Immune checkpoint blockade (ICB) has improved the survival of patients with a number of advanced malignancies. Antibodies targeting PD1 and its ligand PD-L1 are now approved for multiple cancer types, including non-small cell lung cancer (NSCLC) and melanoma.1–3 However, clinical responses to these therapies vary across tumor types and the majority of cancers treated are unresponsive or develop resistance over time. As a result, we need to develop better predictive biomarkers of response and further understand potential mechanisms of primary and adaptive resistance in order to develop novel therapeutic strategies that might improve both the rate and duration of response.

Several biomarker approaches have emerged to improve patient selection for immunotherapeutic agents. These include the characterization of genetic determinants of response, such as the use of somatic tumor mutation burden (TMB)4 and the presence of microsatellite instability,5 along with the use of markers indicative of an inflamed tumor microenvironment, such as assessment of the tumor-immune infiltrate,6–8 and the use of gene expression signatures of activated T cells.9–11 Recently, a gene expression signature including genes that identify T cells, T-cell activation, interferon (IFN)-responsive genes, chemokine expression and adaptive immune resistance, termed the T-cell-inflamed resistance gene expression profile (GEP), was shown to predict response to pembrolizumab in melanoma, gastric, and head and neck squamous cell carcinomas.9 Our group demonstrated that the genes used in this signature were also predictive of response and survival to checkpoint blockade in patients with metastatic lung cancer.12

However, even in the setting of an inflamed tumor microenvironment, a critical step in neoantigen presentation and cytolytic T-cell responses is the interaction between CD8+ T cells and tumor HLA Class I molecules, which present intracellular peptides on the cell surface for recognition by the T-cell receptor (TCR).13–15 Defects in the expression of HLA class I or proteins involved in the antigen processing machinery (APM) result in reduced antigen presentation and thus facilitate immune evasion. For example, in lung cancer, Perea et al showed that reduced expression of HLA Class I was associated with lower levels of CD8 T-cell infiltration.16

Recent work has highlighted the role of genetic losses of HLA Class I APM molecules in mediating response to ICB. Loss of β2-microglobulin (β2m) in melanoma and lung tumors due to mutations has been found to be associated with the development of resistance to ICB.17–19 A recent report by Chowell et al, examining over 1500 patients with advanced malignancies, showed that genetic loss of heterozygosity in the HLA Class I alleles was associated with resistance to checkpoint inhibitors.20 However, most tumors resistant to checkpoint blockade do not harbor readily identifiable genetic causes of resistance,21–23 suggesting epigenetic mechanisms involving HLA Class I molecules or APM may mediate primary resistance.

Defective expression of HLA Class I and APM proteins are common, with a reported incidence of 73% to 90% depending on the tumor type (with most of the changes being due to down regulation of transcription).24–26 Given the essential role of antigen presentation in a successful anti-tumor immune response, we hypothesized that defects in this pathway would be associated with poor response to ICB. To test this hypothesis, we analyzed transcriptional profiles from advanced NSCLC and melanoma patients treated with checkpoint inhibitors. Our results show a strong correlation between expression of genes involved in the APM process with response and survival outcomes.

Antigen presentation is accomplished through a series of coordinated intracellular events involving multiple proteins, the so-called APM proteins.35 36 First, tumor antigens are degraded by the proteasome and the immunoproteosome (specifically, PSMB8 (LMP7), PSMB9 (LMP2), and PSMB10 (LMP10)). These peptides are then transported via the transporter associated with antigen processing (TAP), which is composed of the two non-covalently associated subunits, TAP1 and TAP2 into the endoplasmic reticulum (ER). In the ER, HLA Class I heavy chain and β2m assembly occurs, with the ER chaperones calnexin, calreticulin, and the thiol oxidoreductase ERp57 (PDIA3) allowing for the correct folding of the HLA Class I β2m complex, which then associates with tapasin and subsequently with TAP-peptide complexes. This allows peptide loading onto the HLA Class I β2m molecules, after which, the antigenic peptides may be trimmed by an ER-associated aminopeptidase (ERAP1).29 The HLA Class I β2m-peptide complex then travels to the plasma membrane via the Golgi apparatus. Other genes in our signature that affect the APM machinery include PSME 1, 2, and 3 (20s-26S proteasome proteins that play a role in antigen presentation),30 three heat shock 90 proteins, HSP90AB1, HSP90AAA1, and HSP90B1,31 and the transcription factors NLRC5, RFX5, and CIITA.32 37 We thus used this set of genes in our initial APMS (online supplemental table 1). We did not include the genes for the HLA Class I molecules in our signature as these can be highly polymorphic and subject to post-translational modifications that affect expression levels.38 For example, in a mouse fibrosarcoma cell model with widely different levels of HLA Class I expression measured by flow cytometry, there was no correlation with HLA Class I messenger RNA expression.39

