Research ArticleCancer

MHC proteins confer differential sensitivity to CTLA-4 and PD-1 blockade in untreated metastatic melanoma

See allHide authors and affiliations

Science Translational Medicine  18 Jul 2018:
Vol. 10, Issue 450, eaar3342
DOI: 10.1126/scitranslmed.aar3342

MHC-ing immunotherapy response

Currently, there is no way to predict response to anti–CTLA-4 cancer immunotherapy. Using data from two published independent phase 2 clinical trials, Rodig et al. showed that MHC class I expression in advanced melanoma predicted resistance to anti–CTLA-4, but not anti-PD-1, treatment, which may need MHC class II to be effective. These results may explain why patients on combined therapy do better on average, with one drug overcoming the limitations of the other. The combination is also more toxic than single agents; knowing which drug to administer to which patients could make melanoma immunotherapy less taxing without sacrificing efficacy.

Abstract

Combination anti–cytotoxic T lymphocyte antigen 4 (CTLA-4) and anti–programmed cell death protein 1 (PD-1) therapy promotes antitumor immunity and provides superior benefit to patients with advanced-stage melanoma compared with either therapy alone. T cell immunity requires recognition of antigens in the context of major histocompatibility complex (MHC) class I and class II proteins by CD8+ and CD4+ T cells, respectively. We examined MHC class I and class II protein expression on tumor cells from previously untreated melanoma patients and correlated the results with transcriptional and genomic analyses and with clinical response to anti–CTLA-4, anti–PD-1, or combination therapy. Most (>50% of cells) or complete loss of melanoma MHC class I membrane expression was observed in 78 of 181 cases (43%), was associated with transcriptional repression of HLA-A, HLA-B, HLA-C, and B2M, and predicted primary resistance to anti–CTLA-4, but not anti–PD-1, therapy. Melanoma MHC class II membrane expression on >1% cells was observed in 55 of 181 cases (30%), was associated with interferon-γ (IFN-γ) and IFN-γ–mediated gene signatures, and predicted response to anti–PD-1, but not anti–CTLA-4, therapy. We conclude that primary response to anti–CTLA-4 requires robust melanoma MHC class I expression. In contrast, primary response to anti–PD-1 is associated with preexisting IFN-γ–mediated immune activation that includes tumor-specific MHC class II expression and components of innate immunity when MHC class I is compromised. The benefits of combined checkpoint blockade may be attributable, in part, to distinct requirements for melanoma-specific antigen presentation to initiate antitumor immunity.

INTRODUCTION

Patients with advanced melanoma derive greater benefit from combined treatment with antibodies targeting cytotoxic T lymphocyte antigen 4 (CTLA-4) and programmed cell death protein 1 (PD-1) than from either antibody alone (1, 2). However, the shared and unique biological effects derived from inhibiting the two immune checkpoint proteins are still poorly understood. In vitro and preclinical models show that CTLA-4, expressed by T cells, binds members of the B7 family expressed by antigen-presenting cells (APCs) to inhibit T cell costimulation during the priming and effector phases of T cell activation (3, 4). PD-1, expressed by activated T cells, binds the PD-1 ligands expressed by tumors and APCs to inhibit T cell effector function, a reversible phenotype termed “exhaustion” (5, 6). Combined CTLA-4 and PD-1 blockade relieves both inhibitors of T cell activity and generally results in superior antitumor activity compared with either therapy alone (79). However, antitumor immune responses remain dependent on T cell recognition of tumor-specific antigens in the context of major histocompatibility complex (MHC) proteins to affect tumor regression (1012).

Validation of these mechanisms of action in patients treated with therapeutic anti–CTLA-4 and anti–PD-1 antibodies is ongoing. There are presently no cellular or protein biomarkers predictive of response to anti–CTLA-4 treatment. In contrast, the presence of a T cell infiltrate, PD-1+ immune cells, and programmed cell death ligand 1–positive (PD-L1+) inflammatory and tumor cells in pretreatment biopsy samples are indicators of ongoing but exhausted antitumor immunity and predict better clinical response and overall patient survival with anti–PD-1 treatment (1318). Tumors with a high burden of somatic mutations are also associated with better outcomes among patients treated with anti–CTLA-4 or anti–PD-1 therapy, presumably because of an increased number of neopeptide antigens presented by tumor cells to T cells in the context of MHC (1921).

Emerging data suggest that reduction or loss of proteins associated with antigen presentation can be a mechanism to evade antitumor immunity after immune checkpoint blockade (22). However, the prevalence and significance of altered antigen presentation remain incompletely characterized. We evaluated the expression of MHC class I and class II proteins in pretreatment biopsy samples from patients treated with ipilimumab (IPI; anti–CTLA-4) followed by nivolumab (NIVO; anti–PD-1) (IPI→NIVO), NIVO followed by IPI (NIVO→IPI), IPI alone, or concurrent NIVO + IPI in the clinical trial setting. The results were correlated with transcriptional and genomic profiles available for a subset of the cases and with clinical outcomes.

RESULTS

Expression of MHC proteins in untreated melanoma

Pretreatment biopsies analyzed by immunohistochemistry (IHC) with an antibody recognizing all classical MHC class I proteins revealed positive membrane staining of all cells in normal skin, including SOX10+ melanocytes (fig. S1A). IHC performed on biopsy samples from patients with advanced-stage melanoma enrolled in CheckMate 064 revealed variable expression of MHC class I proteins on SOX10+ melanoma cells compared with nonmalignant cells within the same tissue before immunotherapy (Fig. 1, A and B, fig. S1, and table S1). Malignant cells lacking membrane staining included those with only cytoplasmic staining and those with no staining. Membrane MHC class I expression on melanoma cells ranged from 100 to 0% (median, 70%). Lack of MHC class I expression on most of malignant cells (>50%) was observed in 34 of 92 cases (37%; Fig. 1B).

Fig. 1 Summary of MHC class I and MHC class II data from the CheckMate 064 trial.

