Expression Profiles and Prognostic Value of Multiple Inhibitory Checkpoints in Head and Neck Lymphoepithelioma-Like Carcinoma

Background Inhibitory checkpoints are promising antitumor targets and predictive biomarkers in a variety of cancers. We aimed to identify the expression levels and prognostic value of multiple inhibitory checkpoints supported by preclinical and clinical evidence in head and neck lymphoepithelioma-like carcinoma (HNLELC). Methods The expression of seven inhibitory checkpoints were evaluated in the tumor nest (TN) and tumor stroma (TS) of 102 HNLELC specimens using immunohistochemistry and digital pathology, and an inhibitory checkpoint-based signature (ICS) was subsequently constructed using the LASSO Cox regression model. Results PD-L1, B7H3, and IDO-1 were mostly expressed in the TN, with median H-score of TN vs TS: 63.6 vs 14.6; 8.1 vs 1.0; 61.5 vs 34.7 (all P < 0.001), whereas PD-1, TIM-3, LAG-3, and VISTA were mainly observed in the TS, with median H-score of TN vs TS: 0.2 vs 12.4, 3.4 vs 7.1, 6.2 vs 11.9, 16.4 vs 47.2 (all P < 0.001), respectively. The most common simultaneously expressed combinations consisted of PD-L1 + B7H3 + IDO-1 + TIM-3 + LAG-3 + VISTA and B7H3 + IDO-1 + TIM-3 + LAG-3 in the TN (both occurring in 8.8% of patients) and PD-L1 + B7H3 + IDO-1 in the TS (4.9%). In addition, high-ICS patients had shorter 5-year disease-free (40.6% vs 81.7%; P < 0.001), regional recurrence-free (63.5% vs 88.2%; P = 0.003), and overall survival (73.5% vs 92.9%; P = 0.006) than low-ICS patients. Multivariate analysis revealed that ICS represented an independent predictor, which could significantly complement the predictive performance of TNM stage for 3-year (AUC 0.724 vs 0.619, P = 0.014), 5-year (AUC 0.727 vs 0.640, P = 0.056), and 10-year disease-free survival (AUC 0.815 vs 0.709, P = 0.023). Conclusions The expression of inhibitory checkpoints and ICS classifier may increase the prognostic value of the TNM staging system and guide the rational design of personalized inhibitory checkpoint blockade therapy in HNLELC.


INTRODUCTION
Head and neck lymphoepithelioma-like carcinoma (HNLELC) is a rare malignant neoplasm histologically identical to nonkeratinizing undifferentiated nasopharyngeal carcinoma (NPC), but with origins other than the nasopharynx (1). Generally accepted treatment modalities for HNLELC consist of surgery and/or radiotherapy with or without adjuvant chemotherapy (2)(3)(4). Despite the favorable prognosis associated with HNLELC, over 30% of cases continue to experience recurrence, distant metastasis, and even death (2,3,5). Therefore, there is a need to identify valuable biomarkers that can provide therapeutic insight to further improve the outcome of HNLELC patients.
Emerging studies have demonstrated that inhibitory checkpoints expressed on the surface of tumor cells and/or immune cells negatively regulate the function of effector immune cells, which is an important mechanism of tumor immune escape (6). Antagonist antibodies that block inhibitory checkpoints have been devised to reverse immune resistance, representing the current frontier of cancer immunotherapy (7). To date, various inhibitory checkpoint blockades (e.g., anti-PD-1/PD-L1 and anti-LAG-3 antibodies) have been approved for the clinical application or subjected to preclinical/clinical studies (8). Of note, the results of several encouraging phase 2 trials and practice-changing randomized phase 3 trials exhibit unprecedented clinical efficacy of anti-PD-1 therapy in recurrent/metastatic NPC and other head and neck squamous cell carcinoma (HNSCC) (9,10). HNLELC is characterized by substantial lymphocyte infiltration within the tumor microenvironment (2). Despite the presence of such infiltrates, the tumor remains capable of progressive growth, implying that the antitumor immunity of these lymphocytes may be suppressed. Therefore, inhibitory checkpoint blockades that can revitalize anergic T lymphocytes and alleviate immune suppression may represent an effective treatment strategy for HNLELC.
The expression status of inhibitory checkpoints is critical to guide inhibitory checkpoint blockade therapy. For example, high PD-L1 expression is significantly associated with a favorable clinical response to a PD-1/PD-L1 blockade in certain types of tumors, including HNSCC (10). In addition, a growing number of studies have demonstrated that inhibitory checkpoints present in the tumor microenvironment can predict prognosis in cancer patients (11,12), acting as a complement to the anatomic-based TNM staging system (12). Thus, in the era of immunotherapy, it is essential to clarify the expression levels of inhibitory checkpoints to provide implications for immunotherapies targeting these checkpoints and identify those with prognostic value to improve the current TNM system ability of prognostic prediction in HNLELC.
Based on these considerations, we aimed to delineate the immunohistochemical expression of seven common inhibitory checkpoints in 102 HNLELC patients and develop an inhibitory checkpoint-based signature (ICS) to effectively predict the prognosis of HNLELC patients.