We found that patients with a higher APM score were more likely to respond to ICB and had improved PFS and OS in lung cancer, and the APM score was also associated with improved survival in two independent melanoma cohorts treated with checkpoint blockade. These results are novel, but consistent with a growing body of literature highlighting the importance of MHC Class I and β2m gene mutations in mediating tumor intrinsic resistance to ICB. Loss of heterozygosity at the β2m locus has been associated with primary resistance and poor OS in melanoma patients receiving checkpoint blockade,18 and defects in β2m leading to loss of MHC Class I expression have been associated with acquired resistance to immunotherapy in melanoma.17 40

HLA loss can be classified into two main categories: reversible regulatory or irreversible structural defects. Structural defects are due to mutational events or chromosomal losses (in chromosomes 6p21.3 and 15q21) that disrupt HLA-I heavy chains or β2m.41 Clearly, these patients will likely have poor responses to checkpoint blockade or any type of T-cell-based immunotherapy. However, the frequency of somatic mutations in APM components in epithelial malignancies is reportedly low (less than 10%)24 and, specifically, are relatively rare in lung cancer (2.7% to 6.9%).42 In contrast, reversible defects due to epigenetic, transcriptional, or post-transcription events are common; loss of HLA class I protein occurs in between 30% to 70% of cancer cases.24 A recent study by Pereira et al found somatic mutations in antigen processing components in only 5% of lung cancers, while loss or low level expression of MHC Class I protein was appreciated in 66% of cases.22 This dissociation between genetic defects and protein levels is likely due to at least two factors. First, since many of the defects in HLA Class I and APM proteins are due to epigenetic changes, they are not seen in genetic analyzes.43 Second, it appears that unless the entire antigen presentation machinery is intact, optimal stability and surface expression of MHC Class I molecules is not achieved.30

It is likely that the status of HLA Class I expression and the APM components is tightly linked to the inflammation status of the tumor. Initial T-cell activation requires the ability of tumor cells to present antigen to the T cells. Indeed, loss of HLA Class I expression has been associated with a decrease or absence of CD8 T-cell infiltration within tumors.22 Activation of T cells would result in the subsequent release of IFN-γ and tumor necrosis factor-α, resulting in tumor microenvironmental activation and recruitment of additional T cells. In addition, IFN-γ signaling would be expected to upregulate proteins on tumor cells involved in antigen processing, including MHC Class I and APM molecules,44 45 leading to a positive feedback loop resulting in additional T-cell recruitment and activation. It is thus expected that tumors with higher APM scores would have higher T-cell-inflamed scores, as we observed. It is interesting that the APM score was actually more predictive of response than an inflammatory gene score in our NSCLC cohort, supporting the importance of antigen presentation in response to ICB therapy. Eventually, it has been proposed that the success of CD8 T-cell attack on tumors can result in selection of tumor cells lacking MHC Class I expression making them resistant to further T-cell killing.43 46 This form of tumor editing likely explains the high incidence of loss of MHC Class I and APM molecules on lung cancer and other tumors.22

