The 99 patients included in this study were predominantly male (83%), and represented primarily stage III/IV tumors (83%). Most of the tumors originated from the tonsil (43%) and were moderately (62%) differentiated. The overall HPV prevalence rate was 19% (19/99) with no difference in distribution along all topographic sites (P = 0.5673). More HPV⁺ specimens were detected among poorly differentiated and basaloid tumors (11/32) compared to well/moderately differentiated tumors (8/67; P = 0.0265).

Patients clinicopathological variables are displayed in Table 1.

A positive link was found between HPV status and PD-L1 positivity on TCs for the SP142 clone (P < 0.0001) and 22C3 clone (P = 0.0649). Next, positive SP142 PD-L1 expression was associated with poorly differentiated/basaloid tumors (15/29 versus 7/65; P = 0.0001) whereas positive 22C3 PD-L1 expression on TCs was most common in elder patients (8/12 versus 25/85; P = 0.0234) and in poorly differentiated/basaloid tumors (17/31 versus 16/66; P = 0.0111).

All associations between PD-L1 on TCs and clinicopathological variables are presented in Table 1.

Notably, PD-L1 positivity on TCs was linked with the immune environment, as this parameter was associated with increased TIL count (P < 0.0001 and P = 0.0009), CD3⁺ T cell count (P = 0.0009 and P = 0.0044), CD8⁺ T cell count (P < 0.0001 and P = 0.0001) and FoxP3⁺ T cell count (P = 0.0004 and P = 0.0094) for the SP142 and 22C3 clone, respectively.

All associations with TILs are given in Table 2.

Positive PD-L1 expression using the SP142 clone resulted in a more favorable outcome with median overall survival (OS) of 9.8 years versus 4.1 years for SP142 PD-L1 negativity (hazard ratio [HR] = 0.51 [0.31–0.99], P = 0.0466; Figure 1A) whereas 22C3 PD-L1 expression on TCs showed no significant association with OS (4.0 years versus 4.4 years; HR = 0.99 [0.58–1.72], P = 0.9846; Figure 1B). None of the PD-L1 markers was associated with prolonged disease-free survival (DFS) (HR SP142 = 0.43 [0.22–1.16], P = 0.1070 and HR 22C3 = 1.28 [0.60–2.83], P = 0.5110; Figure 1C–1D). An overview of the HRs on DFS for all tumor markers, IC markers and associated clinicopathological markers is given in Supplementary Table 1.

The relation between TILs and outcome was evaluated as well. High TIL count, high CD3⁺ T cell count and high CD8⁺ T cell count but not FoxP3⁺ T cell count proved to have a favorable influence on OS (Supplementary Figure 1). Particularly, an elevated CD8⁺ T cell count was significantly associated with prolonged OS (9.8 years) versus patients with a low CD8⁺ T cell count (3.5 years; HR = 0.32 [0.22–0.71], P = 0.0020). As CD8⁺ T cell count proved the strongest marker of the immune cell infiltrate, an analysis was performed combining CD8⁺ T cell count and PD-L1 expression on TCs. The combined analysis of SP142 PD-L1 expression on TCs and CD8⁺ T cell count showed that this combination was significantly correlated with OS: median OS of 9.8 years, not reached, 2.4 years and 4.0 years for the PD-L1⁺/high CD8⁺, PD-L1⁻/high CD8⁺, PD-L1⁺/low CD8⁺, and PD-L1⁻/low CD8⁺ cohorts, respectively (P = 0.0321; Figure 2A). Similar, a significance was found for combining 22C3 PD-L1 expression on TCs and CD8⁺ T cell count with median OS of 9.8 years, not reached, 2.1 years and 4.1 years for the respective cohorts (P = 0.0118; Figure 2B). As the majority of patients were stratified as PD-L1⁻/low CD8⁺, patients were bicategorically categorized as ‘at least one variable positive’ versus PD-L1⁻/low CD8⁺. Patients with ‘at least one variable positive’ had a more favorable outcome for the SP142 PD-L1 expression (median OS 9.8 years versus 4.0 years; HR = 0.40 [0.24–0.78], P = 0.0048; Figure 2C) but not for the 22C3 PD-L1 expression (median OS 5.5 years versus 4.1 years; HR = 0.66 [0.38–1.15], P = 0.1415; Figure 2D).

In the multivariate model, which adjusted for HPV status, CD8⁺ T cell count proved to be the independent prognosticator for OS (HR = 0.31 [0.14–0.70]; P = 0.0050) whereas SP142 PD-L1 positivity and other immune cell counts did not contribute to the model. When assessing the combination of SP142 PD-L1 expression and CD8⁺ T cell count in the multivariate model, patients with ‘at least one variable positive’ had an OS benefit compared with PD-L1⁻/low CD8⁺ patients (HR = 0.39 [0.20–0.78]; P = 0.0070).

An overview of all multivariate HR is given in Table 3.

No relation was demonstrated between PD-L1 expression (SP142 or 22C3 clone) on ICs and clinicopathological variables, including TILs, overall and disease-free survival.

Ninety-nine patients with histologically proven OSCC (anno 2004-2013) were selected from the archival database of the Department of Pathology, Ghent University Hospital. Patients were excluded from the study in case of recurrent or distant metastatic disease. When available, samples from resection specimens were preferred for further analysis (n = 27), if not, endoscopic biopsy material was used (n = 72).

The study was approved by the ethics committee of the Ghent University Hospital.

