We retrospectively collected 333 paraffin-embedded NPC specimens for this study. A total of 208 specimens obtained at the Sun Yat-sen University Cancer Center (Guangzhou, China) between January 2011 and December 2013 were designated as the training cohort, while 125 samples obtained at the Affiliated Hospital of Guilin Medical University (Guilin, China) between January 2010 and June 2014 were designated as the validation cohort. All patients from the Guangzhou cohort underwent intensity-modulated radiation therapy (IMRT), and all patients from the Guilin cohort underwent two-dimensional radiotherapy (2D-RT). No patients had received any antitumour therapy before biopsy sampling, and all of the patients were pathologically diagnosed with NPC. All patients were restaged according to the 8th AJCC TNM staging system [12]. This study was approved by the Institutional Ethical Review Boards of both hospitals, and written informed consent was obtained from each patient. This study is reported according to the REMARK criteria [13].

Based on previous studies [14–17], we selected 9 prognostic immune checkpoints for IHC staining: PD-L1, B7-H3, B7-H4, IDO-1, LAG-3, VISTA, TIM-3, ICOS and OX40. IHC was performed as previously described [18]. The following primary antibodies were used: anti-PD-L1 (clone E1L3N, 1:400 dilution; Cell Signaling Technology, CST, Beverly, Massachusetts), anti-B7-H3 (clone D9M2L, 1:400; CST), anti-B7-H4 (clone HPA054200, 1:800; Sigma-Aldrich, Ronkonkoma, NY, USA), anti-IDO-1 (clone D5J4E; 1:800; CST), anti-LAG3 (clone D2G40, 1:100; CST), anti-VISTA (clone D1L2G, 1:800; CST), anti-TIM3 (clone D5D5R, 1:400; CST), anti-ICOS (clone D1K2T, 1:1600; CST), and anti-OX40 (ab119904, 1:1600; Abcam, Cambridge, UK).

A full view of each IHC slide was digitally scanned using a ScanScope Aperio AT2 slide scanner (Leica Microsystems) at 400× magnification. All images were auto-examined using computational pathology analysis, and the expression was quantified as the percentage of tumour cells (TCs) or tumour-associated immune cells (TAICs) expressing the immune checkpoints. As the immune checkpoints PD-L1, B7-H3, B7-H4, and IDO-1 are expressed by both TCs and TAICs, these checkpoints were assessed in both compartments. In contrast, given their predominant expression in TAICs, LAG3, VISTA, TIM3, ICOS, and OX40 were assessed only in the tumour stroma compartment (Additional file 1: Figure S1). In total, there are 13 features.

In brief, the computational pathology analysis consisted of five stages: 1) manual annotation of the individual cell nuclei into TCsand TAICs by two pathologists; 2) stain deconvolution of the IHC staining from the haematoxylin counterstaining; 3) automated segmentation of the nuclei in the haematoxylin channel; 4) automated classification of the cells into TCs, TAICs and other cells using the Xception deep learning model [19]; and 5) quantification of the positive cell percentage for each immune checkpoint (Fig. 1a). Detailed descriptions of the computational pathology analysis are provided in the supplementary materials. Computational pathology analysis showed a high consistency with pathological classification, with an accuracy rate of 83.6% for TC identification and 87.9% for TAIC identification (Fig. 1b).Fig1Fig. 1

Computational pathology analysis. a Histology image analysis pipeline and validation; (b) Precision, recall and F1-score for each of the three cell classes. The scale bar represents 30 μm

We adopted a penalized Cox regression model to select the most useful prognostic features out of all 13 immune checkpoint features [20] and then constructed an ICS for predicting survival in the training cohort. The “glmnet” package was used to perform a least absolute shrinkage and selection operator (LASSO) Cox regression model analysis. Ten-time cross validations with the Lambda.min criteria were used to determine the optimal values of λ, and a value of λ = 0.038 with log (λ) = − 3.269 was chosen. Based on this value, IDO1_TAIC, VISTA _TAIC, B7-H3 _TAIC, ICOS _TAIC and LAG3 _TAIC were selected to construct the prediction model with the coefficients weighted by the penalized Cox model in the training cohort. We then used X-tile software (version 3.6.1; Yale University, New Haven, CT, USA) to generate the optimal cutoff values for the ICS scores based on the associations with patient overall survival (OS) [21]. The thresholds for the scores that were produced by the predictive model were used to separate patients into low-risk and high-risk groups.