We believe there are a number of important potential clinical implications of our study. First, we demonstrate the overall expression levels of APM genes (and the presumed associated decrease in HLA Class I expression) can predict which patients will respond to anti-PD1 therapy with more accuracy than previously published inflammation scores. A very low APM score was particularly powerful in identifying those patients who did not respond. Thus, in our optimized lung cancer data set, an APM score of less than or equal to −3.0 (n=11) was 100% predictive of progressive disease. Although using this cut-off value results in relatively low sensitivity, being able to predict lack of response with very high accuracy would be of great value in sparing a set of patients’ expense, time, and the possibility of immune-related side effects for no benefit. Low scores in the Hugo melanoma data set were also highly predictive of non-response. Indeed, a recent study by Kalaora et al demonstrated that lower expression of immunoproteosome subunits (PSMB8 and PSMB9) were associated with inferior response rates and outcomes in melanoma patients receiving ICB.47

Second, identifying patients with low Class I expression could have important implications in directing more efficacious alternative therapies.15 43 In terms of other immunotherapies, loss of Class I expression would make vaccines and adoptive transfer of tumor infiltrating lymphocytes or T cells expressing transgenic TCRs less likely to be successful, but would not affect sensitivity to chimeric antigen receptor (CAR) T-cell therapy. Loss of Class I expression could actually make the tumors more sensitive to adoptive transfer of NK cells or NK cell-CARs. As mentioned above, most HLA Class I losses are not due to genetic mutations, but secondary to epigenetic or environmental factors. It may thus be possible to alter the tumor cells to re-express Class I molecules, making treatment with ICB or other immunotherapies more effective.23 43 This could be achieved using immunostimulatory agents that induce IFNs (such as cytokines, toll-like receptor agonists, or viral therapies), resulting in increased MHC I expression in tumors. Alternatively, a number of drugs already used in cancer therapy have also been shown to restore MHC I expression. These include mitogen-activated protein kinase inhibitors,48 DNA demethylating agents (eg, 5-azacytidine and decitabine),49 and inhibitors of histone deacetylation (HDAC inhibitors) such as trichostatin A or entanostat.50 One could thus envision screening patients for loss of APM genes or HLA Class I expression and in those with defective expression, initiating a treatment regimen where Class I expression was pharmacologically upregulated. This could then be followed by treatment with ICB, vaccines, or adoptive T-cell transfer.

Our study does have certain limitations. One possible limitation of the genomic signature approach for antigen presentation is that, by nature, the analysis interrogates the entire tumor, including tumor cells and stroma. Although this is of value in understanding the inflammation status and which cell types infiltrate into the tumor, HLA Class I expression is often quite high in stromal tissues, so variable amounts of fibroblasts or leukocytes could dilute signals that are intrinsic to tumor cells (such as Class I or APM expression). The only way to know definitively from which cells our signals came from, would be single cell RNAseq. This is an extremely expensive approach that requires fresh tissue and may not have coverage deep enough for a number of our genes. We do have some data to suggest that our APMS is not simply another surrogate for certain type of leukocyte infiltration. First, in our previous study using these same samples12 (Supplemental figures 2 and 4), we saw no associations of response to anti-PD1 therapy with lymphocyte genes (CD4, CD8A, and CD8B), with macrophage genes (CD68, CD14, CD163, and CSF1R), or to fibroblast genes (ACTA2 and Col4A1). We saw no correlation between the APM score and expression levels of the genes for CD45, CD8A, CD11b, CD68, CD14, and CD163 (online supplemental figure 2). However, despite this potential dilutional effect, we were still able to detect a strong signal. Going forward, we postulate that targeted immunohistochemistry, that would determine expression levels of key surface proteins on only the tumor cells, will be even more specific and valuable. Such studies, staining for HLA Class I expression and key components of the APM are underway. This was a single-center, retrospective study conducted in chemotherapy-refractory patients with metastatic NSCLC. Further study is required to validate our findings in larger data sets. In addition, as checkpoint inhibitors are now approved in the front-line setting in patients with metastatic NSCLC, it will be important to evaluate our APMS in patients receiving checkpoint blockade, both alone and in combination with chemotherapy, in the first-line setting. These studies are currently ongoing. The other major limitations to our study are that we did not consistently have PD-L1 immunohistochemical staining or measurements of TMB; our patients were treated several years ago following progression on standard of care chemotherapy when neither of these tests were routinely performed. Nevertheless, our data demonstrate that defects in antigen presentation may be an important feature in predicting outcomes to ICB in both lung cancer and melanoma and, if substantiated, has the potential to inform therapeutic strategies in this patient population.