Study Specimens
This study retrospectively recruited paraffin-embedded tumor specimens from 102 HNLELC patients treated at Sun Yat-sen University Cancer Center (n = 78), the First People's Hospital of Foshan (n = 21), and the Second People's Hospital of Foshan (n=3) from 2001 to 2019. Inclusion criteria consisted of patients without distant metastasis and antitumor treatment prior to biopsy or surgery. Patients with incomplete clinicopathological characteristics and follow-up data were excluded. Tumor stages were reclassified according to the 8th edition of the American Joint Committee on Cancer (AJCC) TNM staging system (13). The included clinicopathological variables consisted of age, sex, smoking history, drinking history, pretreatment neutrophil-tolymphocyte ratio (NLR), Epstein-Barr virus-encoded RNA (EBER), T stage, N stage, and TNM stage. This study was approved by the institutional research ethical committees of all included institutions (B2020-375).

Immunohistochemistry (IHC)
Slides stained with hematoxylin and eosin (H&E) were reviewed by two experienced pathologists (Dr Feng and Dr Liu) to reconfirm the pathological diagnosis and to select tumor tissue blocks containing tumor nest (TN) and tumor stroma (TS) for IHC staining as described previously (14).

Digital Image Analysis and Selection of Cut-Off Values
IHC-stained slides were scanned at a high power (200×) (0.496 mm/pixel) using a Vectra 2.0.8 multispectral microscope system (Perkin Elmer, Waltham, Massachusetts, USA), and were subsequently quantified using inForm 2.1.1 software (Perkin Elmer). The flow of analysis was performed as follows ( Figure 1A): 1) image preparation: the images were loaded and prepared for segmentation by converting them to optical density (OD); 2) segmenting tissue: tissue regions of interest were manually annotated as TN and TS; 3) segmenting cells: all individual cells and their associated cytoplasm within the segmented tissue categories were identified based on the nuclear signal and the distance to the nucleus using the cell segmentation tool; 4) scoring IHC: positive cells within each region were separately quantified as H-scores, which were calculated by the staining intensity (0, absent; 1+, weak; 2+, moderate; and 3+, strong) multiplied by the percentage (0 − 100) of positive cells at each intensity, which ranged from 0 to 300 (15). The parameters for cell segmentation and the thresholds of the different staining intensities were collectively confirmed by Dr Feng and Dr Liu who were blinded to the clinicopathological characteristics and outcomes of the patients. For each marker per patient, five representative fields were assessed and the average H-score was used as the final score.
The optimum cut-offs for the H-scores of 14 checkpoint features (seven checkpoints in the TN or TS) were selected based on their associations with disease-free survival (DFS) using X-tile software (version 3.6.1; Yale University, New Haven, CT, USA) (16). The patient distribution of each inhibitory checkpoint expression and the cut-off value of high and low expression are shown in Supplementary Figure S1 and Supplementary Table S1. X-tile software is a popular bioinformatics tool for optimizing survival-based cutoff value by log-rank chi-square test. The optimal cut-off value was determined by the minimum P value with the highest c 2 value, and a corrected P value was calculated by Miller-Siegmund model (17).

In Situ Hybridization (ISH) of EBER
EBV infection was assessed using an EBER ISH Kit (ZSGB-BIO, ISH-7001, Beijing, China) according to the manufacturer's instructions. Known EBER-positive and EBER-negative sections were applied as positive and negative controls.