MHC class I and MHC class II expression is shown in melanoma biopsy samples. (A) Double chromogenic IHC with antibodies targeting SOX10 (red coloration) and MHC class I (i, iii, v; brown coloration) in representative cases showing membrane staining of all SOX10+ melanoma cells (i), subset of melanoma cells (iii, scored as 50% tumor cell–positive), and no melanoma cell–positive (v; yellow arrow indicates PAX5-negative lymphocyte with positive membrane staining), or SOX10 (red coloration) and MHC class II (ii, iv, vi; brown coloration) in representative cases showing positive membrane staining MHC class II in all SOX10+ melanoma cells (ii) and no melanoma cell–positive (iv and vi; yellow arrow in iv indicates PAX5-negative lymphocyte with positive membrane staining). (B) Heat map representing the positive membrane staining of tumor cells (0 to 100%) for MHC class I (row 1) and MHC class II (row 2) proteins, the relative RNA expression of HLA-A, HLA-B, and HLA-C (rows 3 to 5, respectively), and the presence and type of alterations in HLA, B2M, STAT1, JAK2, and JAK1 genes (rows 6 to 10, respectively) for all cases (columns) with available data in the series. The threshold of 50% of tumor cells positive for MHC class I is indicated (red arrow and line). HLA, human leukocyte antigen; CN, copy number. (C to F) Kaplan-Meier estimates of overall survival (OS) by MHC class I and MHC class II expression in baseline biopsy samples according to treatment arms. OS according to expression of MHC class I is divided according to the optimum (50%) for the IPI→NIVO arm (C; P = 0.01) and the NIVO→IPI arm (D; P = 0.46). OS according to the expression of MHC class II is divided according to the optimum (1%) for the IPI→NIVO arm (E; P = 0.14) and the NIVO→IPI arm (F; P = 0.01). Patients with incomplete biomarker data are indicated (solid gray lines). Vertical dashed lines indicate the two critical time points for this analysis: 13 weeks (point of therapy switch) and 25 weeks (initiation of maintenance NIVO). The number of evaluated at-risk patients was 42 for the IPI→NIVO arm and 50 for the NIVO→IPI arm. Median OS (in months) and the HR for OS, with the 95% CIs for each, are listed below the curves. NR, not reached.

The percentages of melanoma tumor cells with MHC class I protein expression correlated with the percentages of tumor cells displaying the independently assessed protein biomarker β-2-microglobulin (β2M) (ρ = 0.74; P < 0.001; nonmalignant cells served as an internal positive control for β2M staining; fig. S2A). The distribution of IHC scores for MHC class I and β2M suggested that IHC for MHC class I was more sensitive than IHC for β2M for detecting positive membrane staining and was therefore used for subsequent analyses (fig. S2A). Reduced amounts of MHC class I protein expression on melanoma cells were also associated with reduced amounts of the MHC class I transcripts HLA-A, HLA-B, and HLA-C in the bulk population (P = 0.007, Wilcoxon rank sum test; Fig. 1B and fig. S2B). Reduced HLA-A, HLA-B, and HLA-C transcripts further correlated with reduced B2M transcripts across samples (ρ = 0.8 to 0.87; P < 0.001), consistent with coordinately reduced transcriptional expression of these MHC class I pathway components.

A stop codon in the coding region of B2M was observed at a low allele frequency (6%) for one case with positive membrane staining for MHC class I and β2M proteins in >90% of melanoma cells (Fig. 1B). There were no mutations in the HLA-A, HLA-B, HLA-C, or B2M genes that could explain reduced MHC class I transcript and protein expression in additional cases. Interferon-γ (IFN-γ)–mediated Janus kinase (JAK)–signal transducer and activator of transcription (STAT) signaling is a potent regulator of MHC class I and class II expression, and somatic mutations have been reported in JAK1 and JAK2 in a small number of tumors resistant to immunotherapy (22). We found a disruptive in-frame deletion in JAK2 at low frequency (4%) in one case with MHC class I protein expression in >90% of cells (Fig. 1B). There were no mutations in JAK1, JAK2, or STAT1 that could explain reduced MHC class I transcript and protein expression in additional cases. Together, these data indicate that most or complete loss of MHC class I/β2M protein expression, likely attributable to transcriptional down-regulation of HLA and B2M genes, is common in advanced melanomas before immunotherapy.

IHC performed with an antibody recognizing all classical MHC class II proteins revealed positive membrane staining of cells morphologically consistent with Langerhans cells in normal skin, but no staining of SOX10+ melanocytes or keratinocytes (fig. S1K). More than 1% of melanoma cells expressed membrane MHC class II in 26 of 92 cases (28%) from CheckMate 064 (Fig. 1, A and B, and table S1). In contrast to MHC class I, the percentage of malignant cells with positive staining for MHC class II was generally very low (median, 10%; range, 1 to 100%) and was concentrated at the inflammatory invasive margin of the tumor, consistent with induced local expression (fig. S1, L to N).

MHC protein expression and response to checkpoint blockade in CheckMate 064

We next examined whether tumor MHC class I or MHC class II expression in pretreatment biopsy samples was associated with disease progression with single-agent IPI or single-agent NIVO at week 13, the time of therapy switch, for patients enrolled in CheckMate 064 (fig. S3A shows the trial schematic). Reduced tumor MHC class I expression (≤30%) identified patients who are more likely to have progressive disease than complete response, partial response, or stable disease at week 13 after single-agent IPI (P = 0.02, Fisher’s exact test; Table 1). No amount of MHC class I expression distinguished patients with from those without progressive disease at week 13 after single-agent NIVO (Table 1). Conversely, MHC class II expression (>1%) associated with patients who are more likely to have complete or partial responses or stable disease than progressive disease at week 13 after single-agent NIVO, although the P value was just outside the range of statistical significance (P = 0.0517, Fisher’s exact test; Table 1). No amount of tumor MHC class II expression distinguished patients with complete responses, partial responses, or stable disease from those with progressive disease at week 13 after single-agent IPI (Table 1). Given that single-agent therapy for each arm of the study was of limited duration (13 weeks), we did not further subdivide the response assessment (that is, into complete/partial response and stable disease) for these comparisons. The results were highly significant when best overall response for the trial was also considered (P = 0.03 for MHC class I in IPI→NIVO, P = 0.005 for MHC class II in NIVO→IPI; Fisher’s exact test; table S2).

