Durable responses in metastatic cancers have been achieved with a variety of immunotherapies such as interleukin-2, adoptive cell transfer of tumor-infiltrating lymphocytes, antibodies that block cytotoxic T-lymphocyte–associated antigen 4 (CTLA4),1-5 and antibodies that block programmed death 1 (PD-1).6-10 However, in a recent study, approximately 25% of patients with melanoma who had had an objective response to PD-1 blockade therapy had disease progression at a median follow-up of 21 months.11
The mechanisms of immune-resistant cancer progression are mostly unknown. Previous studies involving humans examined the loss of beta-2-microglobulin as a mechanism of acquired resistance to several forms of cancer immunotherapy.12-14 In preclinical models, defects in the interferon signaling pathway have been proposed as a potential mechanism of cancer escape (insensitivity) to immunotherapy.15,16 In the current study, we assessed the effect of anti–PD-1 therapy on cancer genomic evolution, including acquired mutations in the genes affecting the interferon pathway and antigen-presentation pathway, in an effort to determine genetic mechanisms of acquired resistance to PD-1 blockade therapy.
Patients, Response Assessment, and Tumor Biopsies
Of 78 patients with metastatic melanoma who were treated with the anti–PD-1 antibody pembrolizumab at the University of California, Los Angeles (UCLA), 42 had an objective response, of whom 15 went on to have disease progression. Four of these 15 patients met all three selection criteria for this analysis. First, they must have had an objective tumor response while participating in a clinical trial with single-agent pembrolizumab.6,7,10,11 Tumor responses were evaluated at 12 weeks and confirmed 4 weeks later, and patients were assessed by imaging every 12 weeks thereafter with the use of both the Response Evaluation Criteria in Solid Tumors17 and the immune-related response criteria.18 Second, patients had to have late acquired resistance, defined as in situ recurrence or new lesion development, despite continuous dosing, after more than 6 months of tumor response. Third, patients had to have adequate biopsy material for whole-exome sequencing at two time points: before the initiation of pembrolizumab therapy and after disease progression. We processed tumor biopsy samples as described previously to perform pathological analyses, obtain DNA and RNA, and attempt to establish cell lines.19,20
Immunohistochemical, Immunofluorescence, Western Blot, and Flow-Cytometric Analyses
Immunohistochemical and immunofluorescence analyses19 as well as Western blot and flow-cytometric analyses21 were performed and analyzed as described previously. Full methods are included in the Supplementary Appendix, available with the full text of this article at NEJM.org.
Genetic and Transcriptional-Profiling Analyses
Whole-exome sequencing was performed at the UCLA Clinical Microarray Core with the use of the NimbleGen SeqCap EZ Human Exome Library, version 3.0 (Roche). Mutation calling was performed as described previously.22 Selected gene-expression profiling on interferon exposure was performed with the use of nCounter (NanoString Technologies). Whole-exome sequencing data have been deposited in the National Center for Biotechnology Information Sequence Read Archive under the accession number SRP076315.
Patient-derived and previously established human melanoma cell lines were used to analyze recognition by T-cell receptor transgenic T cells23 with the use of in vitro coculture assays that detect antigen-induced release of interferon-γ assessed by enzyme-linked immunosorbent assay. Cell-proliferation and growth-inhibition assays were performed with the use of an automated live-cell imaging system (IncuCyte, Essen BioScience) with or without exposure to interferons. Full methods are described in the Supplementary Appendix.
Data generated and collected by the study investigators were analyzed by the last author, who vouches for the completeness and accuracy of the data, analyses, and reported results. Summaries of the clinical protocol have been reported by Hamid et al.6 and Robert et al.10
Student’s t-test and a two-way analysis of variance were used for cell-culture experiments, with Dunnett’s correction applied for multiple comparisons with untreated controls.
Clinical Course and Immune Infiltrates
We analyzed paired tumor samples from four (nonconsecutive) selected patients with metastatic melanoma who had had a relapse while receiving PD-1–inhibition therapy with pembrolizumab (Tables S1 and S2 in the Supplementary Appendix). All four patients met objective criteria for a partial response,17,18 though with slightly different kinetics (Figure 1FIGURE 1Clinical Pattern of Acquired Resistance to Anti–Programmed Death 1 (PD-1) Therapy in Patient 1., and Figs. S1, S2, and S3 in the Supplementary Appendix). The mean time to relapse was 624 days (range, 419 to 888). The baseline biopsy samples were obtained just before the initiation of pembrolizumab therapy in Patients 2, 3, and 4, whereas for Patient 1, the only available baseline biopsy sample was obtained before an earlier course of therapy with the BRAF inhibitor vemurafenib. The baseline biopsy samples from Patients 1, 2, and 3 showed preexisting CD8 T-cell infiltrates at the invasive margin that colocalized with programmed death ligand 1 (PD-L1) expression on surrounding macrophages and melanoma cells (Figure 1B, and Figs. S1B and S2B in the Supplementary Appendix). The biopsy samples obtained at the time of response in Patients 2, 3, and 4 showed a marked increase in intratumoral CD8 T-cell infiltrates (Figs. S1C, S2C, and S3C and Table S3 in the Supplementary Appendix; no biopsy sample during therapy was available for Patient 1). At the time of relapse, all four biopsy samples showed CD8 T-cell infiltration and PD-L1 expression concentrated at the tumor margins again (Figure 1C and Figs. S1D, S2D, and S3D in the Supplementary Appendix). Multiplex immunofluorescence assays revealed that melanoma cells at the time of relapse in Patients 1 and 2 were negative for PD-L1 even when directly adjacent to T cells, whereas macrophages and stromal cells were positive for PD-L1.