ICS Construction
Due to the relatively small number of events compared with variables to be filtered, the least absolute shrinkage and selection operator (LASSO) Cox regression model was adopted to select the checkpoint features with a strong prognostic value and low correlation to prevent overfitting (18). The optimal value for l was identified via 10-fold cross validation with 1-SE (standard error) criteria. Next, we constructed a multi-checkpoint featurebased classifier to predict the outcome of HNLELC patients. The ICS score was calculated as the sum of H-scores of the selected features as weighted by their regression coefficients. The LASSO Cox regression analysis was conducted using the R package "glmnet" (19).

Statistical Analysis
Our primary endpoint was DFS, as defined as the time from the first day of treatment to any failures (local, reginal, or distant) or death from any causes. Secondary endpoints included regional recurrence-free survival (RRFS), time to first draining lymph node recurrence; and overall survival (OS), time to death.
Two groups were compared using a c 2 test or Fisher's exact test for categorical variables and a Mann-Whitney U test and Wilcoxon signed rank test for nonpaired and paired continuous variables. Upset plot (https://www.omicstudio.cn/tool/43) was used to assess the co-expression of different inhibitory checkpoints (20,21). A Pearson's correlation test was used to calculate the correlation coefficients. The Kaplan-Meier method with a log-rank test was used to estimate DFS, RRFS, and OS. A univariate Cox regression analysis was used to generate the hazard ratio (HR) and 95% confidence interval (CI). Variables found to be significant in the univariate analysis (P ≤ 0.20) or of clinical significance were included in the multivariate analysis with the backward stepwise method to identify independent prognostic variables. Moreover, we established a combined predictive model using a multivariate Cox regression model. The b coefficient of each variable was divided by that of TNM stage and then rounded to an integer value to generated the risk score (12,22). Next, we compared the predictive performance of the different models using a receiver operating characteristic (ROC) curve and time-dependent area of under the curve (AUC) analysis.
All statistical analyses with a two-sided test were conducted using IBM SPSS software (version 25; Property of IBM Corp., Armonk, NY, USA) and R software (version 4.0.2). A threshold of P < 0.05 was considered significant. Table 1 summarizes the baseline characteristics of 102 HNLELC patients from three independent institutions. Representative IHC images and the paired H-scores of seven inhibitory checkpoints in the TN and TS are presented in Figure 1B Due to a lack of a standard cut-off value for defining the expression of these checkpoints as either positive or negative in HNLELC, we described the distribution of patients using five Hscore cut-offs (> 1, > 5, > 30, > 55, and > 100), which have been frequently used for the evaluation of PD-L1 (23)(24)(25)(26)(27). The median H-score of seven inhibitory checkpoints in the TN and TS was determined. As shown in Table 2, we observed high expression of PD-L1, IDO-1 and VISTA in both the TN and TS and PD-1 and LAG-3 in the TS, with a median H-score greater than 10. In contrast, the expression of B7-H3 in the TS and PD-1 in the TN were extremely low, with a median H-score less than or equal to 1.

Correlation Among Inhibitory Checkpoints
Based to the cut-off values of inhibitory checkpoints generated by X-tile software, patients were classified into high and low expression groups (Supplementary Table S1). Subsequently, we used Upset plots to assess the co-expression of the seven inhibitory checkpoints with the status of high expression in the TN and TS, respectively ( Figure 2). In the TN, the most common co-expression combinations were PD-L1 + B7H3 + IDO-1 + TIM-3 + LAG-3 + VISTA and B7H3 + IDO-1 + TIM-3 + LAG-3, which were both observed in 8.8% of patients (9 of 102). In the TS, the most common co-expression combination was PD-L1 + B7H3 + IDO-1, occurring in 4.9% of patients (5 of 102). Furthermore, in the bivariate correlation analysis for the expression of seven inhibitory checkpoints, the expression of most immune markers was significantly and positively correlated with each other, regardless in the TN or TS (r = 0.20 to 0.90; P = 0.042 to < 0.001). Strikingly, a significantly negative correlation was observed between B7H3 TS and PD-1 TS (r = -0.24; P = 0.017) (Supplementary Figure S2). Considering the efficacy of PD-1 and PD-L1 blockades for the treatment of a wide variety of cancers and the association between PD-L1 expression and the clinical response, we compared the expression status of other checkpoints in patients with low and high PD-L1 expression. Overall, patients with high PD-L1 TN expression also displayed higher levels of expression of six other inhibitory checkpoints compared with those with low PD-L1 TN expression, though the correlation with PD-1 and B7H3 expression did not reach statistical significance (Supplementary Figure S3).