Table 1 Progressive disease at week 13 according to MHC class I and MHC class II thresholds in CheckMate 064.
View this table:

For a small set of patients, matched baseline and week 13 biopsy samples were available for comparison (n = 10 for IPI, n = 11 for NIVO; fig. S4). We did not observe changes in MHC class I or MHC class II expression status (high versus low) between baseline and week 13 samples among patients treated with IPI. We observed changes in MHC class I and/or class II status between baseline and week 13 samples for a minority of patients treated with NIVO (4 of 11), but we found no strict correlation between those patients’ status at week 13 and progressive disease (fig. S4).

When overall survival was examined, low baseline tumor MHC class I expression (≤50%) was associated with inferior overall survival for patients initially treated with IPI, despite an eventual exposure to NIVO for patients continuing in the trial [IPI→NIVO treatment arm; hazard ratio (HR), 0.38; 95% confidence interval (CI), 0.18 to 0.82; P = 0.01; Fig. 1C]. No amount of tumor MHC class I expression, including the 50% threshold (HR, 0.70; 95% CI, 0.27 to 1.18; P = 0.46), identified a population with inferior overall survival when initially treated with NIVO (Fig. 1D). Conversely, baseline tumor MHC class II expression (>1%) was associated with better overall survival for patients initially treated with NIVO, despite eventual exposure to IPI for patients continuing in the trial (NIVO→IPI arm; HR, 0.11; 95% CI, 0.02 to 0.83; P = 0.01; Fig. 1F). No amount of tumor MHC class II expression, including the 1% threshold (HR, 0.50; 95% CI, 0.20 to 1.25; P = 0.14), reached statistical significance for predicting the overall survival of patients initially treated with IPI (Fig. 1E).

MHC protein expression and response to checkpoint blockade in CheckMate 069

To validate the associations between melanoma MHC class expression and survival in a distinct cohort, we evaluated baseline biopsy samples from treatment-naïve patients treated with single-agent IPI (with the option of NIVO at disease progression) or concurrent NIVO + IPI while enrolled in CheckMate 069 (fig. S3B shows the study schematic). Lack of MHC class I expression on most of malignant cells (>50%) was observed in 44 of 89 cases (49%; Fig. 2, A and B, and table S3). More than 1% of melanoma cells expressed membrane MHC class II in 29 of 89 cases (33%) from CheckMate 069 (Fig. 2, C and D). Among patients treated with single-agent IPI (n = 26), reduced tumor MHC class I expression [at the previously defined optimal threshold for overall survival in CheckMate 064 (≤50%)] was associated with a best overall response of progressive disease (P = 0.01; Table 2). Given that most patients received a defined therapy for the course of the trial, we further subdivided the best overall responses according to clinical nonresponse (progressive or stable disease) versus response (complete or partial response) for comparison. Reduced tumor MHC class I expression perfectly predicted a lack of clinical response to IPI (negative predictive value, 100%; 95% CI, 74 to 100%; Table 2). Although the P value was outside the range of significance (P = 0.057; Fig. 2A), reduced MHC class I (≤50%) was additionally associated with inferior overall survival after IPI treatment, with visible separation of the Kaplan-Meier curves (Fig. 2A), and an HR (0.34; 95% CI, 0.11 to 1.03) nearly identical to that found for patients initially treated with IPI in CheckMate 064 (Fig. 1C). Among patients treated with concurrent NIVO + IPI (n = 63), reduced MHC class I expression was not associated with progressive disease, inferior response, or inferior overall survival (Table 2 and Fig. 2B). Positive tumor MHC class II expression, which was associated with superior overall survival among patients initially treated with single-agent NIVO before single-agent IPI in CheckMate 064 (Fig. 1F), did not identify a group with superior clinical response or superior overall survival among patients treated with single-agent IPI or concurrent NIVO + IPI in CheckMate 069 (Table 2 and Fig. 2, C and D). Thus, these data suggest that robust tumor-specific MHC class I expression is essential for response to anti–CTLA-4 therapy but not to anti–PD-1 or combination therapy, whereas tumor-specific MHC class II expression is associated with superior response to anti–PD-1 therapy, but not to anti–CTLA-4 or combination therapy.

Fig. 2 Summary of MHC class I and MHC class II data from the CheckMate 069 trial.

Kaplan-Meier estimates of OS by MHC class I and MHC class II expression in baseline biopsy samples according to treatment arm are shown. OS was divided according to the determined threshold for MHC class I (50%) in the (A) IPI arm (P = 0.057) and (B) NIVO + IPI arm (P = 0.66). OS was also divided according to the determined threshold for MHC class II (1%) in the (C) IPI arm (P = 0.16) and (D) NIVO + IPI arm (P = 0.25). The number of evaluated at-risk patients was 26 for the IPI arm and 63 for the NIVO + IPI arm. The median OS (in months) and the HR, with 95% CIs for each, are listed below the curves.

Table 2 Disease progression according to defined biomarker thresholds in CheckMate 069.

NPV, negative predictive value; PPV, positive predictive value.

