD-1 and high percentages of CD8⁺ TIC expressing PD-1 but not TIM-3 and LAG-3 are positively associated with longer irPFS on nivolumab. The association with higher irORR was strongest for the subset of CD8⁺ TIC expressing PD-1, but not TIM-3 or LAG-3, suggesting that the higher predictive value of these cells might be related to their less exhausted phenotype and their ability to be more efficiently reactivated with PD-1 blockade. In line with these findings, we showed that a high percentage of CD8⁺ TIC expressing PD-1 but not TIM-3 and LAG-3 was associated with high levels of T-cell activation, as assessed by gene expression analysis. The rejuvenation of highly exhausted may require combination therapy including agents targeting PD-1, TIM-3, and LAG-3.

Remarkably, in our cohort, all patients with low percentage of CD8⁺ TIC expressing PD-1 (but not TIM-3 or LAG-3) were also negative for TC PD-L1 expression and did not respond to nivolumab. On the other hand, among patients with a high percentage of CD8⁺ TIC expressing PD-1 (but not TIM-3 or LAG-3), the subset of patients with positive TC PD-L1 expression displayed highest irORR and longest irPFS. Overall, our results suggest that (1) patients with high expression of the therapy target PD-1, on TIC, and of its ligand PD-L1, on TC, are the most likely to benefit from nivolumab; (2) patients with high expression of PD-1 on TIC in the absence of TC PD-L1 expression have an intermediate chance to respond to nivolumab; and (3) patients with neither high PD-1 expression on TIC nor TC PD-L1 expression are resistant to therapy. Validation of this model could help identify patients who would benefit sufficiently from nivolumab monotherapy in the treatment-naïve or adjuvant settings.

The analyses presented here have several potential limitations. First, in our cohort, patients were randomly assigned to receive three different doses of nivolumab. However, it was previously reported that all the tested clinical endpoints (ORR, PFS, and irPFS) were not significantly different in the three randomized arms (17). Moreover, we demonstrated that our candidate biomarkers' expression scores were similar between the three nivolumab treatment groups. Taken together, these data strongly suggest that the randomized drug dosing design of the trial has very limited impact on our results.

Another limitation of the study is the analysis of a single tumor sample for each patient. As intratumor heterogeneity is well-recognized in RCC, biomarker measurement in a single tumor region (mostly from the primary tumor) might not reflect its expression in the patient's metastatic lesions that are the target of systemic therapy (41). The design of clinical trials that include collection of both pretreatment primary tumors and metastatic lesions for correlative studies would be very helpful to address the impact of tumor heterogeneity in biomarker discovery and validation.

Further limitations of our work include the prospective–retrospective nature of the study and the use of samples from a noncontrolled clinical trial where postnivolumab therapy confounded an OS analysis. For this reason, we chose to focus on clinical endpoints more directly affected by the initial therapy (e.g., ORR/irORR and PFS/irPFS). In the CheckMate-025 trial, there was no association between PD-L1 expression and improved OS on nivolumab after a minimum follow-up of 14 months (2). It remains to be tested if the multifactorial model we developed by studying irRECIST endpoints in a mature CheckMate-010 cohort can identify durable responders more accurately than PD-L1 expression alone when the CheckMate-025 trial reaches longer follow-up. Ultimately, our model needs to be validated in prospective biomarker-driven trials such as Hoosier Clinical Research Network GU-260 (NCT03117309).