Association Between Inhibitory Checkpoint Expression and Clinicopathological Features
To explore the association between inhibitory checkpoint expression and patient characteristics, we analyzed differences in the expression of seven inhibitory checkpoints in patients with different clinical risk factors. As shown in Supplementary Table  S2, several checkpoints were significantly associated with smoking history, EBV infection, T stage, and N stage, but not

Development of a Combined Predictive Model
The purely anatomic-based TNM stage provides limited predictive performance for cancer prognosis, implying that it needs additional prognostic factors to improve its prognostic ability. We did a multivariate Cox regression analysis to develop a combined model to predict DFS risk based on its independent predictors: ICS, TNM stage, and NLR ( Figure 3F). The risk score of each independent predictor is shown in Supplementary Table S4. ROC analysis found that the AUCs of ICS alone (AUC 0.687 vs 0.619, P = 0.340) and NLR alone (AUC 0.583 vs 0.619, P = 0.585) were equal to TNM stage alone (Supplementary Figure S8).
To develop a simple model with a good predictive ability, we established three models: model A, TNM stage combining with ICS and NLR; model B, TNM stage combining with ICS; model C, TNM stage combining with NLR. Time-dependent ROC analysis was used to compare the predictive sensitivity and specificity of three models for 3-year, 5-year, and 10-year DFS. Among the three models, these was no significant difference in predictive ability between model A and model B