View this table:

Transcriptional signatures of response to checkpoint blockade in CheckMate 064

To better define the immunological criteria for response to immune checkpoint blockade, we analyzed RNA sequencing (RNA-seq) data from freshly frozen biopsy samples obtained during CheckMate 064. A 10-gene IFN-γ–related transcriptional signature (table S4), previously suggested as predictive of response to the anti–PD-1 agent pembrolizumab (23), was significantly higher in baseline biopsy samples from patients without progressive disease compared to those with progressive disease at week 13 after single-agent NIVO (P = 0.03, Wilcoxon rank sum test; Fig. 3A). This association was not observed among patients treated with single-agent IPI (P = 0.18, Wilcoxon rank sum test; Fig. 3A). Similarly, the IFN-γ signature was significantly higher in baseline biopsy samples from patients without progressive disease (complete response, partial response, and stable disease) as best overall response compared with patients with progressive disease in the NIVO→IPI arm of the trial (P = 0.003, Wilcoxon rank sum test; Fig. 3B). This association was not observed among patients in the IPI→NIVO arm (P = 0.20, Wilcoxon rank sum test; Fig. 3B).

Fig. 3 Predictive value of defined gene set scores derived using RNA-seq data from baseline biopsy samples from CheckMate 064.

(A) Gene set scores and response at week 13 for patients treated with NIVO (left) or IPI (right). (B) Gene set scores and response for the IPI→NIVO arm (left) and NIVO→IPI arm (right). (A) and (B) include gene set scores derived from a published IFN-γ signature (left box plots) (23), the top 25 differentially expressed gene transcripts for distinguishing patients according to best overall response in the NIVO→IPI arm of CheckMate 064 (middle box plots), and a curated set of 13 IFN-γ–related genes derived from the 25-gene set (right box plots). Response is divided according to progressive disease (PD) (red) or nonprogressive disease [complete response (CR), partial response (PR), or stable disease (SD); blue]. Statistical comparisons are based on one-sided Wilcoxon rank sum tests. (C) A heat map shows the relative expression of the top 25 gene transcripts differentially expressed between patients with progressive disease from those without progressive disease as best overall response (BOR) using RNA-seq data from baseline biopsy samples in the NIVO→IPI arm (P = 0.0002 to 0.01; false discovery rates, 0.14 to 0.27). Each column represents a sample. Rows represent response at week 13 (RESW13I, row 1), best overall response (BORI, row 2), and the top 25 differentially expressed transcripts grouped according to unsupervised hierarchical clustering and color-coded to indicate their relative abundance (rows 3 to 27). Gene transcripts encoding IFN-γ, markers of cells responsible for IFN-γ, transcription factors that promote IFN-γ production, or targets expressed in response to IFN-γ signaling are indicated in blue and comprise the 13 IFN-γ–related gene set. (D) Pairwise correlations using Kendall’s τ between indicated protein and transcriptional signature biomarkers across all baseline biopsy samples with available data. The empirically derived IFN-γ gene set score positively correlates with baseline melanoma MHC class II protein expression across all samples (Kendall’s pairwise correlation τ = 0.43, P < 0.001). The IFN-γ gene set score and the previously published IFN-γ signature, respectively, positively correlate with the number of tumor-associated CD3+ (P < 0.0001 and P < 0.0001), CD4+ (P = 0.0012 and P = 0.0005), and CD8+ (P < 0.0001 and P < 0.0001) T cells and with melanoma PD-L1 expression (P < 0.0001 and P < 0.0001). MHC class I protein expression more weakly positively correlates with our IFN-γ gene set (pairwise τ = 0.19, P = 0.02). N.S., not significant; #, number.

To independently uncover the immunological determinants of immunotherapy response, we examined the RNA-seq data for 770 immune-associated genes and identified the top 25 differentially expressed gene transcripts associated with a lack of progressive disease as the best overall patient response in patients receiving NIVO before IPI in CheckMate 064 (Fig. 3C). Thirteen of the top 25 immune transcripts were directly related to IFN-γ (Fig. 3C). These included a cytokine (IFNG), a transcription factor associated with IFN-γ production [TBX21 (Tbet)], markers of adaptive and innate immune cells responsible for IFN-γ production (CD8A, TIGIT, and KLRD1), and major gene targets of IFN-γ–mediated signaling (Fig. 3C) (2426). Six genes identified by this method were shared with the independently discovered, pembrolizumab-derived gene signature (table S4).

We found that a gene set score derived from the top 25 differentially expressed genes was significantly higher in baseline biopsy samples from patients without progressive disease compared with those with progressive disease at week 13 after single-agent NIVO, as expected (P = 0.004, Wilcoxon rank sum test; Fig. 3A). However, we did not find this association among patients receiving single-agent IPI (P = 0.26, Wilcoxon rank sum test; Fig. 3A). In addition, we found that a distilled 13-gene IFN-γ–related gene set score was significantly higher in baseline biopsy samples from patients without progressive disease compared with those with progressive disease at week 13 after single-agent NIVO (P = 0.01, Wilcoxon rank sum test; Fig. 3A). We did not find this association among patients receiving single-agent IPI (P = 0.24, Wilcoxon rank sum test; Fig. 3A). We found similar results when the best overall response of nonprogressive disease versus progressive disease for the full trial was used as an end point (Fig. 3B).

IFN-γ is produced by activated T cells and natural killer (NK) cells and directly and potently induces MHC class II expression on tumor and immune cells (26). The predictive power of baseline IFN-γ transcriptional signatures suggested a biological link between IFN-γ and the focal melanoma MHC class II expression that we observed at the inflammatory margin for a subset of tumors. Consistent with this, we observed a positive correlation between our empirically derived IFN-γ gene set score and baseline melanoma MHC class II protein expression across all samples (Kendall’s pairwise correlation τ = 0.43, P < 0.001; Fig. 3D). The magnitudes of our IFN-γ gene set score and the previously published IFN-γ signature were positively correlated with the number of tumor-associated CD3+, CD4+, and CD8+ T cells, known sources of IFN-γ, and with melanoma PD-L1 expression, a known target of IFN-γ activity (Fig. 3D). Although MHC class I is also a target of IFN-γ, our IFN-γ gene set was more weakly correlated with tumor MHC class I protein expression than the other tissue biomarkers (pairwise τ = 0.19, P = 0.02; Fig. 3D). Together, these data support the notion that baseline IFN-γ and markers of IFN-γ–mediated inflammation, including melanoma-specific MHC class II and PD-L1 expression, are associated with subsequent response to NIVO but not to IPI.