S.A. Shukla holds ownership interest (including patents) in 152 Therapeutics. D.A. Braun is a consultant/advisory board member for Octane Global, Defined Health, Dedham Group, Adept Field Solutions, Slingshot Insights, Blueprint Partnership, Charles River Associates/KC2 Medical, and Schlesinger Trinity Group. C.E. Horak is an employee of and holds ownership interest (including patents) in Bristol-Myers Squibb. M. Wind-Rotolo is an employee of and holds ownership interest (including patents) in Bristol-Myers Squibb. P.J. Catalano is a consultant/advisory board member for Eli Lilly. G.J. Freeman reports receiving commercial research grants from Bristol-Myers Squibb, Roche/Genentech, Novartis, VCB, and Ipsen; holds ownership interest (including patents) in Novartis, Roche/Genentech, Bristol-Myers Squibb, Amplimmune/AstraZeneca, Merck, EMD Serono, Beohringer Ingelheim, and Dako; and is a consultant/advisory board member for Novartis, Lilly, Roche/Genentech, Bristol-Myers Squibb, Bethyl Laboratories, Xios Therapeutics, Quiet Therapeutics, and Seattle Genetics. A.H. Sharpe reports receiving commercial research grants from Novartis, UCB, Ipsen, Roche, Bristol-Myers Squibb, and Quark; holds ownership interest (including patents) in Novartis, Roche/Genentech, Bristol-Myers Squibb, Amplimmune/AstraZeneca, Merck, EMD Serono, Boehringer-Ingelheim, and Dako; and is a consultant/advisory board member for Novartis, Surface Oncology, Sqz Biotech, Adaptimmune, Elstar, Elpiscience, Selecta, Monopteros, Lilly, Roche/Genentech, Bristol-Myers Squibb, Bethyl Laboratories, Xios Therapeutics, Quiet Therapeutics, and Seattle Genetics. F.S. Hodi reports receiving commercial research grants from Bristol-Myers Squibb and Novartis to his institution; holds ownership interest (including patents) in Apricity; is listed as an inventor on a patent regarding MICA-related disorders to be licensed to his institution; and is a consultant/advisory board member for Bristol-Myers Squibb, Merck, Genentech, EMD Serono, Sanofi, Celldex, Novartis, Amgen, Incyte, Bayer, Aduro, Partners Therapeutics, Pfizer, Pionyr, 7 Hills Pharma, Torque, Compass, Takeda, and Surface. R.J. Motzer reports receiving commercial research grants from Bristol-Myers Squibb, Pfizer, Genentech/Roche, and Eisai, and is a consultant/advisory board member for Pfizer, Genentech/Roche, Novartis, Eisai, and Exelixis. T.K. Choueiri reports receiving commercial research grants from AstraZeneca, Alexion, Bayer, Bristol-Myers Squibb, Cerulean, Eisai, Foundation Medicine, Exelixis, Ipsen, Tracon, Genentech, Roche, Hoffman-La Roche, GlaxoSmithKline, Merck, Novartis, Peloton, Pfizer, Prometheus, Corvus, Calithera, Analysis Group, Sanofi/Aventis, and Takeda; and is a consultant/advisory board member for AstraZeneca, Alexion, Sanofi/Aventis, Bayer, Bristol-Myers Squibb, Cerulean, Eisai, Foundation Medicine, Exlixis, Genentech, Heron Therapeutics, Roche, GlaxoSmithKline, Merck, Novartis, Peloton, Pfizer, EMD Serono, Prometheus, Corvus, Ipsen, Up-to-Date, NCCN, and Analysis Group. M.B. Atkins is a consultant/advisory board member for Bristol-Myers Squibb, Merck, Novartis, Pfizer, Genentech, Eisai, Exelixis, Arrohead, COTA, and Agenus. D.F. McDermott reports receiving other commercial research support from Bristol-Myers Squibb, Prometheus Laboratories, Merck, Genentech, Pfizer, Exelixis, Novartis, X4 Pharma, Alkermes, Inc., and Peloton; and is a consultant/advisory board member for Bristol-Myers Squibb, Pfizer, Merck, Exelixis, Array BioPharma, Genentech BioOncology, Alkermes, Inc., Jounce Therapeutics, X4 Pharma, Eli Lilly, Peloton Therapeutics, and EMD Serono. S. Signoretti reports receiving commercial research grants from Bristol-Myers Squibb, AstraZeneca, and Exelixis; is a consultant/advisory board member for Merck, AstraZeneca, AACR, and NCI; and receives royalties from Biogenex. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.-C. Pignon, C.E. Horak, M. Wind-Rotolo, R.J. Motzer, T.K. Choueiri, M.B. Atkins, D.F. McDermott, S. Signoretti

Development of methodology: J.-C. Pignon, A. Flaifel, J.S. Novak, G.J. Freeman, T.K. Choueiri

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.-C. Pignon, C.E. Horak, Y. Ishii, J. Grosha, A. Flaifel, J.S. Novak, K.M. Mahoney, F.S. Hodi, R.J. Motzer, T.K. Choueiri, D.F. McDermott

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.-C. Pignon, O. Jegede, S.A. Shukla, D.A. Braun, M. Wind-Rotolo, P.J. Catalano, J. Grosha, A. Flaifel, F.S. Hodi, R.J. Motzer, C.J. Wu, M.B. Atkins, S. Signoretti

Writing, review, and/or revision of the manuscript: J.-C. Pignon, O. Jegede, D.A. Braun, C.E. Horak, M. Wind-Rotolo, P.J. Catalano, K.M. Mahoney, G.J. Freeman, A.H. Sharpe, F.S. Hodi, R.J. Motzer, T.K. Choueiri, C.J. Wu, M.B. Atkins, D.F. McDermott, S. Signoretti

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.-C. Pignon, J. Grosha, T.K. Choueiri

Study supervision: J.-C. Pignon, T.K. Choueiri, C.J. Wu, S. Signoretti

S. Signoretti, P.J. Catalano, G.J. Freeman, A.H. Sharpe, T.K. Choueiri, M.B. Atkins, and D.F. McDermott have received support from NIH/NCI DF/HCC Kidney Cancer SPORE P50-CA101942. S.A. Shukla has received support from NIH/NCI R50CA211482. T.K. Choueiri has received support from The Trust Family, Michael Brigham, and Loker Pinard Funds for Kidney Cancer Research at Dana-Farber 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.