All the biopsy samples of recurrent tumor (n = 132) were collected. The samples were fixed by formalin and embedded into paraffin using a tissue processor. Among the 132 patients enrolled, there were 106 males and 26 females. The median time to recurrence after the initial treatment was 33.3 months (ranges from 6.6 to 190.2 months). The median age at recurrence was 46-year-old (ranges from 28 to 69 years). According to the 7^th edition of the UICC / AJCC system, 101 patients (76.5 %) were at III-VI stage and 32 patients (24.2 %) were lymph node positive simultaneously. Pathology proven recurrent undifferentiated non-keratinizing carcinoma (World Health Organization [WHO] III) and differentiated non-keratinizing carcinoma (WHO II) were diagnosed in 125 and 7 patients, respectively. The median HB level before salvage therapy was 139.0 g/l (ranges from 101.0 to 171.0 g/l) for all patients, among which 32 patients were anemic (HB < 130 g/l in men; < 120 g/l in women). Tumor necrosis before retreatment presented in 49 out of 132 (37.1 %) patients by pathologic diagnosis. Representative tumor necrosis with PD-L1 staining in recurrent NPC were shown in Figure 1. The average of follow-up time was 38.6 months (ranges from 3.5 to 182.1 months).

A total of 87 out of 132 (65.9 %) patients died during follow-ups. Of these, 35 out of 87 (40.2 %) died due to severe adverse effects, including 16 out of 35 (45.7 %) from mucosa necrosis or massive hemorrhage, 19 out of 35 (54.3 %) from other radiation-related injuries. In addition, 29 out of 87 (33.3 %) patients died due to local-regional failures and 16 out of 87 (18.4 %) due to distant failures. Other causes responsible for 2 out of 87 (2.3 %) deaths included 1 cases of cardiac disease, 1 case of digestive diseases. 5 out of 87 (5.7 %) patients died due to unknown reasons. The 3-year OS rates were 56.1 % in the 132 recurrent patients. Characteristics of the enrolled patients were summarized in Table 1.

PD-1 positive immune cells presented in 50 of 132 patients (37.9%) in a scattered manner, and PD-L1 staining was detectable in 128 tumors (97.0%), which was mainly located at the membrane or in the cytoplasm region (or both) in the tumor cells. PD-L1 was also observed to be stained in tumor-infiltration lymphocytes (TILs) dispersedly. According to ROC curve analysis for OS, the optimal cut-off value of H-score was 190 (AUC: 0.702, sensitivity: 0.759, specificity: 0.511) for PD-L1. The staining of PD-L1 were considered as high level (88 cases) if H-score ≥ 190. Representative staining of PD-L1 and PD-1 in recurrent NPC were shown in Figure 1–2. Regarding the mortality of patients with high level of PD-L1 staining, there were 66 patients with high level of PD-L1 staining died during the follow-ups. Of these, 38 out of 66 patients died due to disease progression with 26 cases from local-regional failures and 12 cases from distant failures, another 28 out of 66 deaths were due to severe adverse effects, the cause of deaths were found to be significantly correlated with PD-L1 level (P = 0.024, Table 1). In this study, high level of PD-L1 staining was significantly correlated with clinical stage, rT classification of recurrent tumor and anemia at diagnose of recurrence. However, no correlation of high PD-L1 level with tumor necrosis, gender or time to recurrence was observed. No clinic-pathological parameters were found related to PD-1 positivity. Detailed data were summarized in Table 1.

Pre-retreatment hematologic biomarkers such as neutrophils, lymphocytes, HB and platelets were measured to identify the potential factors related to PD-1 / PD-L1. Our result showed that no significance (N. S.) between PD-1 positivity and HB level was observed in this cohort of recurrent patients (Table 1, Figure 3A). However, anemia status at diagnoses of recurrence was negatively correlated with PD-L1 staining (P = 0.044, Table 1). Consistently, HB level was significantly reduced in patients with high level of PD-L1 staining (P = 0.029, Figure 3B).

We have also compared PD-1 / PD-L1 expression with numbers of neutrophils, lymphocytes, platelets and neutrophils / lymphocytes ratio before salvage therapy in this exploratory analysis, and no significantly relevant factors were identified (Supplementary Figure 1).