Transcriptional signatures of innate immune cells and response in CheckMate 064

The top 25 immune transcripts associated with a lack of disease progression among patients initially treated with NIVO included transcripts expressed by NK cells (KLRD1 and TIGIT) and those that stimulate NK cells (IL15, also important for memory CD8+ T cells). These data suggested that other baseline effector immune cell populations such as NK cells or γδ T cells may be responsible, in part, for IFN-γ and for response after NIVO among patients with loss of melanoma MHC class I expression. To examine this possibility, we manually selected a set of lineage-specific gene transcripts to comprise an NK cell gene set and a γδ T cell gene set, respectively (table S5). Among patients initially treated with NIVO and who achieved a best overall response of complete response, partial response, or stable disease, we observed a significantly higher baseline transcriptional signature of γδ T cells, NK cells, and interleukin 15 (IL-15) compared with those with a best overall response of progressive disease (P = 0.002, P = 0.01, and P < 0.001, respectively, Wilcoxon rank sum test; Fig. 4). This association was not observed among patients initially treated with IPI. The associations between best overall response and the γδ T cell and IL-15 signatures were also significant for patients initially treated with NIVO and with ≤50% tumor MHC class I expression (P = 0.013 and P = 0.0012, respectively, Wilcoxon rank sum test; Fig. 4). IHC for T cell receptor δ (TCRδ) and CD56 confirmed the presence of γδ T cells and NK cells, respectively, within the tumor microenvironment, including cases with low tumor MHC class I expression (Fig. 4).

Fig. 4 Cell type–specific gene set scores and best overall response for patients treated sequentially with NIVO→IPI or IPI→NIVO in CheckMate 064.

Gene set scores derived from RNA-seq data from baseline biopsy samples were divided according to a best overall response of progressive disease (red) or nonprogressive disease (complete response, partial response, or stable disease; blue) for all samples or for samples with low MHC class I expression (≤50% of cells). (A) γδ T cell gene set scores per arm. Representative image of sample stained for SOX10 (red) and TCRδ (brown) is shown with arrows indicating the presence of γδ T cells in the tumor microenvironment. (B) NK cell gene set scores per arm. Representative image of sample stained for SOX10 (red) and CD56 (brown) is shown with arrows indicating the presence of CD56+ NK cells in the tumor microenvironment. (C) IL-15 gene expression from baseline biopsy samples per arm. Statistical comparisons are based on one-sided Wilcoxon rank sum tests.

DISCUSSION

Combination therapy with NIVO and IPI is known to result in superior overall survival among patients with advanced-stage melanoma compared with either treatment alone (1, 2). These results are consistent with preclinical studies reporting that anti–CTLA-4 and anti–PD-1 treatment promote T cell immunity through distinct and complementary mechanisms. However, T cells must nonetheless recognize tumor cell antigens in the context of MHC to stimulate an antitumor immune response.

We found that the loss of most or all MHC class I protein expression is very common in melanoma before immunotherapy, occurring in 43% of all cases. Attenuated MHC class I protein expression was accompanied by a coordinate reduction in β2M protein and HLA-A, HLA-B, HLA-C, and B2M transcripts. However, we could not attribute reduced expression of these genes to disruptive somatic mutations. Whether dysregulation of the MHC class I “master regulator” such as NLRC5 or, potentially, epigenetic silencing of MHC class I–associated genes underlies diminished MHC class I/β2M expression is under investigation (2729). Regardless of the mechanism, reduction and loss of MHC class I protein expression are common among melanomas before immune checkpoint therapy.

Critically, reduced melanoma MHC class I expression was associated with primary resistance to IPI, but not to NIVO. This association was observed for patients receiving single-agent IPI, was not altered with subsequent switch to NIVO in CheckMate 064, and was validated for patients receiving single-agent IPI until disease progression in CheckMate 069. More specifically, the loss of MHC class I was highly associated with progressive disease and predicted a lack of clinical response (progressive disease or stable disease; negative predictive value, 100%) for patients receiving single-agent IPI in CheckMate 069. The HR for survival in the IPI arm of CheckMate 069 (HR = 0.34) was similar to that of the IPI→NIVO arm of CheckMate 064 (HR = 0.38) despite a smaller sample size. Although loss of MHC class I was associated with progressive disease among patients receiving IPI before NIVO, we observed no such association among patients receiving concurrent NIVO and IPI. Thus, in the absence of robust MHC class I antigen presentation, treatment with NIVO early in the course of therapy appears critical.

These observations refine our understanding of anti–CTLA-4 clinical activity (5, 30, 31). Anti–CTLA-4 allows the costimulatory receptor CD28 on T cells to engage CD80/CD86 on APCs laden with tumor antigens and promotes T cell activation during the priming phase of the adaptive immune response in secondary lymphoid organs. In addition, there is increasing evidence that anti–CTLA-4 depletes CTLA-4–expressing T-regulatory cells to increase the relative number of T effector cells at the tumor site (32). Despite the multiple roles of CTLA-4 in regulating the immune response, our data suggest that the major consequence of anti–CTLA-4 therapy is the development of active CD8+ effector T cells that require intact MHC class I–mediated antigen presentation by tumor cells to affect tumoricidal activity. The high prevalence of reduced MHC class I expression on melanoma may explain why most patients do not respond to single-agent IPI (33).

