This study included patients with R/M HNSCC who received nivolumab treatment across 22 facilities from March 1, 2017, to May 31, 2018. Out of a total of 613 cases, clinicopathological information and formalin-fixed, paraffin-embedded (FFPE) tissue sections were obtained from 53 responders and 47 non-responders as a case-control study ( Supplementary Figures 1A, B ). Patients who achieved complete or partial response according to RECIST version 1.1 were defined as responders, and patients with stable disease or progressive disease as their best overall response were defined as non-responders. Nivolumab was administered every two weeks until disease progression, occurrence of unacceptable adverse events, or death. Patients who had not been previously treated with platinum-containing chemotherapy, used nivolumab as adjuvant chemotherapy, or received nivolumab in combination with other antineoplastic agents were excluded. Age, sex, primary tumor site, progression, differentiation, smoking and alcohol consumption history, serum albumin level, C-reactive protein (CRP) level, lymphocyte count, neutrophil count, best overall response, immune-related adverse events, cycles of nivolumab use, prior cetuximab exposure history, and Eastern Cooperative Oncology Group (ECOG) performance status (PS) were obtained. The cut-off values of cell counts and ratios were determined according to previous reports (13, 14). Smoking history was evaluated using ta smoking index (pack/year) calculated as a number of packs smoked daily multiplied by number of years of smoking. Alcohol consumption history was evaluated using an alcohol consumption index calculated by multiplying the amount of pure alcohol/27 g per day by the number of years of drinking. Human papilloma virus (HPV) status was determined by institutional pathologists using p16 immunohistochemistry (IHC) for oropharyngeal cancers.

Multiplex IHC was performed with 4 or 5 μm of formalin-fixed, paraffin-embedded (FFPE) tissue sections as previously described (15). Briefly, FFPE tumor sections were subjected to sequential immunodetection using antibodies against immune cell lineages. Following the chromogen development of the antibodies, the slides were scanned digitally at a 20x objective magnification using a NanoZoomer S60 scanner (Hamamatsu Photonics). A complete list of antibodies and the conditions used for staining is provided in Supplementary Table 1 .

Following staining, image acquisition and computational processing were performed as previously described (15). Iteratively digitized images were accurately aligned using in-house software (SCREEN Holdings Co., Ltd.). The software calculated the coordinates of each image relative to the reference image. A series of uncompressed TIFF images of each ROI was retrieved based on these coordinates. Visualization was performed using ImageScope Version 12.3.3.5048 and ImageJ. Co-registered images were converted to single-marker images, inverted, and converted to grayscale, followed by pseudo-coloring. Single-cell segmentation and staining intensity quantification were performed for quantitative image assessment, using CellProfiler Version 2.2.0. All pixel intensity and shape-size measurements were saved in a file format compatible with the image cytometry data analysis software, FCS Express 7 Image Cytometry Version 7.06.0015 (De Novo Software).

Tumor and immune cells expressing PD-L1 were quantified according to the lineage identification markers in Supplementary Table 2 . Immune profiles were identified based on the criteria obtained from our previous study on HNSCC: hypo-inflamed group with less than 1500 immune cells per mm², lymphoid-inflamed group with lymphoid immune cell to myeloid immune cell ratio of 2:1 or more, and myeloid-inflamed group with less than a 2:1 ratio (15).

A flowchart of the gene mutation validation process is shown in Supplementary Figures 2A, B . The GeneRead DNA FFPE Kit (Qiagen, Hilden, Germany) was used to extract gDNA from the FFPE tissues including uracil-DNA glycosylase treatment prior to amplification. After quality control exclusions, 62 cases were included for genomic DNA extraction and sequencing. AmpliSeq for the Illumina Comprehensive Cancer Panel (Illumina, California, USA), a target panel for examining the exon regions of 409 cancer-associated genes, was used to prepare the libraries. The prepared libraries were sequenced using the NextSeq 500/550 High Output Kit v2.5 (300 Cycles) (Illumina) on the next-generation sequencer NextSeq (Illumina), and fastq and bam data were obtained by Local Run Manager (Illumina). Hg19 was used for the alignment.

The obtained fastq data were analyzed for gene mutations using the supercomputer, SHIROKANE at the Institute of Medical Science, University of Tokyo, and the analysis pipeline, Genomon 2.6.3 (https://genomon-project.github.io/GenomonPagesR/). Germline mutations were removed on allele frequencies in public databases, i.e., the ToMMo 8.3KJPN-SNV/INDEL Allele Frequency Panel (jmorp.megabank.tohoku.ac.jp), the Human Genetic Variation Database (www.genome.med.kyoto-u.ac.jp/SnpDB/), the 1000 Genomes Project database (www.1000genomes.org), and the NHLBI GO Exome Sequencing Project (esp.gs.washington.edu). Additional review using Integrative Genomics Viewer (IGV, version 2.12.3, Broad Institute, Cambridge, MA, USA) identified and excluded low-mappability regions and sequencing errors. The VAF cutoff was 0.04 according to the previous report (16). Copy number analyses were performed using the Number Targeted Resequencing Analysis (CONTRA, sourceforge.net/projects/contra-cnv/) (17). TMB ≥ 10 was defined as TMB-high (TMB-H), and TMB < 10 as TMB-low (TMB-L).