The plasmatic EBV DNA concentrations were routinely measured before treatment using quantitative polymerase chain reaction, as described in the Additional file 1 [22]. A cut-off level of 2000 copies/mL was chosen to define low and high pretreatment EBV DNA levels [23, 24].

Our primary endpoint was OS, and secondary endpoints included disease-free survival (DFS) and distant metastasis-free survival (DMFS). We calculated OS from the first day of treatment to the date of death from any cause, DFS from the first day of treatment to the date of the first relapse at any site or death from any cause (whichever occurred first) and DMFS from the first day of treatment to the first distant relapse.

The associations between the ICS and clinicopathological variables were calculated using the χ2 test or Fisher’s exact test. Receiver operating characteristic (ROC) curve analysis was used to determine optimal cutoff values separating high and low expression for the 13 immune checkpoint features in the training cohort with respect to OS. The Kaplan-Meier method was used to estimate OS, DFS and DMFS, the log-rank test was used to compare differences, and hazard ratios (HRs) were calculated using univariate Cox regression analysis. Multivariate Cox regression analysis with backward selection was used to test the independent significance of different factors. Significant variables (P ≤ 0.1) were included in the multivariate analysis, and only independent prognostic factors were retained in the multivariate model. In addition, we established a prognostic score model combining the ICS and TNM stage [25, 26]. ROC curves were used to compare its prognostic validity with the TNM stage or ICS alone models. We also did subgroup analysis according to the pre-treatment plasma EBV-DNA levels.

All statistical tests were two-sided and considered significant when the p value was less than 0.05. Statistical analyses were performed using Statistical Package for the Social Sciences (SPSS) v22.0 (IBM, Armonk, NY, USA) and R software (R version 3.2.3; rms package, “rpart” package version 4.1–10, http://www.r-project.org/; “glmnet” package). The authenticity of this article has been validated by uploading the key raw data onto the Research Data Deposit public platform (http://www.researchdata.org.cn), with the approval RDD number as RDDB2019000556.

We collected 333 pretreatment, non-metastatic NPC specimens that were obtained at two academic institutions for this study. Additional file 2: Table S1 shows the clinicopathological characteristics of the patients in the training cohort (n = 208) or validation cohort (n = 125). All patients received radiotherapy, and 307 (92.2%) patients received platinum-based chemotherapy. The median follow-up time was 69.7 months (interquartile range (IQR) 65.1–72.8) for the patients in the training cohort and 58 months (IQR 41–69) for those in the validation cohort.

Representative images of immunohistochemical staining for the 9 immune checkpoints consistent with 13 features tested are shown in Additional file 1: Figure S1. Based on computational pathology analysis, the expression of the immune checkpoints was digitally measured and quantified as the positive expression percentages of TCs and TAICs. Using four cutoff values (> 1, > 5, > 25, and >  50%), which have been frequently used in reports on the expression of PD-L1, we determined the distribution of NPC patients expressing the immune checkpoints in the training cohort. In addition, the median percentages of all immune checkpoints were also determined. With a median percentage greater than 10%, high positive expression of PD-L1 and B7-H4 was observed in TCs, whereas all immune checkpoints except for LAG3 and TIM3 were highly expressed in TAICs (Table 1). In addition, we analysed the co-expression status of four immune checkpoints in TCs and found that PD-L1, B7-H4, and IDO-1 expression was the most common combination of simultaneously expressed markers, as this pattern was seen in 16% of the NPC specimens (Additional file 2: Table S2).Tab1

Expression levels of 13 features regarding 9 immune checkpoint markers in nasopharyngeal carcinoma