Hypoxia-inducible factors (HIFs) have been demonstrated to play an important role in the regulation of erythropoiesis by coordinating a series of graded hypoxic responses [27] and stabilization of HIFs is represented as a novel therapeutic approach for the treatment of anemia [28]. Considering the closely association of HIF with anemia, immunofluorescence staining were performed to investigate the correlation of PD-L1 with anemia (represented as HIF-1α activity) using 30 recurrent tumors. PD-L1 staining was detectable in 28 cases of recurrent tumors with different extent, which was mainly distributed in the cytoplasm region of the tumor cells. HIF-1α staining was detectable in all 30 tumors, and was found to be co-located with PD-L1 staining regardless of the staining intensity or extent. Representative staining of PD-L1 and HIF-1α in recurrent NPC by immunofluorescence was shown in Figure 4A-4L. Pearson correlation coefficient of merged staining with HIF-1α and PD-L1 in Figure 4C, Figure 4F, Figure 4I, and Figure 4L were 0.94, 0.93, 0.76 and 0.93, respectively.

To evaluate the prognostic values PD-L1 expression in recurrent NPC patients, we performed long rank tests in Kaplan-Meier survival analysis. In 132 patients with recurrent NPC, high expression of PD-L1 was correlated with significantly shorter OS (P = 0.001, Figure 5A). Factors significantly correlated with OS by univariate analyses were age at recurrence (P = 0.043, Figure 5B), the level of PD-L1 staining (P = 0.001, Figure 5A), T-stage of recurrence (P ≤ 0.001, Figure 5C), pre-retreatment anemia (P = 0.007, Figure 5D) and tumor necrosis (P = 0.026, Figure 5E). However, PD-1 positivity on TILs was not significantly correlated with OS (P = 0.950, Table 2). Further analysis by dividing patients into 4 groups regarding PD-L1 level and tumor necrosis showed that patients with both tumor necrosis and high level of PD-L1 staining in recurrent NPC had the worst survival outcomes compared with patients in other groups (P = 0.001, Figure 5F). In the multivariate analysis, significantly negative prognostic factors were identified, including age > 50 years, high level of PD-L1 staining, recurrent T3-4 stage and tumor necrosis before salvage treatment (Table 3). Moreover, patients with high level of PD-L1 might have higher risk of death (95 % CI, 0.315 - 0.888, P = 0.016, Table 3) in this cohort of 132 recurrent NPC patients.

All data collection and statistical analysis of this study were conducted in accordance with the Institutional Review Board of Sun Yat-Sen University Cancer Center. Informed consent was obtained from all subjects before their salvage treatment. Individual informed consent was waived based on the Ethical Board requirement for retrospective clinical studies. Patient records / information was anonymized and de-identified prior to analysis.

A total of 179 patients with locally recurrent NPC from Sun Yat-Sen University Cancer Centre between January 2001 and March 2013 were enrolled. The cases were selected based on the following criteria: (1) local recurrence was histologically confirmed with available biopsy specimens; (2) complete clinical data; (3) non-distant metastatic simultaneously; (4) the main treatment of the initial course was radiotherapy; (5) no other malignant diseases; (6) Karnofsky score ≥ 70; (7) received salvage treatment at Sun Yat-Sen University; (8) follow-up regularly. Patients with active hepatitis, diabetes, or who failed the follow-up requirements were excluded. Therefore, we obtained the remaining 132 cases qualified for this study. All patients with recurrent NPC were re-staged according to the 7^th edition of the UICC /AJCC system. Formalin-fixed and paraffin-embedded (FFPE) blocks from fiberscope biopsy of NPC were retrieved from the Department of Pathology. The related clinical data of these patients were retrospectively collected from our database or hospital records, including gender, age, smoking status, family history, clinical stage, pre-therapy laboratory counts of neutrophils, lymphocytes, HB and platelets as well as follow-up records. Tumor necrosis was defined as the necrosis of tumor tissue at diagnose of recurrence and assessed pathologically by two experienced pathologists independently. Anemia was defined according to World Health Organization criteria as hemoglobin < 130 g/l in men and < 120 g/l in women. All patients were followed-up for at least 3 years.

According to the discretion of attending oncologists with the consideration of disease condition and the preference of patients, most of these recurrent patients (114/132, 86.4 %) were given salvage IMRT, 6 patients were treated with conventional radiotherapy, 2 with stereotactic radiotherapy, 1 with brachytherapy and 3 radiofrequency ablation, salvage endoscopic nasopharyngectomy were performed for 6 patients according to the procedure previously reported by Chen MY et al [45]. The prescribed doses of IMRT to the recurrent tumors were 60-70 Gy in 27-35 fractions according to the institutional protocol delineated by Tian YM et al [41]. Cisplatin-based induction, concurrent or adjuvant chemotherapy with radiotherapy was administered to 78 patients with rT3-4 and / or bulky gross tumors. The groups included 29 patients with concurrent chemo-radiotherapy, 32 patients with induction chemotherapy followed by radiotherapy, 13 patients with induction and concurrent chemotherapy, 4 patients with adjuvant chemotherapy. Besides, 2 out of 6 patients who received salvage surgery and 1 out of 3 patients treated with radiofrequency ablation were administered with induction chemotherapy. 13 patients (9.8 %) received targeted agents (cetuximab, nimotuzumab or bevacizumab) combined with their other treatment schedules.