Genetic Changes in Relapse Biopsy Samples
The pattern of a strong initial response, long dormancy, and rapid late progression led us to hypothesize that relapse in these patients resulted from immune-mediated clonal selection and tumor outgrowth.24 To identify mutations that might confer immune resistance, we extracted DNA from bulk-tumor biopsy samples or early-passage primary cell lines (Table S2 in the Supplementary Appendix) and performed whole-exome sequencing to compare baseline and matched relapsed tissues. We achieved a median coverage of 149×, and the percent of tumor cells (as compared with stromal cells) was more than 40% in all samples (Table S2 in the Supplementary Appendix). Nonsynonymous mutations for all samples are shown in Table S4 in the Supplementary Appendix.
JAK Mutations with Concurrent Loss of Heterozygosity at Relapse
We found strong evidence that the relapsed tumors were closely genetically related to their baseline counterparts, despite up to 2 years between biopsies. In the case of Patients 1 and 2, of 1173 and 240 nonsynonymous mutations, respectively, originally identified in the baseline sample, 92.5% and 95.8% were also seen in the resistant tumor (Figure 2AFIGURE 2Acquired JAK1 Loss-of-Function Mutation at Relapse, with Accompanying Loss of Heterozygosity., and Fig. S4 in the Supplementary Appendix). The relapsing tumors also contained the same chromosomal loss-of-heterozygosity events as the baseline tumors, and all differences were due to further loss in the relapse samples. In the relapse biopsy samples from both patients, we identified new homozygous loss-of-function mutations in the kinases associated with the interferon-receptor pathway, with a Q503* nonsense mutation in the gene encoding Janus kinase 1 (JAK1) in Patient 1 (Figure 2A and 2B) and a F547 splice-site mutation in the gene encoding Janus kinase 2 (JAK2) in Patient 2 (Fig. S4 in the Supplementary Appendix). RNA sequencing showed that the JAK2 splice-site mutation caused intron inclusion, producing an in-frame stop codon 10 bp after exon 12 (Fig. S5 in the Supplementary Appendix). Therefore, both mutations were upstream of the kinase domains and probably truncated the protein or caused nonsense-mediated decay. Neither mutation was seen at baseline in the exome sequencing reads, by Sanger sequencing, or by targeted amplicon resequencing (Fig. S6 in the Supplementary Appendix).
The JAK2 mutation was the only homozygous mutation (adjusted variant allele frequency, >85%) of 76 new nonsynonymous mutations in Patient 2, and the JAK1 mutation was 1 of only 3 homozygous mutations among 53 new mutations in Patient 1 (Table S5 in the Supplementary Appendix). To become homozygous, both JAK mutations were acquired in the context of a copy-number–neutral nondisjunction event, resulting in loss of the wild-type chromosome and duplication of the mutated allele. This is seen clearly in Patient 1: at relapse, chromosome 1p (containingJAK1) showed a decrease in minor-allele frequencies for germline single-nucleotide polymorphisms relative to baseline (Fig. S7 in the Supplementary Appendix), was missing 36 heterozygous baseline mutations (presumably on the lost allele), and contained 20 mutations (presumably on the amplified allele) that became homozygous (adjusted variant allele frequency, >85%, with change of >35 percentage points from baseline). A similar loss-of-heterozygosity event occurred for chromosome 9 in Patient 2 (Fig. S8 and Table S5 in the Supplementary Appendix). Together, these data suggest that the tumors resistant to anti–PD-1 are a relatively homogenous population derived directly from the baseline tumor and that acquisition of the JAK mutations was an early founder event before clonal selection and relapse despite the fact that the mutation was not detected in pretreatment tumor tissue.