We did not find attenuated MHC class I expression to be associated with primary resistance to single-agent NIVO. This result may seem inconsistent with a recent report of an inactivating mutation in B2M found in a biopsy sample of recurrent melanoma from a patient with acquired resistance to the anti–PD-1 agent pembrolizumab (22). The reasons for this discrepancy are unclear. First, it is possible that MHC class I expression on a minority of malignant cells is necessary and sufficient for primary response to anti–PD-1 treatment, which can stimulate both innate and adaptive immunity via IFN-γ, but not for primary response to anti–CTLA-4 treatment, which requires high expression of MHC class I to be effective. Second, reduced expression of MHC class I protein reflects general transcriptional repression of multiple genes involved in MHC class I antigen presentation pathway in treatment-naïve melanoma that might be, in part, reversible under the appropriate physiological conditions. Third, it is possible that the MHC dependencies that we observed with initial therapy may change on chronic exposure to immune checkpoint blockade and the eventual emergence of resistance (34). Tumors with acquired resistance to immunotherapy can acquire genetic deficiencies that result in the loss of either constitutive or induced MHC class I (35, 36).

We found that melanoma MHC class II expression was associated with better outcome among patients initially treated with NIVO in CheckMate 064. A switch in therapy to IPI for 13 weeks (followed by maintenance NIVO) did not alter this positive association. This result is both consistent with and validates a previous report that found that focal melanoma MHC class II expression was associated with better outcomes among patients treated with anti–PD-1 or anti–PD-L1 therapy (37).

Tumor antigen presentation via MHC class II may activate tumor-infiltrating CD4+ T cells that have T helper or cytotoxic activities and therefore provide an alternative means to promote an adaptive immune response when MHC class I expression is reduced (38). The most potent stimulus for MHC class II expression is IFN-γ (26). Patients achieving a best overall response of complete response, partial response, or stable disease showed higher expression of IFNG and IFN-γ target gene transcripts compared with patients achieving a best overall of progressive disease when initially treated with NIVO. In contrast, neither a previously published and validated IFN-γ signature, predictive of response to pembrolizumab (23), nor our independently derived IFN-γ gene set predicted a lack of disease progression among patients initially treated with IPI. Although IFN-γ promotes MHC class I expression in most cell types, we observed a weaker correlation between the IFN-γ signature and melanoma-specific MHC class I than between the IFN-γ signature and melanoma-specific MHC class II. These results suggest that the MHC protein classes may be differentially regulated on melanoma cells, at least in part.

CD4+ T cells and NK cells are primary producers of IFN-γ (26). Thus, CD4+ T cells activated via recognition of tumor antigen in the context of MHC class II will produce IFN-γ, which, in turn, will induce additional MHC class II expression on tumor and immune cells at the tumor site. With reduced or absent melanoma MHC class I expression, MHC class II–regulated effector CD4+ T cells may have a more important role in tumoricidal function (38, 39). NK cells and γδ T cells are additional non–MHC class I–restricted sources for IFN-γ. Consistent with a role for these innate immune cells in mediating response, we find that patients with higher baseline expression of NK-cell and γδ T cell–associated gene transcripts have better outcomes after treatment with NIVO. However, in general, we observed fewer numbers of CD56+ NK cells and TCRδ+ T cells than CD3+ T cells within the tumor microenvironment. These data suggest that NIVO exerts its effects through several complementary IFN-γ–related innate and adaptive immune pathways when tumor MHC class I is compromised. We did not observe such an association among patients initially treated with IPI.

Despite our findings, there are limitations inherent to our study that will benefit from the analysis of other clinically annotated cohorts by the current and additional methods. Specifically, it will be important to confirm the associations we have observed between tumor-specific MHC class I and II expression and clinical responses to single-agent IPI and NIVO, respectively, in clinical cohorts that include greater patient numbers and with longer clinical follow-up. A further limitation is our reliance upon retrospective analyses. Ideally, our findings will be further validated in trials in which treatment options are determined in accordance with results of tissue-based biomarker studies (that is, MHC class I and II expression and IFN-γ gene set score), as has been done with tumor-specific PD-L1 expression in lung cancer (40). Finally, our IHC analyses are based on manual scoring by trained pathologists. It will be of interest to determine whether more automated methods of marker quantitation can be developed and implemented for routine use. These may include image analysis algorithms trained on the clinically relevant data sets stained for the MHC proteins or quantitative, targeted transcriptional profiling amenable to formalin-fixed paraffin-embedded (FFPE) tissue.

Together, our data provide insights into the baseline biological determinants of clinical response to single-agent and combination anti–CTLA-4 and anti–PD-1 treatment and, by extension, the immune responses augmented by individual and combined therapies. Clinical efficacy of anti–CTLA-4 therapy is dependent on preexisting, robust expression of MHC class I by tumor cells and is restricted to stimulating an MHC class I–directed cytotoxic T cell immune response. In contrast, efficacy of anti–PD-1 therapy is dependent on at least low-level, preexisting, IFN-γ–mediated inflammation within the tumor microenvironment and is associated with both adaptive and innate immune responses. Combined immune checkpoint blockade provides a further immune stimulus over individual therapies alone and, in addition, overcomes the limitations of each. Patients treated with the combination of NIVO and IPI did not show dependence on MHC class I expression for clinical response in the CheckMate 069 trial, further supporting the notion that alternative antitumor immune mechanisms are stimulated in patients receiving NIVO, either as a single agent or as part of combination therapy. These data provide a basis to further improve our understanding of the effects of immune checkpoint blockade on effector cells beyond CD8+ T cells and suggest future avenues of investigation for manipulating the immune system to effectively treat cancer.