All patients were required to be followed-up every 3-4 months in the first 3 years, every 6 months for 2 additional years, and annually after their salvage treatment. Each follow-up visit included physical examination, indirect nasopharyngoscopy, MRI of the head-and-neck regions, CT of thoracic region and ultrasound or CT of the abdomen.

The samples were fixed by 4 % paraformaldehyde containing 2 % sucrose in PBS at 4 °C for overnight and embedded into paraffin using a tissue processor (EG1150, Leica, Germany). FFPE sections (3 μm thick) were cut with a rotation microtome (RM2255, Leica, Germany).

Paraffin sections were dewaxed and rehydrated by alcohol series, and then were treated with 3% H₂O₂ for 10 min at room temperature. The sections were steam heated for 2.5 min in ethylene diamine tetra acetic acid buffer (PH = 8.0) to retrieve antigen. Subsequent staining was performed with a 50 min incubation period at 37°C with the monoclonal antibodies PD-1 and PD-L1 (Supplementary Table 1). The immunostaining of PD-1 and PD-L1 proteins was performed on two different slides. Tonsils tissue was used as a positive control. Immunoreaction was visualized by a Peroxidase / DAB kit (Cat. K5007, Dako, Denmark). Images were taken with a phase contrast microscope (Eclipse 80i, Nikon, Japan).

Percentages of PD-L1 positive tumor cells and PD-1 positive TILs and staining intensity were evaluated by two experienced pathologists independently who remained blind to clinical information by counting 20 sequential high-power fields (0.54 mm field diameter). Staining intensity was scored by the criteria: 0 as negative or trace, 1 as weak, 2 as moderate and 3 as high. The percentage of positive cells (0 - 100 %) was multiplied by the dominant intensity score of staining ranging from 0 to 3. Therefore, the overall semi-quantitative score, that is H-score, was ranged from 0 to 300 (maximum value of 300 corresponding to 100 % of tumor cells for PD-L1 or TILs for PD-1 staining with an overall staining intensity score of 3).

Paraffin sections were rehydrated and steam heated to retrieve antigen under the same conditions as in IHC procedure, and blocked in normal goat serum for 30 min at room temperature. The sections were incubated with primary antibodies against PD-L1 and HIF-1α (Supplementary Table 1), followed by addition of Alexa-Fluor -555 or Alexa-Fluor-488 labeled secondary antibodies (Supplementary Table 2) respectively in humidified atmosphere. Nuclei were labeled with DAPI in a concentration of 2 μg / ml (Supplementary Table 2). Sections were analyzed by a fluorescence microscope (Olympus, Cat. #BX53). Immunofluorescence analysis was performed by two experienced pathologists independently. Pearson's correlation coefficient was performed by Image-Pro plus 6.0 to evaluate the co-location of PD-L1 with HIF-1α. 0 indicates no significant correlation, -1.0 indicates complete negative correlation, and 1.0 indicates complete positive correlation.

Optimal cut-off point for PD-L1 H-score was determined by the area under the curve (AUC) of the receiver operating characteristic (ROC) curve at the highest positive likelihood ratio point for OS. PD-L1 expression was dichotomized into two groups (high and low), using a cut-off score of ≥ 190. OS was measured from the diagnosis of local recurrence until confirmed death or the date of last follow-up through April 2016.

These statistical analyses were performed with SPSS 16.0. Measurement data were presented as mean ± standard error (S.E.M) Chi square test was used to assess the level of PD-1 and PD-L1 correlated with various clinical parameters such as age, clinical stage, anemia and tumor necrosis before re-therapy. T-test was used to evaluate the association of PD1 / PD-L1 with HB, neutrophil counts and platelet after normality test (Kolmogorov-Smirnov test). Nonparametric Mann-Whitney U test was applied to evaluate the association of PD-1 / PD-L1 with lymphocyte count and neutrophil / lymphocyte ratio. Survival analysis was depicted by Kaplan-Meier method. Univariate analysis and multivariate analysis were performed with log-rank test and Cox regression analysis, respectively. P value < 0.05 was used to denote statistical significant, and all reported P values were two sided. P values were marked as * P < 0.05, ** P <0.01, *** P < 0.001, indicating different level of significance.