DISCUSSION
Inhibitory checkpoint blockades have been proven to be effective antitumor strategies in multiple types of cancer (9,10,29). In this study, we investigated the abundance and coexpression patterns of seven inhibitory checkpoints as well as clinicopathological factors that may affect the expression of inhibitory checkpoints, which provided a basis for inhibitory checkpoint blockade therapy in HNLELC patients. In order to accurately identify patients with high risk who deserve intensified treatment, such as addition of immunotherapy, we developed an eight-feature-based ICS prognostic classifier that substantially improved the predictive ability for DFS when combined with the TNM stage.
Over the past few decades, inhibitory checkpoint blockades, represented by anti-PD-1/PD-L1 antibodies, have achieved unparalleled clinical success in multiple cancer types (30). In addition to PD-1/PD-L1, substantial evidence suggests that a variety of other inhibitory checkpoint molecules might exist in the tumor microenvironment [e.g., B7H3 (31), IDO-1 (32), LAG-3 (33), TIM-3 (34), and VISTA (35)], which also participate in the regulation of tumor immune escape. Thus, we investigated the expression of seven inhibitory checkpoints supported by preclinical and clinical evidence, which might provide a fundamental basis for future individualized immunotherapy targeting these checkpoints in HNLELC. In particular, abundant PD-L1 expression was observed in HNLELC patients. Given that response to a PD-1 blockade is more frequently reported in patients with high PD-L1 expression (36), a PD-1 blockade may be a potential candidate for the treatment of HNLELC.
Unfortunately, anti-PD-1/PD-L1 monotherapy has had a benefit on the long-term survival for only a fraction of cancer patients. One possible explanation is that the simultaneous existence of other inhibitory checkpoints might aggravate lymphocyte exhaustion and dysfunction, blocking the PD-1/PD-L1 signaling pathway alone is insufficient to restore immune function. Recent studies have demonstrated that combination therapies targeting two or more inhibitory checkpoints (e.g., PD-1 and LAG-3) (33), can achieve synergistic antitumor efficacy through nonredundant pathways to activate T cells. Therefore, understanding the co-expression and correlation of various inhibitory checkpoints should be highlighted to design more effective combination immunotherapy. The findings of this study indicate that the most common simultaneous expression patterns consisted of PD-L1 + B7H3 + IDO-1 + TIM-3 + LAG-3 + VISTA and B7H3 + IDO-1 + TIM-3 + LAG-3 in the TN and PD-L1 + B7H3 + IDO-1 in the TS. Moreover, PD-L1 expression was positively correlated with the expression of several other inhibitory checkpoints, suggesting the potential simultaneous involvement of multiple inhibitory checkpoint pathways in the HNLELC immune evasion mechanisms. Thus, the combination of several inhibitory checkpoint blockades (e.g., anti-PD-1 or anti-PD-L1 with other inhibitory checkpoint blockades) may improve clinical efficacy of HNLELC patients.
An accurate prognostic evaluation is critical for the selection of appropriate treatment. The TNM staging system is currently the most important benchmark for defining prognosis and to determine the treatment strategies for HNLELC (13). However, for many cancer patients, the purely anatomic-based staging provides only partial insight, which is useful but insufficient. Recent studies have revealed that the expression of inhibitory checkpoints is not only associated with immunotherapeutic benefits (37), but can also predict the prognosis of cancer patients (12). Integrating multiple biomarkers into a single model would increase the prognostic value (28). The present study constructed an ICS classifier based on the eight features containing critical prognostic information in the LASSO model, which provided a prognostic prediction independent of TNM stage. Furthermore, compared with other clinicopathological variables, the ICS classifier was a better variable that significantly complemented the prognostic performance of TNM stage. These results indicate the importance of the host immune status for predicting the prognosis of HNLELC patients, which may function as a complement to the TNM staging system.
Multiple studies have suggested that chronic infection and cancer often induce T cell exhaustion and the upregulation of multiple inhibitory checkpoints (38,39). HNLELC has been reported to be tightly associated with EBV in regions with a high incidence (3,40). In the present study, 95 of 102 cases were EBER-positive, and these patients displayed high LAG-3 expression, in line with the findings in a previous study of gastric cancer (41). Furthermore, we observed high PD-L1 TS expression in T3-T4 stage patients and low VISTA TN expression in N2-N3 stage patients, suggesting that these markers are probably associated with tumor progression. It known that some lifestyle factors, such as tobacco smoking and alcohol drinking, are significant risk factors for head and neck cancer (42) and are closely associated with its prognosis (43). In order to explore whether smoking and drinking history affects immune microenvironment, we investigated the correlations of inhibitory checkpoint expression with smoking and drinking and found that patients with smoking history had significantly higher B7H3 TN and lower PD-1 TS , suggesting that smoking may affect the expression of B7H3 and PD-1, which was consistent with previous studies (44,45). Although the exact mechanism of how smoking affect the expression of B7H3 and PD-1 is unknown, studies have demonstrated that smoking is characterized by a high frequency of somatic mutation burdens in multiple cancers including head and neck squamous cell carcinomas (46), which may contribute to the improved response to immunotherapy (47,48). Thus, the mechanism of the association between smoking and inhibitory checkpoints deserves to be further explored. There were several limitations in this study that are important to note. First, since this was a retrospective study with a small sample size, external validation is required. However, given the low incidence of HNLELC, the present study represents a relatively large cohort. Second, previous studies have suggested that the load of EBV-DNA in the peripheral blood contains more precise information than the EBER status of the tumor tissue (49). However, due to the limited technological background, data pertaining to the EBV-DNA load in HNLELC was not available. Third, we did not investigate inhibitory checkpoint expression on different immune cell subpopulations and the colocalization of different inhibitory checkpoints using multiple immunohistochemistry staining. Fourth, we only focused on seven inhibitory checkpoints, other important immune checkpoints, such as stimulatory checkpoints (e.g., ICOS and OX40) (50,51), may have been neglected. Clearly, much remains to be done to gain a better understanding of the HNLELC immune microenvironment.
In conclusion, this study demonstrated the heterogeneous distribution of seven inhibitory checkpoints in the TN and TS and developed an associated ICS classifier in HNLELC, which added the prognostic value to the TNM staging system. Together, our findings facilitate a more accurate patient prognostic stratification and pave the way for the development of immunotherapy targeting these checkpoints in HNLELC.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article has been uploaded the Research Data Deposit platform (https://www. researchdata.org.cn) (RDDA2022568951).

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Sun Yat-sen University Cancer Center. Written informed consent to participate in this study was excused by the institutional review board.