MATERIALS AND METHODS

Study design

The current study used data from two previously published clinical trials. CheckMate 064 was a randomized, open-label, phase 2 study (NCT01783938) evaluating NIVO→IPI versus the reverse treatment sequence IPI→NIVO in patients with histologically confirmed unresectable stage III or stage IV melanoma (fig. S3) (41). Patients were randomized (1:1) without bias either into an induction regiment where NIVO was given at 3 mg/kg every 2 weeks for up to six doses, followed by IPI at 3 mg/kg every 3 weeks for up to four doses, or into a second regiment arm where IPI was followed by NIVO (same dosages). Both cohort arms received NIVO maintenance at 3 mg/kg every 2 weeks until progression, unacceptable toxicity, or for a duration of 2 years from first study treatment. Patient inclusion was limited to treatment-naïve patients or those with one failed nonimmunotherapy systemic treatment, Eastern Cooperative Oncology Group performance status of 0 or 1, and patients with available baseline biopsy tissue. The primary end point of CheckMate 064 was the rate of treatment-related grade 3 to 5 adverse events during the induction period. The secondary end points were response rate as determined by RECIST 1.1 at week 25 and progression rates at weeks 13 and 25. Overall survival analysis was an exploratory end point planned after 65% of the subjects died or 2 years of follow-up time from last subject randomized, whichever came first. The sample size of 138 was not based on power considerations but was chosen to achieve a sufficient level of precision for estimating adverse event rates and provide adequate samples of tumor tissue to achieve stable estimates for exploratory biomarker analyses. All IHC results from patients with evaluable baseline biopsies have been reported; no outliers have been omitted. The schema for CheckMate069 is presented in fig. S3. Its design, including patient selection, randomization, and study end points, closely resembled CheckMate 064 (1). In CheckMate 069, patients receiving IPI alone had the option of receiving NIVO at disease progression (1). Biopsy samples (FFPE and fresh-frozen tissues in CheckMate 064 and FFPE tissues in CheckMate 069) were obtained with approval from institutional review boards.

Chromogenic IHC

Dual IHC for MHC class I (HLA-A, HLA-B, and HLA-C, clone EMR8-5, 1:6000; Abcam) or MHC class II (HLA-DP, HLA-DQ, HLA-DR, clone CR3/43, 1:750; Dako) with the melanoma marker SOX10 (EP268, 1:1500; Cell Marque) was performed using an automated staining system (Bond-III; Leica Biosystems), as previously described (42).

IHC for CD3 (LN10; Leica), CD4 (4B12; Dako), CD8 (C8/144B; Dako), PD-1 (NAT105; Abcam), CD56 (clone 123C3; Dako), TCRδ (clone H-41; Santa Cruz Biotechnology), and β2M (A0072, 1:6000; Dako) was performed either manually (CD3, CD4, CD8, and PD-1) or on Bond RX (β2M, CD56, and TCRδ) per standard protocols. IHC for PD-L1 (28-8; Dako) was performed as part of an investigational use–only kit (PD-L1 IHC pharmDx) on Dako Link 48 (2).

IHC evaluation and scoring

MHC class I, MHC class II, and β2M staining was scored in for the percentage of malignant cells in 10% increments (0 to 100%) with positive membrane staining within the entire tissue section, as determined by the consensus of two pathologists, as previously reported (42). Malignant cells were defined by nuclear staining for the melanoma marker SOX10 and were only scored if they were nucleated and viable. Necrotic or fibrotic tissue was excluded from the analysis. Any MHCI/MHCII/β2M expression in nonmalignant (SOX10-negative) inflammatory cells served as an internal positive staining control. CD3, CD4, CD8, and PD-1 staining was evaluated by pathologist-assisted image analysis (Aperio; Leica Biosystems) (2).

RNA-seq and whole-exome sequencing

RNA and DNA were coextracted by Asuragen Inc. using Qiagen AllPrep DNA/RNA Mini Kits (catalog #80204; Qiagen). RNA-seq was performed by Q2 Solutions/EA Genomics with a minimum of 250 ng of total RNA input using the TruSeq Stranded mRNA with 50 million paired-end reads on Illumina HiSeq. Tumor and normal matched DNA samples were processed for whole-exome sequencing with one normal and two tumor libraries, each using 500 ng of DNA input and constructed with individual barcoded adapters. The libraries were combined into an exome-capture sequencing library using the Roche NimbleGen EZ Exome version 3.0 reagent (Roche Sequencing) following the manufacturer’s instructions. Exome sequence data were generated as 2 × 100 base-pair reads on an Illumina HiSeq 2000 instrument (Illumina).

Somatic mutations in the indicated genes were called with MuTect and Strelka algorithms by comparing tumor exome sequencing data with matched normal peripheral blood mononuclear blood samples (43, 44). Mutations were pooled together, annotated with SnpEff software, and further filtered by qualities and matching to known germline polymorphisms from public databases (45). The GATK-CNV best practices pipeline was used to derive deep copy number losses and high copy number gains based on the paired exome sequencing data (46, 47). Gene transcript expression, derived from RNA-seq data, was examined in the space of a curated set of 770 immune-related gene products (48, 49).

Statistical analysis

For clinical response analysis, only patients with complete IHC data (MHC class I, MHC class II, and β2M) were included. Dichotomous cut points for IHC markers were determined to maximize the associations between biomarker high versus low and clinical outcomes (13-week disease progression and overall survival) and were based on 92 patients from CheckMate 064 (IPI→NIVO, n = 42; NIVO→IPI, n = 50) who had complete IHC data. Cut points for 13-week response were estimated using jackknife resampling of receiver operating characteristic curves and Youden’s index (50). The optimal threshold for distinguishing patients developing disease progression from those without at week 13 was 30% for MHC class I with single-agent IPI (there was no threshold for MHC class I expression that reached statistical significance among patients treated with single-agent NIVO) and 1% for MHC class II among patients treated with single-agent NIVO (there was no threshold that reached significance for MHC class II expression among patients treated with single-agent IPI). Cut points for overall survival were estimated using leave-one-out jackknife resampling of the algorithm of Contal-O’Quigley. These were 50% for MHC class I (which was the optimal threshold predicting overall survival among patients initially treated with IPI) and 1% for MHC class II (which was the optimal threshold predicting overall survival among patients initially treated with NIVO). There was no threshold for MHC class I expression and no threshold for MHC class II expression that reached statistical significance for predicting overall survival among patients initially treated with NIVO and IPI, respectively. Only those optimized thresholds for predicting overall survival that reached statistical significance in CheckMate 064 (50%; 1%) were tested for their ability to predict survival and response in CheckMate 069. Cox proportional hazards regression analyses were used to estimate P values and HRs (biomarker high versus low) with 95% Wald CIs. Rates of response and proportions of patients with progressive disease were compared for biomarker high versus low using Fisher’s exact tests. Kendall’s τ was used to show pairwise correlation between biomarkers for the correlation heat map. Additional details on the methods described above are included in the Supplementary Materials.