Chi-square, and Kruskal-Wallis tests were used to compare the frequencies of categorical and continuous variables, respectively. Multivariate analysis was performed using multiple logistic regression analysis with JMP14.2.0 (SAS Institute Inc, Cary, NC, USA). Unsupervised hierarchical clustering was performed using Ward’s minimum variance method (hclust from R) in the R v. 4.3.0 software (http://www.r-project.org). Statistical calculations were performed using JMP14.2.0 (SAS Institute Inc., Cary, NC, USA) and GraphPad Prism 8.3.0. P values < 0.05 were considered statistically significant.

Of the patients studied, 53 were responders (CR/PR) and 47 were non-responders (SD/PD). We examined sex, age, performance status, HPV status, smoking history, primary site, stage at diagnosis, prior cetuximab use, CRP level, lymphocyte count, and neutrophil-to-lymphocyte ratio (NLR) ( Figures 1A–D ). Univariate analysis showed that female sex, non-smoking, p16 positive, low CRP (0.6 ≤), and low NLR (< 7) were significantly associated with a nivolumab response ( Table 1 ). Low smoking index was significantly associated with a therapeutic response, but not with HPV status (22.8 in HPV-positive versus 33.8 in HPV-negative cases on average smoking indexes) ( Figure 1A ). A similar trend was observed with indexed alcohol consumption ( Figure 1B ). HPV positivity was observed in 9% of all cases and significantly correlated with the nivolumab response ( Figure 1C ). Multivariate analysis revealed that smoking history were an independent predictor where non-smokers were more likely to be responders ( Table 1 ). Among the post-treatment clinical factors, the presence of immune-related adverse events was associated with a treatment response ( Figure 1D ). Immune-related adverse events of grade 3 or higher occurred in 5 cases in the responder group, which included arthritis (G3), anemia (G3), adrenal insufficiency (G3), and two cases of hepatobiliary disorders (G3). In the non-responder group, there were 4 cases with adverse events of grade 3 or higher, which included two cases of enterocolitis (G3), one case of enterocolitis (G4), and pneumonitis (G3). A treatment response was not associated with the primary sites ( Supplementary Figure 1B ). Together, these observations indicate that the clinical factors associated with a nivolumab treatment response were the absence of smoking history, and the presence of immune-related adverse events, requiring further investigation of the biological background.

14-marker multiplex immunohistochemistry and image cytometry analysis were performed on FFPE tissue sections, and each single cell was compartmentalized on the image for flow cytometry-like quantification as previously described (15). Immune cell densities and compositions with PD-L1 expression were evaluated quantitatively ( Figure 2A and Supplementary Figure 3A ). Five of the 100 slides (#31, #33, #54, #76, and #96) were excluded due to glass slide incompatibility. In the analysis of the tumor proportion score (TPS), the percentage of tumor cells expressing PD-L1, TPS in the responder group were significantly higher than those of the non-responder group ( Figure 2B and Supplementary Figure 3B ). This trend was further pronounced in the combined positive score (CPS), where PD-L1 expression in both tumor and immune cells was considered. A significantly higher CPS was observed in the responder group than in the non-responder group ( Figure 2C and Supplementary Figure 3C ). The number of responses was particularly low (1 out of 7 cases, 14.4%) in patients with a CPS of less than 1, suggesting a negative predictive ability of negative CPS.

The compositions and cell densities of CD45⁺ immune cell lineages were comparatively analyzed in relation to the therapeutic response and clinicopathological factors ( Figures 2A, D , Supplementary Figure 3A ). The intratumoral immune cell frequency in responders tended to be higher than that in non-responders, where the percentage of immune cells in the total cells was significantly higher in responders than in non-responders ( Supplementary Figure 3D ). In particular, CD8⁺ T cells and macrophages had significantly higher cell densities in responders ( Figure 2D ). No correlation was observed between individual immune cell frequencies and smoking history or the NLR.

Clustering analysis based on immune cell densities revealed that clusters with low immune cell densities were associated with non-responders, whereas clusters with high immune cell densities tended to have favorable responses ( Figure 2A ). Next, we stratified the immune characteristics into three groups of hypo-, lymphoid-, and myeloid-inflamed groups based on criteria from our previous study, where the balance of antitumoral lymphoid versus immunosuppressive myeloid cells was related to prognosis in HNSCC (see Methods) (15). Both the lymphoid- and myeloid-inflamed groups showed a significantly higher frequency of responses than the hypo-inflamed group ( Figure 2E and Supplementary Figure 3E ). The CPS was significantly lower in the hypo-inflamed group, and positively correlated with CD8⁺ T cell density ( Figure 2F ), suggesting an association between high CPS and a favorable response. CD66b⁺ granulocyte/lymphocyte ratios in tumor tissues showed no significant correlation with NLR in the circulating blood or with other standard blood parameters, indicating the significance of analyzing the characteristics of the tissue, not only in the peripheral blood ( Supplementary Figure 3F ).