Furthermore, we explored the prognostic value of the 13 immune checkpoint features in the training cohort. As shown in Fig. 2, eight of the features were significantly associated with patient survival. The patients with high expression of PD-L1 in either their TCs (HR 0.38, 95% confidence interval (CI) 0.20–0.74, P = 0.004) or TAICs (HR 0.47, 95% CI 0.25–0.90, P = 0.023) had better OS than the patients with low expression of PD-L1. Similar results were observed for IDO-1 expression in both the TCs (HR 0.45, 95% CI 0.24–0.85, P = 0.014) and TAICs (HR 0.43, 95% CI 0.23–0.81, P = 0.01). In addition, high expression of LAG3 (HR 0.34, 95% CI 0.16–0.74, P = 0.006), VISTA (HR 0.38, 95% CI 0.19–0.73, P = 0.004), or ICOS (HR 0.41, 95% CI 0.22–0.77, P = 0.006) in the TAICs was associated with better OS than low expression, while high expression of B7-H3 in the TAICs (HR 2.13, 95% CI 1.12–4.03, P = 0.021) was associated with poorer OS than low expression (Fig. 2). The associations between the 13 immune checkpoint features and DFS or DMFS are listed in Additional file 1: Figure S2 and Figure S3.Fig2Fig. 2

Kaplan-Meier curves for overall survival according to the 13 immune checkpoint features. Plots show (a) PD-L1_TC; (b) PD-L1_TAIC; (c) B7-H3_TC; (d) B7-H3_TAIC; (e) B7-H4_TC; (f) B7-H4_TAIC; (g) IDO-1_TC; (h) IDO-1_TAIC; (i) LAG3_TAIC; (j) VISTA_TAIC; (k) TIM-3_TAIC; (l) ICOS_TAIC and (m) OX40_TAIC in the training cohort. Abbreviations: TC, tumour cell; TAIC, tumour-associated immune cell; HR, hazard ratio; and CI, confidence interval

To construct an immune checkpoint-based prognostic model, we identified 5 immune checkpoint features that were significantly associated with OS in the training cohort using penalized LASSO Cox regression models (Additional file 1: Figure S4). Then, a risk score was calculated for each patient using a formula that included 5 features weighted by their regression coefficient: Risk score = (0.6995× positive percentage of B7-H3_TAIC) - (0.0054× positive percentage of IDO-1_TAIC) - (0.4039× positive percentage of VISTA_TAIC) - (1.6908× positive percentage of ICOS_TAIC) - (0.0710× positive percentage of LAG3_TAIC).

After using X-tile plots to generate the optimal cutoff value (− 0.16) for the risk score (Additional file 1: Figure S5), we assigned 159 patients in the training cohort into the low-risk group and 49 patients into the high-risk group. The high-risk group had a shorter 5-year OS rate than the low-risk group (61.2% vs. 88.1%, respectively, HR 3.75, 95% CI 1.98–7.09, P < 0.001). The patients with high-risk scores also had shorter DFS (HR 2.51, 95% CI 1.40–4.50, P = 0.002) and DMFS (HR 2.93, 95% CI 1.41–6.09, P = 0.004) than the patients with low-risk scores (Fig. 3a-c).Fig3Fig. 3

Kaplan-Meier curves for overall, disease-free and distant metastasis-free survival according to the ICS. Plots show (a) overall survival, (b) disease-free survival and (c) distant metastasis-free survival in the training cohort and (d) overall survival, (e) disease-free survival and (f) distant metastasis-free survival in the validation cohort. Abbreviations: ICS, immune checkpoint-based signature; HR, hazard ratio; and CI, confidence interval

To validate whether the ICS has similar prognostic value in different populations, we tested the 5 immune checkpoint features in a validation cohort of 125 NPC patients and then used the formula and cutoff point developed from the training cohort to stratify the patients into low-risk (n = 70) and high-risk (n = 55) groups. The patients with high-risk scores had shorter OS (56.4% vs. 81.4%, respectively, HR 2.58, 95% CI 1.31–5.07, P = 0.006), DFS (HR 2.39, 95% CI 1.32–4.30, P = 0.004) and DMFS (HR 2.55, 95% CI 1.13–5.73, P = 0.024; Fig. 3d-f) than those with low-risk scores. The 5-year OS, DFS, and DMFS rates in each ICS group and the number of patients who had an event in each risk group are listed in Additional file 2: Table S3 and Table S4, respectively.