Clinical, treatment and follow-up data were retrieved from the electronic patient data file. Tumor sites consisted of tonsil, tongue base, other (e.g. tonsillar pillars, posterior wall, vallecula) or multiple subsites involved. The majority of the patients was treated surgically (n = 51) by either tumor resection with or without lymph node dissection (n = 19 and n = 8, respectively) or lymph node dissection alone (n = 24). Forty-six out of 51 surgically patients received adjuvant treatment, that is, concurrent chemoradiotherapy (n = 30) or radiotherapy alone (n = 16). Clinical data and treatment details are presented in Table 1. For all patients, treatment decisions were discussed in a multidisciplinary meeting consisting of dedicated head and neck surgeons, medical and radiation oncologists, radiologists and pathologists.

HPV status was determined on FFPE material by high-risk HPV in situ hybridization (ISH) using the Inform HPV III Family B probe (Ventana Medical Systems) on a BenchMark XT automated stainer (Ventana Medical Systems) under ISO15189:2012 accreditation. The probe cocktail detects HPV types 16, 18, 31, 33, 35, 39, 45, 51, 52, 56, 58 and 66. The high-risk HPV ISH test was considered positive if a discrete, blue colored, precipitated reaction product within the TCs was observed.

Immunohistochemistry for PD-L1 and TILs subtyping was performed using the BenchMark XT instrument (Ventana Medical Systems inc., Arizona, USA). Staining of CD3, CD8 and FoxP3 (DAKO, Glostrup, Denmark) was performed on FFPE tissue slides of 2μm as previously described [29]. For staining of PD-L1, similar antigen retrieval and blocking steps were followed by incubation with the anti-PD-L1 primary Ab (clone 22C3 diluted 1:100 [Agilent-DAKO, United States] and clone SP142 ready-to-use [Roche, Basel, Switzerland]) and DAB detection using the Ultraview kit (Ventana) and the optiview kit (Ventana), respectively, according to the instructions of the manufacturer. Slides were counterstained with hematoxylin.

First, a hematoxylin-eosin slide was evaluated to confirm the presence of invasive squamous cell carcinoma, to determine the grade of tumor differentiation and/or define basaloid subtype. The infiltration of TILs in tumor stroma was scored semi-quantitatively with a 4-tiered scale (1+ to 4+) as previously described [29, 30]. Representative scoring for stromal TILs is given in Supplementary Figure 2.

Elaborate evaluation of the immune infiltrate in the stromal compartment (i.e. lymphocytes within the intertumoral stroma), was done semi-quantitatively by scoring for CD3⁺, CD8⁺ and FoxP3⁺ T cell count. For the subtyping of the immune infiltrate, three representative fields were selected and scored using a similar 4-tiered semi-quantitative method as for TIL scoring (1+ to 4+) [29, 30] (Supplementary Figure 2). The mean of the separate scores from each of the 3 fields was calculated as final score, ranging from 1+ to 4+. Necrotic or ulcerated areas were excluded from evaluation. Criteria to classify samples as ‘impossible to evaluate (ITE)’ were the following: 1) (virtually) no tumor stroma present, 2) presence of only a minimal invasive carcinoma component and 3) the inability to differentiate between the ICs of the pre-existing lymphoid tissue of the tonsil and the tumor-associated IC infiltrate.

Subsequently, samples were scored for PD-L1 expression with both the SP142 clone and the 22C3 clone by two independent observers. IHC was scored 0 if < 1% of TCs were positive, IHC 1 if ≥ 1% but < 5% of TCs were positive, IHC 2 if ≥ 5% but < 10% of TCs were positive or IHC 3 if ≥ 10% of TCs were positive (Figure 3) [31]. Mean scores between both observers were used in the analyses. Cut-off values for PD-L1 positivity on TCs were taken at ≥ 5% (membranous and/or cytoplasmic) for the SP142 and 22C3 clone. PD-L1 expression on ICs was evaluated using Cell^{^}D imaging software (Olympus Corporation, Tokyo, Japan) to ensure objective scoring as the expression of PD-L1 on ICs was low. PD-L1 staining was quantified in the three same representative fields of intertumoral stroma used for scoring of the T cell infiltrate. Scores were given as percentages reflecting the ratio of PD-L1 staining surface area to the whole surface area of the representative field. The final score was determined as the mean of the separate scores from the chosen fields. Again, cut-off values for PD-L1 positivity on ICs were taken at ≥ 5%.

Sample size calculation for a proportional z test of the most fitting immune cell marker (CD8⁺ T cells) has shown that 106 patients are needed to achieve a power of 90% (α = 5%; proportion event group high CD8⁺ = 0.35; proportion event low CD8⁺ = 0.68; sample ratio high CD8⁺/ low CD8⁺ = 0.29). Associations between categorical variables were determined via Chi-square test. In case of 2 × 2 contingency tables with 1 or more observed values lower than 10, a Fisher's Exact test was applied. Secondly, the HR of immunohistochemical parameters on OS and DFS was determined by a log-rank (Mantel–Cox) test. OS was calculated from time from diagnosis on biopsy until day of death or final follow-up whereas DFS was calculated from time of therapy until time of recurrence. For every patient who was not seen in clinic for the preceding 12 months at the point of the survival analysis, the general practitioner was contacted to confirm whether the patient was still alive. Patients that were lost to follow up at either the Ghent University Hospital or the general practitioner were censored in the survival analysis (n = 8). Finally, the covariate effect of the risk factors on survival analysis, that reached a P value lower than 0.1 on univariate log-rank test, was determined by means of the Cox proportional hazard model (backward method). Samples that were classified as ‘ITE’ were not included in the analysis.