Functional Effects of JAK2 Mutation
To assess the functional consequences of the observed JAK mutations, we focused on the JAK2mutation from Patient 2 using two cell lines established at baseline (M420, wild-type JAK2) and at the time of relapse (M464, JAK2 F547 splice-site mutation). Whole-exome sequencing confirmed that the original bulk tumor was well represented by M464 (Fig. S9 in the Supplementary Appendix). Western blot analysis showed that the baseline cell line responded to interferon alfa, beta, and gamma with the expected signal transduction, including an increase in signal transducer and activator of transcription 1 (STAT1) and interferon regulatory factor (IRF) expression, STAT1 phosphorylation (pSTAT1), and the production of downstream interferon targets such as PD-L1, transporter associated with antigen processing 1 (TAP1), and major histocompatibility complex (MHC) class I (Figure 3AFIGURE 3Loss of Interferon Gamma–Induced Signaling and Gene-Expression Changes through AcquiredJAK2 Mutation.). However, the cell line from the progressing lesion showed a total loss of JAK2 protein (Figure 3A), resulting in a lack of response to interferon gamma, without change in sensitivity to interferon alfa or beta. This was true of the pSTAT1 response (Figure 3A) and the expression of PD-L1 and MHC class I molecules (Figure 3A and 3B). The progressing cell line also failed to up-regulate a wider panel of interferon-induced transcripts involved in antigen presentation and T-cell chemotaxis (Figure 3C, and Table S6 in the Supplementary Appendix). Together, these data indicate a total loss of functional response to interferon gamma and are consistent with JAK2 being required for signaling through the interferon-γ receptor, as opposed to the interferon-α/β receptor, which uses TYK2 and JAK1.27-29
Loss of Interferon Gamma–Induced Growth Arrest through Acquired JAK Mutations
We hypothesized that inactivating JAK mutations may result in a functional advantage for the progressive tumors because the lack of interferon signaling either decreased antigen presentation or allowed escape from interferon-induced inhibition of growth. In addition to using M420 and M464, we engineered the human melanoma cell line M407 by means of the CRISPR (clustered regularly interspaced short palindromic repeats)–Cas9 approach to create sublines without expression ofJAK1 or JAK2 (Figs. S10 and S11 in the Supplementary Appendix). These created truncating mutations analogous to those from Patients 1 and 2, and M407 is positive for HLA-A*02:01 and expresses the cancer–testis antigen NY-ESO-1, which allowed us to model T-cell recognition using T cells genetically modified to express an NY-ESO-1–specific T-cell receptor.23 M407 and bothJAK-loss sublines were equally recognized by NY-ESO-1–specific T cells, leading to high levels of interferon-γ production (Figure 4AFIGURE 4Loss of Interferon Gamma–Induced Growth Arrest through Acquired JAKMutations.).
When cultured in recombinant interferon alfa, beta, or gamma, the M420 and M407 parental cell lines showed interferon-induced growth inhibition in a dose-dependent manner (Fig. S12 in the Supplementary Appendix). However, both the JAK2-deficient M464 cell line (from Patient 2 at relapse) and the M407 JAK2-knockout subline were insensitive specifically to interferon gamma–induced growth arrest, yet remained sensitive to type I interferons alfa and beta; in contrast, the M407 JAK1-mutated subline was resistant to all three interferons (Figure 4B). This is again consistent with the specific association of JAK2 with the interferon-γ receptor and the common use of JAK1 by all three interferon receptors.27-29 As an orthogonal test of these effects, we treated our cell lines with 2′3′-cGAMP (cyclic guanosine monophosphate–adenosine monophosphate); this dinucleotide, which is produced in response to cytosolic double-stranded DNA, directly activates the stimulator of interferon genes (STING) and leads to interferon-β production through activation of interferon regulatory factor 3 (IRF-3).30 After 2′3′-cGAMP treatment, we observed growth arrest in all cell lines independent of JAK2 status but no effect in the JAK1-knockout subline (Figure 4C). Therefore, theJAK1 and JAK2 loss-of-function mutations did not decrease in vitro T-cell recognition but selectively blocked the interferon-γ signaling that leads to cell-growth inhibition, which for JAK2loss could be corrected by type I pathway activation or a STING agonist.
Functional Effects of Mutation in the Gene Encoding Beta-2-Microglobulin (B2M)
In Patient 3, whole-exome sequencing of the baseline and progressive lesions showed a 4-bp S14 frame-shift deletion in exon 1 of the beta-2-microglobulin component of MHC class I as 1 of only 24 new relapse-specific mutations and the only such mutation that was homozygous (Fig. S13A and S13B in the Supplementary Appendix). Immunohistochemical analysis for MHC class I heavy chains revealed loss of outer-membrane localization as compared with adjacent stroma or the baseline tumor, even though diffuse intracellular staining indicated continued production of MHC class I molecules (Fig. S14 in the Supplementary Appendix). This finding is in line with the role of beta-2-microglobulin in proper MHC class I folding and transport to the cell surface, and its deficiency has long been recognized as a genetic mechanism of acquired resistance to immunotherapy.12-14 Both the baseline and relapse biopsy samples were negative for MHC class II expression (Fig. S14 in the Supplementary Appendix), which suggests a lack of compensatory MHC up-regulation.
We could not find defined genetic alterations in Patient 4 that had clear potential to result in acquired resistance to T cells, but cancer cells in the baseline and relapse biopsy samples did not express PD-L1 despite proximity to T cells and PD-L1–expressing stroma (Fig. S3D in theSupplementary Appendix). These findings suggest possible nongenetic mechanisms of altered expression of interferon-inducible genes.16