We performed univariate analyses with the training and validation cohorts, and Additional file 2: Table S5, Table S6 and Table S7 shows the associations among the ICS, clinicopathological characteristics and patient clinical outcomes. The ICS was significantly associated with OS, DFS, and DMFS in the two cohorts. Multivariate Cox regression analysis showed that the ICS remained a powerful and independent prognostic factor for OS, DFS, and DMFS in the training cohort (OS: HR 3.62, 95% CI 1.91–6.87, P < 0.001; DFS: HR 2.43, 95% CI 1.35–4.35, P = 0.003; and DMFS: HR 2.77, 95% CI 1.33–5.77, P = 0.007) as well as in the validation cohort (OS: HR 2.59, 95% CI 1.32–5.10, P = 0.006; DFS: HR 2.38, 95% CI 1.32–4.30, P = 0.004; and DMFS: HR 2.55, 95% CI 1.13–5.72, P = 0.024). In addition, the TNM stage and EBV-DNA levels were also significantly associated with OS, DFS and DMFS in the multivariate analysis (Additional file 2: Table S8).

TNM stage is the determinant for predicting prognosis and guiding treatment currently, but its accuracy is limited as it based on anatomical information and it need to be supplement with molecular indicators.

To develop a more sensitive model to predict prognosis of NPC patients, we established a prognostic score model combining the ICS and TNM stage based on the multivariate Cox regression analysis. The regression coefficient of the ICS was divided by the regression coefficient of the TNM stage, and then rounded into an integer value to generate the risk score (Additional file 2: Table S9). We calculated each patient a cumulative risk score, and used ROC analysis to compare the sensitivity and specificity of the prognostic score model with the TNM stage or ICS alone model. Combination of the ICS and TNM stage showed significantly better prognostic value than the TNM stage alone for OS (area under ROC (AUROC) 0.73 [95% CI 0.64–0.82] vs 0.63 [0.55–0.72]; P = 0.003), DFS (0.68 [95% CI 0.59–0.77] vs 0.62 [0.54–0.70]; P = 0.039), and DMFS (0.69 [95% CI 0.58–0.80] vs 0.62 [0.52–0.71]; P = 0.049) in the training cohort, which were confirmed in the validation cohort (OS, 0.72 [95% CI 0.62–0.82] vs 0.62 [0.52–0.72]; P = 0.012; DFS, 0.72 [95% CI 0.62–0.81] vs 0.62 [0.52–0.72]; P = 0.016; DMFS, 0.69 [95% CI 0.58–0.81] vs 0.60 [0.49–0.71]; P = 0.035) (Fig. 4).Fig4Fig. 4

Comparisons of the sensitivity and specificity for the prediction of overall, disease-free and distant metastasis-free survival by the combined ICS and TNM stage model, the TNM stage alone model, and the ICS alone model. Receiver operating characteristics (ROC) curves of (a) overall survival, (b) disease-free survival and (c) distant metastasis-free survival in the training cohort and (d) overall survival, (e) disease-free survival and (f) distant metastasis-free survival in the validation cohort. P values show the area under the ROC (AUROC) of the combined ICS and TNM stage model versus AUROCs of the TNM stage alone model or the ICS alone model

NPC is closely associated with EBV infection, which has been reported to be involved in the regulation of immune-inhibitory biomolecules [27]. We analysed whether the EBV-DNA burden could affect the predictive efficacy of the ICS in 208 NPC patients from the Guangzhou training cohort. After the patients were divided into different subgroups by their pretreatment plasma EBV-DNA level, Kaplan-Meier curves showed that stratification by the ICS resulted in significant differences in OS (HR 4.82, 95% CI 2.22–10.47, P < 0.001), DFS (HR 3.07, 95% CI 1.52–6.19, P = 0.002), and DMFS (HR 4.66, 95% CI 1.92–11.29, P = 0.001) in the patients with an EBV-DNA level > 2000 copy/mL (Fig. 5a-c). However, in the patients with an EBV-DNA level ≤ 2000 copy/mL, we did not find a significant association between the ICS and any of the outcomes (Fig. 5d-f). The 5-year OS, DFS, and DMFS rates in each risk group and the number of patients who had an event in each risk group among the different EBV-DNA burden groups are listed in Additional file 2: Table S3 and Table S4.Fig5Fig. 5

Kaplan-Meier curves for overall, disease-free and distant metastasis-free survival of patients grouped by their EBV-DNA level and then stratified according to the ICS. Plots show (a) overall survival, (b) disease-free survival and (c) distant metastasis-free survival in the EBV-DNA level > 2000 copy/mL subgroup and (d) overall survival, (e) disease-free survival and (f) distant metastasis-free survival in the EBV-DNA level ≤ 2000 copy/mL subgroup. Abbreviations: ICS, immune checkpoint signature; HR, hazard ratio; and CI, confidence interval