SUPPLEMENTARY MATERIALS

www.sciencetranslationalmedicine.org/cgi/content/full/10/450/eaar3342/DC1

Materials and Methods

Fig. S1. Representative IHC images.

Fig. S2. Correlations of MHC class I IHC scores with β2M IHC scores and HLA transcripts.

Fig. S3. Study schemas for the CheckMate 064 and CheckMate 069 trials.

Fig. S4. Comparison of MHC class I and MHC class II expression in paired baseline and week 13 biopsy samples from CheckMate 064.

Table S1. CheckMate 064 MHC class I and MHC class II IHC.

Table S2. Best overall response of PD and non-PD according to optimally defined biomarker thresholds.

Table S3. CheckMate 069 MHC class I and MHC class II IHC.

Table S4. Genes used in the published IFN-γ signature.

Table S5. Genes used in the NK cell and γδ T cell gene sets.

References (5159)

REFERENCES AND NOTES

Acknowledgments: We thank the patients who participated in the CheckMate 064 and CheckMate 069 studies and the staff members at the study sites who cared for them. We thank S. Abdelrahman for expert IHC and M. Lipschitz and A. Lako for editing. We also thank S. Kirov, R. Golhar, and C. Rios at Bristol-Myers Squibb (BMS) for assistance with high-throughput sequencing data processing and management. We thank E. Berghorn, G. Kong, and L. Rollin at BMS for assistance with defining disease progression in CheckMate 064 and CheckMate 069. We thank S. Shukla at the Dana-Farber Cancer Institute for assistance with running Polysolver to call HLA mutations. We thank M. Shipp at the Dana-Farber Cancer Institute for helpful discussions during manuscript development. We thank E. Van Allen at the Dana-Farber Cancer Institute for critical reading of the manuscript. We thank Dako for collaborative development of the PD-L1 IHC 28-8 pharmDx assay. Professional medical writing and editorial assistance were provided by J. DiNieri and C. Hunsberger of StemScientific, an Ashfield Company, funded by BMS. Funding: A portion of this research was funded, in part, through the NIH/National Cancer Institute Cancer Center Support Grant P30 CA008748 (to F.S.H.). Funded by BMS. Author contributions: Patient outcomes and sample collection: J.S.W., J.D.W., M.A.P., A.C.P., J.C., and F.S.H. Study design: S.J.R., D.G., D.G.J., E.G., A.G.-H., S.B.L., C.H., J.S.W., and F.S.H. Primary data collection and interpretation: S.J.R., E.G., S.B.L., C.H., and F.S.H. Bioinformatics and statistics: D.G., D.G.J., A.G.-H., C.J., H.C., J.L.W., and F.S.H. Review and discussion of data: S.J.R., D.G., D.G.J., E.G., A.G.-H., C.J., H.C., S.B.L., C.H., J.S.W., J.L.W., J.D.W., M.A.P., and F.S.H. Writing of the manuscript: S.J.R., D.G., E.G., A.G.-H., and F.S.H. Review and editing of the manuscript: S.J.R., D.G., D.G.J., E.G., A.G.-H., C.H., J.S.W., J.L.W., J.D.W., M.A.P., A.C.P., J.C., and F.S.H. Competing interests: S.J.R. has consulted for PerkinElmer and received research support from Affimed, KITE, Merck, and BMS. J.S.W. has consulted for Celldex, Ichor Medical Systems, cCam Biotherapeutics, Lion Biotechnologies, Pieris Pharmaceuticals, Altor BioScience, BMS, Merck, Genentech, Roche, Amgen, AstraZeneca, GlaxoSmithKline, Daiichi Sankyo, AbbVie, Eisai, CytomX Therapeutics, Nektar, Novartis, and Medivation as well as received research support from BMS, Merck, GlaxoSmithKline, Genentech, Astellas Pharma, Incyte, Roche, and Novartis and honoraria from BMS, Merck, Genentech, AbbVie, AstraZeneca, Daiichi Sankyo, GlaxoSmithKline, Eisai, Altor BioScience, Lion Biotechnologies, Amgen, Roche, Ichor Medical Systems, Celldex, cCam Biotherapeutics, Pieris Pharmaceuticals, CytomX Therapeutics, Nektar, Novartis, and Medivation. J.D.W. reports stock or other ownership in Potenza Therapeutics, Tizona Pharmaceuticals, Adaptive Biotechnologies, Serametrix Corp, Elucida, Imvaq, BeiGene, Ascentage, and Aprea; consulting or advisory role in BMS, MedImmune, Genentech, F-star Biotechnology, Polynoma Pharmaceuticals, Neon, Serametrix Corp, Surface, PsiOxus, Elucida, Syndax, Tizona Pharmaceuticals, Potenza Therapeutics, Chugai, Amgen, Trienza, Inovio, Ascentage, Imvaq, Apricity, and Celgene; and research funding from BMS, MedImmune, Merck, and Genentech. M.A.P. has consulted for BMS, Novartis, Merck, Array BioPharma, Incyte, NewLink Genetics, and Aduro and received honoraria from BMS and Merck. J.C. has consulted for Amgen, Iovance, Replimune, Merck, and BMS. F.S.H. has consulted for BMS, Merck, Novartis, Genentech, EMD Serono, Amgen, and Celldex and received research support from BMS. D.G., D.G.J., E.G., A.G.-H., C.J., H.C., S.B.L., C.H., J.L.W., and A.C.P. declare that they have no competing interests. Data and materials availability: All data associated with this study can be found in the paper or the Supplementary Materials.
View Abstract

Navigate This Article