Next, the association between immune characteristics and clinicopathological factors was explored. Focusing on the history of prior therapy related to a favorable response in multivariate analyses ( Table 1 ), stratified analysis revealed that the frequency of natural killer (NK) cells was associated with a nivolumab response in patients with prior cetuximab use but not those with a cetuximab-naïve status ( Figures 3A, B ). We then stratified the therapeutic response and related immune characteristics according to patient age. A high NLR was significantly associated with a lower response rate in patients aged < 65 ( Figure 3C ) but not in those aged ≥ 65 ( Figure 3D ). In contrast, CPS showed a significant correlation with a nivolumab response in the group aged ≥ 65 but not in the group aged < 65 ( Figures 3E, F ). When we examined the immune cell composition, we found that lymphoid-inflamed immune profiles were highly associated with a nivolumab response only in patients aged ≥ 65 ( Figures 3G, H ), suggesting that nivolumab-related biomarkers may differ by age.

To explore the oncogenic mechanisms related to therapeutic response, genomic DNA was extracted from FFPE specimens and subjected to next-generation sequencing ( Figure 4A and Supplementary Figures 3A, B ). No correlation was observed between TMB and a nivolumab response when stratified by 10 mb as the cut-off for TMB (TMB-high, n = 14; TMB-low, n = 48) ( Figure 4B ). Mutations in mismatch repair (MMR) genes were observed in 9/14 (64.3%) TMB-high cases; however, the presence of MMR gene mutations did not correlate with a nivolumab response (5/8 in non-responders; 4/6 in responders). Among the gene mutations analyzed, TP53 and PIK3CA were frequent, with mutation positivity in 46.8% and 29.0% of cases, respectively ( Figure 4A ). Notably, TP53 mutation was significantly associated with a poor response to nivolumab ( Figure 4B ). There was no correlation between treatment response and TP53 mutation subtypes, including TP53 missense or nonsense mutations ( Figure 4A ). STK11 mutations, known as resistance factors for immunotherapy (18), were observed in 5/62 cases (8.1%), but there was no correlation with responses. Stratification by TMB status did not show that single gene mutations correlated with nivolumab responses ( Supplementary Figures 3A, B ).

To functionally characterize the altered genomic profiles in HNSCC, the set of commonly mutated genes was categorized into representative signaling pathways including the tumor suppressor gene (TSG), PI3 kinase (PI3K) pathway, NOTCH pathway, and receptor tyrosine kinase (RTK) pathway ( Figure 4A and Supplementary Table 3 ). Altered pathways were observed in TSG, NOTCH, PI3K, and RTK at rates of 31.3%, 38.7%, 45.5%, and 50.0%, respectively ( Figure 4B ). Interestingly, we observed that altered TSG pathways were significantly more prevalent in the non-responder group than in the responder group ( Figure 4B ), suggesting that TSG mutations might be associated with resistance to nivolumab in HNSCC.

The potential associations of genomic and immune profiles and their impact on nivolumab responses were analyzed based on 58 cases in which both could be obtained ( Figure 4B and Supplementary Figures 4A, B ). The TMB-high status tended to have higher immune cell densities, TPS and CPS, but the difference was not significant ( Supplementary Figure 5A ). Among the major mutations, the PIK3CA mutation-positive group had significantly higher TPS than the wild-type group ( Supplementary Figure 5B ), but not CPS ( Supplementary Figure 5C ), showing a potential association between PIK3CA mutation and tumor-intrinsic PD-L1 expression. In contrast, the TP53 mutation-positive group had significantly lower CPS and T cell densities than the wild-type group ( Figures 4B–D ), and also tended to have lower immune/tumor cell ratios ( Supplementary Figure 5D ), suggesting the presence of immune-excluded microenvironment characteristics. The patients with altered TSG pathway status, including TP53, CDKN2A, and SMAD4 mutations, had significantly lower CPS and immune/tumor cell rations, and higher smoking index than those without TSG alterations ( Figures 4B, E–G and Supplementary Figure 5E ). These results support our previous observations in which low CPS and high smoking index were associated with poor responses ( Figures 1A , 2C ). Notably, in the group with CPS ≥ 20, altered TSG pathways were highly correlated with non-responders ( Figure 4H ), indicating that immunotherapy candidates with strong PD-L1 expression may benefit from the ability to anticipate treatment resistance by the presence of TSG mutations. These observations suggest that nivolumab treatment effect was influenced by a potential combination of clinical factors, immune environmental characteristics, and gene mutation profiles in HNSCC.